TSINGHUA SCIENCE AND TECHNOLOGY

ISSN 1007-0214 05/31 pp68−86

DOI: 1 0 . 2 6 5 9 9 / T S T . 2 0 2 4 . 9 0 1 0 0 0 4
Volume 30, Number 1, February 2025

Non-Line-of-Sight Multipath Classification Method for BDS Using

Convolutional Sparse Autoencoder with LSTM

Yahang Qin, Zhenni Li*, Shengli Xie, Bo Li, Ming Liu, and Victor Kuzin

Abstract: Multipath signal recognition is crucial to the ability to provide high-precision absolute-position services
by the BeiDou Navigation Satellite System (BDS). However, most existing approaches to this issue involve
supervised machine learning (ML) methods, and it is difficult to move to unsupervised multipath signal
recognition because of the limitations in signal labeling. Inspired by an autoencoder with powerful unsupervised
feature extraction, we propose a new deep learning (DL) model for BDS signal recognition that places a long
short-term memory (LSTM) module in series with a convolutional sparse autoencoder to create a new
autoencoder structure. First, we propose to capture the temporal correlations in long-duration BeiDou satellite
time-series signals by using the LSTM module to mine the temporal change patterns in the time series.
Second, we develop a convolutional sparse autoencoder method that learns a compressed representation of
the input data, which then enables downscaled and unsupervised feature extraction from long-duration BeiDou
satellite series signals. Finally, we add an l1/2 regularizer to the objective function of our DL model to remove
redundant neurons from the neural network while ensuring recognition accuracy. We tested our proposed
approach on a real urban canyon dataset, and the results demonstrated that our algorithm could achieve better
classification performance than two ML-based methods (e.g., 11% better than a support vector machine) and
two existing DL-based methods (e.g., 7.26% better than convolutional neural networks).

Key words: convolutional sparse autoencoder; BeiDou Navigation Satellite System (BDS); long short-term
memory (LSTM); multipath classification

1 Introduction (BDS) is an important element of the Global

Navigation Satellite System (GNSS), which provides
Nowadays, the BeiDou Navigation Satellite System all-weather, all-day, high-precision location services
Yahang Qin and Zhenni Li are with School of Automation, Guangdong University of Technology, and also with Guangdong-Hong
Kong-Macao Joint Laboratory for Smart Discrete Manufacturing, Guangzhou 510006, China. E-mail: [email protected];
[email protected].
Bo Li is with School of Automation, Guangdong University of Technology, and also with Guangdong Key Laboratory of IoT
Information Technology, Guangzhou 510006, China. E-mail: [email protected].
Shengli Xie is with 111 Center for Intelligent Batch Manufacturing Based on IoT Technology, and also with Key Laboratory of
Intelligent Information Processing and System Integration of IoT, Ministry of Education, Guangzhou 510006, China. E-mail:
[email protected].
Ming Liu is with Department of Electronic and Computer Engineering, Hong Kong University of Science and Technolog, Hong Kong
999077, China. E-mail: [email protected].
Victor Kuzin is with Academician of Russian Engineering Academy, Moscow Moscow 125009, Russia. E-mail: [email protected].
* To whom correspondence should be addressed.
Manuscript received: 2023-10-25; revised: 2023-12-08; accepted: 2023-12-26

© The author(s) 2025. The articles published in this open access journal are distributed under the terms of the
Creative Commons Attribution 4.0 International License (http://creativecommons.org/licenses/by/4.0/).
 Yahang Qin et al.: Non-Line-of-Sight Multipath Classification Method for BDS Using Convolutional Sparse... 69

for autonomous driving, intelligent transportation, which not only increases the size of the receiving
robot navigation, and other related fields[1−3]. Usually, device, but also increases its cost. Meanwhile, sidereal
BDS can reach centimeter-level positioning accuracy filtering (SF)[14], multipath hemispherical maps
in open areas[2]. However, multipath interference (MHM)[15], 3D city models[16], and other software
dramatically reduces the positioning accuracy of BDS methods have been proposed to mitigate multipath
in the so-called “urban canyon”[4]. The multipath interference. However, the SF method requires
interference depends on the satellite elevation angle precalculation of the satellite operational period.
(ELE), the environment around the receiving antenna, Because the operational period of each satellite varies
the distance from the reflecting surface to the receiver, slowly over time, there is a consequent decrease in the
the reflection coefficient of the reflecting source, and multipath mitigation effect. The MHM method requires
the antenna characteristics. This leads to multipath a very large matrix of multipath models, which
interference becoming the main error source in the increases the computational load and does not establish
field of BDS and GNSS high-accuracy positioning[5−8]. a fully accurate multipath model. In addition, real-time
In the urban canyon environment, satellite signals updating of 3D city models is difficult.
develop multiple propagation paths by specular Artificial intelligence (AI) is a rapidly developing
reflection and bypassing. These extra signals interfere technology that can learn multipath signals from
with the direct signals from the satellite at the receiver sufficient data without the need to build complex
antenna, generating the multipath effect[9]. For the mathematical models and shows great potential for
BDS, there are two types of signals causing multipath identifying multipath in urban canyon environments.
errors: line-of-sight (LOS) multipath signals and non- These multipath signal-recognition methods can be
line-of-sight (NLOS) signals. NLOS-based errors tend divided into supervised and unsupervised cases.
to be worse than LOS-based errors. Examples of LOS Supervised algorithms for recognizing multipath
and NLOS signals are shown in Fig. 1. The LOS signals include decision tree (DT)[17], support vector
contains only the one direct signal emitted by the machine (SVM)[18], k-nearest neighbor (KNN)[19], and
satellite, and the NLOS contains only the one signal neural network (NN)[20]. They have own advantages,
and which algorithm to choose for a specific task
reflected by the environment near the receiver[10]. LOS
should be considered according to the data
multipath can easily reduce the GNSS accuracy to tens
characteristics. For example, SVM has strong
of meters. NLOS positioning errors can be several
generalization ability for small samples, but it is
hundred meters above[11]. Therefore, some current
relatively slow in processing large data sets. Overall,
studies[11, 12] are focusing on the issue of multipath
supervised learning has two main limitations: the
interference with the aim of mitigating its influence on
construction of databases containing a large number of
satellite positioning accuracy.
labels is difficult, and the recognition accuracy of the
Exploring how to cope with the satellite positioning
models needs to be improved. However, some recent
offset problem caused by multipath interference has
works have attempted to explore the use of
received wide attention. Proposals have mainly been in
unsupervised multipath-signal-recognition methods.
the areas of hardware upgrading and software
For example, the generative adversarial network
processing. Hardware devices such as anti-multipath
considers unsupervised domain adaptation (UDA)-
antennas[13] are used to filter multipath interference,
based models[21] to reduce the discrepancy between
real and simulated data. In Ref. [22], a k-means
LOS
clustering approach was used to achieve unlabeled
NLOS
satellite satellite multipath signal recognition. Unfortunately, this
process includes the problems of high-dimensional data
redundancy and low classification accuracy.
In this paper, we propose an unsupervised
convolutional sparse autoencoder containing a long
short-term memory (LSTM) module to improve the
Direct signal Reflection signal Blocked signal BDS antenna
accuracy of NLOS multipath recognition. First, to
Fig. 1 Two types of BDS signal: LOS and NLOS multipath. tackle the twin issues of the high dimensionality of the
 70 Tsinghua Science and Technology, February 2025, 30(1): 68−86

training samples and the need for a large number of provided and analyzed in section 6. Furthermore, we
labels, we use a convolutional autoencoder to realize present a discussion in this section. Finally,
descending and unsupervised signal feature extraction. conclusions are drawn in Section 7.
Then, to learn more fully the information related to the
time series of BDS signals, we employ multiple LSTM 2 Related Work
modules to capture the time series variation patterns of Numerous works have been studied in the field of
BDS signal features such as carrier-to-noise ratio multipath identification. The research includes three
(C/ N0 ), ELE, and the pseudorange residual (PR). This main kinds of methods: 3D, machine learning (ML),
enables the model to learn the underlying features of and DL-based methods. In the following, we review
the signal. Furthermore, to deal with the high- the closely related works and highlight the novelty of
dimensional data redundancy in convolutional our contributions.
autoencoder models, we propose to add an l1/2
regularizer to the objective function of the 2.1 3D-based methods
unsupervised deep learning (DL) model of a The 3D city models have been widely used in the field
convolutional autoencoder with an LSTM module, of multipath environmental positioning. Typically, a
which can remove redundant neurons from the NN combination of shadow matching and 3D mapping
while obtaining high classification accuracy. Finally, aided (3DMA) was presented in Ref. [23], and the
the performance of our model has been evaluated using method can identify satellite signal types by building
a real urban canyon dataset, demonstrating superior boundaries. The sky and building contours are
performance to that of existing models. estimated by fitting data curves to a smooth spline
The main contributions of this paper can be model, and building boundaries are calculated using an
summarized as follows: adaptive weighting scheme. Reference [24] introduced
(1) We develop a convolutional autoencoder method a shadow-matching classifier and a confidence check to
that learns a compressed representation of the input improve the NLOS classifier accuracy. Moreover,
data for downscaling and unsupervised feature exploring how to build applications over the NLOS
extraction of long-duration time-series signals also achieved wide attention. For example, applying a
containing multipath. 3D model to correct positions so that the NLOS
(2) We employ an LSTM module to capture the contributes to the absolute positioning[25]. Some works
temporal correlation of the long-duration time-series use 3D maps and ray tracing to correct the
signals from BeiDou satellites and thereby mine a localization[26]. However, 3DMA-based methods rely
temporal change law for the time series. on the timely updating of 3D models. This poses a
(3) We add an l1/2 regularizer to the objective great challenge in data-intensive computations. To
function of an unsupervised DL model of a avoid the limitations of this 3D city model, ML was
convolutional autoencoder with an LSTM module to adopted.
achieve the removal of redundant neurons from the NN
while obtaining higher sparsity and recognition 2.2 Machine learning-based methods
accuracy. Due to the complex and variable architectural structure
(4) We develop a novel convolutional sparse of the urban environment in which the receiver is
autoencoder and LSTM network model for NLOS located. Making the establishment of accurate physical
multipath classification. models is difficult. Nevertheless, ML can mine the
The rest of this paper is organized as follows. Section underlying features of large amounts of data. The
2 reviews related work. Section 3 illustrates the BDS gradient boosted the decision tree (GBDT)[17] was used
features with different signal types. Section 4 describes to classify LOS, multipath, and NLOS with three
the proposed DL networks and is explained in detail. In features: C/ N0 , PR, and ELE. Reference [19] compared
section 5, we present the environment where the data various ML algorithms such as KNN, NN, SVM, and
set is collected, followed by signal type labeling DT. LOS/NLOS classifiers based on the signal-to-noise
method using the combination of sky mask and code ratio (SNR), number of received satellites (NRS), ELE,
pseudorange double difference observable. Then, the PR, pseudorange residual percentage (PRP), and
experimental setup and classification results are normalized pseudorange residual (NRP) are
 Yahang Qin et al.: Non-Line-of-Sight Multipath Classification Method for BDS Using Convolutional Sparse... 71

established. In particular, the SVM[18, 20] has played a accuracy, we used the LSTM method to extract the
vital role in multipath detection, where they are used to relevant information from the time series. To solve the
learn the features of national marine electronics problem of high dimensional data redundancy, we use
association (NMEA), receiver independent exchange l1/2 regularizer to maintain the classification
(RINEX), and correlator levels. However, traditional performance while removing redundant neurons and
ML methods can only process information within one increasing the model sparsity. Finally, the NLOS signal
epoch, and it is easy to lose the data features in the time was classified by the constructed convolutional sparse
domain. Thus, some researchers have considered autoencoder with the LSTM model.
applying 3D techniques to identify multipath.
3 Features Selection and Analysis of
2.3 Deep learning-based methods Collected Data
Recently, with the continuous development of AI In this section, we analyze the collected data. By
techniques, the popular DL techniques have been comparing the features of BDS signals, C/ N0 , ELE,
widely adopted in various fields and have achieve and PR[14] are selected as the input data. At the same
outstanding performances. Meanwhile, many works are
time, we provide a data normalization method.
using the DL techniques to identify NLOS. For
example, the fully connected neural networks 3.1 Carrier to noise ratio (C/N0)
(FCNNs)[27] and LSTM[28] were applied to the The C/ N0 is the ratio of the received carrier signal
classification of LOS and NLOS. In Ref. [29], a deep power to the noise signal and is an important indicator
convolutional neural network (CNN) was used to of satellite signal quality. Usually, the receiver in an
detect multipath and further improve the final unobstructed environment receives mostly LOS signals
localisation accuracy. Considering the successful with a high C/ N0 . In the complex urban environment,
application of CNN in the image field. The mapping of the signal is blocked by high-rise buildings as well as
the correlator output signal to a two-dimensional input reflections, making the signal change propagation
image is proposed for classification. A deep CNN was direction and go through multiple paths to reach the
used to detect the correlator output multipath[30], and receiver, so that the C/ N0 of the NLOS is usually
this deep CNN method outperformed the SVM method. lower.
DL can be applied not only for signal classification but
also for positioning correction. Deep Neural Networks 3.2 Elevation angle (ELE)
(DNN) is used to learn a set of PR and the satellite The ELE is the magnitude of the angle between the
LOS vector as the position correction, which receiver’s horizontal plane and the satellite in the sky
outperforms the weighted least square (WLS) baseline and is an important indicator of satellite visibility.
on real data[31]. In contrast to traditional model-based Generally speaking, in the same environment, the
methods, DL has shown great potential in solving satellite signal interference is usually greater at low
complex multipath interference problems. However, angles, so the low angle is more likely to be NLOS.
the training database for multipath studies based on DL Thus, satellites with high ELE are not easily blocked
is difficult to construct and requires a large amount of by high-rise buildings and are usually LOS. The ELE θ
labeled data. In addition, the problems of high can be calculated as
dimensional data redundancy and low classification ( )
accuracy exist in the process of multipath signal θ(i) = − arcsin uiD /r̂ (1)
recognition. where uiD is the component of the distance between the
To tackle the above challenge, in this paper, we satellite and the receiver in the local Cartesian
proposed an unsupervised convolutional sparse coordinates coordinate system in the “Up” direction. r̂
autoencoder with LSTM structures to recognize the is the estimated value of the distance between the
NLOS multipath. To address the difficulty of receiver and the satellite.
constructing a large number of labeled multipath
databases, we proposed an unsupervised deep feature 3.3 Pseudorange residual (PR)
extraction method with a convolutional autoencoder. The PR is the part of the observation after subtracting
Also considering the problem of low classification the useful information, and it is a relatively objective
 72 Tsinghua Science and Technology, February 2025, 30(1): 68−86

standard for evaluating the quality of the signal. widely. Normalization can eliminate the influence of
Normally, the PR of LOS is around zero with less the different ranges of data to ensure the stable
jitter. This is due to the fact that LOS is the signal from convergence of the weights and deviations. In addition,
the satellite that arrives at the receiver via a direct path, data without normalization will slow down the training
without reflections from redundant paths. Conversely, speed of the network. Therefore, the input data features
the NLOS signal quality is poor and the PR is generally C/ N0 , ELE and PR need to be normalized so that they
larger. The position of the receiver is related to the PR remain between [0−1]. We adopt the min-max
measurements and the satellite position. normalization method, which can be expressed as:
Satellite positions can be obtained directly from the { }
xi − min x j
broadcast ephemeris. The state of the receiver can be 0⩽ j⩽n
xi∗ = { } (5)
calculated using a PR equation estimated by the least max {xk } − min x j
0⩽k⩽n 0⩽ j⩽n
squares method, which can be expressed as
−1 where xi is any value in the input sample, max {xk } and
γ = (GTG) GT ρ (2) { } 0⩽k⩽n
min x j denote the maximum and minimum values of
where γ is the receiver state. ρ denotes pseudorange. G 0⩽ j⩽n
denotes the matrix consisting of the unit LOS vector the sample, respectively.
between the satellite and receiver, which can be 4 Non-Line-of-Sight Multipath
expressed as Classification Method for BDS Using
 (1) 
 uN u(1) u(1) −1  Convolutional Sparse Autoencoder with
 E D 
 (2)  LSTM
 uN u(2) u(2) −1 
E D 
G =  . .. .. ..  (3) In this section, we first introduce the convolutional
 .
 . . . . 
  sparse autoencoder network model with an LSTM

u(i)
N u(i)
E u(i)
D −1 module for identifying BDS signals. Next, we describe
Once the pseudorange and satellite states are the objective function of the convolutional sparse
obtained, the PR can be expressed as autoencoder. Furthermore, We also present the details
of this training process of the proposed algorithm.
η = ρ−G ·γ (4)
4.1 Convolutional sparse autoencoder with LSTM
In addition to the above features, the raw data
neural network architecture
consists of the horizontal dilution of precision (HDOP),
vertical dilution of precision (VDOP), azimuth angle In this part, we first present the overall process of the
(AZ), consistency between delta pseudorange and proposed method. Meanwhile, each process of our
pseudorange rate ( ζ ) and Number of visible satellites method from the original signal to the BDS signal
(NS), geometric dilution of precision and the doppler classification is described in detail. Then, the basic
shift frequency in the RINEX format. We explore principles of convolutional sparse autoencoder with
which features contribute more to deep learning based LSTM module are presented.
classifier. When more features are used, the accuracy 4.1.1 Proposed method
decreases greatly for most models. This may be due to The identification of single-moment BDS signal type
the fact that these features are not strongly associated makes it easy to lose time correlation information, and
with the type of signal, e.g., sky occlusion varies in the extraction of BDS signal time series features is of
different environments, and it is not possible to great significance for signal identification. DL usually
determine the type of signal from a high or low requires a large amount of labeled data, which is a
azimuth angle. difficult task. In contrast, unsupervised learning
networks require less labeled data. Autoencoder
3.4 Normalization of features networks, as an important unsupervised learning
Due to we adopted a Convolutional AutoEncoder network can achieve downscaling and feature
(CAE) to extract features, the inputs and outputs of this extraction of signals using only unlabeled data.
structure should have the same range of values. However, the training process of autoencoder networks
Meanwhile, the range of the selected features varies is prone to overfitting and there is a possibility that the
 Yahang Qin et al.: Non-Line-of-Sight Multipath Classification Method for BDS Using Convolutional Sparse... 73

output layer simply copies the input layer. comprises 16 filters, the subsequent convolutional layer
To address the above problems, a convolutional comprises 32 filters, and the final convolutional layer
sparse autoencoder with an LSTM module is proposed comprises 64 filters. The kernel size of all
for the recognition of BDS signals. The proposed convolutional layers is set to 1 × 2.
method structure is shown in Fig. 2. First, a The decoder part of the convolutional sparse
convolutional sparse autoencoder is used to learn the autoencoder replaces the original pooling layers with
compressed representation of the input data. Second, inverse pooling layers and consists of three one-
The output feature data of the encoder is fed to the dimensional convolutional layers and three upsampling
LSTM layer to learn the sequential dependencies layers. The number of filters is the opposite of the
between the feature data. Then, an l1/2 regularizer encoder part, 64, 32, and 16, respectively. The kernels
sparsity constraint is added to the objective function of of all convolutional layers are also set to 1 × 2. The 1D
the method to further reduce the redundancy of the convolutional layers are mainly responsible for
neural network. Unsupervised high-level feature extracting data features from the time series. Maximum
extraction of BDS signals can be achieved with this pooling is a down-sampling operation that retains the
network. Finally, to further improve the classification maximum value in each channel of the output of the
accuracy of the proposed method, we replace the previous layer, which can effectively avoid over-
softmax layer of the last layer of the autoencoder fitting. The maximum pooling means that only the
module with a decision tree to form a hierarchical strongest of these features are retained while discarding
network, which in turn outputs the NLOS/LOS class of other weak features. The network parameters can be
BDS signals. In Fig. 2, the encoder part of the one- effectively reduced to prevent over-fitting of the
dimensional convolutional sparse autoencoder network model. This pyramidal architecture can eliminate
used in this paper consists of three convolutional layers redundant features while learning a compressed
and three maximum pooling layers. Each convolutional representation of the BeiDou satellite series signal.
layer uses Rectified Linear Unit (ReLU) as the 4.1.2 Convolutional autoencoder
activation function. The first convolutional layer The autoencoder is a typical representation learning

BDS

Hidden feature output

u-blox F9 GNSS
receiver
Data
preprocessing Root node Data
RINEX classification

Decision Decision
C/N0 , ELE , and PR node node

Leaf Leaf Leaf Leaf

node node node node
Normalization
Output

Convolutional encoder Satellite signal LOS/NLOS

Model
consturction
LSTM layer

Convolutional decoder

LOS satellite NLOS satellite Direct signal Reflection signal Blocked signal BDS antenna
Sparse weights with l 1/2
Regularizer

Fig. 2 Convolutional sparse autoencoder with LSTM module based BDS series NLOS recognition.
 74 Tsinghua Science and Technology, February 2025, 30(1): 68−86

algorithm that can achieve unsupervised feature can approximately replicate the input data from the
extraction from data. An autoencoder consists of an encoder.
input layer, a hidden layer, and an output layer forming 4.1.3 LSTM module
an encoder and a decoder. The network is trained by a LSTM can effectively solve the problems of gradient
backpropagation algorithm to make the output equal to explosion, gradient disappearance, and the inability to
the input[32]. In the encoder part of a recognition preserve historical information for a long time during
network, the hidden layer is responsible for learning the training process[33]. Therefore, we adopt LSTM to
the compressed representation of the input layer data. solve the problem of long distance dependence of time
In the decoder part of a generative network, the output series data. As shown in Fig. 4, each module contains
layer is responsible for reconstructing the input data the previous module’s hidden state h and cell state C .
from the hidden layer. The structure of the autoencoder Among them, the hidden state h can achieve short-term
is shown in Fig. 3. memory and the cell state C can achieve long-term
The process of reconstructing the input data by the memory. Satellite features have a high degree of
autoencoder can be briefly described as follows. First, temporal correlation, and LSTM can effectively extract
for the input data x ∈ Rn , the hidden layer feature the temporal information of the features, thus
h ∈ Rk (k < n) can be obtained by the encoder. Then the improving the recognition capability. The internal
output layer in the decoder reconstructs the output structure of the LSTM neural network is shown in Fig.
x̂ ∈ Rn from the hidden layer. However, autoencoder 4. The retention and forgetting of historical information
often leads to the degradation of feature learning is achieved by three gating units. The forget gate
performance when dealing with high-dimensional data determines which information needs to be deleted from
due to the excessive amount of parameters. CNN has the cell state. The input gate determines which new
the features of local connectivity and weight sharing, information is added to the cell state. The output gate
which can reduce the network parameters and speed up determines which state features of the cell are output.
the computation of the network. And the unsupervised The computational process of LSTM can be expressed
learning feature of the autoencoder can prevent the as follows:
network from overfitting. Moreover, there are local
ft = σ(W f · [ht−1 , xt ] + b f ) (6)
correlations in continuous data. CNN can perceive
richer local features of the signal using convolutional it = σ(Wi · [ht−1 , xt ] + bi ) (7)
kernels. Convolutional sparse autoencoder combined
with operations such as convolution and pooling in C̃t = tanh(WC · [ht−1 , xt ] + bC ) (8)
CNN can achieve unsupervised training. One-
Ct = ft ∗ Ct−1 + it ∗ C̃t (9)
dimensional convolution is processed on BeiDou
satellite time series data to achieve feature extraction to ot = σ(Wo · [ht−1 , xt ] + bo ) (10)
generate new feature maps. The convolutional sparse
autoencoder network is trained many times so that the ht = ot ∗ tanh(Ct ) (11)
Beidou satellite signal data output from the decoder In the method proposed in this paper, the LSTM
memory module in the forgetting gate ft obtains the
Encoder Decoder
BDS signals at time t . The information that can reflect
x1 w Bottleneck w’ x̂1
ht-1 ht ht+1
x2 h1 x̂2

Input
x3 h2 x̂ 3 Ct-1 Forget gate Input gate Output gate Ct
Error
× ＋
......

x4 x̂4 × tanh
ft it ~
Ct ot
h LSTM σ σ σ LSTM
......
......

tanh ×
xn
m
+1 x̂n
block block
＋ ＋ ＋ ＋
+1 h(x)
ht-1 LSTM block ht
x x ≈ xˆ x̂
xt-1 xt xt+1

Fig. 3 Network structure of autoencoder. Fig. 4 Structure of the LSTM neural network.
 Yahang Qin et al.: Non-Line-of-Sight Multipath Classification Method for BDS Using Convolutional Sparse... 75

the characteristics of the satellite signal type is we have

memorized using the sigmoid function, and the λsign(ωi )
information that cannot reflect the characteristics of the ωi − ai + √ =0 (15)
4 |ωi |
satellite signal type is forgotten. The sigmoid in the
input gate i ∗ C̃t will determine which important for any fixed i , the solution ω∗i must be satisfied
information in the output value of tanh that can reflect ω∗i ai > 0 . Next, we simply consider the case where
ωi ai > 0 for any i .
the type of BDS signal is to be retained. Cell state Ct √
allows the new BDS signal information found by the Case 1: ωi > 0 : At this point, we can make |ωi | = z .
neural network to be updated into the cell state. The Thus we can obtain ωi = z2 . Thus Eq. (15) can be
output gate ot is used to determine the BDS signal written in the form of a cubic algebraic of Eq. (16):
information that the hidden state ht should carry. The λ
z3 − ai z + =0 (16)
hidden state is used as the output of the current cell. At 4
the same time, the new cell state and the new hidden For the cubic algebraic equation of Eq. (16), the
state are passed to the next time step. solution can be found using the Cardano formula[35].
4.2 Objective function with the l1/2 regularizer With the different signs ai , the solution of Eq.
√ (14) is
|ai |
The objective function of the sparse neural network is different. This can be written as r = ,P =
3
constructed by imposing a sparse constraint, which can ai λ
− , and q = .
be expressed as 3 8
q2 p3 3 2 ( q )
When + < 0 , a > λ 3 , ϕλ = arccos , the
1∑
N
4 27 4 2r3
J (ω) = min L (yi , f (xi ; ω)) + λΩ (ω) (12)
ω N i=1 solution
( ) to Eq. ((16) can) be expressed (as z1 = )−2r cos
ϕλ π ϕλ π ϕλ
where y = f (xi ; ω) is represented as a generic deep , z2 = 2r cos + , and z3 = 2r cos − .
3 3 3 3 3
neural network with a vector x ∈ Rd as input. After Since ωi > 0 , so z1 less than zero is not a solution of
propagation through the neural network weight matrix Eq. (16), and z2 , z3 greater than zero is a solution of
ω formed by the hidden layers, the output vector y ∈ R0 Eq. (16). Further, analyzing z2 , z3 , because z2 > z3 , z3
is obtained. L (·, ·) is denoted as the loss function, Ω (·) is the only solution to Eq. (16). Therefore, in this case,
denotes the canonical constraint imposed by the Eq. (16) has a unique solution ω∗i ∈ RN . This unique
network, and λ ∈ (0, 1) denotes the parameter used to solution can be written as
balance the loss and canonical terms. ( ( ))
2 2π 2ϕλ (ai )
The objective function of the l1/2 regularizer method ω∗i = |ai | 1 + cos − (17)
3 3 3
in our study is  
 ( )− 2  3
1∑  λ |ai | 
where ϕλ (ai ) = arccos 
N
 .
1
J (ω) = min (y − x)2 + λ ∥ω∥ 21 (13) 4 3
ω N i=1 2 √
Case 2: ωi < 0 : At this point, we can make |ωi | = z .
where (and henceforth) ∥·∥ denotes the Euclidean norm, Thus we can obtain ωi = −z2 . Thus Eq. (13) can be
ω is the weight matrix, x is the input signal, written in the form of a cubic algebraic Eq. (18):
x = {x1 , x2 , ..., xN }T ∈ RN , y is the output signal, λ > 0 is
λ
a regularization parameter, and λ ∥ω∥ 1 is a z3 + ai z + =0 (18)
2 4
regularization term.
To solve the nonconvex and nonsmooth l1/2 norm Following the Cardano formula for solving cubic
regularized problem, we use the proximal gradient algebraic equations and performing an analysis similar
descent method[34] to solve Eq. (13), find its gradient, to that of the first case for Eq. (18), we obtain the
and set it to zero to get Eq. (14): unique solution:
 1 ( ( ))
  2 2π 2ϕλ (ai )
λ  ω∗i
= − |ai | 1 + cos − (19)
|ωi | 2  = 0
∆
ωi − ai + (14) 3 3 3
2  
3 23
( 1 ) sign(ω ) When ωi = 0 ⇔ |ai | ⩽ λi .
∆ i 4
where ai = ωi − ωi (y − x)2 . Because |ωi | 2 = √
∆
, By analyzing the above two cases, we can obtain the
2 |ωi |
 76 Tsinghua Science and Technology, February 2025, 30(1): 68−86

solution of Eq. (14) as Algorithm 1 Training the convolutional sparse autoencoder

 ( ( )) with LSTM


 2 2π 2ϕλ (ai ) 3 2


 |a i | 1 + cos − , a i > λ3 ; Input: Training input data x = {x1 , x2 , ..., xN }T ∈ RN , positive
∗  3 ( ( 3 3 )) 4
ωi = 
 parameters λ .


 2 2π 2ϕλ (ai ) 3 2

 − 3 |ai | 1 + cos 3 − , ai < − λ 3 Output: The type of the BDS signal.
3 4
(20) hlj = [h1 , h2 , ...h M ] ∈ M l , where hi is the number of hidden units
Therefore, the weight parameter ω in the objective in layer i and l is the number of hidden layers.
[ ]
function will be updated according to Eq. (20). y = y1 , y2 , ...yz ∈ Z l , where yi is the number of output units in
layer i and l is the number of output layers.
4.3 Proposed convolutional sparse autoencoder [ ] { }
1 θ = θ1 , θ2 , ...θz , where θi = Wi j , b j
l l
with LSTM algorithm
[ ] { }
2 ϕ = ϕ1 , ϕ2 , ...ϕz , where ϕi = Ŵi j , b̂ j
l l
We proposed a new method for BDS signal recognition
that combines a convolutional sparse autoencoder and 3 (θ, ϕ) ←− initialize
an LSTM network. During the training phase, the 4 while Stochastic gradient descent not converged do
encoder can learn the compressed representation of the 5 M ∼ X (sample minibatch of data)
input data, and the decoder implements the 6 for xi ∈ M do
reconstruction of the input data. The LSTM layer ∑
7 hlj = f ( i∈M j x ∗ Wil j + blj )
embedded at the end of the encoder learns the temporal ( )
features between the data. We design the mean loss 8 ht ←− LSTM hlj = xt
update parameter with an l1/2 regularizer thereby
9 j = ht
hl−1
suppressing redundant information. We use DT to ∑
complete the classier. Herein, the details of this 10 y = f ( i∈N j hl−1
j ∗ Ŵi j + b̂ j )
l l

training process are summarized in Algorithm 1. First, 11 Let the loss J be defined as (using Eq. (13))
a convolutional sparse autocodeer with LSTM is 1 ∑N 1
12 Ji ←− minω i=1 (y − x)2 + λ ∥ω∥ 21
trained with the original input data x . This allows the N 2

encoder fθ to output hlj and obtain the parameters θi . 13 Compute stochastic gradient of the loss w.r.t each ω
Next, the LSTM learns the temporal correlation of the 14 using l1/2 regularizer trick (Eq. (20)).
( )
hlj generated by the encoder and outputs the 1
15 ∇J̄ = ∇θ,ϕ ∑
dimensionality-reduced data ht . Furthermore, ht |M Ji |
obtains the reconstructed data y after going through the 16 Update network parameters by backpropagation
decoder fϕ function and updating the parameters ϕi . 17 (θ, ϕ) ←− Stochastic gradient descent optimizer
Then, we use the backpropagation algorithm to 18 for h j = [h1 , h2 , . . . , h M ] ∈ M do
l l
compute the gradient of the loss function Ji of the
l
weights and deviation vectors and optimize the 19 To classification on the data h j using DT
autoencoder with a stochastic gradient descent 20 return The type of the BDS signal , θ , ϕ
algorithm. Thereby each sample x can be mapped to a
new feature space ht . Finally, the DT model is used to antennas, two u-blox F9 GNSS modules, a laptop, two
classify and output the BDS signal type. rechargeable batteries, and a small portable foldable
table. Above the base, the fisheye camera can be
5 Experiment installed on the tripod, and other devices can be placed
In this section, we first describe the BDS signal on the table. Due to the double differential approach
collection scheme. Second, we illustrate the urban that can get more accurate positioning results, we use
complex environment of BDS signal acquisition. Then, two GNSS antennas for capturing the BDS signals, one
we depicted the BDS signal labeling method. Finally, of which is used as a base station and the other as a
the data preprocessing is given in Section 5.3. mobile station. The u-blox F9 GNSS module is used
for high-precision positioning, and it requires a
5.1 Data collection
rechargeable battery for continuous power supply. In
Figure 5 shows the BDS signal collected device, which addition, the laptop is used to store the raw BDS data.
consists of a fisheye camera, a tripod, a pair of GNSS The BDS signals were collected at four different
 Yahang Qin et al.: Non-Line-of-Sight Multipath Classification Method for BDS Using Convolutional Sparse... 77

Fisheye camera Antennas Laptop sky occlusion cases in urban canyons, and effectively
avoids the problem of poor model generalisation ability
brought by a single data distribution. Location A
corresponds to a situation where one side is blocked by
a building and the other three sides are open. Similarly,
Location B corresponds to a situation where two sides
u-blox F9
are blocked. Location C corresponds to a situation
where three sides are blocked by the building and the
other sides are open. Location D represents a deep
urban environment blocked by four sides. This location
contains more NLOS due to building blockage.
5.2 Data labeling
Fig. 5 Data collection equipment.
Labeling the true signal type on BDS data collected for
locations in Tianhe, Guangzhou. This environment is DL is necessary. In this study, a fisheye camera is used
densely populated with high-rise buildings, and many to assist in data labeling, and the labeling is achieved
of them are higher than 80 meters. Therefore, we can by the method of oprojections. The flow of data
collect a large number of NLOS. From October 10 to labeling is shown in Fig. 7. First, a sky image with
13 in 2022, we collected data from four Locations, A, building boundaries and sky regions is captured by a
B, C, and D, using the equipment in Fig. 5, and each fisheye camera. Moreover, the sky image is segmented
location collected 4 hours of static data of BDS L1 into sky area and non-sky area using an image
signals in RINEX format with the sampling frequency segmentation algorithm, where the non-sky area
set to 1 s. Thus, four BDS signal datasets (A, B, C, and corresponds to the part obscured by buildings. Finally,
D) were constructed. the ELE and AZ of the satellites acquired by the
Figure 6b shows a sky image of the data collection receiver are projected onto the fisheye camera sky
location, where grey is the area shaded by buildings image using an isotropic projection method, and the
and white is the sky area in the figure. Our data include satellites falling in the non-sky regions are NLOS
four different sky obscuration scenarios. Generally satellites and those falling in the sky region are LOS
speaking, the more sky occlusion area, the easier the satellites.
satellites are occluded by buildings, and there are fewer An example of the method shown above is given in
LOS signals in the dataset. One side, two sides, three Figs. 6b and 6c, where the original sky image taken by
sides and four sides of the sky occlusion cases are the fisheye camera is shown in Fig. 6b, and the sky
selected in the experiment, which includes the typical image after satellite projection is shown on Fig. 6c. The

Fish-eye image Sky plot

Real environment

Location A Location B

Location C Location D

(a) Real environment (b) Fish-eye image (c) Sky plot

Fig. 6 BDS data collection in an urban environments. (a) Real environment; (b) Fish-eye image; (c) Sky plot.
 78 Tsinghua Science and Technology, February 2025, 30(1): 68−86

Fish-eye camera Table 1 Description of the time series data structure.

Sky image Number of epoches Feature
GNSS receiver C/N0
Fish-eye camera image ELE PR
Binarized
1 CNR1 ELE1 PR1
sky image Satellite elevation 2 CNR2 ELE2 PR2
Image segmentation and azimuth angle
... ... ... ...
Elevation and
azimuth n−1 CNRn−1 ELEn−1 PRn−1
Uniform data n CNRn ELEn PRn
Satellite position
projection format
proposed NLOS signal classification method will
LOS or NLOS Camera azimuth angle identify the signal type of the last epoch based on the
characteristics of the signal in each sample. However,
Fig. 7 Flowchart of fish-eye image-aided determination of the value of n is not as small as possible. Too small n
LOS or NLOS satellites. will lead to insignificant sample differentiation in
white area corresponds to the sky, and the black color different types of data, and the model will have
indicates the buildings. Thus, the white areas are LOS difficulty in achieving better recognition results. Here
the input data length is taken as 128 s, and the
satellites, marked with green dots. The black areas are
sampling frequency of this paper is 1 Hz, thus a sample
NLOS satellites, marked with red dots.
contains 128 epochs. In this paper, the sliding window
5.3 Data preprocessing is chosen to be 10, and the sliding window selection
We collected 4 hours of BeiDou satellite data at each process is described in detail in the next section.
location, which in turn created a signal classification 6 Experimental Result and Discussion
dataset that can be used by the proposed unsupervised
DL model. To meet the requirements of the input data In this section, we mainly show the experimental
format of our proposed algorithm, we use a sliding results. First, we describe the effects of different input
window to split the BDS signal data. Specifically, a data lengths and sliding window sizes. Second, the l1/2
window of length n and sliding step m is used to regularizer is compared with several commonly used
regularization, and introduces the parameter adjustment
intercept the time series by sliding to the right one by
of the regularization. Next, several common evaluation
one, with each window n corresponding to one sample.
methods for assessing the performance of classification
Figure 8 shows the time series data format we used.
methods are discussed. Then, we conduct comparative
Green is 4 hours of BDS signal data collected at the
experiments using supervised learning methods such as
same location. In purple is the length of a single
DT, SVM, and CNN. Finally, sparsity analysis
sample, chosen to be 128 s. A single sample of data is experiments are implemented for the l1/2 regularizer.
slid along the yellow arrow to the right to intercept the
time series, and the sliding step is chosen to be 10 s. 6.1 Data parameter analysis
The data is labeled at the last moment. Thus, the To optimize the performance of the the proposed
construction of the BeiDou satellite time series data set method, adjustments to some parameters are necessary.
is completed. The data format we use is the BDS signal time series.
Table 1 shows the specific form of the input data Therefore, we conduct experiments on the selection of
after pre-processing, where n is the time stamp. Our the appropriate BDS signal time series length on the
Sliding window
four Datasets A, B, C, and D. The results of the
size 128 s Sliding experiments are shown in Fig. 9.
Step
10 s
We can find the classification accuracy of the
datasets from Locations A, B, C, and D is comparable
when the input data length is 16 s, and the accuracy is
relatively low. When the input data length is increased
to 32 s, the classification accuracy improves slightly,
Fig. 8 BDS signal data format. albeit not significantly. However, as the length of the
 Yahang Qin et al.: Non-Line-of-Sight Multipath Classification Method for BDS Using Convolutional Sparse... 79

datasets when the sliding step is set to 10 seconds.

Nevertheless, when the sliding step length of the BDS
time series data is increased to 12 seconds, the
classification accuracy decreases slightly. Moreover,
when the sliding step is further increased to 16
seconds, the classification accuracy decreases even
further. Therefore, based on these experimental results,
it can be concluded that 10 seconds is the most
appropriate sliding step for the proposed method. It
Fig. 9 Classification results of different window lengths.
should be noted that the experimental outcomes reveal
BDS signal time series is further increased to 48 s, that if the sliding window of the time series data is too
64 s, 80 s, and 96 s, the classification performance of short, the data fluctuation is small, and the
the proposed method progressively improves. An representation of data types with less distinctive
increase in the input BDS time series from 96 s to features. On the other hand, if the sliding step of time
128 s results in a smaller improvement in the series data is too long, it cannot extract more critical
classification accuracy. Subsequently, as the length of feature information of the BDS signal. Consequently, a
balance must be struck to achieve optimal
the time series data continues to increase to 144 s, the
performance.
classification accuracy begins to decrease. Overall, the
experimental results demonstrate that 128 s is the 6.2 Regularizer parameter analysis
optimal length for the BDS signal time series. The Our proposed model employs the l1/2 regulariser,
input time series data length being too short results in mainly because the l1/2 regulariser is easier to solve
an incomplete expression of the temporal correlation in than the classical l0 regulariser. The l1/2 regulariser
the data. Conversely, the length of the input time series produces sparser solutions than the now popular l1
data is too long, and the data is easily affected by the regulariser. Experiments have shown that the l1/2
environmental noise of the urban canyon. This leads to regulariser can replace the l p ( 0 < p < 1 ) regulariser,
the proposed method of learning noisy information which has an important and wide applications.
about the data features, which reduces the classification We compare the proposed regularizer with some
accuracy. different regularizers, including the LOG
In addition to the time series length, choosing an regularizer [36, 37] [38]
, and the MCP regularizer . We
appropriate sliding window is equally important. We chose MCP and LOG as the comparison methods. This
determined this parameter by the experiences. is due to the fact that the l1 and l2 regularizers
Figure 10 reveals that the accuracy of the method compress not only the smaller weights to 0, but also the
proposed in this paper is poor on all four datasets when larger ones, resulting in biased solutions. To
the sliding step size is set to 5 seconds. However, the compensate for this deficiency, the nonconvex
accuracy of the method on all four datasets shows a regularisers MCP and LOG are proposed, which
significant improvement when the sliding step is compress smaller weight values close to 0 while
increased to 8 seconds. Furthermore, the proposed obtaining approximate unbiased estimates for larger
method attains the highest classification accuracy on all weights. Thus, the non-convex regularisation methods
are more robust and are able to obtain more accurate
features. These regularization methods were used to
sparsely constrain the objective function of the
convolutional sparse autoencoder with the LSTM
module. As shown in Fig. 11, we can see that the use of
the l1/2 regularizer yielded the highest classification
accuracy across all datasets. The LOG regularizer was
slightly less effective than the l1/2 regularizer, while
the MCP regularizer produced the lowest classification
Fig. 10 Classification results of different sliding steps. accuracy. Therefore, our results show that the l1/2
 80 Tsinghua Science and Technology, February 2025, 30(1): 68−86

more objective assessment of the classification method.

Accuracy pertains to the fraction of accurately
classified samples from the entire pool of samples.
Precision represents the ratio of the accurately
classified positive samples to the positive samples
recognized by the classifier. The recall is the measure
of the accurately classified positive samples relative to
the actual positive samples. The F1-score is an average
of the precision and recall rates. By utilizing the
Fig. 11 Classification results of different regularizers.
pertinent parameters, the formulas for accuracy,
regularizer is the most appropriate regularization precision, recall, and F1-score can be derived as
method for this particular model. follows:
Furthermore, the l1/2 regularizer approach involves a TP + TN
Overall accuracy = (21)
regularization parameter λ , that must be adjusted for TP + TN + FP + FN
the optimal performance of the regularization TP
Recall = (22)
technique. In this study, λ was set to 0.08, 0.1, 0.12, TP + FN
0.14, and 0.16, and the resulting classification accuracy TP
of the the proposed method was evaluated across four Precision = (23)
TP + FP
environmental datasets, as depicted in Fig. 12. The
R·P
figure indicates that when λ is set to 0.08, the method's F1-score = 2 (24)
R+P
accuracy is marginally lower than when λ is set to 0.1. where TP , TF , FP , and FN represent the number of
Notably, the classification accuracy of our method is true positives, true negatives, false positives, and false
highest when λ is set to 0.1 across all datasets. negatives, respectively. Furthermore, R and P
However, as λ increases beyond 0.1, the method's represent recall and precision, respectively.
accuracy decreases, suggesting that λ of 0.1 is the most
optimal parameter for the l1/2 regularizer method. 6.4 Performances analysis
To assess the effectiveness of the proposed method for
6.3 Evaluation metrics
recognizing BDS signals, we compare the proposed
We employ accuracy, precision, recall, and F1-score to method with the existing methods, including DT[17],
evaluate the proposed method’s performance. Accuracy SVM[18], CNN[29], and Convolutional Sparse
and recall were used to evaluate the classification Autoencoder (CSAE). The first two comparison
methods concerning a particular aspect of actual or methods are based on traditional ML, and the latter two
predicted values to measure their efficacy. Given that are based on DL architecture. To be specific, the DT
the sum of false positives (FP) and false negatives (FN) method adopts a CART-type structure for
is a constant value, it is not achievable to enhance both classification. The CNN method uses 3 convolutional
precision and recall evaluation metrics. The F1-score, layers to obtain features and a softmax classifier for the
which integrates both precision and recall into a single output layer. The structure of the CSAE method is
metric, is an alternative evaluation metric that enables a similar to the classifier part of this paper but does not
include the LSTM module.
The evaluation results are listed in Tables 2−5. we
can see that our proposed method surpasses other
methods in terms of accuracy, precision, recall, and F1-
score performance metrics across four different
locations. Specifically, our proposed method achieved
accuracy of 92.96 % , 90.85 % , 79.37 % , and 77.51 % on
the datasets of the four locations, respectively.
Compared to traditional ML algorithms and common
Fig. 12 Classification results of different λ. CNN methods, our proposed algorithm yielded
 Yahang Qin et al.: Non-Line-of-Sight Multipath Classification Method for BDS Using Convolutional Sparse... 81

Table 2 Performance comparison of Location A. on the dataset of Location A.

Evaluation index In addition, accuracy is crucial for NLOS
Method
Acc (%) F1-score (%) Prec (%) Recall (%) recognition, as NLOS signals can have a detrimental
DT[17] 81.40 67.56 67.91 67.24 impact on positioning. Thus, we give an in-depth
SVM[18] 82.23 45.12 41.12 50 analysis of accuracy, and the results are illustrated in
CNN[29] 85.70 74.64 75.61 73.80 Fig. 13. This analysis visually demonstrates that our
CSAE 90.69 79.58 82.78 77.20 proposed method outperforms other methods in terms
Proposed method 92.96 84.72 87.29 82.62 of accuracy across four datasets. The convolutional
sparse autoencoder with LSTM method suggested for
Table 3 Performance comparison of Location B. the recognition of BDS signals holds great promise in
Evaluation index its ability to effectively utilize a substantial quantity of
Method
Acc (%) F1-score (%) Prec (%) Recall (%) unlabeled data and a comparatively small amount of
DT[17] 86.50 86.12 86.30 85.98 labeled data, while simultaneously accounting for
SVM[18] 73.04 72.24 72.40 72.13 temporal continuity through the unsupervised training
CNN[29] 86.96 86.63 86.68 86.59 process. In comparison to ML techniques that merely
CSAE 88.89 88.54 88.85 88.31
identify features at a particular point in time, the NN
Proposed method 90.85 90.56 90.83 90.34
method proposed by us can extract brief fluctuations in
the time-series features of the BDS signals. The
Table 4 Performance comparison of Location C.
method is valuable for applying DL in the field of BDS
Evaluation index
Method signal recognition.
Acc (%) F1-score (%) Prec (%) Recall (%)
Finally, we analyze the mean evaluation metrics of
DT[17] 75.63 68.22 68.45 68.02
four location datasets, and the results of the
SVM[18] 73.67 42.42 36.83 50
experiments are presented in Fig. 14. It is observed that
CNN[29] 77.56 69.90 70.04 69.78
SVM, DT, and CNN exhibited high accuracy, but low
CSAE 78.35 71.17 71.59 70.80
performance in other metrics, indicating a lack of
Proposed method 79.37 72.40 72.78 72.06
method robustness. The method in this research
achieved notable accuracy rates of 85.17 % , 81.27 % ,
Table 5 Performance comparison of Location D.
82.08 % , and 80.59 % in terms of accuracy, F1-score,
Evaluation index
Method precision, and recall, respectively, outperforming all
Acc (%) F1-score (%) Prec (%) Recall (%)
other experimental methods. These results validate the
DT[17] 74.69 74.63 74.61 74.65
effectiveness of the proposed method. The CSAE
SVM[18] 64.52 61.36 67.78 63.10
method also performed well in terms of accuracy, F1-
CNN[29] 75.42 75.31 75.30 75.33 score, precision, and recall, but its metrics were lower
CSAE 76.34 76.21 76.20 76.20
than those of the proposed method, suggesting that the
Proposed method 77.51 77.38 77.40 77.35

significant improvements in accuracy, precision, recall,

and F1-score. For instance, our method achieved a
classification accuracy of 92.96 % on location A,
whereas the DT, SVM, CNN, and CSAE method
achieved 81.4 % , 82.23 % , 85.70 % , and 90.69 % ,
respectively. This indicates that our proposed method
is more accurate and stable than other methods, with a
maximum difference in accuracy of 11.56 % and a
minimum difference of 1.86 % . Overall, the proposed
algorithm demonstrates significant improvement
compared to traditional ML algorithms, with high
accuracy, precision, recall, and F1-score reaching
92.96 % , 84.72 % , 87.29 % , and 82.62 % , respectively, Fig. 13 Classification performance of different methods.
 82 Tsinghua Science and Technology, February 2025, 30(1): 68−86

Fig. 14 Average classification performance of different Fig. 15 Sparsity performance of l1/2 regularizer.
methods.
classification, sparsity is ensured. This indicates that
addition of the LSTM module can further enhance the the algorithm is capable of acquiring more effective
method's performance. It is worth noting that SVM features of BDS signals.
approaches achieve such poor performance compared
6.6 Discussions
to other methods. The main reason may be that SVM
can map low-dimensional data into high-dimensional In our experiments, we investigate the performance of
space by kernel function, and then improve the our proposed convolutional sparse autoencoder with
classification accuracy. Therefore, SVM is more the LSTM method in real environments with a variety
suitable for classification problems with low- of sky occlusion scenarios. In contrast to ML methods
dimensional small-sample data. such as SVM and DT, we propose a convolution-based
DL method. In addition, unlike CNN and convolutional
6.5 Sparsity analysis of the weights with the l1/2
sparse autocoder methods, our method incorporates
regularizer layer
LSTM into the method to extract important timing
To analyze the sparsity of the l1/2 regularizer, we use information. We also propose adding an l1/2 regularizer
the sparsity evaluation metric proposed by Hoyer’s to the objective function to induce strong sparsity
sparsity[39] to calculate the sparsity of the feature among the parameters. Specific performance
representation learned by the proposed method. And comparisons of our method against existing methods
this method uses the difference between the l1 and l2 are listed as follows.
parameters of the vector to evaluate the sparsity of the (1) From the detailed comparisons, as shown in
vector. The sparsity can be expressed as Tables 2–5, it is clear that the proposed method is
√ (∑n √∑ ) superior to ML methods such as SVM and DT for all
n − i=1 |xi | / n
x
i=1 i
2

Hoyer’s sparsity(x) = metrics. Our proposed DL method improved the

√ (25)
n−1 accuracy, precision, recall, and F1-scores by 11.56 % ,
where x is the sample vector and n is the 17.16 % , 19.38 % , and 15.38 % , respectively, compared
dimensionality of x . The value of Hoyer’s sparsity is in with the DT method. The results suggest that the
the range of [0, 1] , and the more larger the Hoyer’s proposed method extracts rich features in the training
sparsity, the greater the difference between the process and can further improve performance across a
elements in the vector, indicating that x is more sparse. variety of sky occlusion environments.
In Fig. 15, the correlation between the quantity of (2) Compared with existing DL methods such as the
iterations and the scarcity of the proposed method is CNN and CSAE methods, as shown in Tables 2–5, the
illustrated. The results of the experiment demonstrate proposed method improved accuracy by 7.26 % and
that the l1/2 regularizer progressively approaches a 2.27 % , respectively. The proposed method obtained
consistent level of scarcity as the number of iterations better recognition accuracy in four environments (e.g.,
is raised. While preserving a high standard of 92.96 % for Dataset A, as shown in Fig. 13).
 Yahang Qin et al.: Non-Line-of-Sight Multipath Classification Method for BDS Using Convolutional Sparse... 83

Furthermore, by checking the average values for through seeking deep sparse representations during
accuracy, F1-score, precision, and recall, we note that propagation, the proposed method can improve
all metrics for the method proposed in this paper classification performance by 1 % –2 % . Furthermore,
exceed 80 % , which is not achieved by any of the other the proposed l1/2 regularizer can largely remove
methods, as shown in Fig. 14. unnecessary weights while retaining the classification
(3) In Fig. 11, the performances of various performance. In particular, it can reduce over 20 % of
regularizers when using the corresponding optimal the parameters while maintaining a significantly high
parameters are summarized. Here, the l1/2 regularizer classification accuracy for all urban canyon
achieves the best classification performance across all environments. The main reason may be that the method
datasets, which indicates that the l1/2 regularizer is we proposed method that learns a compressed
robust and can improve sparsity. A comparison of representation of the input data for downscaling and
Hoyer’s sparsity is shown in Fig. 15, where the l1/2 unsupervised feature extraction of long-duration time-
regularizer pruned unnecessary weights from the NN series signals containing multipath. The cited methods
(e.g., around 72 % of the total across the four datasets). also uses an unsupervised approach, but the cited
Furthermore, the l1/2 regularizer could converge faster methods uses a machine learning approach, and the
with fewer training epochs. recognition accuracy is significantly lower than the
deep learning unsupervised feature extraction method
7 Conclusion
used in this paper.
In this paper, we have presented an efficient However, there are two main limitations of our
convolutional sparse autoencoder with an LSTM algorithm: (1) The proposed algorithm only considers
training framework to address the NLOS signal urban canyon; (2) The classification generalisation
classification problem. Specifically, to prevent high- performance of the proposed algorithm needs to be
dimensional data redundancy in DL methods, we improved. In order to address the above problem, in the
implemented dimension reduction and feature future work, we will continue our studies of multipath
extraction through a convolutional sparse autoencoder signal-recognition-based unsupervised DL methods in
module structure. To capture the temporal correlations more diverse and complex urban environments such as
in long-duration time-series signals, we proposed an urban forests, overpasses, and viaducts. We note that
LSTM module that uses an input gate, a forget gate, our method loses information about the satellite
and an output gate to obtain information about the environment, so we are considering learning the
evolution of the BDS signal over time. Then, to reduce representation of indirect environmental information
unnecessary weights, we applied a regularizer to the using a transformer. We are also working on inducing
weights of the hidden layers, which offered results sparsity in DL algorithms that involve a CNN
beyond those for the LOG and MCP regularizers. architecture, which should be more efficient for more
Finally, the trained and simplified DL could provide complex DL environments.
accurate classification of BDS NLOS signals in real
urban canyon environments. Acknowledgment
Experiments with classification and pruning were This research was supported in part by the National
conducted on four datasets with different sky occlusion Natural Science Foundations of China (Nos. 62273106,
scenarios. Compared with ML methods such as DT and 62203122, 62320106008, 62373114, 62203123, and
SVM, our DL method, with its convolutional sparse 62073086), in part by Guangdong Basic and Applied
autoencoder and LSTM module, demonstrated superior Basic Research Foundation (Nos. 2023A1515011480
performance. Moreover, compared with existing DL- and 2023A1515011159), and in part by China
based methods such as CNNs and SCAEs, the Postdoctoral Science Foundation funded project (No.
proposed method achieved 7.26 % and 2.27 % 2022M720840).
improvement in accuracy, respectively. Compared with
the LOG and MCP regularizers, the proposed l1/2 References
regularizer used for NN pruning enabled our proposed [1] C. Ma, S. Pan, W. Gao, F. Ye, L. Liu, and H. Wang,
method to induce strong sparsity in the parameters. The Improving GNSS/INS tightly coupled positioning by using
results demonstrate that, across all environments, BDS-3 four-frequency observations in urban
 84 Tsinghua Science and Technology, February 2025, 30(1): 68−86

environments, Remote Sens., vol. 14, no. 3, p. 615, 2022. and C. Yu, Multipath mitigation in GNSS precise point
[2] F. Li, R. Tu, J. Hong, S. Zhang, M. Liu, and X. Lu, positioning based on trend-surface analysis and multipath
Performance analysis of BDS–5G combined precise point hemispherical map, GPS Solut., vol. 25, no. 3, p. 119,
positioning, Remote Sens., vol. 14, no. 13, p. 3006, 2022. 2021.
[3] Y. Xu, Y. Yang, and J. Li, Performance evaluation of [16] A. El-Mowafy, B. Xu, and L. T. Hsu, Integrity monitoring
BDS-3 PPP-B2b precise point positioning service, GPS using multi-GNSS pseudorange observations in the urban
Solut., vol. 25, no. 4, p. 142, 2021. environment combining ARAIM and 3D city models, J.
[4] H. Zhao, Z. Li, C. Chen, L. Wang, K. Xie, and S. Xie, Spatial Sci., vol. 67, no. 1, pp. 91–110, 2022.
Fusing vehicle trajectories and GNSS measurements to [17] R. Sun, G. Wang, W. Zhang, L. T. Hsu, and W. Y.
improve GNSS positioning correction based on actor-critic Ochieng, A gradient boosting decision tree based GPS
learning, in Proc. 2023 Int. Technical Meeting of The signal reception classification algorithm, Appl. Soft
Institute of Navigation, Long Beach, CA, USA, 2023, pp. Comput., vol. 86, p. 105942, 2020.
82–94. [18] Y. Qin, Z. Li, S. Xie, R. Yuan, and J. Xie, BDS multipath
[5] W. Qiu, Q. Zeng, R. Xu, J. Liu, J. Shi, and Q. Meng, A signal classification using support vector machine, in
multipath mitigation algorithm for GNSS signals based on Proc. 4th Int. Conf. Industrial Artificial Intelligence (IAI),
the steepest descent approach, Satell. Navig., vol. 3, no. 1, Shenyang, China, 2022, pp. 1–6.
p. 14, 2022. [19] G. Zhang, B. Xu, and L. T. Hsu, GNSS shadow matching
[6] P. R. R. Strode and P. D. Groves, GNSS multipath based on intelligent LOS/NLOS classifier, in Proc. 16th
detection using three-frequency signal-to-noise IAIN World Congr., Chiba, Japan, pp. 1–7, 2018.
measurements, GPS Solut., vol. 20, no. 3, pp. 399–412, [20] T. Suzuki and Y. Amano, NLOS multipath classification
2016. of GNSS signal correlation output using machine learning,
[7] Groves, P. D, Jiang, Z, Rudi, M, and Strode, P, A portfolio Sensors, vol. 21, no. 7, p. 2503, 2021.
approach to NLOS and multipath mitigation in dense [21] R. Zawislak, M. Greiff, K. J. Kim, K. Berntorp, S. Di
urban areas, in Proc. the 26th Int. Technical Meeting of the Cairano, M. Konishi, K. Parsons, P. V. Orlik, and Y. Sato,
Satellite Division of The Institute of Navigation, Nashville, GNSS multipath detection aided by unsupervised domain
TN, USA, 2013, pp. 3231–3247. adaptation, in Proc. the 35th Int. Technical Meeting of the
[8] L. T. Hsu, GNSS multipath detection using a machine Satellite Division of The Institute of Navigation, Denver,
learning approach, in Proc. IEEE 20th Int. Conf. CO, USA, pp. 2127–2137, 2022.
Intelligent Transportation Systems (ITSC), Yokohama, [22] H. Wang, S. Pan, W. Gao, Y. Xia, and C. Ma,
Japan, 2017, pp. 1–6. Multipath/NLOS detection based on K-means clustering
[9] Z. Li, K. Zeng, L. Wang, X. Kan, R. Yuan, and S. Xie, for GNSS/INS tightly coupled system in urban areas,
NLOS/LOS classification by constructing indirect Micromachines, vol. 13, no. 7, p. 1128, 2022.
environment interaction from GNSS receiver [23] H. Xu, L. T. Hsu, D. Lu, and B. Cai, Sky visibility
measurements using a transformer-based deep learning estimation based on GNSS satellite visibility: An approach
model, in Proc. 2023 Int. Technical Meeting of The of GNSS-based context awareness, GPS Solut., vol. 24,
Institute of Navigation, Long Beach, CA, USA, 2023, pp. no. 2, p. 59, 2020.
859–870. [24] H. F. Ng, G. Zhang, and L. T. Hsu, Robust GNSS shadow
[10] I. Smolyakov, M. Rezaee, and R. B. Langley, Resilient matching for smartphones in urban canyons, IEEE Sens.
multipath prediction and detection architecture for low- J., vol. 21, no. 16, pp. 18307–18317, 2021.
cost navigation in challenging urban areas, NAVIGATION, [25] S. Zhang, S. Lo, Y. H. Chen, T. Walter, and P. Enge,
vol. 67, no. 2, pp. 397–409, 2020. GNSS multipath detection in urban environment using 3D
[11] Z. Lyu and Y. Gao, An SVM based weight scheme for building model, in Proc. IEEE/ION Position, Location and
improving kinematic GNSS positioning accuracy with Navigation Symp. (PLANS), Monterey, CA, USA, 2018,
low-cost GNSS receiver in urban environments, Sensors, pp. 1053–1058.
vol. 20, no. 24, p. 7265, 2020. [26] S. Miura, L. T. Hsu, F. Chen, and S. Kamijo, GPS error
[12] J. Zidan, E. I. Adegoke, E. Kampert, S. A. Birrell, C. R. correction with pseudorange evaluation using three-
Ford, and M. D. Higgins, GNSS vulnerabilities and dimensional maps, IEEE Trans. Intell. Transport. Syst.,
existing solutions: A review of the literature, IEEE Access, vol. 16, no. 6, pp. 3104–3115, 2015.
vol. 9, pp. 153960–153976, 2021. [27] G. Zhang, P. Xu, H. Xu, and L. T. Hsu, Prediction on the
[13] J. Chen, H. Shi, Z. Fang, C. Yuan and Y. Xu, Performance urban GNSS measurement uncertainty based on deep
analysis of the GNSS instantaneous ambiguity resolution learning networks with long short-term memory, IEEE
method using three collinear antennas, IEEE Sens. J., pp. Sens. J., vol. 21, no. 18, pp. 20563–20577, 2021.
11936–11945, 2023. [28] H. Su, B. Wu, and X. Mao, Non-line-of-sight multipath
[14] G. Chang, C. Chen, Y. Yang, and T. Xu, Tikhonov detection method for BDS/GPS fusion system based on
regularization based modeling and sidereal filtering deep learning, J. Shanghai Jiaotong Univ. Sci., vol. 27, no.
mitigation of GNSS multipath errors, Remote. Sens., vol. 6, pp. 844–854, 2022.
10, no. 11, p. 1801, 2018. [29] Y. Quan, L. Lau, G. W. Roberts, X. Meng, and C. Zhang,
[15] R. Lu, W. Chen, D. Dong, Z. Wang, C. Zhang, Y. Peng, Convolutional neural network based multipath detection
 Yahang Qin et al.: Non-Line-of-Sight Multipath Classification Method for BDS Using Convolutional Sparse... 85

method for static and kinematic GPS high precision pp. 1013–1027, 2012.
positioning, Remote. Sens., vol. 10, no. 12, p. 2052, 2018. [35] Z. Li, S. Ding, W. Chen, Z. Yang, and S. Xie, Proximal
[30] A. Blais, N. Couellan, and E. Munin, A novel image alternating minimization for analysis dictionary learning
representation of GNSS correlation for deep learning and convergence analysis, IEEE Trans. Emerg. Top.
multipath detection, Array, vol. 14, p. 100167, 2022. Comput. Intell., vol. 2, no. 6, pp. 439–449, 2018.
[31] A. V. Kanhere, S. Gupta, A. Shetty, and G. Gao, [36] H. Zhao, J. Wu, Z. Li, W. Chen, and Z. Zheng, Double
Improving gnss positioning using neural-network-based sparse deep reinforcement learning via multilayer sparse
corrections, NAVIGATION-US, vol. 69, no. 4, p. 548, coding and nonconvex regularized pruning, IEEE Trans.
2022. Cybern., vol. 53, no. 2, pp. 765–778, 2023.
[32] Z. He, H. Shao, Z. Ding, H. Jiang, and J. Cheng, Modified [37] Z. Li, H. Zhao, Y. Guo, Z. Yang, and S. Xie, Accelerated
deep autoencoder driven by multisource parameters for Log-regularized convolutional transform learning and its
fault transfer prognosis of aeroengine, IEEE Trans. Ind. convergence guarantee, IEEE Trans. Cybern., vol. 52, no.
Electron., vol. 69, no. 1, pp. 845–855, 2022. 10, pp. 10785–10799, 2022.
[33] K. Greff, R. K. Srivastava, J. Koutnik, B. R. Steunebrink, [38] Z. Li, Z. Yang, H. Zhao, and S. Xie, Direct-optimization-
and J. Schmidhuber, LSTM: A search space odyssey, based DC dictionary learning with the MCP regularizer,
IEEE Trans. Neural Netw. Learn. Syst., vol. 28, no. 10, pp. IEEE Trans. Neural Netw. Learning Syst., vol. 34, no. 7,
2222–2232, 2017. pp. 3568–3579, 2023.
[34] Z. Xu, X. Chang, F. Xu, and H. Zhang, L1/2 regularization: [39] P. O. Hoyer, Nonnegative matrix factorization with
A thresholding representation theory and a fast solver, sparseness constraints, J. Mach. Learn. Res., vol. 5, pp.
IEEE Trans. Neural Netw. Learning Syst., vol. 23, no. 7, 1457–1469, 2004.

Yahang Qin received the MS degree from Zhenni Li received the BSc degree from
Guangxi University of Science and School of Physical Science and
Technology, Liuzhou, China, in 2017. She Electronics, Shanxi Datong University, in
is currently pursuing the PhD degree in 2009, the MSc degree from School of
control science and engineering with Physics and Optoelectronic, Dalian
School of Automation, Guangdong University of Technology, in 2012, and
University of Technology, and also with received the PhD degree from School of
Guangdong-Hong Kong-Macao Joint Computer Science and Engineering,
Laboratory for Smart Discrete Manufacturing, Guangdong University of Aizu, Japan. She is currently an associate
University of Technology. Her research interests include deep professor with School of Automation, Guangdong University of
learning and BDS positioning in challenging environments. Technology, and also with Guangdong-Hong Kong-Macao Joint
Laboratory for Smart Discrete Manufacturing, Guangdong
Shengli Xie received the BS degree in University of Technology. Her research interests include deep
mathematics from Jilin University, China learning, reinforcement learning, and high-precious position.
in 1983, the MS degree in mathematics
from Central China Normal University, Bo Li received the MS degree in
China in 1995, and the PhD degree in mathematics from Central China Normal
control theory and applications from South University, Wuhan, China, in 1992, and
China University of Technology, China in the PhD degree in control theory and
1997. He is currently a full professor in applications from South China University
School of Automation, Guangdong University of Technology, of Technology, Guangzhou, China, in
the head of the 111 Center for Intelligent Batch Manufacturing 1997. He is currently the director of
Based on IoT Technology (GDUT), and also with the Key Laboratory for Intelligent Information
Laboratory of Intelligent Information Processing and System Processing and a full professor with the School of Automation,
Integration of IoT (GDUT), Ministry of Education. He has Guangdong University of Technology, and also with Guangdong
coauthored two books and over 150 research papers in refered Key Laboratory of IoT Information Technology (GDUT). He
journals and conference proceedings, and was awarded Highly has authored or coauthored two monographs, a dozen patents,
Cited Researcher in 2020. His research interests include blind and over 80 papers in journals and conference proceedings. His
signal processing, machine learning, and Internet of Things. He current research interests include automatic controls, signal
was awarded the Second Prize of National Natural Science processing, blind signal processing, and image processing.
Award of China in 2009. He is an associate editor of IEEE
TSMC.
 86 Tsinghua Science and Technology, February 2025, 30(1): 68−86

Victor Kuzin is a leading expert in the Ming Liu received the PhD degree in
field of physical and technical control of robotics from ETH Zurich. He is currently
mining products, physical processes, and the director of Intelligent Autonomous
quality management. He has published Driving Center of HKUST. He published
over 300 articles. He is the author of 18 over 300 articles. His research interests
monographs and booklets, 3 of which were include dynamic environment modeling, 3-
published abroad. He is a full member of D mapping, machine learning, and visual
Academician of Russian Engineering control. He received various IEEE awards,
Academy, Academician of Russian Academy of Natural including the IEEE IROS Young Professional Award in 2018
Sciences, International Academy of Mineral Resources and 15 best paper awards or finalist awards. He is particularly
Sciences, Academician of Academy of Technological Sciences interested in the investigation of novel, real-time online
of the Russian Federation, and Academician of International approaches in solving autonomous driving, mobile robot
Academy of Mineral Resources. He is an honorary doctor and mapping, and navigation. His team developed the first
professor at several universities in Europe and China. In 1979, autonomous vehicle in Hong Kong in 2017. During COVID-19,
he became a recipient of the Lenin Komsomol Prize. He was his autonomous vehicles served several cities and were reported
awarded the Russian Government Prize for Science and by more than 100 media, including IEEE Spectrum and Xinhua.
Technology in 2011 and the Russian Prize for Science and
Technology in 2013.