IEEE SENSORS JOURNAL, VOL. XX, NO.

XX, XXXX 2022 1

Machine Learning Methods for Material

Identification Using mmWave Radar Sensor
Sruthy Skaria, Member, IEEE, Nermine Hendy, Member, IEEE, Akram Al-Hourani, Senior Member, IEEE

Abstract— In recent years, radar sensors are gaining a paramount

role in non-invasive inspection of different objects and materials. In
this paper, we present a framework for using machine learning in
material identification based on their reflected radar signature. We
employ multiple receiving channels of the radar module to capture
the signatures of the reflected signal from different target materi-
als. Within the proposed framework, we present three approaches
suitable for material classification, namely: (i) Convolutional Neural
Networks (CNN), (ii) k-Nearest Neighbor, and (iii) Dynamic Time
Warping (DTW). The proposed framework is tested using extensive
experimentation and found to provide near-ideal classification ac-
curacy in classifying six distinct material types. Furthermore, we
explore the possibility of utilizing the framework to detect the volume of the identified material, where the obtained
classification accuracy is above 98% in distinguishing three different volume levels.
Index Terms— Convolutional neural networks, dynamic time warping, material identification, FMCW, k-NN, machine
learning, material identification, millimeter wave (mmWave) sensors .

I. I NTRODUCTION The characteristic properties of the signal reflected from the

Non-invasive sensing and identification of various objects different layers of materials can hold distinguishable signature
and materials has become one of the prime applications of the materials and thus can be used to identify the subject [7].
of smart sensing technologies. The primary reason for its This signature is detected utilizing different approaches, such
importance is the offered benefits in contamination control, as, (i) model-based parameter measurement, (ii) traditional
infection control, and accident prevention. It also mitigates machine learning with manually extracted features, and (iii)
environmental disruption by identifying dangerous materials deep learning methods with self-learned feature extraction.
or objects. Two of the most popular methods for non-invasive In model-based approaches; the dielectric properties of
detection of materials are: (i) the use of X-ray diffraction [1], the materials are estimated by utilizing various analytic mod-
[2], and (ii) the use of X-ray fluorescence [3], [4], where els [8], [9]. For example, in [8], an FMCW radar-based Syn-
both methods are superior in their quantitative and qualitative thetic Aperture Radar (SAR) is used to generate a 3D image of
analysis of materials. However, X-ray is an ionizing radiation, the target material with a very high resolution. Thereafter, the
and the extended exposure of these radiations has multiple relative permittivity and loss tangent of the materials under test
adverse effects on both the subject under test and the human are estimated using an analytic model for the reflection factor
operator. Other sensing technologies such as optical, infrared, of the dielectric slab, which in-turn gives insights into the
and multispectral imaging are all capable of detecting the target material. Another example using SAR-based approach
reflections from a target, yet they are unable to penetrate can be found in [10], which proposes the extraction of the
the surface of non-opaque materials. MmWave-based radar material dispersion 1 at each pixel of the SAR image by
sensing, on the other hand, is increasingly being used for dividing the received signals into several frequency sub-bands.
sub-surface detection, where the wide commercial availability Thereafter, a rudimentary classification analysis is employed
of this non-ionizing sensor is quite promising for mate- to identify the materials from the dispersion features. The
rial identification [5]. Compact mmWave radar sensors that work in [9] proposes the estimation of the dielectric constant
work reliably are readily available, whether as an integrated utilizing an analytic model relating the dielectric constant and
chipset [6] or as a packaged module, this further enables low- range of the target based on the standard radar equation, the
cost integration of the material identification into production experiment is conducted on two passive printed tags as radio
and processing lines. frequency identification (RFID) sensors. A similar approach
is seen in [12], which employs RFID sensors to monitor the
S. Skaria is with Telstra Corporation Limited, Melbourne, Aus-
tralia. N. Hendy, A. Al-Hourani are with the STEM College,
RMIT University, Melbourne, 3000, Australia, Email: nerminea- 1 Material dispersion manifests as a dependency of reflectivity on the
[email protected],[email protected] illuminating frequency [11].
2 IEEE SENSORS JOURNAL, VOL. XX, NO. XX, XXXX 2022

quality of concrete slabs. these four sets of signals to classify the target material using
In traditional machine learning approaches, a set of three different approaches: (i) two dimensional (2D) CNN,
predefined features are extracted from the reflected signals (ii) k-nearest neighbors k-NN with traditional features, and
which are then utilized for training a machine learning model. (iii) k-NN with dynamic time warping (DTW) distance as the
Compared to the previous model-based, machine learning feature.
techniques typically give real-time classification of materials In 2D CNN approach, we represent the time-domain signals
with higher reliability since analytic approach is vulnerable from the multi receiving channels as a 2D matrix to be
to noisy data and nonlinear relations between the observed fed to a 2D CNN to extract the unique features and utilize
quantities and the corresponding physical parameters [13]. them to classify the target material. In the second k-NN
Machine learning algorithms are able to provide consistent approach, only two predefined features are extracted from the
and accurate classification performance without the need for time-domain signals. In the final k-NN DTW approach, we
task-specific algorithms. These algorithms are designed such utilize the DTW distance between the time-domain signals
that they learn the features from the training dataset itself and reference signals as the input features to train a k-NN
and make classifications based on these learned features. classifier. The parameters of each of the classification models
Therefore, to achieve good classification performance in any are optimized to maximize the classification accuracy. The
practical scenarios the training data need to maintain sufficient developed framework is applied to investigate its performance
diversity in terms of material parameters (size, orientation, in recognizing six types of materials: (i) cardboard, (ii) sand,
shape, etc.) and surrounding environment conditions [14]. (iii) sponge, (iv) Polystyrene, (v) water, and (vi) wood. We
Traditional machine learning algorithms such as support vector further expand the framework to predict the geometric volume
machine (SVM), k-NN, decision trees, etc, require human of the detected materials. The main contributions of the paper
intervention for carefully predefining a set of features from can be summarized as follows,
the input data before feeding these features to the classifier. • It develops a framework to map time-domain signals
An example on the use of traditional machine learning is the captured by an FMCW radar with multiple receiving
work in [15], which utilizes a prototype miniature radar device channels into deep learning using 2D-CNN for material
developed by Google, named Soli. The work uses eight time- identification.
domain signals received by the eight receiving channels of • It creates a minimized predefined feature set, of 2 fea-
the radar to extract a large set of predefined features that are tures, to identify different materials using traditional ML
fed into different traditional machine learning models. Another approach, k-NN classifier.
approach in [14], utilizes a CMOS-based mmWave receiver to • It provides a new method in applying DTW to extract the
capture the signals passing through different target materials, similarity of measurements between the reflected signals
which are transmitted from a W-band transmitter. As the to train a k-NN classifier for material identification.
transmitted frequency sweeps between 75 GHz to 110 GHz, • It extends the identification framework to predict the
the received signal is down-converted by the receiver and fed volume of the target materials.
to a spectrum analyzer to measure the peak power at each
frequency. These measurements are thus utilized to compute
II. F RAMEWORK OVERVIEW
the transmittance and represented as a features vector that
in turn is fed to three different classifiers: SVM, multi-layer Millimeter-wave radars have advantage over other types of
perceptron, and Gaussian process classifier to identify four sensors as they have a small footprint allowing embedding
different materials. them into mobile phones, miniature, and handheld devices.
In deep learning approach, unlike traditional machine Radars support high privacy compared to vision sensors and
learning, there is no need for the manual predefinition of are a user-convenient sensors. Additionally, miniature radars
features before feeding to the classification algorithm. Instead, have low power consumption compared to other sensors.
deep-learning algorithms self-learns the unique features from Therefore, the proposed framework utilizes a multi-receive
the inputs during the supervised training process [16]–[18]. An channel FMCW radar, a Commercial off-the-shelf (COTS)
example of material identification system is proposed in [19] device, to capture the signature reflected from the target mate-
which is based on a portable 3D RF-based imaging system. rials. FMCW radars measures target range R by transmitting
The sensor acquires a 3D radiation map of the target material, a continuous frequency-modulated signal called chirp with a
which is fed to a convolutional neural network (CNN) to sweep bandwidth B, compared to conventional continuous
identify the material. The results are promising, with a high wave (CW) radars.
classification accuracy of 93.3% for 12 different materials. The reflected signal is then down-converted to obtain the
The integration of deep-learning with radar sensors for non- coherent time-domain samples capturing the combined char-
invasive testing of materials is in its early stage and thus opens acteristics of the target material in terms of its transmission
a wide scope for research and development. and reflection properties. The reflectivity of radar signal de-
In this paper, we explore the application of mmWave pends on the material properties such as refractive index and
FMCW radar with multiple antennas to capture the variations thickness of the target material [20]. Similarly, the dielectric
in the received signals which are reflected from the subject properties of the materials affect the absorbed signal power.
material. Each receiving antenna constitute a receiving channel For example, as the dielectric constant increases, the reflection
giving four sets of in-phase and quadrature signals. We utilize coefficient increases which results in more signal reflections
S. SKARIA et al.: PREPARATION OF PAPERS FOR IEEE SENSORS 3

from the materials [21]. Therefore, according to the material features such as the mean and standard deviation from the
properties, the received signals features such as: the mean, signals to train a k-NN classifier. In the third approach (iii)
standard deviation, and power spectral density will vary. These we utilize dynamic time warping distance to measure the
variations in the characteristic properties of the time-domain similarities between the received signals and the reference
signals from different materials are utilized as clues to perform templates which are then utilized as the input features for the
material classification. k-NN classifier. We further experiment on the ability of the
Fig. 1 provides a conceptual block diagram of the pro- framework to classify the materials according to the volume.
posed classification framework with inputs from N receiving
channels, and the output as the prediction of material type III. DATA C OLLECTION
by the classification network. The transmitting antenna emits Fig.3 depicts a simplified block diagram of the utilized
FMCW chirp signals which are partially reflected back to the TI IWR1443 FMCW radar module, developed based on the
radar and captured by the N receiving antennas, the signal TI datasheet [22], with three transmitting antennas and four
is down-converted separately to produce an in-phase (I) and a receiving antennas, where a maximum of two transmitting
quadrature (Q) baseband components. Each of the N baseband antennas can operate simultaneously, while all receiving an-
signals are sampled by an analog-to-digital converter (ADC) tennas can operate at the same time. The radar supports two
and sent to a computer for further processing. frequency bands within 76 to 81 GHz. IWR1443 FMCW is a
radar-on-a-chip that includes RF sub-system, digital front end,
ramp generator, memory, and radar signal processing unit. The
RF sub-system of the radar module consists of binary phase
modulation (BPM) and power amplifiers (PA) connected to
each transmitting (TX) antenna, while each receiving (RX)
antenna is connected to a low-noise amplifier (LNA), mixer,
IF filtering, and analog-to-digital converter (ADC) chains
followed by decimation stage. The ramp generator linearly
sweeps the voltage input of a voltage-controlled oscillator
(VCO), which in-turns ramps the frequency of a local oscilla-
tor. Then an RF synthesizer generates a wave that processed by
an X4 multiplier generating the desired ramp chirp waveform.
The reflected signal, received by the receiving antenna is
coherently down-converted into IF at the mixer and low pass
filtered (LPF) using the same chirp.
In order to produce the in-phase (I) and quadrature (Q)
components of the beat signal; a 90◦ phase shifter is used.
Fig. 1. A block diagram describing the main elements of the proposed This analog beat signals are then sampled by two separate
material classification framework. ADCs and then fed to the processing unit for further signal
processing.
The collected sample size relies on the number of uti-
lized receiving channels, Therefore, N receiving channels
will generate N pairs of I and Q beat signals denoted as
(I) (Q) (I) (Q) (I) (Q)
follows, x1 , x1 , x2 , x2 , . . . , xN , and xN . Typically,
the ADC samples of the beat signal are numerically stored
as a stack of complex vectors, i.e., a complex matrix X, with
dimension equal to (the number of channels × number of time
samples per frame), to be N × L, where N is the number
of utilized receiving channels and L represents the number
of samples per captured frame. Therefore, X, is represented
(I) (Q)
by X[n, l] = x[l]n + j x[l]n , where n ∈ {1, N } is the
antenna index and l ∈ {1, L} is the time index, representing
the size of received raw signal. For each receiving channel
we further standardize the signals by centering the mean to
zero and bringing the standard deviation to unity. This matrix
Fig. 2. The proposed three approaches for feature extraction and
classification of materials. is utilized for the material classification using the employed
classification methods as discussed in the following Section.
We utilize three different approaches for the material clas-
IV. F EATURE E XTRACTION AND C LASSIFICATION
sification which are depicted in Fig. 2. In the first approach (i)
we employ a 2D CNN to classify the target materials utilizing A. Deep-Learning approach
the received signals in the form of a two-dimensional matrix. Deep-learning based on convolutional neural networks CNN
Whereas, in the second approach (ii) we extract predefined have recently become one of the widely used approaches
4 IEEE SENSORS JOURNAL, VOL. XX, NO. XX, XXXX 2022

maximum classification accuracy, which are listed in Table. I.

B. k-NN
k-NN is a supervised learning model which classifies an
input by identifying the classes of its k nearest data samples
and assigns the class based on the most frequent neighbuoring
class. k-NN is one of the simplest machine learning algorithms
with sometimes surprisingly excellent classification outcome,
where it is proven that as the number of training samples → ∞
and k → ∞, this simple method converges into the Bayesian
classifier [23].
We utilize two different approaches to train the k-NN
classifier; (i) based on the mean and standard deviation of the
captured signals, and (ii) based on the dynamic time warping
Fig. 3. A Simplified block diagram for the utilized TI IWR1443 FMCW distance. These two methods are explained as follows:
radar module based on the radar datasheet [22].
1) k-NN with mean and Standard Deviation Approach: The
collected raw data for a single sample has a dimension of
in time-series analysis and pattern detection. The ability of N × L, where N is the number of receiving channels and
CNN to extract the spatial and temporal features from the L is the number of time samples per frame. This data is
inputs makes it more desirable compared to other machine processed to extract the mean and standard deviation. For a
learning algorithms that require manual extraction of features. single material sample and for a single channel, the feature
In the context of multi-antenna radar application, the spatial vector is of dimension 1×L. The mean and standard deviation
features refer to the relation between the signals received by are calculated for this vector, resulting in a new feature vector
the coherent channels, which capture the azimuthal changes in of dimension 1 × 2. Therefore, for each sample, we have N
the material properties. Whereas the temporal features indicate sets of mean and standard deviation from each received frame
the relation between the ADC time samples, which in-turn of the receiving channel constituting N × 2 features, as shown
capture the radial properties of the material. Filters in the in Fig. 2. Accordingly, the feature table has rows representing
convolutional layers slide through the input signal, extracting the samples and columns representing the features. In order to
the unique features from the inputs, which are then utilized maximize the classification accuracy, the parameter k can be
for classification using a fully connected neural network. The further optimized.
standardized matrix X, of dimension N × L, is fed to the 2) k-NN with Dynamic Time Warping Approach: In this ap-
2D CNN for further feature extraction, as shown in Fig. 2, proach we utilize dynamic time warping algorithm to extract
and then a subsequent fully connected layer for material the similarity of samples among the six target materials. DTW
classification. is an algorithm to find the optimal scaling of the time axis of
two time-sequences by minimizing the cost of matching one
TABLE I of them to the other [24]. Suppose we have two signals X =
CNN NETWORK ARCHITECTURE AS UTILIZED IN S ECTION V [x1 , x2 , . . . , xP ] and Y = [y1 , y2 , . . . , yQ ] of length P and Q
respectively. Let D be a P ×Q matrix which represents a point-
Layer Parameters Value
to-point correspondence relationship between X and Y , where
Input shape - 4x256 D(i, j) indicates the Euclidean distance between elements xi
2D convolutional layer Filters 64 and yj . Then the matching between the points in X and Y are
Filter size 3x3
Activation function ReLu represented in a time warping path W = ⟨w1 , w2 , . . . , wK ⟩,
2D convolutional layer Filters 32
max(P, Q) ≤ K ≥ P + Q, where K is the number of warp
Filter size 2x2 paths with each warp path is given as wk = (i, j). The first
Activation function ReLu warp path is w1 = (1, 1) and the last path is at the end of
Flatten layer - - both signals at wK = (P, Q). Furthermore, warp paths should
Final fully connected utilize all the indices of both the signals. That is,
Units 6
layer
Activation function Softmax wk = (i, j), wk+1 = (i′ , j ′ ), (1)

As it will be explained in Sec. V, the proposed classification where i ≥ i′ ≥ i + 1, j ≥ j ′ ≥ j + 1. The optimum warping
architecture consists of two consecutive convolutional layers, path, DTW(i, j) between X(1:i) and Y(1:j), with the mapping
followed by a flatten layer to reshape the data to fit into path starting from (1,1) to (i,j) is computed as,
the next fully connected layer. The first convolutional layer 
consists of 64 filters with size 3×3 and the second convolu- DTW(i − 1, j),

tional layer consists of 32 filters with a filter size 2×2. The DTW(i, j) = d(xi , yj ) + min DTW(1, j − 1) (2)

parameters of the network are empirically set to obtain the DTW(i − 1, j − 1).

S. SKARIA et al.: PREPARATION OF PAPERS FOR IEEE SENSORS 5

TABLE II in signal as compared to the center). The illuminated area

R ADAR PARAMETERS at a distance 12 cm from the radar will be approximately
28.6 cm × 8.7 cm. Since the identification is based on
Parameter Symbol Value
the RF fingerprint of the reflected signal, the near/far field
Start frequency fstart 77 GHz
should not affect the fingerprint as the material is placed
Number of channels N 4
at a consistent distance from the radar within the radar’s
Chirp duration Tc 40 µs illumination area. The size of the utilized container in this
Chirp bandwidth B 4 GHz experiment is 18 cm × 13 cm. Therefore, the container is
IF bandwidth IFmax 15 MHz larger than the radar’s illumination region to maximize the
Ramp chirp rate S 100 MHz/µs utilization of the RF signal.
ADC sampling rate - 18.75 MHz In the developed experimental setup, one transmitting chan-
Detection range Rd 9.6 m nel and four receiving channels are utilized as shown in
ADC samples per frame L 256 Fig. 5. This experimental setup is applied for all materials.
Chirps per frame - 1 While, the signal added by the shelf and glass container
might contribute to the measurement, it will be ignored by
the classifier and will not have a significant effect on the
experimental measurements [26].
Therefore, the returned value of a fully executed DTW algo- The dataset is collected over multiple sessions spanning over
rithm is a single number representing the minimum distance a week approximately to allow a variety of collection envi-
between the two vectors X and Y . ronment. For every type of material, the physical samples are
The samples can be divided into a reference dataset R divided into two portions. One portion is used for collecting
and test dataset T . We compute the minimum DTW distance training data samples, while the second is used for collecting
between each sample in the test dataset and every sample in testing data samples. The test portion has not seen before by
the reference dataset (for each channel). In our case, we have the classifier. For each of the test and training portions, the
P = Q = L. We represent the DTW distances of a test sample physical samples are further divided into sub-portions. For
from all channels with all reference samples R as a vector of example, a training physical portion is divided randomly into
size Z = N ×R, where R is the number of reference samples. 3 sub portions. On collection session, the sample is placed
By computing the minimum DWT distances for all test dataset inside the container for the collection period then removed.
samples T , we finally get a table with dimension T × Z, The following session, the sample is placed again with other
where T represents the number of samples in test dataset random arrangement.
and Z represents the number of DTW distances computed
In order to enrich dataset, three measurements are taken
per sample. By utilizing this features table of size T × N × R,
from each physical sample per material. Those measurements
as shown in Fig. 2, we train the k-NN to classify the target
are with different random orientation. We have maintained
materials.
consistency in measurement of the samples for all the six
material. Therefore, for each material the orientations across
V. E XPERIMENTAL SETUP the vertical axis is similar. Then the material is shuffled
IWR1443 integrated mmWave FMCW radar sensor module randomly inside the container for the same physical sample.
from Texas Instruments [22] is utilized and adjusted to operate
at the band of 77 to 81 GHz. The radar is placed on a flat table
to collect the experimental dataset. Then on top of the radar,
a metal shelf is placed at a distance 12 cm from the radar
as shown in Fig.4. The utilized sample glass container has a
dimension of 18 cm × 13 cm × 15 cm, where the furthest
samples’ range is 27 cm as shown in Fig.4.
According to [25], the radar’s farfield distance is approx-
imated by 2d2 /λ, where d is the largest physical linear
dimension of the array and λ is the radar wavelength. Based on
the datasheet of the radar module [22], the largest dimension
of the MIMO array is ≈ 4.6λ. For the actual operating fre-
quency of 77 GHz, the wavelength is approximately 0.39 cm.
Consequently, the farfield distance is approximately 16.5 cm
at the operating frequency.
Based on the experimental setup, each material is placed
at a distance 12 cm above the radar on top of the metal
shelf. Therefore, the material is at a transition region be-
tween the nearfield and the farfield when considering MIMO
array. While, the utilized radar module has a ±50◦ azimuth
beamwidth and ±20◦ elevation beamwidth (based on 6dB drop Fig. 4. Experimental setup measurements and material orientation.
6 IEEE SENSORS JOURNAL, VOL. XX, NO. XX, XXXX 2022

Fig. 5. The 77 GHz FMCW radar development kit used for data
collection in the experiment.

For the utilized FMCW radar, with a maximum peak power

of 12 dBm at the operating frequency range of 77 to 81 GHz, Fig. 6. The experiment setup to capture signals reflected from the target
we utilize one transmitting channel and four receiving channels materials.
of radar for collecting the data of different materials. The
employed radar module is a complete system-on-chip shown
in Fig. 5. The FMCW radar chipset starting frequency fstart is has different response for the same transmitted waveform, to
set to 77 GHz. Accordingly, the sweep bandwidth B is 4 GHz. be classified using different proposed approaches.
The chipset supports 15 MHz IF bandwidth. The selected In order to verify the performance of the utilized classi-
chirp duration Tc is 40 µs. Therefore, the ramp chirp rate fication framework, we collect 60 samples for each of the
is S = TBc = 100 MHz/µs and accordingly, the maximum three volume levels of each of the six materials. Therefore,
detectable range is given by Rangemax = IF2×S max ×c
,where we obtain a total of 60 × 3 × 6 = 1080 samples. We
IFmax represents the maximum IF bandwidth supported, c is collect the samples which are quarter-filled, half-filled, and
the speed of light, and S is ramp chirp rate, which allows a full-filled, by introducing intentional inconsistencies in the
maximum range of 22.5 m. The ADC sampling rate is chosen position and orientation with respect to the radar sensor. To
as 18.75 MHz. verify the proposed framework, we examined three different
The captured frame is set to contain a single chirp with combinations of the dataset;
detecting range Rd of 9.6 m. The resulting number of time • Case 1: Dataset with only full-filled volume of the six
samples per chirp is 2×Rcd ×B = 256, which equal the time materials. (360 samples)
samples per frame L (as frame includes only one chirp). • Case 2: Dataset with quarter-filled, half-filled, and full-
The resultant raw data from the four receiving channels will filled volume for the six material classes. (1080 samples)
develop a matrix of size 4 × 256, considered as the input • Case 3: Dataset with quarter-filled, half-filled, and full-
for different identification approaches. The utilized parameters filled volume of the six materials as separate classes
of the radar module for collecting the data are described in to detect type and volume of the material. The dataset
Table. II. contains 6 × 3 = 18 classes, which are the three volumes
The experimental setup is shown in Fig. 6. The selected ma- of the six materials. (1080 samples)
terials for this study are; (i) cardboard, (ii) sand, (iii) sponge, Accordingly, the classification results of each of the utilized
(iv) polystyrene, (v) water, and (vi) wood. Those materials are approaches are discussed below:
chosen to examine a wide range of reflective material with
• Approach 1: For the classification process using the deep-
different refractive index, thickness, and dielectric constant.
learning approach, we randomly divide the dataset into
For the volume classification, we collect the samples based
two portions, 80% for training and the remaining 20%
on three-volume levels; (1) quarter-filled, (2) half-filled, and
for testing (unseen dataset). The training dataset is again
(3) full-filled. The utilized materials are shown in Fig. 6. While
divided into 80% for iteration training and 20% for
taking the samples, we made sure that the orientation and
validation. The testing data is rotated five times, and
position of the materials are arbitrary varied with respect to the
for each rotation, the training process is done using 5-
radar sensor. This will enhance the variations in the collected
fold cross-validation. For datasets in Case 1 and Case
data, which will aid the training of the classification network
2 the classification architecture is as shown in Table. I.
in practical scenarios.
For Case 3 the final fully connected layer has 18 units
corresponding to the 18 classes of materials/volumes. The
VI. E XPERIMENTAL R ESULTS AND D ISCUSSION obtained mean classification accuracy for Case 1, 2, and 3
Fig.7 represent a sample of the received time domain are 97.0%, 98.0%, and 98.2% respectively. We compare
signals (ADC samples) from the four receiving channels of the the CNN classifier with the same architecture but the
FMCW radar. From the plot, it could be seen that each material input signals are fed without standardizing explained
S. SKARIA et al.: PREPARATION OF PAPERS FOR IEEE SENSORS 7

TABLE III
P ERFORMANCE OF THE UTILIZED MATERIAL CLASSIFICATION
APPROACHES

Approach Case 1 [%] Case 2 [%] Case 3 [%]

Deep-learning 97.0 ± 2.1 98.0 ± 1.5 98.2 ± 1.4
k-NN 95.5 ± 3.4 96.6 ± 2.4 92.3 ± 5.6
DTW - k-NN 100 ± 0.0 96.2 ± 2.7 90.2 ± 7.3

each material. Case 3 has the least performance due to the

similarities in DTW distances within the same materials
with different volume levels. The classification accuracy
of this approach for Case 1 is significantly higher com-
pared to the previous k-NN-based classification with the
simple mean and standard deviation features.
The error bounds ϵ are calculated based onqthe obtained
success probability estimate β̂ using, ϵ = zp β̂(1− n
β̂)
[7],
where zp denotes the inverse-CDF of Standard Normal Distri-
bution calculated at the probability 1− 1−p2 with p confidence.
The mean classification accuracy obtained for all the three
cases using the proposed classification networks are depicted
in Table. III at a confidence of 95%. This accuracy indicates
the relation between the error bounds and the training data
size.
The classification performance of Case 1 is the highest for
the k-NN with DTW. The reason being is the ability of DTW
to find the best matching between the signals of the same
class. For Case 2 and 3, CNN-based classifier outperforms
the other networks as the signal seems to less structured.
By standardizing the signals, we are able to avoid irrelevant
Fig. 7. An example sample vector of the time domain signals received
by the four receiving channels of the radar module which are reflected absolute variations in the signal level, which results in a better
from the target materials. falsification accuracy.

VII. C ONCLUSION AND F UTURE W ORK

in Section III. The classification performance obtained In this paper, the signal received from a multi-antenna
for Case 1, 2 and 3 are 96.2%, 96.0%, and 90.5% mmWave FMCW radar is utilized for the classification of
respectively, indicating that standardizing the signals prior various types of materials. Three different approaches are
to the CNN enhances the classification performance. presented by utilizing the time domain signals received from
• Approach 2: For the second approach with the k-NN four channels of a miniature mmWave radar module. A near
classifier with mean and standard deviation as inputs, we ideal classification accuracy is obtained to classify six different
vary the k to optimize the classification accuracy. The materials while developing a new way of implementing DWT
feature table is typically divided into 80 % as reference for material identification. Furthermore, the proposed frame-
and 20 % for testing (unseen dataset). For all the three work is implemented to identify the geometric volume of the
cases, the maximum accuracy is obtained for k = 1 target material and to obtain a classification accuracy above
indicating that the training set is sufficiently large to 98% in classifying three different volume levels of the six
produce very close neighbors. The classification accuracy materials. With this high classification accuracy, the proposed
obtained for cases 1, 2, and 3 are 95.5%, 96.6%, and framework shows the potential for practical applications in
92.3% respectively. noninvasive material detection. The proposed framework is
• Approach 3: In the k-NN with DTW method, similar utilizing a COTS device, applicable to any off-the-shelf radar
to the previous classification method, the samples are that can produce range-Doppler measurements. Future work
divided into a reference dataset R and test dataset T might include exploring more materials and liquids while
(unseen dataset) with 80% / 20% split. The classification investigating additional deep learning methods.
accuracy obtained for Case 1, 2, and 3 are 100, 96.2%,
and 90.2% respectively. We observe the highest accuracy R EFERENCES
for Case 1, as DTW distance seem to show minimal
[1] A. A. Bunaciu, E. gabriela Udriştioiu, and H. Y. Aboul-Enein, “X-
variations for materials with the same volume, which ray diffraction: Instrumentation and applications,” Critical Reviews in
gives a distinct pattern of the DTW distance vector of Analytical Chemistry, vol. 45, no. 4, pp. 289–299, Apr 2015.
8 IEEE SENSORS JOURNAL, VOL. XX, NO. XX, XXXX 2022

[2] J. Chan, P. Evans, X. Wang, S. Godber, I. Peatfield, K. Rogers, J. Rogers, [22] IWR1443 Single-Chip 76- to 81-GHz mmWave Sensor, Texas
and A. Dicken, “Scatter enhanced 3D X-ray imaging for materials Instruments, 2018, rev. C. [Online]. Available: https://www.ti.com/lit/
identification,” in 44th Annual 2010 IEEE International Carnahan ds/symlink/iwr1443.pdf
Conference on Security Technology, 2010, pp. 142–147. [23] S. Theodoridis, Machine Learning: A Bayesian and Optimization Per-
[3] T. D. T. Oyedotun, “X-ray fluorescence (XRF) in the investigation of spective, 1st ed. USA: Academic Press, Inc., 2015.
the composition of earth materials: a review and an overview,” Geology, [24] B. Stasiak, M. Skiba, and A. Niedzielski, “FlatDTW – dynamic time
Ecology, and Landscapes, vol. 2, no. 2, pp. 148–154, 2018. warping optimization for piecewise constant templates,” Digital Signal
[4] C. Fiorini, A. Longoni, M. Milazzo, and F. Zaraga, “In-situ, non- Processing, vol. 85, pp. 86–98, Feb 2019.
destructive identification of chemical elements by means of portable [25] C. A. Balanis, Modern antenna handbook. John Wiley & Sons, 2011.
edxrf spectrometer,” in 1998 IEEE Nuclear Science Symposium Con- [26] C. M. Bishop and N. M. Nasrabadi, Pattern recognition and machine
ference Record. 1998 IEEE Nuclear Science Symposium and Medical learning. Springer, 2006, vol. 4, no. 4.
Imaging Conference (Cat. No.98CH36255), vol. 1, 1998, pp. 375–380
vol.1.
[5] G. Dobmann, I. Altpeter, C. Sklarczyk, and R. Pinchuk, “Non-destructive Sruthy Skaria is currently a Data Science Spe-
testing with micro- and MM-waves — where we are — where we go,” cialist at Telstra, Melbourne, Australia. She com-
Welding in the World, vol. 56, no. 1-2, pp. 111–120, Jan 2012. pleted her Ph.D. degree in 2021 from the School
[6] A. Al-Hourani, R. J. Evans, P. M. Farrell, B. Moran, M. Martorella, of Engineering, RMIT University, Melbourne,
S. Kandeepan, S. Skafidas, and U. Parampalli, “Chapter 7 - millimeter- Australia, and her Masters in Electronics Engi-
wave integrated radar systems and techniques,” in Academic Press Li- neering in 2015. Her current research interests
brary in Signal Processing, Volume 7, R. Chellappa and S. Theodoridis, include consumer radars and their applications
Eds. Academic Press, 2018, pp. 317–363. in the fields of automotive, gesture recognition,
[7] S. Skaria, A. Al-Hourani, M. Lech, and R. J. Evans, “Hand-gesture material identification, interference issue in con-
recognition using two-antenna doppler radar with deep convolutional sumer radars, and applications of neural net-
neural networks,” IEEE Sensors Journal, vol. 19, no. 8, pp. 3041–3048, work in radar technology. Her publication ‘Hand-
Apr 2019. Gesture Recognition Using Two-Antenna Doppler Radar with Deep
[8] J. Barowski, J. Jebramcik, J. Wagner, N. Pohl, and I. Rolfes, “Spatial Convolutional Neural Networks’ has received the best paper award
identification of dielectric properties using synthetic aperture radar,” in runner-up from IEEE Sensors Journal in 2020, and one of the top 25
IEEE MTT-S International Microwave Workshop Series on Advanced most downloaded papers since then.
Materials and Processes for RF and THz Applications, 2019, pp. 139–
141.
[9] R. Suwalak, C. Phongcharoenpanich, D. Torrungrueng, and P. Akkaraek-
thalin, “Dielectric material determination using the radar equation in
RFID sensor applications,” in 2015 IEEE Conference on Antenna Nermine Hendy (Member, IEEE) is currently
Measurements Applications, 2015, pp. 1–4. pursuing her PhD degree in the discipline of
[10] P. Sotirelis, J. T. Parker, M. Fu, X. Hu, and R. Albanese, “A study of Electrical and Electronic Engineering, School of
material identification using SAR,” in IEEE Radar Conference, 2012, Engineering, RMIT University, Melbourne, VIC.
pp. 0112–0115. She has completed her master-by-research de-
[11] M. Gilman and S. Tsynkov, “Detection of material dispersion using gree in Electrical and Electronic Engineering
sar,” in EUSAR 2014; 10th European Conference on Synthetic Aperture in 2013. Her current PhD research focuses on
Radar, 2014, pp. 1–4. the applications of Artificial Intelligence and ad-
[12] R. Suwalak, C. Phongcharoenpanich, and D. Torrungrueng, “Chipped vanced signal processing techniques for minia-
and chipless RFID sensors for quality monitoring of light weight ture radars and on interference detection and
concrete using the radar equation,” in International Conference on mitigation for SAR radars.
Electrical Engineering/Electronics, Computer, Telecommunications and
Information Technology, 2016, pp. 1–4.
[13] X. L. Travassos, S. L. Avila, and N. Ida, “Artificial neural networks
and machine learning techniques applied to ground penetrating radar: A
review,” Applied Computing and Informatics, vol. 17, no. 2, pp. 296– Akram Al-Hourani is an Associate Professor
308, 2021. and the Program Manager for the Master of
[14] B. Jamali, D. Ramalingam, and A. Babakhani, “Intelligent material Engineering (Telecommunication and Networks)
classification and identification using a broadband millimeter-wave fre- at the School of Engineering, RMIT University,
quency comb receiver,” IEEE Sensors Letters, vol. 4, no. 7, pp. 1–4, Melbourne, Australia. Dr Al-Hourani completed
2020. the Ph.D. degree in 2016 from RMIT University.
[15] H.-S. Yeo, G. Flamich, P. Schrempf, D. Harris-Birtill, and A. Quigley, He published more than 89 journal articles and
“RadarCat,” in Proceedings of Annual Symposium on User Interface conference proceedings, including 3 book chap-
Software and Technology. ACM, Oct 2016. ters. In 2020, Dr Al-Hourani has won the IEEE
[16] O. Adedeji and Z. Wang, “Intelligent waste classification system using Sensors Council Paper Award for his contribu-
deep learning convolutional neural network,” Procedia Manufacturing, tion in hand-gesture recognition using neural
vol. 35, pp. 607–612, 2019. networks. He has extensive industry / government engagement as a
[17] H. Zheng, L. Fang, M. Ji, M. Strese, Y. Özer, and E. Steinbach, chief investigator in multiple research projects related to The Internet-
“Deep learning for surface material classification using haptic and visual of-Things (IoT), Smart Cities, Satellite / Wireless Communications. As
information,” IEEE Transactions on Multimedia, vol. 18, no. 12, pp. a Lead Chief Investigator, he oversaw the design and deployment of
2407–2416, 2016. the largest open IoT network in Australia in collaboration with 5 local
[18] S. Skaria, A. Al-Hourani, and R. J. Evans, “Deep-learning methods governments “Northern Melbourne Smart Cities Network”, this project
for hand-gesture recognition using ultra-wideband radar,” IEEE Access, has won the 2020 “IoT Awards”, the official awards program of IoT
vol. 8, pp. 203 580–203 590, 2020. Alliance Australia. Prior his academic career, between 2006 - 2013,
[19] G. Agresti and S. Milani, “Material identification using RF sensors Dr Al-Hourani had extensively worked in the ICT industry sector as an
and convolutional neural networks,” in International Conference on R&D engineer, radio network planning engineer and then as an ICT pro-
Acoustics, Speech and Signal Processing. IEEE, may 2019. gram manager for several projects spanning over different technologies;
[20] M. M. Khan, K. M. Iftekharuddin, E. McCracken, K. Islam, S. Bhurtel, including mobile networks deployment, satellite networks, and railway
L. Wang, and R. Kozma, “Autonomous wireless radar sensor mote for ICT systems. Dr Al-Hourani is serving as an Associate Editor in IEEE
target material classification,” Digital Signal Processing, vol. 23, no. 3, Transaction on Aerospace and Electronic Systems, in Frontiers in Space
pp. 722–735, May 2013. Technologies, Frontiers in Communications and Networks, and is an
[21] H. Ku, J. Ball, E. Siores, and B. Horsfield, “Microwave processing Editor in MDPI Remote Sensing. His current research interests include
and permittivity measurement of thermoplastic composites at elevated Satellite and UAV Communications, Spaceborne Radar, Automotive
temperature,” Journal of Materials Processing Technology, vol. 89-90, and mmWave Radar, energy efficiency in Wireless Networks, and the
pp. 419–424, 1999. Internet-of-Things over Satellite.