This full text paper was peer-reviewed at the direction of IEEE Instrumentation and Measurement Society to the acceptance

and publication.

Anomaly Detection of Spacecraft Telemetry Data

Using Temporal Convolution Network
Yuan Wang, Yan Wu Qiong Yang, Jun Zhang
School of Electronics and Information Engineering Institute of Navigation Satellite
Harbin Institute of Technology Innovation Academy for Microsatellites of CAS
Harbin, China Shanghai, China
{wangyuan4003, wuyanhit}@hit.edu.cn {yangq, zhangjun}@microsate.com

Abstract—Spacecraft telemetry data is the only basis for the For on-orbit spacecraft, in order to monitor the internal
2021 IEEE International Instrumentation and Measurement Technology Conference (I2MTC) | 978-1-7281-9539-1/21/$31.00 ©2021 IEEE | DOI: 10.1109/I2MTC50364.2021.9459840

ground management and operation system to determine its on- operating status and obtain real-time data, the information is
orbit status. The rapid development of on-orbit spacecraft measured by sensors in the spacecraft telemetry system, which
generates a large amount of telemetry data with high is converted into electrical signals. After the signals are
dimensionality and complex correlation between variables, combined according to a certain rule, the radio
which brings great challenges for current and future ground communication technology is used to transmit signals to the
management and operation for these spacecrafts. As a result, its ground telemetry equipment which includes receivers,
anomaly detection becomes an essential solution for enhancing antennas, and split demodulators [1]. The ground equipment
the operation safely and reliably and realizing intelligent
restores the original variable information of each channel
maintenance. The above issues make it a hot spot research topic
for realization of high detection rate, low false detection rate,
through the signal demodulation technology. Those variables
and strong interpretability of anomaly detection of telemetry obtained in this case is defined as the spacecraft telemetry data
data. In further, timely detection of abnormalities can reduce [2]. With the increasement of society requirements, the
spacecraft unexpected accidents. Recently, Temporal number of spacecrafts increases gradually so that the
Convolution Network (TCN) has arisen much attention for dimension of telemetry data which needs real-time monitoring
telemetry time series data modeling by its good parallel expands extremely. It is a challenge to monitor massive data
processing ability and temporal characterization. Thus, this by humanity and the experience of experts. Therefore,
paper proposes an anomaly detection model for spacecraft analyzing the dataset is of benefit to the transition from
telemetry data based on TCN time series prediction. Firstly, collecting, transmitting, storing data to mine the value of data.
TCN algorithm is used to extract features of telemetry data and Telemetry data are multidimensional time series [3]. The
make a prediction. Then, the potential anomalies are anomaly detection for multivariate telemetry data draws much
determined by the static threshold judgement method. Actual attention on time series prediction methods in recent years.
satellite telemetry data are used to implement performance
evaluation and comparison. The proposed method is compared At present, the time series prediction methods based on the
with Long Short Term Memory (LSTM) algorithm which is also data-driven model [4, 5] are developed rapidly. The earliest
a widely utilized algorithm in recent time series analysis. Four use of data-driven model is Autoregressive (AR) which acts
metrics are adopted for the experiments: the detection accuracy, on linear time series, but there are many parameter restrictions
detection rate, false detection rate and model inference time. for stationary time series. The introduction of the differential
TCN achieves high accuracy, high detection rate and low false process does not require conditional restrictions to detect
detection rate. In addition, TCN save much more model weakly stationary time series data. Box and Jenkins et al. [6]
inference time than LSTM with long timestep. Generally, TCN proposed Autoregressive Moving Average (ARMA) in 1920s.
is superior to LSTM on operating efficiency. In terms of strongly stationary time series, Box and Jenkins et
al. [7] proposed Autoregressive Integrated Moving Average
Keywords—anomaly detection, telemetry data, TCN, LSTM Model (ARIMA) in 1970s. From the evolution of statistical
I. INTRODUCTION models, the AR model makes a great contribution. As the
diversity of data features increases, machine learning occupies
Spacecraft is a complex system which makes great the main position in prediction in recent years. Back-
contributions to meteorological observation, military Propagation (BP) is the classical algorithm in machine
reconnaissance and resource survey, etc. However, with the learning. Zhao et al. [8] predicted toxic activity on different
gradual improvement of the design complexity and datasets based on Support Vector Machine (SVM).
functionality of spacecraft, it is a challenge to process and Rasmussen et al. [9] proposed Gaussian Process Regression
analyze data for safe and reliable operation of spacecrafts only (GPR) which integrated statistics and machine learning. Both
through traditional reliability engineering methods [1]. SVM and GPR can support not only one-dimensional data, but
Therefore, it is necessary to monitor abnormal data which are also multidimensional data. However, the above models do
different from normal data or not in line with the spacecraft not achieve a good performance with large amount of complex
working mode setting. Timely and effectively discover data. Big data increase computation complexity and occupy
abnormal patterns, and repair or troubleshoot in time remote amount of resources. The propose of Recurrent Neural
command repair, transmission link repair, which is important Network (RNN) [10] effectively decreases the requirement of
for improving the quality of ground service, maintaining the computation resources, but traditional RNN often occurs
maturity, safety and reliability on the design and production vanishing gradient or explosion when analyzing long time
of the spacecraft have significant practical meaning. Thus, the series. Long Short Term Memory (LSTM) [11], as a special
main purpose for anomaly detection is to detect the RNN structure, is proposed to solve the above problems. Also,
components failed, evaluate health statues and warn in LSTM reserves advantages of RNN and has the ability of
advance. remembering long-term series. Due to the fact that LSTM

978-1-7281-9539-1/21/$31.00
Authorized ©2021 Technological
licensed use limited to: Visvesvaraya IEEE University Belagavi. Downloaded on March 30,2024 at 08:45:49 UTC from IEEE Xplore. Restrictions apply.
 merely parses one batch of data, this model consumes long K

time. Bai et al. [12] uses Convolution Neural Network (CNN)  F xT ¦f k 1

x
k T K k  
on time series to gain parallel ability and large receptive field
which is the important factor for temporal characterization.
The proposed method is named as Temporal Convolution output
Network (TCN). Large receptive field and parallel ability
greatly shorten the model inference time.
hidden layer
TCN is widely used in time series in recent years and has
been applied to anomaly detection. However, anomaly
detection of multidimensional telemetry data from spacecraft hidden layer
based on TCN has not been researched. It is meaningful to do
such a research due to the complexity of the correlation
between spacecraft telemetry data. The rest of this paper is input
organized as follows. Section II reviews the related x1 x2 Ċ xT
methodologies of TCN. Section III introduces the proposed
anomaly detection model. Section IV does an experiment with Fig. 2. The illustration of causal convolution.
NASA data by using the proposed algorithm, and then
analyzes the results from two aspects: the detection accuracy C. Dilated Convolution
and the model inference time. Section V concludes this paper The modeling length of simple causal convolution on time
and introduces the future work. is limited by the size of convolution kernel, which is a problem
II. RELATED METHODOLOGIES traditional CNN exists. Stacking layers linearly increases
receptive fields and generates long-time dependence. Dilated
TCN is a specific CNN which is suitable for time series convolution can reduce the depth of simple causal convolution.
analysis. It includes four fundamental structures, namely one- Figure 3 shows the dilated convolution with the dilated factors
dimension (1-D) convolution, causal convolution, dilated d is equal to 1, 2 and 4 respectively between each layer. The
convolution and the residual connection [13]. expression of the dilated convolution is as follows:
A. 1-D Convolution Network K
Under the multi-variable input, the input and output  Fd xT ¦f x
k T  K k d  
usually have different size. Figure 1 intuitively illustrates how k 1
1-D convolution network works with three kernel size. In
order to get the output, take the dot product of the input It is common to expand the dilated factors and increase the
subsequence and the kernel vector of the same length with the kernel size to enhance receptive fields.
learned weights. In order to make the visualization easier, the output
following figures no longer show the dot product with the d=4
kernel vector. hidden layer

d=2
output
hidden layer

d=1
· · · dot product
input

x1 x2 · · · xT

input Fig. 3. The illustration of causal dilated convolution network

kernel size D. Residual Connection

The residual connection is proved to be an effective
Fig. 1. The illustration of 1-D convolution network.
method for training deep networks, which enables the network
B. Causal Convolution to transmit information in a cross-layer manner. The diagram
of the residual connection is given as Fig. 4. This paper
TCN has two essential principles. One is that the size of constructs one residual block with two-layer dilated causal
the input and output are the same. As for the constraint, zero- convolution and ReLU activation function which introduces
padding can ensure each layer have the same length. Another non-linearity. WeightNorm and dropout are used for
is to guarantee no future information leakage occurs. Causal regularization. The output result of residual block can be
convolution can implement the requirement. In TCN, the expressed as 
convolution operation is performed strictly in time sequence,
that is, the convolution operation at time t only occurs on Fd ( xT )  xT 
 F ( xT ) 
the data and the previous layer before the time t 1 . The
illustration of causal convolution is shown in Fig. 2 with three
kernel size and two hidden layers.y Suppose that the input
sequence is X [ x1 , x2 , , xT ] , and convolution kernel is
defined as F [ f1 , f 2 , , f K ] , where K is the size of
convolution kernel. The causal convolution at xT is
expressed as follows:

Authorized licensed use limited to: Visvesvaraya Technological University Belagavi. Downloaded on March 30,2024 at 08:45:49 UTC from IEEE Xplore. Restrictions apply.
 Input for next down
residual block H th* =P e  z G e (5)
+ Residual block (k,d)
where P e is the mean of e . G e is the standard
Dropout
deviation of e . z is the coefficient to trade off mean and
ReLU standard deviation. Generally, z is the confidence level under
the 99.7% confidence interval which obeys normal
WeightNorm up

Optional 1×1
distribution. If the error exceeds the upper threshold H th*
Dilated Causal Conv Conv down
or below the down threshold H th* , the data at this
Dropout
position is considered as an abnormal value.
ReLU
IV. EXPERIMENTS AND RESULTS ANALYSIS
WeightNorm
A. Data Description and Evaluation Metrics
Dilated Causal Conv A set of real telemetry datasets of solid state power
Output from previous amplifier which are transmitted from two different satellites
residual block are used for experiments to evaluate the performance of TCN.
Fig. 4. The diagram of residual connection The variables of experimented data from the first satellite are
voltages and currents, and the data vary in a periodic trend.
III. ANOMALY DETECTION BASED ON TCN TIME SERIES There are enough samples to show the periodicity in the
PREDICTION training datasets. The variables of experimented data from the
other satellite are currents and temperatures, and the data are
The anomaly detection model of spacecraft telemetry data
in a long-term trending variation. Also, there are enough
based on TCN time series prediction is focused on in this
samples to show the rising or descending trend in the training
section. A framework is given in Fig. 5, and there are two main
datasets. In addition, two common types of anomalies are
units: the prediction unit and the anomaly detection unit. The
injected into the real data respectively to simulate anomalies:
main TCN algorithm is described in Section II. Before training
point anomalies and contextual anomalies. Point anomalies
and testing, the datasets need to be normalized. The proposed
are individual data points which are different from remaining
anomaly detection method is illustrated as follows.
normal data points among whole spacecraft data. The
abnormal data where the associated environment between
Train datasets Prediction Unit
Train TCN/LSTM upper and lower of telemetry data is time are contextual
anomalies. Figure 6 gives an intuitive visualization of points
Training
Telemetry
Datasets
Data
Preprocessing Model and contextual anomalies.
Metrics are important factors to evaluate the performance
Test
Test datasets
TCN/LSTM
of models. Accuracy (ACC), True Positive Ratio (TPR) and
False Positive Ratio (FPR) [14] are the common metrics used
Predicted Data
ŷt
for the anomaly detection. The propose of TCN is to save
Error e=ŷt -yt
Threshold
Anomaly interval
Performance model inference time, so the time consume is also an
caculation evaluation
important indicator. Specifically, time consume is the running
Original Data yt
Anomaly detection Unit time to predict the test data.

Fig. 5. The framework of anomaly detection ACC, TPR and FPR can be expressed as follows:

The predicted telemetry values are gained from the TP  TN

ACC (6)
training and testing phase. The statistics method is commonly TN  FN  TP  FP
used to detect anomalies. Because the original data are not
stable, it is not suitable for static threshold method. Thus, the TP
TPR (7)
residual between actual data and predicted data is applicable FN  TP
to the mentioned method due to its stability. At first, calculate
the difference between the actual telemetry values and the FP
FPR (8)
predicted telemetry values to obtain a set of error sequence TN  FP
e(et y t  yt ) . Then, the error sequence is compared with where TP is the number of detected anomaly values in
upper threshold and lower threshold which are constant. The abnormal samples. FP represents the normal data which is
mean and variance of the residual from training data are detected as abnormal data. TP  FN is the number of all
essential statistics of calculating thresholds. Taking an actual abnormal data. FP  TN is the actual normal samples.
t
example of e which is the residual sequence between TN  FN  TP  FP is the total number of all samples for
verification.
original data and predicted data in training phase, its
corresponding threshold H th*
n
is defined as following: B. Experimental Results and Analysis
Due to the fact that LSTM is a widely used algorithm in
up time series anomaly detection with excellent performance, a
H th* =P e  z G e (4) set of comparative experiments is implemented based on the

Authorized licensed use limited to: Visvesvaraya Technological University Belagavi. Downloaded on March 30,2024 at 08:45:49 UTC from IEEE Xplore. Restrictions apply.
 LSTM algorithm. Reference [15] introduced its theories
detailly. In the experiment, considering different types of
telemetry data for universality of results, firstly, take
randomly three sets of telemetry data with short previous
timestep. Specifically, the parameter timestep is set to 20, and
take the average (AVG) of metrics of ACC, TPR, FPR and
model time consume. Take the battery voltage as an example,
its visual results of TCN and LSTM are shown as Fig. 6 and
Fig. 7 separately. The red dotted line represents predicted data,
and the blue solid line represents actual data. These two
figures state the ability of TCN to predict the periodicity of
telemetry data is as good as LSTM. Also, the difference of
predicted data and actual data based on TCN is obvious in
terms of anomalies, contextual anomalies in particular.
In order to obtain anomalies directly, the residuals of
predicted data and actual data are calculated. Then, the upper
and lower thresholds are obtained with the formular (5)
through the residual statistics. The residual curve of predicted Fig. 8. Residual of voltage 1 based on TCN and LSTM respectively
and actual values is shown in Fig. 8. The red line represents
the curve of residuals and the black line represents the upper It can be seen from the Table I, TCN has an excellent
threshold and down threshold. Obviously, TCN obtains performance like LSTM in terms of ACC, TPR and FPR.
prominent anomalies. However, it is difficult to observe the These three metrics indicate that the detection performance is
positions and the specific number of detected anomalies from in high accuracy, high detection rate and low false detection.
the pictorial results. The metrics of telemetry data with short In addition, both methods achieve a good performance in
timestep based on TCN and LSTM are shown on Table I, model inference time, neither of which exceeds 1s. However,
which proves the performance of models convincingly. In the parallel ability of TCN is not reflected from the above
addition, metrics can quantify the points which exceed the experiments. There is no verification on high operating
upper and lower boundaries in Fig.8 to determine actual efficiency of TCN. The reason of this occurrence is that the
anomalies and false detection reflected from TPR and FPR of amount of previous data, i.e., timestep, is not as enough to
voltage1 in Table I. TPR of voltage1 based on TCN is higher demonstrate the parallel feature. In order to explore the
than that based on LSTM, and FPR of voltage1 based on TCN parallel function and the learning ability of TCN, two sets of
is lower than that based on LSTM, which demonstrates all experiments with large amount of data are conducted to ensure
actual anomalies and fewer false detected points break the sufficient training in the case of long timestep. The timestep is
boundaries based on TCN. set to 200 in this experiment. This set of data is current
telemetry data of solar panels with a descent trend. Figure 9
and 10 show the predicted data and actual test data as well as
residuals based on LSTM and TCN respectively. By checking
Fig.9, both methods have a good regression with a descend
trend. In Fig.10, it is obvious that both methods detect two
parts of anomalies which exceed the upper and lower
boundaries. The metrics of large telemetry data is given in
Point anomaly Table II.

TABLE I. METRICS OF ANOMALY DETECTION WITH SHORT

TIMESTEP
Contextual anomaly
Telemetry data
Model Metrics
Voltage1 Voltage2 Current1 AVG
ACC 94.86% 94.86% 97.64% 95.79%
Fig. 6. Anomaly detection of voltage 1 based on TCN
TPR 98.18% 95.45% 100% 97.88%
LSTM
FPR 5.76% 5.18% 2.52% 4.49%

Op.Time 0.098s 0.102s 0.283s 0.161s

ACC 96% 93.4% 100% 96%

TPR 100% 95.45% 100% 98%

TCN
FPR 3.73% 6.7% 0 3.48%

Op.Time 0.198s 0.125s 0.398s 0.240s

Fig. 7. Anomaly detection of voltage 1 based on LSTM

Authorized licensed use limited to: Visvesvaraya Technological University Belagavi. Downloaded on March 30,2024 at 08:45:49 UTC from IEEE Xplore. Restrictions apply.
 Moreover, TCN consumes less model inference time no
matter with short timestep or long timestep due to its parallel
ability. Generally, TCN can be applied to the anomaly
detection of spacecraft telemetry data with good detection
performance even high operating efficiency. Actually, the
proposed method is currently attempted to Beidou satellites
developed by Innovation Academy for Microsatellites of CAS.
More detailed test results will be reported in the next couple
of months to indicates its potential applicable capabilities.
Actual engineering applications generally need online
detection method due to the telemetry data of satellite is
transmitted as a real-time dataflow. The high accuracy and
operating efficiency of TCN are suitable for online detection.
In further, the proposed method will be improved for real
Fig. 9. Anomaly detection of current based on TCN and LSTM respectively satellite online anomaly detection, i.e., multiple variables of
telemetry data, and complex anomaly modes.
REFERENCES
[1] X. Peng, J. Pang, Y. Peng, and D. Liu, "Review on anomaly detection
of spacecraft telemetry data," Chinese Journal of Scientific Instrument,
vol. 37, no. 9, pp. 1929-1945, 2016 (in Chinese).
[2] K. Hundman, V. Constantinou, C. Laporte, I. Colwell, and T.
Soderstrom, "Detecting spacecraft anomalies using lstms and
nonparametric dynamic thresholding," in Proceedings of the 24th ACM
SIGKDD international conference on knowledge discovery & data
mining, 2018, pp. 387-395.
[3] A. Doudkin, Y. Marushko, J. Owsiński, and T. Pawlowski, "Spacecraft
Telemetry Time Series Forecasting With Ensembles of Neural
Networks," in 2019 10th IEEE International Conference on Intelligent
Data Acquisition and Advanced Computing Systems: Technology and
Applications (IDAACS), 2019, vol. 2: IEEE, pp. 752-756.
[4] B. Wang, X. Peng, M. Jiang and D. Liu, "Real-Time Fault Detection
for UAV Based on Model Acceleration Engine," in IEEE Transactions
Fig. 10. Residual of current based on TCN and LSTM respectively
on Instrumentation and Measurement, vol. 69, no. 12, pp. 9505-9516,
Dec. 2020.
TABLE II. METRICS OF ANOMALY DETECTION WITH LONG TIMESTEP
[5] B. Wang, D. Liu, Y. Peng and X. Peng, "Multivariate Regression-
Telemetry data Based Fault Detection and Recovery of UAV Flight Data," in IEEE
Model Metrics Transactions on Instrumentation and Measurement, vol. 69, no. 6, pp.
Current2 Temperature AVG
3527-3537, June 2020.
ACC 99.42% 95.56% 97.49%
[6] G. E. Box, G. M. Jenkins, G. C. Reinsel, and G. M. Ljung, Time series
TPR 99.46% 99.73% 99.60% analysis: forecasting and control. John Wiley & Sons, 2015.
LSTM [7] P. J. Brockwell, R. A. Davis, and M. V. Calder, Introduction to time
FPR 0.58% 4.80% 2.69% series and forecasting. Springer, 2002.
Op.Time 8.198s 1.401s 4.800s [8] C. Zhao, H. Zhang, X. Zhang, M. Liu, Z. Hu, and B. Fan, "Application
of support vector machine (SVM) for prediction toxic activity of
ACC 99.07% 93.82% 96.45% different data sets," Toxicology, vol. 217, no. 2-3, pp. 105-119, 2006.
TPR 97.30% 96.43% 96.865% [9] C. K. Williams and C. E. Rasmussen, Gaussian processes for machine
TCN learning (no. 3). MIT press Cambridge, MA, 2006.
FPR 0.71% 6.40% 3.56% [10] B. Bakker, "Reinforcement learning by backpropagation through an
LSTM model/critic," in 2007 IEEE International Symposium on
Op.Time 0.383s 0.188s 0.571s Approximate Dynamic Programming and Reinforcement Learning,
2007: IEEE, pp. 127-134.
Table II indicates that the average of ACC and TPR based [11] F. Gers, "Long short-term memory in recurrent neural networks," Ph.D.
thesis, Universität Hannover, 2001.
on both models exceed 95% and the average of FPR is below
5%, which remains good performance. The discrepancy of [12] S. Bai, J. Z. Kolter, and V. Koltun, "An empirical evaluation of generic
convolutional and recurrent networks for sequence modeling," arXiv
model inference time between TCN and LSTM is significant. preprint arXiv:1803.01271, 2018.
TCN still keeps excellent property in model inference time, [13] P. Lara-Benítez, M. Carranza-García, J. M. Luna-Romera, and J. C.
but LSTM spends much more time with long previous Riquelme, "Temporal Convolutional Networks Applied to Energy-
timestep than itself in the experiment with short timestep. It Related Time Series Forecasting," Applied Sciences, vol. 10, no. 7, pp.
can be proved that TCN is more significant in saving model 2322, 2020.
processing time and achieve a good detection performance. [14] J. Dong, J. Pang, Y. Peng, and D. Liu, "Spacecraft telemetry data
anomaly detection method based on ensemble LSTM," Chinese Journal
V. CONCLUSION AND FUTURE WORK of Scientific Instrument, vol. 40, no. 7, pp. 22-29, 2019 (in Chinese).
[15] D. Pan, Z. Song, L. Nie, and B. Wang, "Satellite Telemetry Data
This work achieves anomaly detection analysis of Anomaly Detection Using Bi-LSTM Prediction Based Model," in 2020
spacecraft telemetry data transmitted from the on-orbit IEEE International Instrumentation and Measurement Technology
satellite and evaluates the performance of TCN by being Conference (I2MTC), 2020: IEEE, pp. 1-6.
compared with LSTM. Both models perform satisfied
detection performance reflected on ACC, TPR and FPR.

Authorized licensed use limited to: Visvesvaraya Technological University Belagavi. Downloaded on March 30,2024 at 08:45:49 UTC from IEEE Xplore. Restrictions apply.