Unsupervised satellite anomaly detection method, device and equipment based on TDRAE and medium

Abstract

The embodiment of the application discloses a kind of unsupervised satellite anomaly detection methods, device and equipment based on time domain deconvolution reconstruction autoencoder and medium;The method comprises: using the encoder and decoder consisting of causal convolution layer and advanced dilated causal convolution block ADCCB formed based on causal convolution layer constructs time domain deconvolution reconstruction autoencoder TDRAE;Satellite telemetry data is extracted from the satellite telemetry data packet received by ground receiving station, and training set and verification set are generated based on the extracted satellite telemetry data;According to the training parameter set, the training set and verification set are used to train TDRAE, and the trained TDRAE is locally deployed in ground receiving station;The quasi-real-time satellite telemetry data obtained is input into the trained TDRAE locally deployed, to obtain prediction result;The prediction result is evaluated according to the set evaluation strategy, and the evaluation result for indicating whether there is anomaly is obtained.

Description

TECHNICAL FIELD

[0001] The embodiment of the present application relates to the technical field of aerospace signal processing, and particularly relates to an unsupervised satellite anomaly detection method and device based on a temporal deconvolutional reconstruction autoencoder (TDRAE), equipment and a medium. BACKGROUND

[0002] For a spacecraft, for example, a satellite, the structure of each subsystem is very complex, and needs to be operated in the extreme temperature and strong electromagnetic radiation environment in space for a long time. Therefore, through real-time or quasi-real-time anomaly detection of the satellite, the fault can be found and located in time, and the safe and reliable operation of the satellite in orbit is ensured, that is, anomaly detection for the satellite is the most important in satellite health management. Therefore, through anomaly detection of telemetry data, the operation state of the satellite can be directly judged.

[0003] At present, the conventional scheme for satellite anomaly detection mostly adopts a data-driven method. In addition to the satellite in orbit, when the satellite is in the ground test stage, due to the diversity of data, the whole satellite needs to be assembled, integrated and tested (AIT) in 11 different stages. During the AIT process, the high-speed collected 5000 or more satellite time series state data need to be monitored in real time in each stage. The AIT personnel need to judge whether the subsystem or component of the satellite is in a healthy state through the obtained data, and feed back to the chief designer with the concept of health degree or abnormal degree, and the chief designer adjusts and checks the satellite again. Therefore, reliable high-dimensional coupled data anomaly detection is also very important in the ground test stage of the satellite. SUMMARY

[0004] Therefore, the embodiment of the present application expects to provide an unsupervised satellite anomaly detection method and device based on a temporal deconvolutional reconstruction autoencoder (TDRAE), which can improve the detection accuracy of abnormal data and the lightweight of neural network training for data processing, so that network training can be completed with less memory occupation.

[0005] The technical scheme of the embodiment of the present application is as follows:

[0006] In a first aspect, the embodiment of the present application provides an unsupervised satellite anomaly detection method based on a temporal deconvolutional reconstruction autoencoder (TDRAE), which comprises the following steps:

[0007] The encoder and the decoder are constructed by using a causal convolution layer and an advanced dilated causal convolution block ADCCB formed based on the causal convolution layer, and a time domain deconvolution reconstruction autoencoder TDRAE is constructed by using the encoder and the decoder;

[0008] Satellite telemetry data is extracted from a satellite telemetry data packet received by a ground receiving station, and a training set and a verification set are generated based on the extracted satellite telemetry data;

[0009] The TDRAE is trained by using the training set and the verification set according to a set training parameter, and the trained TDRAE is locally deployed at the ground receiving station;

[0010] The acquired quasi-real-time satellite telemetry data is input into the trained TDRAE locally deployed, to obtain a prediction result;

[0011] The prediction result is evaluated according to a set evaluation strategy, to obtain an evaluation result for indicating whether an anomaly exists.

[0012] In a second aspect, an embodiment of the present application provides an unsupervised satellite anomaly detection device based on a time domain deconvolution reconstruction autoencoder TDRAE, the device comprising a construction part, an extraction part, a generation part, a training part, an input part and an evaluation part, wherein,

[0013] The construction part is configured to construct a time domain deconvolution reconstruction autoencoder TDRAE by using an encoder and a decoder composed of a causal convolution layer and an advanced dilated causal convolution block ADCCB formed based on the causal convolution layer;

[0014] The extraction part is configured to extract satellite telemetry data from a satellite telemetry data packet received by a ground receiving station;

[0015] The generation part is configured to generate a training set and a verification set based on the extracted satellite telemetry data;

[0016] The training part is configured to train the TDRAE by using the training set and the verification set according to a set training parameter, and locally deploy the trained TDRAE at the ground receiving station;

[0017] The input part is configured to input acquired quasi-real-time satellite telemetry data into the trained TDRAE locally deployed, to obtain a prediction result;

[0018] The evaluation part is configured to evaluate the prediction result according to a set evaluation strategy, to obtain an evaluation result for indicating whether an anomaly exists.

[0019] In a third aspect, an embodiment of the present application provides a computing device, comprising: a communication interface, a memory and a processor; each component is coupled together through a bus system; wherein,

[0020] The communication interface is configured to receive and send signals in the process of transmitting and receiving information with other external network elements.

[0021] The memory is configured to store a computer program capable of running on the processor.

[0022] The processor is configured to execute the steps of the unsupervised satellite anomaly detection method based on TDRAE in the first aspect when running the computer program.

[0023] In a fourth aspect, an embodiment of the present application provides a computer storage medium, which stores an unsupervised satellite anomaly detection program based on TDRAE, and the steps of the unsupervised satellite anomaly detection method based on TDRAE in the first aspect are implemented when the unsupervised satellite anomaly detection program based on TDRAE is executed by at least one processor.

[0024] The embodiment of the present application provides an unsupervised satellite anomaly detection method, device, equipment and medium based on time domain deconvolution reconstruction autoencoder; the encoder and decoder composed of the causal convolution layer and the advanced dilated causal convolution block ADCCB formed based on the causal convolution layer are used to construct the time domain deconvolution reconstruction autoencoder TDRAE to improve the TCN model in the conventional scheme, and after training by using satellite telemetry data, the acquired quasi-real-time satellite telemetry data is predicted, and finally the prediction result is evaluated to indicate whether there is an anomaly; thereby compared with the conventional TCN model, the detection accuracy of abnormal data is improved, and the neural network training is still lightweight as the TCN model has, so that the neural network training can be completed under the condition of occupying less memory. BRIEF DESCRIPTION OF DRAWINGS

[0025] Figure 1 A causal convolution process schematic diagram is provided for the embodiment of the present application.

[0026] Figure 2 A flowchart of an unsupervised satellite anomaly detection method based on TDRAE is provided for the embodiment of the present application.

[0027] Figure 3 A structure schematic diagram of an advanced dilated causal convolution block ADCCB is provided for the embodiment of the present application.

[0028] Figure 4 A TDRAE structure schematic diagram is provided for the embodiment of the present application.

[0029] Figure 5 A training process schematic diagram in a simulation experiment provided for an embodiment of the present application is shown in FIG. 1.

[0030] Figure 6 A kernel density estimation curve schematic diagram of a histogram and an anomaly index provided for an embodiment of the present application is shown in FIG. 3.

[0031] Figure 7 A data set classification schematic diagram provided for an embodiment of the present application is shown in FIG. 4.

[0032] Figure 8 A TDRAE-based unsupervised satellite anomaly detection device composition schematic diagram provided for an embodiment of the present application is shown in FIG. 5.

[0033] Figure 9 A computing device hardware structure schematic diagram provided for an embodiment of the present application is shown in FIG. 6. DETAILED DESCRIPTION

[0034] The technical solutions in the embodiments of the present application will be described clearly and completely below with reference to the accompanying drawings in the embodiments of the present application.

[0035] For modeling of time series problems, a recurrent neural network architecture is usually used for processing, such as the commonly used method based on a recurrent neural network (RNN) structure and its related variants. A classic convolutional neural network (CNN) model is rarely used for processing of time series problems due to the limitation of the convolution kernel. However, a temporal convolutional network (TCN) model formed by making appropriate modifications to the classic CNN model is mainly composed of causal convolutions (which can also be referred to as causal convolutions for short), so that no information will leak to the past. The TCN model only processes information forward and does not depend on information from the past. In this way, the TCN not only has the advantage of being able to perform high-level parallel computing on parallel processors such as GPUs as the CNN does, but also enables lightweight training of the TCN model with less memory occupation. In addition, causal convolution stacking under different dilation factors can also be realized. Taking a satellite orbit anomaly detection scenario as an example, as shown in the causal convolution process shown in FIG. 2, the dilation factor is set to Dilation = 1, and the input tensor part shown in the dashed box in the lower half part of FIG. 2 represents a time series (X1, X2, …, X Figure 1 Figure 1 T ​​), the longitudinal axis represents the telemetry data number, for example, in the J2000 coordinate system, the longitudinal axis respectively represents the position and velocity of the satellite in the X, Y, Z axis, that is, Figure 1 The element in the first row and the first column of the input tensor represents the telemetry data value of the X-axis coordinate of the satellite in the J2000 coordinate system at the X1 moment. After the input tensor is subjected to the causal convolution of the convolution filter (Convolutional Filters) shown in the solid line box in Figure 1 The filter number of the convolution kernel is 4, and the kernel size is 3. After the causal convolution, the output tensor (Output Tensor) can be obtained. In Figure 1 The output tensor part (Output Tensor) shown in the upper half of the dashed line box in T ), but the longitudinal axis size changes to the filter number of the causal convolution. The above example of the causal convolution process also verifies that the TCN is suitable for time series data problems.

[0036] Based on the above description, the embodiment of the present application expects to use the causal convolution of the TCN to detect anomalies for the high-dimensional coupled telemetry data of the satellite, improve the detection accuracy of the abnormal data, and make the neural network training lightweight, so that the neural network training can be completed in the case of occupying less memory. Based on this, referring to Figure 2 It shows an unsupervised satellite anomaly detection method based on a time domain deconvolution reconstruction autoencoder provided by the embodiment of the present application, and the method comprises:

[0037] S201: An encoder (Encoder) and a decoder (Decoder) composed of a causal convolution layer and an advanced dilated causal convolution block (ADCCB, Advanced Dilated Causal Convolutional Block) formed based on the causal convolution layer are used to construct a time domain deconvolution reconstruction autoencoder (TDRAE, Temporal Deconvolutional Reconsturction AutoEncode);

[0038] S202: Extracting satellite telemetry data from the satellite telemetry data packet received by the ground receiving station, and generating a training set and a validation set based on the extracted satellite telemetry data;

[0039] S203: Training the TDRAE according to the set training parameters using the training set and the validation set, and locally deploying the TDRAE after training in the ground receiving station;

[0040] S204: input the acquired quasi-real-time satellite telemetry data into the locally deployed trained TDRAE to obtain a prediction result;

[0041] S205: evaluate the prediction result according to a set evaluation strategy to obtain an evaluation result for indicating whether there is an anomaly.

[0042] Through the above technical solution, the encoder and the decoder composed of the causal convolution layer and the advanced dilated causal convolution block ADCCB formed based on the causal convolution layer are used to construct the time domain deconvolution reconstruction autoencoder TDRAE to improve the TCN model in the conventional scheme, and after training using satellite telemetry data, the acquired quasi-real-time satellite telemetry data is predicted, and finally the prediction result is evaluated to indicate whether there is an anomaly; thereby compared with the conventional TCN model, the detection accuracy of abnormal data is improved, and the neural network training is still lightweight as the TCN model has, so that the neural network training can be completed with less memory occupation.

[0043] For Figure 2 As shown in the technical solution, in some examples, the TDRAE is constructed using the encoder and the decoder composed of the causal convolution layer and the ADCCB formed based on the causal convolution layer, comprising:

[0044] The advanced dilated causal convolution block ADCCB is constructed using the causal convolution (Causal Convolution) layer, the batch normalization (BN, Batch Normalization) layer, the Gaussian error linear unit (GeLU, Gaussian Error Linear Unit) layer, the spatial random deactivation (Spatial Dropout) layer and the layer normalization (LN, Layer Normalization) layer;

[0045] A first sandwich architecture composed of the causal convolution layer and a plurality of ADCCBs is arranged between the input (Input) layer and the maximum pooling (Max-pooling) layer to form the encoder;

[0046] A second sandwich architecture composed of the causal transpose convolution (Causal Transpose Convolution) layer and a plurality of ADCCBs is arranged between the up-sampling (Up-sampling) layer and the output (Output) layer to form the decoder;

[0047] By arranging the intrinsic vector layer between the maximum pooling layer of the encoder and the up-sampling layer of the decoder, the TDRAE is constructed.

[0048] For the above implementation, in some examples, as shown in Figure 3 The structure of the ADCCB includes a first branch and a second branch coupled in parallel; the first branch includes, in sequence from input to output, a first causal convolution layer, a batch normalization (BN) layer, a first Gaussian error linear unit (GeLU) layer, a first spatial dropout layer, a second causal convolution layer, a layer normalization (LN) layer, a second GeLU layer, a second spatial dropout layer, and a third causal convolution layer; and the second branch includes, from input to output, a fourth causal convolution layer.

[0049] In combination with the above examples and Figure 3 In detail, in Figure 3 The parameters of the causal convolution layer include a parameter "n@s" representing time series information of the input. Understandably, when s = 1, "n@1" in Figure 3 represents that the time series information of the input is a "1" sequence; and a parameter "d" represents a dilated rate, i.e., a number of intervals of a convolution kernel kernel. In detail, since the convolution kernel is discontinuous, not all information participates in the calculation, which easily leads to a loss of information continuity and causes a grid effect. In order to avoid the grid effect, generally, there is no common divisor greater than 1 in the superimposed dilated rate. For example, in the superimposed structure of three causal convolution layers, the dilated rates d corresponding to the three layers cannot have a common divisor greater than 1, such as [1, 2, 4], but not [2, 4, 6].

[0050] Based on the structure of the ADCCB and the causal convolution layer shown in Figure 3 In some examples, the encoder structure includes, in sequence, an input layer, a fifth causal convolution layer, two or more ADCCBs, a sixth causal convolution layer, and a max-pooling layer; and the dilated rates corresponding to the two or more ADCCBs increase in sequence.

[0051] In some examples, the structure of the decoder includes, in sequence, an up-sampling layer, a first causal transposed convolution layer, two or more ADCCBs, a second causal transposed convolution layer, and an output layer; and the dilated rates corresponding to the two or more ADCCBs decrease in sequence.

[0052] For the above two examples, referring to Figure 4 the TDRAE structure shown, specifically, in Figure 4 , the left dashed box shows the encoder Encoder structure adopted by the embodiment of the application, in which the dilated rate d of the fifth causal convolution layer and the sixth causal convolution layer is 1; the number of ADCCB is 3, and the dilated rate d is 1, 2, and 4 in turn. In addition, in Figure 4 , the right dashed box shows the decoder Decoder structure adopted by the embodiment of the application, in which the dilated rate d of the first causal transposed convolution layer and the second causal transposed convolution layer is 1; the number of ADCCB is 3, and the dilated rate d is 4, 2, and 1 in turn. For the encoder Encoder and the decoder Decoder, the connection is made through a latent vector layer, that is, as shown by the solid line in Figure 4 , the output of the Max-pooling layer and the input of the Up-sampling layer are connected through the latent vector layer, which is used to take the latent vector in the data output by the Max-pooling layer as the input data of the Up-sampling layer.

[0053] For the technical solution shown in Figure 2 , in some examples, the satellite telemetry data is extracted from the satellite telemetry data packet received by the ground receiving station, including:

[0054] According to the set data extraction and data conversion rule, the original satellite telemetry data is parsed from the satellite telemetry data packet;

[0055] The original satellite telemetry data is sequentially subjected to a first preprocessing process of wild value removal, data completion, feature selection, and normalization, or a second preprocessing process of wild value removal, data completion, and normalization, to obtain satellite telemetry data applicable to the time-domain deconvolution reconstruction autoencoder TDRAE.

[0056] Based on the above examples, in some examples, the training set and the validation set are generated based on the extracted satellite telemetry data, including:

[0057] For the satellite telemetry data applicable to the time-domain deconvolution reconstruction autoencoder TDRAE, the data is truncated in a sliding window manner to generate a plurality of data sets;

[0058] The generated data sets are divided into a training set and a validation set.

[0059] For the above two examples, in detail, after receiving the satellite telemetry data packet, the ground receiving station can extract data by parsing the software and according to the data extraction and data conversion rules set in the software, so as to obtain the original satellite telemetry data; it can be understood that the original satellite telemetry data has high-dimensional coupling characteristics. Then, the original satellite telemetry data is preprocessed, such as wild value removal, data completion, feature selection and normalization in turn, so as to obtain the satellite telemetry data suitable for the time domain deconvolution reconstruction autoencoder TDRAE proposed in the embodiment of the application; it should be noted that the feature selection in the preprocessing is an optional item and is not a necessary processing means in the data preprocessing process. And for the TDRAE, the input layer and the output layer can be set to be equivalent, that is, the input length of the input layer is equal to the output length of the output layer.

[0060] For the above two examples, in detail, after receiving the satellite telemetry data packet, the ground receiving station can extract data by parsing the software and according to the data extraction and data conversion rules set in the software, so as to obtain the original satellite telemetry data; it can be understood that the original satellite telemetry data has high-dimensional coupling characteristics. Then, the original satellite telemetry data is preprocessed, such as wild value removal, data completion, feature selection and normalization in turn, so as to obtain the satellite telemetry data suitable for the time domain deconvolution reconstruction autoencoder TDRAE proposed in the embodiment of the application; it should be noted that the feature selection in the preprocessing is an optional item and is not a necessary processing means in the data preprocessing process. And for the TDRAE, the input layer and the output layer can be set to be equivalent, that is, the input length of the input layer is equal to the output length of the output layer. Figure 2 For the above two examples, in detail, after receiving the satellite telemetry data packet, the ground receiving station can extract data by parsing the software and according to the data extraction and data conversion rules set in the software, so as to obtain the original satellite telemetry data; it can be understood that the original satellite telemetry data has high-dimensional coupling characteristics. Then, the original satellite telemetry data is preprocessed, such as wild value removal, data completion, feature selection and normalization in turn, so as to obtain the satellite telemetry data suitable for the time domain deconvolution reconstruction autoencoder TDRAE proposed in the embodiment of the application; it should be noted that the feature selection in the preprocessing is an optional item and is not a necessary processing means in the data preprocessing process. And for the TDRAE, the input layer and the output layer can be set to be equivalent, that is, the input length of the input layer is equal to the output length of the output layer.

[0061] According to the prediction result Y pred (i) of the i-th sample and the original output result Y orig (i) of the i-th sample, the abnormal index S(i) of the i-th sample is calculated by using the following formula:

[0062] S(i) = Scaler(N(N(Y pred (i)-Y orig (i), 2), 1)

[0063] Wherein, N(tensor, r) represents the function of L2 criterion of tensor along r axis, and Scaler() represents the (0, 1) normalization function;

[0064] The abnormal index S(i) of the i-th sample is compared with the set evaluation threshold to obtain the evaluation result indicating whether the i-th sample is abnormal.

[0065] For the above example, specifically, the evaluation threshold can be set by the kernel density estimation curve of the abnormal index, and in the embodiment of the application, the threshold can be set to 0.2.

[0066] Based on the aforementioned technical solution, this embodiment of the invention uses telemetry data from a certain type of hyperspectral satellite as an example for simulation experiments, and only a dataset containing 22 types of telemetry data was created based on this telemetry data. In this simulation experiment, the temporal deconvolutional reconstruction autoencoder (TDRAE) proposed in this embodiment of the invention is implemented in Python 3.9 using Tensorflow-Keras 2.8.0. The running environment is a mobile workstation equipped with an RTX 3080Ti GPU with 16GB of video memory; the optimizer for the TDRAE model is Adam; the loss function is the mean square error (MSE); the training epochs are 500, and the batch size is 128. The learning rate is 0.0001. Based on the above simulation conditions, the technical solution proposed in this embodiment of the invention is executed, and its training process is as follows: Figure 5 As shown, in Figure 5 In the diagram, the horizontal axis represents the training epoch, and the vertical axis represents the loss function value. The solid line represents the loss on the training set, and the dashed line represents the loss on the validation set. Figure 5 It can be seen that after 100 training rounds, both the loss on the training set and the loss on the validation set have converged, indicating that the TDRAE model proposed in this embodiment can complete the training process relatively quickly. Furthermore, regarding the histogram used to determine the threshold and the kernel density estimation curve of the anomaly index, as... Figure 6 As shown, in Figure 6 In the histogram, the horizontal axis represents the anomaly score, and the vertical axis represents the kernel density. From the histogram and kernel density estimation curve, it can be seen that the optimal assessment threshold for indicating the presence of anomalies is 0.2. Figure 6 For example, with an evaluation threshold of 0.2, see [link / reference]. Figure 7 The dataset classification diagram shown is in... Figure 7 In the diagram, the horizontal axis represents the number of samples (S / N), with a total sample size of 12,000. The vertical axis represents the anomaly score, and the dotted line represents the evaluation threshold of 0.2. This means that an anomaly score exceeding this threshold (i.e., below a certain threshold) indicates a condition where the anomaly score exceeds a certain threshold. Figure 7 Points (those above the dashed line) are considered abnormal; from Figure 7 It can be seen that the TDRAE model proposed in the embodiments of the present invention and the aforementioned Figure 6 The evaluation threshold of 0.2 shown can be used to classify abnormal samples in the dataset, thereby obtaining abnormal data.

[0067] In order to embody the technical effect of the TDRAE model proposed in the embodiments of the present application, the TDRAE model proposed in the embodiments of the present application (hereinafter referred to as TDRAE) is compared with some more advanced models adopted in the current conventional scheme through simulation experiments; for example, the BiLSTM-Attention model introducing the bidirectional propagation mechanism and the attention mechanism on the basis of the long short-term memory network (LSTM), the GRU-Attetntion model introducing the attention mechanism on the basis of the gated recurrent unit (GRU), the convolutional autoencoder (CAE) model and the classical temporal convolutional network (TCN) model. In the comparison process, in addition to the training round being changed from 500 to 100, the simulation experiment conditions of each model are consistent with the aforementioned simulation experiment conditions, and the comparison experiment results are shown in Table 1.

[0068] Table 1

[0069]

[0070] It can be seen from the above table that if the MSE, the mean absolute error (MAE), the mean squared logarithmic error (MSLE) and the determination coefficient R 2 As the loss function of the training set and the validation set, when the training round reaches 100, the loss function value of the TDRAE is optimal, and the TCN model is second. That is, for the multivariate coupled time series data such as satellite telemetry data, the training effect of the BiLSTM-Attention model, the GRU-Attention model and the CAE model is obviously not as good as that of the TCN model and the TDRAE, which further verifies that the causal convolution type model is more suitable for the multivariate coupled time series data; secondly, although the training effect of the TCN is better than that of the BiLSTM-Attention model, the GRU-Attention model and the CAE model, it is still inferior to the TDRAE. In addition, only by focusing on the determination coefficient R 2 It can be known that when the training round reaches 100, the R 2TDRAE is 65%, 65%, 65% and 3% higher than BiLSTM-Attention model, GRU-Attention model, CAE model and TCN model respectively, which shows that the input tensor of TDRAE has a higher degree of explanation for the output tensor, TDRAE has the highest relative degree of regression contribution, and therefore TDRAE has better regressor properties.

[0071] Based on the same inventive concept of the foregoing technical solutions, see Figure 8 It shows a kind of unsupervised satellite anomaly detection device 80 based on TDRAE provided by the embodiment of the application, the device 80 includes: construction part 801, extraction part 802, generation part 803, training part 804, input part 805 and evaluation part 806, wherein,

[0072] The construction part 801 is configured to construct the time domain deconvolution reconstruction autoencoder TDRAE using the encoder and decoder composed of the causal convolution layer and the advanced dilated causal convolution block ADCCB formed based on the causal convolution layer;

[0073] The extraction part 802 is configured to extract satellite telemetry data from the satellite telemetry data packet received by the ground receiving station;

[0074] The generation part 803 is configured to generate a training set and a validation set based on the extracted satellite telemetry data;

[0075] The training part 804 is configured to train the TDRAE using the training set and the validation set according to the set training parameters, and locally deploy the TDRAE after training in the ground receiving station;

[0076] The input part 805 is configured to input the quasi-real-time satellite telemetry data obtained into the TDRAE after training locally deployed to obtain a prediction result;

[0077] The evaluation part 806 is configured to evaluate the prediction result according to the set evaluation strategy to obtain an evaluation result for indicating whether there is an anomaly.

[0078] In some examples, the construction part 801 is configured to:

[0079] The ADCCB is constructed using the causal convolution layer, batch normalization layer, Gaussian error linear unit layer, spatial random inactivation layer and layer normalization layer;

[0080] A first sandwich architecture composed of a causal convolution layer and a plurality of ADCCBs is arranged between the input layer and the max-pooling layer to form the encoder;

[0081] a second sandwich architecture composed of a causal transposed convolutional layer transposing by causal convolution and a plurality of the ADCCB is arranged between the up-sampling layer and the output layer, forming the decoder;

[0082] By arranging an eigenvector layer between the max-pooling layer of the encoder and the up-sampling layer of the decoder, the TDRAE is constructed.

[0083] In the above example, the structure of the ADCCB includes: a first branch and a second branch coupled in parallel; the first branch includes, in order from input to output: a first causal convolutional layer, a batch normalization layer, a first Gaussian error linear unit layer, a first spatial random dropout layer, a second causal convolutional layer, a layer normalization layer, a second Gaussian error linear unit layer, a second spatial random dropout layer, and a third causal convolutional layer; the second branch includes, from input to output: a fourth causal convolutional layer.

[0084] In the above example, the encoder structure includes, in order: an input layer, a fifth causal convolutional layer, two or more ADCCBs, a sixth causal convolutional layer, and a max-pooling layer; wherein the expansion rates of the two or more ADCCBs in the encoder structure increase in order.

[0085] The decoder structure includes, in order: an up-sampling layer, a first causal transposed convolutional layer, two or more ADCCBs, a second causal transposed convolutional layer, and an output layer; wherein the expansion rates of the two or more ADCCBs in the decoder structure decrease in order.

[0086] In some examples, the extraction part 802 is configured to:

[0087] According to the set data extraction and data conversion rules, the original satellite telemetry data is parsed from the satellite telemetry data packet;

[0088] The original satellite telemetry data is subjected to a first preprocessing process of wild value removal, data completion, feature selection, and normalization in order, or a second preprocessing process of wild value removal, data completion, and normalization in order, to obtain satellite telemetry data applicable to the time-domain deconvolution reconstruction autoencoder TDRAE.

[0089] In some examples, the generation part 803 is configured to:

[0090] For the satellite telemetry data applicable to the time-domain deconvolution reconstruction autoencoder TDRAE, data truncation is performed in a sliding window manner to generate a plurality of data sets;

[0091] The generated data sets are divided into a training set and a validation set.

[0092] In some examples, the evaluating part 806 is configured to:

[0093] According to the prediction result Y pred (i) of the i-th sample orig (i), the abnormality index S(i) of the i-th sample is calculated by using the following formula:

[0094] S(i) = Scaler(N(N(Y pred (i) - Y orig (i), 2), 1)

[0095] Wherein, N(tensor, r) represents the function of L2 norm of tensor along the r axis, and Scaler() represents the (0, 1) normalization function.

[0096] The abnormality index S(i) of the i-th sample is compared with the set evaluation threshold to obtain the evaluation result indicating whether the i-th sample is abnormal.

[0097] It can be understood that in the embodiment, the "part" can be a partial circuit, a partial processor, a partial program or software, etc., and of course can also be a unit, and can also be a module or non-modular.

[0098] In addition, each component in the embodiment can be integrated in a processing unit, or each unit can exist physically alone, or two or more units can be integrated in one unit. The integrated unit can be realized in the form of hardware or in the form of a software function module.

[0099] The integrated unit, if realized in the form of a software function module and not sold or used as an independent product, can be stored in a computer readable storage medium. Based on this understanding, the technical solutions of the embodiment can be embodied in the form of a software product, the computer software product is stored in a storage medium, and includes a plurality of instructions for causing a computer device (which can be a personal computer, a server, or a network device, etc.) or a processor to execute all or part of the steps of the method described in the embodiment. The foregoing storage medium includes: a U disk, a mobile hard disk, a read-only memory (ROM), a random access memory (RAM), a magnetic disk or an optical disk, and various program code storage media.

[0100] Therefore, the embodiment provides a computer storage medium, which stores a TDRAE-based unsupervised satellite anomaly detection program. The TDRAE-based unsupervised satellite anomaly detection program, when executed by at least one processor, implements the TDRAE-based unsupervised satellite anomaly detection method steps in the above technical solutions.

[0101] According to the above TDRAE-based unsupervised satellite anomaly detection device 80 and the computer storage medium, referring to Figure 9 , a specific hardware structure of a computing device 90 capable of implementing the above TDRAE-based unsupervised satellite anomaly detection device 80 is shown, and the computing device 90 can be a wireless device, a mobile or cellular phone (including a so-called smart phone), a personal digital assistant (PDA), a video game console (including a video display, a mobile video game device, a mobile video conference unit), a laptop computer, a desktop computer, a television set-top box, a tablet computing device, an electronic book reader, a fixed or mobile media player, and the like. The computing device 90 includes a communication interface 901, a memory 902, and a processor 903, and each component is coupled together through a bus system 904. It can be understood that the bus system 904 is used to realize the connection communication between the components. In addition to including a data bus, the bus system 904 also includes a power bus, a control bus, and a status signal bus. However, for the purpose of clear illustration, all kinds of buses are marked as the bus system 904 in Figure 9 , wherein,

[0102] The communication interface 901 is configured to receive and send signals in the process of transceiving information with other external network elements;

[0103] The memory 902 is configured to store a computer program capable of running on the processor 903;

[0104] The processor 903 is configured to execute the TDRAE-based unsupervised satellite anomaly detection method steps in the above technical solutions when running the computer program.

[0105] It is to be appreciated that the memory 902 in embodiments of the application can be volatile, nonvolatile, or a combination of both. By way of example, and without limitation, nonvolatile memory can include read-only memory (ROM), programmable ROM (PROM), erasable PROM (EPROM), electrically EPROM (EEPROM), a flash memory, or a combination of these. Volatile memory can include random-access memory (RAM), which acts as external cache. By way of example and without limitation, many forms of RAM are available, such as static RAM (SRAM), dynamic RAM (DRAM), synchronous DRAM (SDRAM), double-data rate SDRAM (DDR SDRAM), enhanced SDRAM (ESDRAM), Synchlink DRAM (SLDRAM), and direct Rambus RAM (DRRAM). The memory 902 of the subject systems and methods herein is intended to include, without being limited to, these and any other suitable types of memory.

[0106] The processor 903 can be an integrated circuit chip logic circuit having a processing capability. In implementation, each step of the above method can be completed by integrated logic circuit of hardware in the processor 903 or instruction in the form of software. The processor 903 described above can be a general processor, a digital signal processor (DSP), an application specific integrated circuit (ASIC), a field programmable gate array (FPGA) or other programmable logic device, a discrete gate or transistor logic device, a discrete hardware component. Each method, step and logic block disclosed in the embodiments of the present application can be implemented or executed. The general processor can be a microprocessor or the processor can also be any conventional processor. The steps of the method disclosed in combination with the embodiments of the present application can be directly embodied as a hardware code processor for execution, or a combination of hardware and software modules in the code processor for execution. The software module can be located in a random access memory, a flash memory, a read only memory, a programmable read only memory or an electrically erasable programmable memory, a register, etc. The storage medium in the art is mature. The storage medium is located in the memory 902, and the processor 903 reads the information in the memory 902 and combines the hardware to complete the steps of the above method.

[0107] It can be understood that the embodiments described herein can be implemented in hardware, software, firmware, middleware, microcode, or a combination thereof. For hardware implementation, the processing unit can be implemented within one or more application specific integrated circuits (ASICs), digital signal processors (DSPs), digital signal processing devices (DSPDs), programmable logic devices (PLDs), field programmable gate arrays (FPGAs), processors, controllers, micro-controllers, microprocessors, other electronic units designed to perform the functions described herein, or a combination thereof.

[0108] For software implementation, the techniques described herein can be implemented with modules (e.g., procedures, functions, and so on) that perform the functions described herein. The software codes can be stored in memory and executed by processors. The memory can be implemented within the processors or external to the processors.

[0109] It can be understood that the exemplary technical solutions of the above-mentioned TDRAE-based unsupervised satellite anomaly detection device 80 and the computing device 90 belong to the same concept as the technical solutions of the aforementioned TDRAE-based unsupervised satellite anomaly detection method, and therefore, the details of the technical solutions of the TDRAE-based unsupervised satellite anomaly detection device 80 and the computing device 90 which are not described in detail can be referred to the description of the technical solutions of the aforementioned TDRAE-based unsupervised satellite anomaly detection method. The present embodiment does not repeat here.

[0110] It should be noted that the technical solutions disclosed in the embodiments of the present application can be combined arbitrarily without conflict.

[0111] The above is only a specific implementation of the present application, but the protection scope of the present application is not limited to this. Any person skilled in the art can easily think of changes or replacements within the technical scope disclosed by the present application, which should be covered within the protection scope of the present application. Therefore, the protection scope of the present application should be subject to the protection scope of the claims.

Claims

1. An unsupervised satellite anomaly detection method based on time-domain deconvolution reconstruction autoencoder (TDRAE), characterized in that, The method comprises: The time domain deconvolution reconstruction autoencoder (TDRAE) is constructed by using an encoder and a decoder composed of causal convolution layers and advanced dilated causal convolution blocks (ADCCB) formed based on the causal convolution layers; the encoder comprises an input layer, a max-pooling layer, and two or more ADCCB arranged between the input layer and the max-pooling layer, and the dilated rates of the two or more ADCCB in the encoder increase in sequence; The decoder comprises an up-sampling layer, an output layer, and two or more ADCCB arranged between the up-sampling layer and the output layer, and the dilated rates of the two or more ADCCB in the decoder decrease in sequence; and The structure of each ADCCB comprises a first branch and a second branch coupled in parallel; the first branch comprises, in sequence from input to output, a first causal convolution layer, a batch normalization layer, a first Gaussian error linear unit layer, a first spatial random dropout layer, a second causal convolution layer, a layer normalization layer, a second Gaussian error linear unit layer, a second spatial random dropout layer, and a third causal convolution layer; and the second branch comprises, from input to output, a fourth causal convolution layer. Satellite telemetry data is extracted from satellite telemetry data packets received by a ground receiving station, and a training set and a verification set are generated based on the extracted satellite telemetry data; The TDRAE is trained by using the training set and the verification set according to set training parameters, and the trained TDRAE is locally deployed at the ground receiving station; The acquired quasi-real-time satellite telemetry data is input into the locally deployed trained TDRAE to obtain a prediction result; The prediction result is evaluated according to a set evaluation strategy to obtain an evaluation result for indicating whether there is an anomaly.

2. The method of claim 1, wherein, The TDRAE is constructed by using an encoder and a decoder composed of causal convolution layers and ADCCB formed based on the causal convolution layers, which comprises: The ADCCB is constructed by using the causal convolution layers, batch normalization layers, Gaussian error linear unit layers, spatial random dropout layers, and layer normalization layers; A first sandwich architecture composed of causal convolution layers and a plurality of ADCCB is arranged between the input layer and the max-pooling layer to form the encoder; A second sandwich architecture composed of causal transposed convolution layers and a plurality of ADCCB is arranged between the up-sampling layer and the output layer to form the decoder; The TDRAE is constructed by arranging an eigenvector layer between the max-pooling layer of the encoder and the up-sampling layer of the decoder.

3. The method of claim 1, wherein, The satellite telemetry data is extracted from satellite telemetry data packets received by a ground receiving station, which comprises: Raw satellite telemetry data is parsed from the satellite telemetry data packets according to set data extraction and data conversion rules; The original satellite telemetry data is sequentially subjected to a first preprocessing process of wild value removal, data completion, feature selection and normalization, or sequentially subjected to a second preprocessing process of wild value removal, data completion and normalization, to obtain satellite telemetry data applicable to the time domain deconvolution reconstruction autoencoder TDRAE.

4. The method of claim 3, wherein, The training set and the validation set are generated based on the extracted satellite telemetry data, which comprises: The satellite telemetry data applicable to the time domain deconvolution reconstruction autoencoder TDRAE is truncated by a sliding window to generate a plurality of data sets; The generated data sets are divided into a training set and a validation set.

5. The method of claim 1, wherein, The prediction result is evaluated according to a set evaluation strategy to obtain an evaluation result for indicating whether there is an anomaly. According to the first i Prediction results for each sample Y pred ( i ) and the i The original output results of each sample Y orig ( i The following formula is used to calculate the first... i Anomaly index of each sample S ( i ): S ( i ) = Scaler （ N （ N （ Y pred ( i )- Y orig ( i )，2），1）） wherein, N ( tensor, r ) denotes a function of the L2 norm of the tensor along the r axis, Scaler () denotes a (0, 1) normalizing function; The abnormality index of the first sample is compared with a set evaluation threshold, and an evaluation result indicating whether the first sample is abnormal is obtained. i S ( i ) with a set evaluation threshold, and an evaluation result indicating whether the first sample is abnormal is obtained. i ​​ 6. An unsupervised satellite anomaly detection apparatus based on a time-domain recurrent autoencoder (TDRAE), characterized in that, The device comprises a construction part, an extraction part, a generation part, a training part, an input part and an evaluation part, wherein, The construction part is configured to construct a time domain deconvolution reconstruction autoencoder TDRAE using an encoder and a decoder composed of a causal convolution layer and an advanced dilated causal convolution block ADCCB formed based on the causal convolution layer; wherein the encoder comprises an input layer, a max-pooling layer, and two or more ADCCBs arranged between the two; wherein the dilution rates of the two or more ADCCBs in the encoder increase in turn; The decoder comprises an up-sampling layer, an output layer, and two or more ADCCBs arranged between the two; wherein the dilution rates of the two or more ADCCBs in the decoder decrease in turn; and The structure of each ADCCB comprises a first branch and a second branch coupled in parallel; the first branch comprises, in the direction from input to output, a first causal convolution layer, a batch normalization layer, a first Gaussian error linear unit layer, a first spatial random inactivation layer, a second causal convolution layer, a layer normalization layer, a second Gaussian error linear unit layer, a second spatial random inactivation layer and a third causal convolution layer; the second branch comprises, in the direction from input to output, a fourth causal convolution layer; The extraction part is configured to extract satellite telemetry data from satellite telemetry data packets received by a ground receiving station; The generation part is configured to generate a training set and a validation set based on the extracted satellite telemetry data; The training part is configured to train the TDRAE using the training set and the validation set according to a set training parameter, and to locally deploy the trained TDRAE at the ground receiving station; The input part is configured to input the obtained quasi-real-time satellite telemetry data to the locally deployed trained TDRAE to obtain a prediction result; The evaluation part is configured to evaluate the prediction result according to a set evaluation strategy to obtain an evaluation result for indicating whether there is an anomaly.

7. A computing device, comprising: The computing device comprises a communication interface, a memory and a processor; each component is coupled together through a bus system; wherein, The communication interface is configured to receive and send signals in the process of transmitting and receiving information with other external network elements. The memory is configured to store a computer program capable of running on the processor. The processor is configured to execute the steps of the unsupervised satellite anomaly detection method based on TDRAE in any one of claims 1 to 5 when running the computer program.

8. A computer storage medium, characterized in that, The computer storage medium stores an unsupervised satellite anomaly detection program based on TDRAE, and the unsupervised satellite anomaly detection program based on TDRAE implements the steps of the unsupervised satellite anomaly detection method based on TDRAE in any one of claims 1 to 5 when executed by at least one processor.

Images