A satellite telemetry parameter anomaly detection system and method based on a neural contrast-enhanced autoencoder

Abstract

The application provides a satellite telemetry parameter anomaly detection system and method based on a neural contrast enhanced autoencoder. The system comprises a neural converter data enhancement module, including a neural converter, for converting original data through the neural converter, then enhancing the converted data to obtain multiple enhanced samples of the original sample, and outputting a sample set including original subsequences and sequences synthesized through data enhancement; and an autoencoder feature module, including two symmetrical LSTM networks of an encoder and a decoder, for mapping the sample to a feature space using the encoder, and reconstructing the feature representation through the decoder to obtain anomaly labels of each time point in the to-be-tested sequence. The application has the advantages that the anomaly detection effect is better than that of the current mainstream deep anomaly detection method, and can effectively detect telemetry parameter temporal context anomalies.

Description

TECHNICAL FIELD

[0001] The application belongs to the field of satellite state anomaly detection, and particularly relates to a satellite telemetry parameter anomaly detection system and method based on a neural contrast-enhanced autoencoder. BACKGROUND

[0002] As a kind of precise spacecraft, artificial satellites play a pivotal role in national defense and security, economic development, ecological protection, and national life. However, due to factors such as adverse space environment and component aging, the satellite state may exhibit abnormal conditions that deviate from expectations, posing a threat to the safety of satellite on-orbit operation and the reliability of target mission execution. Therefore, monitoring the satellite on-orbit state is crucial. Anomaly detection can provide strong decision-making information support for ground control personnel and help ensure the stable and reliable operation of satellites, which has important academic research significance and engineering application value.

[0003] According to the manifestation, satellite on-orbit state anomalies can be divided into three categories: single-point anomaly, correlation relationship anomaly, and time context anomaly. The most complex one is time context anomaly, which refers to the abnormal variation of telemetry parameters in the time dimension. Multiple consecutive single points violate the rules of the original sequence, showing a trend or pattern change that does not exceed the threshold. Time context anomaly may be caused by changes in satellite operating state or events that exceed the predetermined behavior range, such as the slow change in current caused by component aging.

[0004] Telemetry parameters, as a direct reflection of satellite operating mode and operating state, are the key basis for ground control personnel to assess satellite performance and health status. In recent years, with the increasing refinement of satellite instrument components and the rapid growth of on-orbit satellites, the dimension and data volume of telemetry parameters have also increased significantly, posing higher requirements for satellite operating state anomaly detection. Satellite on-orbit state anomalies exhibit diverse phenomena, complex parameter correlations, and time-dependent information superposition. Traditional anomaly detection methods based on thresholds, models, and rules have been unable to meet the current needs of satellite anomaly detection tasks due to the lack of modeling capabilities for the complex characteristics of telemetry parameters.

[0005] Compared with traditional methods, data-driven methods do not rely on prior knowledge and expert experience, can automatically mine potential patterns and rules in historical telemetry parameters, and have strong representation learning and modeling capabilities, and have obvious advantages. Deep learning is one of the commonly used data-driven methods, which has been widely used in telemetry parameter anomaly detection field. However, the time context anomaly detection method based on deep learning still faces problems such as insufficient modeling of time sequence characteristics and lack of representative feature extraction. Satellite telemetry parameters have complex time sequence characteristics. In order to mine time-dependent information, the normal time sequence window is usually taken as input for semi-supervised learning to extract important features. However, compared with the possible abnormal situation in any form, the normal samples used for training are relatively limited, which makes it difficult for machine learning methods to learn representative general features, limits the generalization ability of the model, and further affects the effect of anomaly detection. In-depth study and solution to these problems will help to improve the existing telemetry parameter monitoring method and further realize the intelligent and automatic detection of satellite abnormal state. SUMMARY

[0006] The present application aims to overcome the deficiencies of insufficient modeling of time sequence characteristics and lack of representative feature extraction in the prior art.

[0007] In order to achieve the above-mentioned purpose, the present application proposes a satellite telemetry parameter anomaly detection system based on neural contrast enhanced autoencoder, characterized in that the system comprises:

[0008] A neural converter data enhancement module comprising a neural converter, configured to convert the original data through the neural converter, and then enhance the converted data to obtain a plurality of enhanced samples of the original sample, and output a sample set comprising the original subsequence and the sequence synthesized by data enhancement; and

[0009] An autoencoder feature module comprising two symmetrical LSTM networks of encoder and decoder, configured to map the sample to the feature space by using the encoder, and then reconstruct the feature representation by using the decoder to obtain the abnormal label of each time point in the to-be-detected sequence.

[0010] As an improvement of the above-mentioned system, the neural converter is a recurrent neural network or a Transformer network.

[0011] As an improvement of the above-mentioned system, the neural converter is trained using a deterministic contrast loss method.

[0012] As an improvement of the above-mentioned system, the method of enhancing the converted data comprises: directly outputting the converted sample as an enhanced sample, multiplying the converted sample with the original sample to obtain an enhanced sample output, and adding the converted sample with the original sample to obtain an enhanced sample output.

[0013] As an improvement of the above system, the training of the autoencoder feature module uses reconstruction error as the loss function.

[0014] As an improvement of the above system, the training of the system adopts an end-to-end manner, and the training process includes:

[0015] The input history normal telemetry sequence is preprocessed and divided into a sliding window to obtain a subsequence set for training;

[0016] The subsequence set is input into the neural converter data enhancement module to generate a plurality of enhanced samples of the original sample, obtaining a sample set including the original subsequence and the sequence synthesized by data enhancement;

[0017] The sample set is input into the autoencoder feature module, and the sample is mapped to the feature space through the encoder, and the feature representation is reconstructed through the decoder for feature learning;

[0018] The contrast loss is used to distinguish the samples in the feature space from different neural converters, and the reconstruction loss is used to measure the feature representation capability, and the network parameters of the neural converter data enhancement module and the autoencoder feature module are updated iteratively by jointly optimizing the contrast loss and the reconstruction loss.

[0019] The application also proposes a satellite telemetry parameter anomaly detection method based on neural contrast enhancement autoencoder, which is realized based on the above system, and the method includes:

[0020] The satellite telemetry parameters are preprocessed and divided, and then input into the trained satellite telemetry parameter anomaly detection system based on neural contrast enhancement autoencoder to obtain the anomaly labels of each time point in the satellite telemetry parameters.

[0021] Compared with the prior art, the application has the following advantages:

[0022] Through comparative experiments on two public telemetry parameter datasets SMAP and MSL commonly used in spacecraft anomaly detection field, the superior performance of the NCAAE method proposed in the application is verified, and the anomaly detection effect is better than the current mainstream deep anomaly detection method. The optimal F1* score of the method is obtained on the SMAP and MSL datasets, which is 0.955 and 0.962, respectively, and is improved by 4.83% and 0.94% compared with the suboptimal benchmark method. At the same time, by removing the neural converter and performing ablation experiment, the effectiveness of the self-supervised learning through data enhancement in the proposed method is verified. The results of a series of experiments prove that the proposed method has strong practical ability and can effectively detect the telemetry parameter temporal context anomaly. BRIEF DESCRIPTION OF DRAWINGS

[0023] Figure 1A satellite telemetry parameter anomaly detection system structure diagram based on a neural contrast enhancement autoencoder is shown.

[0024] Figure 2 A matrix stacking diagram of evaluation indexes of results of each method on the SMAP dataset is shown.

[0025] Figure 3 A matrix stacking diagram of evaluation indexes of results of each method on the MSL dataset is shown. DETAILED DESCRIPTION

[0026] The technical solutions of the present application will be described in detail below with reference to the accompanying drawings.

[0027] The present application proposes a satellite telemetry parameter anomaly detection system and method based on a neural contrast enhancement autoencoder for satellite telemetry parameter temporal context anomaly. To solve the problem of lack of representative features learned due to limited normal sample training, neural transformation learning is introduced, and a neural contrast enhancement autoencoder is proposed. The model uses a neural network to learn different transformation methods and performs data augmentation, which expands the diversity of training samples, thereby prompting the autoencoder to learn more general and rich representations. At the same time, the model uses different enhanced samples in the separated feature space as an auxiliary task for self-supervised learning, solving the problem of lack of supervision information in semi-supervised learning. By jointly optimizing the neural transformer and the autoencoder network, the learned model can reconstruct the normal sequence with high quality, so that the abnormal subsequence is distinguished due to the large reconstruction error, and the detection of temporal context anomaly is realized.

[0028] The present application proposes a satellite telemetry parameter anomaly detection system and method based on a neural contrast enhancement autoencoder for satellite telemetry parameter temporal context anomaly. To solve the problem of lack of representative features learned due to limited normal sample training, neural transformation learning is introduced, and a neural contrast enhancement autoencoder is proposed. The model uses a neural network to learn different transformation methods and performs data augmentation, which expands the diversity of training samples, thereby prompting the autoencoder to learn more general and rich representations. At the same time, the model uses different enhanced samples in the separated feature space as an auxiliary task for self-supervised learning, solving the problem of lack of supervision information in semi-supervised learning. By jointly optimizing the neural transformer and the autoencoder network, the learned model can reconstruct the normal sequence with high quality, so that the abnormal subsequence is distinguished due to the large reconstruction error, and the detection of temporal context anomaly is realized.

[0029] Under the background of continuous development of automation and intelligence, the application aims to build an abnormality detection model based on deep learning to improve the ability to mine deep information and rules contained in telemetry parameters and improve the accuracy of abnormality detection. The method provides a theoretical method basis for ground operation and control personnel to monitor the temporal context abnormality of satellite telemetry parameters, helps to improve the existing telemetry parameter monitoring method, and effectively identifies the potential risks faced by the satellite.

[0030] The satellite telemetry parameter abnormality detection system and method based on the neural contrast enhancement autoencoder provided by the application is composed of two stages of detection model training and abnormal state judgment. As shown in the figure, Figure 1 The system contains two main components, a neural converter data enhancement module and an autoencoder feature extraction module, which is called a neural contrast enhancement autoencoder, an end-to-end deep self-supervised anomaly detection model. Through detection model training, the potential rules in the normal telemetry sequence can be learned, and a model based on reconstruction is constructed for abnormal state judgment. For the test sequence, the data rules of the normal sample are similar to the original training data, which can be well reconstructed by the model, while the error of the abnormal sample is larger after model reconstruction, so it can be distinguished.

[0031] The training process of the model is as follows: first, input the historical normal telemetry sequence and perform preprocessing and sliding window division to obtain a set of sub-sequences for training; next, input these sub-sequences into the neural converter to generate multiple enhanced samples of the original sample, and obtain a sample set including the original sub-sequences and the sequences synthesized by data enhancement; then, input the sample set into the LSTM autoencoder, map the sample to the feature space through the encoder, and then reconstruct the feature representation through the decoder to learn the feature; in the above process, the contrast loss is used to distinguish the samples in the feature space from different neural converters, and the reconstruction loss is used to measure the feature representation ability; finally, the network parameters of the neural converter and the autoencoder are updated iteratively by jointly optimizing the contrast loss and the reconstruction loss, so as to improve the performance of the overall model.

[0032] The abnormality detection steps based on the trained model are as follows: first, the test sequence is preprocessed in the same way as the training set, and the original sequence is divided into a set of sub-sequences through a sliding window; then, input these sub-sequences into the LSTM autoencoder to obtain the reconstructed sequence, and calculate the reconstruction error of each time point in the sequence as the abnormality score of the time point; next, the sub-sequences are spliced according to the division method to obtain the abnormality score sequence corresponding to the original test sequence; finally, compare the abnormality score with the preset judgment threshold to obtain the abnormality label of each time point in the test sequence. Specifically, if the abnormality score is greater than the threshold, it is determined to be abnormal, otherwise it is determined to be normal.

[0033] (1) Neural Transformer Data Augmentation

[0034] The anomaly detection method based on self-supervised learning promotes feature representation learning through constructed auxiliary tasks, and builds anomaly discriminators based on the extracted features to detect anomalies. However, for other types of data besides images, such as time series data and tabular data, it is difficult to design manual transformation methods for sample enhancement. Therefore, for satellite telemetry sequences, the present invention proposes a data enhancement method that uses neural network learning transformations, called neural transformers. This method replaces manual transformations with neural network transformations to learn different transformation methods, so that the transformed samples share semantic information with the original samples, and the samples generated by different transformation methods are easy to distinguish. The neural network model can be a recurrent neural network or a Transformer network, etc.

[0035] The pseudo code for the specific method of data augmentation using neural transformers is shown in Table 1. The main part of data augmentation is the proposed neural transformer. After the original sample is input, it is transformed by the neural network to generate intermediate samples, which are then processed through different data processing methods to obtain the final enhanced samples. Common processing methods are divided into three categories: (1) directly outputting the intermediate samples as enhanced samples; (2) multiplying the intermediate samples with the original samples to obtain the enhanced sample output; and (3) adding the intermediate samples to the original samples to obtain the enhanced sample output.

[0036] Table 1. Pseudocode for neural transformation learning data augmentation

[0037]

[0038]

[0039] Building on this foundation, we leverage the concept of contrastive learning to combine neural transformers with feature learning methods for self-supervised anomaly detection. First, we train the neural transformer to facilitate learning of different transformations. Then, using the learned transformers, we generate multiple augmented samples from the original samples, resulting in a collection of samples with pseudo-labeled categories. Next, we perform feature representation learning on the samples, mapping them into a feature space and enabling the distinction between samples from different sources. Finally, we jointly train the neural transformer and feature learning to iteratively optimize network parameters to achieve an anomaly detection model.

[0040] Unlike the method of extracting negative samples from the noise distribution for enhancement, the neural transformer proposed in this application deterministically constructs enhanced samples from the original samples. It is trained using the deterministic contrastive loss (DCL). DCL is a non-stochastic objective that does not require additional regularization. It encourages the neural transformer to learn different transformation methods to generate different enhanced samples, while each sample shares semantic information with the original sample.

[0041] Assume that there are N samples in the original data space χ There are K conversion modes in the neural converter {T1, T2, ..., T K |T k :χ→χ}. DCL encourages the sample x generated by the k-th transformer k =T k (x) is similar to its original sample x, and the sample x generated by other transformers is similar to the original sample x. l =T l (x), l≠k are different. Define the conversion score between two enhanced samples as h(x k x l ):

[0042]

[0043] Where τ is the temperature parameter, f φ (·): χ→γ is a feature extractor that maps samples from the original space χ to the feature space γ. The similarity between enhanced samples is measured by the cosine similarity of the samples in the feature space:

[0044]

[0045] Here, ||·|| represents the modulus of a vector.

[0046] Define the deterministic contrast loss as

[0047]

[0048] Where, Denotes the mathematical expectation E. The numerator promotes alignment of the enhanced samples generated by the neural converter with the original samples in the feature space, while the denominator promotes the separation of enhanced samples generated by different neural converters for the same original sample. During the training process of the neural converter, the network parameters are optimized by minimizing Equation (3).

[0049] (2) Autoencoder feature learning

[0050] Autoencoder is one of the commonly used feature learning models, which aims to learn the most important representation of the original data through neural network, and tries to reconstruct the original data using low-dimensional encoding in the feature space. The semi-supervised anomaly detection based on autoencoder is based on the hypothesis that normal samples can better reconstruct from the feature space than abnormal samples. Training the autoencoder and its variants using normal data can learn the hidden patterns and rules in normal data, and establish a model suitable for the reconstruction of normal data, and then detect anomalies according to the reconstruction error of the data to be tested.

[0051] Different neural network structures are usually used for different types of data. LSTM can capture long-term dependencies of sequences and is suitable for processing time series data. Therefore, for satellite telemetry sequences, the present application selects LSTM autoencoder for feature learning. The encoder and decoder of the LSTM autoencoder are two symmetrical LSTM networks. After the original sequence is input, it is mapped to a fixed-dimensional vector representation in the feature space through the encoder, and then the reconstructed sequence is generated through the decoder.

[0052] In the training process of the autoencoder, the reconstruction error (RL) is used as the loss function. The reconstruction error measures the difference between the reconstructed data and the original data. By minimizing the reconstruction error to optimize the network parameters, the autoencoder can learn and extract the most important features in the data.

[0053] The reconstruction loss is defined as

[0054]

[0055] In the formula, χ train is a training set composed of normal sequences. In the training process of the LSTM autoencoder, the network parameters are optimized by minimizing equation (4).

[0056] (3) Neural contrast enhanced autoencoder

[0057] In semi-supervised anomaly detection based on autoencoder, only normal samples are included in the training set. Compared with a large number of potential anomalies that may be represented in any form, the normal samples as training data are relatively limited, which may cause the anomaly detection model trained to overfit. In addition, due to the inability to maintain the diversity of normal samples, the learned representation may lack representativeness for complex time series data, which limits the generalization ability of the model.

[0058] Based on the foregoing analysis, the present application combines neural transformer data enhancement and LSTM autoencoder feature learning, and proposes a telemetry sequence time context anomaly detection method (NCAAE) based on neural contrast enhancement autoencoder. NCAAE is an end-to-end fully deep anomaly detection method, which belongs to self-supervised learning. In the training process, the original sample is data enhanced by using the neural transformer, and the sample in the separated feature space derived from different transformers is used as an auxiliary task for feature learning, to help the autoencoder learn more general and rich representations, and thus improve the performance of the model.

[0059] The deterministic contrast loss of the neural transformer and the reconstruction loss of the LSTM autoencoder are comprehensively considered, and a single objective function is proposed to perform end-to-end training and jointly optimize the network parameters. The loss function of NCAAE is defined as L:

[0060] L=L RL +λL DCL (5)

[0061] In the formula, λ∈[0,1] is a hyperparameter that balances the deterministic contrast loss L DCL and the reconstruction loss L RL weight. By unifying the two losses together, the optimization of the network parameters can be guided together to promote the neural transformer to learn different transformation methods, and at the same time promote the autoencoder to extract the most important representations.

[0062] The training process pseudo code of NCAAE is shown in Table 2. Among them, the samples in the training set are a set of sub-sequences obtained by randomly sampling and cutting the sliding window from multiple normal time sequences. First, the training sample is input into the neural transformer to generate multiple enhanced samples from the original sample, and a sample set containing the original sample and the enhanced sample is obtained. Then, the sample set is input into the LSTM autoencoder, which first maps the sample to the feature space through the encoder, and then reconstructs the input sequence using the extracted features through the decoder. Finally, by minimizing equation (5), the neural transformer and the LSTM autoencoder are jointly trained, and the network parameters are iteratively optimized to obtain the anomaly detection model.

[0063] Table 2 Training pseudo code of neural contrast enhancement autoencoder

[0064]

[0065] To verify the performance of the method, two public datasets commonly used in spacecraft anomaly detection field are selected for experiment, which are the telemetry parameter datasets of Mars Science Laboratory (MSL) Curiosity rover and Soil Moisture Active Passive (SMAP) satellite. On the SMAP and MSL datasets, the precision, recall and F1* score of the proposed NCAAE method and nine advanced deep anomaly detection benchmark methods are compared, and the results are shown in Table 3. For each algorithm, the hyperparameters are determined by grid search, and the experimental results recorded in the table are the optimal values.

[0066] Table 3 Comparison results of different methods

[0067]

[0068]

[0069] In Table 3, the optimal F1* score is indicated in bold, and the suboptimal F1* score is indicated in underline. As can be seen, on the SMAP and MSL datasets, the proposed NCAAE method obtains the optimal F1* score, and the anomaly detection ability is better than all benchmark methods. Specifically, on the SMAP dataset, the F1* score of the NCAAE method is 0.955, which is 4.83% higher than the suboptimal F1* score 0.911 obtained by the LSTM-NDT method; on the MSL dataset, the F1* score of the NCAAE method is 0.962, which is 0.94% higher than the suboptimal F1* score 0.953 obtained by the DAEMON method, proving the effectiveness of the proposed method.

[0070] For anomaly detection methods, if the precision obtained on the dataset is much higher than the recall, it means that the model tends to judge uncertain samples as normal, resulting in a higher missed detection rate. On the contrary, if the recall is much higher than the precision, it means that the model tends to judge uncertain samples as abnormal, resulting in a higher false alarm rate. The matrix stack diagram of the three evaluation indexes of each method on the SMAP and MSL datasets is shown in Figure 2 and Figure 3

[0071] ​It can be seen that the NCAAE method proposed in the application is obviously better than each benchmark method. At the same time, the precision and recall obtained by the NCAAE method are not much different, which shows that the model can better separate normal samples and abnormal samples, so as to find a better discriminant threshold boundary. In addition, the precision and recall obtained by the NCAAE method on the SMAP dataset are both greater than 0.93, and the precision and recall obtained by the NCAAE method on the MSL dataset are both greater than 0.96, which powerfully proves the robustness and robustness of the NCAAE method.

[0072] The experimental results prove that compared with the benchmark method, the NCAAE method proposed in the application uses the LSTM network to extract the time sequence characteristics of the data, considers the long-time dependence information, and through the joint optimization with the neural transformer used for data enhancement, promotes the self-encoder to learn more general and rich representations, and further increases the distinguishability of the normal and abnormal sequences after being reconstructed by the model, effectively improves the ability of anomaly detection.

[0073] In order to verify the effectiveness of introducing the neural transformer for data enhancement in the NCAAE method, an ablation experiment is performed by removing the module in the algorithm. The comparison method only contains the LSTM self-encoder part in the NCAAE method, which is called LSTM-AE. In the training process, the LSTM-AE optimizes the network parameters by minimizing the reconstruction error between the input subsequence and the reconstructed subsequence. After the model training is completed, the calculation method of the anomaly score in the test set is the same as that of the NCAAE method. On the SMAP and MSL datasets, the optimal precision, recall and F1* score of the NCAAE method and the LSTM-AE method are compared, and the results are shown in Table 4.

[0074] Table 4 Influence of data enhancement on anomaly detection results

[0075]

[0076] As can be seen from Table 4, under the condition that the recall rate has no obvious difference, the precision obtained by the NCAAE method is significantly improved compared with the LSTM-AE method which does not contain the data enhancement module. On the SMAP dataset, the F1* score obtained by the NCAAE method is improved by 0.74% compared with the LSTM-AE method; on the MSL dataset, the F1* score obtained by the NCAAE method is improved by 0.73% compared with the LSTM-AE method, which proves the effectiveness of introducing the neural transformer for data enhancement.

[0077] The application aims at the problem that training with limited normal samples in semi-supervised learning leads to lack of representation of learned representation, introduces neural transformation learning for data enhancement, learns different transformation methods by using a neural network, and generates enhanced samples with class pseudo labels, expands the diversity of training samples, and uses an LSTM autoencoder to extract time sequence features in telemetry parameters, and through joint optimization of the training process, improves the detection ability of the satellite telemetry parameter time context anomaly.

[0078] The application can also provide a computer device, comprising: at least one processor, a memory, at least one network interface and a user interface. The various components in the device are coupled together through a bus system. It can be understood that the bus system is used to realize the connection communication between the components. In addition to including a data bus, the bus system also includes a power bus, a control bus and a status signal bus.

[0079] The user interface can include a display, a keyboard or a pointing device. For example, a mouse, a trackball, a touchpad or a touch screen, etc.

[0080] It can be understood that the memory in the embodiments of the application can be a volatile memory or a non-volatile memory, or can include both volatile and non-volatile memories. The non-volatile memory can be a read-only memory (Read-Only Memory, ROM), a programmable read-only memory (Programmable ROM, PROM), an erasable programmable read-only memory (Erasable PROM, EPROM), an electrically erasable programmable read-only memory (Electrically EPROM, EEPROM) or a flash memory. The volatile memory can be a random access memory (Random Access Memory, RAM) used as an external cache. By way of example and not limitation, many forms of RAM can be used, such as static random access memory (Static RAM, SRAM), dynamic random access memory (Dynamic RAM, DRAM), synchronous dynamic random access memory (Synchronous DRAM, SDRAM), double data rate synchronous dynamic random access memory (Double Data Rate SDRAM, DDR SDRAM), enhanced synchronous dynamic random access memory (Enhanced SDRAM, ESDRAM), synchronous link dynamic random access memory (Synchlink DRAM, SLDRAM) and direct memory bus random access memory (Direct Rambus RAM, DRRAM). The memory described herein is intended to include but not limited to these and any other suitable types of memory.

[0081] In some embodiments, the memory stores elements, executable modules, or data structures, or a subset of them, or an extended set of them, such as an operating system and an application program.

[0082] The operating system includes various system programs, such as a framework layer, a core library layer, a driver layer, etc., for implementing various basic services and processing hardware-based tasks. The application program includes various application programs, such as a media player (Media Player), a browser (Browser), etc., for implementing various application services. The program for implementing the method of the embodiments of the present disclosure can be included in the application program.

[0083] In the above-described embodiments, the processor can be configured to perform the above-described method by invoking the program or instructions stored in the memory, in particular, the program or instructions stored in the application program.

[0084] The steps of the above-described method.

[0085] The above-described method can be applied to a processor or implemented by a processor. The processor can be an integrated circuit chip having a processing capability. In the implementation process, each step of the above-described method can be completed by integrated logic circuits or instructions in the form of software in the processor. The processor described above can be a general-purpose processor, a digital signal processor (DSP), an application specific integrated circuit (ASIC), a field programmable gate array (FPGA) or other programmable logic devices, discrete gate or transistor logic devices, discrete hardware components. Each method, step and logic block disclosed above can be implemented or executed. The general-purpose processor can be a microprocessor or the processor can also be any conventional processor. The steps of the method disclosed above 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, or other mature storage media in the art. The storage medium is located in the memory, and the processor reads the information in the memory and combines the hardware to complete the steps of the above-described method.

[0086] It can be understood that the embodiments described in the present application can be realized by hardware, software, firmware, middleware, microcode or a combination thereof. For hardware implementation, the processing unit can be implemented in 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), general purpose processors, controllers, micro-controllers, microprocessors, other electronic units designed to perform the functions described in the present application, or a combination thereof.

[0087] For software implementation, the present application can be implemented by executing the functional modules (such as processes, functions, etc.) described in the present application. The software code can be stored in a memory and executed by a processor. The memory can be implemented in the processor or outside the processor.

[0088] The present application can also provide a non-volatile storage medium for storing a computer program. When the computer program is executed by a processor, each step of the above method embodiments can be implemented.

[0089] Finally, it should be pointed out that the above embodiments are only used to illustrate the technical solutions of the present application and not to limit. Although the present application is described in detail with reference to the embodiments, those skilled in the art should understand that the technical solutions of the present application can be modified or replaced by equivalents without departing from the spirit and scope of the present application, and they should be covered in the scope of the claims of the present application.

Claims

1. A satellite telemetry parameter anomaly detection system based on neural contrast enhancement autoencoder, characterized in that: The system comprises: A neural converter data enhancement module, comprising a neural converter, configured to convert original data through the neural converter, and then enhance the converted data to obtain a plurality of enhanced samples of the original samples, and output a sample set including the original subsequence and the sequence synthesized by data enhancement; and The autoencoder feature module includes two symmetrical LSTM networks, an encoder and a decoder. The encoder maps samples to the feature space, and then reconstructs the feature representation through the decoder to obtain the anomaly label at each time point in the test sequence. The system is trained in an end-to-end manner. The training process includes: Input the historical normal telemetry sequence, perform preprocessing and sliding window segmentation, and obtain the subsequence set for training; Input the subsequence set into the neural transformer data enhancement module to generate multiple enhanced samples of the original samples, thereby obtaining a sample set including the original subsequence and the sequence synthesized by data enhancement; The sample set is input into the autoencoder feature module, the encoder maps the sample to the feature space, and then the feature representation is reconstructed by the decoder for feature learning; Contrastive loss is used to distinguish samples from different neural converters in the feature space, and reconstruction loss is used to measure the feature representation ability. By jointly optimizing the contrastive loss and reconstruction loss, the network parameters of the neural converter data enhancement module and the autoencoder feature module are iteratively updated; The neural transformer is trained using a deterministic contrastive loss method; The method for enhancing the converted data includes: Directly outputting the converted samples as enhanced samples, multiplying the converted samples with the original samples to obtain enhanced sample outputs, and adding the converted samples with the original samples to obtain enhanced sample outputs.

2. The satellite telemetry parameter anomaly detection system based on neural contrast enhanced autoencoder according to claim 1 is characterized in that: The training of the autoencoder feature module uses reconstruction error as the loss function.

3. A method for detecting satellite telemetry parameter anomalies based on a neural contrast-enhanced autoencoder, implemented based on the system of any one of claims 1-2, comprising: The satellite telemetry parameters are preprocessed and segmented and then input into the trained satellite telemetry parameter anomaly detection system based on neural contrast enhancement autoencoder to obtain the anomaly labels of the satellite telemetry parameters at each time point.

Images