Method and system for detecting time sequence anomaly of satellite telemetry data

Abstract

The invention discloses a method and a system for detecting time sequence anomaly of satellite telemetry data. The method comprises the following steps: normalizing satellite telemetry time sequence data X; dividing the normalized X into a plurality of window data Xw with the length of w by adopting a sliding window strategy; inputting each Xw into a generator G of the generative adversarial network for mask interpolation, and generating covered data with the length of w; comparing Xw with concealed data bit by bit, and if a difference value of a certain bit exceeds a threshold value, determining that the window data is abnormal; the generative adversarial network further comprises a discriminator D in the training stage, the input of the D is window data and concealed data, the output of the D is the probability of whether the concealed data is real data, and the data generated by the G better conforms to the distribution characteristics of the X through training. According to the method, the accuracy, the robustness and the adaptive capacity of anomaly detection are improved, and a new technical scheme is provided for the anomaly detection of the satellite telemetry parameters.

Description

Technical Field

[0001] The present invention belongs to the field of ground operation and control of space exploration equipment, and particularly relates to a method and system for detecting anomalies in the time series of satellite telemetry data. Background Art

[0002] Telemetry parameters are data that the satellite system transmits sensor-recorded information such as voltage and temperature to the ground system in a specified form as data frames. Detecting and analyzing anomalies in telemetry parameters in a timely manner to handle abnormal situations is an effective means to ensure the stable and reliable operation of the satellite.

[0003] The anomaly detection of single-parameter data can be divided into two categories: point-based anomalies and sequence-based anomalies. Point anomalies refer to individual data points in a time series that deviate significantly from the normal values, which can be further divided into single-point anomalies and context anomalies. A single-point anomaly means that an isolated data point is significantly different from other data points; while a context anomaly means that a data point appears abnormal under a specific context or condition. Sequence anomalies involve multiple consecutive data points in a time series that jointly exhibit an abnormal pattern, called collective anomalies, which are manifested as abnormal fluctuations or trends of a series of data points and usually violate the overall behavior or pattern of the time series.

[0004] The anomaly detection problem is essentially a binary classification problem, and its goal is to divide data into two categories: normal and abnormal. In the field of single-parameter time series anomaly detection, this task specifically refers to evaluating whether a given time series follows a normal data distribution pattern and identifying whether it contains the above three types of anomalies. Due to the extreme lack of labels in real satellite telemetry data, unsupervised learning methods that do not rely on labeled data are more needed to detect anomalies.

[0005] There are already some methods for detecting anomalies in satellite telemetry parameters, but it is still a complex problem worthy of in-depth study. First, due to the complex periodicity of satellite motion, it is difficult to model and analyze telemetry parameters, and the effectiveness of existing methods needs to be improved. Second, existing methods rarely consider the phenomena that occur in anomaly datasets, which results in poor robustness of the detection methods and may only perform well on specific datasets. Summary of the Invention

[0006] The purpose of the present invention is to overcome the defects of the prior art and propose a method and system for detecting anomalies in the time series of satellite telemetry data.

[0007] The present invention proposes a method for detecting anomalies in satellite telemetry parameters based on a time series interpolation generative adversarial network. It extracts time series features through a one-dimensional convolutional neural network and uses a generative adversarial network to model the distribution of telemetry parameters. It innovatively adopts an interpolation-based detection method, effectively improving the accuracy of anomaly detection and the adaptability to complex anomaly situations.

[0008] In view of this, the present invention proposes a method for detecting abnormal time series of satellite telemetry data, including:

[0009] A method for detecting abnormal time series of satellite telemetry data, including:

[0010] Normalize the satellite telemetry time series data X;

[0011] Adopt a sliding window strategy to divide the normalized X into several window data X with a length of w w ;

[0012] Input each window data X w into the generator G of the generative adversarial network for mask imputation to generate masked data with a length of w; Compare each bit of the window data X w with the masked data bit by bit. If the difference of a certain bit exceeds the set threshold, the window data is abnormal;

[0013] During the training stage, the generative adversarial network further includes a discriminator D. The input of the discriminator D is the window data and the masked data generated by the generator G, and the output is the probability that the masked data is real data. Through training, the data generated by the generator G is made to better conform to the distribution characteristics of X.

[0014] Preferably, the generator G of the generative adversarial network is implemented based on 1DCNN, including three one-dimensional convolutional layers with different convolutional kernel sizes and a convolutional output layer.

[0015] Preferably, the processing process of the generator G includes:

[0016] Receive the normalized window data X w , judge the difference between each data point in X w and the window mean value, and modify the data points with large deviations to the mask -1 to obtain the modified data Perform a convolution operation on , and then perform an exclusive OR operation with the mask to obtain generated data X w with the same length as the window data X g , that is, the masked data.

[0017] Preferably, the discriminator D is implemented based on 1DCNN, including four one-dimensional convolutional kernels and a fully connected output layer.

[0018] Preferably, the processing process of the discriminator D includes:

[0019] The received window data and the masked data generated by the generator G are analyzed for data distribution through a convolution operation, and a scalar is output through the final fully connected layer and the sigmoid activation function, which is used to represent the probability that the masked data is real data. When the scalar is 0, it is determined that the data is masked data, and when the scalar is 1, it is determined that the data is the original data.

[0020] Preferably, in the training stage of the generative adversarial network, the generator G and the discriminator D are trained simultaneously based on an unsupervised learning method. The training objective is to adjust the parameters of G to minimize 1 - D(G(x)), and at the same time adjust the parameters of D to maximize D(x). The loss function is:

[0021]

[0022] where x represents the real data sample, p data(x) represents the probability distribution of the real data, D(x) represents the probability that the discriminator D believes x is real data, G(x) represents the masked data generated by interpolating from the real data, represents the mathematical expectation.

[0023] On the other hand, the present invention provides a system for satellite telemetry data time series anomaly detection, including: a normalization module, a sliding window module, a mask interpolation module, and an anomaly detection module; wherein,

[0024] The normalization module is used to normalize the satellite telemetry time series data X;

[0025] The sliding window module is used to divide the normalized X into several window data X with a length of w by adopting a sliding window strategy w ;

[0026] The mask interpolation module is used to input each window data X w into the generator G of the generative adversarial network for mask interpolation to generate masked data with a length of w;

[0027] The anomaly detection module is used to compare the window data X bit by bit w with the masked data. If the difference in a certain bit exceeds the set threshold, the window data is abnormal;

[0028] The generative adversarial network also includes a discriminator D in the training stage. The input of the discriminator D is the window data and the masked data generated by the generator G, and the output is the probability that the masked data is real data. Through training, the data generated by the generator G is made to more conform to the distribution characteristics of X.

[0029] Compared with the prior art, the advantages of the present invention are:

[0030] 1. One of the core innovations of the present invention lies in addressing the problem of single-parameter anomaly detection in satellite telemetry parameters. A method is proposed to model normal telemetry parameters using a generative adversarial network (GAN) and a one-dimensional convolutional neural network (1DCNN). This method can solve the limitations of traditional methods in dealing with complex time-series data. Especially in the anomaly detection of satellite telemetry parameters, traditional methods often struggle to accurately capture the time-series characteristics and internal laws of the data. By using GAN, the present invention can learn and simulate the distribution characteristics of normal telemetry parameter data, providing a strong model foundation for anomaly detection. This modeling approach not only improves the accuracy of anomaly detection but also enhances the model's ability to recognize unknown anomaly patterns.

[0031] 2. Another innovation is the time-series interpolation strategy adopted by the present invention. This strategy effectively detects abnormal situations. Especially when faced with complex abnormal situations, it can avoid the deficiencies of traditional reconstruction- or prediction-based methods. In traditional reconstruction-based and prediction-based anomaly detection methods, when faced with dense or complex abnormal data, the performance of the model is often significantly affected, resulting in a decline in the accuracy and stability of the detection results. The present invention pre-identifies potential abnormal points by introducing a time-series interpolation step, replaces the values of these points with a specific marker, and then uses the trained model to interpolate these marker points to generate a complete data sequence. By comparing the differences between the interpolated data and the original data, the present invention can accurately identify abnormal points or abnormal sequences, thereby improving the accuracy and robustness of anomaly detection.

[0032] 3. The present invention not only proposes a new anomaly detection framework theoretically but also demonstrates significant advantages in practical applications. By modeling normal telemetry parameters using a generative adversarial network, the present invention can accurately capture and simulate the distribution characteristics of normal data, providing a solid foundation for anomaly detection. At the same time, through the time-series interpolation strategy, the present invention can effectively handle complex abnormal situations, improving the adaptability and robustness of the model. The comprehensive application of these innovation points makes the present invention have important practical value and broad application prospects in the field of satellite telemetry parameter anomaly detection. Brief Description of the Drawings

[0033] Figure 1 is a schematic diagram of the method framework for time-series anomaly detection of satellite telemetry data in the present invention;

[0034] Figure 2 is a schematic diagram of the detection logic generated by interpolation;

[0035] Figure 3 is the anomaly detection algorithm framework, where Figure 3(a) is the original sequence data and the abnormal interval, Figure 3(b) is the original sequence data and the interpolated generated data, and Figure 3(c) is the original sequence data and the anomaly score;

[0036] Figure 4 It is a comparison of data generated by different detection algorithms;

[0037] Figure 5 It is the influence of different abnormal concentrations on the effectiveness of the algorithm;

[0038] Figure 6 It is the flow chart of model training;

[0039] Figure 7 It is the flow chart of the detection algorithm. Detailed implementation

[0040] The present invention aims to provide a satellite telemetry parameter anomaly detection method based on a time series interpolation generative adversarial network. This method extracts time series features through a one-dimensional convolutional neural network and uses a generative adversarial network to model the distribution of telemetry parameters. It innovatively adopts an interpolation-based detection method, effectively improving the accuracy of anomaly detection and the adaptability to complex anomaly situations.

[0041] 1. Basic framework of the detection algorithm

[0042] The core idea of the method is generative adversarial and interpolation detection. A generative adversarial network is used to model telemetry parameters, and time series interpolation features are used to solve the data generation and detection problems under complex anomaly conditions.

[0043] The one-dimensional convolutional neural network (1DCNN) used in the present invention is a special form of CNN. As Figure 1 shown, its convolutional kernel slides only in a single dimension, that is, on time series or one-dimensional spatial data. For a single-parameter time series x and a convolutional kernel k, its convolutional operation is:

[0044]

[0045] where t is the length of the time series x, is the convolutional operator, n k is the size of the convolutional kernel k, S t is the output of the corresponding one-dimensional convolutional operation. After the convolutional operation, the length of S t is n x -n k +1, where n x is the time step of the sequence x.

[0046] In the time series anomaly detection task, the generator G needs to learn the data distribution p g from the original sequence data X, for generating new data x i from the data sequence x g。The goal of the discriminator D is to maximize the distinction between real data and imputed data. The discriminator D(x; θ d ) outputs a scalar representing the probability that x is real data. When x ∼ p data (x), the discriminator is expected to output 1, and conversely, when x ∼ p g (x), the expected output is 0. In GAN, the generator G and the discriminator D are trained simultaneously. The training goal is to adjust the parameters of G to minimize 1 - D(G(x)), and at the same time adjust the parameters of D to maximize D(x). The loss function is:

[0047]

[0048] where x represents real data samples, p data(x) represents the probability distribution of real data, D(x) represents the probability that the discriminator believes x is real data, G(x) represents the masked data generated by imputing from real data, represents the mathematical expectation.

[0049] For the time - series anomaly detection task, since the proportion of abnormal data is extremely small and has a negligible impact on the overall data distribution characteristics, learning can be carried out through unlabeled data. By combining GAN with 1DCNN, the distribution characteristics of time - series telemetry parameters can be learned, and then anomaly detection can be performed based on this.

[0050] 2. Operation process based on imputation detection

[0051] Deep - learning - based methods in anomaly detection usually start from the original data, assuming that the impact of anomalies is small, to learn the feature distribution of the data. However, when there are many abnormal data or they appear concentratedly, it will cause a deviation in the data distribution and affect the detection effect. To solve the limitations of existing methods in satellite telemetry parameter anomaly detection, this study proposes an innovative imputation strategy.

[0052] The core of this strategy is to introduce the pre - identification and imputation steps of anomaly points in the anomaly detection process to improve the model's ability to identify anomalies and reduce the impact of dense anomaly points on the model performance. As Figure 2 shown, some data in the original sequence are transformed into - 1, and then new sequence data are generated through imputation. Whether it is abnormal is judged according to the deviation of the imputation points.

[0053] Compared with the prediction-based method, the points selected by the imputation strategy proposed in this study are not continuous, effectively avoiding the interference of continuous abnormal data on the detection results. At the same time, compared with the reconstruction-based method, this strategy retains a part of the original data as known points, improving the stability of the imputation process. Based on the error between the imputed data and the original data for anomaly judgment, abnormal patterns in the data can be captured more accurately, improving the accuracy of anomaly detection.

[0054] The technical solution of the present invention will be described in detail below with reference to the accompanying drawings and embodiments.

[0055] Embodiment 1

[0056] Specific implementation process

[0057] To detect anomalies in the time series data X, first perform normalization on it so that all data is within the interval [0, 1], which is convenient for subsequent detection. Then adopt the sliding window strategy to divide X into several time windows X of length w w .

[0058] The overall detection algorithm includes two major parts: feature learning and anomaly detection.

[0059] The feature learning part is overall composed of a generator G and a discriminator D. Both G and D are implemented relying on 1DCNN. The discriminator D is also implemented by 1DCNN and includes 4 one-dimensional convolutional layers and a fully connected output layer. The input of the discriminator D is the window data and the masked data generated by the generator G, and the output is the probability that the masked data is real data. Through training, the data generated by the generator G is made to conform more to the distribution characteristics of X. Specifically, the discriminator D accepts a time series X, analyzes its data distribution through convolutional operations, and outputs a scalar through the final fully connected layer and the sigmoid activation function. 0 indicates that the data is judged to be generated data, and 1 indicates that the data is judged to be original data. The generator G includes 3 one-dimensional convolutional layers with different convolutional kernel sizes and a convolutional output layer, and accepts a normalized time series X w , modifies n data points to -1 according to the degree of deviation of the data points from the sequence, indicating the mask. After the modified data undergoes a series of convolutional operations and then an exclusive OR operation with the mask, that is, only the data at the positions with a value of -1 is modified, and the rest of the data remains the original input, thus obtaining the generated data X with the same length as the input sequence g . The specific process is as Figure 6 shown.

[0060] The anomaly detection part is completed by the generator. As Figure 7 shown, it specifically includes:

[0061] Adopt a sliding window strategy to divide the normalized X into several window data X with a length of w w ;

[0062] Input each window data X w into the generator G of the generative adversarial network for mask imputation to generate masked data with a length of w; Compare each bit of the window data X w with the masked data. If the difference of a certain bit exceeds the set threshold, the window data is abnormal.

[0063] Demonstration of the algorithm detection effect:

[0064] To verify the detection effect of the TiGAN of the present invention, we selected a time series sample containing three known abnormal intervals for analysis. Figure 3 shows the detection results: Figure 3(a) shows the original data and the marks of its known abnormal intervals; Figure 3(b) compares the imputed generated data with the original data. It can be seen that the imputed data generally follows the distribution law of the original data, but there are obvious differences between the imputed data and the original data near some significant abnormal points, which is consistent with the original intention of the algorithm design; Figure 3(c) shows the final abnormal detection results. All known abnormal intervals are successfully identified. In addition, the algorithm also detects some potential abnormal intervals, and the numerical performance of these intervals is significantly different from the surrounding data, which is worthy of further analysis. Generally speaking, in the case of no abnormal labels, the algorithm successfully models the original time series through the generative adversarial learning mechanism. The generated data not only conforms to the distribution characteristics of the original data, but also shows certain differences at potential abnormal points, and is suitable for comparative detection.

[0065] Figure 4 Shows the comparison of the imputation-based method proposed by the present invention with traditional reconstruction or prediction-based methods in terms of generating data. Figure 5 Shows the influence of different abnormal concentrations on the effectiveness of the algorithm. It can be seen that the prediction-based method performs well near single-point anomalies and is basically not affected, but in the collective anomaly area, the predicted generated data will be significantly disturbed; while the reconstruction-based method is the opposite, being less affected by collective anomalies, but single-point anomalies are likely to cause large-scale fluctuations in the generated data, which may lead to misjudgment. In contrast, the imputation-based method of the present invention performs more stably in different abnormal areas. It not only performs well under single-point anomalies, but is also less affected in the collective anomaly area and can better handle complex abnormal situations.

[0066] Example 2

[0067] Example 2 of the present invention provides a system for satellite telemetry data time series anomaly detection, which is implemented based on the method of Example 1. The system includes: a normalization module, a sliding window module, a mask imputation module, and an anomaly detection module; among them,

[0068] The normalization module is used to normalize the satellite telemetry time series data X;

[0069] The sliding window module is used to divide the normalized X into several window data X with a length of w by adopting a sliding window strategy; w ;

[0070] The mask imputation module is used to input each window data X w into the generator G of the generative adversarial network for mask imputation to generate masked data with a length of w; The anomaly detection module is used to compare each bit of the window data X w with the masked data. If the difference of a certain bit exceeds the set threshold, there is an anomaly in the window data;

[0071] In the training stage, the generative adversarial network also includes a discriminator D. The input of the discriminator D is the window data and the masked data generated by the generator G, and the output is the probability that the masked data is real data. Through training, the data generated by the generator G is made to better conform to the distribution characteristics of X.

[0072] Comparison with the effects of other algorithms:

[0073] To quantitatively demonstrate the effects proposed in the present invention, this article selects the Z-score algorithm based on statistical values, the LOF algorithm based on distance clustering, the DBSCAN algorithm based on density clustering, the Bi-LSTM algorithm based on prediction, and the TAnoGAN method based on reconstruction for comparison. The research uses evaluation indicators widely recognized in the field of anomaly detection, including Precision, Recall, and F1 Score. The datasets used come from the real telemetry data of a certain space science satellite and the publicly available device anomaly dataset. The detection results are shown in the following table.

[0074] Table 1 Comparison of detection results of various detection algorithms on the dataset

[0075]

[0076] From the perspective of overall performance, the TiGAN algorithm shows good performance and achieves high-level performance in the selected datasets. Specifically, in the RK, ZK, and AWS datasets, TiGAN obtains the highest F1 score, and in the DY dataset, its F1 score is also close to the optimal value. In particular, on the ZK dataset, compared with the second-ranked algorithm, the F1 score of TiGAN is increased by about 20%, showing a significant performance advantage. In addition, in multiple datasets, TiGAN also achieves a high recall rate.

[0077] From the perspective of the practicality of the algorithm, traditional statistical and clustering-based methods perform well on single-point anomaly data due to their inherent algorithmic mechanism limitations. However, in datasets containing a large number of context anomalies such as ZK and AWS, their detection performance will decline or even fail. Compared with other deep learning-based methods, TiGAN shows better stability in detection effectiveness. Especially in the AWS dataset where anomalies are concentrated, TiGAN not only achieves the best F1 score but also maintains its recall rate within an acceptable range, indicating that the imputation structure has a certain anti-interference ability against concentrated anomalies and proving its potential and reliability in the actual anomaly detection of satellite telemetry parameters.

[0078] Algorithm robustness detection:

[0079] To verify the necessity of introducing adversarial learning in the detection algorithm, we designed two experimental setups: one includes a feature learning module (with the adversarial learning round set to 10), and the other does not include this module (with the adversarial learning round set to 0). The experimental results, as shown in Table 2, prove the importance of the data distribution feature learning module combining 1DCNN and GAN for improving the performance of the detection algorithm. The results show that the synergistic effect of the feature learning module and the detection module is the key to fitting complex data distribution features and achieving efficient anomaly detection.

[0080] Table 2 Comparison of the impact of the feature learning module on the detection results

[0081]

[0082] Table 3 further explores the impact of selecting different numbers of imputation points on the detection results when the fixed sliding time window size is 20. When the imputation number is set to 0, i.e., no imputation is performed, we use it as the control group of the experiment. It should be noted that imputation numbers of 1 and 20 represent prediction and reconstruction methods respectively. The experimental results show that the impact of the imputation number on the algorithm effect is not obvious in a regular pattern. However, when the imputation number is 10, i.e., half of the time window, the algorithm effect is relatively stable, and previous algorithms also used this size of imputation.

[0083] Table 3 Comparison of the impact of the imputation number on the detection results

[0084]

[0085] The proposed TiGAN algorithm in the present invention is based on an unsupervised learning approach, aiming to learn the distribution patterns of data from unlabeled satellite telemetry parameters. The premise of this learning process is that the proportion of abnormal data is relatively low and has little impact on the overall data distribution. Therefore, in this study, datasets with different abnormal concentrations are set to demonstrate the influence of the proportion of abnormal data in the overall data on the effectiveness of the algorithm. The experimental results are as Figure 5 shown. The Z-Score method and the DBSCAN method show a significant decline in effectiveness or even fail when the abnormal concentration is high; while the Bi-LSTM and LOF methods perform poorly when the abnormal concentration is low; although the TanoGAN method performs well overall, its effectiveness fluctuates

[0086] greatly. The proposed TiGAN algorithm in this paper shows good detection effectiveness when the proportion of abnormal data does not exceed 25%. However, when the proportion of abnormal data exceeds 25%, its detection effectiveness begins to decline. Given that the proportion of abnormal data in satellite telemetry parameters is usually much lower than 1%, the TiGAN algorithm can operate stably in this low-abnormal-concentration environment, ensuring its reliability and effectiveness in practical applications.

[0087] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention and not to limit them. Although the present invention has been described in detail with reference to the embodiments, those of ordinary skill in the art should understand that any modification or equivalent replacement of the technical solutions of the present invention does not depart from the spirit and scope of the technical solutions of the present invention, and they should all be covered within the scope of the claims of the present invention.

Claims

1. A method for detecting time series anomalies in satellite telemetry data, comprising: Normalize the satellite telemetry time series data X; Using the sliding window strategy, the normalized X is divided into several window data X with a length of w. w ; Each window data X w Input the generator G of the generative adversarial network for mask interpolation to generate masked data of length w; compare the window data X bit by bit w If the difference between a certain bit and the masked data exceeds the set threshold, the window data is abnormal; The generative adversarial network also includes a discriminator D during the training stage. The input of the discriminator D is the window data and the masked data generated by the generator G, and the output is the probability of whether the masked data is real data. Through training, the data generated by the generator G is more consistent with the distribution characteristics of X.

2. The method for detecting satellite telemetry data time series anomaly according to claim 1, characterized in that: The generator G of the generative adversarial network is implemented based on 1DCNN, including three one-dimensional convolutional layers with different convolution kernel sizes and a convolutional output layer.

3. The method for detecting satellite telemetry data time series anomaly according to claim 2, characterized in that: The processing of the generator G includes: Receive normalized window data X w , judge X w The difference between each data point and the window mean in the window is used to modify the data points with large deviations to mask -1, and the modified data is obtained. right After the convolution operation, the XOR operation is performed with the mask to obtain the window data X w Generate data X of the same length g , that is, the data has been masked.

4. The method for satellite telemetry data time series anomaly detection according to claim 3, characterized in that: The discriminator D is implemented based on 1DCNN, including 4 one-dimensional convolution kernels and a fully connected output layer.

5. The method for satellite telemetry data time series anomaly detection according to claim 4, characterized in that: The processing process of the discriminator D includes: Receive window data and masked data generated by generator G, analyze data distribution through convolution operation, and output a scalar through the final fully connected layer and sigmoid activation function to indicate the probability of whether the masked data is real data. When the scalar is 0, it indicates that it is masked data, and when the scalar is 1, it indicates that it is original data.

6. The method for satellite telemetry data time series anomaly detection according to claim 1, characterized in that: In the training phase of the generative adversarial network, the generator G and the discriminator D are trained simultaneously based on an unsupervised learning method. The training goal is to adjust the parameters of G to minimize 1-D(G(x)) and adjust the parameters of D to maximize D(x). The loss function is: Among them, x represents the real data sample, p data(x) represents the probability distribution of real data, D(x) represents the probability that the discriminator D considers x to be real data, G(x) represents the masked data generated by interpolation from real data, represents mathematical expectation.

7. A system for detecting satellite telemetry data time series anomalies, characterized in that: include: Normalization module, sliding window module, mask interpolation module and anomaly detection module; among them, The normalization module is used to perform normalization processing on the satellite telemetry time series data X; The sliding window module is used to adopt a sliding window strategy to divide the normalized X into a number of window data X with a length of w. w ; The mask interpolation module is used to convert each window data X w Input the generator G of the generative adversarial network for mask interpolation to generate masked data of length w; The abnormality detection module is used to compare the window data X bit by bit. w If the difference between a certain bit and the masked data exceeds the set threshold, the window data is abnormal; The generative adversarial network also includes a discriminator D during the training stage. The input of the discriminator D is the window data and the masked data generated by the generator G, and the output is the probability of whether the masked data is real data. Through training, the data generated by the generator G is more consistent with the distribution characteristics of X.

Images