A method for detecting outliers in MEO orbit electron detection data based on CNN

Abstract

The present invention provides a method for detecting outliers in MEO orbit electron detection data based on CNN, which preprocesses the detection data to remove useless information in the detection data; designs an outlier detection model based on CNN, inputs a training set into the model, uses the training set to complete training and establish a classifier to form a pattern for outlier detection; inputs a test set to test the outlier detection ability of the trained outlier detection model for outliers; obtains the MEO orbit electron detection data to be detected and inputs it into the CNN outlier detection model to carry out outlier detection on the data. The detection process of the present invention is fast, has high recognition accuracy, good reusability and strong robustness, and can provide a highly reliable data set for space environment monitoring and early warning, satellite fault analysis and space environment risk assessment.

Description

Technical Field

[0001] The present invention belongs to the field of space-based space environment exploration, and relates to the detection of outliers in MEO orbit electron detection data. Background Art

[0002] MEO (Medium Earth Orbit) is an orbit with a period of about 12 hours and an inclination of 55-63.43 degrees. Satellites operating in this orbit cross the core area of the outer radiation belt of the earth about four times a day. Its electron environment is characterized by large variation scales and high variation frequencies in terms of time and space. The scientific community has conducted in-depth research on the perturbation mechanism of magnetospheric energetic electrons in the quiet and disturbed states of near-earth space, and analyzed and proved that more than half of the failures of MEO satellites are caused by "killer electrons" in the outer radiation belt. Therefore, the scientific research field and the long-term management and operation control departments of on-orbit spacecraft have put forward higher requirements for the high-precision detection technology of electrons in the spacecraft operation orbit. At the same time, with the development of high-precision energy spectrum detection technology and energy electron pitch angle measurement technology, the high-precision detection of orbit electrons has become possible. Under such conditions, various countries have successively carried different technical systems of orbit electron detection payloads on satellites such as NOAA-POES, GPS, and RBSP and carried out monitoring activities. To meet the needs of orbit electron monitoring, a certain type of MEO satellite in China is equipped with an electron environment detection payload, and its detection data can be applied to space environment monitoring and early warning, satellite fault analysis, and satellite space environment risk assessment, filling the gap in domestic space environment exploration and having great engineering application benefits.

[0003] However, in practical applications, it is found that there are many outliers in such data, which may be caused by abnormal payload detection functions, data transmission, or ground data processing errors. These outliers not only affect the rationality of the monitoring of medium and high-energy electron environments, but also interfere with the accuracy of space environment monitoring and early warning, satellite fault analysis, and satellite space environment risk assessment, and have an adverse impact on subsequent space environment analysis and forecasting, satellite protection effectiveness evaluation, and even the long-term management and operation control of spacecraft. However, among the current methods for detecting outliers in orbit electrons, the physical process analysis method highly depends on physical theories. As shown in its analysis process Figure 1 For the analysis of electron environment detection data, it is necessary to analyze the time-varying characteristics, spatial characteristics, and energy spectrum characteristics of the data for different environmental states respectively. This method highly depends on the background knowledge of space physics, has high professional requirements for the analysis process, is complex and has low reusability, and the analysis results are greatly affected by human subjective factors. The efficiency of outlier detection is low and cannot meet the increasingly refined and rapid requirements for outlier detection.

[0004] Therefore, a method that can quickly and accurately identify outliers in detection data is needed to solve the annotation problem of MEO orbit electron detection data. Summary of the Invention

[0005] To overcome the deficiencies of the prior art, the present invention provides a method for detecting outliers in MEO orbit electron detection data based on a convolutional neural network (CNN), which detects and processes orbit electron monitoring data without labeled data quality information, and provides a high-confidence data set for subsequent space environment monitoring and early warning, satellite fault analysis, and space environment risk assessment through this data quality control method.

[0006] The technical solution adopted by the present invention to solve its technical problems includes the following steps:

[0007] (1) Preprocess the detection data, remove the useless information in the detection data, and convert the remaining valid information into matrix data; regard the numerical values of each element in the matrix data as the pixel values corresponding to the class grayscale image data to obtain the class grayscale image data;

[0008] (2) Design an outlier detection model based on CNN, including an input layer, a convolutional layer, a pooling layer, a fully connected layer, and an output layer;

[0009] (3) Select the detection data within a certain period as the total data set A, and the orbit electron data set of the detector during the abnormal period as B, Then the orbit electron data set of the detector during the normal period is

[0010] (4) Input the training set T train =X∪Y, the data set Use the training set T train Complete the training and establish a classifier to form a pattern for outlier detection; input the test set T test =(B\X)∪(C\Y) to test the outlier detection ability of the trained outlier detection model; the test set and the training set satisfy

[0011] (5) Obtain the MEO orbit electron detection data to be detected and input it into the CNN outlier detection model to perform outlier detection on the data.

[0012] In the process of converting the class grayscale image data in step (1), only 8 energy spectrum detection gear data in the direction of each detection unit in each detection cycle of the data to be processed are retained.

[0013] There are a total of 9 detection units. Each detection unit covers a field of view range of 20°, and merges the detection data within this range into detection data in one direction. The 9 detection units continuously cover a field of view range of 180°. The energy resolution scale of the detection data in the range of 50 - 600 KeV is divided into 8 energy bins, which are [50 KeV, 68 KeV], (68 KeV, 93 KeV], (93 KeV, 130 KeV], (130 KeV, 170 KeV], (170 KeV, 240 KeV], (240 KeV, 320 KeV], (320 KeV, 440 KeV], (440 KeV, 600 KeV).

[0014] The input layer in step (2) is used to control the parameters of the CNN outlier detection model; the convolutional layer is used to extract the image features of the class grayscale map data. This model contains 2 convolutional layers. The convolutional layer 1 is used to extract image features, and the convolutional layer 2 is used to extract image detail features; the pooling layer is used to reduce the model complexity and avoid overfitting; the fully connected layer is used to integrate the extracted image features; the output layer is used to output the model detection results.

[0015] The input of the CNN outlier detection model is a matrix of 9 rows and 8 columns; the model contains 2 convolutional layers for extracting image features. In convolutional layer 1, the kernel size is 3, the number of kernels is 6, and the stride is 1; in convolutional layer 2, the kernel size is 3, the number of kernels is 16, and the stride is 1; the pooling layer adopts the global pooling method; the ReLU function is selected as the activation function for this layer in the fully connected layer; the output layer outputs the model detection results, with normal labeled as 1 and abnormal labeled as 0.

[0016] The data set in step (3) covers data under three situations: the quiet period of the electronic environment, the disturbance period, and the detector anomaly period.

[0017] In step (5), the output result of the CNN outlier detection model is inserted into the data quality identification column of the data to be detected. Among them, the data quality identification of normal values is 1, and that of abnormal values is 0, completing the outlier detection and data annotation.

[0018] The beneficial effects of the present invention are as follows: Outlier detection is carried out based on a convolutional neural network, overcoming the weaknesses of traditional physical analysis methods, such as strong dependence on professional knowledge, complex analysis process, low annotation efficiency, and large subjective influence. It can better meet the requirements of timeliness and accuracy for space-based space environment detection data processing.

[0019] The outlier detection method for MEO orbit electron detection data proposed by the present invention is based on the image feature recognition technology of machine learning. It converts the data features of normal / outlier values under different environmental disturbance states in high-quality labeled data into image features for CNN model training, and converts the physical analysis method into the feature recognition mode of this model, so as to quickly conduct outlier detection and data labeling on the data to be detected. This method has two advantages: First, this method has the advantages of fast detection process and high recognition accuracy. When using 620,000 pieces of data as the dataset for training, the training time of this model is about 15 seconds. When using 630,000 pieces of data as the dataset for testing, the detection time of this model is about 1 second, the precision rate of outlier detection is 99%, and the recall rate is 99%. The above four indicators fully demonstrate its performance advantages. Second, this method has good reusability and strong robustness. During the testing process, this model shows good adaptability to the orbit electron detection data under different environmental disturbance states (both the training set and the test set contain detection data when the environment is calm / disturbed), and there is no need to repeatedly carry out model modification and debugging for different environmental states, showing good method reusability and robustness. Brief Description of the Drawings

[0020] Figure 1 The figure shows the analysis process of outlier detection by the traditional physical analysis method.

[0021] Figure 2 The figure shows the implementation steps of the outlier detection method for MEO orbit electron detection data based on CNN designed by the present invention.

[0022] Figure 3 The figure shows a schematic diagram of MEO orbit electron detection data to be detected for outlier detection. It can be seen from the figure that this data contains information such as time, direction, and electron flux detection values of each energy level. This data is in units of frames, and one frame of data contains 9 directions and 8 energy levels.

[0023] Figure 4 The figure shows the effect after converting Figure 3 one frame of data in it into grayscale-like image data.

[0024] Figure 5 The figure shows a schematic diagram of the physical structure of this detector. It can be seen from the figure the spatial coverage range of this detector.

[0025] Figure 6 The figure shows the MEO orbit electron detection data situation from January 1, 2020 to January 2, 2020. This figure reflects the characteristics of the detection data in terms of time, space, and energy spectrum distribution.

[0026] Figure 7The data of direction 4 of the MEO orbit electron detection data from January 1, 2020 to January 7, 2020 is shown. This figure reflects the characteristics of the detection data of the detector in the detection direction 4 in terms of time, space and energy spectrum distribution, and belongs to Figure 6 Supplementary description.

[0027] Figure 8 The data situation when the detector had abnormal data from July 1, 2019 to July 3, 2019 is shown. This figure reflects the characteristics of the detection data of the detector in the abnormal state in terms of time, space and energy spectrum distribution.

[0028] Figure 9 The characteristics of the detection data of the detector in the detection direction 4 during the abnormal period in terms of time, space and energy spectrum distribution are shown, and belong to Figure 8 Supplementary description.

[0029] Figure 10 The detection data situation of the detector in two states of normal / abnormal is shown. Among them, a) is the characteristics of the detection data of the detector in the normal state for the quiet electron environment, b) is the characteristics of the detection data of the detector in the normal state for the disturbed electron environment, and c) is the characteristics of the detection data of the detector in the abnormal state.

[0030] Figure 11 The structure diagram of the CNN outlier detection model designed by the present invention is shown. This figure truthfully reflects the network structure of the convolutional neural network and the parameter settings of each layer.

[0031] Figure 12 The data set situation for training and testing the CNN outlier detection model is shown. Among them, a) is the situation of the detection data set of the detector in the normal state, b) is the situation of the detection data set of the detector in the abnormal state, and c) is the situation of the overall data set.

[0032] Figure 13 The process of carrying out outlier detection on the MEO orbit electron detection data by using the CNN outlier detection model designed by the present invention is shown. Specific implementation manner

[0033] The present invention will be further described below in conjunction with the accompanying drawings and embodiments. The present invention includes but is not limited to the following embodiments.

[0034] The technical solution adopted by the present invention includes the following 4 steps:

[0035] (1) Detection data preprocessing

[0036] The outlier detection method for the MEO orbit electron detection data studied by the present invention first needs to solve the conversion problem between the detection data and the data of the grayscale-like image.

[0037] The MEO orbit electron detection data is unsigned single-precision floating-point numbers (unsigned float, with a numerical range of 0 - 6.8×10 38 ), while the data of each pixel point in the grayscale image is the grayscale value (gray scale, with a numerical range of 0 - 255). If the values in the direct detection data are normalized to the grayscale values, it will lead to the loss of precision information, resulting in the blurring of the data features of the grayscale image, which is not conducive to the detection of outliers based on image features. The technical solutions to solve this problem are as follows:

[0038] a. Remove the useless information in the detection data;

[0039] b. Convert the valid information in the detection data into matrix data;

[0040] c. Treat the values of each element in the matrix data as the pixel values corresponding to the grayscale-like image data;

[0041] d. Save the converted grayscale-like image data.

[0042] (2) Design of the outlier detection model based on convolutional neural network (CNN)

[0043] The CNN outlier detection model involved in this invention includes an input layer, a convolutional layer, a pooling layer, a fully connected layer, and an output layer. The basic functions of each layer are as follows:

[0044] a. The input layer is used to control the parameters of the CNN model;

[0045] b. The convolutional layer is used to extract the image features of the grayscale-like image data. This model includes 2 convolutional layers. Convolutional layer 1 is used to extract image features, and convolutional layer 2 is used to extract image detail features;

[0046] c. The pooling layer is used to reduce the model complexity and avoid overfitting;

[0047] d. The fully connected layer is used to integrate the extracted image features;

[0048] e. The output layer is used to output the model detection results.

[0049] (3) Model training and testing

[0050] Considering the dynamic development characteristics of electrons in the MEO orbit, the dataset should cover data under three situations: the quiet period of the electron environment, the disturbed period, and the abnormal period of the detector. Therefore, we selected the detection data for a total of 578 days from December 1, 2018 to June 20, 2020 as the total dataset. During this period, there were 493 days of quiet orbital electrons, and 85 days of orbital electron disturbances caused by high-energy electron bursts; during the period, the detector had abnormal detection data for 123 days. During this period, there were 75 days of quiet orbital electrons and 48 days of orbital electron disturbances caused by high-energy electron bursts.

[0051] Let the overall dataset be A, and the orbital electron dataset during the abnormal period of the detector be B. At this time, there is Then the orbital electron dataset during the normal period of the detector is

[0052] To verify the CNN outlier detection model designed by the present invention, a learning sample dataset (training set) needs to be input into the model. At this time, there is the dataset Then the training set T train = X ∪ Y. The model uses the training set T train to complete training and establish a classifier to form a pattern for outlier detection; at this time, a test set needs to be input to test the detection ability of the trained outlier detection model for known outliers. The training set T test = (B\X) ∪ (C\Y). The test set and the dataset should satisfy to prevent the model from having logical self-consistency due to duplicate data in the training set and the test set.

[0053] (4) Outlier detection

[0054] Obtain the MEO orbit electron detection data to be detected and input it into the CNN outlier detection model to perform outlier detection on the data. Insert the output result of the CNN outlier detection model into the data quality identification column of the data to be detected, where the data quality identification of normal values is 1 and that of outliers is 0. Complete outlier detection and data annotation through this method.

[0055] The following example discloses a method for detecting outliers in MEO orbit electron detection data based on CNN, and its steps are as Figure 2 shown.

[0056] (1) Data preprocessing

[0057] The MEO orbit electron detection data to be detected is as Figure 3 shown.

[0058] As Figure 3 shown in the example data, the numerical values of each energy spectrum detection gear are unsigned single-precision floating-point numbers (usignedfloat, numerical range 0 - 6.8×1038 ) while the pixel value of a typical grayscale image is the gray scale value (with a value range of 0 - 255). If the values of each point in the data to be processed are directly normalized to the gray scale value, it will cause loss of precision information, resulting in blurring of the grayscale image features, which is not conducive to carrying out outlier detection based on image features. Therefore, during the conversion of the grayscale-like image data, only the data of 8 energy spectrum detection positions in 9 directions in each detection period of the data to be processed are retained, obtaining grayscale-like image data in which the pixel values correspond one-to-one with the electron detection data. Its specific form is as Figure 4 shown.

[0059] (2) CNN model design

[0060] Before designing the CNN model, it is necessary to first determine the most simplified structure of the model input data according to the detector structure characteristics and data characteristics, so as to determine the CNN model structure and the parameters of the input layer, convolutional layer, and pooling layer.

[0061] (2.1) Detector characteristic analysis

[0062] According to the detector design, the detector and its probe combination structure are as Figure 5 shown.

[0063] From Figure 5 , it can be seen that this detector is composed of three probes, and each probe is composed of 3 SI-PIN position-sensitive detectors (hereinafter referred to as detection units). Each detection unit can cover a 20° field of view range and merge the detection data within this range into the detection data in 1 direction

[18] . Based on the above conclusion, it can be known that this instrument can detect the orbital electrons in 9 directions and its detection data has significant direction anisotropy, that is, the data has spatial characteristics.

[0064] At the same time, according to the theory of the space particle radiation field, the spatial electron distribution has the characteristic of continuous distribution, and its energy spectrum is a continuous spectrum [1 - 6]. Therefore, in terms of energy resolution, the instrument research and development team carried out simulation design on the energy spectrum resolution of each detection unit of the detector through the Monte Carlo method, determined the geometric factor and the center frequency of the detection energy spectrum of each detection unit of the detector through ground calibration of the detector, and finally divided the energy resolution scale of the detection data of this detector in the range of 50 - 600 keV into 8 energy levels and combined them to form the energy spectrum detection data of this detector

[18] . The specific energy level distribution of this detection data is shown in Table 1.

[0065] Table 1 Energy spectrum detection position distribution of a certain MEO orbital electron detector

[0066] Serial number Energy spectrum detection gear Energy spectrum coverage range 1. Energy spectrum detection gear 1 [50KeV, 68KeV] 2. Energy spectrum detection gear 2 (68KeV, 93KeV] 3. Energy spectrum detection gear 3 (93KeV, 130KeV] 4. Energy spectrum detection gear 4 (130KeV, 170KeV] 5. Energy spectrum detection gear 5 (170KeV, 240KeV] 6. Energy spectrum detection gear 6 (240KeV, 320KeV] 7. Energy spectrum detection gear 7 (320KeV, 440KeV] 8. Energy spectrum detection gear 8 (440KeV, 600KeV]

[0067] The direction characteristics of the data within the same time period are as Figure 6 shown.

[0068] The energy spectrum characteristics of the detection data within a certain period in the same direction are as follows Figure 7 shown.

[0069] As can be seen from the above process, the detection data of this instrument has direction anisotropy. The energy spectrum detection gears in each direction are distributed in sequence. The detection data reflects the spatial and physical characteristics of orbital electrons, follows the law of direction and energy spectrum continuity, and the sequence cannot be interchanged.

[0070] (2.2) Abnormal data analysis

[0071] When the detector shows an abnormality, the situation of the detection data of this orbital electron is as follows Figure 8 shown.

[0072] From Figure 8 it can be seen that when such an abnormality occurs, except for the detection data in direction 4, the detection data in all other directions is 0 for this detector. Figure 9 Shown is the characteristic of the abnormal detection data in direction 4.

[0073] From Figure 9 analysis, it can be seen that during the abnormal period, the detection data in detection direction 4 has the saturation situation of multiple energy gear detection values, and the law between the characteristics of the detection data in each energy gear is very different from that in the normal state. According to the different space environment situations, the comparison of the characteristics during the quiet period, disturbance period, and abnormal period of the space environment is as follows Figure 10 shown.

[0074] CNN principle analysis:

[0075] In the field of deep learning, in order to solve the disadvantages of traditional neural networks, such as having a large number of required weights, a large amount of computation, and a large training sample set, scientists invented and designed a convolutional neural network with the characteristics of reducing weights, local connection, and weight sharing, which reduces the number of parameters that need to be trained in the neural network through the receptive field and weight sharing. Currently, CNN is often used to extract features of things with a specific model, and then classify, identify, predict, or make decisions on the things according to the features. The characteristics of this model are more suitable for the abnormal value detection work of orbital electron detection data.

[0076] CNN generally includes an input layer, a convolutional layer, a pooling layer, a fully connected layer, and an output layer. Among them, the input layer mainly completes the function of selecting an appropriate input neuron size and retaining the data itself format.

[0077] Among them, the convolutional layer extracts data features through six steps: receptive field, weight matrix assignment, weight sharing, stride setting, boundary expansion, and feature mapping; the pooling layer uses non-linear downsampling to reduce the amount of computation and control overfitting by reducing the parameters of the network; the fully connected layer refits the features to reduce the loss of feature information; the output layer completes the output of the target result.

[0078] Design of CNN Outlier Detection Model:

[0079] Based on the detector design and its data analysis, the following conclusions are obtained:

[0080] a. The feature image data is a matrix of 9 rows and 8 columns (manifested as 72 pixel points in the class grayscale image data);

[0081] b. There are obvious spatial distribution characteristics between the detection data in different directions, manifested as Figure 10 the brightness information transition characteristics within each column of pixels in, and the characteristic area contains at least 2 pixel points;

[0082] c. There are obvious energy distribution characteristics between the detection data in the same direction but different energy levels, manifested as Figure 10 the brightness information transition characteristics within each row of pixels in, and the characteristic area contains at least 2 pixel points;

[0083] d. There are obvious image characteristics between the calm state and the disturbed state, manifested as Figure 10 the change characteristics of the high-frequency component area of the image in a) and b), and the characteristic area contains at least 2×2 pixel points;

[0084] e. The image characteristics are significantly abnormal under the abnormal state of the detector, manifested as Figure 10 the change of the high-frequency component area of the image in c).

[0085] Based on the above conclusions, the design idea of this CNN model is determined as follows:

[0086] a. The input of the CNN model is a matrix of 9 rows and 8 columns;

[0087] b. The model contains 2 convolutional layers for extracting image features. In convolutional layer 1, the kernel size is 3, the number of kernels is 6, and the stride is 1; in convolutional layer 2, the kernel size is 3, the number of kernels is 16, and the stride is 1;

[0088] c. The pooling layer is used to reduce the model complexity and avoid overfitting, and this step is completed by using the global pooling method;

[0089] d. The fully connected layer is used to integrate the extracted image features, and the ReLU function is selected as the activation function for this layer;

[0090] e. The output layer is used to output the model detection result, that is, normal is labeled as 1 and abnormal is labeled as 0.

[0091] Based on the above CNN model design idea, the structure of the CNN outlier detection model is as Figure 11 shown.

[0092] (3) Training and Testing of Outlier Detection Model

[0093] (3.1) Training of Outlier Detection Model

[0094] Considering the dynamic development characteristics of electrons in the MEO orbit, the training set and test set of the outlier detection model designed in this invention should cover the data under three situations: the quiet period of the electronic environment, the disturbance period, and the abnormal period of the detector. Therefore, we selected the detection data for a total of 578 days from December 1, 2018 to June 20, 2020 as the total data set. During this period, there were 493 days of quiet orbital electrons and 85 days of orbital electron disturbances caused by high-energy electron bursts; during this period, the detector had abnormal detection data for 123 days. During this period, there were 75 days of quiet orbital electrons and 48 days of orbital electron disturbances caused by high-energy electron bursts. The basic situation of the data set is as Figure 12 shown.

[0095] Now let the overall data set be A, and the orbital electron data set during the abnormal period of the detector be B. At this time, there is Then the orbital electron data set during the normal period of the detector is To verify the CNN outlier detection model designed in this invention, a learning sample data set (training set) needs to be input into the model. At this time, there is the data set Then the training set T train = X ∪ Y. The model uses the training set T train to complete the training and establish a classifier to form a pattern for outlier detection; at this time, a test set needs to be input to test the detection ability of the trained outlier detection model for the existing outliers. The training set T test = (B\X) ∪ (C\Y). The test set and the data set should satisfy to prevent the model from having logical self-consistency due to duplicate data in the training set and the test set.

[0096] (3.2) Testing of Outlier Detection Model

[0097] When using 620,000 MEO orbital electron detection data as the training set for training, the training time of this model is about 15 seconds. When using 630,000 MEO orbital electron detection data as the test set for testing, the detection time of this model is about 1 second, the precision rate of outlier detection is 99%, and the recall rate is 99%. The above four indicators prove that this model meets the requirements for outlier detection of current MEO orbital electron detection data and can be used to implement outlier detection of this detection data.

[0098] Using the above-mentioned CNN-based outlier detection method for MEO orbit electron detection data, perform outlier detection on unlabeled detection data, insert a data quality identification column in the orbit electron detection data, where the labeled data quality identified as normal by the CNN model is marked as 1, and the outlier is marked as 0. Complete outlier detection and data annotation according to the identification result of the model for the data to be measured. The process is as Figure 13 shown.

Claims

1. A method for detecting outliers in MEO orbit electron detection data based on CNN, characterized in that, It includes the following steps: (1) Preprocess the detection data, remove the useless information in the detection data, and convert the remaining valid information into matrix data; regard the numerical value of each element in the matrix data as the pixel value corresponding to the class grayscale image data to obtain the class grayscale image data; (2) Design an outlier detection model based on CNN, which includes an input layer, a convolutional layer, a pooling layer, a fully connected layer and an output layer; (3) Select the detection data within a certain period as the total data set A, and the orbital electron data set during the abnormal period of the detector is B. Then the orbital electron data set during the normal period of the detector is (4) Input the training set T into the CNN outlier detection model train = X ∪ Y, the data set Utilize the training set T train Complete the training and establish a classifier to form a pattern for outlier detection; input the test set T test = (B\X) ∪ (C\Y) to test the outlier detection ability of the trained outlier detection model; the test set and the training set satisfy (5) Obtain the MEO orbit electron detection data to be detected and input it into the CNN outlier detection model to carry out outlier detection on the data.

2. The method for detecting outliers in MEO orbit electron detection data based on CNN according to claim 1, characterized in that, In the process of converting the class grayscale image data in step (1), only 8 energy spectrum detection gear data in the direction of each detection unit in each detection period of the data to be processed are retained.

3. The method for detecting outliers in MEO orbit electron detection data based on CNN according to claim 2, wherein There are 9 detection units in total. Each detection unit covers a 20° field of view range and merges the detection data within this range into the detection data in one direction. The 9 detection units continuously cover a 180° field of view range; the energy resolution scale of the detection data in the range of 50 - 600 KeV is divided into 8 energy bins, which are [50 KeV, 68 KeV], (68 KeV, 93 KeV], (93 KeV, 130 KeV], (130 KeV, 170 KeV], (170 KeV, 240 KeV], (240 KeV, 320 KeV], (320 KeV, 440 KeV], (440 KeV, 600 KeV].

4. The method for detecting outliers in MEO orbit electron detection data based on CNN according to claim 1, characterized in that, The input layer in step (2) is used to control the parameters of the CNN outlier detection model; the convolutional layer is used to extract the image features of the class grayscale image data. This model includes 2 convolutional layers. The convolutional layer 1 is used to extract image features, and the convolutional layer 2 is used to extract image detail features; the pooling layer is used to reduce the model complexity and avoid overfitting; the fully connected layer is used to integrate the extracted image features; the output layer is used to output the model detection result.

5. The outlier detection method for MEO orbit electron detection data based on CNN according to claim 1, wherein The input of the CNN outlier detection model is a matrix of 9 rows and 8 columns; the model includes 2 convolutional layers for extracting image features. In the convolutional layer 1, the kernel size is 3, the number of kernels is 6, and the stride is 1; in the convolutional layer 2, the kernel size is 3, the number of kernels is 16, and the stride is 1; the pooling layer adopts the global pooling method; the fully connected layer selects the ReLU function as the activation function of this layer; the output layer outputs the model detection result, with normal labeled as 1 and abnormal labeled as 0.

6. The method for detecting outliers in MEO orbit electron detection data based on CNN according to claim 1, wherein The data set in step (3) covers the data under three situations: the quiet period of the electronic environment, the disturbance period and the abnormal period of the detector.

7. The method for detecting outliers in MEO orbit electron detection data based on CNN according to claim 1, wherein In step (5), insert the output result of the CNN outlier detection model into the data quality identification column of the data to be detected, where the data quality identification of the normal value is 1, and the abnormal value is identified as 0, to complete the outlier detection and data annotation.

Images