Multi-sensor spectrum anomaly detection method based on metric-adversarial learning

Abstract

The invention discloses a multi-sensor spectrum anomaly detection method based on metric-adversarial learning, and the method comprises the steps: taking real-time spectrum data collected by multiple sensors in parallel as input, employing a deep metric learning model to extract the semantic features of the preprocessed spectrum data, building a semantic center feature vector, and obtaining a semantic center feature vector; calculating a mahalanobis distance between the vectors as a first measurement index of spectrum anomaly detection; the generative adversarial network model is used for reconstructing spectrum data, and obtaining a reconstruction error as a second measurement index; and finally, weighting the two metrics to obtain a final metric, calculating a relationship between the metric and a spectrum training exception threshold, and judging an exception condition. According to the method, a multi-sensor parallel acquisition mode is adopted, high-dimension unknown position and frequency illegal frequency spectrums are effectively detected, the detection scale is expanded, and meanwhile, the mode of combining two measurement indexes of mahalanobis distance and reconstruction error is adopted, so that the accuracy and performance of abnormal frequency spectrum detection are improved.

Description

Technical Field

[0001] The present invention relates to the technical field of spectrum detection, and more specifically to a multi-sensor spectrum anomaly detection method based on metric-adversarial learning. Background Art

[0002] With the popularization of Internet of Things technology, the electromagnetic environment in key areas has become increasingly tense and complex. Spectrum resources not only cover the normal activities of legal users, but also some illegal users, which has intensified the competition and abuse of spectrum resources. The surge in wireless devices and communication terminals has made it particularly difficult to ensure spectrum security in key areas and frequency bands, which poses certain challenges to the effective management and utilization of spectrum resources. Based on this, it is particularly important to develop a fast and efficient spectrum anomaly detection technology.

[0003] Traditional spectrum anomaly detection algorithms, such as those based on energy analysis, mainly rely on complex statistical models such as likelihood ratio tests or Bayesian frameworks. However, the prior knowledge that the model relies on, especially when the spectrum anomaly phenomenon presents diverse characteristics, will greatly limit the generalization ability and make it difficult to adapt to the complex and changing actual environment. In wireless communication systems, the use of spectrum resources is highly dynamic and real-time. Therefore, spectrum anomaly detection algorithms need to have good real-time processing capabilities so that they can respond and process anomalies within a short time. Although the current spectrum anomaly detection methods based on deep learning have shown considerable potential, they still have two major limitations. First, the applicability of these methods in actual complex environments is limited, and it is difficult to fully adapt to the changing detection scenarios. Second, most deep learning methods only use a single performance indicator to evaluate the detection effect, and lack the ability to comprehensively consider from multiple dimensions. This weakens the robustness and generalization ability of the detection system to a certain extent, making it possible to face performance fluctuations or instability in actual applications.

[0004] Patent document CN117851947A proposes a spectrum anomaly detection method and system based on time-frequency series difference and single-class support vector machine. This method adopts a universal design to solve the problems of poor universality and repeated design of existing detection methods. Through the customization of interference models, it is easy to realize various forms of RF spectrum anomaly simulation, establish multiple feature quantities for detection, and broaden the abnormal spectrum recognition surface. However, this solution does not consider solving the spectrum anomaly detection problem from a higher dimensional signal feature, and it is difficult to effectively identify spectrum anomalies caused by illegal users whose locations and working frequency bands are uncertain.

[0005] Patent document CN112924749A proposes an unsupervised adversarial learning method for detecting abnormal signals in the electromagnetic spectrum. This method combines the local characteristics of power spectrum data and uses convolutional neural networks to introduce adversarial ideas into the detection of abnormal signals in the electromagnetic spectrum. However, this method only considers the impact of reconstruction errors on electromagnetic anomalies, and does not consider the monitoring of abnormal spectrum ranges. Summary of the invention

[0006] 1. Technical content to be solved

[0007] In view of the problems that existing spectrum anomaly detection technologies are difficult to detect illegal users with unknown locations and frequency bands, and the number of single detections is small and the detection is single, the present invention proposes a multi-sensor spectrum anomaly detection method based on metric-adversarial learning; spectrum acquisition data for multiple sensors is reconstructed to obtain reconstruction accuracy, and then deep metric learning is introduced to obtain the Mahalanobis distance to expand spectrum detection of unknown dimensional bands; weighted processing of the two metrics effectively improves the spectrum detection performance.

[0008] 2. Technical solution

[0009] The purpose of the present invention is achieved through the following technical solutions.

[0010] The present application provides a multi-sensor spectrum anomaly detection method based on metric-adversarial learning, including: S1 using multiple sensors to collect data in the same frequency band at the same time, and dividing and preprocessing the collected spectrum data to obtain training samples and test samples; S2 using the training samples to pre-train the metric-adversarial learning module to obtain a training model for spectrum anomaly detection: the process of pre-training the deep metric learning module is to input the training samples into an encoder to obtain spectral semantic features, introduce a triple loss function to train and optimize the distance between the sample and the class center, control the classification error, improve the reliability of the semantic features, and reduce the mean square error between the input data and the reconstructed data; the process of pre-training the adversarial network module is to further obtain the latent features after the encoder extracts the semantic features. In the feature, upsampling operation is performed through the decoder to complete spectrum reconstruction, and adversarial loss is introduced to improve the accuracy of reconstructed spectrum; S3 collects the input of the sample to be tested, and according to the trained metric-adversarial learning model, the reconstruction error d1 generated by the adversarial network model to reconstruct the spectrum data is used as the first metric; the minimum Mahalanobis distance d3 between the class center and the test sample after deep metric learning optimizes the center loss is used as the second metric; the two metrics are weighted to obtain the final metric: d = γd1 + (1-γ) d3, and the relationship between the final metric d and the spectrum training anomaly score threshold is judged, and the threshold η is selected from the anomaly score of the training spectrum data set; if the final metric is greater than the threshold, the spectrum detection is abnormal; if it is less than or equal to the threshold, the spectrum detection is normal.

[0011] Furthermore, the multi-sensor acquisition process and normalization process include splicing the power spectrum density data received by different spectrum sensors into a spectrum data matrix of size M×N for normalization, where M is the number of sensors and N is the number of frequency points; the sensor’s perception result at s is [T s1 ,T s2 ,…,T sN ], and s=1,2,…M, is a sensor during the acquisition, and the sensor’s perception result at j is [T 1j ,T 2j ,…,T Mj ], j = 1, 2...N, is a certain frequency point, and the final input spectrum matrix is

[0012]

[0013] Furthermore, the deep metric learning module consists of an encoder and a classifier. The encoder contains three convolutional layers, a pooling layer, and two fully connected layers, while the classifier contains two fully connected layers and a Softmax activation function. The preprocessed spectrum data is input into the encoder of the deep metric learning module to obtain semantic features, and the triple loss function training is introduced to optimize the distance between the sample and the class center, control the classification error, improve the reliability of the semantic features, and reduce the mean square error between the input data and the reconstructed data; the triple loss function introduced in the pre-training of deep metric learning includes:

[0014]

[0015] L CL represents the center loss, where z i Represents semantic features, represents the input sample class center vector, R represents the input sample batch size, i represents the sample batch, Represents the norm of the processed class center spectrum vector;

[0016]

[0017] L CE represents the cross entropy loss, where k i represents the input sample label, Represents the predicted label, R represents the input sample batch size, and i represents the sample batch;

[0018]

[0019] L MSE represents the mean square error loss, where F i represents the input spectrum data, represents the reconstructed spectrum data, R represents the input sample batch size, and i represents the sample batch. The final optimization target total loss L obtained by combining the triple loss with training is expressed as: minL=min{λL CL +L CE +L MSE}, where λ is the control coefficient of the center loss.

[0020] Furthermore, the modules of the generative adversarial network module include an encoder, a decoder, and a discriminator. The encoder part is shared with the deep metric learning module, the decoder consists of three upsampling layers and a pooling layer, and a fully connected layer, and the discriminator includes three convolutional layers and two fully connected layers. The adversarial loss function introduced by the discriminator in the adversarial network model is: Where L ad (E,D s ) represents the adversarial training loss, D s represents the discriminator, E represents the encoder, P(F) represents the distribution of the input spectrum data, Represents the mathematical expectation of the input spectral data distribution.

[0021] Furthermore, the first metric reconstruction error of model training is defined as:

[0022] Among them, d1 represents the reconstruction error, F sj Represents the test spectrum data, represents the reconstructed spectrum data; M represents a total of M spectrum sensors, N represents a total of N frequency points, s represents the sth sensor, and j represents the jth frequency point;

[0023] The second metric, Mahalanobis distance, is defined as:

[0024]

[0025] Where d2 represents the Mahalanobis distance between the test sample and the class center, and d3 is the minimum Mahalanobis distance; where E represents the encoder, f represents the test sample, and P k Representative class center, Represents the inverse of the covariance matrix of the test sample.

[0026] The preprocessed test samples are further input into the metric-adversarial model for model verification. The codec is input into each layer of the adversarial network for calculation to obtain the reconstruction error d1. The deep metric learning layers are input to calculate the minimum Mahalanobis distance d3. Finally, the spectrum training anomaly score is obtained, and the model judgment threshold is set according to the anomaly score.

[0027] Furthermore, the two metrics are weighted to obtain the relationship between the final metric and the spectrum training anomaly threshold η for judgment. The weighted processing parameter d is defined as: d = γd1 + (1-γ)d3, where d represents the final metric for spectrum anomaly detection. d1 represents the reconstruction error, d3 represents the minimum Mahalanobis distance, and γ is the control coefficient of the reconstruction error and the Mahalanobis distance. The threshold η is selected from the anomaly score of the training spectrum data set. When d>η, it is judged as abnormal, otherwise it is considered normal.

[0028] 3. Beneficial effects

[0029] (1) The present invention considers spectrum anomaly detection in an electromagnetic environment and uses the correlation of multi-sensor perception results as the raw data for anomaly detection, effectively improving the detection performance of spectrum anomalies caused by illegal users with unknown locations and unknown working frequency bands;

[0030] (2) The present invention proposes a more robust spectrum anomaly detection method, which adopts a dual metric. First, the Mahalanobis distance between the semantic feature vector and the semantic center feature vector is calculated as the first metric, and the reconstruction error is used as the second metric. The dual metric method not only reduces the spectrum missed detection rate in the multi-dimensional space, but also further captures the abnormal points between the feature space and the normal samples to the extreme, effectively making up for the spectrum anomaly detection missed by a single metric. BRIEF DESCRIPTION OF THE DRAWINGS

[0031] Figure 1 It is a flow chart of the multi-sensor spectrum anomaly detection method based on metric-adversarial learning of the present invention;

[0032] Figure 2 This is a schematic diagram for comparing the detection performance of different algorithms of the present invention;

[0033] Figure 3 It is a schematic diagram of the spectrum anomaly detection performance under different numbers of sensors of the present invention. DETAILED DESCRIPTION

[0034] The present invention is described in detail below with reference to the accompanying drawings and specific embodiments.

[0035] A multi-sensor spectrum anomaly detection method based on metric-adversarial learning, the specific steps are:

[0036] 1: Collect spectrum data for preprocessing;

[0037] Through the parallel acquisition method of multiple sensors, the spectrum data of the specified frequency band is collected as the input of model training, and divided into training samples and test samples in a ratio of 8:2. The collected spectrum data is sliced ​​to meet the size of the input neural network model, and the spectrum data is normalized.

[0038] Specifically, the power spectrum density data received by different spectrum sensors are spliced ​​into a spectrum data matrix of size M×N for normalization processing, where M is the number of sensors and N is the number of frequency points; the perception result of the sensor at k is [T s1 ,T s2 ,…,T sN ], and s=1,2,…M, is a sensor during the acquisition, and the sensor’s perception result at j is [T 1j ,T 2j ,…,T Mj ], j = 1, 2...N, is a certain frequency point, and the final input spectrum matrix is

[0039] 2: Pre-training of deep metric learning module and generative adversarial network module;

[0040] Specifically, the deep metric learning module includes an encoder and a classifier. The encoder consists of three convolutional layers, a pooling layer, and two fully connected layers, and the classifier consists of two fully connected layers and a Softmax activation function. The normal spectrum data of the completed slice is input into the encoder of the deep metric learning module to obtain semantic features, and the training process is optimized by introducing triple losses, which include: cross entropy loss, center loss, and reconstruction loss.

[0041] Specifically, we first establish the class center for the normal spectrum data, and optimize the distance between the training samples and the class center through the center loss. By optimizing the distance between each batch of training samples and the class center, we can improve the accuracy of the obtained class center vector. CL As shown below: where z i Represents semantic features, Represents the input sample class center vector, k i represents the sample label, R represents the input sample batch size, i represents the sample batch, Represents the norm of the processed class center spectrum vector.

[0042] Secondly, the cross entropy loss is used to control the effectiveness of obtaining semantic features. By controlling the classification error, the reliability of semantic features is improved, and the accuracy of the center loss clustering effect is improved. The cross entropy loss L CE As shown below: Among them, k i represents the input sample label, Represents the predicted label, R represents the input sample batch size, and i represents the sample batch.

[0043] Reconstruction loss is then used to reduce the mean square error between the input data and the reconstructed data, so that the model can better learn the distribution law of normal spectrum data. The reconstruction loss L MSE As shown below: where F i represents the input spectrum data, Represents the reconstructed spectrum data, R represents the input sample batch size, and i represents the sample batch.

[0044] In order to enhance the robustness of the model and improve its ability to obtain effective features, the triple loss is combined as the training loss, so the final optimization target total loss L is as follows: minL = min{λL CL +L CE +L MSE}, where L CL is the center loss, L CE is the cross entropy loss, L MSE is the reconstruction loss, λ is the control coefficient of the center loss, and its value is set to 0.03.

[0045] The normal spectrum data of the completed slice is input into the generative adversarial network module to obtain potential features. The training process is optimized by introducing adversarial loss to gradually improve the accuracy of spectrum reconstruction.

[0046] Specifically, the generative adversarial network module includes an encoder, a generator (decoder), and a discriminator. The encoder part is shared with the deep metric learning module, the decoder consists of three upsampling layers and a pooling layer, and a fully connected layer, and the discriminator includes three convolutional layers and two fully connected layers. During training, the preprocessed spectrum data is input into the encoder to obtain potential features, and the generator is used to upsample the decoder to complete the spectrum reconstruction. During the adversarial training process, the generator is trained to deceive the discriminator so that the discriminator classifies the reconstructed spectrum data as real spectrum data. On the other hand, the task of the discriminator is to accurately distinguish the source of the spectrum data.

[0047] The optimization goal is that the generator can generate synthetic spectrum data that is more similar to the real spectrum data, and the discriminator tries its best to distinguish. In this adversarial training process, the generative adversarial network module gradually improves the spectrum reconstruction accuracy of the model. The adversarial loss is as follows: Where L ad (E,D s ) represents the adversarial training loss, D s represents the discriminator, F represents the input spectrum data, E represents the encoder, and P(F) represents the distribution of the input spectrum data. Represents the mathematical expectation of the input spectral data distribution.

[0048] The preprocessed test samples are input into the metric-adversarial model for model verification. The codec is input into each layer of the adversarial network for calculation to obtain the reconstruction error d1. The codec is input into each layer of the deep metric learning to calculate the minimum Mahalanobis distance d3. Finally, the spectrum training anomaly score is obtained, and the model judgment threshold is set according to the anomaly score.

[0049] 3: Spectral data anomaly detection;

[0050] The spectrum data sample to be tested is input into the trained generative adversarial network model, the test spectrum data is reconstructed, and the reconstruction error of the test sample is obtained.

[0051] Specifically, the calculation steps of the reconstruction error are as follows:

[0052] Among them, d1 represents the reconstruction error, F sj Represents the test spectrum data, represents the reconstructed spectrum data; M represents a total of M spectrum sensors, N represents a total of N frequency points, s represents the sth sensor, and j represents the jth frequency point.

[0053] Furthermore, the trained deep metric learning module is used to calculate the Mahalanobis distance between the test sample and the class center. The calculation method is as follows:

[0054] Where d2 represents the Mahalanobis distance between the test sample and the class center, E represents the encoder, f represents the test sample, P k Representative class center, Represents the inverse of the covariance matrix of the test sample.

[0055] Choose the minimum distance as the discriminant metric: Among them, d3 represents the smallest Mahalanobis distance in d2;

[0056] In order to magnify the gap between abnormal samples and normal samples, these two metrics are weighted to obtain the final metric for anomaly detection: d = γd1 + (1-γ) d3, where d represents the final metric for spectrum anomaly detection. d1 represents the reconstruction error, d3 represents the minimum Mahalanobis distance, and γ is the control coefficient of the reconstruction error and the Mahalanobis distance, which is set to 0.5. The threshold η is selected from the anomaly score of the training spectrum dataset. When d>η, it is judged as abnormal, otherwise it is considered normal.

[0057] The technical effects of the present invention are further illustrated below through simulation experiments.

[0058] 1. Simulation conditions:

[0059] The simulation scenario was established using MATLAB;

[0060] The specific setting is a 2000×2000m 2 The monitoring frequency band is set to 400-420MHz;

[0061] The noise factor of the spectrum sensor is 11dB, and the propagation model is set to a free propagation model, in which the power of a certain frequency point at a certain location collected by the spectrum sensor is the superposition of the power of all signal sources;

[0062] The noise power spectral density is set to -174dBm / Hz;

[0063] Illegal user intrusion interference anomaly refers to unauthorized or illegal electromagnetic signals. Their transmission power, working bandwidth, and working frequency cannot be obtained in advance, and they will interfere with the normal communication behavior of legitimate users, maliciously seize spectrum resources, and affect the spectrum management of the electromagnetic environment. The presence and power of illegal users cannot be determined in advance. The present invention considers the detection performance of illegal users at multiple locations and different power levels;

[0064] There are three legal users, and the specific settings are shown in Table 1. The transmit power of all legal users is set to 20mW, and channel 3 is a protection channel. Three illegal users are preset, and their locations are (-500m, 500m), (1000m, -1000m), and (900m, 900m).

[0065] The training set contains a total of 3,600 normal training samples, which are divided into four types according to the working mode of legitimate users, and 900 normal samples are collected for each type;

[0066] The test set contains 4400 samples, including 400 normal samples and 4000 abnormal samples caused by illegal users with different power levels.

[0067] Table 1 Legal user parameter settings

[0068] Legal users Legitimate user 1 Legitimate User 2 Legitimate user 3 Working Mode Continuous emission Intermittent emission Intermittent emission Location (-500m, -500m) (500m, -500m) (500m, 500m) Channel Channel 1 Channel 2 Channel 4

[0069] 2. Simulation content

[0070] Figure 2 This is a comparison chart of the detection performance of the present invention and the existing E-GAN and SR2CNN methods. The horizontal axis represents the ratio of the transmission power of illegal users to the transmission power of legal users, and the vertical axis represents the area under the performance curves of the three methods. The triangle, circle and star marks respectively represent the spectrum anomaly detection performance of the three abnormal users under different methods. This figure shows that the detection framework of the present invention has a significant improvement in the detection of intrusion interference of illegal users compared with other algorithms.

[0071] Figure 3The performance comparison corresponding to the simultaneous input detection of different numbers of sensors is shown, such as Figure 3 As shown, the present invention adopts a detection method using 16 sensors, and the detection performance is improved by 11.1% compared with the detection method using a single sensor.

[0072] The above schematically describes the invention and its implementation methods, which is not restrictive. Without departing from the spirit or basic features of the invention, the invention can be implemented in other specific forms. What is shown in the accompanying drawings is only one of the implementation methods of the invention. The actual structure is not limited to this, and any figure mark in the claims should not limit the claims involved. Therefore, if a person of ordinary skill in the art is inspired by it, without departing from the purpose of the invention, a structural method and an embodiment similar to the technical solution are designed without creativity, which should all fall within the scope of protection of this patent. In addition, the word "including" does not exclude other elements or steps, and the word "one" before the element does not exclude the inclusion of "multiple" elements. The multiple elements stated in the product claim can also be implemented by one element through software or hardware. The words first, second, etc. are used to indicate names, and do not indicate any specific order.

Claims

1. A multi-sensor spectrum anomaly detection method based on metric-adversarial learning, characterized in that: The following steps are involved: S1 uses multiple sensors to collect data in the same frequency band at the same time, and divides and preprocesses the collected spectrum data to obtain training samples and test samples; S2 uses training samples to pre-train the metric-adversarial learning module to obtain a training model for spectrum anomaly detection: The pre-training process of the deep metric learning module is to input the training samples into the encoder to obtain the spectral semantic features, introduce the triple loss function training to optimize the distance between the sample and the class center, control the classification error, improve the reliability of the semantic features, and reduce the mean square error between the input data and the reconstructed data; The process of pre-training the adversarial network module is as follows: after the encoder extracts semantic features, it further obtains latent features, performs upsampling operations through the decoder, completes spectrum reconstruction, and introduces adversarial loss to improve the accuracy of reconstructed spectrum; S3 collects the metric-adversarial learning model obtained by inputting the sample to be tested, and uses the reconstruction error d1 generated by the adversarial network model to reconstruct the spectrum data as the first metric; the minimum Mahalanobis distance d3 between the class center and the test sample after the deep metric learning optimizes the center loss as the second metric; the two metrics are weighted to obtain the final metric: d=γd1+(1-γ)d3, Determine the relationship between the final metric d and the spectrum training anomaly score threshold, where the threshold η is selected from the anomaly score of the training spectrum dataset; if the final metric is greater than the threshold, the spectrum detection is abnormal; if it is less than or equal to the threshold, the spectrum detection is normal.

2. The multi-sensor spectrum anomaly detection method based on metric-adversarial learning as claimed in claim 1, characterized in that: The multi-sensor acquisition process and normalization process include splicing the power spectrum density data received by different spectrum sensors into a spectrum data matrix of size M×N for normalization, where M is the number of sensors and N is the number of frequency points; the sensor’s perception result at s is [T s1 ,T s2 ,…,T sN ], s=1,2,…M, is a sensor during the acquisition, and the sensor’s perception result at j is [T 1j ,T 2j …,T Mj ], j = 1, 2, ... N, is a certain frequency point, and the final input spectrum matrix is 3. The multi-sensor spectrum anomaly detection method based on metric-adversarial learning as described in claim 1 is characterized in that: The deep metric learning module consists of an encoder and a classifier. The encoder contains three convolutional layers, a pooling layer, and two fully connected layers, while the classifier contains two fully connected layers and a Softmax activation function.

4. The multi-sensor spectrum anomaly detection method based on metric-adversarial learning as claimed in claim 3, characterized in that: The preprocessed spectrum data is input into the encoder of the deep metric learning module to obtain semantic features, and the triple loss optimization module is introduced. The triple loss function introduced in the pre-training of deep metric learning is as follows: L CL represents the center loss, where z i Represents semantic features, represents the input sample class center vector, R represents the input sample batch size, i represents the sample batch, ‖z i -P ki ‖ represents the norm of the processed class center vector; L CE represents the cross entropy loss, where k i represents the input sample label, Represents the predicted label, R represents the input sample batch size, and i represents the sample batch; L MSE represents the mean square error loss, where F i represents the input spectrum data, Represents the reconstructed spectrum data, R represents the input sample batch size, and i represents the sample batch.

5. The multi-sensor spectrum anomaly detection method based on metric-adversarial learning as claimed in claim 4, characterized in that: The final optimization target total loss L obtained by combining the triple loss with training is as follows: minL=min{λL CL +L CE +L MSE }; Where λ is the control coefficient of the center loss.

6. The multi-sensor spectrum anomaly detection method based on metric-adversarial learning as claimed in claim 1, characterized in that: The generative adversarial network module consists of an encoder, a decoder, and a discriminator; the encoder part is shared with the deep metric learning module, the decoder consists of three upsampling layers and a pooling layer, and a fully connected layer, and the discriminator includes three convolutional layers and two fully connected layers.

7. The multi-sensor spectrum anomaly detection method based on metric-adversarial learning as claimed in claim 6, characterized in that: The adversarial loss function introduced by the discriminator in the adversarial network model is as follows: Where L ad (E,D s ) represents the adversarial training loss, D s represents the discriminator, E represents the encoder, and F represents the test spectrum data. represents the reconstructed spectrum data; P(F) represents the distribution of the input spectrum data, Represents the mathematical expectation of the input spectral data distribution.

8. According to a multi-sensor spectrum anomaly detection method based on metric-adversarial learning as described in claim 1, the test sample obtained by preprocessing is input into the metric-adversarial model for model verification, the codec is input into each layer of the adversarial network for calculation to obtain the reconstruction error d1, and the input is input into each layer of deep metric learning to calculate the minimum Mahalanobis distance d3, and finally the spectrum training anomaly score is obtained, and the model judgment threshold is set according to the anomaly score.

9. The multi-sensor spectrum anomaly detection method based on metric-adversarial learning as claimed in claim 8, characterized in that: The first metric reconstruction error is defined as: Among them, d1 represents the reconstruction error, F sj Represents the test spectrum data, represents the reconstructed spectrum data; M represents a total of M spectrum sensors, N represents a total of N frequency points, s represents the sth sensor, and j represents the jth frequency point; The second metric, Mahalanobis distance, is defined as: Where d2 represents the Mahalanobis distance between the test sample and the class center, and d3 is the minimum Mahalanobis distance; where E represents the encoder, f represents the test sample, and P k Representative class center, Represents the inverse of the covariance matrix of the test sample.

10. The multi-sensor spectrum anomaly detection method based on metric-adversarial learning as claimed in claim 9, characterized in that: The two metrics are weighted to obtain the relationship between the final metric and the spectrum training abnormal threshold η for determination. The weighted processing parameter d is defined as: d=γd1+(1-γ)d3, Among them, d represents the final metric of spectrum anomaly detection, d1 represents the reconstruction error, d3 represents the minimum Mahalanobis distance, and γ is the control coefficient of the reconstruction error and the Mahalanobis distance.

Images