Training method and device applied to time domain convolutional network model and storage medium

Abstract

The embodiment of the invention provides a training method and device applied to a time domain convolutional network model and a storage medium, and the method comprises the steps that the time domain convolutional network model comprises an encoder; acquiring a normal electromagnetic spectrum signal; based on the encoder, according to the normal electromagnetic spectrum signal and a preset weight, obtaining a high-dimensional feature; and optimizing the encoder according to the high-dimensional features based on an anomaly detection identification algorithm. The problems that in the prior art, manpower and material resources are wasted due to the fact that supervision needs to depend on manual annotation of the electromagnetic spectrum, and the training result is not accurate in the non-supervision process are solved.

Description

Technical Field

[0001] This document relates to the field of electromagnetic spectrum monitoring technology, and in particular to a training method, device and storage medium for a time domain convolutional network model. Background Art

[0002] The electromagnetic spectrum is the basis of wireless communications and is used in military and civilian fields. Accurate analysis of abnormal data in the electromagnetic spectrum can promptly detect illegal radio activities and ensure communication security.

[0003] Existing technologies mainly rely on the professional knowledge and experience of staff to determine whether the electromagnetic spectrum is abnormal, or detect abnormal data based on supervised learning algorithms or unsupervised learning in deep learning.

[0004] However, if manual detection is used, this detection method requires the staff to have a high level of professionalism and rich experience, and the detection results are easily affected by subjective factors. Manual monitoring is time-consuming and labor-intensive, affecting work efficiency; if supervised learning is used for detection, this method requires a lot of manpower and material resources to mark abnormal data in order to train a model that can automatically analyze abnormal data, which is a waste of time; if unsupervised learning is used for anomaly detection, this method does not constrain the distribution of the encoded implicit space, resulting in the inability to extract more effective features, making the abnormal data detection inaccurate, affecting work efficiency. Summary of the invention

[0005] In view of the above scheme, the present application aims to propose a training method, device and storage medium for a time domain convolutional network model to solve at least one of the above technical problems.

[0006] In a first aspect, one or more embodiments of this specification provide a training method for a time domain convolutional network model, wherein the time domain convolutional network model includes: an encoder, and the method includes:

[0007] Acquire normal electromagnetic spectrum signals;

[0008] Based on the encoder, obtaining a high-dimensional feature according to the normal electromagnetic spectrum signal and a preset first weight; and

[0009] Based on anomaly detection and discrimination algorithms, the encoder is optimized according to high-dimensional features.

[0010] Furthermore, the time domain convolutional network model also includes a classifier, and the method also includes:

[0011] The classifier is optimized according to the optimized encoder.

[0012] Further, a test data set is obtained, wherein the test data set includes normal electromagnetic spectrum signals, abnormal electromagnetic spectrum signals, and a first classification result;

[0013] Based on the optimized encoder and the optimized classifier in the time-domain convolutional network model, and according to the test data set, determining a second classification result; and

[0014] Based on a preset loss function, the model accuracy is determined according to the first classification result and the second classification result.

[0015] Furthermore, the loss function is calculated using the following formula:

[0016] L=L oc +L C

[0017] Where L represents the loss function;

[0018]

[0019] L oc represents the encoder loss function;

[0020] Es ∈S∪S , represents the expected value of the distribution of electromagnetic signals;

[0021] Represents the mean value of normal electromagnetic signal distribution;

[0022] represents the variance of the normal electromagnetic signal distribution;

[0023]

[0024] L C Represents the classifier loss function

[0025] E s∈S∪S , represents the expected value of the distribution of electromagnetic signals;

[0026] s represents the high-dimensional characteristics of normal electromagnetic signals;

[0027] s′ represents the high-dimensional characteristics of abnormal electromagnetic signals;

[0028] S represents the high-dimensional characteristic distribution of normal electromagnetic signals;

[0029] S′ represents the high-dimensional characteristic distribution of abnormal electromagnetic signals;

[0030] w s represents weight;

[0031] l s∈S represents the distance measurement function;

[0032] Ψ c represents a classifier;

[0033] φ s represents the encoder;

[0034] y + =1 means the classifier result is a normal sample;

[0035] y - =-1 means the classifier result is an abnormal sample.

[0036] Furthermore, the abnormal monitoring and identification algorithm is a One Class SVM algorithm, and the abnormal detection and identification algorithm based on the high-dimensional features, optimizing the encoder includes:

[0037] Based on the One Class SVM algorithm, a hypersphere is determined according to the high-dimensional features;

[0038] According to the hypersphere, determining the center and radius of the hypersphere;

[0039] Calculating the Euclidean distance according to the high-dimensional feature and the center of the hypersphere;

[0040] According to the high-dimensional feature, the center of the hypersphere, the radius and the Euclidean distance, the preset first weight is updated to obtain a second weight;

[0041] Based on the deep time-domain convolutional network, updating high-dimensional features according to the electromagnetic spectrum signal and the second weight;

[0042] If the second weight and the high-dimensional feature remain unchanged, determining the time-domain convolutional network model; and

[0043] If the second weight and the high-dimensional feature change, the preset weight and the high-dimensional feature are updated.

[0044] Furthermore, the Euclidean distance is calculated using the following formula:

[0045]

[0046] Where d(s) represents the Euclidean distance between the normal sample and the center of the hypersphere;

[0047] represents the Euclidean distance between the abnormal sample and the center of the hypersphere;

[0048] f(s) represents outliers;

[0049] ψ * represents a classifier;

[0050] φ *Indicates the encoder;

[0051] s represents the high-dimensional characteristics of normal electromagnetic signals;

[0052] s′ represents the high-dimensional characteristics of abnormal electromagnetic signals;

[0053] c represents the center of the hypersphere;

[0054] ψ′ * Represents a classifier.

[0055] In a second aspect, an embodiment of the present application provides an electromagnetic spectrum anomaly judgment method based on an optimized time domain convolutional network model, wherein the optimized time domain convolutional network model is trained according to the training method applied to the time domain convolutional network model according to any one of the first aspects, including:

[0056] receiving electromagnetic spectrum signals; and

[0057] Taking the electromagnetic spectrum signal as input, based on the optimized time domain convolutional network model, it is determined whether the electromagnetic spectrum signal is a normal electromagnetic spectrum signal or an abnormal electromagnetic spectrum signal.

[0058] In a third aspect, an embodiment of the present application provides a training device for a time-domain convolutional network model, wherein the time-domain convolutional network model includes: an encoder;

[0059] The device comprises:

[0060] A first acquisition module, used for acquiring normal electromagnetic spectrum signals;

[0061] A feature extraction module, configured to obtain a high-dimensional feature based on the encoder, the normal electromagnetic spectrum signal and a preset first weight; and

[0062] The first optimization module is used to optimize the encoder based on an anomaly detection and identification algorithm according to high-dimensional features.

[0063] In a fourth aspect, an embodiment of the present application provides an electromagnetic spectrum anomaly judgment device based on an optimized time domain convolutional network model, the device further comprising:

[0064] A receiving module, configured to receive electromagnetic spectrum signals; and

[0065] The discrimination module is used to take the electromagnetic spectrum signal as input and, based on the optimized time domain convolutional network model, determine whether the electromagnetic spectrum signal is a normal electromagnetic spectrum signal or an abnormal electromagnetic spectrum signal.

[0066] In a fifth aspect, an embodiment of the present application provides a storage medium for storing computer-executable instructions, characterized in that when the computer-executable instructions are executed, the steps of the training method applied to the time-domain convolutional network model described in any one of the first aspects are implemented.

[0067] Compared with the prior art, this application can at least achieve the following technical effects:

[0068] The present application first obtains normal electromagnetic spectrum signals to avoid the use of artificial experience to label electromagnetic spectrum signals in the prior art, reduces manpower and material resources, and improves the learning efficiency of the model; then uses the anomaly detection and identification algorithm to optimize the encoder in the time domain convolutional network model, thereby improving the accuracy of the time domain convolutional network model in detecting abnormal data and ensuring work efficiency. BRIEF DESCRIPTION OF THE DRAWINGS

[0069] In order to more clearly illustrate one or more embodiments of this specification or the technical solutions in the prior art, the drawings required for use in the embodiments or the description of the prior art will be briefly introduced below. Obviously, the drawings described below are only some embodiments recorded in this specification. For ordinary technicians in this field, other drawings can be obtained based on these drawings without paying creative labor.

[0070] Figure 1 A flow chart of a training method applied to a time-domain convolutional network model provided for one or more embodiments of this specification;

[0071] Figure 2 A schematic diagram of the receptive field of a dilated convolution applied to a training method of a time-domain convolutional network model provided in one or more embodiments of this specification;

[0072] Figure 3 A schematic diagram of a classifier for a training method for a time domain convolutional network model provided in one or more embodiments of this specification;

[0073] Figure 4 A schematic diagram of the structure of a training device applied to a time-domain convolutional network model provided in one or more embodiments of this specification. DETAILED DESCRIPTION

[0074] In order to enable those skilled in the art to better understand the technical solutions in one or more embodiments of this specification, the following will be combined with the drawings in one or more embodiments of this specification to clearly and completely describe the technical solutions in one or more embodiments of this specification. Obviously, the described embodiments are only part of the embodiments of this specification, not all of the embodiments. Based on one or more embodiments of this specification, all other embodiments obtained by ordinary technicians in this field without creative work should fall within the scope of protection of this document.

[0075] With the rapid development of radio technology, radio spectrum resources have become an important production factor for the transformation of new and old kinetic energy and the development of the digital economy, and have become one of the most valuable and widely used natural resources. In the military and civilian fields, electromagnetic spectrum control is becoming increasingly important, among which electromagnetic spectrum anomaly detection is an important part of electromagnetic spectrum control. In the early days of radio spectrum monitoring, the discovery of abnormal spectrum mainly relied on the professional knowledge and experience of staff to manually judge whether the authorized signal was abnormal. This requires staff to have a high level of professionalism and rich experience, but the monitoring results are easily affected by subjective factors, and manual monitoring is time-consuming and labor-intensive, which greatly affects the efficiency of radio spectrum monitoring. There are two main traditional electromagnetic environment anomaly detection methods: database comparison method and historical model analysis method. The former requires a relatively complete spectrum license database support, while the latter requires the accumulation of long-term monitoring data and the prior probability distribution knowledge of many frequency parameters. However, these methods perform poorly in electromagnetic environment anomaly detection. The current deep learning-based anomaly detection algorithms are mostly implemented using recurrent neural networks, which have powerful sequence processing capabilities. However, the spectrum data contains many frequency points and long time series. The recurrent neural network needs to convert the spectrum data into column vectors, which weakens the connection between data with the same frequency points and affects the data accuracy.

[0076] In order to solve the above technical problems, this application proposes a training method for a time domain convolutional network model, such as Figure 1 As shown, the specific steps are as follows:

[0077] Step S1, obtaining normal electromagnetic spectrum signals.

[0078] In the embodiment of the present application, the normal electromagnetic spectrum signal is determined and acquired by analyzing the frequency, wavelength, amplitude and application field of the electromagnetic spectrum.

[0079] For example: suppose that the data set is a superset XS, which contains multiple electromagnetic spectrum signals from different frequency bands, and the data set is divided into a training data set and a test data set; the training data set is a normal electromagnetic spectrum signal, which is defined as a normal electromagnetic spectrum signal after analyzing the frequency, wavelength, amplitude and application field of the electromagnetic spectrum; the test data set is data that can distinguish between normal electromagnetic spectrum signals and abnormal electromagnetic spectrum signals, which is defined as normal data after analyzing the frequency, wavelength, amplitude and application field of the electromagnetic spectrum, and data not derived from the distribution p(XS) is abnormal data.

[0080] It should be noted that the frequency, wavelength, amplitude and application field of the electromagnetic spectrum are usually different in different scenarios. Therefore, in the embodiments of the present application, different scenarios correspond to different time domain convolutional network models. Different scenarios actually refer to different data sets. If a batch of data is collected in scenario A, then this network can only be effective in scenario A. If you switch to scenario B, you need to retrain the network using the data in scenario B.

[0081] Step S2, based on the encoder, obtain high-dimensional features according to the normal electromagnetic spectrum signal and a preset first weight.

[0082] In an embodiment of the present application, the time domain convolutional network model includes an encoder, which takes each normal electromagnetic spectrum signal and a preset first weight of each normal electromagnetic spectrum as input, and outputs high-dimensional features of each electromagnetic spectrum signal based on the encoder.

[0083] Specifically, the encoder mainly consists of two parts: a fully connected layer and a ReLU activation function, which are mainly used to obtain the time domain characteristics of electromagnetic signals. First, a given input electromagnetic spectrum signal is given; then the signal is framed with a frame length of L and a frame shift of S=L / 2, and then spliced ​​to obtain electromagnetic signal fragments. Next, the weight matrix of the fully connected layer of the encoder is matrix multiplied, and the ReLU function is used for activation processing. Finally, each fragment of the electromagnetic spectrum signal is projected into the N-dimensional latent space, and the high-dimensional features of each electromagnetic spectrum are output. The feedforward training process of the encoder can be expressed as follows:

[0084] X l+1 [t] = X l [t]W+b

[0085] H = ReLU(X l+1 )

[0086] Among them, l represents the number of fully connected layers;

[0087] t=1,...,T represents a time unit;

[0088] b represents the bias vector of the fully connected layer unit;

[0089] H is the representation vector learned after the electromagnetic signal passes through the activation layer;

[0090] X represents a segment of electromagnetic spectrum signal;

[0091] W represents the fully connected layer weight matrix.

[0092] In this application, the electromagnetic spectrum is mapped into an N-dimensional latent space, which can output features of different dimensions for different data, thereby improving the accuracy of the features.

[0093] Step S3, based on anomaly detection and identification algorithm, optimize the encoder according to high-dimensional features.

[0094] In the embodiment of the present application, based on the abnormality detection and identification algorithm, each high-dimensional feature is mapped to a hypersphere, and the weight of each high-dimensional feature is updated; the updated weight is input into the encoder to optimize each parameter in the encoder. Among them, the abnormality monitoring and identification algorithm is a One Class SVM algorithm. Among them, the weight is the weight of the high-dimensional feature corresponding to each electromagnetic spectrum signal.

[0095] Specifically, during training, the present application uses normal electromagnetic spectrum signals, and obtains high-dimensional features after training according to the time domain convolutional network model, and all of these features are normal high-dimensional features; then, the normal high-dimensional features are mapped to the plane of the hypersphere to obtain weights; the weights are then input into the encoder to update the high-dimensional features; if other parameters such as the high-dimensional features change, the high-dimensional features are mapped to the hypersphere to calculate the weights; if other parameters such as the high-dimensional features remain unchanged, the optimized time domain convolutional network model is determined.

[0096] The present application can utilize the encoder in the time domain convolutional network model to mine the time and space dimensions of the electromagnetic spectrum, and optimize the encoder using anomaly detection and identification algorithms, so that the trained model has higher anomaly detection accuracy and effectively improves work efficiency.

[0097] Further, in the embodiment of the present application, the time domain convolutional network model also includes a classifier, and the classifier is optimized according to the optimized encoder, such as Figure 3 shown.

[0098] Among them, the encoder is used to obtain the time domain characteristics of the electromagnetic spectrum signal, and the classifier is used to distinguish between normal electromagnetic spectrum signals and abnormal electromagnetic spectrum signals; the high-dimensional features in the normal electromagnetic spectrum signal are projected affinely into a unified unit circle hyperplane, and its abnormal sample points are projected into the space outside this hyperplane, which can further improve the accuracy of anomaly detection.

[0099] Specifically, according to the optimized encoder, high-dimensional features are determined; the high-dimensional features and the weights in the One Class SVM algorithm are input into the classifier to optimize the classifier; if the encoder is optimized, the classifier is optimized according to the high-dimensional features in the encoder; if the encoder stops optimizing, the classifier optimization is completed. The classifier includes an uncertain modeling calibration module (Uncertainty Modeling-based Calibration, UMC) and a native anomaly calibration module (Native Anomaly-based Calibration, NAC). Among them, the uncertain modeling calibration module is an anomaly detection algorithm based on a single-class classifier. First, a new single-class loss function is constructed, that is, for a given a priori probability of a single class, it is necessary to maximize the probability of a distance of 0, rather than simply minimize a single distance value. Based on this, the present application designs a new probability density function based on Gaussian distribution:

[0100]

[0101] Where J represents the probability density function of Gaussian distribution;

[0102] e represents the exponential constant;

[0103] u represents the sample mean;

[0104] Then by deduction we can get:

[0105]

[0106] Let ξ=logσ 2 , maximizing the above formula is equivalent to minimizing the following formula:

[0107]

[0108] Based on this, the loss function of the uncertainty modeling calibration module is calculated as follows:

[0109]

[0110] Among them, L oc represents the encoder loss function;

[0111] E s∈S∪S , represents the expected value of the distribution of electromagnetic signals;

[0112] s represents the high-dimensional characteristics of normal electromagnetic signals;

[0113] Represents the mean value of normal electromagnetic signal distribution;

[0114] Represents the variance of the normal electromagnetic signal distribution.

[0115] The existing technology mainly obtains a smaller penalty loss value on the noise, which will cause the output of the single-class classification model to be inconsistent and the prediction of these hidden anomalies to be inaccurate. The loss function in the uncertain modeling calibration module of the present application can successfully cover up the abnormal sample points and ensure the accuracy of the prediction of hidden anomalies.

[0116] An anomaly classifier is designed in the native anomaly calibration module, and then perturbation data is added to the normal sample data. The anomaly classifier is used to identify which data is abnormal and which is normal after the perturbation is added. Finally, the output of this module is sent to the uncertain modeling calibration module to assist in the training of this module. Based on the mean square error loss function, the loss function is calculated according to the perturbation data. The loss function calculation formula is as follows:

[0117] L C =E s∈S∪S′ [w s ·l s∈S (Ψ c (φ s (s))-y - ) 2 +w s ·l s∈S′ (Ψ c (φ s (s′))-y + ) 2 ]

[0118] Among them, L C represents the classifier loss function;

[0119] E s∈S∪S , represents the expected value of the distribution of electromagnetic signals;

[0120] s represents the high-dimensional characteristics of normal electromagnetic signals;

[0121] s′ represents the high-dimensional characteristics of abnormal electromagnetic signals;

[0122] S represents the high-dimensional characteristic distribution of normal electromagnetic signals;

[0123] S′ represents the high-dimensional characteristic distribution of abnormal electromagnetic signals;

[0124] w s represents weight;

[0125] l s∈S represents the distance measurement function;

[0126] Ψ c represents a classifier;

[0127] φ s represents the encoder;

[0128] y + =1 means the classifier result is a normal sample;

[0129] y - =-1 means the classifier result is an abnormal sample.

[0130] Finally, the total loss function is calculated based on the loss function in the uncertainty modeling calibration module and the loss function in the native anomaly calibration module; the electromagnetic spectrum signal is classified based on the total loss function. The calculation formula of the total loss function is:

[0131] L=L oc +L C

[0132] Among them, L oc represents the encoder loss function;

[0133] L C Represents the classifier loss function.

[0134] The mean square error loss function can accurately distinguish whether the electromagnetic spectrum signal is normal or abnormal, thereby improving the resolution accuracy.

[0135] Furthermore, in an embodiment of the present application, a test data set is obtained, wherein the test data set includes normal electromagnetic spectrum signals, abnormal electromagnetic spectrum signals, and a first classification result; based on the optimized encoder and the optimized classifier in the time domain convolutional network model, a second classification result is determined according to the test data set; and based on a preset loss function, the model accuracy is determined according to the first classification result and the second classification result.

[0136] Specifically, during the test, both normal and abnormal electromagnetic spectrum signals are used. According to the model training, the obtained feature representations are evenly distributed on the plane of the hypersphere, and the abnormal samples are distributed outside the hypersphere. Therefore, the distance between the test sample and the center point of the hypersphere after mapping can be used to determine whether it is an abnormal sample point, that is, the points inside the hypersphere are normal samples, and the points outside the hypersphere are abnormal samples.

[0137] For example: obtain a test data set, which contains normal electromagnetic spectrum signals and abnormal electromagnetic spectrum signals, and has classification result labels for distinguishing normal electromagnetic spectrum signals from abnormal electromagnetic spectrum signals; input these data into the optimized time domain convolutional network model, and the classifier in the time domain convolutional network model outputs the classification result; compare the classification result with the initial labeling result to determine the model accuracy; if the model accuracy is within the preset range, determine the model; if the model accuracy is not within the preset range, continue to optimize the model.

[0138] Furthermore, in an embodiment of the present application, the abnormality monitoring and identification algorithm is a One Class SVM algorithm, which is based on the OneClass SVM algorithm and, according to the high-dimensional features, determines a hypersphere; according to the hypersphere, determines the center and radius of the hypersphere; according to the high-dimensional features and the center of the hypersphere, calculates the Euclidean distance; according to the high-dimensional features, the center of the hypersphere, the radius and the Euclidean distance, updates the preset first weight to obtain the second weight; based on the deep time domain convolutional network, updates the high-dimensional features according to the electromagnetic spectrum signal and the second weight; if the second weight and the high-dimensional features are unchanged, determines the time domain convolutional network model; and if the second weight and the high-dimensional features change, updates the preset weight and the high-dimensional features.

[0139] Specifically, high-dimensional features can be understood as spatial points in n-dimensional space, and the hypersphere is the n-dimensional region obtained based on the spatial points. The weight of the high-dimensional feature affects the distribution of the spatial points in the n-dimensional space. Different distributions correspond to different hyperspheres. The Euclidean distance is the distance between each spatial point and the center of the hypersphere. Since all training samples are positive samples, theoretically all spatial points are within the hypersphere, that is, the Euclidean distance is less than the radius of the hypersphere. However, due to errors in the model, the actual situation is that some spatial points are not within the hypersphere. Therefore, by adjusting the weight (changing the first weight to the second weight), the spatial distribution of the spatial points is changed to obtain a new hypersphere. After the above process is repeated many times, all spatial points will be in the hypersphere. For the model, the denser the distribution of spatial points, the higher its accuracy. Therefore, after all spatial points are included in the hypersphere, continue to iterate, that is, by adjusting the weight of the high-dimensional features, the hypersphere is continuously reduced until the second weight and the high-dimensional features remain unchanged.

[0140] Among them, One Class SVM belongs to the category of support vector machines, but it is different from the traditional classification and regression support vector machine based on supervised learning. It is an unsupervised learning method that does not require the output labels of the training set to be marked, reducing manpower and material resources. Assume that the parameters of the generated hypersphere are the center o and the corresponding hypersphere radius r>0, the hypersphere volume V(r) is minimized, and the center o is a linear combination of the support vectors; similar to the traditional SVM method, it can be required that all training data points x i The distance to the center is strictly less than r. At the same time, a slack variable ζ with a penalty coefficient C is constructed i , the optimization problem formula is as follows:

[0141]

[0142] ||x i -o||2≤r+ξ i , i=1,2,3...m

[0143] ξ i ≥0, i=1, 2, ...m

[0144] Among them, i Represents the slack variable

[0145] V (r) represents the volume of a hypersphere;

[0146] C represents the penalty coefficient;

[0147] x i represents the i-th sample;

[0148] o represents the center of the hypersphere;

[0149] r represents the radius of the hypersphere;

[0150] m represents the sample size;

[0151] i represents the i-th sample index.

[0152] Then, the Lagrange dual solution is used to solve the above optimization problem formula. After that, it can be determined whether the new data point z is inside. If the distance from z to the center is less than or equal to the radius r, it is not an outlier. If it is outside the hypersphere, it is an outlier. Based on this, assuming that a trained network is given, the optimal projection network ψ′ is obtained respectively. * and classification network , the calculation method of the anomaly score for each test sample s can be written as follows:

[0153]

[0154] Where d(s) represents the Euclidean distance between the normal sample and the center of the hypersphere;

[0155] represents the Euclidean distance between the abnormal sample and the center of the hypersphere;

[0156] f(s) represents outliers;

[0157] ψ * represents a classifier;

[0158] φ * Indicates the encoder;

[0159] s represents the high-dimensional characteristics of normal electromagnetic signals;

[0160] s′ represents the high-dimensional characteristics of abnormal electromagnetic signals;

[0161] c represents the center of the hypersphere;

[0162] ψ′ * Represents a classifier.

[0163] Based on the Euclidean distance between the normal and abnormal samples and the center of the hypersphere obtained by the One Class SVM, the closer the sample is to the center, the higher the probability that it is a normal sample. Therefore, the samples can be weighted according to the different weights of the samples belonging to normal samples. The closer the sample is to the center, the higher the sample weight. Then the time domain convolutional network model is retrained based on the weight. The weight is calculated as follows:

[0164]

[0165] Among them, w i represents the weight of the i-th sample;

[0166] d i represents the distance of the i-th sample from the center of the hypersphere;

[0167] d max represents the distance from the center of the hypersphere to the farthest sample;

[0168] d min Represents the distance from the center of the hypersphere to the nearest sample.

[0169] In the embodiment of the present application, the feature extraction network in the encoder and the OneClass SVM in the anomaly detection and identification algorithm are alternately optimized using the EM framework. On the one hand, the M step of the EM framework can allow the network based on the deep time domain to extract more detailed features; on the other hand, in the E step of the EM algorithm, extracting better features can enable the OneClassSVM to learn the hypersphere more accurately and the accuracy of anomaly classification is higher.

[0170] Specifically, based on the EM framework, the M step is to input the normal electromagnetic spectrum signal and the initial weight into the deep time domain convolutional network to obtain high-dimensional features. The E step is to input the high-dimensional features into the One Class SVM algorithm to update the initial weights of each electromagnetic spectrum signal. The deep time domain convolutional network model is then updated according to the updated weights. The parameters in the time domain convolutional network model and the One Class SVM algorithm are optimized alternately.

[0171] Furthermore, in an embodiment of the present application, the temporal convolutional network model also includes a decoder and a temporal convolutional network (TCN). When the encoder in the temporal convolutional network model is optimized by the One Class SVM algorithm, certain parameters in the temporal convolutional network will be changed, and there is a correlation between the parameters, and then the decoder and the temporal convolutional module in the temporal convolutional network model are also optimized accordingly; when the encoder in the temporal convolutional network model is optimized, other modules are also optimized, so the optimized temporal convolutional network model is generated.

[0172] Among them, the time domain convolution module TCN mainly includes: gated convolution unit (Gated Conv1D), dilated convolution (Dilated ConvlD), skip connection (Conv Skip) and residual connection (Res). The specific forward propagation process of the model is: first, the encoder encodes the electromagnetic signal into a potential representation with a frame structure, and the potential representation must be reconstructed into a time domain electromagnetic signal by the decoder; then, the input electromagnetic signal is subjected to the modulation domain feature extractor to generate a three-dimensional feature spectrum (modulation domain feature) about the time-frequency-modulation band. The feature representation of the feature spectrum after flattening is combined with the potential representation along the channel dimension to form a fusion feature, which is used as the input of the TCN module; finally, the potential representation after TCN masking is output by the decoder to estimate the time domain electromagnetic signal.

[0173] Specifically, the modulation domain feature extractor is essentially an algorithm for extracting signal features based on the time-frequency domain. The extractor can usually be composed of two time-frequency transforms. First, the electromagnetic spectrum signal is subjected to a short-time Fourier transform (STFT) to obtain the frequency domain representation of the signal, specifically:

[0174]

[0175] Wherein, STFT{} represents the STFT function;

[0176] f=1,...,F represents the frequency band index;

[0177] X represents the frequency domain characteristics of the input electromagnetic signal;

[0178] t represents the timestamp;

[0179] n represents the index of the electromagnetic signal in the time domain;

[0180] N represents the signal length of the electromagnetic signal in the time domain;

[0181] w represents the window function;

[0182] l represents the sampling frequency;

[0183] e represents an exponential constant.

[0184] Then perform STFT operation on each frequency band of the time-frequency vector, and finally obtain the spectrum in the modulation domain, which is expressed as follows:

[0185] X m =Mod{x}=STFT{|X|}

[0186] Among them, X m represents the modulation domain spectrum;

[0187] Mod{x} represents the result of performing the STFT operation on each frequency band of the time-frequency vector.

[0188] In this application, the modulation domain transform, compared with the traditional STFT, additionally examines the sequence relationship between the speech frame structures, and uses FFT technology to project the sequence law into the harmonic space, thereby obtaining a richer electromagnetic feature representation than the time-frequency domain. However, the traditional modulation domain algorithm ignores the perceptual change law of the frame structure in each frequency band during the two STFT transformations, and directly using a fixed transformation matrix for modeling will result in better sparsity of the generated three-dimensional spectrogram, but it is difficult to be understood by the neural network. In response to this limitation, a trainable matrix is ​​used to replace the weight matrix required for the second STFT transformation in the modulation domain transform, so that the extracted electromagnetic spectrum features are more in line with the essential laws of electromagnetic signals, and the dissipation of phase information in the second STFT transformation can be avoided. On the other hand, for the lack of phase information in the first STFT transformation, time domain features can be considered for compensation.

[0189] Based on the above modulation domain correlation analysis, the forward propagation process of the modulation domain feature extractor designed in this application is as follows: First, the time-frequency energy spectrum after STFT processing is input; then, a convolution kernel K with a number of m and a convolution kernel size of L performs the related convolution operation independently on each frequency band. This process can define a one-dimensional modulation domain convolution (ModConv-1D), namely:

[0190]

[0191] in, Used to represent the modulation domain features obtained by the modulation domain feature extractor.

[0192] There is a certain difference between the spectrum formed by the modulation domain feature extractor and the spectrum generated by the standard dual STFT modulation domain, that is, the modulation band has no gradual change pattern. This phenomenon is mainly caused by the computational equivalence of each convolution kernel during the encoding process. This equivalence relaxes the band planning of the standard modulation domain algorithm, allowing the model to automatically track the perceptual sequence law at the frame level under the supervision of the loss function. From the perspective of perceptual characteristics, the proposed TCN model models all frequency bands using weight-shared convolution kernels, imposes similarity constraints on each frequency band, and helps the post-TCN model better learn the harmonic law of the electromagnetic spectrum.

[0193] In this application, in order to avoid the gradient vanishing problem caused by a large number of stacked convolution modules and improve the reconstruction accuracy of electromagnetic waveforms, a novel holey convolution module with skip connections and residual connections is designed, and a time domain convolutional neural network composed of multiple cascaded convolution modules is constructed. The TCN module cooperates with the encoder-decoder pair and the modulation domain feature extractor pair to form an end-to-end time domain network model for electromagnetic spectrum signal processing. The encoder and the modulation domain feature extractor generate the feature H and After shaping, they are concatenated along the channel dimension, and the concatenation of the two is compressed into the standard input tensor H by the weight matrix W convolution. The process is expressed as:

[0194] H c =concat(H,H mod () c )

[0195] Among them, concat() represents the tensor concatenation operation;

[0196] H c represents the standard input tensor;

[0197] H mod represents the modulation domain features obtained by the modulation domain feature extractor;

[0198] c represents the channel dimension.

[0199] The encoded latent features are used as the standard input of the TCN module for subsequent processing. Each convolution module contains four parts: gated convolution layer, holed convolution layer, residual branch convolution layer and skip connection convolution layer. Given the input features of this module:

[0200]

[0201] Among them, ⊙ represents the matrix inner product operation;

[0202] sigmoid represents an S-type saturation function;

[0203] Represents the output result after the input feature passes through the gated convolution unit.

[0204] In this application, the atrous convolution can provide a larger receptive field (e.g. Figure 2 As shown in Figure 1, the hole rate of the middle layer indicates that the sample sampling of each d point is used as the input of the next layer. The size of the sampling effective window increases exponentially with the increase of the number of convolution layers, which can effectively reduce the number of convolution layers required to sample each time series point, and avoid the appearance of the chessboard effect while expanding the receptive field of the model, ultimately enhancing the TCN model's modeling ability for time series.

[0205] After all the above modules complete the forward propagation, all the jump connection outputs in the TCN module need to be summed and processed as the output of the TCN module, and the gated convolution unit outputs in a compressed dimension. The process can be expressed as:

[0206]

[0207] Among them, H out Represents the output of the TCN module;

[0208] B represents the total number of channel dimensions;

[0209] b represents the current channel dimension;

[0210] Indicates jump connection output.

[0211] The final output feature is essentially the integration of all scale features, and each convolutional layer in the model has a certain contribution to the output feature. At the same time, the residual connection mechanism added to the model is conducive to significantly accelerating the convergence of the model, avoiding the gradient vanishing problem caused by a large number of stacked convolutional modules, and effectively improving the overall modeling ability and accuracy of the model. In addition, the TCN structure relies on hierarchical and multi-scale information interaction, and captures the sequence relationship along the time dimension of the input features for modeling. The interactive characteristics between the convolutional layers enable the model to effectively integrate and process the electromagnetic spectrum features from both the modulation domain and the time domain, thereby fitting highly nonlinear waveform signals.

[0212] In the embodiment of the present application, the decoder structure essentially covers two steps of overlap-add and linear transformation. out After linear transformation, the overlap-add algorithm is used to obtain the estimated time-domain electromagnetic spectrum waveform. The specific process can be expressed as:

[0213] x=OLA(H out , H dec )

[0214] x represents the estimated time-domain electromagnetic spectrum waveform;

[0215] H out Indicates the TCN output result;

[0216] H dec Represents the result after linear transformation of TCN output.

[0217] In order to verify the effectiveness of the algorithm of this application, this method is compared with the currently commonly used time series anomaly detection algorithms on different data sets. The data sets include ASD, SMD, SWaT, WaQ, DSADS and Epilepsy. Among them, the ASD data set comes from a large Internet company, which uses a set of indicators to characterize the status of different servers (entities); SMD is a public data set for multivariate time series anomaly detection; SWaT is a network security data set, including 11 days of continuous operation data. 7 days of normal operation data and 4 days of attack scenario data were collected; the WaQ data set contains pollution level data of various water sources, mainly used to monitor water quality; the DSADS data set contains motion sensor data of 19 daily sports, each of which is performed by 8 subjects in their own way for 5 minutes; the Epilepsy data set is generated by healthy participants simulating classroom activities. The data was collected by 6 participants using a three-axis accelerometer on the main wrist while performing 4 different activities.

[0218] The comparison algorithms include OCSVM, GOAD, ECOD, DAGMM, DSVDD, MSCRED and TranAD. Among them, OCSVM is the abbreviation of One Class SVM mentioned in this section; GOAD unifies the current method of training only with normal samples, and is specifically used for anomaly detection of image and table data sets; ECOD first estimates the underlying distribution of the input data in a non-parametric way by calculating the empirical cumulative distribution of each dimension of the data, and then uses these empirical distributions to estimate the probability of whether each data point is an anomaly; DAGMM combines the autoencoder with the Gaussian mixture model and is an unsupervised anomaly detection algorithm; DSVDD is a One Class SVM anomaly detection algorithm based on deep learning; MSCRED is a new method for multi-indicator anomaly detection using graph neural networks; TranAD is an anomaly detection model based on a deep network (Transformer), which uses an attention-based sequence encoder to quickly reason about data outliers by taking advantage of the broader temporal trends in the data.

[0219] The evaluation indicators are F1 and AUC-PR. The larger the values ​​of the above two values, the better the performance. Specifically, the calculation method of the F1 score is based on the harmonic mean of precision and recall. The score ranges between 0 and 1. The closer to 1, the better the balance between precision and recall. AUC-PR is an indicator for evaluating the performance of classification models. It is calculated based on precision and recall, and is more applicable when dealing with class imbalance or focusing on positive example identification. Usually, the value range of AUC-PR is 0 to 1. The closer the value is to 1, the better the model performs in the precision-recall space.

[0220] It can be seen from the table below that the method proposed in this application is better than the comparison algorithm, which shows that the algorithm proposed in this application can obtain good results in the anomaly detection task.

[0221] Table 1 Summary of anomaly detection comparison algorithms

[0222]

[0223] An embodiment of the present application provides an electromagnetic spectrum anomaly judgment method based on an optimized time domain convolutional network model, wherein the optimized time domain convolutional network model is trained according to the training method applied to the time domain convolutional network model described in any one of the above embodiments, and includes: receiving an electromagnetic spectrum signal; and taking the electromagnetic spectrum signal as input, and judging whether the electromagnetic spectrum signal is a normal electromagnetic spectrum signal or an abnormal electromagnetic spectrum signal based on the optimized time domain convolutional network model.

[0224] The present application embodiment provides a training device for a time domain convolutional network model, such as Figure 4 As shown, the time domain convolutional network model includes: an encoder;

[0225] The device comprises:

[0226] A first acquisition module 101 is used to acquire normal electromagnetic spectrum signals;

[0227] A feature extraction module 102, configured to obtain a high-dimensional feature based on the encoder, the normal electromagnetic spectrum signal and a preset first weight; and

[0228] The first optimization module 103 is used to optimize the encoder based on an anomaly detection and identification algorithm according to high-dimensional features.

[0229] The embodiment of the present application provides an electromagnetic spectrum anomaly judgment device based on an optimized time domain convolutional network model, and the device also includes:

[0230] A receiving module, configured to receive electromagnetic spectrum signals; and

[0231] The discrimination module is used to take the electromagnetic spectrum signal as input and, based on the optimized time domain convolutional network model, determine whether the electromagnetic spectrum signal is a normal electromagnetic spectrum signal or an abnormal electromagnetic spectrum signal.

[0232] An embodiment of the present application provides a storage medium for storing computer-executable instructions, characterized in that when the computer-executable instructions are executed, the steps of the training method applied to the time-domain convolutional network model described in any one of the embodiments are implemented.

[0233] It should be noted that the embodiments of the storage medium in this specification and the embodiments of the blockchain-based service provision method in this specification are based on the same inventive concept. Therefore, the specific implementation of this embodiment can refer to the aforementioned corresponding implementation of the blockchain-based service provision method, and the repeated parts will not be repeated.

[0234] The above is a description of a specific embodiment of the specification. Other embodiments are within the scope of the appended claims. In some cases, the actions or steps recorded in the claims can be performed in an order different from that in the embodiments and still achieve the desired results. In addition, the processes depicted in the drawings do not necessarily require the specific order or continuous order shown to achieve the desired results. In some embodiments, multitasking and parallel processing are also possible or may be advantageous.

[0235] In the 1930s, improvements to a technology could be clearly distinguished as hardware improvements (for example, improvements to the circuit structure of diodes, transistors, switches, etc.) or software improvements (improvements to the method flow). However, with the development of technology, many improvements to the method flow today can be regarded as direct improvements to the hardware circuit structure. Designers almost always obtain the corresponding hardware circuit structure by programming the improved method flow into the hardware circuit. Therefore, it cannot be said that an improvement in a method flow cannot be implemented using a hardware entity module. For example, a programmable logic device (PLD) (such as a field programmable gate array (FPGA)) is such an integrated circuit whose logical function is determined by the user's programming of the device. Designers can "integrate" a digital system on a PLD by programming it themselves, without having to ask a chip manufacturer to design and produce a dedicated integrated circuit chip. Moreover, nowadays, instead of manually making integrated circuit chips, this kind of programming is mostly implemented by "logic compiler" software, which is similar to the software compiler used when developing and writing programs, and the original code before compilation must also be written in a specific programming language, which is called hardware description language (HDL). There is not only one HDL, but many kinds, such as ABEL (Advanced Boolean Expression Language), AHDL (Altera Hardware Description Language), Confluence, CUPL (Cornell University Programming Language), HDCal, JHDL (Java Hardware Description Language), Lava, Lola, MyHDL, PALASM, RHDL (Ruby Hardware Description Language), etc. The most commonly used ones are VHDL (Very-High-Speed ​​Integrated Circuit Hardware Description Language) and Verilog. Those skilled in the art should also know that it is only necessary to program the method flow slightly in the above-mentioned hardware description languages ​​and program it into the integrated circuit, and then it is easy to obtain the hardware circuit that implements the logic method flow.

[0236] The controller can be implemented in any appropriate manner, for example, the controller can take the form of a microprocessor or processor and a computer-readable medium storing a computer-readable program code (such as software or firmware) that can be executed by the (micro)processor, a logic gate, a switch, an application-specific integrated circuit (ASIC), a programmable logic controller, and an embedded microcontroller. Examples of controllers include, but are not limited to, the following microcontrollers: ARC 625D, Atmel AT91SAM, Microchip PIC18F26K20, and Silicone Labs C8051F320. The memory controller can also be implemented as part of the control logic of the memory. Those skilled in the art also know that in addition to implementing the controller in a purely computer-readable program code manner, the controller can be implemented in the form of a logic gate, a switch, an application-specific integrated circuit, a programmable logic controller, and an embedded microcontroller by logically programming the method steps. Therefore, this controller can be considered as a hardware component, and the devices included therein for implementing various functions can also be regarded as structures within the hardware component. Or even, the devices for implementing various functions can be regarded as both software modules for implementing the method and structures within the hardware component.

[0237] The systems, devices, modules or units described in the above embodiments may be implemented by computer chips or entities, or by products with certain functions. A typical implementation device is a computer. Specifically, the computer may be, for example, a personal computer, a laptop computer, a cellular phone, a camera phone, a smart phone, a personal digital assistant, a media player, a navigation device, an email device, a game console, a tablet computer, a wearable device, or a combination of any of these devices.

[0238] For the convenience of description, the above devices are described in terms of functions and are divided into various units. Of course, when implementing the embodiments of this specification, the functions of each unit can be implemented in the same or multiple software and / or hardware.

[0239] It should be understood by those skilled in the art that one or more embodiments of this specification may be provided as a method, system or computer program product. Therefore, one or more embodiments of this specification may take the form of a complete hardware embodiment, a complete software embodiment, or an embodiment combining software and hardware. Moreover, this specification may take the form of a computer program product implemented on one or more computer-usable storage media (including but not limited to disk storage, CD-ROM, optical storage, etc.) containing computer-usable program code.

[0240] This specification is described with reference to the flowcharts and / or block diagrams of the methods, devices (systems), and computer program products according to the embodiments of this specification. It should be understood that each process and / or box in the flowchart and / or block diagram, as well as the combination of the processes and / or boxes in the flowchart and / or block diagram, can be implemented by computer program instructions. These computer program instructions can be provided to a processor of a general-purpose computer, a special-purpose computer, an embedded processor, or other programmable data processing device to produce a machine, so that the instructions executed by the processor of the computer or other programmable data processing device generate instructions for implementing the processes in the flowchart and / or block diagram. Figure 1 A process or multiple processes and / or boxes Figure 1 A device that provides the functions specified in a block or multiple blocks.

[0241] These computer program instructions may also be stored in a computer-readable memory capable of directing a computer or other programmable data processing device to operate in a specific manner, so that the instructions stored in the computer-readable memory produce an article of manufacture comprising an instruction device, which implements the process Figure 1 A process or multiple processes and / or boxes Figure 1 A function specified in one or more boxes.

[0242] These computer program instructions can also be loaded onto a computer or other programmable data processing device so that a series of operating steps are executed on the computer or other programmable device to produce a computer-implemented process, thereby providing instructions for implementing the process. Figure 1 A process or multiple processes and / or boxes Figure 1 The steps for the functions specified in one or more boxes.

[0243] In a typical configuration, a computing device includes one or more processors (CPU), input / output interfaces, network interfaces, and memory.

[0244] The memory may include non-permanent storage in a computer-readable medium, random access memory (RAM) and / or non-volatile memory in the form of read-only memory (ROM) or flash RAM. The memory is an example of a computer-readable medium.

[0245] Computer readable media include permanent and non-permanent, removable and non-removable media that can be implemented by any method or technology to store information. Information can be computer readable instructions, data structures, program modules or other data. Examples of computer storage media include, but are not limited to, phase change memory (PRAM), static random access memory (SRAM), dynamic random access memory (DRAM), other types of random access memory (RAM), read-only memory (ROM), electrically erasable programmable read-only memory (EEPROM), flash memory or other memory technology, compact disk read-only memory (CD-ROM), digital versatile disk (DVD) or other optical storage, magnetic cassettes, magnetic tape magnetic disk storage or other magnetic storage devices or any other non-transmission media that can be used to store information that can be accessed by a computing device. As defined herein, computer readable media does not include temporary computer readable media (transitory media), such as modulated data signals and carrier waves.

[0246] It should also be noted that the terms "include", "comprises" or any other variations thereof are intended to cover non-exclusive inclusion, so that a process, method, commodity or device including a series of elements includes not only those elements, but also other elements not explicitly listed, or also includes elements inherent to such process, method, commodity or device. In the absence of more restrictions, the elements defined by the sentence "comprises a ..." do not exclude the existence of other identical elements in the process, method, commodity or device including the elements.

[0247] One or more embodiments of the present specification may be described in the general context of computer-executable instructions executed by a computer, such as program modules. Generally, program modules include routines, programs, objects, components, data structures, etc. that perform specific tasks or implement specific abstract data types. One or more embodiments of the present specification may also be practiced in distributed computing environments where tasks are performed by remote processing devices connected through a communication network. In a distributed computing environment, program modules may be located in local and remote computer storage media, including storage devices.

[0248] Each embodiment in this specification is described in a progressive manner, and the same or similar parts between the embodiments can be referred to each other, and each embodiment focuses on the differences from other embodiments. In particular, for the system embodiment, since it is basically similar to the method embodiment, the description is relatively simple, and the relevant parts can be referred to the partial description of the method embodiment.

[0249] The above description is only an embodiment of this document and is not intended to limit this document. For those skilled in the art, this document may have various changes and variations. Any modification, equivalent replacement, improvement, etc. made within the spirit and principle of this document should be included in the scope of the claims of this document.

Claims

1. A training method for a time-domain convolutional network model, the time-domain convolutional network model comprising: The encoder is characterized in that the method comprises: Acquire normal electromagnetic spectrum signals; Based on the encoder, a high-dimensional feature is obtained according to the normal electromagnetic spectrum signal and a preset first weight; and Based on anomaly detection and discrimination algorithms, the encoder is optimized according to high-dimensional features.

2. The method according to claim 1, characterized in that The time domain convolutional network model also includes a classifier, and the method further includes: The classifier is optimized according to the optimized encoder.

3. The method according to claim 2, characterized in that The method further comprises: Acquire a test data set, wherein the test data set includes normal electromagnetic spectrum signals, abnormal electromagnetic spectrum signals, and a first classification result; Based on the optimized encoder and the optimized classifier in the time-domain convolutional network model, and according to the test data set, determining a second classification result; and Based on a preset loss function, the model accuracy is determined according to the first classification result and the second classification result.

4. The method according to claim 3, characterized in that The loss function is calculated using the following formula: L=L oc +L C Where L represents the loss function; L oc represents the encoder loss function; E s∈S∪S′ represents the expected value of the distribution of electromagnetic signals; Represents the mean value of normal electromagnetic signal distribution; represents the variance of the normal electromagnetic signal distribution; L C =E s∈S∪S′ [w s ·l s∈S (P c (f s (s))-y - ) 2 +w s ·l s∈S′ (P c (f s (s′))-y + ) 2 ] L C represents the classifier loss function; s represents the high-dimensional characteristics of normal electromagnetic signals; s′ represents the high-dimensional characteristics of abnormal electromagnetic signals; S represents the high-dimensional characteristic distribution of normal electromagnetic signals; S′ represents the high-dimensional characteristic distribution of abnormal electromagnetic signals; w s represents weight; l s∈S represents the distance measurement function; Ψ c represents a classifier; φ s represents the encoder; y + =1 means the classifier result is a normal sample; y_=-1 means that the classifier result is an abnormal sample.

5. According to the method of claim 1, the abnormal monitoring and identification algorithm is a One Class SVM algorithm, characterized in that: The method of optimizing the encoder based on the anomaly detection and identification algorithm according to the high-dimensional features includes: Based on the One Class SVM algorithm, a hypersphere is determined according to the high-dimensional features; According to the hypersphere, determining the center and radius of the hypersphere; Calculating the Euclidean distance according to the high-dimensional feature and the center of the hypersphere; According to the high-dimensional feature, the center of the hypersphere, the radius and the Euclidean distance, the preset first weight is updated to obtain a second weight; Based on the deep time-domain convolutional network, updating high-dimensional features according to the electromagnetic spectrum signal and the second weight; If the second weight and the high-dimensional feature remain unchanged, determining the time-domain convolutional network model; and If the second weight and the high-dimensional feature change, the preset weight and the high-dimensional feature are updated.

6. The method according to claim 5, characterized in that The Euclidean distance is calculated using the following formula: Where d(s) represents the Euclidean distance between the normal sample and the center of the hypersphere; represents the Euclidean distance between the abnormal sample and the center of the hypersphere; f(s) represents outliers; ψ* represents the classifier; φ* indicates the encoded code; s represents the high-dimensional characteristics of normal electromagnetic signals; s ′ High-dimensional features representing abnormal electromagnetic signals; c represents the center of the hypersphere; ψ ′* Represents a classifier.

7. A method for judging an anomaly in an electromagnetic spectrum based on an optimized time-domain convolutional network model, wherein the optimized time-domain convolutional network model is trained according to the training method for a time-domain convolutional network model according to any one of claims 1 to 6, characterized in that include: receiving electromagnetic spectrum signals; as well as Taking the electromagnetic spectrum signal as input, based on the optimized time domain convolutional network model, it is determined whether the electromagnetic spectrum signal is a normal electromagnetic spectrum signal or an abnormal electromagnetic spectrum signal.

8. A training device for a time domain convolutional network model, characterized in that The time domain convolutional network model includes: encoder; The device comprises: A first acquisition module, used for acquiring normal electromagnetic spectrum signals; A feature extraction module, configured to obtain a high-dimensional feature based on the encoder, the normal electromagnetic spectrum signal and a preset first weight; and The first optimization module is used to optimize the encoder based on an anomaly detection and identification algorithm according to high-dimensional features.

9. An electromagnetic spectrum anomaly judgment device based on an optimized time domain convolutional network model, characterized in that The device also includes: A receiving module, configured to receive electromagnetic spectrum signals; and The discrimination module is used to take the electromagnetic spectrum signal as input and, based on the optimized time domain convolutional network model, determine whether the electromagnetic spectrum signal is a normal electromagnetic spectrum signal or an abnormal electromagnetic spectrum signal.

10. A storage medium for storing computer executable instructions, characterized in that: When executed, the computer executable instructions implement the steps of the training method applied to the time domain convolutional network model described in any one of claims 1-6.

Images