Electromagnetic anomaly signal detection method and system

Abstract

The application discloses an electromagnetic abnormal signal detection method and system, and relates to the technical field of signal detection and identification. The method comprises the following steps: obtaining a target electromagnetic signal; performing feature extraction on the target electromagnetic signal to obtain a signal feature; determining a signal detection result of the target electromagnetic signal according to the signal feature and a signal detection model; the signal detection model comprises an identity mapping module and a convolution layer module connected in sequence; when the signal detection result is that the target electromagnetic signal is a suspicious abnormal signal, performing signal time-frequency domain statistical distribution feature calculation on the target electromagnetic signal, and performing signal identification on the calculation result according to a signal distribution feature map to determine whether the target electromagnetic signal is abnormal; the signal distribution feature map is determined according to signal time-frequency domain statistical distribution feature calculation on a measured electromagnetic signal. The application can improve the applicability and accuracy of abnormal signal detection in a complex environment.

Description

Technical Field

[0001] This invention relates to the field of signal detection and recognition technology, and in particular to a method and system for detecting electromagnetic anomalies. Background Technology

[0002] With the development of wireless communication technology, the electromagnetic environment has become increasingly harsh, posing a threat to electromagnetic security. Exploring effective methods for detecting abnormal signals is therefore crucial. The detection and identification of abnormal electromagnetic signals plays a vital role in judging and providing early warning of abnormal behavior of electromagnetic target equipment (systems) in combat environments and complex electromagnetic environments. It is also highly significant in areas such as space spectrum monitoring, malicious electromagnetic attacks, and electromagnetic interference identification. Different types of electromagnetic signal data have different characteristics in their time-frequency forms, especially when data types differ from normal signals appear. To avoid the leakage of critical information and electromagnetic signal intrusion, timely detection and identification of abnormal electromagnetic signals are necessary.

[0003] Currently, commonly used electromagnetic anomaly detection methods both domestically and internationally fall into two categories: maximum likelihood ratio (MMR) methods based on hypothesis testing and anomaly identification methods based on feature extraction. MMR methods, based on hypothesis testing, use the likelihood function of the received electromagnetic signal as a basis, comparing the likelihood ratio with an appropriate threshold to detect anomalies. However, their disadvantages lie in the fact that the threshold value depends on experience and lacks universality for different types of signals. Furthermore, the high computational cost limits the widespread adoption of this method. Feature extraction-based anomaly identification methods extract predefined features from the received signal and detect anomalies based on the differences between feature values. Many parameters are used for feature analysis, including not only common features such as amplitude, phase, and frequency, but also more complex parameters such as higher-order cumulants, time-frequency domain features, and IQ waveform features. However, using these features alone suffers from insufficient feature selection specificity, high computational complexity, and a cumbersome identification process, especially for signals with complex components where the identification effect is unsatisfactory.

[0004] With the increasing maturity of artificial intelligence technology, it has become possible to apply deep learning methods to the detection and recognition of electromagnetic anomaly signals. The electromagnetic anomaly signal detection method based on deep learning consists of three processes: (1) signal preprocessing; (2) signal feature extraction; and (3) anomaly detection and recognition. Signal preprocessing and feature extraction are completed by relying on the deep learning framework, which reduces the computational complexity of signal recognition to a certain extent. Deep learning is a representation learning method that achieves accurate characterization of the features of the original signal data by constructing a classification model with multiple hidden feature extraction layers, which significantly improves the performance of signal recognition. Deep learning uses the backpropagation algorithm to guide the classifier to adjust the parameters of the feature extraction layer in real time based on the results calculated by the previous layer network, thereby achieving parameter optimization of the entire deep learning framework. Compared with the large amount of design experience and theoretical foundation required by machine learning in signal detection and recognition, as well as the problem of weak applicability of the extracted features, the deep learning method obtains the internal relationship between data features from the original data through a multi-layer structure, accurately characterizes the small feature signals, and thus achieves accurate detection and recognition of electromagnetic anomaly signals in complex environments with high noise and indistinct signal features. However, electromagnetic signal detection methods based on deep learning frameworks are all black-box models in terms of feature extraction, which have poor interpretability and limited applicability, thus failing to achieve high accuracy in detecting abnormal signals in complex electromagnetic environments. Summary of the Invention

[0005] The purpose of this invention is to provide a method and system for detecting electromagnetic anomalies, which can improve the applicability and accuracy of anomaly signal detection in complex electromagnetic environments.

[0006] To achieve the above objectives, the present invention provides the following solution:

[0007] A method for detecting electromagnetic anomaly signals, comprising:

[0008] Acquire the target electromagnetic signal;

[0009] Feature extraction is performed on the target electromagnetic signal to obtain signal features;

[0010] The signal detection result of the target electromagnetic signal is determined based on the signal characteristics and the signal detection model; the signal detection model includes an identity mapping module and a convolutional layer module connected in sequence; the signal detection result is that the target electromagnetic signal is a normal signal or the target electromagnetic signal is a suspicious abnormal signal; the type of the normal signal is within the set carrier type range; the type of the suspicious abnormal signal is outside the set carrier type range.

[0011] When the signal detection result indicates that the target electromagnetic signal is a suspicious abnormal signal, the time-frequency domain statistical distribution characteristics of the target electromagnetic signal are calculated, and the calculation result is used to identify the signal based on the signal distribution characteristic map to determine whether the target electromagnetic signal is abnormal; the signal distribution characteristic map is determined based on the time-frequency domain statistical distribution characteristics of the measured electromagnetic signal.

[0012] Optionally, the step of extracting features from the target electromagnetic signal to obtain signal features specifically includes:

[0013] The target electromagnetic signal is clipped using a sliding window to obtain clipped data;

[0014] The cropped data is subjected to continuous wavelet transform to obtain a two-dimensional time-frequency graph.

[0015] Feature extraction is performed on the two-dimensional time-frequency graph to obtain signal features.

[0016] Optionally, the step of performing continuous wavelet transform to two-dimensionalize the cropped data to obtain a two-dimensional time-frequency graph specifically includes:

[0017] Perform a one-dimensional continuous wavelet transform operation on the cropped data to obtain a one-dimensional signal;

[0018] The one-dimensional signal is encoded in two dimensions to obtain a two-dimensional time-frequency diagram.

[0019] Optionally, the method for determining the signal detection model is as follows:

[0020] Acquire training data; the training data includes measured electromagnetic signals and corresponding tags; the measured electromagnetic signals include normal signals and suspected abnormal signals;

[0021] Construct a deep convolutional neural network model;

[0022] The training data is input into the deep convolutional neural network model, and the deep convolutional neural network model is trained using cross-validation. The trained deep convolutional neural network model is then determined as the signal detection model.

[0023] Optionally, the step of calculating the time-frequency domain statistical distribution characteristics of the target electromagnetic signal and identifying the signal based on the calculation results according to the signal distribution characteristic map to determine whether the target electromagnetic signal is abnormal specifically includes:

[0024] Calculate the time-frequency domain statistical distribution characteristics of the target electromagnetic signal to obtain suspicious abnormal signal feature points;

[0025] A probability calculation is performed on the feature points within the first range to obtain the anomaly probability; the first range feature points are the feature points of the suspected abnormal signal feature points located outside the signal distribution feature map.

[0026] When the probability of an anomaly is greater than a preset anomaly threshold, the target electromagnetic signal is determined to be abnormal.

[0027] The present invention also provides an electromagnetic anomaly signal detection system, comprising:

[0028] The signal acquisition module is used to acquire the target electromagnetic signal;

[0029] The feature extraction module is used to extract features from the target electromagnetic signal to obtain signal features;

[0030] A signal detection module is used to determine the signal detection result of the target electromagnetic signal based on the signal characteristics and a signal detection model; the signal detection model includes an identity mapping module and a convolutional layer module connected in sequence; the signal detection result is that the target electromagnetic signal is a normal signal or the target electromagnetic signal is a suspicious abnormal signal; the type of the normal signal is within a set carrier type range; the type of the suspicious abnormal signal is outside the set carrier type range.

[0031] The signal recognition module is used to calculate the time-frequency domain statistical distribution characteristics of the target electromagnetic signal when the signal detection result indicates that the target electromagnetic signal is a suspicious abnormal signal, and to perform signal recognition based on the calculation result according to the signal distribution feature map to determine whether the target electromagnetic signal is abnormal; the signal distribution feature map is determined based on the calculation of the time-frequency domain statistical distribution characteristics of the measured electromagnetic signal.

[0032] Optionally, the feature extraction module specifically includes:

[0033] The clipping unit is used to perform sliding window clipping on the target electromagnetic signal to obtain clipped data;

[0034] A two-dimensional processing unit is used to perform continuous wavelet transform two-dimensionalization operation on the cropped data to obtain a two-dimensional time-frequency diagram.

[0035] The feature extraction unit is used to extract features from the two-dimensional time-frequency graph to obtain signal features.

[0036] Optionally, the two-dimensional processing unit specifically includes:

[0037] The wavelet transform subunit is used to perform a one-dimensional continuous wavelet transform operation on the cropped data to obtain a one-dimensional signal.

[0038] A two-dimensional encoding subunit is used to perform two-dimensional encoding on the one-dimensional signal to obtain a two-dimensional time-frequency diagram.

[0039] Optionally, it also includes:

[0040] A model determination module is used to determine the signal detection model;

[0041] The model determination module specifically includes:

[0042] A training data determination unit is used to acquire training data; the training data includes measured electromagnetic signals and corresponding tags; the measured electromagnetic signals include normal signals and suspected abnormal signals;

[0043] Network building blocks are used to build deep convolutional neural network models;

[0044] The network training unit is used to input the training data into the deep convolutional neural network model, and to train the deep convolutional neural network model using cross-validation, thereby determining the trained deep convolutional neural network model as a signal detection model.

[0045] Optionally, the signal recognition module specifically includes:

[0046] The feature point determination unit is used to calculate the time-frequency domain statistical distribution characteristics of the target electromagnetic signal to obtain suspicious abnormal signal feature points.

[0047] An anomaly probability determination unit is used to perform probability calculations on feature points within a first range to obtain an anomaly probability; the first range feature points are feature points located outside the signal distribution feature map among the suspected anomaly signal feature points.

[0048] The signal recognition unit is used to determine that the target electromagnetic signal is abnormal when the abnormal probability is greater than a preset abnormal threshold.

[0049] According to specific embodiments provided by the present invention, the present invention discloses the following technical effects:

[0050] This invention discloses a method and system for detecting electromagnetic anomalies. The method includes extracting features from a target electromagnetic signal to obtain signal features, inputting these features into a signal detection model for detection, distinguishing between normal signals and suspected abnormal signals, calculating the time-frequency domain statistical distribution characteristics of suspected abnormal target electromagnetic signals, and then identifying the signal based on the calculation results using the signal distribution feature map to determine whether an abnormal signal exists in the target area. The signal distribution feature map is determined based on the calculation of the time-frequency domain statistical distribution characteristics of the measured electromagnetic signal. This invention integrates the calculation of the time-frequency domain statistical distribution characteristics with the signal detection model, achieving high accuracy and applicability for detecting abnormal signals in complex electromagnetic environments through the signal distribution feature map. Attached Figure Description

[0051] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the drawings used in the embodiments will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0052] Figure 1 This is a flowchart of the electromagnetic anomaly signal detection method of the present invention;

[0053] Figure 2 This is a logic flowchart of the electromagnetic anomaly signal detection method of the present invention;

[0054] Figure 3 This is a schematic diagram of the acquisition of a single type of electromagnetic signal in an embodiment of the present invention;

[0055] Figure 4 This is a schematic diagram illustrating the acquisition of multiple types of electromagnetic signals in an embodiment of the present invention;

[0056] Figure 5 This is a time-domain waveform diagram of the electromagnetic signal of symbol 1 in five carrier frequency bands in an embodiment of the present invention; wherein, Figure 5 (a) shows the time-domain waveform of the symbol 1 electromagnetic signal on a 50MHz carrier wave. Figure 5 (b) is the time-domain waveform of the electromagnetic signal of symbol 1 on an 80MHz carrier wave; Figure 5 (c) is the time-domain waveform of the electromagnetic signal of symbol 1 on a 160MHz carrier wave; Figure 5 (d) is the time-domain waveform of the electromagnetic signal of symbol 1 on a 300MHz carrier wave; Figure 5 (e) is the time-domain waveform of the electromagnetic signal of symbol 1 on a 500MHz carrier wave;

[0057] Figure 6 This is a time-domain waveform diagram of the electromagnetic signal of symbol 2 in five carrier frequency bands in an embodiment of the present invention; wherein, Figure 6 (a) shows the time-domain waveform of the symbol 2 electromagnetic signal on a 50MHz carrier wave. Figure 6 (b) shows the time-domain waveform of the symbol 2 electromagnetic signal on an 80MHz carrier wave. Figure 6 (c) shows the time-domain waveform of the symbol 2 electromagnetic signal on a 160MHz carrier wave. Figure 6 (d) is the time-domain waveform of the electromagnetic signal of symbol 2 on a 300MHz carrier wave; Figure 6 (e) is the time-domain waveform of the symbol 2 electromagnetic signal on a 500MHz carrier wave;

[0058] Figure 7This is a time-domain signal waveform diagram of the mixed electromagnetic signal of symbol 1 and symbol 2 in five carrier frequency bands in an embodiment of the present invention; wherein, Figure 7 (a) shows the time-domain waveform of the mixed electromagnetic signal on a 50MHz carrier wave. Figure 7 (b) shows the time-domain waveform of the mixed electromagnetic signal on an 80MHz carrier wave. Figure 7 (c) shows the time-domain waveform of the mixed electromagnetic signal on a 160MHz carrier wave. Figure 7 (d) shows the time-domain waveform of the mixed electromagnetic signal on a 300MHz carrier wave. Figure 7 (e) is the time-domain waveform of the mixed electromagnetic signal on a 500MHz carrier wave;

[0059] Figure 8 These are time-domain signal waveforms of other types of electromagnetic signals in five carrier frequency bands in embodiments of the present invention; wherein, Figure 8 (a) shows the time-domain waveforms of other types of electromagnetic signals on a 900-940MHz carrier wave. Figure 8 (b) shows the time-domain waveforms of other types of electromagnetic signals on a carrier wave of 1820-1860MHz; Figure 8 (c) shows the time-domain signal waveforms of other types of electromagnetic signals on a 2480-2520MHz carrier wave. Figure 8 (d) shows the time-domain waveforms of other types of electromagnetic signals on a 2560-2600MHz carrier wave. Figure 8 (e) shows the time-domain waveforms of other types of electromagnetic signals on a 2600-2640MHz carrier wave.

[0060] Figure 9 This is a two-dimensional time-frequency diagram of the electromagnetic signal in an embodiment of the present invention; wherein, Figure 9 (a) is a two-dimensional time-frequency diagram of symbol 1; Figure 9 (b) is a two-dimensional time-frequency diagram of symbol 2; Figure 9 (c) is a two-dimensional time-frequency diagram of a mixture of symbol 1 and symbol 2; Figure 9 (d) represents other types of two-dimensional time-frequency diagrams;

[0061] Figure 10 This refers to the detection results of symbol 1 in five carrier frequency bands in this embodiment of the invention; wherein, Figure 10 (a) shows the waveform of the detection result at a 50MHz carrier wave; Figure 10 (b) shows the detection waveform of symbol 1 on an 80MHz carrier wave. Figure 10 (c) shows the detection result waveform of symbol 1 on a 160MHz carrier wave; Figure 10 (d) is the waveform of the detection result of symbol 1 on a 300MHz carrier wave; Figure 10(e) is the waveform of the detection result of symbol 1 on a 500MHz carrier wave;

[0062] Figure 11 This refers to the detection results of symbol 2 in five carrier frequency bands in this embodiment of the invention; wherein, Figure 11 (a) shows the detection waveform of symbol 2 on a 50MHz carrier wave; Figure 11 (b) shows the detection waveform of symbol 2 on an 80MHz carrier wave; Figure 11 (c) shows the detection waveform of symbol 2 on a 160MHz carrier wave; Figure 11 (d) is the waveform of the detection result of symbol 2 with a 300MHz carrier wave; Figure 11 (e) is the waveform of the detection result of symbol 2 on a 500MHz carrier wave;

[0063] Figure 12 This is the detection result of the mixed electromagnetic signal of code element 1 and code element 2 in five carrier frequency bands in an embodiment of the present invention; wherein, Figure 12 (a) shows the waveform of the detection result of the mixed electromagnetic signal at a 50MHz carrier wave; Figure 12 (b) shows the waveform of the detection result of the mixed electromagnetic signal with an 80MHz carrier wave; Figure 12 (c) shows the waveform of the detection result of the mixed electromagnetic signal at a 160MHz carrier wave; Figure 12 (d) shows the waveform of the detection result of the mixed electromagnetic signal with a 300MHz carrier wave; Figure 12 (e) is the waveform of the detection result of the mixed electromagnetic signal at a 500MHz carrier wave;

[0064] Figure 13 These are the detection results of other types of electromagnetic signals in five carrier frequency bands in this embodiment of the invention; wherein, Figure 13 (a) shows the detection waveforms of other types of electromagnetic signals with a carrier frequency of 900-940MHz; Figure 13 (b) shows the detection waveforms of other types of electromagnetic signals with a carrier frequency of 1820-1860MHz; Figure 13 (c) shows the detection waveforms of other types of electromagnetic signals with a carrier frequency of 2480-2520MHz; Figure 13 (d) shows the detection waveforms of other types of electromagnetic signals with a carrier frequency of 2560-2600MHz; Figure 13 (e) shows the detection waveforms of other types of electromagnetic signals with a carrier frequency of 2600-2640MHz;

[0065] Figure 14 This is a signal distribution feature map and the final recognition result in an embodiment of the present invention; wherein, Figure 14 (a) shows the signal distribution characteristics. Figure 14 (b) shows the distribution of characteristic points of electromagnetic signals without abnormalities in the signal distribution characteristic map; Figure 14 (c) shows the distribution of anomalous electromagnetic signal feature points in the signal distribution feature map;

[0066] Figure 15 This is a schematic diagram of the electromagnetic anomaly signal detection system of the present invention. Detailed Implementation

[0067] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0068] The purpose of this invention is to provide a method and system for detecting electromagnetic anomalies, which can improve the applicability and accuracy of anomaly signal detection in complex electromagnetic environments.

[0069] To make the above-mentioned objects, features and advantages of the present invention more apparent and understandable, the present invention will be further described in detail below with reference to the accompanying drawings and specific embodiments.

[0070] like Figure 1 As shown, an embodiment of the present invention provides a method for detecting electromagnetic anomalies, comprising:

[0071] Step 100: Acquire the target electromagnetic signal.

[0072] Step 200: Extract features from the target electromagnetic signal to obtain signal features.

[0073] As a preferred embodiment of step 200, it specifically includes:

[0074] Step 201: Perform sliding window clipping on the target electromagnetic signal to obtain clipped data.

[0075] Step 202: Perform a continuous wavelet transform on the cropped data to obtain a two-dimensional time-frequency graph. Specifically, the process involves: performing a one-dimensional continuous wavelet transform on the cropped data to obtain a one-dimensional signal; and then performing two-dimensional encoding on the one-dimensional signal to obtain a two-dimensional time-frequency graph.

[0076] Step 203: Extract features from the two-dimensional time-frequency graph to obtain signal features.

[0077] Step 300: Determine the signal detection result of the target electromagnetic signal based on the signal characteristics and the signal detection model; the signal detection model includes an identity mapping module and a convolutional layer module connected in sequence; the signal detection result is that the target electromagnetic signal is a normal signal or the target electromagnetic signal is a suspicious abnormal signal; the type of the normal signal is within the set carrier type range; the type of the suspicious abnormal signal is outside the set carrier type range.

[0078] The method for determining the signal detection model is as follows:

[0079] Step 301: Obtain training data; the training data includes measured electromagnetic signals and corresponding tags; the measured electromagnetic signals include normal signals and suspected abnormal signals.

[0080] Step 302: Construct a deep convolutional neural network model.

[0081] Step 303: Input the training data into the deep convolutional neural network model, and train the deep convolutional neural network model using cross-validation. The trained deep convolutional neural network model is then determined as the signal detection model.

[0082] Step 400: When the signal detection result indicates that the target electromagnetic signal is a suspicious abnormal signal, the time-frequency domain statistical distribution characteristics of the target electromagnetic signal are calculated, and the calculation result is used to identify the signal based on the signal distribution characteristic map to determine whether the target electromagnetic signal is abnormal; the signal distribution characteristic map is determined based on the time-frequency domain statistical distribution characteristics of the measured electromagnetic signal.

[0083] As a preferred embodiment of step 400, it specifically includes:

[0084] Step 401: Calculate the time-frequency domain statistical distribution characteristics of the target electromagnetic signal to obtain the feature points of the suspicious abnormal signal.

[0085] Step 402: Perform probability calculations on the feature points within the first range to obtain the anomaly probability; the first range feature points are the feature points located outside the signal distribution feature map among the suspected abnormal signal feature points; based on the 13 time-frequency domain statistical distribution features proposed in this embodiment, such as Figure 14 As shown, when the number of signal feature points located outside the range of the signal distribution feature map is greater than 2, that is, when the abnormal probability is greater than 15%, it is judged as a suspicious abnormal signal.

[0086] Step 403: When the anomaly probability is greater than a preset anomaly threshold, the target electromagnetic signal is determined to be abnormal. In this embodiment, the preset anomaly threshold is 15%, and the preset anomaly threshold can be set according to the actual application.

[0087] To verify the method of the present invention, such as Figure 2 A specific embodiment is presented as shown:

[0088] First, we will determine the electromagnetic signal data acquisition method and perform data preprocessing.

[0089] The first step is to acquire the signal. The acquisition method used is as follows: Figures 3-4 As shown.

[0090] The electromagnetic signal source is a programmable source (carrier and symbol information are adjustable), outputting multiple types of electromagnetic signals by changing the carrier frequency and symbol information. Five carrier frequencies are selected: 50MHz, 80MHz, 160MHz, 300MHz, and 500MHz. The actual data collected are different types of electromagnetic signals radiated into the electromagnetic space by the transmitting antenna, and then the electromagnetic signals radiated in the electromagnetic space are collected by a data acquisition device connected to the receiving antenna. The sampling frequency is 50MHz, and the sampling time is 2ms. Other signal frequencies include 900-940MHz, 1820-1860MHz, 2480-2520MHz, 2560-2600MHz, and 2600-2640MHz, etc. Figures 5-8 As shown in Table 1, the specific parameters of the electromagnetic signals in this embodiment are as follows.

[0091] Table 1 Specific parameters of various types of electromagnetic signals

[0092]

[0093] The second step is to process the collected signals.

[0094] The acquired target electromagnetic signals are cropped. The number of data points acquired for a single-type carrier electromagnetic signal is 50MHz × 2ms = 100,000. To construct sufficient data samples, this embodiment uses a sliding window method to crop the original data. The original data is saved in a sliding window period of 50,000 points, becoming a new data sample. The data is acquired in 2ms increments, with each slide interval of 642 data points, for a total of 80 slides and saves. The data is then saved and a dataset is constructed. The signal types in the constructed dataset include symbol 1, symbol 2, a mixture of symbol 1 and symbol 2, and others. Among them, symbol 1, symbol 2, and the mixed signal of symbol 1 and symbol 2 contain 5 carrier types, and other signals include communication signals from 5 frequency bands. Using a sliding window to crop the signal has the advantages of avoiding data redundancy and reducing computational complexity.

[0095] The clipped signal data is subjected to two-dimensional temporal representation based on continuous wavelet transform (CWT). Many types of electromagnetic signals exhibit strong stationary properties in both the time and frequency domains. Wavelet analysis allows for refined operations on different frequency components, acting as a time-frequency domain magnifier to fully reveal the characteristics of electromagnetic signals. Therefore, one-dimensional continuous wavelet transform (CWT) can directly represent the time-frequency domain information of the signal simultaneously as a two-dimensional image. The clipped electromagnetic time-domain signal is defined as f(t)∈L. 2 If (R), then the expression for the continuous wavelet transform of the signal f(t) is:

[0096]

[0097] In equation (1), the sub-wavelet ψ with scale a and translation τ is used. a,τ The inner product of f(t) and the signal f(t) yields each wavelet transform coefficient WT. f (a,τ), <f,ψ a,τ (t)> represents the inner product of two functions. The more similar the signal is to the wavelet, the larger the coefficient value, and the characteristic components will gather into a high-amplitude energy block at some point on the phase plane of the time scale; conversely, the energy will diverge. Through different scale transformations, the coefficient matrix WT is finally obtained. f (a,τ) is used to characterize the two-dimensional time-scale features of a time-domain electromagnetic signal. The original electromagnetic signal undergoes a two-dimensional transformation using continuous wavelet transform (CWT), and the image size is 512×512, as shown below. Figure 9 As shown.

[0098] Secondly, the construction and training of the signal detection model.

[0099] The first step is to use the ResNet50 deep convolutional neural network as the basis for building the signal detection model.

[0100] The performance of traditional neural networks is influenced by both network width and depth, with depth having a more significant impact than width. However, simply stacking convolutional layers can lead to gradient explosion during backpropagation, making it difficult for the network to converge and directly affecting the accuracy of training and testing. To address this issue, this invention utilizes the ResNet50 deep convolutional neural network framework and proposes a deep convolutional neural network model with added identity mapping, effectively solving the performance degradation problem of deep networks.

[0101] The ResNet50 deep convolutional neural network consists of two parts: an identity mapping part and a convolutional layer structure. The identity mapping function is designed as H(x) = F(x) + x, and the residual function is further expressed as F(x) = H(x) - x, where F(x) is the network mapping before summation, and H(x) is the network mapping from the input to the network after summation. When F(x) = 0, an identity mapping H(x) = x is formed, which makes it easier to fit the residual. For backpropagation, let the loss function be ε, and the chain rule of backpropagation is expressed as equation (2):

[0102]

[0103] in, To perform partial derivative operations; x l Features of shallow layer unit l in ResNet50 network; x L Features of deep unit L in ResNet50 network; x i For any layer in the backpropagation of the residual network; w i is the updated parameter value in the backpropagation of the residual network; i is the i-th layer in the backpropagation of the residual network.

[0104] In equation (2), Without passing through the weight layer, Through the transmission of weight layers, the former ensures that the electromagnetic signal features of the input network can be directly transmitted back to any structural layer, while the latter ensures that the network structure will not experience gradient vanishing, because its value cannot be -1. This allows the performance of deep convolutional neural networks to remain in an optimal state as the depth increases.

[0105] The convolutional layers of the ResNet50 deep convolutional neural network involve two basic convolutional blocks, named ConBlock and Identity Block, which are connected in parallel with different dimensions. The Identity Block has the same input and output dimensions and can be concatenated to increase the network depth; the Con Block has different input and output dimensions and cannot be concatenated, its function is to change the network's dimensions.

[0106] The ResNet50 deep convolutional neural network is divided into three parts: input, intermediate layers, and output. The input part includes one 7×7 convolutional layer and one 3×3 pooling layer. The 7×7 kernel size is chosen to increase the receptive field, allowing the model's input features to retain global information and semantic relationships with adjacent features, as well as establishing complex spatial relationships, thus improving model performance. The intermediate layers are mainly divided into four blocks, each constructed primarily of three bottleneck structures. The bottleneck structure in this invention changes the dimension of the input feature vector through 1×1 convolutions, performs feature learning using small 3×3 convolutions, and then adds an identity mapping to the residual term. This structure can more fully utilize features at network depth, establishing strong correlations between multiple feature maps in each layer. Furthermore, through the linear addition of different feature maps within the same layer, the model can quickly generate high responses and activations to discriminative regions.

[0107] Specifically, in this embodiment of the invention, the two-dimensional time-frequency graph of the input network model has a size of 512×512. After convolution operation with a 7×7 window and a stride of 2, and then downsampling operation by a pooling layer with a 3×3 window and a stride of 2, the output feature size is 64×64. The subsequent intermediate layers consist of 4 blocks (each block has 3 bottleneck residual structures), for a total of 12 bottleneck residual structures, compressing the output feature size from 32×32 to 16×16. The output layer outputs through a fully connected layer, and finally, a softmax layer is used to detect multiple types of electromagnetic signals.

[0108] The second step is to train the ResNet50 deep convolutional neural network.

[0109] 1) The collected signals of various types were preprocessed using continuous wavelet transform (CWT). One-dimensional electromagnetic signals were encoded into two-dimensional time-frequency graphs. The dataset was divided into four categories: symbol 1, symbol 2, a mixture of symbol 1 and 2, and other categories, with a total of 20,000 training data examples for each category. The dataset was divided into training, test, and validation sets in a ratio of 8:1:1. Labels were added to the four datasets, which were then used as input to the ResNet50 deep convolutional neural network model for feature extraction.

[0110] 2) During model training, the recognition model iteratively predicts data in the test set based on training rounds. Cross-validation is used to add correctly identified results to the training set. Through multiple iterations, the SGD optimizer is used during training. The loss function value and validation accuracy are used as convergence metrics until the recognition model's accuracy on both the training and test sets, as well as its recognition rate for the four types of electromagnetic signals, reaches its optimal level, thus completing the construction of the electromagnetic signal recognition model based on a deep convolutional neural network.

[0111] 3) To ensure the training efficiency of the ResNet50 deep convolutional neural network model, the parameters to be tuned in the network include: learning rate, batch size, number of iterations, and activation function of the network connection layer.

[0112] The third step is to use the detection results based on the trained ResNet50 deep convolutional neural network model.

[0113] To evaluate the model's detection performance, this embodiment uses the validation set in the dataset to conduct signal detection experiments on the ResNet50 deep convolutional neural network model. The waveform of the experimental detection results is shown below. Figures 10-13 As shown, the detection accuracy for all types of signals can reach 100%. The validation set includes three types of electromagnetic signals: code element 1, code element 2, and a mixed signal of code element 1 and 2, each containing 5 carrier types. Other signal types include electromagnetic signals in 5 frequency bands. The 5 different carrier (frequency band) signals contained in each signal type belong to their respective categories. This has the following advantages: First, it increases the applicability of the recognition model and improves the model's recognition accuracy. Second, when the signal detection result is code element 1, code element 2, or a mixed signal of code element 1 and 2, the detected signal is determined to be a normal signal; when the signal detection result is other types, it indicates that an abnormal signal may exist, i.e., it is judged as a suspicious abnormal signal. Classifying suspicious abnormal signals into other categories facilitates further calculation and analysis of the signal's time-frequency domain statistical distribution characteristics, which can reduce the false alarm rate.

[0114] Finally, the model detection results are identified by calculating the statistical distribution characteristics of the signal in the time and frequency domain.

[0115] To enhance the interpretability of feature extraction based on the ResNet50 deep convolutional neural network, 13 time-frequency domain statistical distribution features were calculated for electromagnetic signal data classified as other classes by the recognition model. The specific features and corresponding calculation methods are shown in Table 2. The calculation results were saved to an Excel spreadsheet in a specified location. Based on the data in the table, the fluctuation range of the feature value probability distribution was plotted, i.e., a feature map. The fluctuation range was then tested using feature calculations for electromagnetic signals classified as other classes by the model. Figure 14As shown, based on the probability distribution feature map of the predicted signal, the probability that the feature points of the suspicious abnormal signal fall into the feature map interval is used as the evaluation index for abnormal signal identification, thus completing the detection and identification of abnormal signals.

[0116] Table 2 shows the statistical distribution characteristics of the 13 time-frequency domain types.

[0117]

[0118] The beneficial effects of this invention are:

[0119] 1. This invention proposes for the first time an electromagnetic anomaly signal detection method based on the time-frequency statistical feature distribution of signals and deep learning. This method leverages the advantages of deep learning in signal detection and feature extraction, proposing a ResNet50 deep convolutional neural network model for the detection and identification of multiple types of electromagnetic signals. To enhance the interpretability of deep learning feature extraction, 13 time-frequency domain features are proposed for signal statistical distribution calculation based on the ResNet50 deep convolutional neural network model. Features of electromagnetic signals classified as other types by the model are calculated, and feature maps are generated based on the feature calculation results. The probability that a suspicious anomaly signal feature point falls within the feature map interval is used as an evaluation index for anomaly signal identification, thus completing the detection and identification of anomaly signals.

[0120] 2. This invention employs a sliding window approach to crop the original data. Using a sliding window to crop the signal avoids data redundancy and reduces computational complexity. The cropped electromagnetic signal time-series data is preprocessed using Continuous Wavelet Transform (CWT) to encode the one-dimensional signal into a two-dimensional time-frequency graph, constructing an electromagnetic signal dataset with a size of 512×512. A total of 20,000 training data examples are provided, divided into training, testing, and validation sets in a ratio of 8:1:1.

[0121] 3. This invention relates to the acquisition of multiple types of electromagnetic signals. The specific method is as follows: The electromagnetic signal source uses a programmable signal source (carrier and symbol information are adjustable), and multiple types of electromagnetic signals are output by changing the carrier frequency and symbol information. The actual acquired data consists of different types of electromagnetic signals radiated into the electromagnetic space through a transmitting antenna, and then the electromagnetic signals radiated in the electromagnetic space are acquired through a data acquisition device connected to a receiving antenna. The sampling frequency is 50MHz, and the sampling time is 2ms. The carrier frequencies for the signal types are symbol 1, symbol 2, and mixed signals of symbols 1 and 2, selected from five options: 50MHz, 80MHz, 160MHz, 300MHz, and 500MHz. Other signal frequencies include 900-940MHz, 1820-1860MHz, 2480-2520MHz, 2560-2600MHz, and 2600-2640MHz.

[0122] 4. The performance of traditional neural networks is affected by both network width and depth. Simply stacking convolutional layers leads to gradient explosion during backpropagation, making convergence difficult and directly impacting training and testing accuracy. To address this issue, this invention utilizes the ResNet50 deep convolutional neural network framework and proposes a deep convolutional neural network model incorporating an identity mapping method, effectively solving the performance degradation problem of deep networks. The chain-like backpropagation ensures that the performance of the deep convolutional neural network remains optimal as depth increases, effectively avoiding gradient vanishing. During model training, the recognition model iteratively predicts data in the test set based on training rounds. Cross-validation is used to add correctly identified results to the training set, iterating through multiple rounds until the recognition model achieves its optimal recognition rate. To ensure the training efficiency of the ResNet50 deep convolutional neural network model, the parameters tuned include: learning rate, batch size, number of iterations, and activation functions of network connection layers.

[0123] like Figure 15 As shown, the present invention also provides an electromagnetic anomaly signal detection system, including: a signal acquisition module 1, a feature extraction module 2, a signal detection module 3, and a signal recognition module 4.

[0124] Specifically, signal acquisition module 1 is used to acquire the target electromagnetic signal. Feature extraction module 2 is used to extract features from the target electromagnetic signal to obtain signal features. Signal detection module 3 is used to determine the signal detection result of the target electromagnetic signal based on the signal features and the signal detection model; the signal detection model includes an identity mapping module and a convolutional layer module connected in sequence; the signal detection result is that the target electromagnetic signal is a normal signal or a suspicious abnormal signal; the type of the normal signal is within a set carrier type range; the type of the suspicious abnormal signal is outside the set carrier type range. Signal recognition module 4 is used to calculate the time-frequency domain statistical distribution characteristics of the target electromagnetic signal when the signal detection result is that the target electromagnetic signal is a suspicious abnormal signal, and to perform signal recognition on the calculation result based on the signal distribution feature map to determine whether the target electromagnetic signal is abnormal; the signal distribution feature map is determined based on the calculation of the time-frequency domain statistical distribution characteristics of the measured electromagnetic signal.

[0125] As a preferred embodiment, the feature extraction module 2 specifically includes: a cropping unit, a two-dimensional processing unit, and a feature extraction unit.

[0126] Specifically, the clipping unit performs sliding window clipping on the target electromagnetic signal to obtain clipped data. The two-dimensional processing unit performs continuous wavelet transform on the clipped data to obtain a two-dimensional time-frequency graph. The feature extraction unit extracts features from the two-dimensional time-frequency graph to obtain signal features.

[0127] In a preferred embodiment, the two-dimensional processing unit specifically includes a wavelet transform subunit and a two-dimensional coding subunit.

[0128] Specifically, the wavelet transform subunit is used to perform a one-dimensional continuous wavelet transform operation on the cropped data to obtain a one-dimensional signal. The two-dimensional encoding subunit is used to perform two-dimensional encoding on the one-dimensional signal to obtain a two-dimensional time-frequency diagram.

[0129] The electromagnetic anomaly signal detection system further includes:

[0130] The model determination module is used to determine the signal detection model.

[0131] The model determination module specifically includes: a training data determination unit, a network construction unit, and a network training unit.

[0132] Specifically, the training data determination unit is used to acquire training data; the training data includes measured electromagnetic signals and corresponding labels; the measured electromagnetic signals include normal signals and suspected abnormal signals. The network construction unit is used to construct a deep convolutional neural network model. The network training unit is used to input the training data into the deep convolutional neural network model and train the deep convolutional neural network model using cross-validation, and determine the trained deep convolutional neural network model as the signal detection model.

[0133] In a preferred embodiment, the signal recognition module 4 specifically includes: a feature point determination unit, an anomaly probability determination unit, and a signal recognition unit.

[0134] Specifically, the feature point determination unit is used to calculate the time-frequency domain statistical distribution characteristics of the target electromagnetic signal to obtain suspicious abnormal signal feature points. The anomaly probability determination unit is used to perform probability calculations on the feature points within a first range to obtain an anomaly probability; the first range feature points are the feature points located outside the signal distribution feature map among the suspicious abnormal signal feature points. The signal identification unit is used to determine that the target electromagnetic signal is abnormal when the anomaly probability is greater than a preset anomaly threshold.

[0135] The various embodiments in this specification are described in a progressive manner, with each embodiment focusing on its differences from other embodiments. Similar or identical parts between embodiments can be referred to interchangeably. For the systems disclosed in the embodiments, since they correspond to the methods disclosed in the embodiments, the descriptions are relatively simple; relevant parts can be referred to the method section.

[0136] This document uses specific examples to illustrate the principles and implementation methods of the present invention. The descriptions of the above embodiments are only for the purpose of helping to understand the method and core ideas of the present invention. Furthermore, those skilled in the art will recognize that, based on the ideas of the present invention, there will be changes in the specific implementation methods and application scope. Therefore, the content of this specification should not be construed as a limitation of the present invention.

Claims

1. A method for detecting electromagnetic anomaly signals, characterized in that, include: Acquire the target electromagnetic signal; Feature extraction is performed on the target electromagnetic signal to obtain signal features, specifically including: The target electromagnetic signal is clipped using a sliding window to obtain clipped data; The cropped data is subjected to continuous wavelet transform to obtain a two-dimensional time-frequency graph. Feature extraction is performed on the two-dimensional time-frequency graph to obtain signal features; The step of performing continuous wavelet transform to two-dimensionalize the cropped data to obtain a two-dimensional time-frequency graph specifically includes: Perform a one-dimensional continuous wavelet transform operation on the cropped data to obtain a one-dimensional signal; The one-dimensional signal is encoded in two dimensions to obtain a two-dimensional time-frequency diagram; The clipped electromagnetic signal is defined as f(t)∈L 2 If (R), then the expression for the continuous wavelet transform of the signal f(t) is: ； In the formula, the sub-wavelet ψ has a scale of a and a translation amount of τ. a,τ The inner product of f(t) and the signal f(t) yields each wavelet transform coefficient WT. f (a,τ), <f,ψ a,τ (t) > represents the inner product of two functions; The signal detection result of the target electromagnetic signal is determined based on the signal characteristics and the signal detection model; the signal detection model includes an identity mapping module and a convolutional layer module connected in sequence; the signal detection result is that the target electromagnetic signal is a normal signal or the target electromagnetic signal is a suspicious abnormal signal; the type of the normal signal is within the set carrier type range; the type of the suspicious abnormal signal is outside the set carrier type range. When the signal detection result indicates that the target electromagnetic signal is a suspicious abnormal signal, the time-frequency domain statistical distribution characteristics of the target electromagnetic signal are calculated, and the calculation result is used to identify the signal based on the signal distribution characteristic map to determine whether the target electromagnetic signal is abnormal; the signal distribution characteristic map is determined based on the time-frequency domain statistical distribution characteristics calculation of the measured electromagnetic signal; The step of calculating the time-frequency domain statistical distribution characteristics of the target electromagnetic signal and identifying the signal based on the signal distribution characteristic map to determine whether the target electromagnetic signal is abnormal specifically includes: Calculate the time-frequency domain statistical distribution characteristics of the target electromagnetic signal to obtain suspicious abnormal signal feature points; A probability calculation is performed on the feature points within the first range to obtain the anomaly probability; the first range feature points are the feature points of the suspected abnormal signal feature points located outside the signal distribution feature map. When the probability of an anomaly is greater than a preset anomaly threshold, the target electromagnetic signal is determined to be abnormal.

2. The electromagnetic anomaly signal detection method according to claim 1, characterized in that, The method for determining the signal detection model is as follows: Acquire training data; the training data includes measured electromagnetic signals and corresponding tags; the measured electromagnetic signals include normal signals and suspected abnormal signals; Construct a deep convolutional neural network model; The training data is input into the deep convolutional neural network model, and the deep convolutional neural network model is trained using cross-validation. The trained deep convolutional neural network model is then determined as the signal detection model.

3. An electromagnetic anomaly signal detection system, characterized in that, include: The signal acquisition module is used to acquire the target electromagnetic signal; The feature extraction module is used to extract features from the target electromagnetic signal to obtain signal features, specifically including: The clipping unit is used to perform sliding window clipping on the target electromagnetic signal to obtain clipped data; A two-dimensional processing unit is used to perform continuous wavelet transform two-dimensionalization operation on the cropped data to obtain a two-dimensional time-frequency diagram. The feature extraction unit is used to extract features from the two-dimensional time-frequency graph to obtain signal features; The two-dimensional processing unit specifically includes: The wavelet transform subunit is used to perform a one-dimensional continuous wavelet transform operation on the cropped data to obtain a one-dimensional signal. A two-dimensional encoding subunit is used to perform two-dimensional encoding on the one-dimensional signal to obtain a two-dimensional time-frequency diagram; The clipped electromagnetic signal is defined as f(t)∈L 2 If (R), then the expression for the continuous wavelet transform of the signal f(t) is: ； In the formula, the sub-wavelet ψ has a scale of a and a translation amount of τ. a,τ The inner product of f(t) and the signal f(t) yields each wavelet transform coefficient WT. f (a,τ), <f,ψ a,τ (t) > represents the inner product of two functions; A signal detection module is used to determine the signal detection result of the target electromagnetic signal based on the signal characteristics and a signal detection model; the signal detection model includes an identity mapping module and a convolutional layer module connected in sequence; the signal detection result is that the target electromagnetic signal is a normal signal or the target electromagnetic signal is a suspicious abnormal signal; the type of the normal signal is within a set carrier type range; the type of the suspicious abnormal signal is outside the set carrier type range. The signal recognition module is used to calculate the time-frequency domain statistical distribution characteristics of the target electromagnetic signal when the signal detection result indicates that the target electromagnetic signal is a suspicious abnormal signal, and to perform signal recognition based on the calculation result according to the signal distribution feature map to determine whether the target electromagnetic signal is abnormal; the signal distribution feature map is determined based on the calculation of the time-frequency domain statistical distribution characteristics of the measured electromagnetic signal; The signal recognition module specifically includes: The feature point determination unit is used to calculate the time-frequency domain statistical distribution characteristics of the target electromagnetic signal to obtain suspicious abnormal signal feature points. An anomaly probability determination unit is used to perform probability calculations on feature points within a first range to obtain an anomaly probability; the first range feature points are feature points located outside the signal distribution feature map among the suspected anomaly signal feature points. The signal recognition unit is used to determine that the target electromagnetic signal is abnormal when the abnormal probability is greater than a preset abnormal threshold.

4. The electromagnetic anomaly signal detection system according to claim 3, characterized in that, Also includes: A model determination module is used to determine the signal detection model; The model determination module specifically includes: A training data determination unit is used to acquire training data; the training data includes measured electromagnetic signals and corresponding tags; the measured electromagnetic signals include normal signals and suspected abnormal signals; Network building blocks are used to build deep convolutional neural network models; The network training unit is used to input the training data into the deep convolutional neural network model, and to train the deep convolutional neural network model using cross-validation, thereby determining the trained deep convolutional neural network model as a signal detection model.

Images