Electromagnetic signal clustering processing method and apparatus, electronic device, and readable storage medium

Abstract

The application discloses an electromagnetic signal clustering processing method and device, electronic equipment and a readable storage medium. The method comprises the following steps: acquiring an electromagnetic signal set to be clustered; using an unsupervised clustering model to perform clustering processing on a plurality of electromagnetic signals in the electromagnetic signal set, to obtain a plurality of clustered electromagnetic signals; inputting each clustered electromagnetic signal into a deep autoencoder to obtain a plurality of target feature maps of a target hidden layer output of a decoder and corresponding joint loss values; obtaining an importance weight corresponding to each target feature map according to the joint loss values, weighting and fusing the plurality of target feature maps according to the importance weight, and performing saliency feature extraction on the fused feature map after weighting and fusion to obtain a saliency map of each clustered electromagnetic signal; and superimposing the saliency map of each clustered electromagnetic signal into a waveform map of the electromagnetic signal to obtain a visual representation of the saliency feature in each clustered electromagnetic signal.

Description

Technical Field

[0001] This application relates to the field of signal processing technology, and in particular to an electromagnetic signal clustering processing method, apparatus, electronic device, and readable storage medium. Background Technology

[0002] Unsupervised electromagnetic signal clustering plays a crucial role in fields such as electromagnetic spectrum monitoring, cognitive radio, and cyberspace security. Unsupervised electromagnetic signal clustering is not only a classification method for unsupervised situations but can also be used for automatic annotation of large-scale existing electromagnetic signals. Compared to supervised electromagnetic signal clustering methods, unsupervised electromagnetic signal clustering methods have a wider range of applications, eliminating the need for manual data annotation and better meeting the practical application needs of recognition systems.

[0003] However, our understanding of the internal logic and behavior of unsupervised clustering models is currently very limited. During the clustering process, we only need to input the set of electromagnetic signals to be clustered into the unsupervised clustering model, but we cannot know what knowledge the unsupervised clustering model learns from the original electromagnetic signals or how it arrives at the final decision result. We cannot explain how the unsupervised clustering model achieves such good performance. Without a clear understanding of how unsupervised clustering models work effectively, developing better unsupervised clustering models can only involve continuous parameter tuning and trial and error. Therefore, it is urgent to improve the interpretability of electromagnetic signal clustering methods and make unsupervised clustering models transparent.

[0004] To intuitively understand the target features and conceptual representations extracted by unsupervised clustering models, existing techniques visualize the weight parameters and hidden features learned internally by the unsupervised clustering model, transforming them into results with practical meaning that humans can understand. For example, saliency maps are used to visualize salient features in the input data that are meaningful for decision-making. The saliency map is superimposed on the waveform of the input electromagnetic signal of the unsupervised clustering model; the features in the red highlighted areas reflect the salient features relied upon by the unsupervised clustering model in the decision-making process. However, this visualization method in existing techniques requires changing the parameters of the unsupervised clustering model, and even altering its network structure. In other words, existing unsupervised clustering models need to be retrained to visualize salient features. Summary of the Invention

[0005] This application provides an electromagnetic signal clustering processing method, apparatus, electronic device, and readable storage medium to visualize the salient features of electromagnetic signals in the decision-making process of an unsupervised clustering model without modifying the parameters of the unsupervised clustering model.

[0006] The embodiments of this application adopt the following technical solutions:

[0007] In a first aspect, embodiments of this application provide an electromagnetic signal clustering processing method, including:

[0008] Obtain a set of electromagnetic signals to be clustered, wherein the set of electromagnetic signals includes multiple electromagnetic signals;

[0009] Multiple electromagnetic signals in the electromagnetic signal set are clustered using an unsupervised clustering model to obtain multiple clustered electromagnetic signals. The unsupervised clustering model includes a deep autoencoder and a clustering network. The deep autoencoder includes an encoder and a decoder.

[0010] Each clustered electromagnetic signal is input into the deep autoencoder to obtain multiple target feature maps and the joint loss value corresponding to the multiple target feature maps output by the target hidden layer of the decoder.

[0011] The importance weights corresponding to each target feature map are obtained based on the joint loss value. The multiple target feature maps are then weighted and fused according to the importance weights. The saliency features of the fused feature map are extracted to obtain the saliency map of each electromagnetic signal after clustering.

[0012] The saliency map of each clustered electromagnetic signal is superimposed on the waveform map of that electromagnetic signal to obtain a visual representation of the saliency features in each clustered electromagnetic signal.

[0013] Secondly, embodiments of this application provide an electromagnetic signal clustering processing apparatus, comprising:

[0014] A signal acquisition unit is used to acquire a set of electromagnetic signals to be clustered, wherein the set of electromagnetic signals includes multiple electromagnetic signals;

[0015] A clustering processing unit is used to perform clustering processing on multiple electromagnetic signals in the electromagnetic signal set using an unsupervised clustering model to obtain multiple clustered electromagnetic signals. The unsupervised clustering model includes a deep autoencoder and a clustering network. The deep autoencoder includes an encoder and a decoder.

[0016] The target feature map acquisition unit is used to input each clustered electromagnetic signal into the deep autoencoder to obtain multiple target feature maps and the joint loss value corresponding to the multiple target feature maps output by the target hidden layer of the decoder.

[0017] The saliency map calculation unit is used to obtain the importance weight corresponding to each target feature map based on the joint loss value, perform weighted fusion of the multiple target feature maps based on the importance weight, and extract saliency features from the fused feature map obtained after weighted fusion to obtain the saliency map of each electromagnetic signal after clustering.

[0018] The visualization processing unit is used to overlay the saliency map of each clustered electromagnetic signal onto the waveform map of the electromagnetic signal to obtain a visual representation of the saliency features in each clustered electromagnetic signal.

[0019] Thirdly, embodiments of this application provide an electronic device, including a processor; and a memory arranged to store computer-executable instructions, which, when executed, cause the processor to perform the methods of the above embodiments.

[0020] Fourthly, embodiments of this application provide a computer-readable storage medium that stores one or more programs, which, when executed by a processor, implement the methods of the above embodiments.

[0021] The above-mentioned technical solutions adopted in this application embodiment can achieve the following beneficial effects: This application embodiment uses an unsupervised clustering model including a deep autoencoder and a clustering network to cluster multiple electromagnetic signals in the electromagnetic signal set to be clustered. Each clustered electromagnetic signal is then input into the deep autoencoder. Taking the multiple target feature maps output by the target hidden layer of the decoder as objects, the importance weight of each target feature map is calculated through the joint loss value corresponding to the target feature map. The multiple target feature maps are weighted and fused using the importance weight. Since the features in the fused feature map have been distinguished based on the importance weight, the salient features that play a decisive role in the decision-making process of the unsupervised clustering model of the clustered electromagnetic signals can be accurately extracted from the fused feature map. Finally, the salient map formed by the salient features is superimposed on the waveform of the corresponding electromagnetic signal after clustering to complete the visualization representation of the salient features.

[0022] The electromagnetic signal processing in this embodiment of the application can obtain a visual representation of the salient features of each clustered electromagnetic signal without modifying the parameters of the unsupervised clustering model. People can directly observe which part of the electromagnetic signal features the unsupervised clustering model relies on when performing clustering, which helps people understand and analyze the working principle and decision-making process of the unsupervised clustering model, and provides research directions for developing better unsupervised clustering models. Attached Figure Description

[0023] The accompanying drawings, which are included to provide a further understanding of this application and form part of this application, illustrate exemplary embodiments and are used to explain this application, but do not constitute an undue limitation of this application. In the drawings:

[0024] Figure 1This is a flowchart illustrating an electromagnetic signal clustering processing method in an embodiment of this application.

[0025] Figure 2 This is a schematic diagram of an electromagnetic signal acquisition process in an embodiment of this application;

[0026] Figure 3 This is a schematic diagram of the structure of a depth autoencoder in an embodiment of this application;

[0027] Figure 4 This is a schematic diagram illustrating the visualization process of a salient feature in an embodiment of this application;

[0028] Figure 5 This is a schematic diagram of the structure of an electromagnetic signal clustering processing device according to an embodiment of this application;

[0029] Figure 6 This is a schematic diagram of the structure of an electronic device according to an embodiment of this application. Detailed Implementation

[0030] To make the objectives, technical solutions, and advantages of this application clearer, the technical solutions of this application will be clearly and completely described below in conjunction with specific embodiments and corresponding drawings. Obviously, the described embodiments are only a part of the embodiments of this application, and not all of them. Based on the embodiments in this application, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of this application.

[0031] The technical solutions provided by the various embodiments of this application are described in detail below with reference to the accompanying drawings.

[0032] This application provides an electromagnetic signal clustering processing method, such as... Figure 1 The diagram shows a flowchart of an electromagnetic signal clustering processing method according to an embodiment of this application. The method includes at least the following steps S110 to S150:

[0033] Step S110: Obtain the set of electromagnetic signals to be clustered, wherein the set of electromagnetic signals includes multiple electromagnetic signals.

[0034] The set of electromagnetic signals to be clustered includes a sufficient number of electromagnetic signals to facilitate clustering of the electromagnetic signals in the set based on a deep clustering algorithm. In this application, the set of electromagnetic signals may include one or more categories of electromagnetic signals.

[0035] To facilitate the unsupervised clustering model of this application in clustering multiple electromagnetic signals, this embodiment uses vector form to represent electromagnetic signals. For example... Figure 2As shown, an analog-to-digital converter (A / D) is used to convert the analog electromagnetic signal X received by the antenna into digital signal X. IF (t) is converted into a digital electromagnetic signal X IF (n), for digital electromagnetic signal X IF (n) After undergoing a digital orthogonal transformation, the I-channel signal x is obtained. I (n) and Q-path signal x Q (n), thus obtaining the electromagnetic signal vector form x(n)=[x I (n); x Q (n)]. Repeating this process yields multiple electromagnetic signals from the electromagnetic signal set. Of course, in practical applications, the I-channel signal x can also be... I (n) and Q-path signal x Q (n) Perform linear combination, and represent the linearly combined signal as a vector. Those skilled in the art can flexibly set the digital representation of the electromagnetic signal, and this application does not make specific limitations in this regard.

[0036] Step S120: Cluster the multiple electromagnetic signals in the electromagnetic signal set using an unsupervised clustering model to obtain multiple clustered electromagnetic signals. The unsupervised clustering model includes a deep autoencoder and a clustering network. The deep autoencoder includes an encoder and a decoder. The clustering network uses an unsupervised clustering algorithm for clustering, such as the K-means clustering algorithm.

[0037] The unsupervised clustering model in this application consists of a trained deep autoencoder and a clustering network. The encoder of the deep autoencoder reduces the dimensionality of multiple electromagnetic signals in the electromagnetic signal set to latent variable, and outputs the latent variable to the decoder and the clustering network respectively. The decoder reconstructs the corresponding multiple electromagnetic signals based on the latent variable. The clustering network performs unsupervised clustering processing on the latent variable to obtain the clustering result of multiple electromagnetic signals. The clustering result includes the category of each electromagnetic signal, that is, the category of each electromagnetic signal after clustering is known.

[0038] Figure 3 This is a schematic diagram of the structure of a depth autoencoder in an embodiment of this application, as shown below. Figure 3As shown, the encoder includes L first basic layers and a first fully connected layer. The first basic layer includes a convolutional layer, a batch normalization (BN) layer, a ReLU activation function layer, and a pooling layer. The convolutional layer performs convolution processing on the signal to extract features; the BN layer performs regularization processing on the features; the ReLU layer provides coefficients for nonlinear transformation; the pooling layer compresses the input feature map, extracts the main features, and reduces the computational complexity of the network; the fully connected layer connects all features and outputs a latent feature vector.

[0039] The decoder consists of a second fully connected layer and M second basic layers. The second fully connected layer connects all the latent feature vectors. The second basic layers consist of an upsampling layer, a ReLU layer, a BN layer, and a deconvolution layer. The upsampling layer upsamples the feature map. The ReLU layer provides coefficients for non-linear transformation. The BN layer regularizes the features. The deconvolution layer deconvolves the features.

[0040] It should be noted that the specific number of the L first basic layers and M second basic layers mentioned above can be set according to the minimum number of basic layers required for depth autoencoders in the existing technology. For example, at present, when the number of the first basic layers and the second basic layers are both greater than 5, the autoencoder constructed is a depth autoencoder.

[0041] Step S130: Each clustered electromagnetic signal is input into the deep autoencoder to obtain multiple target feature maps and the joint loss value corresponding to the target hidden layer output of the decoder.

[0042] In order to obtain the salient features of each clustered electromagnetic signal that play a decisive role in the decision-making process of the unsupervised clustering model, this application inputs each clustered electromagnetic signal into a deep autoencoder for forward propagation processing. Since the features of the feature map output by the second basic layer are closer to the original electromagnetic signal than the decoder output, in an optional embodiment of this application, the Mth second basic layer in the decoder is used as the target hidden layer, and salient features are extracted based on the feature map output by the target hidden layer. In practical applications, the (M-1)th second basic layer in the decoder can also be used as the target hidden layer.

[0043] Since the features in multiple target feature maps include both features that have a positive impact on the target class and features that have a negative impact on the target class, this application embodiment also obtains the joint loss value corresponding to multiple target feature maps, learns the importance weight corresponding to each target feature map based on the joint loss value, and uses the importance weight to increase the strength of features that have a positive impact on improving clustering accuracy, so as to accurately detect the significant features in each electromagnetic signal after clustering.

[0044] Step S140: Obtain the importance weight corresponding to each target feature map based on the joint loss value, perform weighted fusion on the multiple target feature maps based on the importance weight, and extract saliency features from the fused feature map obtained after weighted fusion to obtain the saliency map of each electromagnetic signal after clustering.

[0045] After obtaining the importance weights, this embodiment performs weighted fusion of multiple target feature maps according to the importance weights. The weighted multiple target feature maps distinguish between positive and negative features. Specifically, it strengthens the features in the target feature maps that have a positive impact on improving clustering accuracy. Thus, the features in the fused feature map that have a positive impact on improving clustering accuracy have higher intensity. In this application, the features that have a positive impact on improving clustering accuracy can be understood as salient features.

[0046] Step S150: The saliency map of each clustered electromagnetic signal is superimposed on the waveform map of the electromagnetic signal to obtain a visual representation of the saliency features in each clustered electromagnetic signal.

[0047] The saliency map of each clustered electromagnetic signal is superimposed with the waveform map of that electromagnetic signal to obtain the desired waveform map. In the desired waveform map, the features of the red highlighted areas reflect the saliency features of the clustered electromagnetic signals that play a decisive role in the decision-making process of the unsupervised clustering model. At this time, the colors in the desired waveform map can be used to intuitively observe which part of the electromagnetic signal features the unsupervised clustering model relies on when performing clustering. This can help to understand and analyze the working principle and decision-making process of the unsupervised clustering model, and provide research directions for developing better unsupervised clustering models.

[0048] based on Figure 1As shown in the electromagnetic signal processing clustering method, this embodiment utilizes an unsupervised clustering model including a deep autoencoder and a clustering network to cluster multiple electromagnetic signals in the electromagnetic signal set to be clustered. Each clustered electromagnetic signal is then input into the deep autoencoder. Taking the multiple target feature maps output by the target hidden layer of the decoder as objects, the importance weight of each target feature map is calculated through the joint loss value corresponding to the target feature map. The multiple target feature maps are then weighted and fused using the importance weights. Since the features in the fused feature map have been distinguished based on the importance weights, the salient features that play a decisive role in the decision-making process of the unsupervised clustering model can be accurately extracted from the fused feature map. Finally, the salient map formed by the salient features is superimposed on the waveform of the corresponding clustered electromagnetic signal to complete the visualization representation of the salient features. The processing in this embodiment does not require modification of the parameters of the unsupervised clustering model to obtain a visual representation of the salient features in each clustered electromagnetic signal. People can directly observe which part of the electromagnetic signal features the unsupervised clustering model relies on when performing clustering, which helps people understand and analyze the working principle and decision-making process of the unsupervised clustering model, and provides research directions for developing better unsupervised clustering models.

[0049] In one embodiment of this application, the unsupervised clustering model is trained through the following steps:

[0050] Construct a training sample set, which includes multiple electromagnetic signal samples;

[0051] Multiple electromagnetic signal samples from the training sample set are input into the deep autoencoder to obtain the latent feature vector output by the encoder, the multiple electromagnetic signal samples reconstructed by the decoder based on the latent feature vector, and the category label of each electromagnetic signal sample obtained by the clustering network after performing unsupervised clustering on the latent feature vector.

[0052] The reconstructed sample vectors of the multiple electromagnetic signal samples, the sample vectors of the multiple electromagnetic signal samples, the category label of each electromagnetic signal sample, and the latent feature vectors of the multiple electromagnetic signal samples are input into the joint loss function to obtain the joint loss value for this round.

[0053] The parameters of the deep autoencoder are updated based on the joint loss value of the current round. Multiple electromagnetic signal samples from the training sample set are input into the deep autoencoder after parameter update to obtain the joint loss value of the next round. The parameters of the deep autoencoder are updated based on the joint loss value of the next round until the model training stopping condition is met.

[0054] In this embodiment, the conditions for stopping model training include, for example, the convergence of the joint loss function, the joint loss value being less than a preset loss threshold, and the number of parameter updates reaching a preset number of updates.

[0055] In the process of training unsupervised clustering models, some existing technologies use the K-means loss function to optimize the parameters of the unsupervised clustering model to achieve the purpose of feature extraction. However, in practical applications, it has been found that even if the K-means loss during the training process can be reduced to a very small value, the trained classifier will still cause the feature space to collapse, resulting in a decrease in feature extraction energy.

[0056] Some other solutions propose using the reconstruction loss function as the loss function for training unsupervised clustering models. However, the reconstruction loss function only considers the quality of data reconstruction and does not take into account the distribution of features used for clustering. Unsupervised clustering models trained solely based on the reconstruction loss function have the problem of good electromagnetic signal reconstruction results but poor clustering results.

[0057] Unlike existing technologies, the joint loss function in this embodiment includes a reconstruction loss function and a clustering loss function. The reconstruction loss function is used to minimize the difference between the reconstructed electromagnetic signal and the original electromagnetic signal, while the clustering loss function is used to increase intra-class similarity and decrease inter-class similarity.

[0058] In one embodiment, the mathematical expression for the joint loss function is as follows:

[0059]

[0060] In equation (1), {C i Let i = 1, 2, ..., K, represent the cluster centers, and N be the number of samples. Let B be the reconstructed electromagnetic signal corresponding to electromagnetic signal x, B be the reconstruction process of the decoder, and A be the dimensionality reduction process of the encoder. md=(mind i,1 ,...,mind i,N ) T ,i∈[1,K], ||C K A(x) N || 2 Let A(x) be the latent feature vector corresponding to the Nth electromagnetic signal sample. N With C K The distance between cluster centers.

[0061] By training an unsupervised clustering model based on the above embodiments, an unsupervised clustering model with good clustering performance can be obtained.

[0062] After obtaining the trained unsupervised clustering model, it can be used to cluster multiple electromagnetic signals in the set of electromagnetic signals to be clustered. For example, it can be used to cluster electromagnetic signals X into category C. c The input is fed into the trained deep autoencoder to obtain the latent feature vectors output by the encoder. The clustering network uses an unsupervised clustering algorithm, such as the K-means algorithm, to cluster the latent feature vectors and outputs the category label for each electromagnetic signal after clustering.

[0063] After obtaining electromagnetic signal X c After the category label, then the electromagnetic signal X c The input is fed into the deep autoencoder for forward propagation processing. Multiple feature maps output from the deconvolution layer in the final base layer of the decoder are used to obtain multiple target feature maps output from the hidden target layer. Figure 3 Taking the deep autoencoder shown as an example, the Mth second basic layer is used as the target hidden layer, and the K feature maps output by the deconvolution layer in the Mth second basic layer are used. Here, W and H represent the length and width of the target feature map, respectively.

[0064] After obtaining K target feature maps, and the electromagnetic signal X c If the category labels are known, the joint loss value L of these K target feature maps can be calculated according to the above equation (1). c .

[0065] In one embodiment of this application, the joint loss value L corresponding to the plurality of target feature maps is obtained through the following steps. c :

[0066] Obtain a joint loss function for training the unsupervised clustering model, the joint loss function including a reconstruction loss function and a clustering loss function;

[0067] The feature vectors of multiple target feature maps corresponding to each clustered electromagnetic signal and the signal vector of the electromagnetic signal are input into the reconstruction loss function to obtain the reconstruction loss value;

[0068] The latent feature vector corresponding to each clustered electromagnetic signal and the category label of the electromagnetic signal are input into the clustering loss function to obtain the clustering loss value;

[0069] Based on the clustering loss value and the reconstruction loss value, the joint loss value L corresponding to the multiple target feature maps is obtained. c .

[0070] After obtaining the joint loss value L corresponding to multiple target feature maps c Then, the joint loss value L c The deep autoencoder is input for backpropagation to obtain multiple gradient feature maps of the target hidden layer output of the decoder. These multiple gradient feature maps are then compared with multiple target features. Figure 1 One-to-one correspondence; with Figure 3 The deep autoencoder shown is an example of a decoder that takes the K gradient feature maps output from the deconvolution layer in the Mth second base layer of the decoder. As the gradient feature maps required in this embodiment, global average pooling is performed on the multiple gradient feature maps to obtain the importance weight corresponding to each target feature map, wherein the importance weight of the k-th target feature map is... The expression is as follows:

[0071]

[0072] In equation (2), Z = W × H, This reflects the importance of the k-th target feature map to the decision-making process of the unsupervised clustering model.

[0073] After obtaining the importance weights corresponding to each target feature map, the multiple target feature maps are weighted and fused according to the importance weights, and the saliency features are extracted from the fused feature map after weighted fusion to obtain the saliency map of each electromagnetic signal after clustering.

[0074] This embodiment can perform linear weighted fusion of multiple target feature maps according to their importance weights to obtain a fused feature map.

[0075] Due to the feature vector of the fused feature map The vector elements include positive and negative vector elements, where positive vector elements are features that positively impact clustering accuracy, and negative vector elements are features that negatively impact clustering accuracy. Based on this, this embodiment extracts the positive vector elements from the feature vector of the fused feature map as the saliency features. For example, the negative vector elements in the feature vector of the fused feature map are set to 0, resulting in a saliency map that is a grayscale image with the same size as the fused feature map.

[0076] The saliency map in this embodiment is also called a class activation map, class heatmap, etc. In some embodiments, the ReLU activation function can be used to apply the electromagnetic signal X. c The corresponding fusion feature map is activated to obtain the saliency map M. c :

[0077]

[0078] As can be seen, the saliency map M obtained using the ReLU activation function c In this study, the strength of salient features that have a positive impact on improving clustering accuracy was enhanced.

[0079] Based on the above embodiments, a saliency map corresponding to each clustered electromagnetic signal can be obtained. Subsequently, the saliency map of each clustered electromagnetic signal can be superimposed onto the waveform of that electromagnetic signal. Specifically, this includes:

[0080] First, obtain the waveform of each clustered electromagnetic signal, and then use tools such as MATLAB to plot the waveform of each clustered electromagnetic signal.

[0081] Next, the saliency map of the electromagnetic signal is converted into a pseudo-color map to obtain the pseudo-color map of the electromagnetic signal.

[0082] The significance obtained in this embodiment Figure 1 Generally, the image is grayscale. Therefore, it is necessary to convert the saliency map into a pseudocolor image. This can be done using the Color() function in MATLAB to obtain the pseudocolor image corresponding to the grayscale saliency map.

[0083] Then, the pseudo-color image is upsampled so that the size of the upsampled pseudo-color image is the same as the size of the waveform image. Upsampling can also be performed using the upsample() function in MATLAB.

[0084] Finally, the upsampled pseudo-color image is superimposed onto the waveform image to obtain the desired waveform image.

[0085] Using the electromagnetic signal X in the aforementioned embodiment c For example, the electromagnetic signal X c The desired waveform is S(X) c )=F(X c )+upsample(Color(M c )), where F(X) c ) is the electromagnetic signal X c The waveform diagram. In the desired waveform diagram F(X) c In the diagram, the red area represents the electromagnetic signal X. c Significant features that are important for clustering.

[0086] like Figure 4As shown, this application uses a joint loss function consisting of a reconstruction loss function and a clustering loss function to train a deep autoencoder, making the target clusters of electromagnetic signals of the same type more compact in the latent feature space, thus improving the clustering effect of the unsupervised clustering model. Then, the K-means clustering method is used to perform unsupervised clustering on the latent feature vectors to obtain the category label of each electromagnetic signal after clustering. The deep autoencoder is used to extract salient features from the clustered electromagnetic signals, visualize the salient features, and establish the causal relationship between the salient features and the clustering results, thereby realizing the visualization of the salient features of the signals on which the unsupervised clustering model is based in the decision-making process.

[0087] This application enables the visualization of salient features extracted by deep autoencoders without modifying the parameters of the unsupervised clustering model, achieving good clustering results. It indicates which part of the input signal features the unsupervised clustering model relies on when performing clustering, thus helping to understand and analyze the working principle and decision-making process of the unsupervised clustering model.

[0088] The electromagnetic signal clustering processing method described in the foregoing embodiments belongs to the same technical concept. This application also provides an electromagnetic signal clustering processing apparatus for implementing the electromagnetic signal clustering processing method described in the foregoing embodiments.

[0089] Figure 5 A schematic diagram of an electromagnetic signal clustering processing apparatus according to an embodiment of this application is shown, as follows: Figure 5 As shown, the electromagnetic signal clustering processing device 500 includes: a signal acquisition unit 510, a clustering processing unit 520, a target feature map acquisition unit 530, a saliency map calculation unit 540, and a visualization processing unit 550.

[0090] The signal acquisition unit 510 is used to acquire a set of electromagnetic signals to be clustered, wherein the set of electromagnetic signals includes multiple electromagnetic signals;

[0091] Clustering processing unit 520 is used to perform clustering processing on multiple electromagnetic signals in the electromagnetic signal set using an unsupervised clustering model to obtain multiple clustered electromagnetic signals. The unsupervised clustering model includes a deep autoencoder and a clustering network. The deep autoencoder includes an encoder and a decoder.

[0092] The target feature map acquisition unit 530 is used to input each clustered electromagnetic signal into the deep autoencoder to obtain multiple target feature maps and joint loss values ​​corresponding to the target hidden layer output of the decoder.

[0093] The saliency map calculation unit 540 is used to obtain the importance weight corresponding to each target feature map according to the joint loss value, perform weighted fusion of the multiple target feature maps according to the importance weight, and extract saliency features from the fused feature map obtained after weighted fusion to obtain the saliency map of each electromagnetic signal after clustering.

[0094] The visualization processing unit 550 is used to overlay the saliency map of each clustered electromagnetic signal onto the waveform map of the electromagnetic signal to obtain a visual representation of the saliency features in each clustered electromagnetic signal.

[0095] In one embodiment of this application, the decoder includes a fully connected layer and a plurality of cascaded basic layers. The fully connected layer receives the latent feature vector output by the encoder, and the basic layers reconstruct the output feature map of the fully connected layer. The basic layers include sequentially connected upsampling layers, activation function layers, batch normalization layers, and deconvolution layers.

[0096] The target feature map acquisition unit 530 is used to obtain multiple target feature maps output by the target hidden layer through multiple feature maps output by the deconvolution layer in the last basic layer of the decoder.

[0097] In one embodiment of this application, each clustered electromagnetic signal includes a category label for each electromagnetic signal. The target feature map acquisition unit 530 is further configured to acquire a joint loss function for training the unsupervised clustering model. The joint loss function includes a reconstruction loss function and a clustering loss function. The feature vectors of multiple target feature maps corresponding to each clustered electromagnetic signal and the signal vector of the electromagnetic signal are input into the reconstruction loss function to obtain a reconstruction loss value. The latent feature vectors of each clustered electromagnetic signal and the category label of the electromagnetic signal are input into the clustering loss function to obtain a clustering loss value. The joint loss value corresponding to the multiple target feature maps is obtained based on the clustering loss value and the reconstruction loss value.

[0098] In one embodiment of this application, the saliency map calculation unit 540 is used to input the joint loss value into the deep autoencoder for backpropagation to obtain multiple gradient feature maps of the target hidden layer output of the decoder, and the multiple gradient feature maps and multiple target features Figure 1 One-to-one correspondence; global average pooling is performed on the multiple gradient feature maps to obtain the importance weight corresponding to each target feature map.

[0099] In one embodiment of this application, the vector elements of the feature vector of the fused feature map include positive vector elements and negative vector elements. The saliency map calculation unit 540 is further configured to extract the positive vector elements in the feature vector of the fused feature map as the saliency features.

[0100] In one embodiment of this application, the visualization processing unit 550 is used to acquire the waveform of each clustered electromagnetic signal; convert the saliency map of the electromagnetic signal into a pseudo-color map to obtain a pseudo-color map of the electromagnetic signal; upsample the pseudo-color map so that the size of the upsampled pseudo-color map is the same as the size of the waveform map; and superimpose the upsampled pseudo-color map onto the waveform map to obtain the desired waveform map.

[0101] In one embodiment of this application, the electromagnetic signal clustering processing apparatus 500 further includes: a preprocessing unit;

[0102] A preprocessing unit is used to construct a training sample set, which includes multiple electromagnetic signal samples. The multiple electromagnetic signal samples from the training sample set are input into the deep autoencoder to obtain the latent feature vector output by the encoder, the multiple electromagnetic signal samples reconstructed by the decoder based on the latent feature vector, and the class label of each electromagnetic signal sample obtained after unsupervised clustering of the latent feature vector by the clustering network. The reconstructed sample vectors of the multiple electromagnetic signal samples, the sample vectors of the multiple electromagnetic signal samples, the class label of each electromagnetic signal sample, and the latent feature vectors of the multiple electromagnetic signal samples are input into the joint loss function to obtain the joint loss value for this round. The parameters of the deep autoencoder are updated based on the joint loss value for this round. The multiple electromagnetic signal samples from the training sample set are input into the parameter-updated deep autoencoder to obtain the joint loss value for the next round. The parameters of the deep autoencoder are updated based on the joint loss value for the next round until the model training stopping condition is met.

[0103] It is understood that the above-described electromagnetic signal clustering processing device can implement each step of the electromagnetic signal clustering processing method provided in the foregoing embodiments. The relevant explanations of the electromagnetic signal clustering processing method are applicable to the electromagnetic signal clustering processing device, and will not be repeated here.

[0104] Figure 6 A schematic diagram of an electronic device according to one embodiment of this application is shown. Please refer to... Figure 6At the hardware level, the electronic device includes a processor and memory, and optionally also includes an internal bus and a network interface. The memory may include main memory, such as high-speed random-access memory (RAM), or non-volatile memory, such as at least one disk drive. Of course, the electronic device may also include other hardware required for other business operations.

[0105] The processor, interface module, communication module, and memory can be interconnected via an internal bus. This internal bus can be an ISA (Industry Standard Architecture) bus, a PCI (Peripheral Component Interconnect) bus, or an EISA (Extended Industry Standard Architecture) bus, etc. The bus can be divided into address bus, data bus, control bus, etc. For ease of representation, Figure 6 The symbol is represented by a single double-headed arrow, but this does not mean that there is only one bus or one type of bus.

[0106] Memory is used to store executable instructions for a computer. Memory provides these instructions to the processor via an internal bus.

[0107] The processor executes computer-executable instructions stored in memory and specifically performs the following operations:

[0108] Obtain a set of electromagnetic signals to be clustered, wherein the set of electromagnetic signals includes multiple electromagnetic signals;

[0109] Multiple electromagnetic signals in the electromagnetic signal set are clustered using an unsupervised clustering model to obtain multiple clustered electromagnetic signals. The unsupervised clustering model includes a deep autoencoder and a clustering network. The deep autoencoder includes an encoder and a decoder.

[0110] Each clustered electromagnetic signal is input into the deep autoencoder to obtain multiple target feature maps and the joint loss value corresponding to the multiple target feature maps output by the target hidden layer of the decoder.

[0111] The importance weights corresponding to each target feature map are obtained based on the joint loss value. The multiple target feature maps are then weighted and fused according to the importance weights. The saliency features of the fused feature map are extracted to obtain the saliency map of each electromagnetic signal after clustering.

[0112] The saliency map of each clustered electromagnetic signal is superimposed on the waveform map of that electromagnetic signal to obtain a visual representation of the saliency features in each clustered electromagnetic signal.

[0113] The above is as stated in this application. Figure 1 The electromagnetic signal clustering processing method disclosed in the illustrated embodiment can perform functions that can be applied to a processor, or implemented by a processor. The processor may be an integrated circuit chip with signal processing capabilities. In implementation, each step of the above method can be completed through integrated logic circuits in the processor's hardware or through software instructions.

[0114] This application also proposes a computer-readable storage medium that stores one or more programs, which, when executed by a processor, perform the following operations:

[0115] Obtain a set of electromagnetic signals to be clustered, wherein the set of electromagnetic signals includes multiple electromagnetic signals;

[0116] Multiple electromagnetic signals in the electromagnetic signal set are clustered using an unsupervised clustering model to obtain multiple clustered electromagnetic signals. The unsupervised clustering model includes a deep autoencoder and a clustering network. The deep autoencoder includes an encoder and a decoder.

[0117] Each clustered electromagnetic signal is input into the deep autoencoder to obtain multiple target feature maps and the joint loss value corresponding to the multiple target feature maps output by the target hidden layer of the decoder.

[0118] The importance weights corresponding to each target feature map are obtained based on the joint loss value. The multiple target feature maps are then weighted and fused according to the importance weights. The saliency features of the fused feature map are extracted to obtain the saliency map of each electromagnetic signal after clustering.

[0119] The saliency map of each clustered electromagnetic signal is superimposed on the waveform map of that electromagnetic signal to obtain a visual representation of the saliency features in each clustered electromagnetic signal.

[0120] Those skilled in the art will understand that embodiments of this application can be provided as methods, systems, or computer program products. Therefore, this application can take the form of a completely hardware embodiment, a completely software embodiment, or an embodiment combining software and hardware aspects. Furthermore, this application can take the form of a computer program product embodied 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.

[0121] These computer program instructions may also be stored in a computer-readable storage medium that can direct a computer or other programmable data processing device to function in a particular manner, such that the instructions stored in the computer-readable storage medium produce an article of manufacture including instruction means, which are implemented in a process Figure 1 One or more processes and / or boxes Figure 1 The function specified in one or more boxes.

[0122] These computer program instructions may also be loaded onto a computer or other programmable data processing apparatus to cause a series of operational steps to be performed on the computer or other programmable apparatus to produce a computer-implemented process, thereby providing instructions that execute on the computer or other programmable apparatus for implementing a flow chart of one or more flows and / or blocks. Figure 1 The steps that specify the function are in one or more boxes.

[0123] In a typical configuration, a computing device includes one or more processors (CPU), input / output interfaces, network interfaces, and memory.

[0124] Memory may include non-persistent storage in computer-readable media, such as random access memory (RAM) and / or non-volatile memory, such as read-only memory (ROM) or flash RAM. Memory is an example of computer-readable media.

[0125] Computer-readable media includes both permanent and non-permanent, removable and non-removable media, and information storage can be achieved by any method or technology. Information can be computer-readable instructions, data structures, program modules, or other data. As defined herein, computer-readable media does not include transient media, such as modulated data signals and carrier waves.

[0126] It should also be noted that the terms "comprising," "including," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or apparatus that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such process, method, article, or apparatus. Unless otherwise specified, an element defined by the phrase "comprising one..." does not exclude the presence of other identical elements in the process, method, article, or apparatus that includes that element.

[0127] It should be understood that although the terms first, second, third, etc., may be used in this invention to describe various information, this information should not be limited to these terms. These terms are only used to distinguish information of the same type from one another. For example, without departing from the scope of this invention, first information may also be referred to as second information, and similarly, second information may also be referred to as first information.

[0128] The above are merely embodiments of this application and are not intended to limit the scope of this application. Various modifications and variations can be made to this application by those skilled in the art. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of this application should be included within the scope of the claims of this application.

Claims

1. A method for clustering electromagnetic signals, characterized in that, The method includes: Obtain a set of electromagnetic signals to be clustered, wherein the set of electromagnetic signals includes multiple electromagnetic signals; Multiple electromagnetic signals in the electromagnetic signal set are clustered using an unsupervised clustering model to obtain multiple clustered electromagnetic signals. The unsupervised clustering model includes a deep autoencoder and a clustering network. The deep autoencoder includes an encoder and a decoder. Each clustered electromagnetic signal is input into the deep autoencoder to obtain multiple target feature maps and the joint loss value corresponding to the multiple target feature maps output by the target hidden layer of the decoder. The importance weights corresponding to each target feature map are obtained based on the joint loss value. The multiple target feature maps are then weighted and fused according to the importance weights. The saliency features of the fused feature map are extracted to obtain the saliency map of each electromagnetic signal after clustering. The saliency map of each clustered electromagnetic signal is superimposed on the waveform map of the electromagnetic signal to obtain a visual representation of the saliency features in each clustered electromagnetic signal. The step of obtaining the importance weight corresponding to each target feature map based on the joint loss value includes: The joint loss value is input into the deep autoencoder for backpropagation to obtain multiple gradient feature maps of the target hidden layer output of the decoder. The multiple gradient feature maps correspond one-to-one with the multiple target feature maps. Global average pooling is performed on the multiple gradient feature maps to obtain the importance weights corresponding to each target feature map; The saliency feature is a feature that has a positive impact on improving clustering accuracy, and the importance weight is used to increase the strength of the feature that has a positive impact on improving clustering accuracy.

2. The method according to claim 1, characterized in that, The decoder includes a fully connected layer and multiple cascaded basic layers. The fully connected layer receives the latent feature vector output by the encoder. The basic layers reconstruct the output feature map of the fully connected layer. The basic layers include sequentially connected upsampling layers, activation function layers, batch normalization layers, and deconvolution layers. The step of inputting each clustered electromagnetic signal into the deep autoencoder to obtain multiple target feature maps output by the target hidden layer of the decoder includes: Multiple target feature maps output by the target hidden layer are obtained from multiple feature maps output by the deconvolution layer in the last basic layer of the decoder.

3. The method according to claim 2, characterized in that, Each clustered electromagnetic signal includes a category label for that electromagnetic signal. The joint loss value corresponding to the multiple target feature maps is obtained through the following steps: Obtain a joint loss function for training the unsupervised clustering model, the joint loss function including a reconstruction loss function and a clustering loss function; The feature vectors of multiple target feature maps corresponding to each clustered electromagnetic signal and the signal vector of the electromagnetic signal are input into the reconstruction loss function to obtain the reconstruction loss value; The latent feature vector corresponding to each clustered electromagnetic signal and the category label of the electromagnetic signal are input into the clustering loss function to obtain the clustering loss value; The joint loss value corresponding to the multiple target feature maps is obtained based on the clustering loss value and the reconstruction loss value.

4. The method according to claim 1, characterized in that, The feature vector of the fused feature map includes positive and negative vector elements. The extraction of saliency features from the fused feature map obtained after weighted fusion includes: The positive vector elements in the feature vector of the fused feature map are extracted as the salient features.

5. The method according to claim 1, characterized in that, The step of overlaying the saliency map of each clustered electromagnetic signal onto the waveform map of that electromagnetic signal to obtain a visual representation of the saliency features in each clustered electromagnetic signal includes: Obtain the waveform of each electromagnetic signal after clustering; The saliency map of the electromagnetic signal is converted into a pseudo-color map to obtain the pseudo-color map of the electromagnetic signal. The pseudo-color image is upsampled so that the size of the upsampled pseudo-color image is the same as the size of the waveform image. The upsampled pseudo-color image is superimposed onto the waveform image to obtain the desired waveform image.

6. The method according to claim 3, characterized in that, The unsupervised clustering model is trained using the following steps: Construct a training sample set, which includes multiple electromagnetic signal samples; Multiple electromagnetic signal samples from the training sample set are input into the deep autoencoder to obtain the latent feature vector output by the encoder, the multiple electromagnetic signal samples reconstructed by the decoder based on the latent feature vector, and the category label of each electromagnetic signal sample obtained by the clustering network after performing unsupervised clustering on the latent feature vector. The reconstructed sample vectors of the multiple electromagnetic signal samples, the sample vectors of the multiple electromagnetic signal samples, the category label of each electromagnetic signal sample, and the latent feature vectors of the multiple electromagnetic signal samples are input into the joint loss function to obtain the joint loss value for this round. The parameters of the deep autoencoder are updated based on the joint loss value of the current round. Multiple electromagnetic signal samples from the training sample set are input into the deep autoencoder after parameter update to obtain the joint loss value of the next round. The parameters of the deep autoencoder are updated based on the joint loss value of the next round until the model training stopping condition is met.

7. An electromagnetic signal clustering processing device, characterized in that, The device includes: A signal acquisition unit is used to acquire a set of electromagnetic signals to be clustered, wherein the set of electromagnetic signals includes multiple electromagnetic signals; A clustering processing unit is used to perform clustering processing on multiple electromagnetic signals in the electromagnetic signal set using an unsupervised clustering model to obtain multiple clustered electromagnetic signals. The unsupervised clustering model includes a deep autoencoder and a clustering network. The deep autoencoder includes an encoder and a decoder. The target feature map acquisition unit is used to input each clustered electromagnetic signal into the deep autoencoder to obtain multiple target feature maps and the joint loss value corresponding to the multiple target feature maps output by the target hidden layer of the decoder. The saliency map calculation unit is used to obtain the importance weight corresponding to each target feature map according to the joint loss value, perform weighted fusion of the multiple target feature maps according to the importance weight, and extract saliency features from the fused feature map obtained after weighted fusion to obtain the saliency map of each electromagnetic signal after clustering. A visualization processing unit is used to overlay the saliency map of each clustered electromagnetic signal onto the waveform map of the electromagnetic signal to obtain a visual representation of the saliency features in each clustered electromagnetic signal. The saliency map calculation unit is used for: The joint loss value is input into the deep autoencoder for backpropagation to obtain multiple gradient feature maps of the target hidden layer output of the decoder. The multiple gradient feature maps correspond one-to-one with the multiple target feature maps. Global average pooling is performed on the multiple gradient feature maps to obtain the importance weights corresponding to each target feature map; The saliency feature is a feature that has a positive impact on improving clustering accuracy, and the importance weight is used to increase the strength of the feature that has a positive impact on improving clustering accuracy.

8. An electronic device, characterized in that, include: processor; as well as A memory configured to store computer-executable instructions, which, when executed, cause the processor to perform the method of any one of claims 1-6.

9. A computer-readable storage medium storing one or more programs that, when executed by a processor, implement the method of any one of claims 1-6.

Images