Weight learning method for signal fusion recognition

Abstract

This invention discloses a weighted learning method for signal fusion recognition, belonging to the field of signal recognition technology. This invention fuses the recognition results of BP neural networks, three-channel convolutional neural networks, and four-channel convolutional neural networks using a weighted learning convolutional neural network (WLCNN) model. The WLCNN model automatically learns feature merging coefficients, merges feature data according to these coefficients, and outputs the signal type recognition result. This invention achieves higher recognition accuracy and maintains high recognition performance even at low noise-to-interference ratios. Furthermore, the proposed WLCNN network is not limited to specific input data types and is applicable to the recognition of modulated and interference signals in multi-antenna scenarios, exhibiting good versatility.

Description

Technical Field

[0001] This invention belongs to the field of signal recognition technology, specifically relating to a weight learning method for signal fusion recognition. Background Technology

[0002] As signal complexity increases, the recognition capabilities of single-node signal recognition algorithms become limited, leading to growing interest in distributed signal recognition methods based on multi-node collaboration. A common processing structure based on information fusion reveals that centralized structures primarily involve the direct transmission of information acquired by individual nodes to a central node for processing. While this structure ensures data integrity, it places high demands on the transmission system's capacity, and the central node's processing load is excessive due to the need for all data processing to be completed at the central node. Distributed structures, on the other hand, first process and detect the data acquired by each sensor at its own node, then send the results to the central node for a comprehensive evaluation of the processing decisions made by each node. Distributed structures overcome the shortcomings of centralized structures, significantly reducing the amount of data processed by the central node, while also offering higher system stability.

[0003] Distributed multi-node signal recognition algorithms based on information fusion can be mainly divided into three-layer fusion models from the perspective of spatial information fusion: data layer, feature layer, and decision layer. Figure 1 As shown in the diagram. Data layer fusion fuses the raw signals acquired by each node, enhancing signal strength and significantly improving performance. However, the large amount of data transmitted between nodes makes it unsuitable for practical wireless applications. Furthermore, unavoidable parameter differences exist between the received signals, and compensating for these differences places a significant computational burden on the central node. Feature layer fusion extracts features containing signal information, preserving more signal information. The parameter differences between signals may become less sensitive after the signal transformation to the feature layer, and the signal features compress a larger amount of data compared to the signal itself, making it more suitable for wireless transmission. Decision layer fusion utilizes decision values ​​extracted by different nodes, improving performance through well-designed and effective collaborative strategies. Its advantages include significantly reducing the amount of data transmitted and eliminating the need to consider parameter differences between signals. Its disadvantages include the loss of considerable information in the decision results compared to the signal itself, limited performance improvement, and poor adaptability to different environments. Summary of the Invention

[0004] To overcome the shortcomings of traditional methods in signal recognition processing, this invention proposes a weight learning method for signal fusion recognition based on a weight-learning convolutional neural network (WLCNN). This invention uses the WLCNN network to perform fusion recognition processing on signals, thereby effectively improving the accuracy of signal recognition.

[0005] The technical solution adopted in this invention is as follows:

[0006] A weight learning method for signal fusion recognition, comprising the following steps:

[0007] A predetermined signal recognition algorithm is used to perform preliminary signal recognition processing on the signal sequence to be recognized acquired by each type of sensing node, and a preliminary recognition result corresponding to the current type of sensing node is obtained. This preliminary recognition result is the classification probability of the predetermined signal category of the signal sequence to be recognized; and the number of signal categories in each path is consistent.

[0008] A weighted learning convolutional neural network (WLCNN) model is constructed, comprising a first convolutional neural network (CNN1), a weighted learning module (WLM), and a second convolutional neural network (CNN2). The preliminary recognition results from each path are input into a WLCNN model. The first convolutional neural network of the WLCNN model extracts features from the input preliminary recognition results to obtain first feature information. The weighted learning module of the WLCNN model performs weight learning on the input preliminary recognition results, outputting feature merging coefficients for each signal in the signal sequence to be recognized. These feature merging coefficients are normalized and then fused with the first feature information by coefficient weighting to obtain second feature information for each preliminary recognition result. All second feature information is fused and the fused features are input into the second convolutional neural network of the WLCNN model. The second convolutional neural network is used to predict the final signal fusion recognition result.

[0009] Each preliminary identification result corresponds to a WLCNN model, and the weight learning modules of each WLCNN model share the network structure and network parameters.

[0010] The WLCNN model is trained using pre-set training data. When the pre-set training termination condition is met, a fusion recognition model for target signal category identification is obtained based on the trained WLCNN model. Then, the weight parameters (i.e., feature merging coefficients) for signal fusion recognition can be obtained based on the weight learning module in the fusion recognition model.

[0011] Furthermore, the structure of the first convolutional neural network in the WLCNN model includes, in sequence, a convolutional block, a max pooling layer, and several residual modules. The convolutional block includes, in sequence, a convolutional layer, a batch normalization layer, and an activation function layer. Each residual module includes two convolutional layers, two batch normalization layers, and two activation function layers. The input of the residual block is sequentially processed by convolution and batch normalization, activation function, convolution and batch normalization, and then added to the input of the residual block. The result of the addition is processed by the activation function to obtain the output of the residual block.

[0012] Furthermore, the first convolutional neural network of the WLCNN model includes 3 residual modules, and the kernel size of the convolutional layer in the convolutional block is 3×3 with a convolution stride of 2; the kernel size of the max pooling layer is 3×3 with a pooling stride of 1, and the kernel size of the convolutional layer in the residual module is 3×3 with a convolution stride of 1.

[0013] Furthermore, the structure of the second convolutional neural network in the WLCNN model includes, in sequence: several residual blocks, an average pooling layer, and a fully connected layer with a Softmax function, wherein the residual blocks are composed of several stacked residual modules.

[0014] Furthermore, the second convolutional neural network of the WLCNN model includes three residual blocks. The first residual block includes four stacked residual modules, the second residual block includes six stacked residual modules, and the third residual block includes three stacked residual modules. The convolutional kernel size of the convolutional layers in each residual block is 3×3, and the convolution stride is 2. The average pooling layer has a pooling kernel size of 3×3 and a pooling stride of 1.

[0015] Furthermore, the network structure of the weight learning module of the WLCNN model includes, in sequence: convolutional blocks, several residual modules, and fully connected layers.

[0016] Furthermore, the convolutional kernel size of the convolutional blocks in the weight learning module of the WLCNN model is 3×3, the convolutional stride is 2, and it includes 3 residual modules, in which the convolutional layer of each residual module has a convolutional kernel size of 3×3 and a convolutional stride of 1.

[0017] Furthermore, the signal recognition algorithms include: signal recognition algorithms based on BP (Back Propagation) neural networks, signal recognition algorithms based on three-channel convolutional neural networks, and signal recognition algorithms based on four-channel convolutional neural networks.

[0018] The technical solution provided by this invention brings at least the following beneficial effects:

[0019] This invention fuses the recognition results of different neural networks (such as BP neural networks, three-channel convolutional neural networks, and four-channel convolutional neural networks) through a weighted learning convolutional neural network (WLCNN) model. The WLCNN model automatically learns feature merging coefficients, merges feature data according to the coefficients, and outputs the signal type recognition result. This invention has higher recognition accuracy and still maintains high recognition performance at low noise-to-interference ratios. At the same time, the WLCNN network proposed in this invention is not limited to the input data type and can be applied to the recognition of modulated signals and interference signals in multi-antenna scenarios, thus having good versatility. Attached Figure Description

[0020] To more clearly illustrate the technical solutions in the embodiments of the present invention, the accompanying drawings used in the description of the embodiments will be briefly introduced below. Obviously, the accompanying 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.

[0021] Figure 1 This is a multi-node distributed fusion diagram;

[0022] Figure 2 This is a schematic diagram of a multi-node signal recognition model;

[0023] Figure 3 This is a schematic diagram of the WLCNN signal recognition and fusion network architecture;

[0024] Figure 4 This is a schematic diagram of the Residual Block structure;

[0025] Figure 5 This is a comparison chart of the overall signal recognition accuracy under different algorithms.

[0026] Figure 6 A comparison chart showing the recognition accuracy of different interference signals under different algorithms;

[0027] in,

[0028] Figure 6-a shows a comparison of the recognition accuracy of different algorithms for STJ (single tone jamming) interference.

[0029] Figure 6-b shows a comparison of the recognition accuracy of different algorithms for MTJ (multi-tone jamming) interference.

[0030] Figure 6-c shows a comparison of the recognition accuracy of different algorithms for LFMJ (linear frequency modulation jamming) linear frequency sweeping interference;

[0031] Figure 6-d shows a comparison of the recognition accuracy of different algorithms for NFMJ (noise frequency modulation jamming) noise interference;

[0032] Figure 6-e shows a comparison of the accuracy of different algorithms in identifying partial bandwidth noise jamming (PBNJ).

[0033] Figure 6-f shows a comparison of the recognition accuracy of different algorithms for MSK (Minimum Shift Keying);

[0034] Figure 6-g is a comparison of the recognition accuracy of different algorithms for PDradar (pulse doppler radar) signals;

[0035] Figure 6-h shows a comparison of the recognition accuracy of different algorithms for TGPJ (time gauss pulse jamming) time-domain Gaussian impulse interference. Detailed Implementation

[0036] To make the objectives, technical solutions, and advantages of the present invention clearer, the embodiments of the present invention will be described in further detail below with reference to the accompanying drawings.

[0037] To overcome the limitations of traditional methods for signal recognition fusion, this invention utilizes a convolutional neural network to automatically learn variable feature merging coefficients for intermediate layer feature merging. End-to-end model training is performed based on the classification cross-entropy loss function. This allows for the comprehensive processing of recognition information acquired from multiple heterogeneous sensing nodes before signal recognition. This approach offers high fault tolerance and achieves superior performance compared to a single algorithm (signal recognition algorithm, i.e., a neural network-based recognition model), thereby improving signal recognition effectiveness.

[0038] In this embodiment of the invention, it is assumed that the application scenario contains multiple types of signals, including communication signals, radar signals, and various interference signals. After preprocessing the signal features, a distributed multi-node identification method is used for signal type identification. This method requires each node to process the signal features, and then the processing results are sent to the fusion center node for comprehensive evaluation of the processing decisions of each node. Different nodes refer to different signal identification algorithms, such as... Figure 2The BP neural network, three-channel convolutional neural network, and four-channel convolutional neural network shown in the diagram exhibit different signal recognition capabilities and yield varying results when used individually. Finally, a WLCNN model (including CNN1 and CNN2 networks) is employed to fuse the outputs of the three signal recognition algorithms, achieving more accurate recognition. In other words, the multi-node signal recognition algorithm used in this embodiment belongs to the decision-level fusion model, which integrates the preliminary recognition results from the three algorithms. The multi-node signal recognition model is as follows: Figure 2 As shown, the input data used in the fusion part is the probability of each signal category. First, the feature is extracted by the CNN1 network, and the signal feature tensor obtained by each algorithm is output. Then, the feature is fused by the CNN2 network to obtain the final signal recognition result.

[0039] When performing multi-node signal recognition, the feature combination coefficients of different algorithms may be different, and the merging coefficients need to be adapted to different inputs. Therefore, the weight-learning convolutional neural network learns the feature merging coefficients of the three algorithms through the weight-learning module (WLM) and merges the feature tensors according to the coefficients to obtain the signal recognition result.

[0040] As one possible implementation, the architecture of a Weight Learning Convolutional Neural Network (WLCNN) model for multi-node signal recognition fusion is as follows: Figure 3 As shown, the signals to be identified acquired by each type of sensing node are first preprocessed (based on a preset signal recognition algorithm, preliminary recognition is performed, and preliminary recognition results are output, i.e., the classification probability of each signal category). The preliminary recognition results are then input into the WLCNN model. For each preliminary recognition result (the preliminary recognition branch corresponding to each signal recognition algorithm), the CNN1 network is input for feature extraction to obtain the first feature information of each preliminary recognition result. At the same time, the first feature information is input into the weight learning module (WLM) for weight learning. The weight coefficients output by the WLM module are then normalized and fused with the feature information extracted by the CNN1 network (feature information is accumulated according to the coefficients) to obtain the second feature information of each preliminary recognition result. Then, the second feature information of all preliminary recognition results is fused and input into the CNN2 network. The final recognition result, i.e., the fused recognition result, is obtained based on the output of the CNN2 network.

[0041] As one possible implementation, in this embodiment of the invention, the WLCNN model is a 34-layer ResNet network, consisting of one convolutional layer, 16 residual blocks, and one fully connected layer, as shown in Table 1. Since the probability values ​​of the input signal range from (0,1), ReLU is used as the activation function to reduce gradient explosion and vanishing effects. The SoftMax function is used as the activation function of the last layer to normalize the output probability distribution of the network.

[0042] Table 1 ResNet Structure

[0043]

[0044] The structure of each residual block is as follows: Figure 4 As shown, it consists of two convolutional layers, batch normalization, and an activation function. The input of the residual block is sequentially processed by convolution and batch normalization, activation function, convolution and batch normalization, and then added to the input of the residual block. The result of the addition is then processed by the activation function to obtain the output of the residual block.

[0045] In this embodiment of the invention, taking three types of heterogeneous sensing nodes as an example, the signal recognition information output by them (in this example, the probability of the recognized category) is used as the input data of the WLCNN model. After passing through CNN1, their respective feature tensors are obtained. The features are merged through the combination coefficients learned by WLM. The merged features are further input into CNN2 for the feature classification process, and the probability prediction distribution of each signal category is output.

[0046] Given the probability sample values ​​of each signal recognition algorithm, the nth output of CNN1... r Feature tensor of each sample Merging feature F c It is a tensor that is added by coefficients, represented as:

[0047]

[0048] in, It is the nth r Normalized feature weighting coefficients for N samples, r This indicates the number of samples (i.e., the number of samples fused during signal fusion recognition).

[0049] From the nth r Sample values The parameter is θ w The single value output by WLM represents the nth value. r The feature pooling coefficients estimated for each sample can be expressed as:

[0050]

[0051] Where g(·) represents the output of WLM.

[0052] The feature merging weights estimated by WLM can vary according to the input samples. The WLM-estimated feature merging weight coefficients are further normalized using the SoftMax function as follows:

[0053]

[0054] Considering the number of parameters and computational cost of WLM, and to obtain a better network structure, the coefficient learning modules for each path share the network structure and parameters. Through experimental testing, the following WLM network structure was designed: WLM consists of one convolutional layer, three residual blocks, and one fully connected layer. All convolutions use 16 convolutional kernels. The specific structure is shown in Table 2.

[0055] Table 2 WLM Network Architecture

[0056]

[0057] To further verify the recognition performance of the method in this embodiment, simulation experiments were conducted. This embodiment requires the identification of eight different types of signals, including: MSK modulated communication signals (MSK), pulse Doppler radar signals (PDradar), single-tone interference (STJ), multi-tone interference (MTJ), partial-band noise interference (PBNJ), sweep frequency interference (LFMJ), noise-frequency modulation interference (NFMJ), and time-domain Gaussian impulse interference (TGPJ). The interference-to-noise ratio (JNR) of the signals ranges from -20dB to 20dB, with values ​​taken in 2dB increments, resulting in 21 JNR types. Each JNR generates 1200 samples. The ratio of the training set to the test set is set to 3:1. Three types of heterogeneous sensing nodes are used: a BP neural network, a three-channel convolutional neural network, and a four-channel neural network. They output the recognition probabilities of each type of signal, which are then input as feature data into the WLCNN model of this embodiment.

[0058] During the training of the WLCNN model, the Adam optimizer was used with an initial learning rate of 0.0005 and a batch size of 256. Training lasted for 20 epochs, with the learning rate adjusted every 5 epochs to accelerate the training of the multi-antenna model. Cross-entropy loss was used as the training loss function. Similarly, all JNR data were used for co-training, and accuracy was tested separately on each JNR.

[0059] To test the performance of the WLCNN model in signal recognition, the recognition results of the BP neural network, the three-channel convolutional neural network, and the four-channel convolutional neural network were compared with the recognition results of the fusion using the WLCNN model. Figure 5 The overall recognition accuracy curves of four algorithms under different JNRs are presented, where CNN1 represents a three-channel convolutional neural network and CNN2 represents a four-channel convolutional neural network. Figure 5 It can be seen that, compared with traditional BP neural networks, three-channel convolutional neural networks, and four-channel convolutional neural networks, the method of fusing recognition results using the WLCNN model in this embodiment achieves a higher recognition accuracy. At a JNR of -15dB, the overall recognition accuracy of the WLCNN model reaches 45%, while the overall accuracy of the other three algorithms is no higher than 35%. The WLCNN model can improve the overall accuracy by 10%, indicating that the WLCNN model has higher accuracy than the other three algorithms at low JNR. At a JNR of -18dB, the WLCNN model achieves the same accuracy as the three-channel and four-channel convolutional neural networks (-16dB) and the BP neural network (-14dB), indicating that the WLCNN model can improve the noise margin of signal recognition by approximately 2-4dB. At a JNR of -5dB, the overall recognition rate of the WLCNN model is close to 100% and remains stable, demonstrating that the multi-node signal recognition fusion method has higher performance.

[0060] To further analyze the recognition of different types of signals, Figure 6 The recognition accuracy curves for various signal types under four algorithms are presented. Figure 6It can be seen that the WLCNN model outperforms the BP neural network, three-channel convolutional neural network, and four-channel convolutional neural network in recognizing STJ, MTJ, NFMJ, and MSK signals. For STJ, the accuracy reaches 100% when the JNR is -12dB, showing a significant performance improvement. At -14dB, the accuracy of the four signal types fused using the WLCNN algorithm is 20% higher than other single algorithms. When the JNR is less than -14dB, the WLCNN model in this embodiment has lower recognition accuracy for LFMJ and PDradar than the BP neural network. These two signal types have digital features that distinguish them from other signals when recognized by the BP neural network, resulting in better recognition performance than the three-channel and four-channel convolutional neural networks. However, its performance is poor when the JNR is greater than -14dB. This is because the performance of the BP neural network is unstable, while the performance of the WLCNN fusion is more stable with changes in JNR, leading to a performance lower than that of the BP neural network. The network optimizes based on the overall loss. When the recognition performance of the three algorithms differs significantly, it has a negative effect, hindering fusion and potentially causing the performance of certain signal types to be lower than the original signal, as seen in the recognition of PBNJ and TGPJ. From an overall perspective, this embodiment uses the WLCNN model to fuse the outputs of the three algorithms, resulting in a smoother accuracy curve, more stable recognition performance, and greater fault tolerance, thereby improving signal recognition performance.

[0061] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention, and not to limit them; although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some of the technical features; and these modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the spirit and scope of the technical solutions of the embodiments of the present invention.

[0062] The above descriptions are merely some embodiments of the present invention. Those skilled in the art can make various modifications and improvements without departing from the inventive concept of the present invention, and these all fall within the scope of protection of the present invention.

Claims

1. A weight learning method for signal fusion recognition, characterized in that, Includes the following steps: A predetermined signal recognition algorithm is used to perform preliminary signal recognition processing on the signal sequence to be recognized acquired by each type of sensing node, and a preliminary recognition result corresponding to the current type of sensing node is obtained. This preliminary recognition result is the classification probability of the predetermined signal category of the signal sequence to be recognized; and the number of signal categories in each path is consistent. A weighted learning convolutional neural network (WLCNN) model is constructed. The WLCNN model includes a first convolutional neural network, a weight learning module, and a second convolutional neural network. The preliminary recognition results of each path are input into a weighted learning convolutional neural network (WLCNN) model. The first convolutional neural network of the WLCNN model extracts features from the input preliminary recognition results to obtain the first feature information. The WLCNN model's weight learning module performs weight learning on the input preliminary recognition results, outputs the feature merging coefficients of each signal in the signal sequence to be recognized, normalizes the feature merging coefficients, and then performs coefficient-weighted fusion with the first feature information to obtain the second feature information of each preliminary recognition result. All second feature information is fused and the fused features are input into the second convolutional neural network of the WLCNN model. The second convolutional neural network is used to predict the final signal fusion recognition result. Each preliminary identification result corresponds to a WLCNN model, and the weight learning modules of each WLCNN model share the network structure and network parameters. The WLCNN model is trained using deep learning based on pre-set training data. When the pre-set training termination condition is met, a fusion recognition model for target signal category recognition is obtained based on the trained WLCNN model. The structure of the first convolutional neural network in the WLCNN model includes, in sequence: convolutional blocks, max pooling layers, and several residual modules; The convolutional block consists of, in sequence: a convolutional layer, a batch normalization layer, and an activation function layer; Each residual module includes two convolutional layers, two batch normalization layers, and two activation function layers. The input of the residual module is sequentially processed by convolution and batch normalization, activation function, convolution and batch normalization, and then added to the input of the residual module. The result of the addition is processed by the activation function to obtain the output of the residual module. The second convolutional neural network structure of the WLCNN model includes, in sequence, several residual blocks, an average pooling layer, and a fully connected layer with a softmax function. Each residual block is composed of stacked residual modules. Each residual module includes two convolutional layers, two batch normalization layers, and two activation function layers. The input to each residual module is sequentially processed by convolution and batch normalization, followed by activation functions, and then added to the input of the residual module. The result of this addition is then processed by activation functions to obtain the output of the residual module. The network structure of the weight learning module of the WLCNN model includes, in sequence: convolutional blocks, several residual modules, and fully connected layers; The convolutional block consists of, in sequence: a convolutional layer, a batch normalization layer, and an activation function layer; The residual module consists of two convolutional layers, two batch normalization layers, and two activation function layers. The input of the residual module is sequentially processed by convolution and batch normalization, activation function, convolution and batch normalization, and then added to the input of the residual module. The result of the addition is then processed by the activation function to obtain the output of the residual module.

2. The method as described in claim 1, characterized in that, The first convolutional neural network of the WLCNN model includes three residual modules. The kernel size of the convolutional layer in the convolutional block is 3×3, and the stride is 2. The kernel size of the max pooling layer is 3×3, and the stride is 1. The kernel size of the convolutional layer in the residual module is 3×3, and the stride is 1.

3. The method as described in claim 1, characterized in that, The second convolutional neural network of the WLCNN model includes three residual blocks. The first residual block includes four stacked residual modules, the second residual block includes six stacked residual modules, and the third residual block includes three stacked residual modules. The convolutional kernel size of the convolutional layers in each residual block is 3×3, and the convolution stride is 2. The average pooling layer has a pooling kernel size of 3×3 and a pooling stride of 1.

4. The method as described in claim 1, characterized in that, The weight learning module of the WLCNN model has a 3×3 kernel size and a stride of 2 for all convolutional blocks. It also includes 3 residual modules, where the kernel size of the convolutional layers in each residual module is 3×3 and the stride is 1.

5. The method according to any one of claims 1 to 4, characterized in that, Signal recognition algorithms include: signal recognition algorithms based on BP neural networks, signal recognition algorithms based on three-channel convolutional neural networks, and signal recognition algorithms based on four-channel convolutional neural networks.

Images