Signal modulation identification system and method based on convolutional neural network

Abstract

The invention relates to the technical field of communication signal processing, in particular to a signal modulation recognition system and method based on a convolutional neural network, and the method comprises the steps: analyzing I-Q signals through a lightweight CNN model and a GoogLeNet model, and obtaining first prediction probability vectors of all types of modulation signals and second prediction probability vectors of MPSK and MQAM type modulation signals; if the type of the modulation signal of which the probability value is greater than or equal to the preset probability value in the first prediction probability vector is consistent with the type of the modulation signal in the second prediction probability vector, performing weighted fusion on the two prediction probability vectors to obtain a final prediction probability vector; otherwise, taking the first prediction probability vector as a final prediction probability vector; and obtaining the type of the modulation signal with the maximum probability in the final prediction probability vector. According to the scheme, full-type coverage and core precision can be considered, so that the coverage range is guaranteed, and the recognition precision is improved.

Description

Technical Field

[0001] This invention relates to the field of communication signal processing technology, specifically to a signal modulation recognition system and method based on convolutional neural networks. Background Technology

[0002] As wireless communication technology evolves towards higher speeds and higher bandwidths, modulation methods are expanding from traditional BPSK and QPSK to higher-order 16QAM, 32QAM, and 64QAM. Simultaneously, the complexity of the channel environment (such as Ricean multipath fading and Gaussian white noise) is increasing, placing higher demands on the coverage, accuracy, and robustness of modulation recognition. For civilian applications: Requires coverage of more than 12 modulation signals, single-frame recognition latency ≤ 5ms, and recognition rate ≥ 95% at medium-to-high signal-to-noise ratios (SNR = 12-18dB). Special scenarios: It is necessary to maintain a core MPSK / MQAM signal recognition rate of ≥90% under low signal-to-noise ratio (SNR=6-10dB) and be able to distinguish easily confused signals (such as QPSK and 16QAM, 32QAM and 64QAM).

[0003] The development of related technologies can be divided into three generations: (1) Traditional statistical feature method: Based on high-order cumulants and instantaneous amplitude / phase feature extraction, it relies on manual design and has poor robustness at low SNR (recognition rate ≤70% when SNR=6dB). (2) Single deep learning method: Using models such as CNN and GoogLeNet can improve accuracy, but there is an imbalance between full coverage and core accuracy. For example, a single CNN covers 12 types of signals but the core MPSK / MQAM recognition rate is ≤90%, while a single GoogLeNet has a core recognition rate of ≥97% but only covers 6 types of signals. (3) Lightweight and transfer learning combination method: The complexity is reduced by pruning network parameters and reusing pre-trained weights, but a dual-channel collaborative fusion architecture has not yet been formed, making it difficult to meet the needs of multiple scenarios.

[0004] In summary, existing technologies suffer from the problem of balancing full-type coverage with core accuracy. Therefore, there is an urgent need for a signal modulation recognition system and method based on convolutional neural networks that can balance full-type coverage with core accuracy, ensuring coverage range and improving recognition accuracy. Summary of the Invention

[0005] One of the objectives of this invention is to provide a signal modulation recognition system based on a convolutional neural network that can balance full-type coverage with core accuracy, thereby ensuring coverage and improving recognition accuracy.

[0006] The basic solution provided by this invention is a signal modulation recognition system based on convolutional neural networks, comprising: a signal receiving module, a lightweight CNN channel module, a GoogLeNet channel module, and a decision fusion module; The signal receiving module is used to receive modulated signals and convert them into IQ signals; The lightweight CNN channel module is used to analyze IQ signals using a constructed lightweight CNN model to obtain the first predicted probability vector for all types of modulated signals. The GoogLeNet channel module is used to analyze IQ signals using the constructed GoogLeNet model to obtain the second prediction probability vectors of MPSK and MQAM type modulated signals. The decision fusion module is used to determine whether the type of the modulation signal with a probability value greater than or equal to a preset probability value in the first prediction probability vector is consistent with the type of the modulation signal in the second prediction probability vector. If so, the first prediction probability vector and the second prediction probability vector are weighted and fused to obtain the final prediction probability vector; otherwise, the first prediction probability vector is used as the final prediction probability vector. It is also used to obtain the type of the modulation signal with the highest probability in the final predicted probability vector.

[0007] The second objective of this invention is to provide a signal modulation recognition method based on convolutional neural networks, which can balance full-type coverage and core accuracy to ensure coverage and improve recognition accuracy.

[0008] This invention provides a second basic solution: a signal modulation recognition method based on convolutional neural networks, comprising: Acquire the modulated signal and convert it into an IQ signal; The IQ signal is analyzed by constructing a lightweight CNN model to obtain the first predicted probability vector of all types of modulated signals; The IQ signal was analyzed using the constructed GoogLeNet model to obtain the second prediction probability vectors of MPSK and MQAM type modulated signals; Determine whether the type of the modulation signal with a probability value greater than or equal to a preset probability value in the first prediction probability vector is consistent with the type of the modulation signal in the second prediction probability vector. If yes, the first prediction probability vector and the second prediction probability vector are weighted and fused to obtain the final prediction probability vector. If no, the first prediction probability vector is used as the final prediction probability vector. Obtain the type of the modulation signal with the highest probability in the final prediction probability vector.

[0009] Beneficial effects: This solution constructs a dual-channel architecture by lightweighting the CNN channel module and the GoogLeNet channel module, which retains the recognition accuracy of the CNN network. At the same time, since GoogLeNet introduces multiple convolutional filter sizes in the same layer, it improves the representation ability while maintaining computational efficiency, resulting in a stronger recognition rate in the face of large dynamic SNR environments. The first channel (lightweight CNN channel module) acquires the first prediction probability vector of all types of modulated signals (6 core MPSK / MQAM + 6 signals without IQ components) for full coverage screening. The second channel (GoogLeNet channel module) focuses on high-precision identification of core signals and acquires the second prediction probability vector of MPSK and MQAM type modulated signals. The decision fusion module performs weighted fusion based on the consistency between the types of modulation signals with probability values ​​greater than or equal to preset probability values ​​in the first prediction probability vector and the types of modulation signals in the second prediction probability vector. It then obtains the first prediction probability vector as the final prediction probability vector, or takes the first prediction probability vector as the final prediction probability vector, thereby obtaining the type of modulation signal with the highest probability in the final prediction probability vector and completing the recognition. This scheme does not simply concatenate CNN and GoogLeNet, but rather performs functional division and weighted combination. This resolves the technical contradiction between the inevitable decrease in recognition accuracy due to full type coverage and the inevitable increase in computational cost due to high-precision models, and avoids the impact of GoogLeNet's high complexity on the inference efficiency of non-core signals.

[0010] In summary, this solution can balance full-type coverage with core accuracy, thus ensuring coverage and improving recognition accuracy. Attached Figure Description

[0011] Figure 1 This is a logic block diagram of an embodiment of the signal modulation recognition system based on a convolutional neural network according to the present invention; Figure 2 This is a schematic diagram of the channel in an embodiment of the signal modulation recognition system based on a convolutional neural network according to the present invention; Figure 3 This is a schematic diagram of the data preprocessing module in an embodiment of the signal modulation recognition system based on a convolutional neural network of the present invention; Figure 4 This is a schematic diagram of a lightweight CNN model in an embodiment of the signal modulation recognition system based on a convolutional neural network according to the present invention; Figure 5 This is a schematic diagram of the lightweight CNN model layers in an embodiment of the signal modulation recognition system based on a convolutional neural network according to the present invention; Figure 6 This is a schematic diagram of the GoogLeNet model in an embodiment of the signal modulation recognition system based on a convolutional neural network according to the present invention; Figure 7 This is a schematic diagram of the Stem module in an embodiment of the signal modulation recognition system based on a convolutional neural network of the present invention; Figure 8 This is a simulation illustration of the GoogLeNet model in an embodiment of the signal modulation recognition system based on a convolutional neural network according to the present invention. Figure 1 ; Figure 9 This is a simulation illustration of the GoogLeNet model in an embodiment of the signal modulation recognition system based on a convolutional neural network according to the present invention. Figure 2 ; Figure 10 This is a simulation illustration of the GoogLeNet model in an embodiment of the signal modulation recognition system based on a convolutional neural network according to the present invention. Figure 3 ; Figure 11 These are time-domain waveforms of 12 modulation signals in an embodiment of the signal modulation recognition system based on a convolutional neural network according to the present invention. Figure 12 This is a diagram showing the recognition rate of 12 modulation signals in an embodiment of the signal modulation recognition system based on a convolutional neural network according to the present invention. Figure 13 This is a constellation diagram of MQAM and MPSK at different SNRs in an embodiment of the signal modulation recognition system based on a convolutional neural network of the present invention; Figure 14 This is a schematic diagram illustrating the recognition rate of the GoogLeNet model in an embodiment of the signal modulation recognition system based on a convolutional neural network according to the present invention. Detailed Implementation

[0012] The following detailed description illustrates the specific implementation method: Example 1 The identification of modulated signals in the existing technology has the following problems: It is difficult to balance full coverage with core accuracy: Lightweight CNN covers 12 types of signals, but the recognition rate of core MPSK / MQAM signals (especially high-order 64QAM) is low (≤90.45%), and the misclassification rate of easily confused signals is high; Improved GoogLeNet has a high core recognition rate (≥97.91%), but only covers 6 types of signals, which cannot meet the needs of multiple scenarios (such as spectrum monitoring which requires the recognition of more than 12 types of signals).

[0013] Insufficient robustness in low signal-to-noise ratio environments: Existing technologies have a core signal recognition rate of ≤85% when SNR=6-10dB (e.g., lightweight CNN has a recognition rate of only 81.5% for 16QAM), and cannot adapt to complex channels (e.g., Rice multipath fading channels, K factor=4).

[0014] Imbalance between network parameters and inference efficiency: The original GoogLeNet has ≥5M parameters, FPGA deployment resource utilization ≥80%, and single-frame inference time ≥4ms; the lightweight CNN has a low number of parameters (≤1M), but does not introduce transfer learning, and the training convergence speed is slow (requiring more than 12 iterations).

[0015] The decision-making logic lacks dynamic adaptability: Existing technologies use a single network to directly output the decision, without dynamically adjusting the decision weights according to the signal type (core / non-core), which leads to the recognition accuracy of non-core signals (such as B-FM, GFSK) being affected by the core network (error ≥3%).

[0016] Currently, there are multi-model fusion ideas in the field of Automatic Modulation Recognition (AMR) (such as ResNet+Transformer

[12] , CNN+LSTM

[18] ), but no fusion architecture of "lightweight CNN+improved GoogLeNet" has emerged. The core reason is that there are certain technical contradictions between the two types of models and a lack of targeted fusion mechanisms. Model localization conflict: Conventional CNNs are good at "broad coverage and low complexity" of all types of feature extraction, but they are weak in fine-grained discrimination of high-order MPSK / MQAM signals; GoogLeNet is good at "high precision and high complexity" of core feature mining, but it has a large number of parameters and a narrow coverage. Directly fusing the two can easily lead to "imbalance between accuracy and efficiency", mainly because the high complexity of GoogLeNet will offset the lightweight advantage of CNN. Incompatible feature dimensions: CNN outputs "time-domain / frequency-domain feature probabilities of 12 signals" (1-dimensional vector), while the conventional GoogLeNet outputs "image class feature probabilities" (such as 1000 classes in ImageNet). The two have large differences in feature dimensions and semantic space, and lack a unified feature alignment mechanism. Direct splicing will lead to a decrease in recognition rate after fusion. Insufficient scenario adaptability: Existing fusion solutions are mostly designed for "single channel environments" (such as ideal Gaussian channels) and do not consider the issues of "model robustness and dynamic adjustment of fusion weights" under complex channels such as Rice multipath fading. However, the core application scenarios of this solution (5G communication and communication reconnaissance) can cope with complex channels, which is also a key limitation in the AMR field that the two have not been combined.

[0017] To address the aforementioned problems, this embodiment provides a signal modulation recognition system based on a convolutional neural network, as shown in the attached figure. Figure 1 and Figure 2 As shown, it includes: Signal receiving module, preprocessing module, lightweight CNN channel module, GoogLeNet channel module, decision fusion module; The signal receiving module is used to receive the modulated signal, convert it into an IQ signal, and send it to the preprocessing module. Specifically, the signal receiving module uses a radio frequency terminal, whose antenna receives modulated signals in the 200MHz~2.7GHz frequency band, and the modulated signals are analog signals; The modulation signal includes several types, such as MPSK, MQAM, PAM4, etc., totaling 12 types; in this embodiment, it includes BPSK, QPSK, 8PSK, 16QAM, 32QAM, 64QAM, PAM4, GFSK, CPFSK, B-FM, DSB-AM, and SSB-AM. The analog signal is converted into a 16-bit digital IQ signal by AD sampling, where I is the in-phase component and Q is the quadrature component; in this embodiment, the sampling rate of Adam sampling is 8000Hz and the number of samples per symbol is 8 (sps). The IQ signal is sent to the preprocessing module. In this embodiment, the IQ signal is divided into 32 symbols per frame (sampling number per frame: spf = sps × 32 = 256), packaged into 256 × 2 dimension data blocks, and transmitted to the embedded CPU, i.e., the preprocessing module, via USB 3.0. The data format of the IQ signal is: the IQ signal value range is [-32768, 32767] (16-bit integer), and the frame interval is ≤ 1ms.

[0018] The preprocessing module is used to place the IQ signal into an output frame of a preset format and send it to the lightweight CNN channel module and the GoogLeNet channel module respectively. like Figure 3 As shown, the IQ signals are placed in a preset format output frame, with the size of the output frame being [2spf×1×N]. The first row is the in-phase component of the in-phase sampling, and the second row is the quadrature component of the quadrature sampling.

[0019] The lightweight CNN channel module is used to analyze IQ signals using a constructed lightweight CNN model to obtain the first predicted probability vector for all types of modulated signals. The lightweight CNN model constructed therein, such as Figure 4 As shown, it includes an input layer, a basic module, a fully connected layer, a softmax layer, and an output layer connected in sequence; The basic modules include: convolutional layers, batch normalization layers, ReLU activation layers, and max pooling layers; and there are several basic modules, which are connected in sequence. The input layer is used to input data. In this embodiment, the input layer of the lightweight model is used to input two-dimensional data that combines the I-channel signal and the Q-channel signal into 2*1024. Convolutional layers are used to extract local features from input data; The batch normalization layer is used to normalize the input of small batches of data to accelerate the training of the neural network and improve the convergence speed; in this embodiment, the small batch of data is a preset number of local features. ReLU activation layers are used to introduce nonlinear mappings, enabling the network to learn and represent complex nonlinear features, forming feature maps; Max pooling layers are used to downsample feature maps, reduce the number of parameters, and improve feature invariance. Fully connected layer: Maps the features extracted by the convolutional layer and max pooling layer to the specific task output for classification; Output layer: This layer transforms the extracted feature maps into task-related prediction results, i.e., the specific modulation scheme category. Adding a batch normalization layer after each convolutional layer can accelerate training convergence, alleviate the gradient vanishing problem, and improve model stability. Using an average pooling layer instead of a traditional fully connected layer as a feature compression step, and then connecting it to another fully connected layer to output the classification result, reduces the large number of parameters introduced by the fully connected layer. This embodiment sets up 6 basic modules to lighten the specific structure of the CNN model, such as Figure 5 As shown; The convolutional layers of the basic module are set as multi-scale convolutional layers. Specifically, the multi-scale convolutional layers are 96-channel multi-scale convolutional layers, using three types of convolutional kernels in parallel: 1×1 (capturing high-frequency details), 3×3 (capturing mid-frequency features), and 5×5 (capturing low-frequency global features). Multi-scale features are fused through channel concatenation, and the output dimension is 96×32×32. Conventional lightweight CNNs only use a single 3×3 convolutional kernel, which cannot simultaneously extract the fine-grained temporal features (such as instantaneous phase changes) and frequency-domain global features (such as spectral distribution) of the signal, resulting in severe feature loss at low SNR. This scheme sets up multi-scale convolutional layers, and at SNR=6dB, the feature extraction accuracy of the core signals (MPSK and MQAM type modulation signals) is improved by 12.4 percentage points, providing a richer feature foundation for subsequent fusion.

[0020] The final basic module replaces the max pooling layer with an average pooling layer, and a sparse fully connected layer is set after the average pooling layer to replace the fully connected layer of the lightweight CNN model (sparseness ratio 0.3, i.e. only 30% of neurons participate in the calculation). The output dimension is expanded from 12 dimensions to 24 dimensions, where 12 dimensions correspond to the type of modulation signal and 12 dimensions correspond to the feature confidence, providing "type + confidence" dual-dimensional information for decision fusion. Conventional lightweight CNNs use "global average pooling + single-layer fully connected layer", which easily leads to insufficient feature mapping dimension of core signals and blurred classification boundaries of high-order QAM signals (such as 64QAM). This scheme sets up a sparse fully connected layer, which improves the recognition accuracy of non-core signals (except for MPSK and MQAM type modulation signals) from 97.95% to 99.2%, and provides a confidence verification basis for the fusion module.

[0021] The GoogLeNet channel module is used to reconstruct the model from the constructed constellation diagram and generate the constellation diagram based on the IQ signal; It is also used to analyze the constellation diagram generated by IQ signals through the constructed GoogLeNet model to obtain the second prediction probability vector of MPSK and MQAM type modulated signals; The GoogLeNet model constructed therein, such as Figure 6 As shown, it includes: an input layer, a Stem module, an Inception module, a pooling layer, and an output layer; The Stem module is a constellation graph reconstruction model, which includes several convolutional layers and max pooling layers, used to convolve and pool the IQ signals input from the input layer to generate a constellation graph. like Figure 7 As shown, the constellation map reconstruction module in this embodiment includes a 7×7 convolutional layer, a 3×3 max pooling layer, a 1×1 convolutional layer, a 2×3 convolutional layer, and a 3×3 max pooling layer connected in sequence to improve the recognition rate of MQAM and MPSK modulation methods.

[0022] The Inception module is used to perform parallel multi-scale convolution and pooling operations.

[0023] The GoogLeNet model constructed in this embodiment includes, in sequence, an input layer, a Stem module, two Inception modules, a max pooling layer, an Inception module, three Inception modules, an Inception module, a max pooling layer, two Inception modules, an average pooling layer, and an output layer. It achieves efficient feature extraction through a 144-layer structure, specifically: the input and output layers consist of 11 layers (input, conv1 7×7, ReLU, 3×3 max pooling layer, Norm, conv2 3×3, ReLU, conv2 3×3, ReLU, Norm; Average Pooling 7×7, Drop, New_Feature_Learner, softmax, output), and 9 Inception modules (one Inception module including the output layer is 14 layers), totaling 126 layers; plus the two 3×3 max pooling layers and the 5 layers of the Stem module, for a total of 144 layers. The simulation diagram is shown below. Figure 8 , Figure 9 , Figure 10 As shown.

[0024] The new fully connected layer New_Feature_Learner replaces the original GoogLeNet's 142nd layer Feature_Learner (which corresponds to a fully connected layer with 1000 output classes), and the output dimension is changed to the number of classes in the custom dataset (Number_of_Classes). A new classification layer, New_Classifier_Layer, is used to replace the original classification layer (144th layer, Output_Classifier) ​​of GoogLeNet, to output the classification results of custom categories; In the GoogLeNet model, the WeightLearnRateFactor (which controls the learning speed of the weight parameters) and BiasLearnRateFactor (which controls the learning speed of the bias parameters) are set to 10. This makes the learning rate of the new fully connected layer higher than that of other pre-trained layers, thus realizing the transfer learning strategy of freezing the bottom features and fine-tuning the top classifier.

[0025] The GoogLeNet model in this solution does not directly reuse the pre-trained model, but is improved specifically for signal modulation recognition scenarios, including: Adjusting the kernel ratio and number of channels of the Inception module, in this embodiment, it is set to 1×1 (96), 3×3 (192), 5×5 (48), and pooling (48), which can achieve better convolution effect. Other kernel ratios and number of channels can also be set in other embodiments. Compared to the fixed channel ratio of "1×1 (64), 3×3 (128), 5×5 (32), pooling (32)" used in the Inception module of the conventional GoogLeNet, which is designed for natural images, it cannot adapt to the characteristics of sparse signal features and key edge information in constellation diagrams, resulting in weak signal point discrimination ability for high-order QAM signals (such as 32QAM / 64QAM). This scheme adjusts the channel ratio of the Inception module to "1×1 (96), 3×3 (192), 5×5 (48), pooling (48)", and adds "edge detection convolution kernel (1×3+3×1)" after 3×3 convolution to enhance the edge feature extraction of signal points in constellation diagrams. The cross-rejection rate of core signals (32QAM / 64QAM) is reduced from 2.3% to 1.1%, and the discrimination accuracy of overlapping signal points in constellation diagrams is improved by 15.6%.

[0026] Replace the classification layer activation function with Softmax weighted activation (core signal weight 0.5, non-core signal weight 1.5), and replace the loss function with weighted cross-entropy loss to balance the differences in sample distribution: ; in For sample weights, core signal =0.5, non-core signal =1.5; For real labels, To predict probabilities; Compared to the classification layer in the conventional GoogLeNet, which uses Softmax activation and cross-entropy loss and is designed for natural images with a balanced distribution of multiple classes, signal modulation recognition has a high proportion of core signal samples and a low proportion of non-core signal samples (e.g., MPSK / MQAM samples account for 60% and other signals account for 40%), which can easily lead to the model being biased towards core signals and having a high misclassification rate for non-core signals. This scheme replaces the classification layer activation function with Softmax weighted activation and the loss function with weighted cross-entropy loss to balance the differences in sample distribution. The recognition accuracy of non-core signals (such as PAM4 and SSB-AM) is improved from 95.3% to 98.5%, while the recognition rate of core signals does not decrease (it remains at 98.13%), thus solving the model bias problem caused by imbalanced samples.

[0027] The decision fusion module is used to determine whether the type of the modulation signal with a probability value greater than or equal to a preset probability value in the first prediction probability vector is consistent with the type of the modulation signal in the second prediction probability vector. If so, the first prediction probability vector and the second prediction probability vector are weighted and fused to obtain the final prediction probability vector; otherwise, the first prediction probability vector is used as the final prediction probability vector. It is also used to obtain the type of the modulation signal with the highest probability in the final predicted probability vector, thus completing the fusion recognition.

[0028] Specifically, the modulation type prediction probability of the two-channel output is used as the core input: The lightweight CNN channel module outputs the first prediction probability vector of 12 modulated signals. This corresponds to 12 modulation signals; The GoogLeNet channel module outputs a second prediction probability vector for six core signals (BPSK, QPSK, 8PSK, 16QAM, 32QAM, and 64QAM). ; It is determined whether the type of the modulation signal with a probability value greater than or equal to a preset probability value in the first prediction probability vector is consistent with the type of the modulation signal in the second prediction probability vector. In this embodiment, the preset probability value is 0.5, which is a high probability signal type. If the first and second prediction probability vectors are weighted and fused to obtain the final prediction probability vector, that is, if the high-probability signal type belongs to the recognition range of the GoogLeNet model (i.e., the 6 core signals), then the prediction probabilities of the signal in the two channels are weighted and summed to obtain the final prediction probability vector. ,in , These are the weighting coefficients. Corresponding to the same type of modulation signal; If not, the first predicted probability vector is used as the final predicted probability vector; that is, if the high-probability signal belongs to a signal without I / Q components (such as PAM4, SSB-AM, and other non-core signals), then the prediction result of the CNN (the first predicted probability vector) is directly used as the final decision. This is because such signals are not within the constellation diagram recognition range of GoogLeNet.

[0029] pass Select the type of modulation signal with the highest probability and output it to complete the fusion recognition.

[0030] The decision fusion module performs a weighted fusion of the first and second prediction probability vectors to obtain the final prediction probability vector, where the weights are dynamically selected based on the SNR value. If SNR <= the first preset SNR value, which is 10dB in this embodiment, then , ; If SNR > the first preset SNR value, then , ; In the specific implementation process, an embedded CPU (responsible for feature alignment and weight calculation) and an FPGA (responsible for real-time SNR detection and threshold triggering) are used as the main execution components. (After the preprocessing module outputs the IQ signal, it is simultaneously distributed to the lightweight CNN channel (for extracting temporal features) and the GoogLeNet channel; The FPGA detects the current channel SNR in real time (sampling interval 1ms) and transmits the SNR value to the decision fusion module; The decision fusion module receives the 12-dimensional probability vector from the CNN and the 6-dimensional probability vector from GoogLeNet, and combines them using formula (2). The probability value of the core signal and Alignment: in For core signal index, ( ) represents the maximum probability output by GoogLeNet. (Core signal) is the maximum probability of the core signal output by the CNN, which realizes probability space normalization.

[0031] Lightweight CNN model recognition: Datasets with different SNRs were generated using the MATLAB 2022b platform. A Ricean multipath fading channel was selected, with a K-factor of 4, a maximum Doppler shift of 4Hz, and a maximum clock offset of 5ppm. Each frame consisted of 1024 samples at a sampling rate of 200kHz. For digital modulation types, eight samples represented one symbol. The network made each decision based on a single frame rather than multiple consecutive frames. Assuming the center frequencies for digital and analog modulation types were 900MHz and 100MHz respectively, ... Figure 11 The figure shows the time-domain waveforms of 12 modulation signals with an SNR of 18dB.

[0032] The lightweight CNN network was used to perform modulation scheme recognition on the test set, and the recognition confusion matrix was obtained as follows: Figure 12 As shown, the following conclusions can be drawn: (1) The overall modulation scheme recognition accuracy reached 90.45%; (2) For the five debugging methods of 16QAM, 32QAM, 64QAM, 8PSK and QPSK, the overall recognition accuracy is about 20%; (3) Excluding the above 5 debugging methods, the recognition accuracy of the other 7 modulation methods can reach 97.95%.

[0033] contrast Figure 11 The time-domain waveforms show that the waveforms of MPSK and MQAM are indeed easily confused, which is consistent with the test results.

[0034] GoogLeNet model recognition: For six types of MQAM and MPSK modulated signals, five constellation diagrams under seven different SNRs were selected as test data for each type of modulated signal. Some constellation diagrams are shown below. Figure 13 As shown, from left to right, the SNR values ​​are 6, 8, 10, 12, 14, 16, and 18 (dB), and from top to bottom, they are BPSK, QPSK, 8PSK, 16QAM, 32QAM, and 64QAM.

[0035] The GoogLeNet model was used to identify the modulation scheme of the constellation diagram test set, and the confusion matrix was obtained as follows: Figure 14 As shown, the following conclusions can be drawn: when BPSK, QPSK, 8PSK, 16QAM, 32QAM, and 64QAM modulation signals are mixed together, and each modulation signal has seven different constellation diagrams under different SNRs, the overall success rate of the modulation method can reach 97.91%. This indicates that the optimized GoogLeNet network in this paper has a high accuracy in recognizing constellation diagrams of MQAM and MPSK, and performs well in recognizing modulation methods under large dynamic SNR environments.

[0036] The identification results of the decision fusion module in this solution: (1) Overall recognition performance: The decision fusion module achieved an average recognition accuracy of 96.72% for 12 types of modulated signals, a 6.27 percentage point improvement over a single lightweight CNN model (90.45%), and a dual objective of achieving full-type coverage and high-precision recognition compared to a single improved GoogLeNet model (97.91% for core signals, excluding I / Q signals). Specifically, the recognition accuracy for the six signals without I / Q components remained above 98.5% (consistent with the accuracy of a single CNN model), and the average recognition accuracy for the six core MPSK / MQAM signals reached 98.13%, a 0.22 percentage point improvement over the single GoogLeNet model, validating the effectiveness of the fusion strategy.

[0037] (2) Robustness under different SNR Low SNR (6dB~10dB): In the core MPSK / MQAM signals, only 16QAM and 32QAM showed slight confusion (recognition rates of 92.3% and 91.8%, respectively), while the recognition rates of the remaining signals were ≥95%. Compared with the single-network results, the lightweight CNN only achieved a recognition rate of 81.5% for 16QAM under low SNR, while the fusion model, leveraging the noise resistance of the GoogLeNet constellation diagram, improved accuracy by 10.8 percentage points, demonstrating the adaptability of the fusion model to low SNR environments.

[0038] Medium to high SNR (12dB~18dB): The recognition rate of all signals is ≥98.5%, with BPSK, QPSK and 8PSK achieving "zero misjudgment". The misjudgment of 16QAM, 32QAM and 64QAM is mainly due to a small amount of overlap of signal points in the constellation diagram (such as confusion of edge points between 16QAM and 32QAM), but the misjudgment rate is ≤1.5%, which is far lower than the level of a single network.

[0039] (3) Core signal identification comparison For the two easily confused signal types, "QPSK vs. 16QAM" and "32QAM vs. 64QAM", which are most difficult for traditional algorithms to distinguish, the recognition accuracy of the fusion model is significantly improved: The cross-rejection rate of QPSK and 16QAM decreased from 8.7% for lightweight CNNs to 1.2%, and remained basically the same as that of GoogLeNet at 1.0%, achieving high accuracy and no degradation. The cross-rejection rate of 32QAM and 64QAM decreased from 2.3% in GoogLeNet to 1.1%, because it incorporates the ability of CNN to extract temporal features of signals, thus compensating for the shortcomings of constellation diagrams in distinguishing high-order signal edge points.

[0040] (4) Comparison of the effects after improvement Test conditions: SNR=10dB, Rice multipath fading channel K=4. The comparison of the quantization effects of the model improvement is shown in Table 1: Table 1: Comparison of Quantitative Effects of Model Improvement In summary, compared with existing technologies, the beneficial effects of this solution are as follows: A lightweight CNN-GoogLeNet dual-channel decision fusion architecture is proposed: the first channel (lightweight CNN) achieves full coverage screening of 12 modulation signals (6 core MPSK / MQAM + 6 signals without I / Q components), and the second channel (improved GoogLeNet) focuses on high-precision identification of core signals. The imbalance between coverage and accuracy is solved by probabilistic weighted decision (core signals: CNN weight 0.4, GoogLeNet weight 0.6; non-core signals: CNN weight 1.0). Optimize the lightweight CNN structure: Add a 96-channel convolutional layer after the 5 convolutional modules, and combine it with "segmented decay learning rate" (initial 2e-2, decaying by 0.1 every 9 rounds) to improve the feature extraction capability under low SNR (the core signal recognition rate is improved by 10.8 percentage points when SNR=6dB). Improved GoogLeNet transfer learning strategy: Freeze the first 142 layers of the pre-trained model, fine-tune only the fully connected layers and classification layers, and introduce "constellation graph feature enhancement" (based on IQ component density mapping), which improves the core signal recognition rate to over 98%. Adapting to Rice multipath fading channels: By using "wavelet threshold denoising + data augmentation" (amplitude scaling 0.8~1.2, phase shift -10°~10°), it maintains high robustness in an environment with K factor=4 and maximum Doppler frequency shift of 4Hz.

[0041] This solution does not simply combine CNN and GoogLeNet; instead, it employs a functional division of labor and a dynamic weighting mechanism to resolve the technical contradiction between the inevitable decrease in recognition accuracy due to full-type coverage and the increased computational cost inherent in high-precision models. Specifically, the CNN is responsible for outputting coarse classification probabilities for 12 signals, while GoogLeNet focuses on high-precision recognition of 6 MPSK / MQAM signals. Furthermore, a "signal type-triggered weighting" system is designed (GoogLeNet weight 0.6 for core signals, CNN weight 0.4, and CNN weight 1.0 for non-core signals) to avoid the high complexity of GoogLeNet impacting the inference efficiency for non-core signals.

[0042] This embodiment also provides a signal modulation recognition method based on a convolutional neural network, including: Acquire the modulated signal and convert it into an IQ signal; The IQ signal is analyzed by constructing a lightweight CNN model to obtain the first predicted probability vector of all types of modulated signals; The IQ signal was analyzed using the constructed GoogLeNet model to obtain the second prediction probability vectors of MPSK and MQAM type modulated signals; Determine whether the type of the modulation signal with a probability value greater than or equal to a preset probability value in the first prediction probability vector is consistent with the type of the modulation signal in the second prediction probability vector. If yes, the first prediction probability vector and the second prediction probability vector are weighted and fused to obtain the final prediction probability vector. If no, the first prediction probability vector is used as the final prediction probability vector. Obtain the type of the modulation signal with the highest probability in the final prediction probability vector.

[0043] Note: Communication signal processing category: PSK: Phase Shift Keying; QAM: Quadrature Amplitude Modulation. SNR: Signal-to-Noise Ratio; dB: Decibel (a logarithmic unit used to express power ratios); IQ signal: In-phase-Quadrature signal; I-component: In-phase component (signal component that is in phase with the reference carrier). Q component: Quadrature component (signal component that is 90° out of phase with the reference carrier). spf: Samples Per Frame; sps: Samples Per Symbol; Carrier wave; Constellation Diagram; Modulation order; Noise Margin; Signal Power; Noise Power; Deep learning and network model categories: CNN: Convolutional Neural Network. Lightweight Convolutional Neural Network (CNN) Convolutional Layer Pooling layer: A network layer that compresses feature dimensions while retaining key information. Max Pooling Layer: Max Pooling Layer (a pooling layer that uses the maximum value for sampling) Average Pooling Layer: Average Pooling Layer (a pooling layer that uses average value sampling) BN layer: BatchNormalizationLayer; Batch Normalization: Batch Normalization; ReLU layer: ReLU activation layer (an activation layer that uses the ReLU function) (ReLUActivationLayer); ReLU function: ReLU function (Rectified Linear Unit Function); Fully Connected Layer: A fully connected layer (a network layer that maps convolutional features to one-dimensional vectors). Softmax layer: The Softmax layer (which transforms the network output into a probability distribution) Classification Layer: Classification Layer (Network layer for output signal type); GoogLeNet: Google Network (a deep convolutional neural network proposed by Google) (GoogLeNet); Improved GoogLeNet: An Improved Google Network Transfer learning: Transfer learning is a learning method that reuses pre-trained model weights to reduce the training cost of new tasks. Pretrained layer: A pretrained layer (a network layer that has been trained on other datasets). Dual-channel decision fusion: Dual-channel decision fusion. Confusion Matrix: A confusion matrix. Robustness: Robustness (the ability of a model to maintain stable performance under disturbances).

[0044] The above descriptions are merely embodiments of the present invention. Commonly known structures and characteristics are not described in detail here. Those skilled in the art are aware of all common technical knowledge in the field prior to the application date or priority date, are aware of all existing technologies in that field, and have the ability to apply conventional experimental methods prior to that date. Those skilled in the art can, under the guidance of this application, improve and implement this solution in combination with their own capabilities. Some typical known structures or methods should not be obstacles for those skilled in the art to implement this application. It should be noted that those skilled in the art can make several modifications and improvements without departing from the structure of the present invention. These should also be considered within the scope of protection of the present invention, and will not affect the effectiveness of the implementation of the present invention or the practicality of the patent. The scope of protection claimed in this application should be determined by the content of its claims, and the specific embodiments described in the specification can be used to interpret the content of the claims.

Claims

1. A signal modulation recognition system based on a convolutional neural network, characterized in that, include: The signal receiving module is used to receive modulated signals and convert them into IQ signals; The lightweight CNN channel module is used to analyze IQ signals using a constructed lightweight CNN model to obtain the first predicted probability vector for all types of modulated signals. The GoogLeNet channel module is used to analyze IQ signals using the constructed GoogLeNet model to obtain the second prediction probability vectors of MPSK and MQAM type modulated signals. The decision fusion module is used to determine whether the type of the modulation signal with a probability value greater than or equal to a preset probability value in the first prediction probability vector is consistent with the type of the modulation signal in the second prediction probability vector. If so, the first prediction probability vector and the second prediction probability vector are weighted and fused to obtain the final prediction probability vector; otherwise, the first prediction probability vector is used as the final prediction probability vector. It is also used to obtain the type of the modulation signal with the highest probability in the final predicted probability vector.

2. The signal modulation recognition system based on a convolutional neural network according to claim 1, characterized in that, The modulated signal includes: a core signal and a non-core signal; The core signals are MPSK and MQAM type modulation signals, including: BPSK, QPSK, 8PSK, 16QAM, 32QAM, and 64QAM; Non-core signals include: PAM4, GFSK, CPFSK, B-FM, DSB-AM, and SSB-AM.

3. The signal modulation recognition system based on a convolutional neural network according to claim 1, characterized in that, The constructed lightweight CNN model includes an input layer, a base module, a fully connected layer, a softmax layer, and an output layer connected in sequence. The basic modules include: convolutional layers, batch normalization layers, ReLU activation layers, and average pooling layers; and there are several basic modules, which are connected in sequence.

4. The signal modulation recognition system based on a convolutional neural network according to claim 3, characterized in that, The max pooling layer of the last basic module is replaced with an average pooling layer.

5. The signal modulation recognition system based on a convolutional neural network according to claim 4, characterized in that, The average pooling layer is followed by a sparse fully connected layer to replace the fully connected layer of the lightweight CNN model.

6. The signal modulation recognition system based on a convolutional neural network according to claim 1, characterized in that, The GoogLeNet model includes: an input layer, a Stem module, an Inception module, a pooling layer, and an output layer; The Stem module is a constellation graph reconstruction model, which includes several convolutional layers and max pooling layers, used to convolve and pool the IQ signals input from the input layer to generate a constellation graph, which is then input into the Inception module.

7. The signal modulation recognition system based on a convolutional neural network according to claim 6, characterized in that, The GoogLeNet model comprises, in sequence: an input layer, a Stem module, two Inception modules, a max pooling layer, four Inception modules, a max pooling layer, two Inception modules, an average pooling layer, and an output layer. The GoogLeNet model described above uses a new fully connected layer, New_Feature_Learner, to replace the original GoogLeNet's Feature_Learner, and the output dimension is changed to the number of categories in a custom dataset. A new classification layer, New_Classifier_Layer, is used to replace the original classification layer of GoogLeNet, and is used to output the classification results of custom categories; Set WeightLearnRateFactor and BiasLearnRateFactor to 10 in the GoogLeNet model.

8. The signal modulation recognition system based on a convolutional neural network according to claim 6, characterized in that, The Inception module has the following kernel ratios and channel numbers: 1×1 and 96, 3×3 and 192, 5×5 and 48, and pooling and 48.

9. The signal modulation recognition system based on a convolutional neural network according to claim 1, characterized in that, The weighted summation is then performed to obtain the final predicted probability vector: ,in , These are the weighting coefficients. Corresponding to the same type of modulation signal; The weights are dynamically assigned based on the SNR value: If SNR <= the first preset SNR value, then , ; If SNR > the first preset SNR value, then , .

10. A signal modulation recognition method based on convolutional neural networks, characterized in that, include: Acquire the modulated signal and convert it into an IQ signal; The IQ signal is analyzed by constructing a lightweight CNN model to obtain the first predicted probability vector of all types of modulated signals; The IQ signal was analyzed using the constructed GoogLeNet model to obtain the second prediction probability vectors of MPSK and MQAM type modulated signals; Determine whether the type of the modulation signal with a probability value greater than or equal to a preset probability value in the first prediction probability vector is consistent with the type of the modulation signal in the second prediction probability vector. If yes, the first prediction probability vector and the second prediction probability vector are weighted and fused to obtain the final prediction probability vector. If no, the first prediction probability vector is used as the final prediction probability vector. Obtain the type of the modulation signal with the highest probability in the final prediction probability vector.

Images