Electromagnetic signal identification technology based on pulse neural network
By constructing a SCNN model based on LIF neurons, integrating tdBN, CBAM and residual structures, and using the pulse coding layer to process I/Q signals, the high energy consumption and high parameter problems of electromagnetic signal recognition in edge hardware are solved, and efficient electromagnetic signal recognition is achieved.
Patent Information
- Application Number
- CN202510577002.5
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2025-05-06
- Publication Date
- 2025-08-15
AI Technical Summary
Existing deep learning models have high energy consumption and large parameters under edge hardware conditions, making it difficult to operate efficiently in resource-constrained devices.
A SCNN model based on LIF neurons is constructed, tdBN normalization technology, CBAM attention mechanism and residual structure are integrated, and the I/Q signal is converted into pulse sequence signals using the pulse coding layer to perform electromagnetic signal recognition.
While ensuring recognition accuracy, it reduces the calculation energy consumption and parameter quantity, and is suitable for energy-constrained edge hardware scenarios, improving the robustness and efficiency of signal recognition.
Smart Images
Figure CN120492971A_ABST
Abstract
Description
Technical Field
[0001] The present invention relates to the field of electromagnetic signal recognition technology, and in particular to an electromagnetic signal recognition technology based on a pulse neural network. Background Art
[0002] Electromagnetic signal recognition plays a crucial role in electromagnetic environment perception and processing. Its core purpose is to extract valuable target signals from the complex electromagnetic spectrum while eliminating interference and background noise. The key to achieving this goal lies in the accurate extraction and analysis of signal attributes, including core metrics such as frequency, bandwidth, and duration. With the widespread deployment of next-generation wireless communication technologies and IoT devices, the density of electromagnetic signal usage continues to rise, with a wide variety of signals and frequent overlap. Electromagnetic interference between communication terminals and base stations is becoming increasingly prominent. Therefore, effectively distinguishing legitimate communication signals from potential anomalies has become a crucial prerequisite for improving the performance of electromagnetic signal processing systems.
[0003] Signal recognition technology is a key research area in the field of information and communication technology. Currently, mainstream recognition methods can be roughly divided into two categories: recognition strategies based on the maximum likelihood criterion and recognition models built on signal characteristics. The first method offers high theoretical recognition accuracy because it fully utilizes unknown parameters in the signal to optimize the recognition process and is often used as a theoretical performance benchmark for modulation mode recognition. However, these methods typically have high computational complexity, especially in practical applications with significant frequency drift or time offset. The demand for computing resources is extremely demanding, making it difficult to operate efficiently on resource-constrained devices. The second method also presents similar issues.
[0004] In recent years, the development of feature-based deep learning has provided a new approach to electromagnetic signal recognition. Compared to traditional methods, deep neural networks possess superior self-learning capabilities, enabling automatic feature extraction from large amounts of data, significantly improving the model's adaptability and robustness in diverse signal environments. Deep learning methods can maintain high recognition accuracy, particularly in high-noise environments, and exhibit a certain degree of interference resistance. Convolutional neural networks, particularly due to their exceptional performance in image and audio signal processing, have also been introduced to the field of wireless signal analysis.
[0005] While deep learning improves recognition performance, its model structures are often complex, requiring high resource consumption for training and inference, limiting its applicability in real-time processing scenarios. Current technical bottlenecks primarily arise from the conflict between model size and operational efficiency: large models offer excellent accuracy but struggle to run in real time, while small models, while computationally inexpensive, have limited recognition accuracy and struggle to process complex or high-order modulated signals. Furthermore, deep models generally have high requirements for computing power and energy consumption, making them particularly limited when deployed at the edge, making them difficult to directly adapt to low-power hardware platforms like neuromorphic chips.
[0006] Therefore, a new technical solution is urgently needed to solve the above technical problems. Summary of the Invention
[0007] The purpose of the present invention is to overcome the above-mentioned problems of the prior art and provide an electromagnetic signal recognition technology based on a pulse neural network to solve the problem of high computational energy consumption and high parameter quantity in the prior art when using deep neural networks for electromagnetic signal recognition under edge hardware conditions.
[0008] The above objectives are achieved through the following technical solutions:
[0009] An electromagnetic signal recognition technology based on a pulse neural network, comprising:
[0010] Step (1) Integrate tdBN normalization technology, CBAM attention mechanism and residual structure to build a SCNN model based on LIF neurons;
[0011] Step (2) uses the I / Q signal as basic data and converts it into a pulse sequence signal that can be processed by the SCNN model through a pulse coding layer;
[0012] Step (3) inputs the pulse sequence signal into the SCNN model to perform an electromagnetic signal recognition simulation experiment and output the recognition accuracy.
[0013] Furthermore, the step (1) is specifically as follows:
[0014] Step (101) introduces the LIF model as the model neuron function, and its discrete form dynamic equation is as follows:
[0015]
[0016] Step (102) uses tdBN with an additional time dimension instead of traditional BN;
[0017] Step (103) uses CBAM attention with an additional time dimension after each convolutional layer to improve the feature extraction ability of the convolutional layer;
[0018] Step (104) uses a residual structure to prevent the model from experiencing gradient vanishing problems.
[0019] Furthermore, the step (102) is specifically as follows: tdBN optimizes the standard BN from two aspects:
[0020] The first aspect is the optimization of the time dimension. Compared with ANN, SNN has an additional time dimension, and information is transmitted from one moment to the next. Therefore, tdBN normalizes the input in both time and space dimensions, to a certain extent avoiding the time dependency that may occur in traditional BN when processing time series.
[0021] The second aspect is that tdBN modifies the normalized variance. In BN, the value before activation is normalized to N(0,1); in tdBN, the value before activation will be normalized to N(0,(αV th ) 2 ); where V th As the threshold, it helps to capture the statistical characteristics of the signal on the local time scale and identify the dynamic changes of the signal.
[0022] Furthermore, the step (103) is specifically as follows: using a modified CBAM attention mechanism to enhance the feature extraction capability of the convolutional layer; the modified CBAM is composed of CAM and SAM, and all layers have an additional time dimension;
[0023] CAM first performs maximum pooling and average pooling on the upper layer output while preserving the channel dimension. It then uses a multi-layer perceptron and neuron function to extract valuable information from the channel. Finally, the two are added together to obtain the channel weight and multiplied with the original data.
[0024] SAM first performs size-invariant maximum pooling and average pooling on the upper layer output and concatenates them, then uses convolutional layers and neuron functions to obtain spatial weights, and finally multiplies them with the original data.
[0025] Furthermore, the step (104) is specifically as follows: each two groups of neuron-convolution-CBAM-tdBN structures are connected by a shortcut structure before and after, and the two groups of pre-convolution data and post-convolution data are added to form a residual structure.
[0026] Furthermore, the step (2) is specifically as follows:
[0027] Step (201) constructs a pulse coding layer;
[0028] Step (202) pulse encodes the I / Q signal.
[0029] Furthermore, the step (201) is specifically as follows: first, a convolution layer, a CBAM and a tdBN layer are used to extract features of the I / Q signal; then, the output result is copied into two copies, one of which is subjected to a neuron-convolution-CBAM-tdBN operation to fuse the I component and the Q component, and the other is subjected to only a convolution operation for size change; finally, the two output results are added and fed into the neuron function to complete the construction of the pulse coding layer.
[0030] Furthermore, the pulse coding layer in step (202) receives floating-point data input, performs multiplication and addition operations on synapses, and finally generates pulse outputs through the neuron layer, and can update parameters as the network is trained.
[0031] Furthermore, the recognition accuracy described in step (3) is measured by the average accuracy, that is, the percentage of correctly predicted samples to the total number of samples; the model computing energy consumption is obtained by multiplying the number of synaptic operations by the computing energy consumption of a single addition in the hardware.
[0032] The electromagnetic signal recognition technology based on a spiking neural network provided by this invention achieves comparable recognition accuracy to other commonly used models while reducing computational energy consumption and parameter requirements. This technology addresses the challenges of performing electromagnetic signal recognition tasks on edge hardware and provides reliable support for electromagnetic signal recognition tasks in energy-constrained scenarios. Furthermore, it offers the following benefits:
[0033] By using the pulse coding layer to convert the original continuous-valued I / Q signal into a pulse train signal, it is updated as the network trains, and the features of the I and Q components are fused in the pulse coding layer, so that the entire network can better extract the data in the signal and provide more useful input data for the lower-layer network;
[0034] By integrating tdBN technology, the model's ability to capture the statistical characteristics of signals at local time scales and identify dynamic changes in signals is enhanced. Integrating CBAM technology, which adds an additional time dimension, allows the model to focus more on key signal features, thereby improving overall network performance. Integrating a residual structure prevents problems such as vanishing gradients. Using LIF functions as network neurons allows for inter-network transmission of spike data with relatively simple computational requirements. BRIEF DESCRIPTION OF THE DRAWINGS
[0035] Figure 1 This is a flow chart of an electromagnetic signal recognition technology based on a pulse neural network according to the present invention;
[0036] Figure 2 This is a schematic diagram of the working of tdBN in the electromagnetic signal recognition technology based on pulse neural network described in the present invention;
[0037] Figure 3 This is a schematic diagram of the structure of the CBAM in the electromagnetic signal recognition technology based on the pulse neural network described in the present invention;
[0038] Figure 4 This is a structural diagram of the SCNN model in the electromagnetic signal recognition technology based on pulse neural network described in the present invention. DETAILED DESCRIPTION
[0039] The present invention will be further described in detail below with reference to the accompanying drawings and examples. The described embodiments are only some embodiments of the present invention, not all embodiments. All other embodiments derived by persons of ordinary skill in the art based on the embodiments of the present invention without inventive effort shall fall within the scope of protection of the present invention.
[0040] like Figure 1 As shown, this solution provides an electromagnetic signal recognition technology based on a pulse neural network, including:
[0041] Step (1) Integrate the tdBN normalization technology, the CBAM attention mechanism with an additional time dimension, and the residual structure to construct a SCNN model based on LIF neurons;
[0042] Step (2) uses the I / Q signal as basic data and converts it into a pulse sequence signal that can be processed by the SCNN model through a pulse coding layer;
[0043] Step (3) inputs the pulse sequence signal into the SCNN model to perform an electromagnetic signal recognition simulation experiment and output the recognition accuracy.
[0044] like Figure 2 、 3 As shown in Figure 4, in step (1), this solution chooses to use the LIF model as the neuron function. The operating mechanism of the LIF neuron model mainly includes two stages: membrane potential update and pulse response. In the absence of external stimulation, the membrane potential inside the neuron will gradually decrease due to the leakage effect and approach its resting state. When the signal released by other neurons is input in the form of synaptic current, the current will cause the membrane potential to rise. If the input is strong enough to raise the membrane potential to the preset threshold potential, the neuron will emit pulses, simulating the action potential release behavior in the biological nervous system. Once the neuron completes the discharge, the membrane potential will be quickly reset to a value lower than the resting potential to simulate the "absolute refractory period" state that the biological neuron experiences after discharge. During this stage, the neuron temporarily no longer responds to new stimuli. Subsequently, the membrane potential slowly recovers under the combined influence of leakage and external input, gradually returning to the resting potential level, preparing for processing subsequent inputs.
[0045] If a capacitor C is used to represent the neuron membrane, I(t) represents the input current of the neuron, and q represents the charge of the membrane, then we have
[0046] q=∫I(t)dt
[0047] The charge on the cell membrane will leak over time and can be replaced by a resistor R. According to Kirchhoff's current law, the following formula can be obtained:
[0048]
[0049] where u reset is the resting potential, I R and I C is the current corresponding to the resistance and capacitance. Here, a decay constant τ=RC is set, and we can get
[0050]
[0051] If the current is a constant current I0 and the membrane potential is initially the resting potential, the above differential equation can be solved to obtain
[0052]
[0053] The LIF model simulates the working process of biological neurons well and has low computational complexity, making it suitable for use as a pulse neural network model in different scenarios.
[0054] Therefore, step (1) in this embodiment includes:
[0055] Step (101) introduces the LIF model as the model neuron function, and its discrete form dynamic equation is as follows:
[0056]
[0057] Step (102) uses tdBN with an additional time dimension instead of traditional BN;
[0058] Specifically, tdBN optimizes the standard BN from two aspects. The first aspect is the optimization of the time dimension. Compared with ANN, SNN has an additional time dimension, and information is transmitted from one moment to the next. Therefore, tdBN normalizes the input in both time and space dimensions, which to a certain extent avoids the time dependence that may occur in traditional BN when processing time series. The second aspect is that tdBN modifies the normalized variance. In BN, the value before activation is normalized to N(0,1). In tdBN, the value before activation will be normalized to N(0,(αV th ) 2 ); where V th As the threshold, it helps to capture the statistical characteristics of the signal on the local time scale and identify the dynamic changes of the signal.
[0059] Mathematically, tdBN can be seen as a process of modifying the parameters of the convolutional layer. The relationship between the weight W' and bias B' after tdBN and the weight W and bias B of the upper convolutional layer can be expressed by the following formula:
[0060] Mathematically, tdBN can be seen as a process of modifying the parameters of the convolutional layer. The relationship between the weight W' and bias B' after tdBN and the weight W and bias B of the upper convolutional layer can be expressed by the following formula: ;in represents the expectation of the data, α 2 represents the variance of the data, λ represents the weight of the tdBN layer, β represents the bias of the tdBN layer, and ε is a very small number, which defaults to 1e-5 to prevent division by zero errors.
[0061] where μ represents the expectation of the data, σ 2 represents the variance of the data, λ represents the weight of the tdBN layer, β represents the bias of the tdBN layer, ∈ is a very small number, the default value is 1e-5, which is used to prevent zero division errors.
[0062] Step (103) uses CBAM attention after each convolutional layer to improve the feature extraction ability of the convolutional layer;
[0063] Specifically, a modified CBAM attention mechanism is used to enhance the feature extraction capabilities of the convolutional layer. The modified CBAM consists of CAM and SAM, and all layers have an additional time dimension. CAM first performs max pooling and average pooling on the upper layer output, keeping the channel dimension unchanged. Then, a multi-layer perceptron and neuron functions are used to extract valuable information from the channel. Finally, the two are added together to obtain the channel weights, which are then multiplied with the original data. SAM first performs max pooling and average pooling on the upper layer output, keeping the size unchanged, and then concatenates them. Then, a convolutional layer and neuron functions are used to obtain spatial weights, which are finally multiplied with the original data.
[0064] CAM can increase the weight of channels that are useful to the results and reduce the weight of channels that are useless to the results. SAM can increase the weight of features that are useful to the results and reduce the weight of features that are useless to the results. By using the combination of CAM and SAM, SCNN can pay more attention to the key features that can identify the signal, thereby improving the perception ability of the SCNN model and improving the recognition accuracy.
[0065] Step (104) uses a residual structure to prevent the model from experiencing a vanishing gradient problem;
[0066] Specifically, each pair of neuron-convolution-CBAM-tdBN structures is connected by a shortcut structure. The pre-convolution data and post-convolution data of the two groups are added together to form a residual structure. By introducing a residual structure, data from deeper layers of the network can be fused with data from shallower layers, avoiding the vanishing gradient problem without affecting feature extraction. It also allows for a more stable increase in the number of network layers, thereby improving model performance.
[0067] Through the above series of methods, we constructed the SCNN model.
[0068] like Figure 2 and 4 As shown, step (2) in this embodiment includes:
[0069] Step (201) constructs a pulse coding layer;
[0070] Step (202) pulse encodes the I / Q signal.
[0071] Specifically, step (201) is as follows: first, a convolution layer, a CBAM layer, and a tdBN layer are used to extract features from the I / Q signal. Then, the output result is copied into two copies, one of which is subjected to the neuron-convolution-CBAM-tdBN operation. In this process, the I component and the Q component are fused, and the original 2×128 I / Q signal is converted into a 1×128 output through this part of the network. The other copy is only subjected to the convolution operation for resizing to ensure that the output result is 1×128. Finally, the two output results are added and fed into the neuron function to obtain the encoded pulse sequence signal, completing the construction of the pulse coding layer.
[0072] The specific step (202) of this embodiment is as follows: the pulse coding layer does not rely on manually set filters for feature extraction. It can update parameters together with the pulse neural network model during the training process and perform different encodings for different input signals, thereby improving the robustness of the encoding strategy. The pulse coding layer is highly flexible in design and can integrate convolution operations, fully connected structures or other differentiable function modules. It generates corresponding pulse outputs by performing weighted summation and other operations on continuous numerical inputs. Since the pulse coding layer directly processes the original continuous value data, it can retain the information in the signal more completely than traditional encoding schemes. In addition, the embedded convolution or fully connected structure not only has the encoding function, but also can realize feature extraction during the training process, effectively improving the learning efficiency and recognition accuracy of subsequent hierarchical networks. Therefore, the pulse coding layer has been widely used in many types of complex tasks, such as image recognition, speech processing and other artificial intelligence fields. In view of its superior representation ability, the present invention also selects this type of coding structure to complete the signal coding task.
[0073] The recognition result described in step (3) of this embodiment is measured by the average accuracy, that is, the percentage of correctly predicted samples to the total number of samples. This is also one of the most commonly used measurement standards in electromagnetic signal recognition tasks. After experiments, the average recognition accuracy of the present invention on the open source dataset RadioML2016.10b dataset is 64.05%. When the signal-to-noise ratio is 18dB, the recognition accuracy is the highest, which is 93.38%. The recognition accuracy is comparable to that of commonly used artificial neural network models. The model calculation energy consumption is obtained by multiplying the number of synaptic operations by the calculation energy consumption of a single addition in the hardware. The parameter amount of the present invention is 0.176M, which is lower than that of commonly used artificial neural network models, and the energy consumption of single signal recognition is reduced by more than 60% compared with other commonly used artificial neural network models.
[0074] In summary, this solution addresses the problems of high parameter count and high computational energy consumption in traditional deep learning methods, explores the application of brain-like pulse neural networks in electromagnetic signal recognition tasks, uses pulse coding layers as a temporal coding method, integrates multiple technologies to improve network recognition capabilities, and proposes an electromagnetic signal recognition architecture based on backpropagation pulse neural networks. While ensuring recognition accuracy, it reduces the number of network parameters, saves computational energy consumption, and solves the problem of performing electromagnetic signal recognition tasks in energy-constrained edge hardware scenarios.
[0075] The above description is only for explaining the embodiments of the present invention and is not intended to limit the present invention. For those skilled in the art, any modifications, equivalent substitutions, improvements, etc. made within the spirit and principles of the present invention should be included in the scope of protection of the present invention.
Claims
1. An electromagnetic signal recognition technology based on pulse neural network, characterized in that: include: Step (1) Integrate tdBN normalization technology, CBAM attention mechanism and residual structure to build a SCNN model based on LIF neurons; Step (2) uses the I / Q signal as basic data and converts it into a pulse sequence signal that can be processed by the SCNN model through a pulse coding layer; Step (3) inputs the pulse sequence signal into the SCNN model to perform an electromagnetic signal recognition simulation experiment and output the recognition accuracy.
2. The electromagnetic signal recognition technology based on pulse neural network according to claim 1 is characterized in that: The step (1) is specifically as follows: Step (101) introduces the LIF model as the model neuron function, and its discrete form dynamic equation is as follows: Step (102) uses tdBN with an additional time dimension instead of traditional BN; Step (103) uses CBAM attention with an additional time dimension after each convolutional layer to improve the feature extraction ability of the convolutional layer; Step (104) uses a residual structure to prevent the model from experiencing gradient vanishing problems.
3. The electromagnetic signal recognition technology based on pulse neural network according to claim 2 is characterized in that: The step (102) is specifically as follows: tdBN optimizes the standard BN from two aspects: The first aspect is the optimization of the time dimension. Compared with ANN, SNN has an additional time dimension, and information is transmitted from one moment to the next. Therefore, tdBN normalizes the input in both time and space dimensions, to a certain extent avoiding the time dependency that may occur in traditional BN when processing time series. The second aspect is that tdBN modifies the normalized variance. In BN, the value before activation is normalized to N(0,1); in tdBN, the value before activation will be normalized to N(0,(αV th ) 2 ); where V th As the threshold, it helps to capture the statistical characteristics of the signal on the local time scale and identify the dynamic changes of the signal.
4. The electromagnetic signal recognition technology based on pulse neural network according to claim 2 is characterized in that: The step (103) is specifically as follows: using a modified CBAM attention mechanism to enhance the feature extraction capability of the convolutional layer; the modified CBAM is composed of CAM and SAM, and all layers have an additional time dimension; CAM first performs maximum pooling and average pooling on the upper layer output while preserving the channel dimension. It then uses a multi-layer perceptron and neuron function to extract valuable information from the channel. Finally, the two are added together to obtain the channel weight and multiplied with the original data. SAM first performs size-invariant maximum pooling and average pooling on the upper layer output and concatenates them, then uses convolutional layers and neuron functions to obtain spatial weights, and finally multiplies them with the original data.
5. The electromagnetic signal recognition technology based on pulse neural network according to claim 2 is characterized in that: The step (104) is specifically as follows: each two groups of neuron-convolution-CBAM-tdBN structures are connected by a shortcut structure before and after, and the data before and after the convolution of the two groups are added to form a residual structure.
6. The electromagnetic signal recognition technology based on pulse neural network according to claim 1 is characterized in that: The step (2) is specifically as follows: Step (201) constructs a pulse coding layer; Step (202) pulse encodes the I / Q signal.
7. The electromagnetic signal recognition technology based on pulse neural network according to claim 6 is characterized in that: The step (201) is specifically as follows: first, a convolution layer, a CBAM and a tdBN layer are used to extract features of the I / Q signal; then, the output result is copied into two copies, one of which is subjected to a neuron-convolution-CBAM-tdBN operation to fuse the I component and the Q component, and the other is subjected to only a convolution operation for size change; finally, the two output results are added and fed into a neuron function to complete the construction of the pulse coding layer.
8. The electromagnetic signal recognition technology based on pulse neural network according to claim 6 or 7, characterized in that: The pulse coding layer in step (202) receives floating-point data input, performs multiplication and addition operations on synapses, and finally generates pulse outputs through the neuron layer, and can update parameters as the network is trained.
9. The electromagnetic signal recognition technology based on pulse neural network according to claim 1 is characterized in that: The recognition accuracy described in step (3) is measured by the average accuracy, that is, the percentage of correctly predicted samples to the total number of samples; the model calculation energy consumption is obtained by multiplying the number of synaptic operations by the calculation energy consumption of a single addition in the hardware.