A low-power automatic modulation recognition method based on pulse neural network

Abstract

The application discloses a low-power automatic modulation recognition method based on a pulse neural network, and belongs to the field of software radio.The method adopts normalized pulse coding based on a quantile threshold, normalizes a floating-point signal and encodes the floating-point signal into a pulse sequence signal, enhances neural network generalization, and reduces pulse neural network calculation complexity and energy consumption; a dynamic timing training test scheme is adopted, a long timing training method is used to improve the memory capacity and convergence speed of pulse neural network neurons, a short timing test method is used to reduce network reasoning time delay and energy consumption in actual application; a pulse neural network technology oriented to residual connection is adopted, the degradation problem of a deep pulse neural network is relieved, portability is enhanced, and quick conversion of a traditional neural network and the pulse neural network is facilitated.The application is suitable for the field of software radio, and realizes low-power, high-precision and real-time automatic modulation recognition.

Description

Technical Field

[0001] This invention relates to a low-power automatic modulation identification method based on a pulse neural network, belonging to the field of software radio. Background Technology

[0002] Automatic modulation identification (AMI) algorithms are used to automatically identify the modulation type of received signals. Their applications are wide-ranging, including but not limited to spectrum management, interference identification, and electronic warfare. In military applications, the electromagnetic spectrum environment is complex, with a wide range of signal frequencies and diverse modulation methods. Communication processes must ensure the secure transmission and reception of friendly signals while simultaneously locating, identifying, and jamming enemy signals. In this context, intelligent signal processing technology, with AMII algorithms at its core, is crucial for tactical operations such as electronic warfare. In civilian scenarios, identifying the modulation method of signals with unknown operating modes is also a significant issue in communication systems, especially in software-defined radio and cognitive radio systems.

[0003] In recent years, with the rise of deep learning algorithms, automatic modulation recognition algorithms based on deep learning have gradually become dominant. Modulation recognition can be viewed as a classification problem, and therefore, deep learning methods can be used to train a classifier to solve this problem. The core of deep learning methods lies in automatic feature learning. Its learning process is based on a large amount of data, and the network parameters are gradually adjusted through the backpropagation algorithm to enhance the network's ability to represent features. Deep learning methods can automatically extract features with significant differences and helpful for classification from the original input, eliminating the need for manual feature selection or complex signal preprocessing. Compared with traditional methods, deep learning methods have significant advantages because the network can learn low-level features that are transparent to humans, sometimes achieving better results than traditional methods.

[0004] However, existing automatic modulation recognition (AM) algorithms are based on second-generation neural networks. Their inference process involves numerous floating-point multiplication and addition operations, relying heavily on high-performance graphics processing units (GPUs). Therefore, they typically employ a client-server architecture, transmitting signals collected by edge devices to a computing center, which then processes and returns the modulation recognition results. This client-server architecture not only increases the processing latency of AM algorithms and raises the risk of data privacy breaches, but also easily exposes our computing center, failing to meet the real-time and security requirements of AM algorithms in both military and civilian applications. To address these issues, AM algorithms need to be deployed directly on edge devices. However, the high inference speed of second-generation neural networks places high demands on computing resources and power consumption, imposing a significant burden on edge devices. Therefore, low-power and low-resource-consumption AM algorithms are gradually becoming a new development requirement in signal processing and edge device deployment. The third-generation Spike Neural Network (SNN) based on neuromorphic computing uses the spatiotemporal information of pulse trains and simplifies floating-point multiplication and addition operations to floating-point addition operations by utilizing the binary nature of pulse trains. Compared with the second-generation traditional neural network, it has lower energy consumption and less computing power requirements, and has been proven to have similar accuracy to traditional neural networks in the field of modulation recognition. Summary of the Invention

[0005] To address the issues of existing automatic modulation recognition algorithms relying on high-performance graphics processing units and high power consumption, the main objective of this invention is to propose a low-power automatic modulation recognition method based on spiking neural networks. By employing normalized pulse coding based on quantile thresholds, a dynamic time-series training and testing scheme, and spiking neural network technology oriented towards residual connections, this method reduces the power consumption, inference speed, and computational resource requirements of the neural network, which is beneficial for the deployment of edge devices and enables low-power, high-precision real-time automatic modulation recognition.

[0006] The objective of this invention is achieved through the following technical solution:

[0007] This invention discloses a low-power automatic modulation recognition method based on spiking neural networks. It employs normalized pulse coding based on quantile thresholds to normalize and encode floating-point signals into pulse sequence signals. While ensuring information integrity and accuracy, the normalization method enhances the generalization ability of the neural network, helping it learn the intrinsic characteristics of signals under different modulation modes rather than invalid amplitude features. Pulse coding reduces the computational complexity and energy consumption of spiking neural networks. A dynamic time-series training and testing scheme is adopted, using long-series training to improve the memory capacity and convergence speed of spiking neural network neurons, and short-series testing to reduce network inference latency and energy consumption in practical applications. The method utilizes residual-connection-oriented spiking neural network technology to alleviate the degradation problem of deep spiking neural networks, exhibiting strong portability and facilitating rapid conversion between traditional neural networks and spiking neural networks. This achieves low-power, high-precision real-time automatic modulation recognition.

[0008] This invention discloses a low-power automatic modulation identification method based on a pulse neural network, comprising the following steps:

[0009] Step 1: Acquire signals from known operating modes and downconvert them to IQ signals;

[0010] Radio frequency signal data in known operating modes are acquired using radio frequency signal acquisition equipment. The acquired raw data is multiplied by a carrier wave with a 90° phase difference, and then passed through a low-pass filter to obtain in-phase and quadrature (IQ) signals.

[0011] Step 2: Encode the IQ signal using normalized pulse coding based on quantile thresholds to obtain a binary pulse sequence, and then divide it to obtain the basic input unit;

[0012] The basic principle of the normalized pulse coding based on quantile thresholds is binary quantization coding, which converts the floating-point IQ signal obtained in step one into a binary pulse sequence. The specific steps are as follows:

[0013] 2.1 Calculate the positive and negative K quantiles x for all IQ signals. 1-K With x K , where K takes values ​​in the range [0,1]. 1-K Let x be the 1-K quantile of all negative IQ signals. 1-K Only IQ signals with negative values ​​that are greater than (1-K). K The K quantile of all positive IQ signals, i.e., x K Only IQ signals with positive values ​​greater than K are considered. The IQ signal normalization method is as follows: The [0,1] constraint indicates that after normalization... If it is less than 0, then let If it is 0, If it is greater than 1, then let =1; x i This represents all IQ signals, where i is a natural number.

[0014] 2.2. Normalize the IQ signal x i Through the N-bit binary quantizer Q N ((x i -x min ) / (x max -x min )*2 N ):R→B N Encoded as N-bit binary data, where x max x represents the maximum value of the normalized IQ signal. min This represents the minimum value of the normalized IQ signal, and the output of the binary quantizer is the encoding result of this step. The basic input unit of the network is M consecutive pairs of encoded IQ signals, with a dimension of N*2*M.

[0015] Step 3: Improve the traditional residual module to obtain the pulse residual connection module, and build a pulse neural network based on residual connection;

[0016] First, to simplify network computation, the LIF (Leaky Integrated and Fired) neuron model is chosen as the basic unit of the spiking neural network. The input and output of the LIF neuron model are defined as follows: when the input is input, the membrane potential threshold is TH, and the decay hyperparameter is β, the output of the current time t is determined by O(t) = g(input + βv(t-1)), and the membrane potential of the current time t is determined by v(t) = input + βv(t-1) - O(t), where g(.) is a 0-1 step function, and the output is 1 when the input is greater than TH.

[0017] Secondly, an improved pulse residual connection module is obtained based on the traditional residual connection module. The specific improvement methods are as follows: the activation function layer of the traditional residual module is replaced with a LIF neuron model; the LIF neuron model between layers is placed in the residual branch, and the layers are directly connected; the addition operation of the traditional residual module is replaced with an element-wise AND operation.

[0018] Finally, the multi-layer stacked pulse residual connection module yields a spiking neural network based on residual connections.

[0019] Step 4: Train the spiking neural network based on residual connections using the encoded basic input units and the gradient substitution backpropagation method;

[0020] 4.1 Forward Propagation Process of the Spiking Neural Network: The forward propagation process of the spiking neural network consists of T forward propagations, where T represents the temporal sequence of the spiking neural network. The input for each forward propagation is the basic input unit obtained in step two, and the spiking neural network is the residual-connected spiking neural network obtained in step three. The parameters of the spiking neural network are identical within each forward propagation, relying on the membrane potential of the LIF neuron model to transmit information. The output of the spiking neural network forward propagation is the sum of the outputs of all temporal forward propagations. The value of the temporal sequence T has a linear relationship with the inference speed of the spiking neural network; as the value of T increases, the inference latency of the spiking neural network increases linearly. During the training phase, T = 2N is set to obtain better training results; during the testing phase, T = N / 2 is set to reduce the inference latency of the spiking neural network.

[0021] 4.2 Backpropagation Process of Spiking Neural Networks: Backpropagation in spiking neural networks is an improvement on traditional backpropagation. Except for backpropagation to the LIF neuron model output O(t) = 1, the other backpropagation operations are the same as in traditional neural networks. The following only describes the steps of backpropagating from a spiking neural network to the LIF neuron model.

[0022] When a spiking neural network backpropagates to the LIF neuron model output O(t) = 1, this process is represented by the impulse function f(in) = δ(in), where δ(in) is 1 only when in = 0, and in is the input of the current layer. Since the function f(in) is not differentiable, backpropagation cannot proceed, and training cannot be performed using traditional neural network methods. To address the problem of non-differentiability preventing backpropagation, the gradient substitution method keeps the impulse function f(in) unchanged during forward propagation and manually defines a function during backpropagation. The graph of the function S(in) approximates the impulse function by replacing the impulse function f(in), and it has the property of being continuously differentiable. Its differential... The gradient substitution method can replace the differentiation of the impulse function during backpropagation, solving the problem of backpropagation failure caused by the non-differentiability of the impulse function. In the backpropagation process, the gradient substitution method uses a differentiable function that approximates the impulse to replace the pulse sending process, making the computational flow of the backpropagation process the same as that of traditional neural networks.

[0023] 4.3 Training process of spiking neural network: In the training stage of spiking neural network, based on the basic input unit after pulse coding in step two, the above forward propagation and backward propagation processes are performed sequentially until the network parameters converge, and a spiking neural network that can be used for automatic modulation pattern recognition is obtained.

[0024] It also includes step five, which uses the pulse neural network for automatic modulation pattern recognition obtained in step four to be applied in signal processing scenarios. This can identify the modulation pattern of unknown radio frequency signals, realize low power consumption, high energy efficiency and low latency automatic modulation recognition tasks, and provide signal working mode information for the subsequent processing of unknown radio frequency signals.

[0025] Beneficial effects:

[0026] 1. This invention discloses a low-power automatic modulation recognition method based on a pulse neural network. Based on quantile threshold normalized pulse coding, it calculates the positive and negative quantiles of the input data, normalizes the floating-point IQ signal in the [0,1] interval, and then encodes it into a pulse sequence signal with multiple discrete 0 / 1 values, ensuring information integrity and accuracy while reducing computational complexity. The quantile threshold normalization method enhances the generalization ability of the neural network, helping the network learn the intrinsic characteristics of signals under different modulation modes rather than invalid amplitude characteristics. Through binary conversion, a single floating-point signal with continuous values ​​is represented by a multi-bit 0 / 1 pulse signal, achieving near-lossless pulse conversion of the original data. Compared with existing rate coding, delay coding, and quantization coding methods, it offers higher coding accuracy, significantly reduced delay, and advantages of low power consumption and high energy efficiency.

[0027] 2. This invention discloses a low-power automatic modulation recognition method based on spiking neural networks (SLNNs). It employs a residual connection structure for SLNNs, where the SLNN consists of numerous interconnected LIF neuron models. Information is transmitted between network layers via pulse trains. The residual connection structure combines the input and output pulse signals of a certain layer as the input to the next layer. The merging methods include, but are not limited to, AND, XOR, and addition operations. To better mitigate network degradation, the spiking residual connection module adjusts the position of the LIF neuron models within the network. Compared to existing traditional residual connection modules and SLNNs without residual connections, the spiking residual connection module increases the maximum number of layers in deep SLNNs, enhancing their expressive and generalization capabilities, while retaining the low-power and high-efficiency advantages of SLNNs.

[0028] 3. This invention discloses a low-power automatic modulation recognition method based on spiking neural networks. It uses spiking neural networks as classifiers and trains them through gradient substitution methods. It makes full use of the low power consumption, high energy efficiency and high accuracy characteristics of spiking neural networks, which is more conducive to deploying automatic modulation recognition algorithms on resource-constrained edge devices. Compared with existing automatic modulation recognition algorithms based on traditional neural networks, it reduces resource consumption and power consumption while maintaining similar recognition accuracy.

[0029] 4. This invention discloses a low-power automatic modulation recognition method based on a spiking neural network. The method employs a general spiking neural network architecture that can be converted from a traditional neural network architecture at a low cost. It only requires adding a pulse coding layer, replacing the traditional residual module with a pulse residual connection module, and replacing the activation function with a LIF neuron model. Subdividing the complete spiking neural network into combinations of pulse residual connection modules facilitates task transfer and fine-tuning of the spiking neural network, achieving a low-power, high-efficiency spiking neural network.

[0030] 5. This invention discloses a low-power automatic modulation recognition method based on spiking neural networks. By setting the training latency of the spiking neural network to twice the pulse code length, the memory capacity of the neurons in the spiking neural network can be improved, training data can be fully utilized, and the network convergence speed can be increased. It can also compensate for the accuracy loss caused by using a shorter latency in the inference stage. Setting the inference and testing latency to half the pulse code length can reduce the inference latency to one-quarter of that in the training stage, significantly reducing the inference latency while only reducing the inference accuracy by 1%. This achieves the goal of improving training effect and reducing the inference latency and energy consumption of the spiking neural network. Attached Figure Description

[0031] Figure 1 This is a flowchart illustrating a low-power automatic modulation recognition method based on a pulse neural network disclosed in this invention.

[0032] Figure 2 This is a schematic diagram of the spiking neural network based on residual connections disclosed in this embodiment;

[0033] Figure 3 This embodiment discloses a modulation recognition algorithm based on a pulse neural network.

[0034] Figure 4 This is a structural diagram of the pulse residual connection module disclosed in this embodiment;

[0035] Figure 5 This example compares the errors of different pulse coding methods. Detailed Implementation

[0036] The present invention will now be described in detail with reference to the accompanying drawings and embodiments. The technical problems solved by the present invention and its beneficial effects are also described. It should be noted that the described embodiments are only intended to facilitate understanding of the present invention and do not constitute any limitation thereof.

[0037] In commercial communication systems, modulation scheme identification of signals with unknown modulation schemes is a core technology for signal demodulation. This embodiment discloses a low-power automatic modulation identification method based on a pulse neural network, which was implemented on the RadioML2016 10a dataset. This dataset was obtained by simulating channel data with 11 modulation schemes and 20 signal-to-noise ratios based on real channels using GNU radio software radio, and then down-converting the data to obtain an IQ quadrature modulation dataset. The accuracy of modulation scheme identification on this dataset confirms the performance of this method in the automatic modulation identification task.

[0038] The specific flowchart of this embodiment is as follows: Figure 1 As shown, the RadioML2016 10a dataset is divided into training, validation, and test sets in a 6:1:3 ratio. The training and validation sets are used as... Figure 1 The input for training a spiking neural network is the test set. Figure 1 Unknown operating mode signal input.

[0039] Step 1: Acquire signals from known operating modes and downconvert them to IQ signals;

[0040] The RML2016 10a dataset uses music data as its raw data and modulates it into radio frequency signals using eleven modulation schemes: 8PSK, AM-DSB, AM-SSB, BPSK, CPFSK, GFSK, 4-PAM, 16-QAM, 64-QAM, QPSK, and WBFM. After adding noise through a simulated channel, each signal is multiplied by a carrier wave with a 90° phase difference and then passed through a low-pass filter to obtain in-phase and quadrature (IQ) signals. The modulation schemes of all IQ signals in the training and validation sets are known, while the modulation schemes of all IQ signals in the test set are unknown.

[0041] Step 2: Encode the IQ signal using normalized pulse coding based on quantile thresholds to obtain a binary pulse sequence, which is then segmented to obtain the basic input unit;

[0042] The basic principle of the normalized pulse coding based on quantile threshold is binary quantization coding, which converts the floating-point IQ signal obtained in step one into a binary pulse sequence.

[0043] First, calculate the positive and negative 95th percentiles x for all IQ signals. 5％ With x 95％ x 5％ x is the 5th percentile for all negative input data. 95％ The 95th percentile for all positive input data. The IQ signal normalization method is as follows: x iThis represents the IQ signal, where i is a natural number; the [0,1] constraint indicates that after normalization... If it is less than 0, then let If it is 0, If it is greater than 1, then let =1;

[0044] Secondly, normalize the IQ signal Through an 8-bit binary quantizer Q8((x) i -x min ) / (x max -x min )*2 8 ):R→B 8 Encoded as 8-bit binary data, where x max This represents the maximum value of the normalized IQ signal, which is set to 1 here. min This represents the minimum value of the normalized IQ signal, which is set to 0 here. The output of the binary quantizer is the encoding result of this step. The basic input unit of the network is 128 consecutive pairs of encoded IQ signals, with a dimension of 8*2*128.

[0045] Step 3: Improve the traditional residual module to obtain the pulse residual connection module, and build a pulse neural network based on residual connection;

[0046] The LIF (Leaky Integrated and Fired) neuron model is chosen as the basic unit of the spiking neural network. The input and output of the LIF neuron model are defined as follows: when the input is input, the membrane potential threshold is TH = 1, and the decay hyperparameter is β = 0.7, the output of the current time t is determined by O(t) = g(input + βv(t-1)), and the membrane potential of the current time t is determined by v(t) = input + βv(t-1) - O(t), where g(.) is a 0-1 step function, and the output is 1 when the input is greater than TH.

[0047] The spiking neural network used for the RML2016 10a dataset has the same architecture as traditional neural networks, such as... Figure 2 As shown, the basic input unit outputs the prediction result after passing through a pulse coding layer, a spiking neural network intermediate layer, and a fully connected layer. Information is transmitted between layers via pulse sequences; the pulse residual connection module is as follows... Figure 4 As shown, the LIF neuron model between layers is placed within the residual branch; the traditional residual module addition operation is replaced with an element-wise AND operation; with the spiking residual connection module as the core of the spiking neural network, multiple layers are stacked to obtain a spiking neural network based on residual connections. The overall network architecture is as follows. Figure 3 As shown in the figure, conv(m,n) represents a convolutional layer with m input channels and n output channels.

[0048] Step 4: Train the residual connection-based spiking neural network using the encoded basic input units and the gradient substitution backpropagation method.

[0049] 4.1 Forward Propagation Process of the Spiking Neural Network: The forward propagation process of the spiking neural network consists of T forward propagations, where T represents the temporal sequence of the spiking neural network. The input for each forward propagation is the basic input unit obtained in step two, and the spiking neural network is the residual-connected spiking neural network obtained in step three. The parameters of the spiking neural network are identical within each forward propagation, relying on the membrane potential of the LIF neuron model to transmit information. The forward propagation output of the spiking neural network is the sum of the outputs of all temporal forward propagations. The value of the temporal sequence T has a linear relationship with the inference speed of the spiking neural network; as T increases, the inference latency of the spiking neural network increases linearly. During the training phase, T = 16 is set to obtain better training results; during the testing phase, T = 4 is set to reduce the inference latency of the spiking neural network.

[0050] 4.2 Backpropagation process of spiking neural network: The backpropagation of spiking neural network is an improvement on the traditional backpropagation operation. Except for backpropagation to the output O(t) = 1 of the LIF neuron model, the other backpropagation operations are the same as those of traditional neural network.

[0051] The traditional backpropagation (BP) algorithm for neural networks is a learning algorithm suitable for multi-layered neural networks. It is based on the gradient descent method. Gradient descent is a common method for training neural networks. It calculates the gradient of the loss function with respect to the network parameters and feeds this gradient back to the optimization function to update the network parameters and minimize the loss function.

[0052] When the spiking neural network backpropagates to the LIF neuron model output O(t) = 1, this process is represented by the impulse function f(in) = δ(in), where δ(in) is 1 only when in = 0, and in represents the input of the current layer. Since the function f(in) is non-differentiable, gradient calculation cannot be performed, making training impossible using traditional neural network methods. To address the problem of non-differentiability preventing backpropagation, this embodiment employs a gradient substitution method that keeps the forward propagation impulse function f(n) unchanged, while manually defining a function during backpropagation. The graph of the function S(in) approximates the impulse function by replacing the impulse function f(in), and it has the property of being continuously differentiable. Its differential... The gradient substitution method can replace the differentiation of the impulse function during backpropagation, solving the problem of backpropagation failure caused by the non-differentiability of the impulse function. In the backpropagation process, the gradient substitution method uses a differentiable function that approximates the impulse to replace the pulse sending process, making the computational flow of the backpropagation process the same as that of traditional neural networks.

[0053] 4.3 Spiking Neural Network Training Process: During the training phase, the basic input units from the training and validation sets obtained in step two (pulse encoding) are repeatedly input into the network according to the time delay requirements. The membrane potential of each delayed neuron is calculated based on neuronal dynamics and the LIF neuron model, and forward propagation of the spiking neural network is performed. After the forward propagation process is completed, the time delay dimension of the final result is accumulated to obtain the network output. The highest value of the network output is taken as the modulation mode recognition result. The spiking neural network repeats the forward and backward propagation process on the training set until the recognition accuracy of the spiking neural network obtained from the forward propagation on the validation set converges, ultimately yielding a neural network that can be used to recognize modulation modes.

[0054] Step 5: Based on the trained spiking neural network, determine the modulation method of the signal in the unknown operating mode.

[0055] The basic input units of the test set are used as data for unknown operating modes and input into the spiking neural network trained in step four for recognition. The accuracy is then compared with that of a traditional neural network. The traditional neural network, represented by ResNet-CNN, is not directly trained without encoding the input. The spiking neural network, represented by ResNet-SNN, is trained using normalized pulse coding based on quantile thresholds, pulse rate coding (pulse frequency related to the input signal), and pulse delay coding (pulse occurrence time related to the input signal), with test delays of 4 and 8 seconds. The final accuracy is shown in Table 1. Figure 5 As shown, pulse rate coding and delay coding have higher coding errors than normalized pulse coding based on quantile thresholds under the same time delay (the figure uses binary quantization coding to represent normalized pulse coding based on quantile thresholds). Therefore, at the same time delay, the accuracy of pulse rate coding and delay coding is reduced by approximately 30% and 23% respectively compared to normalized pulse coding based on quantile thresholds. The accuracy of binary pulse coding at times of 4 and 8 is only 2.5% and 1.1% lower than that of traditional neural networks, respectively, and its performance is close to that of traditional neural networks.

[0056] During network training, we observed high sparsity in the release of pulses by neurons, defining the pulse rate. Spike represents the total number of pulses during the forward propagation, and Neuron represents the total number of neurons in the spiking neural network. Therefore, the spiking rate (SP) represents the number of pulses generated by each neuron per time delay. A batch of data was randomly selected from the dataset to estimate the spiking rate, and the results are shown in Table 2. For a spiking neural network with a time delay of 4, SR*T = 57.6%, meaning that only about one pulse is generated between every two neurons in one complete forward propagation. For a spiking neural network with a time delay of 8, SR*T = 116.8%, meaning that only about one pulse is generated per neuron in one complete forward propagation. The spiking rate experiment demonstrates the high sparsity of the spiking neural network. To further demonstrate its energy efficiency, we estimated the energy cost based on the number and type of 32-bit floating-point operations in the 45nm technology [Horowitz M.1.1 computing's energy problem (and what we can do about it) [C] / / 2014 IEEE International Solid-State Circuits Conference Digest of Technical Papers (ISSCC). IEEE, 2014: 10-14.]. One multiply-accumulate (MAC) operation consumes 4.9 pJ of energy, and one add (AC) operation consumes 0.9 pJ of energy. The energy cost estimation focuses on the convolutional and fully connected layers in the residual path, which constitute the network's floating-point operations (FLOPs). This estimation does not include the quantization coding layer, batch normalization layer, and pooling layer that process the input data, as they do not involve multiply-accumulate or add operations; the energy cost is the same for spiking neural networks and traditional neural networks. Table 2 shows the energy consumption of a single image in different networks. Given the sparsity and short latency of spiking neural networks (SNNs), ResNet-SNN can achieve computation using almost the same number (T=8) or about half (T=4) of synaptic operations (SyOPs) instead of FLOPs. This corresponds to each neuron emitting an average of only about one spiking signal or two neurons emitting about one spiking signal. For a single input signal, when the latency is 8, the energy cost of a traditional neural network is 4–5 times that of a spiking neural network; when the latency is reduced to 4, the energy cost is 8–9 times that of a spiking neural network with the same structure. Different encoding methods have little impact on energy efficiency. When processing spatiotemporal streaming data, traditional neural networks need to introduce recurrent connections to remember previous states. Therefore, both networks will spend multiple latency periods to complete the inference task, so latency no longer affects the energy efficiency ratio, and the low energy consumption advantage of spiking neural networks becomes more prominent. The above conclusions highlight the low power consumption and high energy efficiency characteristics of spiking neural networks.

[0057] Table 1 Comparison of recognition accuracy between CNN and SNN

[0058]

[0059] Note: T represents the test delay.

[0060] Table 2. Energy efficiency comparison between CNN and SNN (processing one frame of image)

[0061]

[0062]

[0063] In summary, this embodiment discloses a modulation recognition algorithm based on a spiking neural network, mainly illustrating the encoding and training methods and comparing their energy consumption with that of traditional neural networks. First, normalized pulse coding based on quantile thresholds is used to encode the input IQ signal, reducing network energy consumption. Second, to obtain a low-latency spiking neural network, gradient substitution and dynamic temporal quantization encoding methods are used for direct training. Third, to address the degradation problem of deep convolutional networks, residual connections for spiking neural networks are used to achieve accuracy similar to that of equivalent traditional neural networks, with energy consumption only one-eighth to one-quarter of that of traditional neural networks. Furthermore, the method disclosed in this invention does not require special network design; simply replacing the activation function with neuron layers allows for good accuracy using a general network architecture. It is highly portable, fully leveraging the low-power advantage of spiking neural networks, and possesses advantages such as simple design, high energy efficiency, and ease of edge deployment.

[0064] The above detailed description further illustrates the purpose, technical solution, and beneficial effects of the invention. It should be understood that the above description is merely a specific embodiment of the present invention and is not intended to limit the scope of protection of the present invention. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the scope of protection of the present invention.

Claims

1. A low-power automatic modulation recognition method based on a spiking neural network, characterized in that: Includes the following steps, Step 1: Acquire signals from known operating modes and downconvert them to IQ signals; Step 2: Encode the IQ signal using normalized pulse coding based on quantile thresholds to obtain a binary pulse sequence, and then divide it to obtain the basic input unit; Step 3: Improve the traditional residual module to obtain the pulse residual connection module, and build a pulse neural network based on residual connection; Step 4: Train the spiking neural network based on residual connections using the encoded basic input units and the gradient substitution backpropagation method; The implementation method for step four is as follows: 4.1 Forward Propagation Process of Spiking Neural Network: The forward propagation process of the spiking neural network consists of T forward propagations, where T represents the timing sequence of the spiking neural network. The input for each forward propagation is the basic input unit obtained in step two, and the spiking neural network is the residual connection-based spiking neural network obtained in step three. The parameters of the spiking neural network are the same in each forward propagation, and information is transmitted by membrane potentials of the LIF neuron model. The output of the spiking neural network forward propagation is the sum of the outputs of all time-series forward propagations. The value of the timing sequence T is linearly related to the inference speed of the spiking neural network. As the value of T increases, the inference delay of the spiking neural network increases linearly. During the training phase, let T=2N to achieve better training results; During the testing phase, T=N / 2 is set to reduce the inference latency of the spiking neural network; 4.2 Backpropagation Process of Spiking Neural Networks: Backpropagation in spiking neural networks is an improvement on the traditional backpropagation operation, except that it backpropagates to the output of the LIF neuron model. Apart from that, the other backpropagation operations are the same as those of traditional neural networks; the following only gives the steps for backpropagating from a spiking neural network to a LIF neuron model. When the spiking neural network backpropagates to the LIF neuron model output At that time, the process uses the impulse function express, Only When it is 1, ; due to the function Non-differentiability prevents backpropagation and training using traditional neural network methods. To address this issue, gradient substitution methods maintain forward propagation by sending impulse functions. The function remains unchanged and is manually defined during backpropagation. Alternate impulse function ,function The image approximates the impulse function, which has the property of being continuously differentiable, and its differential... It can replace the differential of the impulse function during backpropagation, solving the problem of backpropagation failure caused by the non-differentiability of the impulse function; the gradient substitution method uses a differentiable function of an approximate impulse to replace the pulse sending process during backpropagation, making the calculation process of backpropagation the same as that of traditional neural networks. 4.3 Training process of spiking neural network: In the training stage of spiking neural network, based on the basic input unit after pulse coding in step two, the above forward propagation and backward propagation processes are performed sequentially until the network parameters converge, and a spiking neural network that can be used for automatic modulation pattern recognition is obtained.

2. The low-power automatic modulation recognition method based on a spiking neural network as described in claim 1, characterized in that: It also includes step five, which uses the pulse neural network for automatic modulation pattern recognition obtained in step four to be applied in signal processing scenarios. This can identify the modulation pattern of unknown radio frequency signals, realize low power consumption, high energy efficiency and low latency automatic modulation recognition tasks, and provide signal working mode information for the subsequent processing of unknown radio frequency signals.

3. The low-power automatic modulation recognition method based on a spiking neural network as described in claim 1, characterized in that: The implementation method for step one is as follows: Radio frequency signal data in a known operating mode is acquired by radio frequency signal acquisition equipment. The acquired raw data is multiplied by a carrier wave with a 90° phase difference, and then passed through a low-pass filter to obtain in-phase and quadrature signals, i.e., IQ signals.

4. The low-power automatic modulation recognition method based on a spiking neural network as described in claim 1, characterized in that: The implementation method for step two is as follows: The basic principle of the normalized pulse coding based on quantile thresholds is binary quantization coding, which converts the floating-point IQ signal obtained in step one into a binary pulse sequence; the specific steps are as follows: 2.1 Calculate the positive and negative K quantiles of all IQ signals. and K takes the value [0,1]. The 1-K quantile of all negative IQ signals; The K quantiles of all positive IQ signals; the IQ signal normalization method is as follows: The [0,1] constraint indicates that after normalization... If it is less than 0, then let =0; if If it is greater than 1, then let =1; This represents all IQ signals, where i is a natural number; 2.2 Normalizing the IQ signal Through an N-bit binary quantizer Encoded as N bits of binary data, where This represents the maximum value of the normalized IQ signal. This represents the minimum value of the normalized IQ signal, and the output of the binary quantizer is the encoding result of this step; the basic input unit of the network is a continuous set of M pairs of encoded IQ signals, with a dimension of N*2*M.

5. The low-power automatic modulation recognition method based on a spiking neural network as described in claim 1, characterized in that: The implementation method for step three is as follows: First, to simplify network computation, the LIF neuron model is chosen as the basic unit of the spiking neural network. The input and output of the LIF neuron model are defined as follows: when the input is... The membrane potential threshold is The attenuation hyperparameter is At that time, the current time sequence The output is from Confirmed, current timing The membrane potential is determined by Confirmed, among which It is a 0-1 step function, and the output is 1 when the input is greater than TH; Secondly, an improved pulsed residual connection module is obtained based on the traditional residual connection module. The specific improvement methods are as follows: the activation function layer of the traditional residual module is replaced with a LIF neuron model; the LIF neuron model between layers is placed in the residual branch, and the layers are directly connected; the addition operation of the traditional residual module is replaced with an element-wise AND operation. Finally, the multi-layer stacked pulse residual connection module yields a spiking neural network based on residual connections.

Images