Automatic modulation identification method and system based on modular pulse neural network

Abstract

The invention discloses an automatic modulation identification method and system based on a modular spiking neural network. The method comprises the following steps: converting input I / Q signal data into a pulse sequence and inputting the pulse sequence into the modular spiking neural network; in the first stage, performing coarse classification on the pulse sequence by using the spiking neural network to confirm a large category to which a modulation signal belongs; in the second stage, the pulse neural network corresponding to the large category is activated, the amplitude and the phase of the input I / Q signal data are extracted through a signal feature enhancement and state fusion module, the extracted amplitude and phase features and the features extracted by the activated pulse neural network are spliced to generate a fusion feature vector, and a modulation recognition result is obtained through classification. The invention aims to realize accurate recognition of complex modulation signals, solve the problem that the recognition accuracy and the power consumption requirement of a traditional deep learning model in edge equipment such as an unmanned aerial vehicle are difficult to balance, and provide a lightweight and efficient automatic modulation recognition scheme for application scenes such as unmanned aerial vehicle communication and intelligent monitoring.

Description

Technical Field

[0001] The present invention relates to automatic modulation recognition (AMR) technology for embedded communication equipment such as unmanned aerial vehicles (UAVs), and in particular to an automatic modulation recognition method and system based on a modular pulse neural network. Background Art

[0002] The widespread adoption of resource-constrained devices, such as drone communications, the Internet of Things, and intelligent surveillance, has placed higher efficiency demands on automatic modulation recognition (AMR). As a critical function of embedded devices, AMR technology needs to quickly and efficiently identify multiple signal modulation types in complex electromagnetic environments to ensure communication security and smooth mission execution. However, these devices, constrained by factors such as power consumption and computing resources, struggle to meet the requirements of traditional deep learning (DL) methods. Existing convolutional neural network (CNN) and recurrent neural network (RNN) models typically require significant computing resources and high energy consumption while achieving high recognition rates, making them difficult to effectively apply in drones and other edge devices.

[0003] Spiking neural networks (SNNs), with their biomimetic computing and event-driven characteristics, offer a promising solution for efficient automatic modulation recognition in resource-constrained environments. SNNs are capable of processing information at low power consumption and are suitable for handling complex modulated signals. However, the application of spiking neural networks still faces challenges, especially in effectively encoding temporal information and reducing training complexity. Liquid state machines (LSMs), a spiking neural network architecture, achieve low-complexity computation through sparsely connected reservoir layers. However, traditional approaches fail to fully utilize the temporal characteristics of communication signals, and their static characteristics limit their performance in complex automatic modulation recognition tasks. Summary of the Invention

[0004] Technical problem to be solved by the present invention: In response to the above-mentioned problems of the prior art, an automatic modulation recognition method and system based on a modular pulse neural network are provided. The present invention aims to achieve accurate recognition of complex modulated signals, solve the problem that traditional deep learning models are difficult to balance recognition accuracy and power consumption requirements in edge devices such as drones, and provide a lightweight and efficient automatic modulation recognition solution for application scenarios such as drone communications and intelligent monitoring.

[0005] In order to solve the above technical problems, the technical solution adopted by the present invention is:

[0006] An automatic modulation recognition method based on a modular pulse neural network comprises the following steps: converting input I / Q signal data into a pulse sequence; inputting the pulse sequence into the modular pulse neural network, wherein the modular pulse neural network processes the pulse sequence in two stages, wherein the first stage uses the pulse neural network to roughly classify the pulse sequence to confirm the broad category to which the signal belongs, and activates the corresponding pulse neural network in multiple pulse neural networks in the second stage; and the second stage activates the corresponding pulse neural network, extracts the amplitude and phase of the input I / Q signal data through a signal feature enhancement and state fusion module, concatenates the extracted amplitude and phase with the extracted features of the activated pulse neural network to generate a fused feature vector, and uses the classifier of the pulse neural network for classification to obtain a modulation recognition result.

[0007] Optionally, converting the input I / Q signal data into a pulse sequence includes:

[0008] S1, initialize time step t to 0;

[0009] S2, determine that the time step t is less than the preset time window t l Is it true? If so, jump to step S3; otherwise, end;

[0010] S3, determine whether the current value v satisfies the condition v≥v th Is it true, where v th If is the threshold value of the current time step, then return to time step t and jump to step S4; otherwise, update the threshold value v of the current time step according to the following formula th :

[0011]

[0012] Among them, v th (t) is the threshold v of the current time step th , v th (0) is the initial threshold, t c is the constant that controls the decay rate;

[0013] S4, add 1 to the time step t and jump to step S2.

[0014] Optionally, before inputting the pulse sequence into the modular pulse neural network, the method also includes dividing the multiple modulation modes corresponding to the modulation identification results into N major categories according to similarity, thereby initializing the pulse neural network of the two stages of the modular pulse neural network, and the second stage includes N pulse neural networks, and each pulse neural network corresponds to a major category, and each major category corresponds to one or more modulation modes.

[0015] Optionally, the pulse neural network in the first stage is composed of a liquid state machine and a classifier, and the liquid state machine is composed of an input layer, a reservoir layer and a readout layer.

[0016] Optionally, the pulse neural network in the second stage is composed of a liquid state machine and a classifier, and the liquid state machine is composed of an input layer, a reservoir layer and a readout layer.

[0017] Optionally, the function expression for extracting amplitude and phase from the input I / Q signal data through the signal feature enhancement and state fusion module is:

[0018]

[0019] Wherein, A is the amplitude, θ is the phase, I and Q are the real signal data and imaginary signal data in the I / Q signal data respectively; the function expression for generating a fused feature vector by concatenating the extracted amplitude and phase with the extracted features of the activated spiking neural network is:

[0020] f SFEF =[f A / P ;f reservoir ],

[0021] Among them, f SFEF is the fusion feature vector, f A / P is the eigenvector composed of amplitude and phase, f reservoir represents the reservoir layer state vector of the liquid state machine, and [·;·] represents vector concatenation.

[0022] Optionally, the modulation identification result is part or all of amplitude keying, phase keying, differential phase shift keying, quadrature amplitude modulation, frequency keying, staggered quadrature phase shift keying, π / 4 quadrature phase shift keying and Gaussian filtered minimum frequency shift keying.

[0023] In addition, the present invention also provides an automatic modulation recognition system based on a modular pulse neural network, comprising a microprocessor and a memory connected to each other, wherein the microprocessor is programmed or configured to execute the automatic modulation recognition method based on a modular pulse neural network.

[0024] In addition, the present invention also provides a computer-readable storage medium, which stores a computer program or instruction, and the computer program or instruction is programmed or configured to execute the automatic modulation recognition method based on modular pulse neural network through a processor.

[0025] In addition, the present invention also provides a computer program product, including a computer program or instructions, which are programmed or configured to execute the automatic modulation recognition method based on modular pulse neural network through a processor.

[0026] Compared with the prior art, the present invention mainly has the following advantages: The automatic modulation recognition method based on modular pulse neural network of the present invention includes converting the input I / Q signal data into a pulse sequence and inputting it into the modular pulse neural network: in the first stage, the pulse sequence is roughly classified using the pulse neural network to confirm the broad category to which the signal belongs; in the second stage, the pulse neural network corresponding to the broad category is activated, and the amplitude and phase of the input I / Q signal data are extracted through the signal feature enhancement and state fusion module, and the fused feature vector is generated by splicing with the extracted features of the activated pulse neural network and classified to obtain the modulation recognition result, thereby realizing accurate recognition of complex modulated signals, solving the problem that traditional deep learning models are difficult to balance recognition accuracy and power consumption requirements in edge devices such as drones, and providing a lightweight and efficient automatic modulation recognition solution for application scenarios such as drone communications and intelligent monitoring. BRIEF DESCRIPTION OF THE DRAWINGS

[0027] Figure 1 Schematic diagram of the basic principle of the method of the embodiment of the present invention.

[0028] Figure 2 Schematic diagram of the activity principle of a single spiking neuron in an embodiment of the present invention.

[0029] Figure 3 Schematic diagram of the network structure of the pulse neural network in an embodiment of the present invention.

[0030] Figure 4 Schematic diagram of the network structure of a modular spiking neural network in an embodiment of the present invention.

[0031] Figure 5 Schematic diagram of the network structure of signal feature enhancement and state fusion (SFEF) in an embodiment of the present invention.

[0032] Figure 6 The figure compares the recognition accuracy of the proposed method and the baseline solution under different signal-to-noise ratios (SNRs).

[0033] Figure 7 This is the recognition accuracy of the method proposed in the embodiment of the present invention under different signal-to-noise ratios on the additional dataset, where (a) is the result on the RML2016.10a dataset; (b) is the result on the RML2016.10b dataset.

[0034] Figure 8Figure 1 shows the ablation experiment results of the method proposed in the embodiments of the present invention. (a) compares the recognition accuracy of the proposed method, a single SNN (single LSM) model, and the proposed method model without the SFEF module at different signal-to-noise ratios. (b) compares the recognition accuracy of different automatic modulation recognition schemes (the proposed method, AMR without SFEF, ResNet, and a single SNN) at different sample lengths.

[0035] Figure 9 The confusion matrix comparison of modulation recognition under a signal-to-noise ratio of 24 for the method proposed in an embodiment of the present invention is shown in Figure 1, where (a) is the result of a single LSM network with an average recognition accuracy of 50.95%; (b) is the result of the model of this method with an average recognition accuracy of 85.6%. DETAILED DESCRIPTION

[0036] In order to enable those skilled in the art to better understand the technical solution of the present invention, the technical solution of the present invention will be further described in detail below with reference to the accompanying drawings in the embodiments of the present invention.

[0037] like Figure 1 As shown, the automatic modulation recognition method based on modular pulse neural network in this embodiment includes the following steps: converting input I / Q signal data into a pulse sequence; inputting the pulse sequence into the modular pulse neural network, and the modular pulse neural network processing the pulse sequence is divided into two stages. In the first stage, the pulse neural network is used to roughly classify the pulse sequence to confirm the large category to which the signal belongs to activate the corresponding pulse neural network in the second stage of multiple pulse neural networks; in the second stage, the corresponding pulse neural network is activated, and the amplitude and phase of the input I / Q signal data are extracted through the signal feature enhancement and state fusion module, and the extracted amplitude and phase are spliced ​​with the extracted features of the activated pulse neural network to generate a fusion feature vector and the modulation recognition result is obtained by classification using the classifier of the pulse neural network. Figure 1As shown, to achieve efficient and accurate automatic modulation recognition through a structured workflow, the modular spiking neural network-based automatic modulation recognition method of this embodiment comprises three main components: a pulse encoding module, a signal feature enhancement and state fusion (SFEF) module, and a multi-liquid state machine (LSM) framework. The process begins with the pulse encoding module, which takes raw I / Q (in-phase and quadrature data, specifically the in-phase and quadrature data of the baseband signal) signal data as input and converts it into a pulse train using the TTFS encoding method. The encoded pulse train is then input into the multi-reservoir framework, which consists of two stages: the first stage performs coarse classification to determine the broad category to which the signal belongs and activates the corresponding LSM classifier in the second stage; the second stage activates the corresponding LSM classifier for fine recognition, interacting with the SFEF module during this stage: the pulse emission state of the reservoirs in the LSM is input to the SFEF module. The SFEF module receives the raw I / Q data and the reservoir status data from the second-stage activated LSM. It first converts the I / Q data into A / P data, then concatenates the reservoir status and A / P (amplitude / phase) data to generate a more comprehensive feature vector. Finally, this concatenated data is output to the MLP classifier corresponding to the second-stage LSM for processing, resulting in modulation recognition results. This structural design leverages the efficiency advantages of pulse coding and the feature fusion capabilities of the SFEF module to effectively improve the model's recognition accuracy and robustness in complex signal environments, making it particularly suitable for resource-constrained edge device applications.

[0038] In order to meet the challenge of real-time modulation recognition, this embodiment adopts the TTFS encoding method and applies it to the modulation recognition task for the first time. This method directly encodes the original I / Q signal data into a pulse train, skipping the computationally complex constellation diagram generation step in the traditional method. By focusing on the precise timing of the pulse rather than the pulse frequency, the TTFS method can more effectively capture the timing relationship of the signal and reduce the number of pulses, thereby improving computational efficiency. In this encoding method, each data point of the I / Q signal is mapped to a pulse, and its timing represents the amplitude of the signal - earlier pulses correspond to larger values. The encoding process is carried out in a fixed time window t l The window contains multiple time steps. In each time step t, the algorithm determines whether to trigger a pulse based on the data value and the threshold of dynamic attenuation. Specifically, in this embodiment, converting the input I / Q signal data into a pulse sequence includes:

[0039] S1, initialize time step t to 0;

[0040] S2, determine that the time step t is less than the preset time window t l Is it true? If so, jump to step S3; otherwise, end;

[0041] S3, determine whether the current value v satisfies the condition v≥vth Is it true, where v th If is the threshold value of the current time step, then return to time step t and jump to step S4; otherwise, update the threshold value v of the current time step according to the following formula th :

[0042]

[0043] Among them, v th (t) is the threshold v of the current time step th , v th (0) is the initial threshold, t c is the constant that controls the decay rate;

[0044] This encoding mechanism is based on the "virtual leaky neuron" model, where the threshold decays exponentially in each time step and the pulse triggering condition depends on whether v ≥ v th (t) and t≤t th If there is no trigger pulse within the time window, the data value is considered insufficient to generate a response;

[0045] S4, add 1 to the time step t and jump to step S2.

[0046] The TTFS encoding method avoids the traditional image conversion-based processing method by directly mapping the timing dynamics and amplitude phase characteristics of the signal, greatly improving the computational efficiency.

[0047] One of the challenges of automatic modulation recognition is the similarity between different modulation schemes, particularly common modulation schemes such as amplitude shift keying (ASK), phase shift keying (PSK), and quadrature amplitude modulation (QAM), which exhibit high similarity in their I / Q mapping features. This similarity can easily lead to confusion during classification. For example, the amplitude patterns of high-order ASK and QAM signals are very similar, making them prone to misclassification. Meanwhile, the smaller phase variations of PSK signals lead to overlap in the feature space, increasing classification difficulty. Distinguishing these highly similar modulation categories can be challenging. Modulation recognition tasks often involve multiple groups of highly similar categories, and the relatively simple structure of LSM makes it difficult to accurately distinguish between multiple groups of similar modulation schemes simultaneously. Building on the analysis and insights presented in the previous section on modulation schemes, this embodiment proposes a grouping recognition approach: by grouping the modulation schemes' intrinsic features, the modulation schemes are divided into several groups. The modulation schemes within each group have a high degree of confusion (similarity), while the distinction between groups is higher. This grouping strategy ensures that each LSM classifier focuses only on its assigned subset, allowing it to focus on a narrower, more clearly defined feature space. In this way, a corresponding LSM structure can be designed for each group, which can achieve the dual benefits of higher recognition accuracy and a more streamlined LSM structure, not only improving training efficiency but also improving classification accuracy. Specifically, in this embodiment, before inputting the pulse sequence into the modular pulse neural network, the multiple modulation modes corresponding to the modulation recognition results are also classified into N major categories based on similarity, thereby initializing the two-stage pulse neural network of the modular pulse neural network, and the second stage includes N pulse neural networks, and each pulse neural network corresponds to a major category, and each major category corresponds to one or more modulation modes. Specifically, the specific design of the multi-LSM framework used in the automatic modulation recognition task in this embodiment is as follows: The framework design includes two stages of sequential processing. The first stage uses LSM to achieve coarse classification, dividing the input I / Q signal into one of N major categories (for example, in the RML2016.10a and RML2016.10b datasets, N is 3; in the RML2018.01 dataset, N is 6), narrowing the signal space for subsequent classification. The second stage further refines each modulation category by activating the corresponding specialized LSM classifier, thereby improving classification accuracy and efficiency. This multi-reservoir design not only reduces the amount of data each classifier needs to process, but also effectively reduces interference between classifiers, especially when dealing with similar modulation schemes. In the first stage, the output of the coarse classifier significantly reduces the computational complexity of subsequent fine-classification tasks, while the specialized classifiers in the second stage further optimize the recognition accuracy of each category.

[0048] Figure 2This is a schematic diagram of the activity principle of a single spiking neuron (LIF spiking neuron) in this embodiment. Spiking neural networks (SNNs) are inspired by biological nervous systems and transmit information through pulse trains. In this embodiment, the spiking neural network (SNN) is composed of multiple LIF (leaky integration and fire) neurons. When the voltage accumulation of a neuron exceeds a set threshold, it will generate a pulse and reset the voltage to the resting potential. This characteristic of spiking neurons allows SNNs to retain spatiotemporal information, enabling them to efficiently process complex time series data.

[0049] Figure 3 Schematic diagram of the network structure of the pulse neural network in this embodiment. Figure 3 In this embodiment, the spiking neural network in the first stage is composed of an LSM and a classifier, and the LSM is composed of an input layer, a reservoir layer, and a readout layer. In this embodiment, the spiking neural network in the second stage is composed of an LSM and a classifier, and the LSM is composed of an input layer, a reservoir layer, and a readout layer.

[0050] In this embodiment, the LSM model is used to classify the modulation recognition scheme. Figure 3 The LSM shown in the figure consists of three main components: the input layer, the liquid layer, and the readout layer, with the liquid layer being the core. The liquid layer contains two types of neurons: excitatory neurons and inhibitory neurons. Excitatory neurons increase the membrane potential of connected neurons by firing pulses, while inhibitory neurons decrease the membrane potential of connected neurons by firing pulses. These two types of neurons are arranged in a certain ratio and interconnected through synapses. During LSM operation, input pulses are transmitted from the input layer to the liquid layer, with each input neuron corresponding to a pulse input channel. After receiving the pulses, the membrane potential of the spiking neurons in the liquid layer gradually accumulates. When the membrane potential reaches a threshold, they fire a pulse to the connected neurons. The recurrent connections within the liquid layer enable continuous propagation of the pulse signal, demonstrating the memory property of the LSM. After the sample input is completed, the pulses transmitted between the neurons in the liquid layer form a liquid state. This state can be represented by the number of spikes fired by the neuron or the membrane potential at the time the input was completed. The readout layer then extracts this state and passes it to the subsequent classifier to complete the classification and recognition. The state of the liquid layer at time t+1 can be calculated from the input at the current time and the state at the previous time, as shown below:

[0051] x M (t+1)=A·I(t)+B·x M (t),

[0052] Among them, I(t) represents the input pulse at the current moment, x M (t) is the state of the liquid layer at the current moment, and A and B are the coefficients of state transition.

[0053] For a given input pulse pattern I(s), the response state x of the liquid layer at time t is M (t) can be expressed as:

[0054] x M (t) = L M (I(s)),

[0055] Among them L M represents the response function of the liquid layer to the input pulse. Different spatiotemporal pulse patterns will produce different states in the liquid layer, and each liquid state corresponds to a specific input pattern.

[0056] The function of the readout layer is to convert the liquid state into an output feature vector y(t), which can be expressed as the liquid state x M Function of (t):

[0057] y(t)=f M (x M (t)),

[0058] Combined with the above formula, the output feature vector y(t) of LSM can be further expressed as a function of the input pulse pattern I(s):

[0059] y(t)=f M (L M (I(s))),

[0060] Therefore, LSM can map each input pulse pattern to a corresponding output feature vector, thereby converting an otherwise linearly inseparable spatiotemporal pulse train into a linearly separable feature vector. This property is similar to the kernel method in support vector machines (SVMs). By transforming the spatiotemporal pulse train using LSM and then classifying the feature vectors using a subsequent classifier, classification and recognition of the spatiotemporal pulse train can be achieved.

[0061] Figure 4The following is a schematic diagram of the network structure of the modular spiking neural network in this embodiment. The modular spiking neural network is divided into two stages: the first stage (stage 1) is coarse classification by a general classifier, and the second stage (stage 2) is fine classification by a specialized classifier. In the first stage, this embodiment uses an LSM as a general classifier to perform coarse classification on the input I / Q signal. This classifier classifies the signal into one of six major categories, reducing the computational complexity of subsequent fine classification tasks. By narrowing the range of modulation modes, the general classifier effectively reduces the computational burden of subsequent processing and ensures efficient signal processing. The sparse dynamic reservoir structure of the LSM enables it to effectively process time series signals and extract key features from the pulse train generated by the TTFS encoding module. In this stage, the role of the LSM is to perform preliminary pattern recognition on the signal, extract the signal's time series features, and provide coarse classification input for subsequent specialized classifiers. Based on the output of the general classifier in the first stage, a specialized LSM classifier is activated in the second stage. Each specialized classifier focuses on processing its assigned modulation category, optimizing the recognition performance of that specific category. In these specialized classifiers, modulation modes are rationally assigned to different groups based on similarity. Modulation modes within each group are relatively easy to distinguish, while inter-group distinction is relatively high. This allows the specialized classifiers to focus on the unique characteristics of each class, reducing confusion within the group and improving overall classification accuracy. The specialized classifiers in the second stage further process the input signal to produce even more accurate classification results. These optimized results significantly improve recognition accuracy under low SNR (signal-to-noise ratio) conditions, particularly for high-order ASK, QAM, and PSK signals. The output of the activated LSM classifier in the second stage is passed to the Signal Feature Enhancement and State Fusion (SFEF) module. The SFEF module combines complementary features from the I / Q signal and the LSM reservoir state to provide an enhanced feature representation for final classification. The SFEF module combines the amplitude-phase (A / P) features of the I / Q signal with the temporal features extracted by the LSM to generate a high-quality feature vector. This fused feature vector is ultimately passed to a multi-layer perceptron (MLP) classifier for final automatic modulation recognition. By grouping modulation modes according to their intrinsic characteristics and processing them using a multi-reservoir framework, this embodiment effectively improves the accuracy and efficiency of automatic modulation recognition. The general classifier in the first stage significantly reduces the computational complexity of subsequent sub-classification tasks, while the specialized classifier in the second stage optimizes the unique characteristics of each category, effectively improving recognition performance. Furthermore, the SFEF module further improves recognition accuracy in complex scenarios such as low SNR, ensuring the model's robustness in practical applications. This multi-reservoir framework provides an efficient and scalable solution for automatic modulation recognition tasks.

[0062] In order to further improve the recognition accuracy, especially under low signal-to-noise ratio (SNR) conditions and limited signal length, this embodiment introduces a signal feature enhancement and state fusion (SFEF) module. Figure 5 As shown in the figure, the signal feature enhancement and state fusion (SFEF) module works in parallel with the LSM to process two complementary sources of information: the original I / Q signal data and the reservoir state from the LSM. By explicitly splicing these features, the SFEF module generates a comprehensive feature representation for the final classification task, thereby improving the robustness and recognition accuracy of the model. Specifically, in this embodiment, the function expression for extracting the amplitude and phase of the input I / Q signal data by the signal feature enhancement and state fusion (SFEF) module is:

[0063]

[0064] Wherein, A is the amplitude, θ is the phase, I and Q are the real signal data and imaginary signal data in the I / Q signal data respectively; the function expression for generating a fused feature vector by concatenating the extracted amplitude and phase with the extracted features of the activated spiking neural network is:

[0065] f SFEF =[f A / P ;f reservoir ],

[0066] Among them, f SFEF is the fusion feature vector, f A / P is the eigenvector composed of amplitude and phase, f reservoir The reservoir-layer state vector of the liquid state machine is represented by [·;·], and vector concatenation is used. This explicit concatenation ensures that both signal-level features (A / P) and temporal reservoir state information are preserved in the fused representation. By simultaneously leveraging these two sources of information, the SFEF module significantly improves model performance, particularly in challenging scenarios with low signal-to-noise ratios (SNRs) or limited signal length. Its design overcomes the limitations of traditional feature extraction methods and provides a robust and efficient mechanism for feature fusion and classification.

[0067] In addition, for the pulse neural network in the first and second stages, the Bayesian optimization (BO) method is used in this embodiment. The Bayesian optimization method regards the search for optimal hyperparameters as an optimization problem to ensure that the optimal model configuration is obtained in different tasks. In the model, this embodiment uses a 30-round optimization process. In each round, the hyperparameters are iteratively adjusted by training and validating the model to maximize the recognition accuracy on the validation set. This optimization method ensures that the optimal hyperparameter configuration for the automatic modulation recognition task can be obtained and maintains efficient computing performance in different tasks. Bayesian optimization (BO) is a model-based sequential optimization method that efficiently searches the hyperparameter space by balancing exploration and exploitation. This embodiment uses a tree-like Parsons estimator (TPE) as a proxy model for BO. TPE is widely used in BO because it exhibits excellent robustness and efficiency when processing high-dimensional parameter spaces. The mathematical framework of the TPE model is as follows.

[0068] TPE defines two probability density functions P(x|y), which represent the probability that the objective function value is less than or greater than a threshold y. * Hyperparameter distribution when :

[0069]

[0070] According to Bayes' theorem, we have:

[0071] p(y|x)·p(x)=p(x|y)·p(y),

[0072] Further expanded to:

[0073] p(x)=∫ Θ p(x|y)p(y)dy=γl(x)+(1-γ)g(x),

[0074] Where (γ=p(y<y * )), the default value in the HyperOpt library is 0.25.

[0075] In the TPE framework, the expected improvement (EI) function can be expressed as:

[0076]

[0077] By deduction, we can get:

[0078]

[0079] To maximize the expected improvement at a specific hyperparameter point x, l(x) requires a large value while λ(x) requires a small value. By iteratively adjusting hyperparameters, this embodiment can efficiently optimize model performance and find the hyperparameter configuration that best suits the current automatic modulation recognition task. Through this optimization framework, this embodiment can identify the optimal hyperparameter configuration, enabling the automatic modulation recognition model to achieve a good balance between accuracy and computational efficiency in the automatic modulation recognition task. Experimental results demonstrate that this optimization method is very effective in improving model performance.

[0080] In this embodiment, the modulation recognition results are amplitude keying (ASK), phase keying (PSK), differential phase shift keying (DPSK), quadrature amplitude modulation (QAM), frequency keying (FSK), alternating quadrature phase shift keying (OQPSK), π / 4 quadrature phase shift keying (π / 4-DQPSK) and Gaussian filtered minimum frequency shift keying (GMSK). In addition, the automatic modulation recognition method based on modular pulse neural network in this embodiment can also be applied to some of the above modulation recognition results.

[0081] In order to verify the automatic modulation recognition method based on modular pulse neural network in this embodiment, the method of this embodiment is implemented using the SNN simulator brian2 and the scikit-learn library, while the baseline model is built using TensorFlow. The experiment is conducted on a HP Omen laptop equipped with a GeForce GTX 1050Ti GPU and an Intel i7-7700HQ CPU. The experimental dataset uses the DeepSig RadioML 2018.01A dataset, which covers simulated and real signals in the signal-to-noise ratio range from -10dB to 20dB with a step length of 2dB. Each signal-to-noise ratio level contains 98,304 frames of data. The dataset is divided into training set, validation set and test set in a ratio of 6:2:2. In addition, RML2016.10a and RML2016.10b[2] are used as supplementary datasets. RML2016.10a includes 11 modulation types (8 digital modulations and 3 analog modulations), with a signal-to-noise ratio ranging from -20dB to +18dB. Several existing automatic modulation recognition models are used as comparisons for the method of this embodiment. The existing automatic modulation recognition models include CNN2, DAE, CGDNet, CNNS and ResNet, among which the ResNet method is used as the benchmark scheme for recognition accuracy. Experimental design: The experiment aims to verify the following: 1) By comparing the recognition accuracy and model complexity with the existing models, the feasibility of multiple methods on edge devices is demonstrated. 2) The proposed method is evaluated compared with the model of a single SNN (single LSM) to evaluate the effectiveness of modular SNN. 3) Compare pulse coding using constellation diagrams to verify the advantages of the proposed coding scheme. 4) Test the performance under different symbol sampling lengths on ResNet, this method and single LSM models to verify the effectiveness of the SFEF module. In this example, the performance of the method is evaluated from two aspects: model complexity and recognition accuracy. The experiment is mainly conducted on the RML2018.01 dataset, and supplemented with verification on the RML2016.10a and RML2016.10b datasets.

[0082] Table 1 Comparison of model complexity on the RML2018.01 dataset

[0083]

[0084] Table 1 compares our method with a baseline method on the RML2018.01 dataset at a signal-to-noise ratio of 24dB. Compared to the state-of-the-art ResNet-based method, our method reduces model size by 63.91%, trainable parameters by 94.22%, and training time by 99.95%, while only losing 1.54% in top recognition accuracy. These results highlight the efficiency and practicality of our method on resource-constrained edge devices.

[0085] Figure 6 The comparison of the recognition accuracy of the proposed method and the baseline solution under different signal-to-noise ratios (SNRs) is shown in Figure 2. The x-axis represents the signal-to-noise ratio in dB, and the y-axis represents the recognition accuracy in percentage. Figure 6 As can be seen, our method achieves a peak accuracy of 86.5%, which is comparable to the state-of-the-art methods with minimal loss. However, its accuracy drops under low SNR conditions, highlighting the importance of the SFEF module for improving robustness.

[0086] Experiments on RML2016.10a and RML2016.10b further verify the robustness of this method. Figure 7 The recognition accuracy of the method proposed in the embodiment of the present invention under different signal-to-noise ratios on the additional dataset, where (a) is the result on the RML2016.10a dataset; (b) is the result on the RML2016.10b dataset. Figure 7 As shown, the accuracy of our method is comparable to that of the state-of-the-art methods at different signal-to-noise ratios, demonstrating its generalization capability. Across all datasets, our method reduces trainable parameters by 94.2% and reduces training time by three orders of magnitude compared to the state-of-the-art methods, establishing it as a highly effective solution for modulation recognition.

[0087] Figure 8 These are the ablation experiment results of the method proposed in the embodiment of the present invention, where (a) is the comparison of the recognition accuracy between the present method, a single SNN (single LSM) model, and the present method model without the SFEF module under different signal-to-noise ratios. (b) is the comparison of the recognition accuracy of different automatic modulation recognition schemes (this method, \ac{MR-AMR} without SFEF, ResNet, and a single SNN) under different sample lengths. The standard sample length of the RML2018.01 dataset is 1024 samples per symbol, and the results are shown as 25%, 50%, 75%, and 100% of the standard length. It can be seen that the multi-reservoir design in this method significantly improves the recognition accuracy, as shown in Figure 2. Figure 8 Compared with a single SNN model, our method achieves higher accuracy at all signal-to-noise ratio levels, especially on the 24-class dataset, which highlights the effectiveness of dividing the task into specialized classifiers for targeted feature extraction.

[0088] Figure 9 The confusion matrix comparison of modulation recognition under a signal-to-noise ratio of 24 for the method proposed in this embodiment is shown in Figure 1, where (a) is the result of a single LSM network with an average recognition accuracy of 50.95%; (b) is the result of the model of this method with an average recognition accuracy of 85.6%. Figure 9(a) in the figure shows the performance of a single LSM network, while Figure 9 Panel (b) shows the results of the proposed model, highlighting its improved recognition accuracy. However, due to the similar signal characteristics of the APSK and MQAM schemes, a higher confusion rate was observed. These results demonstrate the scalability and reliability of the proposed method and point out areas for future improvement in distinguishing similar schemes.

[0089] In addition, the comparison of different pulse coding schemes in terms of pulse number and coding time during the operation of the method in this embodiment is shown in Table 2.

[0090] Table 2 Comparison of different pulse coding schemes in terms of pulse number and coding time during operation

[0091] Comparative indicators Single LSM method This method Number of pulses (per data) 719.97 490.40 ↓31.89％ Encoding time (seconds) 0.087 0.006 ↓93.1％

[0092] Table 2 highlights the efficiency of the proposed TTFS encoding scheme, compared to the Poisson-based encoding used in the single LSM approach. This approach reduces the number of pulses by 31.89% and encoding time by 93.1%, while losing less than 1% in accuracy. These results emphasize the suitability of TTFS encoding for real-time edge applications, enabling faster and more efficient processing.

[0093] In summary, the automatic modulation recognition method based on modular pulse neural network in this embodiment can be applied to embedded communication equipment such as drones. This method realizes accurate recognition of complex modulated signals by introducing a modular pulse neural network (SNN) structure. First, the input I / Q signal is time-encoded, and the continuous signal is directly converted into a pulse sequence, bypassing the traditional image conversion step and reducing the amount of coding calculation. Then, multiple liquid state machine (LSM) modules are used to extract features of the coded signal, so that the SNN can more efficiently utilize the spatiotemporal information of the signal, reducing the model complexity and power consumption. At the same time, the designed signal feature enhancement and state fusion module (SFEF) maintains a high recognition rate in low signal-to-noise ratio scenarios and effectively improves the robustness of signal processing. The method of this embodiment effectively solves the technical problem that traditional deep learning models are difficult to balance the recognition accuracy and power consumption requirements in edge devices such as drones, and provides a lightweight and efficient AMR solution for application scenarios such as drone communications and intelligent monitoring. This embodiment proposes a modular SNN architecture to meet the needs of embedded devices such as drones. This architecture not only adopts a modular design based on SNN, but also innovatively adopts a time coding method to directly convert I / Q signals into pulses, reducing the computational burden in the conversion step and significantly improving the system response efficiency and energy efficiency. Combined with the LSM module and the signal feature enhancement and state fusion module (SFEF), the method of this embodiment has efficient signal recognition capabilities in low signal-to-noise ratio scenarios, and is particularly suitable for the needs of real-time and efficient processing in drone communications. This innovative technology solves the problems of limited computing resources and high energy consumption requirements of existing AMR methods in embedded edge devices, providing strong support for future intelligent communications and drone applications.

[0094] In addition, this embodiment also provides an automatic modulation recognition system based on a modular pulse neural network, comprising a microprocessor and a memory connected to each other, wherein the microprocessor is programmed or configured to execute the automatic modulation recognition method based on a modular pulse neural network.

[0095] In addition, this embodiment also provides a computer-readable storage medium, which stores a computer program or instruction. The computer program or instruction is programmed or configured to execute the automatic modulation recognition method based on modular pulse neural network through a processor.

[0096] In addition, this embodiment also provides a computer program product, including a computer program or instructions, which are programmed or configured to execute the automatic modulation recognition method based on modular pulse neural network through a processor.

[0097] Those skilled in the art should understand that the technical solution provided by the present invention may be in the form of a method, a system, or a computer program product. Therefore, the present invention may take the form of a complete hardware embodiment, a complete software embodiment, or an embodiment combining software and hardware aspects. Moreover, the present invention may take the form of a computer program product implemented on one or more computer-readable storage media (including but not limited to disk storage, CD-ROM, optical storage, etc.) containing computer-usable program code. The present invention is described with reference to the flowcharts and / or block diagrams of the methods, devices (systems), and computer program products according to the embodiments of the present invention. It should be understood that each process and / or box in the flowchart and / or block diagram, as well as the combination of processes and / or boxes in the flowchart and / or block diagram, may be implemented by computer program instructions. These computer program instructions may be provided to a processor of a general-purpose computer, a special-purpose computer, an embedded processor, or other programmable data processing device to produce a machine, so that the instructions executed by the processor of the computer or other programmable data processing device generate instructions for implementing the functions described in the process. Figure 1 a process or multiple processes and / or boxes Figure 1 These computer program instructions can also be stored in a computer-readable memory that can guide a computer or other programmable data processing device to work in a specific way, so that the instructions stored in the computer-readable memory produce a product including the instruction device, which implements the function specified in the process. Figure 1 a process or multiple processes and / or boxes Figure 1 These computer program instructions can also be loaded onto a computer or other programmable data processing device, so that a series of operation steps are executed on the computer or other programmable device to produce a computer-implemented process, thereby providing instructions for implementing the process in the process. Figure 1 a process or multiple processes and / or boxes Figure 1 A step that specifies a function in one or more boxes.

[0098] The above description is merely a preferred embodiment of the present invention. The scope of protection of the present invention is not limited to the above embodiment. All technical solutions based on the concept of the present invention are within the scope of protection of the present invention. It should be noted that for those skilled in the art, various improvements and modifications that do not depart from the principles of the present invention should also be considered within the scope of protection of the present invention.

Claims

1. An automatic modulation recognition method based on modular pulse neural network, characterized in that: The method comprises the following steps: converting input I / Q signal data into a pulse sequence; The pulse sequence is input into a modular pulse neural network, and the modular pulse neural network processes the pulse sequence in two stages. In the first stage, the pulse neural network is used to perform coarse classification of the pulse sequence to confirm the broad category to which the modulation signal belongs, and activate the corresponding pulse neural network in multiple pulse neural networks in the second stage; in the second stage, the corresponding pulse neural network is activated, and the amplitude and phase of the input I / Q signal data are extracted through the signal feature enhancement and state fusion module, and the extracted amplitude and phase are concatenated with the features extracted by the activated pulse neural network to generate a fused feature vector, and the modulation recognition result is obtained by classification using the classifier of the pulse neural network.

2. The automatic modulation recognition method based on modular pulse neural network according to claim 1 is characterized in that: The step of converting the input I / Q signal data into a pulse sequence includes: S1, initialize time step t to 0; S2, determine that the time step t is less than the preset time window t l Is it true? If so, jump to step S3; otherwise, end; S3, determine whether the current value v satisfies the condition v≥v th Is it true, where v th If is the threshold value of the current time step, then return to time step t and jump to step S4; otherwise, update the threshold value v of the current time step according to the following formula th : Among them, v th (t) is the threshold v of the current time step th , v th (0) is the initial threshold, t c is the constant that controls the decay rate; S4, add 1 to the time step t and jump to step S2.

3. The automatic modulation recognition method based on modular pulse neural network according to claim 1, characterized in that: Before inputting the pulse sequence into the modular pulse neural network, the method also includes dividing the multiple modulation methods corresponding to the modulation recognition results into N major categories according to similarity, thereby initializing the pulse neural network of the two stages of the modular pulse neural network, and the second stage includes N pulse neural networks, each pulse neural network corresponds to a major category, and each major category corresponds to one or more modulation methods.

4. The automatic modulation recognition method based on modular pulse neural network according to claim 1, characterized in that: The spiking neural network in the first stage is composed of a liquid state machine and a classifier, and the liquid state machine is composed of an input layer, a reservoir layer and a readout layer.

5. The automatic modulation recognition method based on modular pulse neural network according to claim 1 is characterized in that: The pulse neural network in the second stage is composed of a liquid state machine and a classifier, and the liquid state machine is composed of an input layer, a reservoir layer and a readout layer.

6. The automatic modulation recognition method based on modular pulse neural network according to claim 5, characterized in that: The function expression for extracting amplitude and phase of the input I / Q signal data through the signal feature enhancement and state fusion module is: Wherein, A is the amplitude, θ is the phase, I and Q are the real signal data and imaginary signal data in the I / Q signal data respectively; the function expression for generating the fusion feature vector by splicing the extracted amplitude and phase with the extracted features of the activated pulse neural network is: f SFEF =[f A / P ;f reservoir ], Among them, f SFEF is the fusion feature vector, f A / P is the eigenvector composed of amplitude and phase, f reservoir represents the reservoir layer state vector of the liquid state machine, and [·;·] represents vector concatenation.

7. The automatic modulation recognition method based on modular pulse neural network according to claim 1, characterized in that: The modulation identification result is part or all of amplitude keying, phase keying, differential phase shift keying, quadrature amplitude modulation, frequency keying, staggered quadrature phase shift keying, π / 4 quadrature phase shift keying and Gaussian filtered minimum frequency shift keying.

8. An automatic modulation recognition system based on a modular pulse neural network, comprising a microprocessor and a memory connected to each other, characterized in that: The microprocessor is programmed or configured to execute the automatic modulation recognition method based on modular pulse neural network according to any one of claims 1 to 7.

9. A computer-readable storage medium having a computer program or instruction stored therein, characterized in that: The computer program or instruction is programmed or configured to execute the automatic modulation recognition method based on modular pulse neural network according to any one of claims 1 to 7 through a processor.

10. A computer program product comprising a computer program or instructions, characterized in that The computer program or instruction is programmed or configured to execute the automatic modulation recognition method based on modular pulse neural network according to any one of claims 1 to 7 through a processor.

Images