Low signal-to-noise ratio OTFS signal identification and classification method and identification and classification system

Abstract

The invention discloses a low-signal-to-noise-ratio OTFS signal identification and classification method and an identification and classification system. The identification and classification method comprises the following steps: S1, acquiring a wireless communication signal and constructing a data set; s2, classifying high and low signal-to-noise ratios of signals in the data set by using an LSTM network; s3, performing noise reduction processing on the low-signal-to-noise-ratio signal through a two-dimensional input DTLN model; and S4, using a convolutional neural network as a modulation signal classification module, and inputting the data subjected to noise reduction processing of the DTLN model and the high signal-to-noise ratio signal into the modulation signal classification module for processing. According to the method, the multi-level feature extraction capability of the cascaded neural network under the low signal-to-noise ratio condition can be matched, efficient recognition and classification of the low signal-to-noise ratio OTFS signals are completed, efficient processing of the whole process from signal preprocessing to final classification output is achieved, the signal classification effect can be remarkably improved even in the low signal-to-noise ratio environment, and the method is suitable for popularization and application. And the requirements of real-time signal processing on high accuracy and low delay can be met.

Description

Technical Field

[0001] The present invention relates to the technical field of signal processing and recognition, and particularly relates to a method, a system, a device and a storage medium for identifying and classifying low signal-to-noise ratio OTFS signals. Background Art

[0002] With the development of communication technologies, the complexity and diversity of communication scenarios have increased significantly, and the demand for high-rate, low-latency, and high-reliability transmission has become even more urgent. However, when facing time-varying multipath channels, the performance of traditional orthogonal frequency division multiplexing (OFDM) technology is limited. Especially in high-speed mobile or harsh communication environments, it is difficult to meet the high-performance requirements of communication systems.

[0003] To solve these problems, orthogonal time-frequency space (OTFS) technology is adopted in the prior art for waveform modulation. By mapping information symbols to the delay-Doppler domain, OTFS technology enables signals to effectively cope with changes in time and frequency, showing higher anti-interference ability and robustness.

[0004] Although OTFS has significant advantages, in practical applications, low signal-to-noise ratio (SNR) scenarios remain a major problem in signal processing. In complex communication environments, noise, interference, multipath effects, high-dynamic Doppler effects, etc. will all cause serious interference to the identification and classification of OTFS signals.

[0005] Traditional signal processing methods, such as those based on Fourier transform, wavelet transform, etc., although they can play a role in certain specific scenarios, often have difficulty extracting effective features in low SNR environments, and both the identification accuracy and classification performance decrease significantly.

[0006] In addition, traditional methods have a strong dependence on parameters and lack flexibility when facing complex channel conditions, and it is difficult to adapt to dynamically changing communication environments.

[0007] Therefore, this application specifically proposes a method for identifying and classifying low SNR OTFS signals to solve the above technical problems. Summary of the Invention

[0008] The main objective of the present invention is to provide a method for identifying and classifying low SNR OTFS signals. Under low SNR conditions, by utilizing the multi-level feature extraction ability of a cascaded neural network, the key features of OTFS signals can be effectively separated from noise to complete efficient identification and classification, so as to solve the technical problems proposed in the background art.

[0009] The present invention adopts the following technical solutions to solve the above technical problems:

[0010] A method for identifying and classifying low SNR OTFS signals, comprising the following steps:

[0011] S1. Obtain a wireless communication signal and construct a data set;

[0012] S2. Use an LSTM network to classify the high and low signal-to-noise ratios of the signals in the data set;

[0013] S3. The low signal-to-noise ratio signals are denoised by a DTLN model with two-dimensional input;

[0014] S4. Use a convolutional neural network as a modulation signal classification module, and input the data after denoising by the DTLN model and the high signal-to-noise ratio signals into the modulation signal classification module for processing.

[0015] Preferably, the data set constructed in step S1 includes OTFS, OFDM, 2ASK, 4FSK, and 8PSK signals with different signal-to-noise ratios;

[0016] The OTFS signal and OFDM are modulated and generated with consistent channel parameter settings;

[0017] The signal-to-noise ratio ranges of the wireless communication signals inside the data set are set consistently.

[0018] Preferably, the specific operation process of classification in step S2 includes:

[0019] S21. Define the input signal data dimensions including the data frame length and batch size, and use the in-phase I dimension and quadrature Q dimension of the signal as independent features;

[0020] S22. Use an LSTM network to process the data in the data set through a hierarchical structure, extract the specified time features using the hidden layer, perform data modeling according to the time series characteristics of the signal, and use a stochastic gradient descent optimizer and binary cross-entropy loss as the loss function for model training;

[0021] S23. Through the Sigmoid activation function, map the features of the last time step to a probability value for judging the signal-to-noise ratio category;

[0022] S24. Preset a classification threshold, classify the signals with a signal-to-noise ratio greater than the threshold as high signal-to-noise ratio signals, and classify the signals less than or equal to the threshold as low signal-to-noise ratio signals.

[0023] Preferably, the specific operation process of the DTLN model for denoising in step S3 includes:

[0024] S31. Represent the low signal-to-noise ratio signals as time domain signals including the in-phase I dimension and quadrature Q dimension, including batch size, sample length, and in-phase and quadrature IQ channel dimensions, where the batch size is set to be adjustable, and the IQ channel represents the real and imaginary parts of the signal;

[0025] After the low signal-to-noise ratio signal is input into the DTLN model for processing, a denoised signal with the same dimension as the input is output, and the DTLN model uses the high signal-to-noise ratio signal as the target clean signal and the low signal-to-noise ratio signal as the input for supervised training;

[0026] S33. The loss function is calculated by combining the time domain and frequency domain joint errors, and the data is optimized by combining the mean square error and the signal-to-noise ratio gain.

[0027] Preferably, the DTLN model includes:

[0028] Encoder module: Extract the local features of the time domain signal through one-dimensional convolution and represent the encoding as high-dimensional features;

[0029] Time dimension separation network: Use a dual-path separation structure to capture the time context characteristics of the signal, and use LSTM or GRU as the time dimension separation path to model the sequence characteristics;

[0030] Frequency dimension separation network: Extract the spectral features of the signal through short-time Fourier transform, and combine convolutional neural network or self-attention mechanism to enhance the ability to distinguish noise from signal;

[0031] Decoder module: Map the enhanced high-dimensional features back to the time domain signal through deconvolution or fully connected reconstruction module to output a denoised signal with the same dimension as the input.

[0032] Preferably, the data processing flow of the modulation signal classification module in step S4 includes:

[0033] S41. Input the low signal-to-noise ratio signal after denoising processing into the convolutional neural network model. The input data includes batch size, time step, and in-phase quadrature IQ channel dimension, where the batch size is an adjustable parameter, and the IQ channel represents the real and imaginary parts of the signal;

[0034] S42. The first layer uses a Conv1D convolutional layer with a kernel size of 3, an input channel number of 2 for the IQ signal, and an output channel number parameter that is adjustable. The ReLU activation function is used to introduce non-linearity and extract the local time features of the signal;

[0035] S43. In the second layer, use a max pooling layer with a pooling kernel size of 2 to gradually reduce the feature dimension to reduce the computational complexity;

[0036] S44. The features after pooling processing are flattened into a one-dimensional vector by using a Flatten layer in the third layer;

[0037] S45. Input the processed vector data into a fully connected network for classification processing.

[0038] Preferably, the fully connected network consists of two sets of fully connected layers:

[0039] The first fully connected layer contains multiple neurons and enhances the feature expression ability through the ReLU activation function;

[0040] The second fully connected layer is used to output the classification result. It contains multiple nodes, corresponding to multiple modulation types respectively, and uses the Softmax activation function to calculate the probabilities of all modulation types in the fully connected network. The classification result is the modulation type with the highest probability.

[0041] A low signal-to-noise ratio OTFS signal recognition and classification system is used to execute the low signal-to-noise ratio OTFS signal recognition and classification method described above. It consists of a signal-to-noise ratio classification module, a signal enhancement module, and a signal classification module to form a cascaded neural network structure, where:

[0042] The signal-to-noise ratio classification module is used to preliminarily classify the input signal according to the signal-to-noise ratio;

[0043] The signal enhancement module is used to perform noise reduction processing on low signal-to-noise ratio signals. The signal enhancement module is constructed based on a dual-path real-time denoising network, and while reducing noise interference, it retains the core features of the signal, which helps the subsequent classification module to more accurately identify the signal type;

[0044] The signal classification module is used to classify the modulation type of the enhanced signal. The signal classification module is constructed based on a convolutional neural network, and its structure design is simple and efficient, and high-precision classification can be achieved in a short training time.

[0045] Preferably, the specific signal processing flow of the signal-to-noise ratio classification module includes:

[0046] a1. Capture the time series features of the signal in the LSTM layer, and the output is a high-dimensional feature representation;

[0047] a2. Extract the features of the last time step to obtain a low-dimensional feature vector;

[0048] a3. Pass through the fully connected layer and use the Sigmoid activation function to output the signal-to-noise ratio classification result.

[0049] On the other hand, the present invention also discloses a computer-readable storage medium storing a computer program. When the computer program is executed by a processor, the processor is caused to execute the steps of the above method.

[0050] On yet another aspect, the present invention also discloses a computer device including a memory and a processor. The memory stores a computer program. When the computer program is executed by the processor, the processor is caused to execute the steps of the above method.

[0051] As can be seen from the above technical solutions, the present invention provides a method for identifying and classifying low signal-to-noise ratio OTFS signals. Compared with the prior art, the present invention has the following advantages:

[0052] 1. The present invention uses an LSTM network for high and low signal-to-noise ratio classification, classifies and processes different signal-to-noise ratio signals after effective differentiation, can cooperate with the multi-level feature extraction ability of the cascaded neural network under low signal-to-noise ratio conditions, effectively separates the key features of the OTFS signal from noise, completes efficient identification and classification, not only breaks through the limitations of traditional algorithms, but also can adapt to complex dynamic environments, improves the signal processing performance in low signal-to-noise ratio scenarios, and can also improve the signal processing efficiency and accuracy of the entire system.

[0053] 2. The present invention uses an improved DTLN model to denoise low signal-to-noise ratio signals, can optimize signal feature extraction simultaneously in the time domain and frequency domain, thereby significantly improving signal clarity and quality, and enhancing the recognition accuracy of the subsequent modulation signal classification module.

[0054] 3. The present invention improves the structure of the DTLN model including an encoder, a time dimension separation network, a frequency dimension separation network, and a decoder. It can capture the time context characteristics and spectral features of the signal, thereby efficiently removing noise interference and retaining the core features of the signal, which is helpful for subsequent classification tasks.

[0055] 4. The present invention uses a convolutional neural network as the modulation signal classification module and combines the ReLU activation function to introduce non-linearity, can extract local time and frequency features of the signal, thereby maintaining high classification accuracy while reducing computational complexity, and realizing fast and accurate modulation type classification.

[0056] 5. The present invention organically combines the signal-to-noise ratio classification module, the signal enhancement module, and the signal classification module through a cascaded neural network structure, can efficiently process the entire process from signal preprocessing to final classification output, and can significantly improve the signal classification effect even in a low signal-to-noise ratio environment, and can meet the requirements of high accuracy and low latency for real-time signal processing.

[0057] It should be understood that the content described in this part is not intended to identify the key or important features of the embodiments of the present invention, nor is it used to limit the scope of the present invention. Other features of the present invention will become easily understood through the following description. Of course, any product implementing the present invention does not necessarily need to achieve all the above advantages simultaneously. BRIEF DESCRIPTION OF THE DRAWINGS

[0058] The specification drawings constituting a part of this application are used to provide a further understanding of the present invention. The schematic embodiments of the present invention and their descriptions are used to explain the present invention and do not constitute an improper limitation to the present invention. In the drawings:

[0059] Figure 1 This is a schematic diagram of the overall method flow based on the system framework of the present invention;

[0060] Figure 2 This is a schematic diagram of the data processing flow of the present invention;

[0061] Figure 3 This is a schematic diagram of the network structure of the high and low signal-to-noise ratio signal discrimination network based on LSTM of the present invention;

[0062] Figure 4 This is a schematic diagram of the network structure of the two-dimensional input DTLN denoising model of the present invention;

[0063] Figure 5 This is a schematic diagram of the network structure of the CNN modulation type classification model of the present invention;

[0064] Figure 6 This is a schematic diagram of the comparison of the test accuracy results of the present invention;

[0065] Figure 7 This is a confusion matrix diagram of the test results of the present invention at a signal-to-noise ratio of 10 dB;

[0066] Figure 8 This is a confusion matrix diagram of the test results of the present invention at a signal-to-noise ratio of 6 dB;

[0067] Figure 9 This is a confusion matrix diagram of the test results of the present invention at a signal-to-noise ratio of 2 dB;

[0068] Figure 10 This is a confusion matrix diagram of the test results of the present invention at a signal-to-noise ratio of -2 dB;

[0069] Figure 11 This is a comparison diagram of the radio signal before and after noise reduction of the present invention. Detailed implementation manners

[0070] Next, the technical solutions in the embodiments of the present invention will be clearly and completely described in conjunction with the accompanying drawings in the embodiments of the present invention. Obviously, the described embodiments are only a part of the embodiments of the present invention, rather than all the embodiments. Without conflict, the embodiments in the present application and the features in the embodiments can be combined with each other. All other embodiments obtained by those of ordinary skill in the art based on the embodiments of the present invention without creative efforts shall fall within the protection scope of the present invention.

[0071] In the embodiments, please refer in detail to Figures 1 to 11 .

[0072] Such as Figure 1 and Figure 2As shown in the figure, a method for identifying and classifying low signal-to-noise ratio OTFS signals proposed in an embodiment of the present invention includes the following steps:

[0073] S1. Obtain a wireless communication signal and construct a data set;

[0074] The constructed data set includes OTFS, OFDM, 2ASK, 4FSK, and 8PSK signals with different signal-to-noise ratios;

[0075] The OTFS signal and OFDM are modulated and generated with consistent channel parameter settings;

[0076] The signal-to-noise ratio ranges of the wireless communication signals within the data set are set consistently.

[0077] It should be added that in terms of the generation of the data set, the present invention uses an efficient simulation platform to generate a variety of typical modulation signals, including signals of five modulation types such as OTFS, OFDM, 2ASK, 4FSK, and 8PSK. The generation process of these signals fully considers common channel characteristics in wireless communication, including parameters such as path delay, path gain, and Doppler frequency shift. In addition, these channel parameter designs are based on typical multipath propagation scenarios, ensuring the authenticity and representativeness of the data set;

[0078] Furthermore, in terms of signal-to-noise ratio design, the data set covers a variety of communication environments from low signal-to-noise ratio to high signal-to-noise ratio, ensuring the adaptability of model training and testing to different complex environments;

[0079] In addition, each sample data contains two paths of signals, namely in-phase (I) and quadrature (Q). Parameters such as the time step and frequency characteristics of the sample conform to the physical characteristics of actual communication signals.

[0080] To improve the generalization ability of the model, the data set is divided into a training set and a testing set according to a certain ratio, which can provide diverse and high-quality input data for subsequent model training, ensuring that the model can better adapt to complex channel conditions.

[0081] S2. Use an LSTM network to classify the high and low signal-to-noise ratios of the signals in the data set, which can effectively distinguish signals with different signal-to-noise ratios, thereby providing an accurate basis for subsequent processing and improving the signal processing efficiency and accuracy of the entire system;

[0082] For the specific operation process, refer to Figure 3 and includes:

[0083] S21. Define the input signal data dimensions including the data frame length and batch size, and use the in-phase I dimension and quadrature Q dimension of the signal as independent features;

[0084] S22. Process the data in the dataset through a hierarchical structure using an LSTM network, extract the specified time features using a hidden layer, perform data modeling based on the time series characteristics of the signal, and use a stochastic gradient descent optimizer. Use binary cross-entropy loss as the loss function to train the model;

[0085] S23. Through the Sigmoid activation function, map the features of the last time step to a probability value for determining the signal-to-noise ratio category;

[0086] S24. Preset a classification threshold (the classification threshold can be set by oneself), classify the signals with a signal-to-noise ratio greater than the threshold as high signal-to-noise ratio signals, and classify the signals less than or equal to the threshold as low signal-to-noise ratio signals.

[0087] At this time, by using the LSTM network for high and low signal-to-noise ratio classification, classifying and processing the different signal-to-noise ratio signals after effective discrimination, it can cooperate with the multi-level feature extraction ability of the cascaded neural network under low signal-to-noise ratio conditions, effectively separate the key features of the OTFS signal from the noise, complete efficient recognition and classification, not only break through the limitations of traditional algorithms, but also adapt to complex dynamic environments, improve the signal processing performance in low signal-to-noise ratio scenarios, and improve the signal processing efficiency and accuracy of the entire system.

[0088] S3. Denoise the low signal-to-noise ratio signal through the DTLN model with two-dimensional input;

[0089] At this time, the network structure of the DTLN denoising model refers to Figure 4 , and the specific operation process of the DTLN model for denoising includes:

[0090] S31. Represent the low signal-to-noise ratio signal as a time-domain signal including the in-phase I dimension and the quadrature Q dimension, including the batch size, sample length, and in-phase and quadrature IQ channel dimensions, where the batch size is set to be adjustable, and the IQ channel represents the real and imaginary parts of the signal;

[0091] S32. After the low signal-to-noise ratio signal is input into the DTLN model for processing, a denoised signal with the same dimension as the input is output, and the DTLN model uses the high signal-to-noise ratio signal as the target clean signal and the low signal-to-noise ratio signal as the input for supervised training. At this time, by using the DTLN model to denoise the low signal-to-noise ratio signal, it can optimize the signal feature extraction in both the time domain and the frequency domain, thereby significantly improving the signal clarity and quality and enhancing the recognition accuracy of the subsequent modulation signal classification module;

[0092] S33. Adopt a loss function that calculates the joint error in the time domain and the frequency domain, and perform data optimization in combination with the mean square error and the signal-to-noise ratio gain.

[0093] Specifically, the DTLN model includes:

[0094] Encoder module: Extract local features of time-domain signals through one-dimensional convolution and represent the encoding as high-dimensional features;

[0095] Time dimension separation network: Use a dual-path separation structure to capture the time context characteristics of signals, and adopt LSTM or GRU as the time dimension separation path to model sequence characteristics;

[0096] Frequency dimension separation network: Extract the spectral features of signals through short-time Fourier transform, and combine convolutional neural networks or self-attention mechanisms to enhance the ability to distinguish noise from signals;

[0097] Decoder module: Map the enhanced high-dimensional features back to time-domain signals through deconvolution or fully connected reconstruction modules to output a denoised signal with the same dimension as the input.

[0098] By designing the DTLN model structure including an encoder, a time dimension separation network, a frequency dimension separation network, and a decoder, it is possible to capture the time context characteristics and spectral features of signals, efficiently remove noise interference, retain the core features of signals, and contribute to subsequent classification tasks.

[0099] S4. Use a convolutional neural network CNN as the modulation signal classification module, and input the data after denoising by the DTLN model and the high signal-to-noise ratio signal into the modulation signal classification module for processing. At this time, the model structure refers to Figure 5 ;

[0100] Among them, the data processing process of the modulation signal classification module includes:

[0101] S41. Input the low signal-to-noise ratio signal after denoising into the convolutional neural network model. The input data includes batch size, time step, and in-phase quadrature IQ channel dimension, where the batch size is an adjustable parameter, and the IQ channel represents the real and imaginary parts of the signal;

[0102] S42. The first layer uses a Conv1D convolutional layer with a kernel size of 3, an input channel number of 2 for the IQ signal, and an adjustable output channel number parameter. Introduce non-linearity through the ReLU activation function to extract local time features of the signal;

[0103] S43. Use a max-pooling layer in the second layer with a pooling kernel size of 2 to gradually reduce the feature dimension to reduce computational complexity;

[0104] S44. The features after pooling are flattened into a one-dimensional vector by using a Flatten layer in the third layer;

[0105] S45. Input the processed vector data into a fully connected network for classification processing;

[0106] It should be noted at this time that the fully connected network consists of two sets of fully connected layers:

[0107] The first fully connected layer contains multiple neurons, and the ReLU activation function is used to enhance the feature expression ability;

[0108] The second fully connected layer is used to output the classification result. It contains multiple nodes, corresponding to multiple modulation types respectively, and the Softmax activation function is used to calculate the probabilities of all modulation types in the fully connected network. The classification result is the modulation type with the highest probability.

[0109] At this time, by using a convolutional neural network as the modulation signal classification module and combining the ReLU activation function to introduce non-linearity, local time and frequency features of the signal are extracted, and the computational complexity is reduced while maintaining high classification accuracy, realizing fast and accurate modulation type classification.

[0110] In summary, the low signal-to-noise ratio OTFS signal recognition and classification method realizes signal classification processing by cascading the LSTM network, DTLN model and CNN network. This cascaded structure realizes the full-process processing from signal classification at high and low signal-to-noise ratios, noise reduction optimization to accurate classification, and significantly improves the signal recognition ability and classification accuracy in a low signal-to-noise ratio environment.

[0111] In addition, what needs to be noted during the use of this method is:

[0112] (1) This method models the time series characteristics of the input signal through the LSTM network and extracts the key time features of the signal using the hidden layer;

[0113] (2) After the signal is divided into high signal-to-noise ratio signals and low signal-to-noise ratio signals, the high signal-to-noise ratio signals directly enter the CNN network for classification, and the low signal-to-noise ratio signals are transmitted to the DTLN model for noise reduction processing;

[0114] (3) The encoder module extracts high-dimensional feature representations through Conv1D, and the dual-path separation network separates and enhances the time-dimensional and frequency-dimensional features respectively;

[0115] (4) The noise-reduced signal reconstructed by the decoder module retains the key modulation information and serves as the input to the CNN classification module;

[0116] (5) The input signal and the classified output signal include but are not limited to five modulation types: 2ASK, 4FSK, 8PSK, OTFS, and OFDM.

[0117] On the other hand, referring to Figure 1, the present invention also discloses a low signal-to-noise ratio OTFS signal recognition and classification system for implementing the low signal-to-noise ratio OTFS signal recognition and classification method described above. It consists of a signal-to-noise ratio classification module, a signal enhancement module, and a signal classification module to form a cascaded neural network structure, where:

[0118] The first module is the signal-to-noise ratio classification module, mainly used for the preliminary classification of input signals according to the signal-to-noise ratio. This module is constructed using a Long Short-Term Memory (LSTM) network. The shape of the input signal includes batch size, time step length, and IQ channels, where the batch size is an adjustable parameter, the time step length is determined according to the specific application scenario, and the IQ channels represent the real and imaginary parts of the signal. The signal passes through the following structure in sequence: the LSTM layer, which is used to capture the time series characteristics of the signal and outputs a high-dimensional feature representation; then the features of the last time step are extracted to obtain a low-dimensional feature vector; and then through a fully connected layer, using the Sigmoid activation function to output the signal-to-noise ratio classification result. This module can accurately distinguish high signal-to-noise ratio signals and low signal-to-noise ratio signals, providing an effective basis for the processing of subsequent modules. Through this processing method, signals with different signal-to-noise ratios can be input into subsequent modules respectively for more targeted processing.

[0119] The second module is the signal enhancement module, specifically used for noise reduction processing of low signal-to-noise ratio signals. This module is constructed based on an improved Dual-Path Real-Time Denoising Network (DTLN). By combining the dual-path feature extraction methods in the time domain and frequency domain, it significantly improves the feature representation ability of low signal-to-noise ratio signals. The structure of the DTLN module includes an encoder, a dual-path separation network, and a decoder. The encoder extracts the high-dimensional feature representation of the time-domain signal through one-dimensional convolution (Conv1D); the dual-path separation network is divided into a time-dimensional separation path and a frequency-dimensional separation path. The former uses a recurrent neural network (such as LSTM or GRU) to capture the time context characteristics of the signal, and the latter extracts the frequency characteristics of the signal through frequency-domain modeling (such as STFT) to enhance the ability to separate noise; finally, through the decoder, the enhanced high-dimensional features are reconstructed back into the time-domain signal, and the processed low signal-to-noise ratio signal is output. The shape of the input signal includes batch size, time step length, and IQ channels, and the output signal shape is the same as the input, realizing low-latency processing of one-frame input and one-frame output. While reducing noise interference, this module retains the core features of the signal, which helps the subsequent classification module to more accurately identify the signal type.

[0120] The third module is the signal classification module, which is mainly used for classifying the modulation types of the enhanced signals. This module is constructed based on a Convolutional Neural Network (CNN), and its structure includes a convolutional layer, a pooling layer, a flattening layer, and a fully connected layer. Specifically, the input signal first undergoes processing in the convolutional layer. The number of input channels is 2, the number of output channels is an adjustable parameter, the convolutional kernel size is 3, and the activation function is ReLU, which is used to extract local time and frequency features. Subsequently, the features are dimensionally reduced through the max pooling layer, with a pooling kernel size of 2 to reduce the computational complexity. Then, the convolutional features are flattened into a one-dimensional vector through the flattening layer. Finally, classification is completed through the fully connected layer. The number of output nodes in the first layer is an adjustable parameter, and the number of output nodes in the second layer is multiple nodes, corresponding to the classification probabilities of different modulation types respectively. The structure design of this module is simple and efficient, and high-precision classification can be achieved within a short training time.

[0121] Through the collaborative work of the signal-to-noise ratio classification module, the signal enhancement module, and the signal classification module, the entire cascaded network can achieve an efficient process from signal preprocessing to classification output. In a low signal-to-noise ratio environment, the model can significantly improve the classification effect of low signal-to-noise ratio signals by first performing signal-to-noise ratio classification and noise reduction enhancement; while in a high signal-to-noise ratio condition, classification is directly carried out to save computational resources. This modular design method not only improves the performance of the model but also provides convenience for future extended applications.

[0122] In summary, the present invention organically combines the signal-to-noise ratio classification module, the signal enhancement module, and the signal classification module through dataset design and a modular cascaded neural network structure, efficiently processes the entire process from signal preprocessing to final classification output, and achieves the goal of significantly improving the signal classification effect even in a low signal-to-noise ratio environment, realizing the efficient classification of various signals in a low signal-to-noise ratio environment. In particular, the recognition accuracy of OTFS signals has been significantly improved, meeting the requirements of high accuracy and low latency for real-time signal processing, and having a wide range of application prospects.

[0123] In addition, the modular design and efficient architecture of the model enable it to exhibit superior generalization ability and adaptability in complex channel environments, while having low computational complexity and strong scalability.

[0124] At the same time, the invention provides an innovative solution for the classification requirements of more types of signals in future wireless communication systems. It can cooperate with the multi-level feature extraction ability of the cascaded neural network under low signal-to-noise ratio conditions to complete the efficient recognition and classification of low signal-to-noise ratio OTFS signals, and efficiently processes the entire process from signal preprocessing to final classification output. Even in a low signal-to-noise ratio environment, it can achieve the goal of significantly improving the signal classification effect, and can meet the requirements of high accuracy and low latency for real-time signal processing, having a wide range of practical application prospects.

[0125] Based on this system, the operation process of the method of the present invention includes:

[0126] L1. Input the generated wireless signal dataset into the cascaded neural network architecture; first, preliminarily classify the signals through the LSTM neural network, distinguishing them into high signal-to-noise ratio and low signal-to-noise ratio signals; for the complex characteristics of low signal-to-noise ratio signals, input them into the improved DTLN model for noise reduction processing; for high signal-to-noise ratio signals, directly input them into the convolutional neural network (CNN) for modulation type recognition;

[0127] L2. Input the denoised signals into the convolutional neural network (CNN) to classify and recognize multiple modulation types;

[0128] L3. By jointly using the LSTM, DTLN, and CNN models, efficient signal recognition and classification in complex low signal-to-noise ratio environments are achieved.

[0129] This method comprehensively considers the complexity of the characteristics of low signal-to-noise ratio signals. By introducing a deep learning-based denoising and classification mechanism, the accuracy of signal classification and the anti-interference ability are significantly improved; and through a flexible neural network architecture design, the signal processing error in low signal-to-noise ratio environments is effectively reduced, enhancing the robustness and reliability of the system. Experimental results show that in a low signal-to-noise ratio environment, the present invention can still achieve a high classification accuracy, maximizing the guarantee of the stability of the communication system and the user experience. This method is applicable to fields such as next-generation wireless communication, the Internet of Things, and radar signal processing.

[0130] In a specific embodiment, the operation process of this method includes:

[0131] P1. To ensure that the model can adapt to different complex channel environments, a wireless signal dataset of five typical modulation methods is generated using the Matlab2024 simulation platform, including OTFS, OFDM, 2ASK, 4FSK, and 8PSK.

[0132] When generating these signals, common channel characteristics in wireless communication are fully considered, simulating a real multipath propagation scenario, providing rich data support for model training, including the following four aspects.

[0133] (1) Channel parameter setting

[0134] The channel parameter design follows the typical characteristics in wireless communication. Specifically, three multipath paths are used to simulate the real channel environment:

[0135] Path delay: Set to [0, 4, 8] μs.

[0136] Path gain: Corresponding to [1, 0.8, 0.5] dB.

[0137] Doppler shift index: Set to [0, -3, 5], covering the frequency shift changes from stationary to high-speed moving scenarios. The selection of these channel parameters can well reflect the multipath propagation characteristics and dynamic channel changes in real communication.

[0138] (2) SNR setting

[0139] To comprehensively cover the communication environment from low SNR to high SNR, the present invention sets the SNR range to -10 dB to 10 dB, with a step of 2 dB, and is divided into 11 SNR points in total. 3000 samples are generated at each SNR point, and a total of 33000 samples are generated for the five modulation methods.

[0140] (3) Modulation methods and parameters

[0141] In the selection of modulation methods, OTFS and OFDM signals adopt the same subcarrier mapping method, including BPSK, QPSK, 8PSK, 16QAM, 64QAM, and 256QAM, to ensure the comparability between different signal types. A fixed number of samples are generated under each subcarrier mapping method. In addition, 2ASK, 4FSK, and 8PSK signals adopt classical modulation schemes, further expanding the diversity of the dataset.

[0142] (4) Data sample design

[0143] Each sample contains two I / Q signals. The sampling frequency is set to 960 kHz, and the total time step is 2220, where the number of symbols is 30 and the number of subcarriers is 64. The sample length formed by adding the padding length is (64 + 10) * 30 = 2220. The generated dataset is divided into a training set and a test set according to a ratio of 7:3 to ensure the distribution consistency of the model in the training and test stages, providing reliable data support for subsequent performance evaluation.

[0144] P2. The system of the present invention is constructed based on a cascaded neural network, which consists of three parts: an SNR classification module, a signal enhancement module, and a signal classification module. Through task division and cooperation among the modules, the signal classification performance is significantly improved, especially excellent under low SNR conditions, where:

[0145] (1) Signal-to-Noise Ratio Classification Module (LSTM): The signal-to-noise ratio classification module is mainly used to distinguish high signal-to-noise ratio and low signal-to-noise ratio signals. Through the time series feature extraction ability of the Long Short-Term Memory Network (LSTM), this module can effectively identify the signal-to-noise ratio category of the input signal. Input signal: Shape (64, 2220, 2), where 64 is the batch size, 2220 is the time step length, and 2 represents the two IQ signals. Network structure: LSTM layer: Captures the time dynamic features of the signal, with an output shape of (64, 2220, 64). Time step extraction: Takes the output features of the last time step, and the shape becomes (64, 64). Fully connected layer: Uses the Sigmoid activation function to output the signal-to-noise ratio classification probability. Function: Classifies signals with a signal-to-noise ratio higher than 4 dB as high signal-to-noise ratio category, and signals lower than or equal to 4 dB as low signal-to-noise ratio category. Performance: After 50 Epochs of training, the loss value converges to 0.14, and the classification accuracy reaches 90%.

[0146] (2) Signal Denoising Module (DTLN): The signal enhancement module is based on the improved Dual-Path Temporal Localization Network (DTLN) to perform denoising processing on low signal-to-noise ratio signals. The DTLN module significantly improves the feature representation ability of low signal-to-noise ratio signals by combining time-domain and frequency-domain feature extraction. Network structure: Encoder: Module type: 1D convolutional layer (Conv1d). Function: Converts the time-domain signal into frequency-domain features and extracts local time correlations. Configuration parameters: Number of input channels: 2 (dual-channel). Number of output channels: fft_size (512). Convolution kernel size: 512. Stride: 256. Padding: 256 (to ensure that the length of the convolutional sequence matches the input). Dual-path separation network: Time-dimensional separation path: Uses a recurrent neural network (LSTM / GRU) to capture the time context features of the signal. Frequency-dimensional separation path: Combines Fourier transform and frequency-domain modeling methods to extract frequency features. Decoder: Restores the enhanced features back to the time domain signal through deconvolution or reconstruction module. Input and output: Input shape is (64, 2220, 2), and the output shape remains the same, achieving frame-by-frame input and output. Performance: Significantly improves the signal quality of low signal-to-noise ratio signals and enhances the recognition ability of subsequent classification modules. Aiming at the problem of low recognition accuracy in the low signal-to-noise ratio region, the denoising model based on the improved Dual-Path Temporal Localization Network (DTLN) proposed in this paper has the specific structure as Figure 4 shown.

[0147] This application uses the generated dataset to train the model, and tests the denoising and reconstruction effect of the model through the modulation type recognition model. White noise obeying Gaussian distribution is added to high signal-to-noise ratio signals, and these signals with additive noise and the original high signal-to-noise ratio signals are input into the denoising model for training. Using the trained denoising model, the denoising and reconstruction processing of low signal-to-noise ratio signals is realized. The effects before and after reconstruction are asFigure 11 as shown

[0148] (3) Signal Classification Module (CNN): The signal classification module is based on a Convolutional Neural Network (CNN) and is used to classify the modulation types of the enhanced signals. Network Structure: Convolutional Layer: The convolutional kernel size is 3, the number of input channels is 2, the number of output channels is 32, and the activation function is ReLU. Max Pooling Layer: The pooling kernel size is 2. Flattening Layer: Flattens the convolutional features into a one-dimensional vector. Fully Connected Layer: First Layer: The number of output nodes is 128, and the activation function is ReLU. Second Layer: The number of output nodes is 5, corresponding to five modulation types. Performance: After 5 Epochs of training, the classification accuracy reaches 99% under high signal-to-noise ratio conditions.

[0149] P3. Execute the method of the present invention in the system framework based on the above data, test verification and result analysis:

[0150] To verify the actual performance of the model of the present invention, the classification accuracies of a single CNN model and a cascaded network model under different signal-to-noise ratio conditions were tested respectively.

[0151] Refer to Figures 7 to 10 , the test signal-to-noise ratio range is from -10 dB to 10 dB, with a step of 2 dB. Each signal-to-noise ratio is repeatedly tested 20 times, and the average value is taken.

[0152] Test Results: The test results show that the cascaded model shows significant advantages under low signal-to-noise ratio conditions. For example: at a signal-to-noise ratio of -10 dB: the classification accuracy of the single CNN model is 49.21%; the classification accuracy of the cascaded model is 50%. At a signal-to-noise ratio of 4 dB: the classification accuracy of the single CNN model is 94.16%; the classification accuracy of the cascaded model is 96%. At a signal-to-noise ratio of 10 dB: the classification accuracies of both models are close to 100%.

[0153] Combined with Figure 6 the experimental results, it can be seen from the results that the classification performance of the cascaded model is significantly better than that of the single CNN model under low signal-to-noise ratio conditions, verifying its superiority in complex channel environments.

[0154] The present invention realizes the efficient classification of signals of multiple modulation types in a low signal-to-noise ratio environment through a finely designed data set and a modular cascaded neural network architecture. The signal-to-noise ratio classification module improves the processing efficiency, the signal enhancement module significantly improves the quality of low signal-to-noise ratio signals, and the signal classification module completes the classification task with high accuracy. The test results show that this method not only performs excellently under complex channel conditions, but also has low computational complexity and strong scalability, providing a reliable solution for the real-time signal processing requirements of future wireless communication systems.

[0155] In another aspect, the present invention also discloses a computer-readable storage medium storing a computer program, which, when executed by a processor, causes the processor to execute the steps of the above method.

[0156] In yet another aspect, the present invention also discloses a computer device including a memory and a processor, where the memory stores a computer program, which, when executed by the processor, causes the processor to execute the steps of the above method.

[0157] In yet another embodiment provided by the present application, there is also provided a computer program product containing instructions, which, when running on a computer, causes the computer to execute any of the low signal-to-noise ratio OTFS signal recognition and classification methods in the above embodiments.

[0158] It can be understood that the system provided by the embodiments of the present invention corresponds to the method provided by the embodiments of the present invention. Explanations, examples, and beneficial effects of related content can refer to the corresponding parts in the above method.

[0159] The embodiments of the present application also provide an electronic device including a processor, a communication interface, a memory, and a communication bus. Among them, the processor, the communication interface, and the memory complete communication with each other through the communication bus.

[0160] The memory is used to store a computer program.

[0161] The processor is used to implement the above low signal-to-noise ratio OTFS signal recognition and classification method when executing the program stored in the memory.

[0162] The communication bus mentioned in the above electronic device may be a peripheral component interconnect standard (PCI) bus or an extended industry standard architecture (EISA) bus, etc. This communication bus can be divided into an address bus, a data bus, a control bus, etc.

[0163] The communication interface is used for communication between the above electronic device and other devices.

[0164] The memory may include a random access memory (RAM), and may also include a non-volatile memory (NVM), such as at least one disk memory. Optionally, the memory may also be at least one storage device located far from the aforementioned processor.

[0165] The above-mentioned processor may be a general-purpose processor, including a central processing unit (CPU), a network processor (NP), etc.; it may also be a digital signal processor (DSP), an application-specific integrated circuit (ASIC), a field-programmable gate array (FPGA), or other programmable logic devices, discrete gate or transistor logic devices, discrete hardware components.

[0166] It should also be noted that the electronic device further includes a terminal device, which can also be referred to as a terminal, user equipment, mobile station, mobile terminal, etc. The terminal device can be a mobile phone, smart TV, wearable device, tablet computer, computer with wireless transceiver function, virtual reality terminal device, augmented reality terminal device, wireless terminal in industrial control, wireless terminal in driverless, wireless terminal in remote surgery, wireless terminal in smart grid, wireless terminal in transportation safety, wireless terminal in smart city, wireless terminal in smart home, and so on. The embodiments of the present application do not limit the specific technologies and specific device forms adopted by the terminal device.

[0167] In the above embodiments, it can be implemented in whole or in part by software, hardware, firmware, or any combination thereof. When implemented using software, it can be implemented in whole or in part in the form of a computer program product. The computer program product includes one or more computer instructions. When the computer program instructions are loaded and executed on a computer, the processes or functions described in the embodiments of the present application are generated in whole or in part. The computer can be a general-purpose computer, a special-purpose computer, a computer network, or other programmable devices. The computer instructions can be stored in a computer-readable storage medium, or transmitted from one computer-readable storage medium to another computer-readable storage medium. For example, the computer instructions can be transmitted from a website, computer, server, or data center to another website, computer, server, or data center in a wired manner (such as coaxial cable, optical fiber, digital subscriber line) or a wireless manner (such as infrared, wireless, microwave, etc.). The computer-readable storage medium can be any available medium that can be accessed by a computer, or a data storage device such as a server or data center that includes one or more available media integrated. The available medium can be a magnetic medium (such as a floppy disk, hard disk, magnetic tape), an optical medium, or a semiconductor medium (such as a solid-state disk, Solid State Disk).

[0168] The above are only the preferred embodiments of the present invention and are not intended to limit the present invention. Any modifications, equivalent replacements, improvements, etc. made within the spirit and principles of the present invention shall be included in the protection scope of the present invention.

[0169] In addition, it should be noted that if there are directional indications (such as up, down, left, right, front, back...) involved in the embodiments of the present invention, the directional indications are only used to explain the relative position relationship and movement conditions between components in a specific posture. If the specific posture changes, the directional indications will also change accordingly.

[0170] In addition, if the descriptions such as "first" and "second" are involved in the embodiments of the present invention, the descriptions of "first", "second", etc. are for descriptive purposes only, and cannot be construed as indicating or implying their relative importance or implicitly indicating the quantity of the indicated technical features. Thus, the features defined with "first" and "second" may explicitly or implicitly include at least one such feature. In addition, the meaning of "and / or" appearing throughout the text includes three parallel scenarios. Taking "A and / or B" as an example, it includes Scenario A, or Scenario B, or the scenario where both A and B are satisfied simultaneously. In addition, in the embodiments of the present invention, "a plurality of" means two or more. In addition, the technical solutions between various embodiments can be combined with each other, but it must be based on the ability of those of ordinary skill in the art to implement. When the combination of technical solutions results in contradictions or cannot be implemented, it should be considered that such a combination of technical solutions does not exist and is not within the scope of protection required by the present invention.

Claims

1. A low signal-to-noise ratio OTFS signal recognition and classification method, characterized in that: The following steps are involved: S1. Acquire wireless communication signals and construct data sets; S2. Use LSTM network to classify high and low signal-to-noise ratios of signals in the dataset; S3. Low signal-to-noise ratio signals are denoised by a two-dimensional input DTLN model; S4. Use the convolutional neural network as the modulation signal classification module, and input the data after denoising by the DTLN model and the high signal-to-noise ratio signal into the modulation signal classification module for processing.

2. The low signal-to-noise ratio OTFS signal recognition and classification method according to claim 1, characterized in that: The data set constructed in step S1 includes OTFS, OFDM, 2ASK, 4FSK and 8PSK signals with different signal-to-noise ratios; The OTFS signal and OFDM are modulated and generated using the same channel parameter settings; The signal-to-noise ratio range of the wireless communication signals within the data set is set to be consistent.

3. The low signal-to-noise ratio OTFS signal recognition and classification method according to claim 1, characterized in that: The specific operation process of classification in step S2 includes: S21. Define the input signal data dimensions including data frame length and batch size, and take the in-phase I dimension and orthogonal Q dimension of the signal as independent features; S22. Use LSTM network to process data set data through hierarchical structure, use hidden layer to extract specified time features, model data according to the time series characteristics of signals, and use stochastic gradient descent optimizer and binary cross entropy loss as loss function to train the model; S23. Through the Sigmoid activation function, the features of the last time step are mapped to a probability value for determining the signal-to-noise ratio category; S24. Preset a classification threshold, classify signals with a signal-to-noise ratio greater than the threshold as high signal-to-noise ratio signals, and classify signals less than or equal to the threshold as low signal-to-noise ratio signals.

4. The low signal-to-noise ratio OTFS signal recognition and classification method according to claim 3, characterized in that: The specific operation process of the DTLN model for noise reduction in step S3 includes: S31. Represent the low signal-to-noise ratio signal as a time domain signal including an in-phase I dimension and an orthogonal Q dimension, including a batch size, a sample length, and an in-phase orthogonal IQ channel dimension, wherein the batch size setting is adjustable, and the IQ channel represents the real part and the imaginary part of the signal; S32. After the low signal-to-noise ratio signal is input into the DTLN model for processing, a noise reduction signal with the same input dimension is output, and the DTLN model uses the high signal-to-noise ratio signal as the target clean signal and the low signal-to-noise ratio signal as the input for supervised training; S33. The loss function is calculated by joint error calculation in time domain and frequency domain, and data optimization is performed by combining mean square error and signal-to-noise ratio gain.

5. The low signal-to-noise ratio OTFS signal recognition and classification method according to claim 4, characterized in that: The DTLN model includes: Encoder module: extracts local features of time domain signals through one-dimensional convolution and represents the encoding as high-dimensional features; Time dimension separation network: Use a dual-path separation structure to capture the temporal context characteristics of the signal, and use LSTM or GRU as the time dimension separation path to model the sequence characteristics; Frequency separation network: extracts the spectral features of the signal through short-time Fourier transform, and combines convolutional neural network or self-attention mechanism to enhance the ability to distinguish between noise and signal; Decoder module: The enhanced high-dimensional features are mapped back to the time domain signal through a deconvolution or fully connected reconstruction module to output a denoised signal with the same dimension as the input.

6. The low signal-to-noise ratio OTFS signal recognition and classification method according to claim 4, characterized in that: The data processing flow of the modulation signal classification module in step S4 includes: S41. Inputting the low signal-to-noise ratio signal after denoising into the convolutional neural network model, the input data includes batch size, time step, and in-phase orthogonal IQ channel dimension, wherein the batch size is an adjustable parameter, and the IQ channel represents the real part and imaginary part of the signal; S42. The first layer uses the Conv1 D convolution layer with a convolution kernel size of 3, an IQ signal with 2 input channels, and an adjustable output channel number parameter. The nonlinear capability is introduced through the ReLU activation function to extract the local time characteristics of the signal; S43. Use the maximum pooling layer in the second layer with a pooling kernel size of 2 to gradually reduce the feature dimension to reduce the computational complexity; S44. The pooled features are flattened into a one-dimensional vector using the Flatten layer in the third layer; S45. Input the processed vector data into the fully connected network for classification processing.

7. The low signal-to-noise ratio OTFS signal recognition and classification method according to claim 4, characterized in that: The fully connected network consists of two sets of fully connected layers: The first fully connected layer contains multiple neurons and uses the ReLU activation function to enhance feature expression capabilities; The second fully connected layer is used to output the classification results. It contains multiple nodes, each corresponding to a multiple modulation type. The Softmax activation function is used to calculate the probabilities of all modulation types in the fully connected network. The classification result is the modulation type with the highest probability.

8. A low signal-to-noise ratio OTFS signal recognition and classification system, used to execute the low signal-to-noise ratio OTFS signal recognition and classification method according to any one of claims 1 to 6, characterized in that: The cascade neural network structure is composed of a signal-to-noise ratio classification module, a signal enhancement module, and a signal classification module, where: The signal-to-noise ratio classification module is used to preliminarily classify the input signal according to the signal-to-noise ratio; A signal enhancement module, used for performing noise reduction processing on low signal-to-noise ratio signals, wherein the signal enhancement module is constructed based on a dual-path real-time denoising network; The signal classification module is used to classify the modulation type of the enhanced signal, and the signal classification module is constructed based on a convolutional neural network.

9. The low signal-to-noise ratio OTFS signal recognition and classification system according to claim 8, characterized in that: The specific signal processing flow of the signal-to-noise ratio classification module includes: a1. Capture the time series features of the signal in the LSTM layer and output it as a high-dimensional feature representation; a2. Extract the features of the last time step to obtain a low-dimensional feature vector; a3. Through the fully connected layer, use the Sigmoid activation function to output the signal-to-noise ratio classification result.

Images