OFDM high-order modulation signal recognition method based on adaptive data preprocessing and adaptive wavelet M-band algorithm

Abstract

The application discloses an OFDM high-order modulation signal recognition method based on adaptive data preprocessing and an adaptive wavelet Mamba algorithm, solves how to classify OFDM signals under unknown channel parameters, Doppler frequency shift and complex multipath propagation conditions, and belongs to the field of modulation carriers; the method comprises the following steps: storing OFDM modulation signals as IQ files by simulating actual communication environments through a physical channel simulator; calculating a stable interval of the IQ files by using a steady-state distribution method, cleaning and retaining target samples whose amplitudes meet a threshold; dividing the target samples into sub-blocks, reconstructing the sub-blocks into a structured three-dimensional matrix, and dividing the three-dimensional matrix into a training set and a test set; constructing an AWM from sequentially connected convolution layers, a wavelet transform module, a frequency band perception module, an adaptive gating fusion mechanism module and a fully connected layer; obtaining the trained AWM from the training set; and identifying OFDM high-order modulation signals through the test set; and the application realizes OFDM signal high-order modulation classification.

Description

Technical Field

[0001] This invention belongs to the field of modulation carrier technology and is specifically related to the field of wireless signal modulation recognition technology. It relates to a method for recognizing high-order OFDM modulation signals based on adaptive data preprocessing and adaptive wavelet Mamba algorithm. Background Technology

[0002] Automatic modulation recognition (AMR) plays a crucial role in modern wireless communication and electromagnetic control. In the future 6G communication field, AMR will play a vital role; essentially, it is an adaptive signal recognition system that drives the entire transmission process. In the 6G era, AMR will become a core supporting technology for building low-latency, high-efficiency, and high-security communication systems. Simultaneously, utilizing control signals to achieve adaptive and intelligent communication will become a key research and development direction.

[0003] Orthogonal Frequency Division Multiplexing (OFDM) technology, with its excellent spectral efficiency and superior performance in high-speed data transmission and multipath interference resistance, has become a fundamental component of fifth-generation (5G) and subsequent wireless communication systems. In adaptive OFDM transceiver systems, the transmitter dynamically selects an appropriate modulation format based on the real-time communication environment, while the receiver accurately decodes the received data by identifying this modulation format. To improve the high data rate of wireless communication systems, more and more modulation schemes and higher-order modulation formats are needed. Because future 6G will encounter signal attenuation and distortion in unknown channel parameters, complex multipath channels, and high-speed mobile conditions, adaptive signal recognition tasks face significant challenges and complexity. Therefore, the designed AMR algorithm needs to accurately capture dynamic changes in the channel under Doppler effects, unknown channel parameters, and complex multipath conditions to ensure a high recognition rate in this environment.

[0004] Raw OFDM modulated signals are often affected by various noises, interferences, and environmental factors. Data cleaning, standardization, and preprocessing can significantly reduce the impact of noise and interference on the recognition results, ensuring that the extracted features are closer to the real signal. Data segmentation and sampling techniques can effectively capture the instantaneous modulation characteristics of the signal, ensuring sufficient representative samples for training and testing across different time periods, thereby improving the model's generalization ability and robustness. Proper data preprocessing and feature engineering also enable subsequent machine learning or deep learning algorithms to more accurately capture key differences in modulation formats, which is particularly important for improving automatic classification accuracy and real-time response capabilities.

[0005] In early studies based on the RML2016 and RML2018 datasets, the data processing workflow typically involved preprocessing the raw IQ data to extract and normalize the real and imaginary parts. Then, the signal was segmented into fixed-length samples, and each sample was labeled according to its modulation scheme at the time of acquisition. This method played a crucial role in constructing automatic modulation classification benchmark datasets. However, when processing OFDM signals, due to the characteristics of OFDM signals, such as multiple subcarriers and cyclic prefixes, traditional fixed-length segmentation often implicitly assumes that the signal is stationary and homogeneous within each segment. This simple segmentation process easily disrupts important dependencies between subcarriers, resulting in extracted features that cannot fully capture the actual modulation characteristics. Simultaneously, labeling the entire signal segment with a single modulation label fails to reflect the subtle differences within the OFDM signal caused by resource allocation and channel effects, potentially obscuring certain crucial information. Furthermore, since these datasets are typically generated using idealized or simplified channel models, preprocessing methods struggle to fully simulate complex factors in real-world wireless environments, such as multipath effects, frequency shifts, and nonlinear distortions. Even if significant noise and anomalous samples are filtered out during the data cleaning phase, they cannot accurately handle edge effects or slight distortions caused by cyclic prefixes, thus losing some potential discriminative information.

[0006] In summary, while the RML2016 and RML2018 datasets provide valuable benchmark data for automatic modulation classification research, they have shortcomings in areas such as OFDM signal segmentation, labeling, and simulation of complex channel effects. These shortcomings limit the representativeness of the data to real-world complex communication environments and the robustness of the AMR algorithm. Future improvements should focus on designing more reasonable segmentation strategies for OFDM signal characteristics, optimizing labeling mechanisms, and introducing more realistic channel models to provide more representative and fine-grained data support for automatic modulation identification. Figure 1 The diagram illustrates a general method for processing communication modulation signal datasets.

[0007] O'Shea et al. O'Shea TJ, West N. Radio Machine Learning DatasetGeneration with GNU Radio[C] / / Proceedings of the GNU Radio Conference. 2016. and TJ O'Shea, T. Roy and TC Clancy, "Over-the-Air Deep Learning Based Radio Signal Classification," in IEEE Journal of Selected Topics in SignalProcessing, vol. 12, no. 1, pp. 168-179, Feb. 2018, doi: 10.1109 / JSTSP.2018.2797022., as Figure 2 As shown, the data is normalized to eliminate non-realistic features that may be introduced during the simulation process. Each sampled data vector (128 sampling points in this paper) is normalized to unit energy, thereby removing numerical amplitude differences that may simply be caused by the simulation process. This allows subsequent machine learning algorithms to focus on learning the valuable features of the signal itself, rather than biases at various scales. Finally, the normalized signal data is randomly selected for time periods and packaged according to a specific format, saving the data in a format that meets the requirements of the machine learning environment.

[0008] Tekbıyık, Kürşat, et al. "Robust and fast automatic modulation classification with CNN under multipath fading channels." 2020IEEE 91st Vehicular Technology Conference (VTC2020-Spring). IEEE, 2020. As in Figure 3 As shown, each signal in the fixed slice dataset contains 1024 I / Q samples to ensure consistent input data length. Each fixed-length data segment is normalized, such as unit energy normalization or zero-mean unit variance normalization, to eliminate amplitude differences that may be introduced during the generation process. The data is then labeled according to the modulation type before being stored.

[0009] The aforementioned datasets typically employ fixed-length I / Q sample slices and normalization, inputting the signal as continuous time-domain data into the model. However, OFDM signals exhibit a distinct frame structure, including cyclic prefixes, symbol spacing, and subcarrier orthogonality. This processing method may fail to adequately preserve the time-frequency characteristics of OFDM signals and the relevant information between different subcarriers, thus affecting the model's ability to capture their unique multicarrier structure.

[0010] Data slices are often truncated to a predetermined length, such as 128 or 1024 I / Q samples. This fixed-length design may not match the frame structure of OFDM signals. For example, an OFDM symbol may contain a cyclic prefix and a data portion, causing the cyclic prefix information to be truncated or mixed into the data portion, making it impossible to extract the structural characteristics that are crucial for synchronization and frequency domain processing.

[0011] During data generation and preprocessing, simulation models or acquisition devices may generate some abnormal or distorted samples in actual applications. If the data cleaning process does not include a rigorous detection and removal step for these abnormal data, outliers may be mixed into the training set, thus affecting the robustness and generalization ability of the model.

[0012] Despite the progress made in AMR technology for OFDM systems in recent years, many challenges remain. Current methods mainly focus on low-order or single modulation schemes and relatively simple multipath channels, with limited performance when dealing with high-order modulation and complex multipath classification. For example, application number 202010102659.3, publication number CN111343115A, entitled "A Method and System for Identifying Modulation Signals in 5G Communication," describes a method comprising: extracting features from a digital signal to obtain high-order cumulant features; performing dimensionality reduction on the high-order cumulant features to obtain dimensionality-reduced high-order cumulant features; assigning weights to the dimensionality-reduced high-order cumulant features to obtain a weighted feature combination; and, based on the KNN algorithm, obtaining the modulation type identification result of the digital signal according to the weighted feature combination. This method has limitations in feature extraction. For example, cross-models based on KNN or CNN, LSTM and other algorithms mainly emphasize local features in the time or frequency domains, and cannot clearly capture information in the global frequency and time domains, which is crucial for the accurate classification of OFDM signals. The existing model structure exhibits black-box characteristics in decision-making, with poor interpretability, which to some extent limits its value in practical applications and theoretical extensions. Summary of the Invention

[0013] To address the technical challenge of classifying high-order OFDM modulation signals under conditions of unknown channel parameters, Doppler shift, and complex multipath propagation in Internet access and related services, this invention provides a method for identifying high-order OFDM modulation signals based on adaptive data preprocessing and an adaptive wavelet Mamba algorithm. The method utilizes the adaptive wavelet Mamba algorithm to perform adaptive modulated signal (AMR) on the high-order OFDM modulation signal after adaptive preprocessing of the received signal. The input signal is explicitly decomposed using an adaptive wavelet network via a wavelet Mamba network to extract the optimal approximation and detail components. Subsequently, leveraging the Mamba's ability to handle long sequences and its adaptive selection capability, the method effectively focuses on the most discriminative time-frequency components, thereby improving the accuracy of modulation identification and enhancing the model's effectiveness and interpretability. The effectiveness of the proposed method in high-order modulation identification is verified using modulation schemes such as Binary Phase Shift Keying (BPSK), Quadrature Phase Shift Keying (QPSK), 8PSK, 4PAM, and M-QAM with M = 16, 64, 256, 512, and 1024. The proposed AWM was validated on real-time transceiver datasets generated by an RF test bench and a physical channel simulator. The AWM disclosed in this invention enables high-order modulation classification of OFDM signals under conditions of unknown channel parameters, Doppler shift, and complex multipath propagation in Internet access and related services.

[0014] The objective of this invention is specifically achieved through the following technical solutions:

[0015] This invention discloses a method for identifying high-order OFDM modulation signals based on adaptive data preprocessing and adaptive wavelet Mamba algorithm. The method includes:

[0016] Step 1: Use the FPGA software in the USRP device to transmit an OFDM modulated signal of a preset modulation type; configure the impulse response of the multipath channel, normalized carrier frequency deviation, phase deviation and Gaussian additive white noise of the received OFDM modulated signal through the physical channel simulator to simulate the actual communication environment, obtain the simulated OFDM modulated signal, and save it as an IQ file.

[0017] Step 2: Calculate the stable interval of the IQ file using the steady-state distribution method, collect sample data within the stable interval, and take the odd-numbered data as the real part and the even-numbered data as the imaginary part; calculate the amplitude after normalizing each sample data, and clean and retain the target samples whose amplitude meets the threshold.

[0018] Step 3: Divide the target sample into sub-blocks according to the sample length and sub-block interval, stack all sub-blocks in the receiving order, and reconstruct a structured three-dimensional matrix; store the structured three-dimensional matrix as a dataset after forming key-value pairs according to modulation type and signal-to-noise ratio, and divide the dataset into training set and test set;

[0019] Step four involves constructing a wavelet Mamba classifier network (AWM) using sequentially connected convolutional layers, wavelet transform modules, frequency band sensing modules, adaptive gated fusion mechanism modules, and fully connected layers. The training set is input into the AWM, where features are extracted through convolutional layers. The wavelet transform module decomposes the features into high- and low-frequency components. The frequency band sensing module performs frequency band sensing time-series modeling on the high- and low-frequency components separately. After adaptive gated fusion by the adaptive gated fusion mechanism module, the classifier model is trained through fully connected layers, resulting in the trained AWM. The test set is then input into the trained AWM for testing, identifying high-order OFDM modulation signals.

[0020] In step one, the IQ file is:

[0021] ;

[0022] In the formula, Here, n represents the analog OFDM modulated signal in the IQ file, n is the sampling point number of the analog OFDM modulated signal in time, e is the base of the natural logarithm, and j is the imaginary unit. The normalized carrier frequency offset is given, where N is the number of subcarriers in the OFDM symbol, and represents the number of points used in the Inverse Fast Fourier Transform (IFFT) to generate the time-domain signal. For phase deviation, For the summation index, The impulse response is given by a frequency-selective fading channel of length L. OFDM modulated signals with a cyclic prefix CP transmitted by the FPGA software in the USRP. Preset modulation types for OFDM modulated signals, including 8PSK, BPSK, 4PAM, 16QAM, 64QAM, 512QAM, 1024QAM, QPSK, and 256QAM. It is Gaussian additive white noise.

[0023] In step two, the steady-state distribution method is used to calculate the stable interval of the IQ file. Sample data is collected within the stable interval, and the odd-numbered data points in the sample data are taken as the real part, and the even-numbered data points as the imaginary part.

[0024] Using the steady-state distribution method to read the number of pre-skipped sampling points in the IQ file and the number of skipped sampling points ;

[0025] The starting index S and ending index E, determined by the number of pre-skipped sampling points and the number of post-skipped sampling points, constitute the stable interval; the calculation methods for the starting index S and ending index E are as follows: ;

[0026] Collect sample data within the stable interval The real and imaginary parts of the sample data are obtained; the calculation methods for the real and imaginary parts are as follows: ; ;

[0027] In the formula, The virtual part, This refers to even-numbered positions in the sample data. For the real part, These are the odd-numbered positions in the sample data, where n is the sampling point number of the analog OFDM modulated signal in time. .

[0028] In step two, the calculation method for calculating the amplitude after normalizing each sample data, and cleaning and retaining target samples whose amplitude meets the threshold is as follows:

[0029] [n] = , [n] = ;

[0030] ;

[0031] Cleaning satisfies , or sample data containing NaN;

[0032] In the formula, [n] is the normalized imaginary part. [n] is the normalized real part. For amplitude, Minimum power threshold, This represents the maximum power threshold; NaN indicates an invalid operation.

[0033] In step three, the structured three-dimensional matrix is:

[0034] ,[2,320,6]);

[0035] In the formula, Y is a structured three-dimensional matrix. These are array manipulation functions used to change the shape of an array; [2,320,6] represents the sub-blocks stacked sequentially according to the receiving order, and [2,320,6] is the shape of the structured three-dimensional matrix.

[0036] In step four, the method for extracting features from the training set using convolutional layers is as follows:

[0037] ;

[0038] In the formula, The features extracted by the convolutional layers are represented by ConvBlock, which consists of a 2D convolutional layer conv1 with a kernel size of (2,7), a 1D convolutional layer conv2, and a 1D convolutional layer conv3 with a stride of 2, connected sequentially. This is the training set.

[0039] In step four, the method for decomposing the features into high- and low-frequency components using the wavelet transform module includes:

[0040] S1 divides the features into non-overlapping odd components. and even components ;in, ; ; For odd-numbered positions in the features, For even-numbered positions in the features;

[0041] S2, the prediction error d[n] of the even-numbered components predicting the neighboring odd-numbered components, is taken as the high-frequency component; where, In the formula, P(·) is the prediction operator, which predicts the odd components with the even components;

[0042] S3, using the high-frequency components to update the even-numbered components to obtain the low-frequency component c[n]; where... In the formula, U(·) is the update operator, which updates the even-numbered components with the high-frequency components.

[0043] In step four, the method for performing frequency band sensing time series modeling on the high and low frequency components using the frequency band sensing module is as follows:

[0044] The frequency band sensing module consists of a parallel low-frequency frequency band sensing Mamba module and a high-frequency frequency band sensing Mamba module.

[0045] The low-frequency band-aware Mamba module transposes the low-frequency components and assigns a fixed band ID to each time step to obtain a low-frequency band embedding vector. This low-frequency band embedding vector is then summed with the low-frequency components to obtain the low-frequency band vector. A bias term is added to drive the evolution of the hidden state, resulting in the constructed low-frequency band-aware timing module. Simultaneously,

[0046] The high-frequency band sensing Mamba module is used to concatenate high-frequency components of all scales in the time dimension to form a high-frequency sequence; and to assign a frequency band ID corresponding to the source level to each time step in the high-frequency sequence to obtain a high-frequency band embedding vector. After adding the high-frequency band embedding vector to the high-frequency components, the high-frequency band vector is obtained. After adding the bias term, the evolution of the hidden state is driven to obtain the constructed high-frequency band sensing time sequence module.

[0047] In step four, the adaptive gating fusion method performed by the adaptive gating fusion mechanism module is as follows:

[0048] The low-frequency band sensing timing module and the high-frequency band sensing timing module are compressed separately in the time dimension and then concatenated along the time dimension to obtain a concatenated sequence. The average value of the concatenated sequence is calculated to obtain the descriptor, and the gating weight of the descriptor is output. The gating weight is then used to perform a weighted summation of the compressed low-frequency band sensing timing module and the high-frequency band sensing timing module to obtain the fused feature.

[0049] In step four, the method for training the classifier model through fully connected layers is as follows:

[0050] ;

[0051] ;

[0052] In the formula, The predicted probability is the output of the fully connected layer. Regularization functions are used to prevent overfitting in neural networks. For activation function, This is the weight matrix of the fully connected layer. As a feature of fusion, This is the bias vector of the fully connected layer. This is the classifier model output by the fully connected layer. This is the weight matrix of the classifier model. This is the bias vector of the classifier model.

[0053] The beneficial effects of this invention are:

[0054] 1. Random sampling is used to extract stable intervals in the IQ file to ensure that the sample data is diverse and retains the complete time-domain frame structure, thereby avoiding the loss of cyclic prefix and symbol interval information in the analog OFDM modulated signal due to fixed-length slices;

[0055] 2. In the sample data cleaning process, a refined outlier detection was added. By calculating the amplitude of each sample data and combining it with the set threshold, it is ensured that the slicing operation can distinguish between the cyclic prefix and the data part, retain the symbol interval and subcarrier orthogonality information, improve the existing fixed-length extraction method, make it more suitable for the inherent frame synchronization characteristics of OFDM signals, effectively remove abnormal and distorted samples caused by equipment errors or simulation model defects, avoid mixing into the training set, and improve data quality and model robustness.

[0056] 3. When reconstructing the data, the original fixed-length data is reorganized into a data form that conforms to the OFDM time-frequency structure. While maintaining the data shape to adapt to the deep learning model, local time-frequency transformation and preservation of symbol boundary information are considered to ensure that the model can capture the intrinsic correlation between subcarriers. The fixed-length sub-blocks are reconstructed into multi-dimensional tensors that meet the input requirements of the deep learning network based on the preset subcarrier number and OFDM symbol number parameters. This more completely preserves the relevant information between subcarriers as well as synchronization and frequency domain characteristics. The reconstructed structured three-dimensional matrix fully preserves the time-frequency characteristics of the OFDM signal and the intrinsic correlation information between subcarriers, improving the deep learning model's ability to capture multi-carrier structures.

[0057] 4. The technical solution disclosed in this invention constructs an end-to-end bridge for cross-time and cross-frequency information interaction, enabling the state-space model to "see the frequency band," thereby achieving efficient and interpretable modulation classification under complex interference channels. This method not only achieves a significant improvement in classification performance but also has obvious advantages in terms of parameter quantity and computational complexity. Attached Figure Description

[0058] The present invention will now be described in further detail with reference to the accompanying drawings and embodiments.

[0059] Figure 1 This is a schematic diagram of the communication modulation signal data processing flowchart provided in this invention.

[0060] Figure 2 This is a schematic diagram of the data processing flowchart for the RML2016a and RML2018.01a datasets provided in this invention example.

[0061] Figure 3 This is a schematic diagram of the HisarMod2019.1 dataset data processing flowchart provided in this invention example.

[0062] Figure 4 This is a schematic diagram of the overall flowchart for the creation and data processing of the dataset provided in this invention example.

[0063] Figure 5 This is a schematic diagram of the software configuration of OFDM transmitted by USRP provided in this invention example.

[0064] Figure 6 This is a schematic diagram of the physical channel simulator configuration provided in this invention example.

[0065] Figure 7 This is a schematic diagram of the USRP data retention software provided in this invention example.

[0066] Figure 8This is a schematic diagram of the OFDM modulated signal data processing flowchart of the steady-state block random sampling and data reconstruction algorithm provided in this invention.

[0067] Figure 9 This is a schematic diagram of the AWM network structure provided in this invention example.

[0068] Figure 10 This is a schematic diagram illustrating the classification and recognition rate of OFDM high-order modulation signals under a deep learning model, as provided in this invention example.

[0069] Figure 11 This is a schematic diagram of the confusion matrix of the AWM model after data preprocessing provided in this invention at 10dB.

[0070] Figure 12 This is a schematic diagram of the confusion matrix of the AWM model without data preprocessing provided in this invention at 10dB. Detailed Implementation

[0071] To enable those skilled in the art to better understand the present invention, the technical solutions of the present invention will be clearly and completely described below with reference to the accompanying drawings of the embodiments of the present invention. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort should fall within the scope of protection of the present invention.

[0072] like Figure 4 As shown, this embodiment of the invention provides a method for identifying high-order OFDM modulation signals based on adaptive data preprocessing and adaptive wavelet Mamba algorithm. The method includes:

[0073] Step 1: Use the FPGA software in the USRP device to transmit an OFDM modulated signal of a preset modulation type, such as... Figure 5 As shown, in the FPGA software, the frequency band, power, and bandwidth corresponding to the task requirements are set, and an OFDM modulated signal of the preset modulation type is transmitted; through, as shown... Figure 6 The physical channel simulator shown simulates a real communication environment by configuring the impulse response of the multipath channel, normalized carrier frequency offset, phase offset, and Gaussian additive white noise on the received OFDM modulated signal, thus obtaining a simulated OFDM modulated signal. Figure 7 The USRP data retention software at the receiving end receives and retains the data as an IQ file;

[0074] Step 2: Calculate the stable interval of the IQ file using the steady-state distribution method, collect sample data within the stable interval, and take the odd-numbered data as the real part and the even-numbered data as the imaginary part; calculate the amplitude after normalizing each sample data, and clean and retain the target samples whose amplitude meets the threshold.

[0075] Step 3: Divide the target sample into sub-blocks according to the sample length and sub-block interval, stack all sub-blocks in the receiving order, and reconstruct a structured three-dimensional matrix; store the structured three-dimensional matrix as a dataset after forming key-value pairs according to modulation type and signal-to-noise ratio, and divide the dataset into training set and test set;

[0076] Step four, as Figure 9 As shown, a wavelet Mamba classifier network (AWM) is constructed by sequentially connecting convolutional layers, wavelet transform modules, frequency band sensing modules, adaptive gated fusion mechanism modules, and fully connected layers. The training set is input into the AWM, where features are extracted through convolutional layers. The wavelet transform module decomposes the features into high- and low-frequency components. The frequency band sensing module performs frequency band sensing time-series modeling on the high- and low-frequency components respectively. After adaptive gated fusion by the adaptive gated fusion mechanism module, the classifier model is trained through a fully connected layer, resulting in a trained AWM. The test set is then input into the trained AWM for testing, identifying high-order OFDM modulation signals.

[0077] In step one, the IQ file is:

[0078] ;

[0079] In the formula, Here, n represents the analog OFDM modulated signal in the IQ file, n is the sampling point number of the analog OFDM modulated signal in time, e is the base of the natural logarithm, namely Euler's number, which is a mathematical constant approximately equal to 2.71828, and j is the imaginary unit, satisfying... , The normalized carrier frequency offset is given, where N is the number of subcarriers in the OFDM symbol, and represents the number of points used in the Inverse Fast Fourier Transform (IFFT) to generate the time-domain signal. For phase deviation, For summation index, also representing the discrete-time index of I[k], indicating different paths in a multipath channel; The impulse response is given by a frequency-selective fading channel of length L. OFDM modulated signals with a cyclic prefix CP transmitted by the FPGA software in the USRP. Preset modulation types for OFDM modulated signals, including 8PSK, BPSK, 4PAM, 16QAM, 64QAM, 512QAM, 1024QAM, QPSK, and 256QAM. It is Gaussian additive white noise.

[0080] The IQ file integrates multiple influencing factors in OFDM systems, including the impulse response of multipath channels, carrier frequency deviation, phase deviation, and additive noise interference, and can more accurately describe the characteristics of received signals in real-world environments.

[0081] In step two, the steady-state distribution method is used to calculate the stable interval of the IQ file. Sample data is collected within the stable interval, and the odd-numbered data points in the sample data are taken as the real part, and the even-numbered data points as the imaginary part.

[0082] Using the steady-state distribution method to read the number of pre-skipped sampling points in the IQ file and the number of skipped sampling points ;

[0083] The starting index S and ending index E, determined by the number of pre-skipped sampling points and the number of post-skipped sampling points, constitute the stable interval; the calculation methods for the starting index S and ending index E are as follows: ;

[0084] Collect sample data within the stable interval The real and imaginary parts of the sample data are obtained; the calculation methods for the real and imaginary parts are as follows: ; ;

[0085] In the formula, The virtual part, This refers to even-numbered positions in the sample data. For the real part, These are the odd-numbered positions in the sample data, where n is the sampling point number of the analog OFDM modulated signal in time. .

[0086] In step two, the calculation method for calculating the amplitude after normalizing each sample data, and cleaning and retaining target samples whose amplitude meets the threshold is as follows:

[0087] [n] = , [n] = ;

[0088] ;

[0089] Cleaning satisfies This process helps eliminate temporal discontinuities and noise in the sample data, providing a clearer and cleaner input.

[0090] In the formula, [n] is the normalized imaginary part. [n] is the normalized real part. For amplitude, Minimum power threshold, This represents the maximum power threshold; NaN indicates an invalid operation.

[0091] Each sample is of a predetermined length. The interval between blocks can be determined by random or sequential sampling, dividing the continuous data in the cleaned target sample into blocks; Data of each sub-block [n], [n] In the formula, The range of values ​​is determined by the total length of the sample. After stacking all sub-blocks in the order of reception, they are reconstructed into a structured three-dimensional matrix to capture the correlation between symbols and between symbols.

[0092] In step three, the structured three-dimensional matrix is:

[0093] ,[2,320,6]);

[0094] In the formula, Y is a structured three-dimensional matrix. These are array manipulation functions used to change the shape of an array; The sub-blocks are stacked sequentially according to the receiving order. [2,320,6] represents the shape of the structured three-dimensional matrix, indicating that the two-dimensional matrix [2,1920] with a signal sampling length of 1920 is reconstructed into [2,320,6]; 2 represents the I / Q channel dimension.

[0095] After completing all data cleaning and necessary reconstruction processes, the entire dataset is randomly shuffled to enhance data randomness. Key-value pairs extracted from the filenames based on modulation type and signal-to-noise ratio are stored in a data dictionary. Finally, the dictionary is serialized and saved to a specified directory, thus fully realizing the entire preprocessing workflow from extraction, cleaning, and format reconstruction of raw IQ data to final data storage. The processed structured three-dimensional matrix, in 2x320x6 format, is then stored as a dataset named "Dataset".

[0096] In step four, the method for extracting features from the training set using convolutional layers is as follows:

[0097] ;

[0098] In the formula, The features extracted by the convolutional layers are represented by ConvBlock, which consists of a 2D convolutional layer conv1 with a kernel size of (2,7), a 1D convolutional layer conv2, and a 1D convolutional layer conv3 with a stride of 2, connected sequentially. The training set is used as the network input. The training signal dimension of the convolutional layer network is [B, L, M], where B is the batch size, L is the number of I and Q channels of the signal, and M is the signal sampling length. The OFDM signal first passes through a two-dimensional convolutional layer Conv1 with a kernel size of (2, 7). This design can process both I and Q data simultaneously, effectively capturing their local spatiotemporal correlations and mapping the input channels to M feature channels. Subsequently, the signal flows through two one-dimensional convolutional layers, Conv2 and Conv3, for deeper feature extraction in the time dimension. Conv3 has a stride of 2, which enhances features while achieving downsampling, reducing the computational complexity for subsequent multi-scale analysis.

[0099] One-dimensional convolution operation can be represented as:

[0100] ;

[0101] In the formula, Here, c represents the feature extracted by the one-dimensional convolutional layer, and i represents the index of the signal sequence. For activation function, It is the first The weights of each convolutional kernel, It is the first The bias of each convolution kernel It refers to the kernel size; the downsampling effect is reflected in the index. The use of this technology reduces the number of time steps. The signal shape after passing through the convolutional layer is [B, 64, L'].

[0102] In step four, the method for decomposing the features into high- and low-frequency components using the wavelet transform module includes:

[0103] S1 divides the features into non-overlapping odd components. and even components Used to separate different frequency components of a signal; among which, ; ; For odd-numbered positions in the features, For even-numbered positions in the features;

[0104] S2, the prediction error d[n] of the even-numbered components predicting the neighboring odd-numbered components, is taken as the high-frequency component; where, In the formula, P(·) is the prediction operator that predicts odd components with even components; the prediction error d[n] captures the rapidly changing parts of the signal that cannot be predicted from its smooth trend (represented by even points), thus constituting high-frequency details.

[0105] S3, using the high-frequency components to update the even-numbered components to obtain the low-frequency component c[n]; where... In the formula, U(·) is the update operator, which updates the even-numbered components with the high-frequency components to ensure that the new low-frequency components can better maintain the global statistical characteristics of the original signal. The generated low-frequency component c[n] is smoother and accurately expresses the low-frequency trend information of the original signal.

[0106] The wavelet transform module decomposes features into low-frequency and high-frequency components across multiple frequency bands. The low-frequency components capture the global trend of the signal, while the high-frequency components highlight its detailed features, thus significantly improving the ability to extract signal features at different time resolutions. Compared to traditional wavelet basis functions, this adaptive method better adapts to specific signal features, making it suitable for data-driven learning tasks while enhancing interpretability.

[0107] In step four, the method for performing frequency band sensing time series modeling on the high and low frequency components using the frequency band sensing module is as follows:

[0108] The frequency band sensing module consists of a parallel low-frequency frequency band sensing Mamba module and a high-frequency frequency band sensing Mamba module, used to capture the long-range dependencies inherent in the low-frequency and high-frequency components respectively.

[0109] The low-frequency band sensing Mamba module is used to transpose low-frequency components, i.e. A fixed frequency band ID is assigned to each time step to obtain the low-frequency band embedding vector. 'b' is the sample index in the batch dimension. The position index is defined in the sequence length dimension. The low-frequency band embedding vector is summed with the low-frequency component to obtain the low-frequency band vector. A bias term is added to drive the evolution of the hidden state, resulting in the constructed low-frequency band sensing temporal module. Simultaneously...

[0110] The high-frequency band sensing Mamba module is used to handle high-frequency components at all scales. By splicing them together along the time dimension, a high-frequency sequence is formed. ; in, And assign a frequency band ID corresponding to the source level to each time step in the high-frequency sequence to obtain the high-frequency band embedding vector. The high-frequency band embedding vector is summed with the high-frequency components to obtain the high-frequency band vector. After adding the bias term, the evolution of the hidden state is driven to obtain the constructed high-frequency band sensing timing module.

[0111] The calculation methods for the low-frequency band vector and the high-frequency band vector are as follows:

[0112] ;

[0113] Where t is the time step; Let t be the t-th input vector. This is the frequency band ID, an integer used to identify the frequency band to which xt belongs. For example, 0 represents low-frequency components, 1 represents high-frequency details in the first layer of decomposition, 2 represents the second layer, and so on. This is a key label for implementing "frequency band awareness". For the frequency band embedding matrix, learn a dedicated, dense vector representation for each discrete frequency band ID (e.g., 0, 1, 2...). For frequency band embedding vectors, from According to the matrix The specific vector retrieved. It is optimized during training to capture the unique characteristics of that frequency band. To augment the feature vector, this involves taking the original features... and its frequency band identity The new feature obtained after addition.

[0114] Core parameters of the Mamba module , , and ;exist During generation, an additional learnable bias term related to the frequency band ID is added:

[0115] ;

[0116]

[0117]

[0118] The activation function Softplus is a smooth, always positive activation function that ensures the discretization step size. It is always a positive number; , and The first, second, and third projection layers are standard fully connected layers responsible for augmenting the feature vectors. Mapping to the generation of Mamba kernel parameters ( , , ) Required dimensions; , and Here are the first parameter bias embedding matrix, the second parameter bias embedding matrix, and the third parameter bias embedding matrix; there is also a learnable lookup table that does not directly affect the features but rather the parameter generation process. Each matrix learns a unique bias vector for each frequency band ID. for The frequency band offset vector, which is for the frequency band The learned specific bias. It is added to proj_ The generated dynamic value is for this frequency band. Set a default adjustment baseline. , and These are frequency band-aware dynamic parameters, and they are doubly dynamic: they depend on both the current input and... The content also depends on the frequency band it belongs to. .

[0119] The newly generated parameters, now incorporating frequency band characteristics, are used to actually drive the evolution of the hidden state. This step follows the standard State-Space Model (SSM) computation process, but all its key components have now been infused with frequency band information:

[0120]

[0121] Here and Both are frequency band sensing. Therefore, the discrete-time state transition matrix... and input matrix The properties of A will change dynamically depending on whether the signal being processed is low-frequency or high-frequency. A is a continuous-time state transition matrix.

[0122]

[0123] Hidden state The rate of evolution and the degree of absorption of new inputs, and the final output. The calculation method ultimately depends on . and The hidden state is for the model's memory. It is a sequence information summary up to the previous time step, while It combines the current input The updated memory afterwards. For the discretized state matrix, use the formula It is calculated. It is not a directly learned parameter, but rather based on A and frequency band sensing. Dynamically derived, it represents the actual evolution of the system state within a discrete time step t. To discretize the input matrix, use the formula The calculation yields the result. It is also dynamically derived. It represents the current input within the discrete time step t. For hidden state The actual impact. The output vector of the signal after passing through the Mamba module, the final output at time step t, is obtained through... Calculation. It is based on frequency band sensing. From the updated memory The result obtained from the reading.

[0124] In step four, the adaptive gating fusion method performed by the adaptive gating fusion mechanism module is as follows:

[0125] Low-frequency band sensing timing module and high-frequency band sensing timing module After compression in the time dimension, the compressed low-frequency band sensing timing module is obtained. and compressed high-frequency band sensing timing module Each represents its own global information; they are concatenated along the time dimension to obtain a concatenated sequence. The average value of the concatenated sequence is calculated to obtain the descriptor z, and then the gating weight g of the descriptor is output. The gating weight is used to weight and sum the compressed low-frequency band sensing time-series module and the high-frequency band sensing time-series module to obtain the fused feature. The calculation method includes:

[0126] ; ;

[0127]

[0128]

[0129] in, The length of the low-frequency sequence. The length of the high-frequency sequence. To connect two tensors along a specified dimension, It is the Sigmoid function, ensuring . For the gated weight matrix, The gated bias vector is used with the gated weights. Weighted summation yields the final fusion characteristics. :

[0130] ;

[0131] This gating mechanism is adaptive: the network learns autonomously to determine whether it represents a macro trend based on the characteristics of the input signal. ) or local details ( The contribution ratio of the two is more important to the final task, thus dynamically adjusting the contribution ratio of the two.

[0132] In step four, the method for training the classifier model through fully connected layers is as follows:

[0133] ;

[0134] ;

[0135] In the formula, The predicted probability is the output of the fully connected layer. Regularization functions are used to prevent overfitting in neural networks. For activation function, This is the weight matrix of the fully connected layer. As a feature of fusion, This is the bias vector of the fully connected layer. This is the classifier model output by the fully connected layer. This is the weight matrix of the classifier model. This is the bias vector of the classifier model.

[0136] To provide a detailed explanation of the technical solution of this invention, specific examples are provided below:

[0137] Example 1, Dataset Creation:

[0138] like Figure 4 As shown, the FPGA software in the USRP used in this study was set to a frequency band of 3GHz, a power of 15dB, and a bandwidth of 20MHz. Figure 5 As shown; the OFDM modulation types of the signals transmitted this time include 8PSK, BPSK, 4PAM, 16QAM, 64QAM, 512QAM, 1024QAM, QPSK and 256QAM.

[0139] The transmitted OFDM modulated signal passes through a physical channel simulator, such as... Figure 6 As shown, a propagation environment was simulated using one Ricean fading channel and nine Rayleigh fading channels with the relative power and phase delay distribution of a standard 5G RAx channel. This channel simulates a test channel for a person traveling on a high-speed train in a rural environment. To provide a more realistic simulation, we also considered additional impairments of 1 kHz Doppler shift and 1.65 μs channel delay.

[0140] like Figure 7 As shown, the OFDM signal transmitted through the physical channel simulator is received and stored as an IQ file using the data retention software of the receiver USRP.

[0141] Example 2, Data Processing of the Dataset:

[0142] The IQ files are processed according to methods commonly found in the literature, such as... Figure 2 , 3 The data is processed as shown, and the processed OFDM data file is saved as "OFDM Modulation Signal Dataset 1";

[0143] Then, the IQ file is processed according to the method proposed in this invention, as shown in the flowchart below. Figure 8 The data processing is performed as shown, and the processed OFDM data file is saved as "OFDM Modulation Signal Dataset 2".

[0144] Example 3, Classifier Training and Results:

[0145] The obtained "OFDM Modulated Signal Dataset 1" and "OFDM Modulated Signal Dataset 2" were preprocessed with modulation schemes including BPSK, QPSK, 8PSK, 4PAM, 16QAM, 64QAM, 256QAM, 512QAM, and 1024QAM. For each modulation, examples were generated for data subcarriers with K=256, with signal-to-noise ratio (SNR) values ​​ranging from -10 dB to 20 dB in 2 dB increments. There were 6000 examples for each SNR value in each modulation category. The datasets used in this invention were divided into two groups: 70% of the data served as the training set, and the remaining 30% served as the test set for validation. The test set contained 750 examples for each modulation category for each SNR value. To improve the representation of the datasets, an oversampling factor of ρ=4 was applied, and a CP of Ncp=64 samples was used.

[0146] The training sets of the two datasets, "OFDM Modulated Signal Dataset 1" and "OFDM Modulated Signal Dataset 2", are fed into a database using the following method: Figure 9Two classifier models, "Classifier Model 1" and "Classifier Model 2," were trained using the constructed AWM model. The "OFDM Modulated Signal Dataset 1" was then used for training with CVCNN, MCLDNN, Transformer, and MCNet models. During training, data samples with a signal-to-noise ratio (SNR) ranging from -10 dB to 20 dB were used for 100 epochs. The optimization algorithm employed was Adam, which leverages the strengths of AdaGrad and RMSProp, making it particularly suitable for handling noisy or sparse gradients. A learning rate of 0.001 was used over 100 iterations. Training was performed in batches of 512, while the model architecture remained independent of the number of samples in the input signal, ensuring a consistent sampling length of 1920 samples. Classification cross-entropy was used as the loss function for classification tasks involving multiple classes due to its computational efficiency and fast convergence. The loss function measures the difference between the predicted probability distribution and the actual probability distribution of the output class. Minimizing classification cross-entropy can improve the model's accuracy during training. All models were built on PyTorch 2.0.1 and trained on an NVIDIA GeForce GTX TeSla V100 using the Adam optimizer.

[0147] The test sets of two datasets, "OFDM Modulated Signal Dataset 1" and "OFDM Modulated Signal Dataset 2", are fed into the trained AWM "Classifier Model 1" and "Classifier Model 2" for testing and detection of the two datasets after data processing. "OFDM Modulated Signal Dataset 1" is then fed into CVCNN, MCLDNN, Transformer, and MCNet to obtain the following results: Figure 10 The recognition rate and Figure 11 , Figure 12 The confusion matrix of AWM with and without data preprocessing;

[0148] Figure 10 In the diagram, the horizontal axis represents the signal-to-noise ratio (SNR), the vertical axis represents the overall recognition rate, the title is modulation recognition rate, the red line represents the recognition effect of the proposed data preprocessing method and adaptive wavelet Mamba, the orange line represents the recognition effect of adaptive wavelet Mamba, and the other colored lines represent the recognition effects of the baseline models CVCNN, MCLDNN, Transformer, and MCNet, respectively.

[0149] Figure 11 , Figure 12In the diagram, rows represent actual labels, and columns represent predicted labels. Diagonal elements with higher numbers indicate the accurate prediction for each modulation scheme. The recognition rate curve shows that the DPAWM of the AWM model after data preprocessing is consistently 2% to 5% higher than the AWM, peaking at around −5 dB with a difference of approximately 5%. Almost all errors in this range originate from higher-order QAMs. The confusion matrix shows that 64, 256, 512, and 1024 QAMs are misclassified as lower-order neighboring QAMs. After preprocessing, the recall rate of 256 QAM increases from 0.77 to 0.79, 512 QAM from 0.57 to 0.66, and 1024 QAM from 0.60 to 0.63, while other modulations remain almost unchanged. This indicates that preprocessing primarily enhances the discriminative power of higher-order QAMs under strong noise, thus raising the overall curve. A 2–4% advantage is still observed in the 0–8 dB range, after which the difference gradually converges. At 10 dB, the difference between the two is approximately 2%, and they essentially overlap at 15 dB and above. The low-order modulation in the confusion matrix has reached its optimal level, and the high-order QAM has also improved to 0.79–0.87, no longer becoming a bottleneck. With sufficient signal-to-noise ratio, each model can fully learn constellation features, and the structural advantages brought by preprocessing are masked by the "saturation accuracy." Compared to other deep learning models, AWM and DPAWM consistently achieve recognition rates higher than 1% to 3%.

[0150] The beneficial effects of this invention are:

[0151] 1. Random sampling is used to extract stable intervals in the IQ file to ensure that the sample data is diverse and retains the complete time-domain frame structure, thereby avoiding the loss of cyclic prefix and symbol interval information in the analog OFDM modulated signal due to fixed-length slices;

[0152] 2. In the sample data cleaning process, a refined outlier detection was added. By calculating the amplitude of each sample data and combining it with the set threshold, it is ensured that the slicing operation can distinguish between the cyclic prefix and the data part, retain the symbol interval and subcarrier orthogonality information, improve the existing fixed-length extraction method, make it more suitable for the inherent frame synchronization characteristics of OFDM signals, effectively remove abnormal and distorted samples caused by equipment errors or simulation model defects, avoid mixing into the training set, and improve data quality and model robustness.

[0153] 3. When reconstructing the data, the original fixed-length data is reorganized into a data form that conforms to the OFDM time-frequency structure. While maintaining the data shape to adapt to the deep learning model, local time-frequency transformation and preservation of symbol boundary information are considered to ensure that the model can capture the intrinsic correlation between subcarriers. The fixed-length sub-blocks are reconstructed into multi-dimensional tensors that meet the input requirements of the deep learning network based on the preset subcarrier number and OFDM symbol number parameters. This more completely preserves the relevant information between subcarriers as well as synchronization and frequency domain characteristics. The reconstructed structured three-dimensional matrix fully preserves the time-frequency characteristics of the OFDM signal and the intrinsic correlation information between subcarriers, improving the deep learning model's ability to capture multi-carrier structures.

[0154] 4. The technical solution disclosed in this invention constructs an end-to-end bridge for cross-time and cross-frequency information interaction, enabling the state-space model to "see the frequency band," thereby achieving efficient and interpretable modulation classification under complex interference channels. This method not only achieves a significant improvement in classification performance but also has obvious advantages in terms of parameter quantity and computational complexity.

[0155] The above description is merely a specific embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any variations or substitutions that can be easily conceived by those skilled in the art within the technical scope disclosed in the present invention should be included within the scope of protection of the present invention. Therefore, the scope of protection of the present invention should be determined by the scope of the claims.

Claims

1. An OFDM high-order modulation signal recognition method based on adaptive data pre-processing and adaptive wavelet Mamba algorithm, characterized in that, The method comprises: Step one, using the FPGA software in the USRP device to transmit an OFDM modulated signal of a preset modulation type; configuring the impulse response of a multipath channel, a normalized carrier frequency deviation, a phase deviation and a Gaussian additive white noise to the received OFDM modulated signal through a physical channel simulator to simulate an actual communication environment, obtaining a simulated OFDM modulated signal, and storing the simulated OFDM modulated signal as an IQ file; Step two, calculating a stable interval of the IQ file by using a steady-state distribution method, collecting sample data in the stable interval, taking odd-position data in the sample data as real parts and even-position data as imaginary parts, normalizing each sample data to calculate an amplitude, and cleaning and retaining target samples whose amplitudes meet a threshold value; Step three, dividing the target samples into sub-blocks according to a sample length and a sub-block interval, stacking all the sub-blocks in a receiving order to reconstruct a structured three-dimensional matrix, storing the structured three-dimensional matrix as a dataset in a key-value pair according to a modulation type and a signal-to-noise ratio, and dividing the dataset into a training set and a test set; Step four, constructing a wavelet Mamba classifier network AWM from sequentially connected convolutional layers, a wavelet transform module, a frequency band perception module, an adaptive gate fusion mechanism module and fully connected layers, inputting the training set into the AWM, extracting features of the training set by the convolutional layers, decomposing the features into high-frequency components and low-frequency components by the wavelet transform module, respectively modeling the high-frequency components and the low-frequency components in a frequency band perception time sequence by the frequency band perception module, adaptively fusing the high-frequency components and the low-frequency components by the adaptive gate fusion mechanism module, classifying the training set by the fully connected layers to train a classifier model, obtaining the trained AWM, and inputting the test set into the trained AWM to test and identify an OFDM high-order modulated signal.

2. The method of claim 1, wherein, In step one, the IQ file is: ； wherein, is an analog OFDM modulated signal in the IQ file, n is the serial number of the sampling point of the analog OFDM modulated signal in time, e is the base number of the natural logarithm, and j is the imaginary unit, is a normalized carrier frequency offset, N is the number of subcarriers in the OFDM symbol, and represents the number of points of an inverse fast Fourier transform (IFFT) used to generate the time domain signal, is a phase offset, is a summation index, is an impulse response of a frequency-selective fading channel with a length of L, is an OFDM modulated signal with a cyclic prefix (CP) transmitted by the FPGA software in the USRP, is a preset modulation type of the OFDM modulated signal, including 8PSK, BPSK, 4PAM, 16QAM, 64QAM, 512QAM, 1024QAM, QPSK, and 256QAM, is a Gaussian additive white noise.

3. The method of claim 2, wherein, In step two, the method for calculating a stable interval of the IQ file by using a steady-state distribution method, collecting sample data in the stable interval, taking odd-position data in the sample data as real parts and even-position data as imaginary parts is: Pre-skip sample point number of IQ file read by steady-state distribution method and post-skip sample point number ; The starting index S and the ending index E determined by the pre-skip sample point number and the post-skip sample point number constitute a stable interval; wherein the calculation method of the starting index S and the ending index E is: ; Collecting sample data in a stable interval , to obtain the real part and the imaginary part in the sample data; wherein the calculation method of the real part and the imaginary part is: ; ; wherein is an imaginary part, is even position data in the sample data; is a real part, is odd position data in the sample data, n is the sequence number of the sampling points in time of the analog OFDM modulated signal, .

4. The method of claim 3, wherein, In step two, the calculation method for normalizing each sample data to calculate an amplitude, cleaning and retaining target samples whose amplitudes meet a threshold value is: [n] = 1 , [n] = 1 ; ； Cleaning meets or sample data comprising NaNs; wherein, [n] is a normalized imaginary part, [n] is a normalized real part, is an amplitude, is a minimum power threshold, is a maximum power threshold, and NaN is not a number.

5. The method of claim 4, wherein, In step three, the structured three-dimensional matrix is: ,[2,320,6])； wherein Y is a structured three-dimensional matrix, is an array manipulation function that changes the shape of an array; is a sub-block stacked in reception order, and [2, 320, 6] is the shape of a structured three-dimensional matrix.

6. The method of claim 5, wherein, In step four, the method for extracting features of the training set by the convolutional layers is: ； In the formula, is the feature extracted by the convolution layer, ConvBlock is a two-dimensional convolution layer convl with a kernel size of (2, 7), a one-dimensional convolution layer conv2, and a one-dimensional convolution layer conv3 with a step size of 2 connected in sequence, is the training set.

7. The method of claim 6, wherein, In step four, the method for decomposing the features into high-frequency components and low-frequency components by the wavelet transform module comprises: S1, split the feature into non-overlapping odd components and even components ; wherein, ; ; is odd position data in the feature, is even position data in the feature; S2, predicting the prediction error d[n] of the neighboring odd component from the even component as a high frequency component; wherein ; where P( ) is a prediction operator to predict the odd component from the even component. S3, updating the even components with the high frequency components to obtain the low frequency components c[n]; wherein, wherein U(·) is an update operator that updates the even components with the high frequency components.

8. The method of claim 7, wherein, In step four, the method for respectively modeling the high-frequency components and the low-frequency components in a frequency band perception time sequence by the frequency band perception module is: The frequency band perception module is a parallel low-frequency frequency band perception Mamba module and a high-frequency frequency band perception Mamba module; The low-frequency frequency band perception Mamba module is used for transposing the low-frequency components, assigning a fixed frequency band ID to each time step, obtaining a low-frequency frequency band embedding vector, adding the low-frequency frequency band embedding vector and the low-frequency components to obtain a low-frequency frequency band vector, adding a bias term to drive the evolution of a hidden state, and obtaining a constructed low-frequency frequency band perception time sequence module; meanwhile, The high-frequency band-aware Mamba module is used for splicing high-frequency components of all scales in a time dimension to form a high-frequency sequence; and a frequency band ID of a corresponding source level is assigned to each time step in the high-frequency sequence to obtain a high-frequency band embedding vector; and the high-frequency band embedding vector is added to the high-frequency components to obtain a high-frequency band vector; and the high-frequency band vector is added to a bias term to drive evolution of a hidden state to obtain a constructed high-frequency band-aware time sequence module.

9. The method of claim 8, wherein, In step four, the adaptive gating fusion method performed by the adaptive gating fusion mechanism module is as follows: After the low-frequency band-aware time sequence module and the high-frequency band-aware time sequence module are compressed in the time dimension respectively, the compressed low-frequency band-aware time sequence module and the high-frequency band-aware time sequence module are spliced along the time dimension to obtain a spliced sequence; an average value of the spliced sequence is calculated to obtain a descriptor; and a gating weight of the descriptor is output; and the gating weight is used to weight-sum the compressed low-frequency band-aware time sequence module and the high-frequency band-aware time sequence module to obtain a fused feature.

10. The method of claim 9, wherein, In step four, the method for training the classifier model through the fully connected layer classification is as follows: ； ； wherein, is a prediction probability output by the fully connected layer, is a regularization function to prevent overfitting of the neural network, is an activation function, is a weight matrix of the fully connected layer, is a fused feature, is a bias vector of the fully connected layer, is a classifier model output by the fully connected layer, is a weight matrix of the classifier model, is a bias vector of the classifier model.

Images