Noise signal identification method based on deep learning

Abstract

The invention discloses a noise signal recognition method based on deep learning. The method comprises the steps of original signal collection, signal preprocessing, noise recognition feature extraction, noise signal recognition model establishment and noise signal intelligent recognition. The invention relates to the technical field of digital signal data processing, in particular to a noise signal identification method based on deep learning, and innovatively provides a self-adaptive signal mode decomposition method based on a combined judgment mechanism of energy threshold value judgment and energy range value judgment. Decoupling and reservation of different frequency bandwidth information in an original signal can be effectively completed, and the recognition accuracy of non-stationary communication noise is improved; a frequency band filter bank constructed based on a Gaussian function is introduced, so that the accuracy of communication signal identification is improved; a method for constructing a noise recognition model by fusing graph neural network structure modeling and an original type reference vector matching mechanism is innovatively proposed, and the recognition robustness of the model to non-stationary communication interference signals is improved.

Description

Technical Field

[0001] This invention relates to the field of digital signal data processing technology, specifically to a noise signal recognition method based on deep learning. Background Technology

[0002] A noise signal recognition method based on deep learning refers to an intelligent signal processing method that utilizes digital signal acquisition and processing technology, combined with deep neural networks and pattern recognition, to efficiently classify and analyze the features of massive, multi-source, and unstructured signal and noise data collected in complex environments. This enables automatic identification and accurate classification of noise types, interference patterns, and background signals. It can be widely applied in multi-interference environments to improve the accuracy and response efficiency of noise recognition.

[0003] However, traditional communication signal recognition methods suffer from technical problems such as rapid frequency drift, complex modulation modes, and severe mode aliasing in communication signals. These problems easily lead to high dimensionality and redundancy in subsequent feature extraction, resulting in a decrease in the accuracy of noise recognition results. In addition, feature extraction in traditional communication signal recognition methods relies solely on a single cepstral structure, which cannot effectively cover global and local spectral change information. This results in a decrease in recognition accuracy when the model is dealing with non-stationary noise such as short-term interference and frequency shift modulation. Furthermore, existing noise recognition models generally cannot adapt to non-fixed type input signals, and their recognition accuracy decreases when type boundaries are blurred and signals overlap. This leads to weak recognition ability for new noise types and low accuracy in recognizing known types. Summary of the Invention

[0004] To address the aforementioned issues and overcome the shortcomings of existing technologies, this invention provides a noise signal recognition method based on deep learning. Addressing the technical problems of traditional communication signal recognition methods, such as rapid frequency drift, complex modulation modes, and severe mode aliasing in communication signals, which easily lead to high dimensionality and redundancy in subsequent feature extraction and consequently decreased accuracy in noise recognition, this invention innovatively proposes an adaptive signal mode decomposition method based on a joint judgment mechanism of energy threshold judgment and energy range judgment. This method allows for dynamic adjustment and optimization of the number of modes and control over the decomposition depth, extracting multi-mode signal components with reconstructibility, low redundancy, and high-frequency interpretability. It effectively decouples and preserves different frequency bandwidth information in the original signal, providing a structurally robust foundational data representation for subsequent feature extraction and graph modeling, ultimately significantly improving the recognition accuracy of non-stationary communication noise. Furthermore, addressing the technical problem that traditional communication signal recognition methods rely solely on a single cepstral structure for feature extraction, failing to effectively cover global and local spectral changes, thus causing decreased recognition accuracy when dealing with short-term interference, frequency shifts, and other non-stationary noise, this invention innovatively introduces a deep learning-based... By constructing a frequency band filter bank using Gaussian functions and jointly extracting spectral morphology features and frequency band dynamic change features, the sensitivity to frequency changes and local energy disturbances is enhanced, improving the stability and information coverage of feature expression. This provides the noise recognition model with input signal feature vectors that are more capable of distinguishing signal types, thereby improving the accuracy of communication signal recognition in complex noise interference scenarios. Addressing the technical problems of existing noise recognition models generally being unable to adapt to non-fixed type input signals and experiencing decreased recognition accuracy in cases of blurred type boundaries and signal overlap, resulting in weak recognition ability for new noise types and low accuracy in recognizing known classes, this invention innovatively proposes a method for constructing a noise recognition model that integrates graph neural network structure modeling with the original type reference vector matching mechanism. This method supports dual modeling based on feature structure and category semantics, and improves the robustness of the model in recognizing non-stationary communication interference signals by constructing a joint loss function for collaborative training and optimization of model parameters. It also possesses intelligent detection capabilities for unknown noise types, while maintaining the accuracy of known class recognition, achieving efficient, robust, and scalable recognition of noise signals in complex communication environments.

[0005] The technical solution adopted by this invention is as follows: This invention provides a noise signal recognition method based on deep learning, which includes the following steps:

[0006] Step S1: Acquisition of raw signals;

[0007] Step S2: Signal preprocessing;

[0008] Step S3: Noise recognition feature extraction;

[0009] Step S4: Establish a noise signal recognition model;

[0010] Step S5: Intelligent identification of noise signals.

[0011] Further, in step S1, the original signal acquisition specifically involves acquiring signals through a wireless signal receiving terminal to obtain original signal identification data; the original signal identification data includes historical signal data and real-time signal data, both of which include timestamps, signal amplitude sequences, sampling rate parameters, and acquisition channel information; the historical signal data also includes signal type.

[0012] Further, in step S2, the signal preprocessing specifically includes the following steps:

[0013] Step S21: Signal data cleaning, specifically, involves effectively filtering the original signal identification data by setting sampling amplitude thresholds, time length ranges, and redundancy detection mechanisms;

[0014] Step S22: Signal standardization, specifically, the frame is divided by a sliding window mechanism, and the signal value of each frame is numerically standardized by a maximum-minimum normalization algorithm to obtain the standardized signal;

[0015] Step S23: Adaptive signal mode decomposition, specifically, using an improved variational mode decomposition algorithm to adaptively decompose the standardized signal to obtain the final set of mode components; including the following steps:

[0016] Step S231: Parameter initialization, specifically setting the range of decomposition layers and the energy ratio threshold. Threshold for judging range ;

[0017] Step S232: Initial modal component extraction, specifically, involves performing multivariate variational mode decomposition on the standardized input signal, including constructing the multi-channel mode decomposition objective, designing the constraint-enhanced optimization function, and iteratively updating the multi-channel modal components to obtain the initial modal components;

[0018] Step S233: Modal energy ratio calculation, specifically, calculating the energy of each initial modal component and determining its proportion in the total signal energy, using the following formula:

[0019] ;

[0020] In the formula, This represents the proportion of the k-th mode component in the total energy, where K represents the number of modes. This represents the time-domain signal of the i-th initial modal component. This represents the time-domain signal of the k-th initial modal component;

[0021] Step S234: Energy threshold judgment, specifically, judging the energy ratio of each modal component under the current decomposition level one by one. If any modal component satisfies If the energy percentage of this mode is low, it will be considered in the energy range judgment. If the modal energy is redundant, the number of modes K is increased, and multivariate variational mode decomposition is continued.

[0022] Step S235: Energy range determination, specifically, sorting the energy proportions of all modal components under the current decomposition level by magnitude, and calculating the energy range between the maximum and minimum values. ,like If the current decomposition level is considered to have reached a relative equilibrium, the modal increment is terminated; otherwise, the number of modes K is increased, and the multivariate variational mode decomposition and subsequent judgment steps are repeated until the termination condition is met.

[0023] Step S236: Output the final modal components. Specifically, after all the energy threshold judgment and energy range judgment conditions are met, the modal components obtained from the current modal decomposition are taken as the final modal components.

[0024] Further, in step S3, the noise identification feature extraction specifically includes the following steps:

[0025] Step S31: Time-frequency power spectrum acquisition, specifically, performing a short-time Fourier transform on the final modal component set to obtain the spectral distribution of each time frame, and calculating the corresponding power spectrum. ;

[0026] Step S32: Obtaining the spectral morphology feature vector. Specifically, this involves constructing a set of band filter banks based on Gaussian functions, applying them to the power spectrum, calculating the energy response value of each filter channel, then performing a logarithmic transform on the filter energy and applying a discrete cosine transform to obtain a set of spectral morphology feature vectors; the formula used is as follows:

[0027] ;

[0028] In the formula, This represents the center frequency of the m-th filter. This represents the standard deviation of the m-th filter. Indicates the m-th filter in The output energy value, where f represents the frequency index. Indicates the time frame index;

[0029] Step S33: Obtain the frequency band dynamic change feature vector. Specifically, the power spectrum is divided into several frequency band intervals, and the logarithmic ratio of the maximum and minimum power values ​​is calculated for each frequency band to obtain the frequency band dynamic change feature vector.

[0030] Step S34: Multi-feature fusion vector acquisition, used to fuse spectral morphology features and frequency band dynamic change features to construct comprehensive noise identification features. Specifically, the spectral morphology feature vector and the frequency band dynamic change feature vector are concatenated to obtain the noise identification feature vector.

[0031] Further, in step S4, establishing the noise signal recognition model specifically includes the following steps:

[0032] Step S41: Frame-level graph modeling and construction, specifically including node definition, edge definition and initialization of the adjacency matrix;

[0033] The node definition specifically involves mapping the noise identification feature vector of each frame to a node in a graph structure;

[0034] The edge definition specifically adopts a full connection strategy, where an edge is established between any two nodes in the graph;

[0035] The initialization of the adjacency matrix specifically involves storing all edge connections as an adjacency matrix and performing symmetric normalization on the adjacency matrix to obtain a normalized adjacency matrix. ;

[0036] Step S42: Graph feature encoding enhancement, specifically, using the node feature matrix and normalized adjacency matrix as input, a graph convolutional neural network is used to calculate the structural feature matrix, and a multi-head attention mechanism is introduced to generate an attention-enhanced feature matrix from the node feature matrix. The structural feature matrix and the attention-enhanced feature matrix are added and fused, and the feature nonlinear mapping is performed through a multilayer perceptron to obtain updated node features. Finally, the structural enhancement noise recognition vector is obtained by extracting through various pooling operations.

[0037] Step S43: Constructing the original type reference vector, specifically setting the original type reference vector set. Where h represents the number of known signal types, and each original type reference vector This represents the original type reference vector corresponding to the j-th signal type, with dimension d and the structure-enhanced noise identification vector. Consistent;

[0038] Step S44: Noise signal identification output, specifically, the structure-enhanced noise identification vector. Reference vectors of each original type The data is then concatenated and fed into the fully connected network scoring function with shared weights. Calculate its matching score with each signal type. Generate a set of matching score vectors And match scores for all signal types with signal type detection thresholds; Perform unknown signal type detection and noise signal identification output, if If so, the input is determined to be an unknown signal type. If so, the corresponding signal type label will be output to obtain the noise signal identification result;

[0039] Step S45: Construct the joint loss function, specifically by comprehensively constructing the graph structure optimization loss term, signal type identification loss term, and signal type differentiation loss term, and then weighting and combining them to form the final joint loss function. The formula used is as follows:

[0040] ;

[0041] ;

[0042] ;

[0043] In the formula, This represents the value of the loss term for graph structure optimization. and This represents the control term weight hyperparameter. This indicates the number of nodes in the graph. Represents the time-location weight matrix. Describing the L1 norm, Describing the Frobenius norm, This represents Hadamard multiplication. This indicates the signal type identification loss term value. This represents the number of training samples, and tn represents the index of a training sample. This indicates the type represented by the original type reference vector. This represents the true class label of the nth input sample. This represents a temperature coefficient used to adjust the smoothness of the softmax output distribution. This indicates the loss term value that distinguishes signal types. This represents the original type reference vector corresponding to the r-th type of signal. Represents the L2 norm;

[0044] Step S46: Construct and train the model. Specifically, this involves constructing the noise signal recognition model through the frame-level graph modeling, graph feature encoding enhancement, original type reference vector construction, and noise signal recognition output. The model is constructed based on the historical signal data as training input data, and the joint loss function is used as the optimization objective. The backpropagation algorithm is used to jointly optimize the parameters in the model. The model is continuously iterated and updated during training until the loss function converges, thus completing the training process of the recognition model and obtaining the trained noise signal recognition model.

[0045] Further, in step S5, the intelligent noise signal identification specifically involves inputting the real-time signal data to be identified into the trained noise signal identification model to obtain the real-time noise signal identification result. Based on the real-time noise signal identification result, noise type judgment and response processing are performed. If the real-time noise signal identification result is a normal signal, no abnormal recording or alarm operation is performed. If the real-time noise signal identification result is one of the known noise types, the corresponding abnormal recording process is triggered, and a warning message is output. If the real-time noise signal identification result is a new type of noise, a high-priority alarm is immediately executed, thereby realizing intelligent identification and graded response processing for different types of noise signals.

[0046] The beneficial effects achieved by the present invention using the above solution are as follows:

[0047] (1) In view of the technical problems of traditional communication signal recognition methods in dealing with communication signals with fast frequency drift, complex modulation mode and severe mode mixing, which easily leads to high dimensionality and large redundancy of subsequent feature extraction, resulting in a decrease in the accuracy of noise recognition results, this invention proposes an adaptive signal mode decomposition method based on the joint judgment mechanism of energy threshold judgment and energy range judgment. It can dynamically adjust the number of modes and optimize the decomposition depth control, extract multi-mode signal components with reconstructibility, low redundancy and high frequency domain interpretability, effectively complete the decoupling and preservation of different frequency bandwidth information in the original signal, provide a strong basic data representation for subsequent feature extraction and graph modeling, and ultimately significantly improve the recognition accuracy of non-stationary communication noise.

[0048] (2) In view of the technical problem that the feature extraction of traditional communication signal recognition methods relies only on a single cepstral structure, which cannot effectively cover global and local spectral change information, thus causing the model to lose recognition accuracy when dealing with non-stationary noise such as short-term interference and frequency shift modulation, this invention innovatively introduces a frequency band filter bank based on Gaussian function to jointly extract spectral morphology features and frequency band dynamic change features, which enhances the sensitivity to frequency changes and local energy disturbances, improves the stability and information coverage of feature expression, and provides the noise recognition model with input signal feature vectors that are more capable of distinguishing signal types, thereby improving the accuracy of communication signal recognition in complex noise interference scenarios.

[0049] (3) In view of the technical problems that existing noise recognition models generally cannot adapt to non-fixed type input signals and the recognition accuracy decreases when the type boundary is blurred and the signal overlaps, resulting in weak recognition ability for new noise types and low recognition accuracy of known classes, this invention innovatively proposes a method to construct a noise recognition model by integrating graph neural network structure modeling and the original type reference vector matching mechanism. It supports dual modeling based on feature structure and category semantics, and improves the robustness of the model in recognizing non-stationary communication interference signals by constructing a joint loss function. It also has the ability to intelligently detect unknown noise types, while taking into account the recognition accuracy of known categories, and realizes efficient, robust and scalable recognition of noise signals in complex communication environments. Attached Figure Description

[0050] Figure 1 A flowchart illustrating a noise signal recognition method based on deep learning provided by this invention;

[0051] Figure 2 This is a flowchart illustrating step S2;

[0052] Figure 3 This is a flowchart illustrating step S23;

[0053] Figure 4 This is a flowchart illustrating step S3;

[0054] Figure 5 This is a flowchart illustrating step S4;

[0055] The accompanying drawings are provided to further illustrate the invention and form part of the specification. They are used together with the embodiments of the invention to explain the invention and do not constitute a limitation thereof. Detailed Implementation

[0056] 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. All other embodiments obtained by those skilled in the art based on the embodiments of the present invention without creative effort are within the scope of protection of the present invention.

[0057] In the description of this invention, it should be understood that the terms "upper", "lower", "front", "rear", "left", "right", "top", "bottom", "inner", "outer", etc., indicate the orientation or positional relationship based on the orientation or positional relationship shown in the accompanying drawings. They are only for the convenience of describing this invention and simplifying the description, and do not indicate or imply that the system or element referred to must have a specific orientation, or be constructed and operated in a specific orientation. Therefore, they should not be construed as limitations on this invention.

[0058] Example 1, see Figure 1 The technical solution adopted by this invention is as follows: This invention provides a noise signal recognition method based on deep learning, which includes the following steps:

[0059] Step S1: Raw signal acquisition, specifically, signal acquisition is performed through a wireless signal receiving terminal to obtain raw signal identification data;

[0060] Step S2: Signal preprocessing, used to improve signal data quality and enhance the separability of key signal features. Specifically, it involves cleaning and standardizing the signal data, and using an improved variational mode decomposition method to perform adaptive mode decomposition on the standardized signal to obtain the final set of mode components.

[0061] Step S3: Noise identification feature extraction, used to extract structural features with strong characterization and discriminative power from the final modal component set. Specifically, it involves performing short-time Fourier transform to obtain the time-frequency power spectrum, introducing a frequency band filter bank based on Gaussian function, and combining logarithmic transform and discrete cosine transform to generate a spectral morphology feature vector. At the same time, it extracts the frequency band dynamic change feature vector, and finally concatenates the multi-dimensional features to generate a noise identification feature vector.

[0062] Step S4: Establish a noise signal recognition model to build a deep recognition model with the ability to distinguish multiple types of signals. Specifically, the feature vector sequence is mapped to a frame-level graph structure, and a graph convolutional network and attention mechanism are introduced to obtain structure enhancement features. The original type reference vector is combined to perform type matching and discrimination. A joint loss function consisting of graph structure optimization loss, type recognition loss and type discrimination loss is constructed. Finally, the trained noise signal recognition model is obtained through joint training.

[0063] Step S5: Intelligent noise signal recognition, used to realize the intelligent recognition of the trained model in actual communication scenarios. Specifically, real-time signal data is input into the trained recognition model, the real-time noise signal recognition result is output, and hierarchical response processing is performed according to the recognition result to achieve effective discrimination and processing of normal signals, known interference and new noise types.

[0064] Example 2, see Figure 1This embodiment is based on the above embodiment. In step S1, the original signal acquisition specifically involves acquiring signals through a wireless signal receiving terminal to obtain original signal identification data. The original signal identification data includes historical signal data and real-time signal data. Both the historical signal data and the real-time signal data include timestamps, signal amplitude sequences, sampling rate parameters, and acquisition channel information. The historical signal data also includes signal type.

[0065] The signal types include normal signals, thermal noise signals, shot noise signals, electromagnetic interference signals, mechanical noise signals, natural noise signals, and background noise.

[0066] Example 3, see Figure 1 , Figure 2 and Figure 3 This embodiment is based on the above embodiment. In step S2, the signal preprocessing specifically includes the following steps:

[0067] Step S21: Signal data cleaning, used to remove abnormal data, invalid data and frame loss data in the original signal recognition data. Specifically, it involves filtering the original signal recognition data for effective signals by setting sampling amplitude threshold, time length range and redundancy detection mechanism.

[0068] Step S22: Signal standardization, which is used to divide the continuous signal into analysis units of fixed length and unify the signal amplitude range. Specifically, it involves dividing the frame using a sliding window mechanism and using a maximum-minimum normalization algorithm to standardize the signal value of each frame to obtain the standardized signal.

[0069] Step S23: Adaptive signal mode decomposition, used to decouple the mixed components in the original signal according to different frequency bandwidths. Specifically, it employs an improved variational mode decomposition algorithm to adaptively decompose the standardized signal, obtaining the final set of mode components; including the following steps:

[0070] Step S231: Parameter initialization, specifically setting the range of decomposition layers and the energy ratio threshold. Threshold for judging extreme value difference ;

[0071] The energy ratio threshold Used to filter out invalid modes with too low energy;

[0072] The extreme value difference judgment threshold Used to evaluate whether the distribution of each mode is balanced;

[0073] Step S232: Initial modal component extraction, used to obtain a set of initial modal components with energy concentration and decoupling interpretability; specifically, by performing multivariate variational mode decomposition on the standardized input signal, including multi-channel mode decomposition objective construction, constraint-enhanced optimization function design, and iterative update of multi-channel modal components, the initial modal components are obtained; including the following steps:

[0074] Step S2321: Multi-channel mode decomposition objective construction, used to establish a mathematical optimization objective for mode decomposition in multi-channel communication noise signals. Specifically, based on the input multi-channel original signal, a function model is constructed with minimizing the sum of the bandwidths of each mode component as the optimization index, and signal reconstruction constraints are set so that the superposition of each mode component can recover the overall waveform of the original signal; the formula used is as follows:

[0075] ;

[0076] In the formula, This represents the time-domain signal value of the k-th mode component of the r-th channel in the multi-channel signal at time t. This represents the k-th modal component of the r-th channel. express The analytical form is obtained by the Hilbert transform. This represents the center frequency of the k-th modal component. This represents the original input signal of the r-th channel. This represents the number of channels, and K represents the number of modes. This represents the set of all channels for the k-th modal component, where j represents the imaginary unit, and is used to construct the analytic signal. This represents the first derivative operation with respect to time t;

[0077] Step S2322: Design of a constraint-enhanced optimization function, used to unify the optimality of bandwidth compression and the accuracy of signal reconstruction into the same optimization system based on the multi-channel mode decomposition objective; specifically, by introducing an augmented Lagrangian function, combined with a preset penalty factor and multi-channel Lagrangian multipliers, the constraint condition for summing modal components is strengthened into an executable optimization term; the formula used is as follows:

[0078] ;

[0079] ;

[0080] In the formula, The modal bandwidth compression term is represented by L, which represents the augmented Lagrangian function. This represents the bandwidth control penalty factor, used to balance the convergence speed of the modal component bandwidth. This represents the Lagrange multiplier corresponding to the r-th channel;

[0081] Step S2323: Iterative update of multi-channel modal components. Specifically, based on the constructed modal optimization objective, the original high-dimensional optimization problem is divided into multiple sub-optimization problems that can be solved in parallel using the alternating direction multiplier method. In multiple alternating iterations, each modal component and its frequency center parameter are gradually converged, ultimately outputting a set of initial modal components that satisfy frequency domain separability and time domain reconfigurability. The formula used is as follows:

[0082] ;

[0083] In the formula, This represents the frequency domain of the k-th mode component in the r-th channel during the (n+1)-th iteration. Represents the original signal Fourier transform, Represents frequency variables. Represent the frequency domain form of the r-th channel Lagrange multiplier. This represents the frequency domain of the i-th mode component in the r-th channel during the n-th iteration. This represents the frequency domain of the i-th mode component in the r-th channel during the (n+1)-th iteration;

[0084] Step S233: Modal energy ratio calculation, used to evaluate the contribution of each initial modal component to the overall energy of the original signal at the current decomposition level; specifically, energy calculation is performed on each initial modal component, and its proportion in the total signal energy is determined accordingly. The formula used is as follows:

[0085] ;

[0086] In the formula, This represents the proportion of the k-th mode component in the total energy, where K represents the number of modes. This represents the time-domain signal of the i-th initial modal component. This represents the time-domain signal of the k-th initial modal component;

[0087] Step S234: Energy threshold judgment, used to determine whether there are redundant modes in the current decomposition level based on the modal energy ratio; specifically, it judges the energy ratio of each modal component in the current decomposition level one by one, and if any modal component satisfies If the energy percentage of this mode is low, it will be considered in the energy range judgment. If the modal energy is redundant, the number of modes K is increased, and multivariate variational mode decomposition is continued.

[0088] Step S235: Energy range judgment, used to assess the balance of the current modal energy distribution and determine whether the decomposition is excessively dispersed; specifically, sorting the energy proportions of all modal components under the current decomposition level by magnitude and calculating the energy range between their maximum and minimum values. ,like If the current decomposition level is considered to have reached a relative equilibrium, the modal increment is terminated; otherwise, the number of modes K is increased, and the multivariate variational mode decomposition and subsequent judgment steps are repeated until the termination condition is met.

[0089] Step S236: Output the final modal components. Specifically, after all the energy threshold judgment and energy range judgment conditions are met, the modal components obtained from the current modal decomposition are taken as the final modal components.

[0090] By performing the above operations, this invention addresses the technical problems of traditional communication signal recognition methods in handling communication signals with rapid frequency drift, complex modulation modes, and severe mode aliasing, which easily leads to high dimensionality and redundancy in subsequent feature extraction, resulting in decreased accuracy of noise recognition results. It proposes an adaptive signal mode decomposition method based on a joint judgment mechanism of energy threshold judgment and energy range judgment. This method can dynamically adjust the number of modes and optimize the decomposition depth, extracting multimodal signal components with reconstructibility, low redundancy, and high-frequency domain interpretability. It effectively decouples and preserves different frequency bandwidth information in the original signal, providing a structurally strong basic data representation for subsequent feature extraction and graph modeling, ultimately significantly improving the recognition accuracy of non-stationary communication noise.

[0091] Example 4, see Figure 1 and Figure 4 This embodiment is based on the above embodiment. In step S3, the noise recognition feature extraction is used to extract feature vectors with discriminative ability from the final modal component set; specifically, it includes the following steps:

[0092] Step S31: Time-frequency power spectrum acquisition, used to convert the time-domain signal of the final modal components into a time-frequency domain representation. Specifically, it involves performing a short-time Fourier transform on the final modal component set to obtain the spectral distribution of each time frame and calculating the corresponding power spectrum; the formula used is as follows:

[0093] ;

[0094] In the formula, Represents the final set of modal components. This indicates the short-time Fourier transform operation. This represents the power spectrum, and f represents the frequency index. Indicates the time frame index;

[0095] Step S32: Obtaining spectral morphology feature vectors, used to extract low-dimensional cepstral features representing the overall distribution trend of the signal spectrum from the power spectrum. Specifically, a set of band filter banks based on Gaussian functions is constructed and applied to the power spectrum. The energy response value of each filter channel is calculated. Subsequently, a logarithmic transformation is performed on the filter energy, and a discrete cosine transform is applied to obtain a set of spectral morphology feature vectors. The formula used is as follows:

[0096] ;

[0097] ;

[0098] In the formula, This represents the center frequency of the m-th filter. This represents the standard deviation of the m-th filter. Indicates the m-th filter in The output energy value, Denotes the nth discrete cosine transform coefficient. Indicates the first The nth spectral morphological feature vector value in the frame, Indicates the first filter in The output energy value, Indicates that the Mth filter is in The output energy value;

[0099] Step S33: Obtaining the dynamic change feature vector of the frequency band, used to capture the energy fluctuations and dynamic change characteristics of local frequency bands in the signal spectrum. Specifically, the power spectrum is divided into several frequency band intervals, and the logarithmic ratio of the maximum and minimum power values ​​is calculated for each frequency band to obtain the dynamic change feature vector of the frequency band; the formula used is as follows:

[0100] ;

[0101] In the formula, This indicates that the b-th frequency band is in the... The characteristic vector of dynamic frequency band change of the frame. This indicates that the b-th frequency band is in the... The power spectrum segment corresponding to the frame This indicates a small positive value that should be avoided when dividing by zero;

[0102] Step S34: Multi-feature fusion vector acquisition, used to fuse spectral morphology features and frequency band dynamic change features to construct comprehensive noise identification features. Specifically, the spectral morphology feature vector and the frequency band dynamic change feature vector are concatenated to obtain the noise identification feature vector.

[0103] By performing the above operations, this invention addresses the technical problem in traditional communication signal recognition methods where feature extraction relies solely on a single cepstral structure, failing to effectively cover global and local spectral variation information. This leads to a decrease in recognition accuracy when dealing with non-stationary noise such as short-term interference and frequency shifts. The invention innovatively introduces a band filter bank based on a Gaussian function to jointly extract spectral morphology features and band dynamic variation features. This enhances the sensitivity to frequency changes and local energy disturbances, improves the stability and information coverage of feature representation, and provides noise recognition models with input signal feature vectors that are more capable of distinguishing signal types. This ultimately improves the accuracy of communication signal recognition in complex noise interference scenarios.

[0104] Example 5, see Figure 1 and Figure 5 This embodiment is based on the above embodiment. In step S4, the establishment of the noise signal recognition model specifically includes the following steps:

[0105] Step S41: Frame-level graph modeling and construction, used to capture the temporal evolution relationship and modal coupling characteristics between frames, specifically including node definition, edge definition and initialization of adjacency matrix;

[0106] The node definition is used to construct the node set in the graph, specifically by mapping the noise recognition feature vector of each frame to a node in the graph structure.

[0107] The edge definition is used to construct the set of edges between nodes in the graph. Specifically, a full connection strategy is adopted, in which an edge is established between any two nodes in the graph.

[0108] The initialization of the adjacency matrix specifically involves storing all edge connections as an adjacency matrix, initializing it as a fully connected matrix, and then performing symmetric normalization on the adjacency matrix to obtain a normalized adjacency matrix. ;

[0109] Step S42: Graph Feature Encoding Enhancement. This step performs deep feature encoding on the constructed frame-level graph structure to obtain a graph embedding representation that integrates spatial and temporal structures. Specifically, it uses a graph convolutional neural network to calculate the structural feature matrix using the node feature matrix and the normalized adjacency matrix as input. A multi-head attention mechanism is then introduced to generate an attention-enhanced feature matrix from the node feature matrix. The structural feature matrix and the attention-enhanced feature matrix are added and fused. A multilayer perceptron is used for nonlinear feature mapping to obtain updated node features. Finally, various pooling operations are used to extract the structure-enhanced noise recognition vector. The formulas used are as follows:

[0110] ;

[0111] ;

[0112] ;

[0113] ;

[0114] In the formula, Represents the normalized adjacency matrix. This represents the feature matrix of the k-th layer nodes. This represents the trainable weight matrix of the graph convolution kernel in the k-th layer. Represents the structural feature matrix. Represents the attention feature matrix, This indicates the operation of a multilayer sensor. This represents the feature matrix of the nodes in the (k+1)th layer. This indicates a max pooling operation. This indicates the average pooling operation. This indicates an accumulation pooling operation. This represents the feature fusion mapping function. This represents the structure-enhanced noise identification vector. This represents the feature matrix of the last layer of nodes. This indicates a bullish attention-based operation. Indicates the feature fusion mapping parameters;

[0115] Step S43: Constructing the original type reference vector, used to establish a discriminative feature reference center for each known communication noise category, in order to achieve type differentiation and identification of new noise types. Specifically, this involves setting up a set of original type reference vectors. Where h represents the number of known signal types, and each original type reference vector This represents the original type reference vector corresponding to the j-th signal type, with dimension d and the structure-enhanced noise identification vector. Consistent with this, the original type reference vectors are all trainable parameters, supporting continuous adaptive updates during model training to gradually learn and fit the feature distribution center of each type of communication noise;

[0116] Step S44: Noise signal identification output, used to determine the category of communication noise signal and detect unknown categories, specifically by converting the structure-enhanced noise identification vector... Reference vectors of each original type The data is then concatenated and fed into the fully connected network scoring function with shared weights. Calculate its matching score with each signal type. Generate a set of matching score vectors And match scores for all signal types with signal type detection thresholds; Perform unknown signal type detection and noise signal identification output, if If so, the input is determined to be an unknown signal type. If so, the corresponding signal type label will be output to obtain the noise signal identification result;

[0117] Step S45: Construct a joint loss function to jointly optimize the graph connectivity structure, signal type discrimination ability, and signal type distinguishability in the noise recognition model, thereby improving the robustness of the noise signal recognition model to non-stationary, multi-modal interference signals and its ability to identify novel noise types. Specifically, this involves comprehensively constructing a graph structure optimization loss term, a signal type recognition loss term, and a signal type distinguishability loss term, and then weighting and combining them to form the final joint loss function. The formula used is as follows:

[0118] ;

[0119] ;

[0120] ;

[0121] ;

[0122] In the formula, This represents the value of the loss term for graph structure optimization. and This represents the control term weight hyperparameter. This indicates the number of nodes in the graph. Represents the time-location weight matrix. Describing the L1 norm, Describing the Frobenius norm, This represents Hadamard multiplication. This indicates the signal type identification loss term value. This represents the number of training samples, and tn represents the index of a training sample. This indicates the type represented by the original type reference vector. This represents the true class label of the nth input sample. This represents a temperature coefficient used to adjust the smoothness of the softmax output distribution. This indicates the loss term value that distinguishes signal types. This represents the original type reference vector corresponding to the r-th type of signal. , and These represent the weighting coefficients for the contributions of each sub-loss term. Represents the L2 norm;

[0123] Step S46: Construct and train the model. Specifically, this involves constructing the noise signal recognition model through the frame-level graph modeling, graph feature encoding enhancement, original type reference vector construction, and noise signal recognition output. The model is constructed based on the historical signal data as training input data, and the joint loss function is used as the optimization objective. The backpropagation algorithm is used to jointly optimize the parameters in the model. The model is continuously iterated and updated during training until the loss function converges, thus completing the training process of the recognition model and obtaining the trained noise signal recognition model.

[0124] By performing the above operations, this invention addresses the technical problems of existing noise recognition models being generally unable to adapt to non-fixed type input signals and experiencing decreased recognition accuracy under conditions of blurred type boundaries and signal overlap, resulting in weak recognition ability for new noise types and low accuracy in recognizing known classes. This invention innovatively proposes a method for constructing a noise recognition model that integrates graph neural network structure modeling with the original type reference vector matching mechanism. It supports dual modeling based on feature structure and category semantics, and improves the model's robustness in recognizing non-stationary communication interference signals by constructing a joint loss function. It also possesses intelligent detection capabilities for unknown noise types, while maintaining the accuracy of known class recognition, thus achieving efficient, robust, and scalable recognition of noise signals in complex communication environments.

[0125] Example 6, see Figure 1 This embodiment is based on the above embodiment. In step S5, the intelligent noise signal identification specifically involves inputting the real-time signal data to be identified into the trained noise signal identification model to obtain the real-time noise signal identification result. Based on the real-time noise signal identification result, noise type judgment and response processing are performed. If the real-time noise signal identification result is a normal signal, no abnormal recording or alarm operation is performed. If the real-time noise signal identification result is one of the known noise types, the corresponding abnormal recording process is triggered, and a warning message is output. If the real-time noise signal identification result is a new type of noise, a high-priority alarm is immediately executed, and the waveform data and feature vector information of the signal are recorded and archived for subsequent model expansion training and knowledge base updates, thereby realizing intelligent identification and hierarchical response processing for different types of noise signals.

[0126] It should be noted that, in this document, relational terms such as "first" and "second" are used only to distinguish one entity or operation from another, and do not necessarily require or imply any such actual relationship or order between these entities or operations. Furthermore, the terms "comprising," "including," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or apparatus that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such process, method, article, or apparatus.

[0127] Although embodiments of the invention have been shown and described, it will be understood by those skilled in the art that various changes, modifications, substitutions and alterations can be made to these embodiments without departing from the principles and spirit of the invention.

[0128] The present invention and its embodiments have been described above. This description is not restrictive, and the accompanying drawings are only one embodiment of the present invention; the actual structure is not limited thereto. In conclusion, if those skilled in the art are inspired by this description and design similar structures and embodiments without departing from the spirit of the invention, such designs should fall within the protection scope of the present invention.

Claims

1. A noise signal recognition method based on deep learning, characterized in that: The method includes the following steps: Step S1: Raw signal acquisition. Through signal acquisition operations, the raw data for signal recognition is obtained. Step S2: Signal preprocessing, including signal data cleaning and standardization, and using a joint judgment mechanism based on energy threshold judgment and energy range judgment to improve the variational mode decomposition algorithm, thereby achieving adaptive multimodal decomposition of the standardized signal and obtaining the final set of mode components. Step S3: Noise identification feature extraction. First, a short-time Fourier transform is performed on the final modal component set to obtain its time-frequency power spectrum. Then, a frequency band filter bank based on the Gaussian function is constructed to weight the time-frequency power spectrum and calculate the energy response value. Then, the spectral morphology feature vector is generated by combining logarithmic transform and discrete cosine transform. At the same time, the frequency band dynamic change feature vector is extracted. Finally, the two types of feature vectors are concatenated to generate the noise identification feature vector. Step S4: Establish a noise signal recognition model to build a deep recognition model with the ability to distinguish multiple types of signals. By mapping the noise recognition feature vector sequence to a frame-level graph structure, a graph convolutional network combined with a multi-head attention mechanism is used to encode the graph structure with deep features, extract structure-enhanced noise recognition vectors, and introduce the original type reference vectors for type matching. A joint loss function is constructed based on graph structure optimization loss, signal type recognition loss, and signal type discrimination loss. The recognition model is then jointly trained to obtain the trained noise signal recognition model. Step S5: Intelligent noise signal identification, specifically, inputting real-time signal data into the trained noise signal identification model, outputting real-time noise signal identification results, and performing hierarchical response processing based on the identification results to achieve effective discrimination and processing of normal signals, known interference, and new noise types.

2. The noise signal recognition method based on deep learning according to claim 1, characterized in that: In step S2, the signal preprocessing specifically includes the following steps: Step S21: Signal data cleaning, specifically, involves effectively filtering the original signal identification data by setting sampling amplitude thresholds, time length ranges, and redundancy detection mechanisms; Step S22: Signal standardization, specifically, the frame is divided by a sliding window mechanism, and the signal value of each frame is numerically standardized by a maximum-minimum normalization algorithm to obtain the standardized signal; Step S23: Adaptive signal mode decomposition.

3. The noise signal recognition method based on deep learning according to claim 1, characterized in that: In step S23, the adaptive signal mode decomposition specifically includes the following steps: Step S231: Parameter initialization, specifically setting the range of decomposition layers and the energy ratio threshold. Threshold for judging range ; Step S232: Initial modal component extraction, specifically, involves performing multivariate variational mode decomposition on the standardized input signal, including constructing the multi-channel mode decomposition objective, designing the constraint-enhanced optimization function, and iteratively updating the multi-channel modal components to obtain the initial modal components; Step S233: Modal energy ratio calculation, specifically, calculating the energy of each initial modal component and determining its proportion in the total signal energy, using the following formula: ； In the formula, This represents the proportion of the k-th mode component in the total energy, where K represents the number of modes. This represents the time-domain signal of the i-th initial modal component. This represents the time-domain signal of the k-th initial modal component; Step S234: Energy threshold judgment, specifically, judging the energy ratio of each modal component under the current decomposition level one by one. If any modal component satisfies If the energy percentage of this mode is low, it will be considered in the energy range judgment. If the modal energy is redundant, the number of modes K is increased, and multivariate variational mode decomposition is continued. Step S235: Energy range determination, specifically, sorting the energy proportions of all modal components under the current decomposition level by magnitude, and calculating the energy range between the maximum and minimum values. ,like If the current decomposition level is considered to have reached a relative equilibrium, the modal increment is terminated; otherwise, the number of modes K is increased, and the multivariate variational mode decomposition and subsequent judgment steps are repeated until the termination condition is met. Step S236: Output the final modal components. Specifically, after all the energy threshold judgment and energy range judgment conditions are met, the modal components obtained from the current modal decomposition are taken as the final modal components.

4. The noise signal recognition method based on deep learning according to claim 1, characterized in that: In step S3, the noise identification feature extraction specifically includes the following steps: Step S31: Time-frequency power spectrum acquisition, specifically, performing a short-time Fourier transform on the final modal component set to obtain the spectral distribution of each time frame, and calculating the corresponding power spectrum. ; Step S32: Obtaining the spectral morphology feature vector. Specifically, this involves constructing a set of band filter banks based on Gaussian functions, applying them to the power spectrum, calculating the energy response value of each filter channel, then performing a logarithmic transform on the filter energy and applying a discrete cosine transform to obtain a set of spectral morphology feature vectors; the formula used is as follows: ； In the formula, This represents the center frequency of the m-th filter. This represents the standard deviation of the m-th filter. Indicates the m-th filter in The output energy value, where f represents the frequency index. Indicates the time frame index; Step S33: Obtain the frequency band dynamic change feature vector. Specifically, the power spectrum is divided into several frequency band intervals, and the logarithmic ratio of the maximum and minimum power values ​​is calculated for each frequency band to obtain the frequency band dynamic change feature vector. Step S34: Obtaining multi-feature fusion vector, specifically by concatenating the spectral morphology feature vector with the frequency band dynamic change feature vector to obtain the noise identification feature vector.

5. The noise signal recognition method based on deep learning according to claim 1, characterized in that: In step S4, establishing the noise signal recognition model specifically includes the following steps: Step S41: Frame-level graph modeling and construction, specifically including node definition, edge definition and initialization of the adjacency matrix; The node definition specifically involves mapping the noise identification feature vector of each frame to a node in a graph structure; The edge definition specifically adopts a full connection strategy, where an edge is established between any two nodes in the graph; The initialization of the adjacency matrix specifically involves storing all edge connections as an adjacency matrix and performing symmetric normalization on the adjacency matrix to obtain a normalized adjacency matrix. ; Step S42: Graph feature encoding enhancement, specifically, using the node feature matrix and normalized adjacency matrix as input, a graph convolutional neural network is used to calculate the structural feature matrix, and a multi-head attention mechanism is introduced to generate an attention-enhanced feature matrix from the node feature matrix. The structural feature matrix and the attention-enhanced feature matrix are added and fused, and the feature nonlinear mapping is performed through a multilayer perceptron to obtain updated node features. Finally, the structural enhancement noise recognition vector is obtained by extracting through various pooling operations. Step S43: Constructing the original type reference vector, specifically setting the original type reference vector set. Where h represents the number of known signal types, and each original type reference vector This represents the original type reference vector corresponding to the j-th signal type, with dimension d and the structure-enhanced noise identification vector. Consistent; Step S44: Noise signal identification output, specifically, the structure-enhanced noise identification vector. Reference vectors of each original type The data is then concatenated and fed into the fully connected network scoring function with shared weights. Calculate its matching score with each signal type. Generate a set of matching score vectors And match scores for all signal types with signal type detection thresholds; Perform unknown signal type detection and noise signal identification output, if If so, the input is determined to be an unknown signal type. If so, the corresponding signal type label will be output to obtain the noise signal identification result; Step S45: Construct the joint loss function, specifically by comprehensively constructing the graph structure optimization loss term, signal type identification loss term, and signal type differentiation loss term, and then weighting and combining them to form the final joint loss function. The formula used is as follows: ； ； ； In the formula, This represents the value of the loss term for graph structure optimization. and This represents the control term weight hyperparameter. This indicates the number of nodes in the graph. Represents the time-location weight matrix. Describing the L1 norm, Describing the Frobenius norm, This represents Hadamard multiplication. This indicates the signal type identification loss term value. This represents the number of training samples, and tn represents the index of a training sample. This indicates the type represented by the original type reference vector. This represents the true class label of the nth input sample. This represents a temperature coefficient used to adjust the smoothness of the softmax output distribution. This indicates the loss term value that distinguishes signal types. This represents the original type reference vector corresponding to the r-th type of signal. Represents the L2 norm; Step S46: Construct and train the model. Specifically, this involves constructing the noise signal recognition model through the frame-level graph modeling, graph feature encoding enhancement, original type reference vector construction, and noise signal recognition output. Based on historical signal data as training input data, and using the joint loss function as the optimization objective, the backpropagation algorithm is used to jointly optimize the parameters in the model. During training, iterative updates are continuously performed until the loss function converges, thus completing the training process of the recognition model and obtaining the trained noise signal recognition model.

6. The noise signal recognition method based on deep learning according to claim 1, characterized in that: In step S5, the intelligent noise signal identification specifically involves inputting the real-time signal data to be identified into the trained noise signal identification model to obtain the real-time noise signal identification result. Based on the real-time noise signal identification result, noise type judgment and response processing are performed. If the real-time noise signal identification result is a normal signal, no abnormal recording or alarm operation is performed. If the real-time noise signal identification result is one of the known noise types, the corresponding abnormal recording process is triggered and a warning message is output. If the real-time noise signal identification result is a new type of noise, a high-priority alarm is immediately executed, thereby realizing intelligent identification and graded response processing for different types of noise signals.

7. The noise signal recognition method based on deep learning according to claim 1, characterized in that: The original signal acquisition specifically involves acquiring signals through a wireless signal receiving terminal to obtain original signal identification data. The original signal identification data includes historical signal data and real-time signal data, both of which include timestamps, signal amplitude sequences, sampling rate parameters, and acquisition channel information. The historical signal data also includes signal type.

Images