A deep learning-based method for noise signal recognition

Abstract

This invention discloses a noise signal recognition method based on deep learning. The method includes raw signal acquisition, signal preprocessing, noise recognition feature extraction, establishment of a noise signal recognition model, and intelligent noise signal recognition. This invention relates to the field of digital signal data processing technology, specifically a noise signal recognition method based on deep learning. This solution innovatively proposes an adaptive signal mode decomposition method based on a joint judgment mechanism of energy threshold judgment and energy range judgment, which can effectively decouple and retain information on different frequency bandwidths in the original signal, improving the recognition accuracy of non-stationary communication noise. It introduces a frequency band filter bank based on Gaussian functions, thereby improving the accuracy of communication signal recognition. Furthermore, it innovatively proposes a method that integrates graph neural network structure modeling with the original type reference vector matching mechanism to construct a noise recognition model, improving the model's robustness in recognizing non-stationary communication interference signals.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of digital signal data processing, in particular to a noise signal recognition method based on deep learning. BACKGROUND

[0002] The noise signal recognition method based on deep learning refers to using digital signal acquisition and processing technology, combining deep neural network and pattern recognition, to efficiently classify and analyze the characteristics of massive, multi-source and unstructured signal noise data collected in complex environments, so as to realize the automatic recognition and accurate classification of noise types, interference patterns and background signals. This intelligent signal processing method can be widely applied in multi-interference environments to improve the recognition accuracy and response efficiency of noise.

[0003] However, the traditional communication signal recognition method has the technical problems of fast frequency drift, complex modulation mode and serious mode aliasing of communication signals, which easily leads to high dimension and large redundancy of subsequent feature extraction, and further causes the accuracy of noise recognition results to decrease; the feature extraction in the traditional communication signal recognition method only relies on a single cepstrum structure, which cannot effectively cover the global and local frequency spectrum change information, thus leading to the technical problem of decreased recognition accuracy of the model when processing non-stationary noise such as short-time interference and frequency shift modulation; the existing noise recognition model generally cannot adapt to non-fixed type input signals, and the recognition accuracy decreases in the case of fuzzy type boundary and signal overlap, which leads to the consequences of weak recognition ability for new noise types and low recognition accuracy for known types. SUMMARY

[0004] In view of the above, in order to overcome the defects of the prior art, the present application provides a noise signal recognition method based on deep learning, which is aimed at the technical problems of traditional communication signal recognition methods in processing communication signal frequency drift fast, modulation mode complex and mode aliasing serious, which easily leads to high dimension and redundancy of subsequent feature extraction, and further causes the accuracy of noise recognition result to decrease, the present application innovatively proposes an adaptive signal mode decomposition method based on energy threshold judgment and energy range value judgment combined judgment mechanism, which can dynamically regulate and optimize the decomposition depth control of the mode number, extract multi-modal signal components with reconstruction, low redundancy and high frequency domain interpretability, effectively complete the decoupling and reservation of different frequency bandwidth information in the original signal, provide strong structural basic data representation for subsequent feature extraction and graph modeling, and finally significantly improve the recognition accuracy of non-stationary communication noise, for the technical problems that the feature extraction of the traditional communication signal recognition method only depends on a single cepstrum structure, which cannot effectively cover the global and local spectral change information, thereby leading to the decrease of the recognition accuracy of the model in processing short-time interference, frequency shift modulation and other non-stationary noise, the present application innovatively introduces a frequency band filter bank based on Gaussian function construction, jointly extracts frequency spectrum feature and frequency band dynamic change feature, enhances the sensitivity to frequency change and local energy disturbance, improves the stability and information coverage of feature expression, provides a more signal type distinguishing ability input signal feature vector for the noise recognition model, and realizes the accuracy improvement of communication signal recognition in complex noise interference scene, for the technical problems that the existing noise recognition model generally cannot adapt to non-fixed type input signal, and the recognition accuracy decreases in the type boundary fuzzy and signal overlapping condition, which leads to the consequences of weak recognition ability of new noise type and low recognition accuracy of known class, the present application innovatively proposes a method of constructing a noise recognition model by fusing graph neural network structure modeling and original type reference vector matching mechanism, supports dual modeling based on feature structure and category semantics, and through constructing a joint loss function, the model parameters are collaboratively trained and optimized, the recognition robustness of the model to non-stationary communication interference signal is improved, the intelligent detection ability of unknown noise type is also possessed, the known class recognition accuracy is taken into account, and efficient, robust and scalable recognition of noise signal in complex communication environment is realized.

[0005] The technical scheme adopted by the present application is as follows: the noise signal recognition method based on deep learning provided by the present application comprises the following steps:

[0006] Step S1: original signal acquisition;

[0007] Step S2: signal preprocessing;

[0008] Step S3: noise recognition feature extraction;

[0009] Step S4: establishing 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, the energy proportion of each modal component at the current decomposition level is judged one by one, if any one modal component meets , then the modal energy proportion is low, energy range value judgment is entered, if , then the modal energy is redundant, the modal number K is increased, and the multivariate variational modal decomposition is continued to be executed;

[0022] Step S235: Energy range value judgment, specifically, the energy proportions of all modal components at the current decomposition level are sorted according to size, and the energy range value between the maximum value and the minimum value is calculated , if , it is considered that the current decomposition level has reached a relative balance state, the modal increment is terminated, otherwise the modal number K is continued to be increased, the multivariate variational modal decomposition and the subsequent judgment steps are repeatedly executed until the termination condition is met;

[0023] Step S236: Output the final modal component, specifically, when the energy threshold judgment and the energy range value judgment conditions are met, the modal component obtained by the current modal decomposition is taken as the final modal component.

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

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

[0026] Step S32: Frequency spectrum morphology feature vector acquisition, specifically, a group of frequency band filter banks based on Gaussian function is constructed, and is applied to the power spectrum, the energy response value of each filter channel is calculated, then the filter energy is logarithmically transformed, and the discrete cosine transform is applied, to obtain a group of frequency spectrum morphology feature vectors; the used formula is as follows:

[0027] ;

[0028] In the formula, m represents the center frequency of the mth filter, m represents the standard deviation of the mth filter, m represents the output energy value of the mth filter at , f represents the frequency index, m represents the time frame index;

[0029] Step S33: frequency band dynamic change feature vector acquisition, specifically, dividing the power spectrum into several frequency band intervals, calculating the logarithmic ratio of maximum and minimum power values for each frequency band, and obtaining the frequency band dynamic change feature vector;

[0030] Step S34: multi-feature fusion vector acquisition, used for fusing the frequency spectrum morphology feature and the frequency band dynamic change feature to construct comprehensive noise recognition features, specifically, splicing the frequency spectrum morphology feature vector and the frequency band dynamic change feature vector to obtain a noise recognition feature vector.

[0031] Further, in step S4, the noise signal recognition model is established, specifically including the following steps:

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

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

[0034] The edge definition specifically adopts a full connection strategy to establish edge connections between any two nodes in the graph;

[0035] The initialization of the adjacency matrix specifically stores all edge connections in the adjacency matrix, and performs symmetric normalization processing on the adjacency matrix to obtain a normalized adjacency matrix ;

[0036] Step S42: graph feature coding enhancement, specifically taking the node feature matrix and the normalized adjacency matrix as inputs, using a graph convolutional neural network to calculate a structure feature matrix, introducing a multi-head attention mechanism to generate an attention enhanced feature matrix from the node feature matrix, adding and fusing the structure feature matrix and the attention enhanced feature matrix, performing feature nonlinear mapping through a multilayer perceptron to obtain updated node features, and finally extracting through various pooling operations to obtain a structure enhanced noise recognition vector;

[0037] Step S43: original type reference vector construction, specifically setting an original type reference vector set wherein h represents the number of known signal types, each original type reference vector represents an original type reference vector corresponding to the jth signal type, and the dimension d is consistent with the structure enhanced noise recognition vector ;

[0038] Step S44: noise signal recognition output, specifically splicing the structure enhanced noise recognition vector and each original type reference vector and inputting it into a fully connected network scoring function with shared weights , calculate its matching score with each signal type , generate a set of matching score vectors ; and all signal type matching scores with signal type detection threshold unknown signal type detection and noise signal identification output, if , the input is determined as an unknown signal type, if , the corresponding signal type label is output, and the noise signal identification result is obtained;

[0039] Step S45: construct a joint loss function, specifically, construct a graph structure optimization loss term, a signal type identification loss term and a signal type distinction loss term, and combine them with weights to form a final joint loss function ; the formula used is as follows:

[0040] ;

[0041] ;

[0042] ;

[0043] In the formula, represents the value of the graph structure optimization loss term, and represent the control term weight hyperparameter, represents the number of nodes in the graph, represents the time position weight matrix, represents the L1 norm, represents the Frobenius norm, represents the Hadamard multiplication, represents the value of the signal type identification loss term, represents the number of training samples, and tn represents the index of the training sample, represents the type represented by the original type reference vector, represents the true class label of the tn input sample, represents the temperature coefficient, which is used to adjust the distribution smoothing degree of the softmax output, represents the value of the signal type distinction loss term, represents the original type reference vector corresponding to the rth signal type, represents the L2 norm;

[0044] Step S46: constructing and training the model, specifically, constructing the frame-level graph modeling, the graph feature coding enhancement, the original type reference vector construction and the noise signal identification output, completing the construction of the noise signal identification model, taking the historical signal data as the training input data and taking the joint loss function as the optimization objective, using the back propagation algorithm to jointly optimize the parameters in the model, continuously updating during the training process until the loss function converges, completing the training process of the identification model, and obtaining the trained noise signal identification model.

[0045] Further, in step S5, the noise signal intelligent identification, specifically, inputting the real-time signal data to be identified into the trained noise signal identification model to obtain a noise signal real-time identification result, performing noise type judgment and response processing according to the noise signal real-time identification result, if the noise signal real-time identification result is a normal signal, no abnormal record and alarm operation is performed, if the noise signal real-time identification result is one of the known noise types, the corresponding abnormal record process is triggered and the early warning information is output, if the noise signal real-time identification result is a new noise type, a high-priority alarm is immediately performed, realizing intelligent identification and hierarchical response processing of different categories of noise signals.

[0046] The beneficial effects obtained by the above-mentioned scheme are as follows:

[0047] (1) For the technical problem of traditional communication signal identification method in processing communication signal frequency drift fast, modulation mode complex and mode aliasing serious, which easily leads to high dimension and large redundancy of subsequent feature extraction, and further causes the accuracy of noise identification result to decrease, the present application proposes an adaptive signal mode decomposition method based on energy threshold judgment and energy range value judgment joint judgment mechanism, which can dynamically control the number of modes and optimize the decomposition depth control, extract multi-modal signal components with reconstruction, low redundancy and high frequency domain explainability, effectively complete the decoupling and reservation of different frequency bandwidth information in the original signal, provide structural strong basic data representation for subsequent feature extraction and graph modeling, and finally significantly improve the recognition accuracy of non-stationary communication noise.

[0048] (2) For the technical problem that the feature extraction in the traditional communication signal identification method only depends on a single cepstrum structure, which cannot effectively cover the global and local spectral change information, resulting in the recognition accuracy of the model to decrease when processing short-time interference, frequency shift modulation and other non-stationary noise, the present application innovatively introduces a frequency band filter bank constructed based on a Gaussian function, jointly extracts frequency spectrum feature and frequency band dynamic change feature, enhances the sensitivity to frequency change and local energy disturbance, improves the stability and information coverage of feature expression, provides a more signal type distinguishing ability input signal feature vector for the noise identification model, and realizes the accuracy improvement of communication signal identification in complex noise interference scene.

[0049] (3) In view of the technical problems that existing noise recognition models cannot generally adapt to non-fixed type input signals, and the recognition accuracy is reduced in the case of fuzzy type boundary and signal overlap, resulting in weak recognition ability for new noise types and low recognition accuracy for known classes, the present application innovatively proposes a method of constructing a noise recognition model by fusing a graph neural network structure modeling and an original type reference vector matching mechanism, supporting dual modeling based on feature structure and category semantics, and through the construction of a joint loss function, the model parameters are collaboratively trained and optimized, improving the recognition robustness of the model for non-stationary communication interference signals, and also having intelligent detection ability for unknown noise types, taking into account the recognition accuracy of known classes, realizing efficient, robust and scalable recognition of noise signals in complex communication environments. BRIEF DESCRIPTION OF DRAWINGS

[0050] Figure 1 A flowchart of a noise signal recognition method based on deep learning provided by the present application is shown in the figure;

[0051] Figure 2 A flowchart of step S2 is shown in the figure;

[0052] Figure 3 A flowchart of step S23 is shown in the figure;

[0053] Figure 4 A flowchart of step S3 is shown in the figure;

[0054] Figure 5 A flowchart of step S4 is shown in the figure;

[0055] The accompanying drawings are used to provide a further understanding of the present application, and constitute a part of the specification, together with the embodiments of the present application, to explain the present application, and do not constitute a limitation on the present application. DETAILED DESCRIPTION

[0056] The technical solutions in the embodiments of the present application will be described clearly and completely below with reference to the accompanying drawings in the embodiments of the present application. Obviously, the described embodiments are only a part of the embodiments of the present application, not all the embodiments. Based on the embodiments in the present application, all other embodiments obtained by those skilled in the art without creative labor fall within the scope of protection of the present application.

[0057] In the description of the present application, it should be understood that the terms "upper", "lower", "front", "back", "left", "right", "top", "bottom", "inner", "outer" and the like indicate the orientation or positional relationship based on the orientation or positional relationship shown in the drawings, and are only for the convenience of describing the present application and simplifying the description, and do not indicate or imply that the systems or elements referred to must have a particular orientation, be constructed and operated in a particular orientation, and therefore cannot be understood as a limitation on the present application.

[0058] Embodiment one, refer to Figure 1 The technical scheme adopted by the present application is as follows: the present application provides a noise signal recognition method based on deep learning, which comprises the following steps:

[0059] Step S1: original signal acquisition, specifically signal acquisition by a wireless signal receiving terminal to obtain signal recognition original data;

[0060] Step S2: signal preprocessing, used to improve signal data quality and enhance the separability of key signal features, specifically signal data cleaning and signal standardization are performed, and an improved variational mode decomposition method is used to adaptively decompose the standardized signal to obtain a final modal component set;

[0061] Step S3: noise recognition feature extraction, used to extract strong and discriminative structural features from the final modal component set, specifically short-time Fourier transform is performed to obtain time-frequency power spectrum, a frequency band filter bank based on a Gaussian function is introduced, and a frequency spectrum morphological feature vector is generated by combining a logarithmic transformation and a discrete cosine transformation, while a frequency band dynamic change feature vector is extracted, and finally a noise recognition feature vector is generated by splicing multiple dimensional features;

[0062] Step S4: establishing a noise signal recognition model, used to construct a deep recognition model with multi-type signal discrimination ability, specifically mapping the feature vector sequence into a frame-level graph structure, introducing a graph convolution network and an attention mechanism to obtain structure-enhanced features, combining the original type reference vector for type matching discrimination, and constructing a joint loss function composed of three items of graph structure optimization loss, type recognition loss and type discrimination loss, and finally obtaining the trained noise signal recognition model through joint training;

[0063] Step S5: noise signal intelligent recognition, used to realize intelligent recognition of the trained model in actual communication scenarios, specifically inputting real-time signal data into the trained recognition model, outputting noise signal real-time recognition results, and performing hierarchical response processing according to the recognition results to realize effective discrimination and processing of normal signals, known interference and new noise types.

[0064] Embodiment two, refer to Figure 1The embodiment is based on the above embodiment, and in step S1, the original signal collection is specifically signal collection by a wireless signal receiving terminal to obtain signal identification original data; the signal identification original data includes historical signal data and real-time signal data, and the historical signal data and the real-time signal data both include a timestamp, a signal amplitude sequence, a sampling rate parameter and acquisition channel information; the historical signal data further includes a signal type;

[0065] The signal type includes a normal signal, a thermal noise signal, a shot noise signal, an electromagnetic interference signal, a mechanical noise signal, a natural noise signal and background noise.

[0066] Embodiment three, refer to Figure 1 、 Figure 2 and Figure 3 The embodiment is based on the above embodiment, and in step S2, the signal preprocessing specifically includes the following steps:

[0067] Step S21: signal data cleaning, used for eliminating abnormal data, invalid data and lost frame data in the signal identification original data, and specifically for performing effective signal screening on the signal identification original data by setting a sampling amplitude threshold value, a time length range and a redundancy detection mechanism;

[0068] Step S22: signal standardization, used for dividing continuous signals into fixed-length analysis units and unifying signal amplitude ranges, and specifically for performing frame division by a sliding window mechanism and performing numerical standardization on signal values of each frame by a maximum-minimum value normalization algorithm to obtain a standardized signal;

[0069] Step S23: adaptive signal modal decomposition, used for decoupling mixed components in the original signal according to different frequency bandwidths, and specifically for performing adaptive signal modal decomposition on the standardized signal by an improved variational modal decomposition algorithm to obtain a final modal component set; including the following steps:

[0070] Step S231: parameter initialization, specifically setting a decomposition layer number range, an energy proportion threshold value and an extreme value difference judgment threshold value ;

[0071] The energy proportion threshold value is used for screening out invalid modes with too low energy;

[0072] The extreme value difference judgment threshold value is used for evaluating 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 modal decomposition on the input normalized signal, including multi-channel modal decomposition target construction, constraint enhanced optimization function design and multi-channel modal component iterative update, the initial modal components are obtained; including the following steps:

[0074] Step S2321: multi-channel modal decomposition target construction, used to establish a mathematical optimization target for modal decomposition in a multi-channel communication noise signal; specifically, according to the input multi-channel original signal, a function model is constructed with the optimization index of minimizing the total bandwidth of each modal component, and a signal reconstruction constraint condition is set to make each modal component superimposed to restore the overall waveform of the original signal; the formula used is as follows:

[0075] ;

[0076] In the formula, represents the signal value of the kth modal component of the rth channel in the multi-channel signal at the tth time in the time domain, represents the kth modal component of the rth channel, represents is obtained by Hilbert transform, represents the center frequency of the kth modal component, represents the original input signal of the rth channel, represents the number of channels, and K represents the number of modes, represents the set of the kth modal component on all channels, and j represents the imaginary unit used to construct the analytical signal, represents the first derivative operation with respect to time t;

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

[0078] ;

[0079] ;

[0080] In the formula, represents the modal bandwidth compression term, and L represents the augmented Lagrangian function, represents the bandwidth control penalty factor, used to balance the convergence speed of the modal component bandwidth, denotes the Lagrange multiplier corresponding to the rth channel;

[0081] Step S2323: iterative update of multi-channel modal components, specifically, based on the constructed modal optimization target, the original high-dimensional optimization problem is divided into multiple sub-optimization problems that can be solved in parallel through the alternating direction multiplier method, and the modal components and their frequency center parameters are gradually converged in multiple alternating iterations, and finally a set of initial modal components that satisfy the frequency domain separation and time domain reconstructability are output; the formula used is as follows:

[0082] ;

[0083] In the formula, denotes the frequency domain of the kth modal component in the rth channel in the n+1th iteration, denotes the Fourier transform of the original signal , denotes the frequency variable, denotes the frequency domain form of the Lagrange multiplier of the rth channel, denotes the frequency domain of the ith modal component in the rth channel in the nth iteration, denotes the frequency domain of the ith modal component in the rth channel in the n+1th iteration.

[0084] Step S233: modal energy proportion 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, the energy of each initial modal component is calculated, and the proportion of each initial modal component in the total energy is determined, the formula used is as follows:

[0085] ;

[0086] In the formula, denotes the proportion of the kth modal component in the total energy, and K denotes the number of modes, denotes the time domain signal of the ith initial modal component, denotes the time domain signal of the kth initial modal component.

[0087] Step S234: energy threshold judgment, used to judge whether there is a redundant modal at the current decomposition level according to the modal energy proportion; specifically, the energy proportions of the modal components at the current decomposition level are judged one by one, if any one of the modal components satisfies , the modal energy proportion is low, and the energy difference value judgment is entered, if , the modal energy is redundant, the number of modes K is increased, and the multi-variable variational modal decomposition is continued.

[0088] Step S235: energy range value judgment, used to evaluate the balance of the current modal energy distribution, judge whether the decomposition is over-dispersed; Specifically, sort all modal component energy proportions under the current decomposition level by size, calculate the energy range value between the maximum and minimum values , if , it is considered that the current decomposition level has reached a relative balance state, the modal increment is terminated, otherwise the number of modes K is increased, and the steps of multivariate variational modal decomposition and subsequent judgment are repeated until the termination condition is met.

[0089] Step S236: output the final modal component, specifically, when the energy threshold value judgment and the energy range value judgment conditions are met, the modal component obtained by the current modal decomposition is taken as the final modal component.

[0090] By performing the above operation, in view of the technical problems of fast frequency drift, complex modulation mode and serious modal aliasing in the traditional communication signal recognition method, which easily leads to high dimension and large redundancy of subsequent feature extraction, and further causes the accuracy of noise recognition result to decrease, the application proposes an adaptive signal modal decomposition method based on energy threshold value judgment and energy range value judgment joint judgment mechanism, which can dynamically control the number of modes and optimize the decomposition depth control, extract multi-modal signal components with reconstruction, low redundancy and high frequency domain interpretability, effectively decouple and retain different frequency bandwidth information in the original signal, provide structural strong basic data representation for subsequent feature extraction and graph modeling, and finally significantly improve the recognition accuracy of non-stationary communication noise.

[0091] Embodiment four, refer to 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, the following steps are included:

[0092] Step S31: time-frequency power spectrum acquisition, used to convert the time domain signal of the final modal component into time-frequency domain representation, specifically, perform short-time Fourier transform on the final modal component set to obtain the frequency spectrum distribution of each time frame, and calculate the corresponding power spectrum; The formula used is as follows:

[0093] ;

[0094] In the formula, represents the final modal component set, represents the short-time Fourier transform operation, represents the power spectrum, f represents the frequency index, represents the time frame index.

[0095] Step S32: Spectrum morphology feature vector acquisition, for extracting low-dimensional cepstrum 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, then the filter energy is logarithmically transformed and discrete cosine transform is applied to obtain a set of spectrum morphology feature vectors; the formula used is as follows:

[0096] ;

[0097] ;

[0098] In the formula, m represents the center frequency of the mth filter, m represents the standard deviation of the mth filter, m represents the output energy value of the mth filter at m represents the n th discrete cosine transform coefficient, m represents the n th spectrum morphology feature vector value in the m th frame, m represents the output energy value of the 1st filter at m represents the output energy value of the Mth filter at m represents the output energy value of the 1st filter at m represents the output energy value of the Mth filter at m represents the output energy value of the 1st filter at m represents the output energy value of the Mth filter at

[0099] Step S33: Band dynamic change feature vector acquisition, for capturing the energy fluctuation and dynamic change characteristics of the local band in the signal spectrum, specifically, the power spectrum is divided into several band intervals, the logarithmic ratio of the maximum and minimum power values is calculated for each band to obtain the band dynamic change feature vector; the formula used is as follows:

[0100] ;

[0101] In the formula, m represents the band dynamic change feature vector of the b th band in the m th frame, m represents the power spectrum segment corresponding to the b th band in the m th frame, m represents a small positive value to avoid division by zero; Step S34: Multi-feature fusion vector acquisition, for fusing the spectrum morphology feature and the band dynamic change feature to construct comprehensive noise recognition features, specifically, the spectrum morphology feature vector and the band dynamic change feature vector are spliced to obtain a noise recognition feature vector.

[0102] Step S34: Multi-feature fusion vector acquisition, for fusing the spectrum morphology feature and the band dynamic change feature to construct comprehensive noise recognition features, specifically, the spectrum morphology feature vector and the band dynamic change feature vector are spliced to obtain a noise recognition feature vector.

[0103] ​By performing the above operation, in order to solve the technical problem that in the conventional communication signal recognition method, feature extraction only depends on a single cepstrum structure, which cannot effectively cover the global and local spectral change information, thereby causing the model to have a low recognition accuracy when processing non-stationary noise such as short-time interference, frequency shift and modulation, the present application innovatively introduces a frequency band filter bank based on a Gaussian function, jointly extracts spectral morphological features and frequency band dynamic change features, enhances the sensitivity to frequency changes and local energy disturbances, improves the stability and information coverage of feature expression, provides a noise recognition model with a more signal type distinguishing capability input signal feature vector, and realizes the accuracy improvement of communication signal recognition in a complex noise interference scene.

[0104] Embodiment five, refer to Figure 1 and Figure 5 This embodiment is based on the above-mentioned embodiment, and in step S4, the noise signal recognition model is established, specifically including the following steps:

[0105] Step S41: frame-level graph modeling construction, used for capturing inter-frame time sequence evolution relationship and modal coupling characteristics, 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 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 edge set between nodes in the graph, specifically using a full connection strategy to establish edge connection between any two nodes in the graph;

[0108] The initialization of the adjacency matrix specifically stores all the edge connections into an adjacency matrix, initializes it as a full connection matrix, and performs symmetric normalization processing on the adjacency matrix to obtain a normalized adjacency matrix ;

[0109] Step S42: graph feature coding enhancement, used for deep feature coding of the constructed frame-level graph structure, obtaining a graph embedding representation fused with space and time sequence structure, specifically taking the node feature matrix and the normalized adjacency matrix as input, using a graph convolutional neural network to calculate a structure feature matrix, introducing a multi-head attention mechanism to generate an attention enhanced feature matrix from the node feature matrix, adding and fusing the structure feature matrix and the attention enhanced feature matrix, performing feature nonlinear mapping through a multilayer perceptron to obtain updated node features, and finally extracting through various pooling operations to obtain a structure enhanced noise recognition vector; the used formula is as follows:

[0110] ;

[0111] ;

[0112] ;

[0113] ;

[0114] wherein, denotes a normalized adjacency matrix, denotes a k-th layer node feature matrix, denotes a trainable weight matrix of a k-th layer graph convolution kernel, denotes a structure feature matrix, denotes an attention feature matrix, denotes a multi-layer perception operation, denotes a k+1-th layer node feature matrix, denotes a max-pooling operation, denotes an average-pooling operation, denotes an accumulative-pooling operation, denotes a feature fusion mapping function, denotes a structure-enhanced noise recognition vector, denotes a last layer node feature matrix, denotes a multi-head attention operation, denotes a feature fusion mapping parameter;

[0115] Step S43: original type reference vector construction, for establishing a discriminative feature reference center for each known communication noise category to realize type differentiation and new noise type recognition, specifically setting an original type reference vector set wherein h denotes the number of known signal types, each original type reference vector denotes a j-th signal type corresponding original type reference vector, dimension d is consistent with the structure-enhanced noise recognition vector The original type reference vector is a trainable parameter, which supports continuous adaptive update in the model training process to gradually learn and fit the feature distribution center of each category of communication noise;

[0116] Step S44: noise signal recognition output, for realizing the category determination and unknown category detection of communication noise signals, specifically concatenating the structure-enhanced noise recognition vector with each original type reference vector and inputting into a shared weight fully connected network score function to calculate the matching score of each signal type, generating a matching score vector set ; and performing unknown signal type detection and noise signal recognition output on all signal type matching scores and signal type detection thresholds , if If the input is determined as an unknown signal type, if the corresponding signal type label is outputted, and a noise signal recognition result is obtained.

[0117] Step S45: A joint loss function is constructed for jointly optimizing the graph connection structure in the noise recognition model, the signal type discrimination ability, and the signal type distinguishability, so as to improve the robustness of the noise signal recognition model to non-stationary and multi-modal interference signals and the recognition ability of new noise types. Specifically, a graph structure optimization loss term, a signal type recognition loss term, and a signal type distinguishability loss term are comprehensively constructed, and are combined with weights to form a final joint loss function. The formula used is as follows:

[0118] ;

[0119] ;

[0120] ;

[0121] ;

[0122] In the formula, represents the value of the graph structure optimization loss term, and represents the control term weight hyperparameter, represents the number of nodes in the graph, represents the time position weight matrix, represents the L1 norm, represents the Frobenius norm, represents the Hadamard multiplication, represents the value of the signal type recognition loss term, represents the number of training samples, and tn represents the index of the training sample, represents the type represented by the original type reference vector, represents the true class label of the tn input sample, represents the temperature coefficient for adjusting the distribution smoothness of the softmax output, represents the value of the signal type distinguishability loss term, represents the original type reference vector corresponding to the rth signal type, 、 and respectively represent the weighting coefficients of the contributions of each sub-loss term, represents the L2 norm;

[0123] Step S46: constructing and training the model, specifically, constructing the noise signal recognition model through the frame-level graph modeling, the graph feature coding enhancement, the original type reference vector construction, and the noise signal identification output, based on the historical signal data as the training input data and the joint loss function as the optimization objective, using the back propagation algorithm to jointly optimize the parameters in the model, continuously updating during the training process until the loss function converges, completing the training process of the recognition model, and obtaining the trained noise signal recognition model.

[0124] By performing the above operations, the technical problems that existing noise recognition models cannot generally adapt to non-fixed type input signals and recognition accuracy decreases in the case of type boundary ambiguity and signal overlap, resulting in weak recognition ability for new noise types and low recognition accuracy for known classes, the present application innovatively proposes a method of constructing a noise recognition model by fusing graph neural network structure modeling and original type reference vector matching mechanism, supporting dual modeling based on feature structure and category semantics, and through the construction of a joint loss function for collaborative training and optimization of model parameters, improving the recognition robustness of the model to non-stationary communication interference signals, also having intelligent detection ability for unknown noise types, taking into account the recognition accuracy of known classes, achieving efficient, robust and scalable recognition of noise signals in complex communication environments.

[0125] Embodiment six, see Figure 1 This embodiment is based on the above-mentioned embodiments, in step S5, the noise signal intelligent recognition, specifically, inputting the real-time signal data to be recognized into the trained noise signal recognition model, obtaining the noise signal real-time recognition result, performing noise type judgment and response processing according to the noise signal real-time recognition result, if the noise signal real-time recognition result is normal signal, no abnormal record and alarm operation is performed, if the noise signal real-time recognition result is one of the known noise types, the corresponding abnormal record process is triggered, and the warning information is output, if the noise signal real-time recognition result is a new noise type, 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 updating, realizing intelligent recognition and hierarchical response processing of different categories of noise signals.

[0126] It is to be understood that the terminology used herein is for the purpose of describing particular embodiments only and is not intended to be limiting; it is not intended to exclude myriad other embodiments of the present application that other inventors can develop based on the same general inventive concepts embodied by the described embodiments. That is, although the present application is described in terms of particular embodiments and implementations, it is to be understood that the terminology used is for the purpose of descriptive clarity and that it is intended to be limited only by the claims.

[0127] While the embodiments of the application have been shown and described, it is to be understood that the embodiments described are only by way of example and that modifications, changes, substitutions and variations can be made by those skilled in the art without departing from the spirit and scope of the application.

[0128] The above description of the application and its embodiments is not intended to limit the application, as described by the appended claims, to the embodiments described above. Rather, it is intended to cover all adaptations, modifications and variations of the specific embodiments of the application chosen by the inventors as coming within the scope of the application.

Claims

1. A method for identifying a noise signal based on deep learning, characterized in that: The method comprises the following steps: Step S1: original signal acquisition, through signal acquisition operation, obtain signal recognition original data; Step S2: signal preprocessing, signal data cleaning and signal standardization are performed, and a combined judgment mechanism based on energy threshold judgment and energy range value judgment is adopted to improve the variational mode decomposition algorithm, realize adaptive multi-modal decomposition of the standardized signal, and obtain a final modal component set; Step S3: noise recognition feature extraction, first, 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 Gaussian function is constructed to weight process the time-frequency power spectrum, calculate the energy response value, and combine the logarithmic transformation and discrete cosine transformation to generate a frequency spectrum morphology feature vector, while extracting a frequency band dynamic change feature vector, and finally splicing the two types of feature vectors to generate a noise recognition feature vector; specifically comprising the following steps: Step S31: Time-frequency power spectrum acquisition, specifically, performing short-time Fourier transform on the final modal component set to obtain the frequency spectrum distribution of each time frame, and calculating the corresponding power spectrum ; Step S32: frequency spectrum morphology feature vector acquisition, specifically constructing a group of frequency band filter banks based on Gaussian function, and applying it to the power spectrum to calculate the energy response value of each filter channel, then performing logarithmic transformation on the filter energy and applying discrete cosine transformation to obtain a group of frequency spectrum morphology feature vectors; the formula used is as follows: ； wherein denotes the mth filter center frequency, denotes the mth filter standard deviation, denotes the output energy value of the mth filter at f denotes the frequency index, denotes the time frame index; Step S33: frequency band dynamic change feature vector acquisition, specifically dividing the power spectrum into several frequency band intervals, calculating the logarithmic ratio of the maximum and minimum power values of each frequency band to obtain a frequency band dynamic change feature vector; Step S34: multi-feature fusion vector acquisition, specifically splicing the frequency spectrum morphology feature vector and the frequency band dynamic change feature vector to obtain a noise recognition feature vector; Step S4: establishing a noise signal recognition model, for constructing a deep recognition model with multi-type signal discrimination ability, mapping the noise recognition feature vector sequence into a frame-level graph structure, using a graph convolution network combined with a multi-head attention mechanism to encode the graph structure in depth, extracting a structure-enhanced noise recognition vector, and introducing an original type reference vector for type matching, constructing a joint loss function based on graph structure optimization loss, signal type recognition loss and signal type distinction loss, jointly training the recognition model, and obtaining a trained noise signal recognition model; Step S5: noise signal intelligent recognition, specifically inputting real-time signal data into the trained noise signal recognition model, outputting noise signal real-time recognition results, and performing hierarchical response processing according to the recognition results to realize effective discrimination and processing of normal signals, known interference and new noise types. 2.The noise signal identification method based on deep learning according to claim 1, characterized in that: In step S2, the signal preprocessing specifically comprises the following steps: Step S21: signal data cleaning, specifically performing effective signal screening on the signal recognition original data by setting the sampling amplitude threshold, time length range and redundancy detection mechanism; Step S22: signal standardization, specifically dividing frames by a sliding window mechanism, and using the maximum and minimum value normalization algorithm to normalize the signal value of each frame to obtain the standardized signal; Step S23: adaptive signal modal decomposition. 3.The noise signal identification method based on deep learning according to claim 2, characterized in that: In step S23, the adaptive signal modal decomposition specifically includes the following steps: Step S231: parameter initialization, specifically setting the decomposition layer range, energy ratio threshold and range difference judgment threshold and range difference judgment threshold ; Step S232: initial modal component extraction, specifically by performing multivariate variational modal decomposition on the input normalized signal, including multi-channel modal decomposition target construction, constraint enhanced optimization function design and iterative update of multi-channel modal components, to obtain 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 energy of the signal according to the formula as follows: ； wherein represents the proportion of the kth modal component in the total energy, K represents the number of modes, represents the time-domain signal of the ith initial modal component, represents the time-domain signal of the kth initial modal component; Step S234: Energy threshold judgment, specifically, the energy proportion of each modal component at the current decomposition level is judged one by one, if any one modal component satisfies , the modal energy proportion is low, energy range value judgment is entered, if , the modal energy is redundant, the modal number K is increased, and the multivariate variational modal decomposition is continued to be executed. Step S235: Energy range value judgment, specifically, the energy proportion of all modal components at the current decomposition level is sorted by size, and the energy range value between the maximum value and the minimum value is calculated , if , it is considered that the current decomposition level has reached a relative balance state, the modal increment is terminated, otherwise the number of modes K is increased, and the steps of multivariate variational modal decomposition and subsequent judgment are repeated until the termination condition is met; Step S236: output final modal component, specifically when the energy threshold condition and the energy range condition are both met, the modal component obtained by the current modal decomposition is taken as the final modal component. 4.The noise signal identification method based on deep learning according to claim 1, characterized in that: In step S4, the noise signal recognition model is established, specifically including the following steps: Step S41: frame-level graph modeling construction, specifically including node definition, edge definition and initialization of the adjacency matrix; The node definition specifically maps the noise recognition feature vector of each frame to a node in the graph structure; The edge definition specifically adopts a full connection strategy to establish edge connections between any two nodes in the graph; The initialization of the adjacency matrix specifically refers to storing all edge connection systems as an adjacency matrix, and performing symmetric normalization processing on the adjacency matrix to obtain a normalized adjacency matrix ; Step S42: graph feature coding enhancement, specifically taking the node feature matrix and the normalized adjacency matrix as input, using a graph convolutional neural network to calculate the structure feature matrix, introducing a multi-head attention mechanism to generate an attention enhanced feature matrix from the node feature matrix, adding and fusing the structure feature matrix and the attention enhanced feature matrix, performing feature nonlinear mapping through a multilayer perceptron, obtaining updated node features, and finally extracting through various pooling operations to obtain a structure enhanced noise recognition vector; Step S43: original type reference vector construction, specifically, setting the original type reference vector set Wherein, h is expressed as the known signal type number, each original type reference vector The original type reference vector corresponding to the jth signal type, the dimension d is consistent with the structure enhanced noise recognition vector ​ Step S44: noise signal identification output, specifically, the structural enhancement noise identification vector is spliced with each original type reference vector to enter the shared weight full connection network score function , calculate its matching score with each signal type , generate a matching score vector set ; and all signal type matching scores are compared with the signal type detection threshold unknown signal type detection and noise signal identification output, if , it is determined that the input is an unknown signal type, if , the corresponding signal type label is output, and the noise signal identification result is obtained; Step S45: construct a joint loss function, specifically, comprehensively construct a graph structure optimization loss term, a signal type identification loss term and a signal type distinction loss term, and combine them to form a final joint loss function The formula used is as follows: ； ； ； In the formula, denotes the graph structure optimization loss term value, and denotes the control term weight hyperparameter, denotes the number of nodes in the graph, denotes the time position weight matrix, denotes the L1 norm, denotes the Frobenius norm, denotes the Hadamard multiplication, denotes the signal type identification loss term value, denotes the number of training samples, and tn denotes the index of the training sample, denotes the type represented by the original type reference vector, denotes the true class label of the nth input sample, denotes the temperature coefficient, used to adjust the distribution smoothing degree of the softmax output, denotes the signal type distinction loss term value, denotes the original type reference vector corresponding to the rth signal type, denotes the L2 norm; Step S46: model construction and training, specifically completing the construction of the noise signal recognition model through the frame-level graph modeling construction, the graph feature coding enhancement, the original type reference vector construction and the noise signal recognition output, based on historical signal data as training input data and a joint loss function as the optimization objective, using a backpropagation algorithm to jointly optimize the parameters in the model, continuously updating during the training process until the loss function converges, completing the training process of the recognition model, and obtaining the trained noise signal recognition model.

5. The noise signal identification method based on deep learning according to claim 1, characterized in that: In step S5, the noise signal intelligent recognition, specifically inputting the real-time signal data to be recognized into the trained noise signal recognition model to obtain a noise signal real-time recognition result, performing noise type judgment and response processing according to the noise signal real-time recognition result, if the noise signal real-time recognition result is a normal signal, not performing abnormal record and alarm operation, if the noise signal real-time recognition result is one of the known noise types, triggering the corresponding abnormal record process and outputting a warning information, if the noise signal real-time recognition result is a new noise type, immediately performing high-priority alarm, realizing intelligent recognition and hierarchical response processing of different categories of noise signals. 6.The noise signal identification method based on deep learning according to claim 1, characterized in that: The original signal collection, in particular, through the wireless signal receiving terminal signal acquisition, get signal identification original data; The signal identification original data includes historical signal data and real-time signal data, The historical signal data and real-time signal data all include timestamp, signal amplitude sequence, sampling rate parameter and acquisition channel information;The historical signal data also includes signal type.

Images