Broadband wide frequency spectrum analysis method and system based on deep learning

Abstract

The invention discloses a broadband wide frequency spectrum analysis method and system based on deep learning. The method comprises the steps of spectrum signal acquisition, spectrum preprocessing, abnormal interference detection, spectrum trend prediction and spectrum scheduling management. The invention relates to the technical field of wide-frequency spectrum analysis, in particular to a broadband wide-frequency spectrum analysis method and system based on deep learning, which adopts a three-step method of abnormal interference detection, spectrum trend prediction and spectrum scheduling management, and adopts a layer-by-layer progressive deep learning and reinforcement learning technology. The real-time detection of spectrum abnormity, the accurate prediction of future trend and the optimal allocation of dynamic resources can be realized; carrying out abnormal interference detection by adopting a joint improvement multi-classification support vector machine method combined with feature enhancement; performing spectrum trend prediction by adopting a time sequence integrated network combined with one-dimensional convolution feature extraction; and performing spectrum scheduling management by adopting a reinforcement learning method enhanced by a near-end strategy optimization algorithm.

Description

Technical Field

[0001] The present invention relates to the technical field of wide-bandwidth frequency spectrum analysis, and specifically to a wide-bandwidth frequency spectrum analysis method and system based on deep learning. Background Art

[0002] The wide-bandwidth frequency spectrum analysis method based on deep learning is an innovative technology that combines deep learning technology with spectrum analysis, and is designed to perform efficient and accurate frequency analysis on wide-bandwidth signals. This method automatically extracts features from the signal's spectrum by training a deep neural network model, identifies and classifies different frequency components, and greatly improves the speed and accuracy of spectrum analysis. It can process complex and nonlinear signal features, identify frequency changes or potential interference that are difficult to detect with traditional methods, and is widely used in communications, radar, acoustics, medical diagnosis and other fields. Through the adaptive characteristics of the deep learning model, it can perform real-time spectrum monitoring in a dynamic environment, provide accurate frequency prediction and interference identification, and thus optimize system performance and reliability.

[0003] However, among the existing wide-bandwidth frequency spectrum analysis methods, there are existing intelligent methods that often focus on single signal processing or spectrum analysis, and pay more attention to the static characteristic extraction of the global spectrum. However, in the wide-frequency spectrum analysis, due to the wider range of wide-frequency data and greater data fluctuations, traditional methods cannot accurately locate abnormal interference, lack the ability to understand the long-term trend of the spectrum, and lack intelligent optimization mechanisms in spectrum resource scheduling. Technical problems; among the existing abnormal interference detection methods, there is a relatively common technical method for abnormal interference detection, which is the support vector machine method. However, the traditional support vector machine is a basic model of binary classification, and the classification accuracy, feature extraction depth and classification effect are difficult to meet the complex abnormal interference type detection needs in spectrum analysis, and stacking multiple support vector machines The method is more likely to increase the computational difficulty in the complex integration process, while it is difficult to effectively improve the spectrum analysis performance; in the existing spectrum trend prediction methods, there is a technical problem that the spectrum trend prediction is technically processed by the timing model in the existing technical solution, but the existing two-way timing network has better long-term feature capture capability, but has poor ability to capture the trend changes of spectrum signals at the micro level; in the existing spectrum scheduling management methods, there is a technical problem that the existing spectrum scheduling management methods usually rely on traditional rule-based optimization algorithms (such as greedy algorithms or linear programming). Although these methods perform well in small-scale or simple scenarios, they lack global adaptability and real-time optimization capabilities when faced with dynamic changes in spectrum resource distribution, complex multi-dimensional interference factors and nonlinear resource demands. Summary of the invention

[0004] In view of the above situation, in order to overcome the defects of the prior art, the present invention provides a wide-bandwidth frequency spectrum analysis method and system based on deep learning. This scheme creatively adopts a three-step method of abnormal interference detection-spectrum trend prediction-spectrum scheduling management. Through the progressive deep learning and reinforcement learning technology, it can realize real-time detection of spectrum anomalies, accurate prediction of future trends, and dynamic resource optimization allocation. This method can not only comprehensively capture the complex characteristics of the wide-bandwidth frequency spectrum, but also significantly improve the spectrum utilization efficiency and environmental adaptability, making it more advantageous in complex and changeable scenarios; creatively adopt a joint improved multi-classification support vector machine method combined with feature enhancement to perform abnormal interference detection. Through feature enhancement and multi-dimensional joint improvement, the classifiable categories and overall classification performance of the support vector machine are improved, and the training ideas of weight decomposition and projection parameter optimization are used to improve the degree of performance optimization, providing better data guarantee for abnormal interference detection; creatively adopt a joint improved multi-classification support vector machine method combined with feature enhancement to detect abnormal interference. Through feature enhancement and multi-dimensional joint improvement, the classifiable categories and overall classification performance of the support vector machine are improved, and the training ideas of weight decomposition and projection parameter optimization are used to improve the degree of performance optimization, providing better data guarantee for abnormal interference detection; creatively adopt a three-step method of abnormal interference detection-spectrum trend prediction-spectrum scheduling management. Through the progressive deep learning and reinforcement learning technology, it can realize real-time detection of spectrum anomalies, accurate prediction of future trends, and dynamic resource optimization allocation, which can comprehensively capture the complex characteristics of the wide-bandwidth frequency spectrum, and significantly improve the spectrum utilization efficiency and environmental adaptability, making it more advantageous in complex and changeable scenarios; creatively adopt a joint improved multi-classification support vector machine method combined with feature enhancement to detect abnormal interference. Through feature enhancement and multi-dimensional joint improvement, the classifiable categories and overall classification performance of the support vector machine are improved, and the training ideas of weight decomposition and projection parameter optimization are used to improve the degree of performance optimization, providing better data guarantee for abnormal interference detection; creatively adopt a three-step method of abnormal interference detection-spectrum trend prediction-spectrum scheduling management. A time series integration network combined with one-dimensional convolutional feature extraction is creatively used to predict spectrum trends. By constructing an integration of a bidirectional long short-term memory network and a unidirectional gated recurrent network, the two model structures complement each other through time series integration, which not only optimizes the prediction effect of macro spectrum trend characteristics, but also improves the ability to capture short-term trend changes, and comprehensively improves the effect of spectrum trend prediction; a reinforcement learning method enhanced by a proximal strategy optimization algorithm is creatively adopted. Through the adaptive characteristics of reinforcement learning and the stability of the proximal strategy optimization algorithm, the optimal strategy for spectrum scheduling can be efficiently explored and utilized, and the spectrum environment changes can be dynamically responded to while minimizing the instability in the strategy update process. The clipping mechanism of the proximal strategy optimization algorithm also effectively balances exploration and utilization, thereby achieving efficient and robust spectrum resource allocation. This method can not only improve spectrum utilization, but also better adapt to complex and dynamic application scenarios, and provide exploration and practical experience for the effective application of wide-frequency spectrum analysis.

[0005] The technical solution adopted by the present invention is as follows: The present invention provides a wide-bandwidth frequency spectrum analysis method based on deep learning, the method comprising the following steps:

[0006] Step S1: spectrum signal acquisition;

[0007] Step S2: spectrum preprocessing;

[0008] Step S3: abnormal interference detection;

[0009] Step S4: spectrum trend prediction;

[0010] Step S5: spectrum scheduling management.

[0011] Further, in step S1, the spectrum signal acquisition is used to collect the original spectrum signal data set required for the wide-bandwidth frequency spectrum analysis, specifically, by performing wide-bandwidth frequency spectrum data acquisition to obtain the original data set for the wide-bandwidth frequency spectrum analysis;

[0012] The wide frequency spectrum analysis original data set specifically includes timestamp data, frequency data, signal amplitude data, signal noise data, signal sampling rate data and sampling device information data.

[0013] Furthermore, in step S2, the spectrum preprocessing is used to optimize and enhance the original spectrum signal data, specifically by performing signal denoising, data normalization and data quality optimization operations to obtain spectrum analysis optimization data;

[0014] The spectrum analysis optimization data specifically includes denoising optimized spectrum data, standardized spectrum data, smoothing optimized spectrum data and frequency domain feature data set.

[0015] Further, in step S3, the abnormal interference detection is used to detect abnormal data in the spectrum signal and perform abnormal classification, specifically, based on the spectrum analysis optimization data, a joint improved multi-classification support vector machine method combined with feature enhancement is used to perform abnormal interference detection to obtain abnormal interference detection data, specifically including the following steps:

[0016] Step S31: construct a basic model of a multi-classification support vector machine, specifically by constructing a stacked binary support vector machine objective function, performing multi-classification calculations by calculating the maximum category score of the objective function, and applying a resampling method to optimize the category imbalance problem, thereby constructing a basic model of a multi-classification support vector machine. The calculation formula of the stacked binary support vector machine objective function is:

[0017] ;

[0018] In the formula, F TSVM is the stacked binary support vector machine objective function, w is the support vector weight, b is the support vector bias term, is a slack variable used to represent the degree of misclassification, C is the support vector regularization parameter, x is the original data input, used to represent the spectrum analysis optimization data, i is the category index, and w i is the support vector weight corresponding to the i-th category, n is the total number of data samples, j is the data sample index, is the slack variable corresponding to the jth data sample of the ith category, is the maximum selected category score, which is used to represent the final category output detected after stacking, b i is the support vector bias term corresponding to the i-th category;

[0019] Step S32: linear discriminant feature dimensionality reduction optimization, specifically constructing a linear discriminant analysis operator, amplifying the variance of samples between classes and reducing the variance of samples within classes, and performing feature dimensionality reduction optimization on the frequency domain feature data set in the spectrum analysis optimization data through orthogonal linear discriminant analysis to obtain dimensionality reduction optimized feature data;

[0020] The linear discriminant analysis operator specifically includes total scatter matrix construction, inter-class matrix construction and orthogonal linear discriminant analysis modeling;

[0021] The calculation formula for constructing the total scatter matrix is:

[0022] ;

[0023] In the formula, S t is the total scatter matrix, which is used to represent the overall variance distribution of all samples, X is the original feature matrix, X T is the transpose of the original feature matrix;

[0024] The calculation formula for constructing the inter-class matrix is:

[0025] ;

[0026] In the formula, S b is the inter-class matrix, which is used to represent the distribution of the mean values ​​of each fault interference class. X is the original feature matrix, X T is the transpose of the original feature matrix, Y is the classification information matrix, which is used to provide the category feature information of signal fault interference, Y T is the transpose of the classification information matrix;

[0027] The calculation formula of the orthogonal linear discriminant analysis modeling is:

[0028] ;

[0029] In the formula, F LDA is the objective function of orthogonal linear discriminant analysis, A is the projection matrix, which is used as the optimal linear discriminant parameter, Tr is the matrix trace parameter, which is used to represent the sum of the matrix diagonal matrix, S t is the total divergence matrix, S b is the between-class matrix;

[0030] Step S33: multi-dimensional linear joint improvement, specifically, three-dimensional iterative linear joint improvement, including abnormal fault multi-classification decision function improvement, iterative objective function improvement and linear discriminant improvement, multi-classification support vector machine joint improvement, to obtain a multi-dimensional linear joint improved multi-classification support vector machine model, including the following steps:

[0031] Step S331: improving the objective function of multi-classification of abnormal faults, specifically, decomposing the weight matrix of the basic model of the multi-classification support vector machine into a first projection matrix and a second projection matrix, constructing the original model of multi-classification of abnormal faults, and constructing the objective function of multi-classification of abnormal faults, and improving the loss optimization objective function by introducing Lagrange multiplier parameters;

[0032] The calculation formula of the abnormal fault multi-classification objective function is:

[0033] ;

[0034] In the formula, F SPQ is the abnormal fault multi-classification objective function, w is the support vector weight, b is the support vector bias term, is a slack variable used to indicate the degree of misclassification, P is the first projection matrix parameter, Q is the second projection matrix parameter, C is the support vector regularization parameter, x is the original data input, i is the category index, and w i is the support vector weight corresponding to the i-th category, n is the total number of data samples, j is the data sample index, is the slack variable corresponding to the jth data sample of the ith category, is the regularization parameter, Y is the classification information matrix, X T is the transpose of the original feature matrix, P i is the first projection matrix of the ith category, Q i is the second projection matrix of the ith category, is the abnormal fault multi-classification improvement to maximize the selected category score, is the transpose of the first projection matrix of the i-th category;

[0035] The improved calculation formula of the loss optimization objective function is:

[0036] ;

[0037] Where, L i is the improved loss optimization objective function corresponding to the i-th data category, is the Lagrange multiplier parameter for the jth data sample of the ith category, is the Lagrange multiplier penalty parameter for the jth data sample corresponding to the i-th category;

[0038] Step S332: Linear discriminant improvement, specifically, introducing orthogonal constraints in the first projection matrix and the second projection matrix through linear discriminant improvement, and constructing a first projection matrix constraint function and a second projection matrix constraint function;

[0039] The calculation formula of the first projection matrix constraint function is:

[0040] ;

[0041] Where P i is the first projection matrix of the i-th category, where X is the original feature matrix, X is used to represent the first projection matrix after linear discriminant improvement and the introduction of orthogonal constraints. T is the transpose of the original feature matrix, Y is the classification information matrix, which is used to provide the category feature information of signal fault interference, Y T is the transpose of the classification information matrix, is the transpose of the second projection matrix of the ith category, is the regularization parameter, v i is the Lagrange multiplier weight corresponding to the i-th category, is the spatial dimension identity matrix, s is the subspace dimension, which is used to represent the total number of category dimensions;

[0042] The calculation formula of the second projection matrix constraint function is:

[0043] ;

[0044] In the formula, Q i is the second projection matrix of the i-th category, where is specifically used to represent the second projection matrix after linear discriminant improvement and introduction of orthogonal constraints;

[0045] Step S333: alternating iterative joint optimization, specifically, performing joint optimization of a multi-classification support vector machine by alternately updating the Lagrange multiplier parameters, the first projection matrix, and the second projection matrix, and performing global optimal solution calculation of the first projection matrix and the second projection matrix by fixing the Lagrange multiplier parameters, so as to obtain an optimal multi-classification support vector machine model;

[0046] Step S34: Improve multi-classification support vector machine iterative training, specifically, through the multi-classification support vector machine basic model, the linear discriminant feature dimensionality reduction optimization and the multi-dimensional linear joint improvement, improve the multi-classification support vector machine iterative training, and obtain the abnormal interference detection model Model CS ;

[0047] Step S35: abnormal interference detection, specifically using the abnormal interference detection model Model CS , performing abnormal interference detection based on the spectrum analysis optimization data to obtain abnormal interference detection data;

[0048] The detection categories of the abnormal interference detection data specifically include sudden interference, continuous narrowband interference, broadband noise interference, transient frequency drift interference and nonlinear abnormal distortion interference.

[0049] Further, in step S4, the spectrum trend prediction is used to predict and construct the change trend of the spectrum data, specifically, based on the spectrum analysis optimization data and the abnormal interference detection data, a time series integration network combined with one-dimensional convolution feature extraction is used to perform spectrum trend prediction to obtain spectrum trend prediction data, specifically including the following steps:

[0050] Step S41: Convolution feature extraction, specifically constructing a one-dimensional convolution feature extraction layer, extracting local trend features from the spectrum analysis optimization data and abnormal interference detection data, and constructing average pooling to perform feature dimensionality reduction to obtain convolution feature data;

[0051] Step S42: constructing a bidirectional long short-term memory subnet, specifically constructing a standard bidirectional long short-term memory subnet, extracting long-term trend features based on the convolution feature data through a bidirectional structure, and obtaining bidirectional temporal network feature data;

[0052] Step S43: constructing a gated recurrent subnet, specifically constructing a standard gated recurrent unit including a reset gate and an update gate as a gated recurrent subnet, and extracting short-term change trend features according to the convolution feature data through the gated recurrent subnet to obtain short-term temporal network feature data;

[0053] Step S44: time series network integration, specifically, integrating the bidirectional time series network feature data and the short-term time series network feature data through feature splicing to obtain a time series integrated network feature output, and performing trend prediction according to the time series integrated network feature output to obtain trend prediction output data;

[0054] Step S45: Spectrum trend prediction model training, specifically, performing spectrum trend prediction model training through the convolution feature extraction, the bidirectional long short-term memory subnet, the gated recurrent subnet and the timing network integration to obtain the spectrum trend prediction model Model SP ;

[0055] Step S46: Spectrum trend prediction, specifically using the spectrum trend prediction model Model SP , performing spectrum trend prediction based on the spectrum analysis optimization data and abnormal interference detection data to obtain spectrum trend prediction data;

[0056] The spectrum trend prediction data specifically includes spectrum signal time period prediction values, spectrum signal trend characteristics and abnormal trend identifiers.

[0057] Further, in step S5, the spectrum scheduling management is used to perform spectrum dynamic optimization and spectrum resource allocation optimization in combination with the results of abnormal interference detection and spectrum trend prediction, specifically, based on the spectrum analysis optimization data, the abnormal interference detection data and the spectrum trend prediction data, a reinforcement learning method enhanced by a proximal strategy optimization algorithm is used to perform spectrum scheduling management to obtain wide bandwidth frequency analysis scheduling reference data, specifically including the following steps:

[0058] Step S51: reinforcement learning environment construction, specifically, constructing a reinforcement learning environment through reinforcement learning modeling, and performing reinforcement learning training according to the reinforcement learning environment, wherein the reinforcement learning modeling specifically includes state parameter construction, action parameter construction and reward function construction;

[0059] The state parameters specifically include spectrum signal occupancy parameters, historical spectrum trend parameters, abnormal fault interference parameters and predicted trend parameters;

[0060] The action parameters specifically include bandwidth allocation adjustment actions, spectrum resource time slot scheduling actions, and high-load area spectrum usage planning actions;

[0061] The reward function is specifically calculated by weighted combination of resource utilization maximization target, interference minimization target and system delay minimization target;

[0062] Step S52: constructing a proximal policy gradient optimization algorithm, specifically by limiting the policy update amplitude and constructing a clipping objective function to construct a proximal policy gradient optimization algorithm;

[0063] Step S53: Proximal policy reinforcement learning enhancement, specifically, by applying the proximal policy gradient optimization algorithm, according to the reinforcement learning environment, performing reinforcement learning training enhanced by the proximal policy optimization algorithm until the training converges, and performing reinforcement learning enhancement;

[0064] Step S54: spectrum scheduling management model training, specifically, training the spectrum scheduling management model through the reinforcement learning environment construction, the proximal policy gradient optimization algorithm and the proximal policy reinforcement learning enhancement, to obtain the spectrum scheduling management model Model SM ;

[0065] Step S55: spectrum scheduling management, specifically using the spectrum scheduling management model Model SM , performing spectrum scheduling management according to the spectrum analysis optimization data, the abnormal interference detection data and the spectrum trend prediction data, and obtaining wide bandwidth frequency analysis scheduling reference data;

[0066] The wide bandwidth frequency analysis scheduling reference data specifically includes spectrum resource allocation scheme, spectrum utilization, abnormal area interference information reference, prediction trend optimization results and dynamic scheduling optimization results.

[0067] The present invention provides a wide-bandwidth frequency spectrum analysis system based on deep learning, comprising a spectrum signal acquisition module, a spectrum preprocessing module, an abnormal interference detection module, a spectrum trend prediction module and a spectrum scheduling management module;

[0068] The spectrum signal acquisition module is used for spectrum signal acquisition, obtains a wide-frequency spectrum analysis original data set through spectrum signal acquisition, and sends the wide-frequency spectrum analysis original data set to the spectrum preprocessing module;

[0069] The spectrum preprocessing module is used for spectrum preprocessing, obtains spectrum analysis optimization data through spectrum preprocessing, and sends the spectrum analysis optimization data to the abnormal interference detection module, the spectrum trend prediction module and the spectrum scheduling management module;

[0070] The abnormal interference detection module is used for abnormal interference detection, obtains abnormal interference detection data through abnormal interference detection, and sends the abnormal interference detection data to the spectrum trend prediction module;

[0071] The spectrum trend prediction module is used for spectrum trend prediction, obtains spectrum trend prediction data through spectrum trend prediction, and sends the spectrum trend prediction data to the spectrum scheduling management module;

[0072] The spectrum scheduling management module is used for spectrum scheduling management, and obtains wide bandwidth frequency analysis scheduling reference data through spectrum scheduling management.

[0073] The beneficial effects achieved by the present invention using the above scheme are as follows:

[0074] (1) In the existing wide-bandwidth frequency spectrum analysis methods, the existing intelligent methods often focus on single signal processing or spectrum analysis, and pay more attention to the extraction of static characteristics of the global spectrum. However, in the wide-bandwidth spectrum analysis, due to the wider range of wide-bandwidth data and greater data changes, traditional methods cannot accurately locate abnormal interference, lack the ability to understand the long-term trend of the spectrum, and lack intelligent optimization mechanisms in spectrum resource scheduling. This solution creatively adopts a three-step method of abnormal interference detection-spectrum trend prediction-spectrum scheduling management. Through layer-by-layer deep learning and reinforcement learning technology, it can realize real-time detection of spectrum anomalies, accurate prediction of future trends, and dynamic resource optimization allocation. This method can not only fully capture the complex characteristics of wide-bandwidth frequency spectrum, but also significantly improve spectrum utilization efficiency and environmental adaptability, making it more advantageous in complex and changing scenarios;

[0075] (2) Among the existing abnormal interference detection methods, the most commonly used technical method is the support vector machine method. However, the traditional support vector machine is a basic binary classification model, and its classification accuracy, feature extraction depth and classification effect are difficult to meet the complex abnormal interference type detection requirements in spectrum analysis. The method of stacking multiple support vector machines is more likely to increase the computational difficulty in the complex integration process, and it is difficult to effectively improve the spectrum analysis performance. This solution creatively adopts a joint improved multi-classification support vector machine method combined with feature enhancement to perform abnormal interference detection. Through feature enhancement and multi-dimensional joint improvement, the classifiable categories and overall classification performance of the support vector machine are improved. The training ideas of weight decomposition and projection parameter optimization are used to improve the degree of performance optimization, providing better data guarantee for abnormal interference detection.

[0076] (3) In the existing spectrum trend prediction methods, there is a technical problem that the spectrum trend prediction adopts the time series model for technical processing in the existing technical scheme, but the existing bidirectional time series network has better long-term feature capture ability, but has poor ability to capture the trend changes of spectrum signals at the micro level. This scheme creatively adopts the time series integration network combined with one-dimensional convolution feature extraction to predict the spectrum trend. By constructing an integration of a bidirectional long short-term memory network and a unidirectional gated recurrent network, the two model structures complement each other through time series integration, which not only optimizes the prediction effect of macro spectrum trend characteristics, but also improves the ability to capture short-term trend changes, and comprehensively improves the effect of spectrum trend prediction;

[0077] (4) In view of the fact that the existing spectrum scheduling management methods usually rely on traditional rule-based optimization algorithms (such as greedy algorithms or linear programming), these methods perform well in small-scale or simple scenarios, but lack global adaptability and real-time optimization capabilities when faced with dynamic changes in spectrum resource distribution, complex multi-dimensional interference factors and nonlinear resource demands. This solution creatively adopts a reinforcement learning method enhanced by a proximal policy optimization algorithm. Through the adaptive characteristics of reinforcement learning and the stability of the proximal policy optimization algorithm, it can efficiently explore and utilize the optimal strategy for spectrum scheduling, dynamically respond to changes in the spectrum environment, and minimize the instability in the strategy update process. The clipping mechanism of the proximal policy optimization algorithm also effectively balances exploration and utilization, thereby achieving efficient and robust spectrum resource allocation. This method can not only improve spectrum utilization, but also better adapt to complex and dynamic application scenarios, providing exploration and practical experience for the effective application of wide-frequency spectrum analysis. BRIEF DESCRIPTION OF THE DRAWINGS

[0078] Figure 1A schematic diagram of a flow chart of a wide-bandwidth frequency spectrum analysis method based on deep learning provided by the present invention;

[0079] Figure 2 A schematic diagram of a wide-bandwidth frequency spectrum analysis system based on deep learning provided by the present invention;

[0080] Figure 3 This is a flow chart of abnormal interference detection in step S3;

[0081] Figure 4 This is a schematic diagram of the improved multi-dimensional linear combination process of step S33;

[0082] Figure 5 A schematic diagram of the process of spectrum trend prediction in step S4;

[0083] Figure 6 This is a flowchart of spectrum scheduling management in step S5.

[0084] The accompanying drawings are used to provide further understanding of the present invention and constitute a part of the specification. They are used to explain the present invention together with the embodiments of the present invention and do not constitute a limitation of the present invention. DETAILED DESCRIPTION

[0085] The technical solutions in the embodiments of the present invention will be clearly and completely described below in conjunction with the drawings in the embodiments of the present invention. Obviously, the described embodiments are only part of the embodiments of the present invention, rather than all the embodiments; based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without making creative work are within the scope of protection of the present invention.

[0086] In the description of the present invention, it should be understood that terms such as “upper”, “lower”, “front”, “back”, “left”, “right”, “top”, “bottom”, “inside” and “outside” indicating directions or positional relationships are based on the directions or positional relationships shown in the accompanying drawings, and are only for the convenience of describing the present invention and simplifying the description, rather than indicating or implying that the device or element referred to must have a specific direction, be constructed and operated in a specific direction, and therefore should not be understood as limiting the present invention.

[0087] Example 1, see Figure 1 The present invention provides a wide bandwidth frequency spectrum analysis method based on deep learning, which comprises the following steps:

[0088] Step S1: spectrum signal acquisition;

[0089] Step S2: spectrum preprocessing;

[0090] Step S3: abnormal interference detection;

[0091] Step S4: spectrum trend prediction;

[0092] Step S5: spectrum scheduling management.

[0093] By performing the above operations, in the existing wide-bandwidth frequency spectrum analysis methods, there is a problem that the existing intelligent methods often focus on a single signal processing or spectrum analysis, and pay more attention to the static characteristics extraction of the global spectrum. However, in the wide-frequency spectrum analysis, due to the wider range of wide-frequency data and greater data changes, the traditional methods cannot accurately locate abnormal interference, lack the ability to understand the long-term trend of the spectrum, and lack the technical problems of intelligent optimization mechanism in spectrum resource scheduling. This solution creatively adopts a three-step method of abnormal interference detection-spectrum trend prediction-spectrum scheduling management. Through layer-by-layer progressive deep learning and reinforcement learning technology, it can realize real-time detection of spectrum anomalies, accurate prediction of future trends, and dynamic resource optimization allocation. This method can not only comprehensively capture the complex characteristics of the wide-bandwidth frequency spectrum, but also significantly improve the spectrum utilization efficiency and environmental adaptability, making it more advantageous in complex and changeable scenarios.

[0094] Example 2, see Figure 1 and Figure 2 In step S1, the spectrum signal acquisition is used to collect the original spectrum signal data set required for the wide-bandwidth frequency spectrum analysis, specifically, by performing wide-bandwidth frequency spectrum data acquisition to obtain the original data set for the wide-bandwidth frequency spectrum analysis;

[0095] The broadband spectrum data acquisition specifically includes spectrum signal source setting, acquisition equipment setting, spectrum acquisition, data storage and preliminary optimization of data quality;

[0096] The spectrum signal source setting is specifically to collect spectrum data of the 5G communication system, and select the Sub-6 GHz frequency band as the collection range;

[0097] The acquisition device is specifically set to use a wideband spectrum analyzer to collect spectrum, and the sampling rate is set to 10 MHz and the sampling bandwidth is set to 100 MHz;

[0098] The spectrum acquisition specifically includes performing spectrum acquisition operation by setting the spectrum acquisition range, acquisition time interval and data output form;

[0099] The spectrum acquisition range is specifically to set the device to spectrum analysis mode and to collect sub-segment spectrum data for a frequency band ranging from 3.3 GHz to 3.8 GHz;

[0100] The acquisition time interval is specifically to perform multi-segment time interval acquisition, including a 1s time interval, a 5s time interval and a 10s time interval;

[0101] The data output form specifically refers to the amplitude spectrum and power spectrum density, in dBm / Hz, which are used to represent the signal strength at each frequency point;

[0102] The data storage specifically includes performing initial feature extraction on the collected data to obtain an original spectrum feature data set, and storing the data in a CSV format;

[0103] The data quality is initially optimized, specifically by detecting the signal loss data, electromagnetic interference data and equipment failure abnormal data in the original video spectrum feature data set through manual screening, and performing preliminary optimization of the data quality through manual data marking and filling;

[0104] The wide frequency spectrum analysis original data set specifically includes timestamp data, frequency data, signal amplitude data, signal noise data, signal sampling rate data and sampling device information data.

[0105] Example 3, see Figure 1 and Figure 2 This embodiment is based on the above embodiment. In step S2, the spectrum preprocessing is used to optimize and enhance the original spectrum signal data, specifically by performing signal denoising, data normalization and data quality optimization operations to obtain spectrum analysis optimization data, which specifically includes the following steps:

[0106] Step S21: signal denoising, specifically removing high-frequency noise through a wavelet transform method, and performing complex signal denoising through a deep autoencoder to obtain denoised optimized spectrum data;

[0107] Step S22: data standardization, specifically, performing a Z-score standardization operation on the feature data in the denoised and optimized spectrum data to obtain standardized spectrum data;

[0108] Step S23: filling optimization, specifically, filling the missing data in the standardized spectrum data by using the mean filling method to obtain filled optimized spectrum data;

[0109] Step S24: smoothing optimization, specifically, smoothing the padded optimized spectrum data by using a polynomial fitting filter to obtain smoothed optimized spectrum data;

[0110] Step S25: feature extraction, specifically extracting frequency domain statistical features from the smoothed optimized spectrum data to obtain a frequency domain feature data set, wherein the frequency domain feature data set specifically includes a signal strength peak feature, a frequency distribution feature, a bandwidth feature, a noise baseline feature, and a spectrum density feature;

[0111] Step S26: spectrum preprocessing, specifically, performing spectrum preprocessing through the signal denoising, the data standardization, the padding optimization, the smoothing optimization and the feature extraction to obtain spectrum analysis optimization data;

[0112] The spectrum analysis optimization data specifically includes denoising optimized spectrum data, standardized spectrum data, smoothing optimized spectrum data and frequency domain feature data set.

[0113] Example 4, see Figure 1 , Figure 2 , Figure 3 and Figure 4 This embodiment is based on the above embodiment. In step S3, the abnormal interference detection is used to detect abnormal data in the spectrum signal and perform abnormal classification. Specifically, based on the spectrum analysis optimization data, a joint improved multi-classification support vector machine method combined with feature enhancement is used to perform abnormal interference detection to obtain abnormal interference detection data. Specifically, the following steps are included:

[0114] Step S31: construct a basic model of a multi-classification support vector machine, specifically by constructing a stacked binary support vector machine objective function, performing multi-classification calculations by calculating the maximum category score of the objective function, and applying a resampling method to optimize the category imbalance problem, thereby constructing a basic model of a multi-classification support vector machine. The calculation formula of the stacked binary support vector machine objective function is:

[0115] ;

[0116] In the formula, F TSVM is the stacked binary support vector machine objective function, w is the support vector weight, b is the support vector bias term, is a slack variable used to represent the degree of misclassification, C is the support vector regularization parameter, x is the original data input, used to represent the spectrum analysis optimization data, i is the category index, and w i is the support vector weight corresponding to the i-th category, n is the total number of data samples, j is the data sample index, is the slack variable corresponding to the jth data sample of the ith category, is the maximum selected category score, which is used to represent the final category output detected after stacking, b i is the support vector bias term corresponding to the i-th category;

[0117] Step S32: linear discriminant feature dimensionality reduction optimization, specifically constructing a linear discriminant analysis operator, amplifying the variance of samples between classes and reducing the variance of samples within classes, and performing feature dimensionality reduction optimization on the frequency domain feature data set in the spectrum analysis optimization data through orthogonal linear discriminant analysis to obtain dimensionality reduction optimized feature data;

[0118] The linear discriminant analysis operator specifically includes total scatter matrix construction, inter-class matrix construction and orthogonal linear discriminant analysis modeling;

[0119] The calculation formula for constructing the total scatter matrix is:

[0120] ;

[0121] In the formula, S t is the total scatter matrix, which is used to represent the overall variance distribution of all samples, X is the original feature matrix, X T is the transpose of the original feature matrix;

[0122] The calculation formula for constructing the inter-class matrix is:

[0123] ;

[0124] In the formula, S b is the inter-class matrix, which is used to represent the distribution of the mean values ​​of each fault interference class. X is the original feature matrix, X T is the transpose of the original feature matrix, Y is the classification information matrix, which is used to provide the category feature information of signal fault interference, Y T is the transpose of the classification information matrix;

[0125] The calculation formula of the orthogonal linear discriminant analysis modeling is:

[0126] ;

[0127] In the formula, F LDA is the objective function of orthogonal linear discriminant analysis, A is the projection matrix, which is used as the optimal linear discriminant parameter, Tr is the matrix trace parameter, which is used to represent the sum of the matrix diagonal matrix, S t is the total divergence matrix, S b is the between-class matrix;

[0128] Step S33: multi-dimensional linear joint improvement, specifically, three-dimensional iterative linear joint improvement, including abnormal fault multi-classification decision function improvement, iterative objective function improvement and linear discriminant improvement, multi-classification support vector machine joint improvement, to obtain a multi-dimensional linear joint improved multi-classification support vector machine model, including the following steps:

[0129] Step S331: improving the objective function of multi-classification of abnormal faults, specifically, decomposing the weight matrix of the basic model of the multi-classification support vector machine into a first projection matrix and a second projection matrix, constructing the original model of multi-classification of abnormal faults, and constructing the objective function of multi-classification of abnormal faults, and improving the loss optimization objective function by introducing Lagrange multiplier parameters;

[0130] The calculation formula of the abnormal fault multi-classification objective function is:

[0131] ;

[0132] In the formula, F SPQ is the abnormal fault multi-classification objective function, w is the support vector weight, b is the support vector bias term, is a slack variable used to indicate the degree of misclassification, P is the first projection matrix parameter, Q is the second projection matrix parameter, C is the support vector regularization parameter, x is the original data input, i is the category index, and w i is the support vector weight corresponding to the i-th category, n is the total number of data samples, j is the data sample index, is the slack variable corresponding to the jth data sample of the ith category, is the regularization parameter, Y is the classification information matrix, X T is the transpose of the original feature matrix, P i is the first projection matrix of the ith category, Q i is the second projection matrix of the ith category, is the abnormal fault multi-classification improvement to maximize the selected category score, is the transpose of the first projection matrix of the i-th category;

[0133] The improved calculation formula of the loss optimization objective function is:

[0134] ;

[0135] Where, L i is the improved loss optimization objective function corresponding to the i-th data category, is the Lagrange multiplier parameter for the jth data sample of the ith category, is the Lagrange multiplier penalty parameter for the jth data sample corresponding to the i-th category;

[0136] Step S332: Linear discriminant improvement, specifically, introducing orthogonal constraints in the first projection matrix and the second projection matrix through linear discriminant improvement, and constructing a first projection matrix constraint function and a second projection matrix constraint function;

[0137] The calculation formula of the first projection matrix constraint function is:

[0138] ;

[0139] Where P i is the first projection matrix of the i-th category, where X is the original feature matrix, X is used to represent the first projection matrix after linear discriminant improvement and the introduction of orthogonal constraints. Tis the transpose of the original feature matrix, Y is the classification information matrix, which is used to provide the category feature information of signal fault interference, Y T is the transpose of the classification information matrix, is the transpose of the second projection matrix of the ith category, is the regularization parameter, v i is the Lagrange multiplier weight corresponding to the i-th category, is the spatial dimension identity matrix, s is the subspace dimension, which is used to represent the total number of category dimensions;

[0140] The calculation formula of the second projection matrix constraint function is:

[0141] ;

[0142] In the formula, Q i is the second projection matrix of the i-th category, where is specifically used to represent the second projection matrix after linear discriminant improvement and introduction of orthogonal constraints;

[0143] Step S333: alternating iterative joint optimization, specifically, performing joint optimization of a multi-classification support vector machine by alternately updating the Lagrange multiplier parameters, the first projection matrix, and the second projection matrix, and performing global optimal solution calculation of the first projection matrix and the second projection matrix by fixing the Lagrange multiplier parameters, so as to obtain an optimal multi-classification support vector machine model;

[0144] Step S34: Improve multi-classification support vector machine iterative training, specifically, through the multi-classification support vector machine basic model, the linear discriminant feature dimensionality reduction optimization and the multi-dimensional linear joint improvement, improve the multi-classification support vector machine iterative training, and obtain the abnormal interference detection model Model CS ;

[0145] Step S35: abnormal interference detection, specifically using the abnormal interference detection model Model CS , performing abnormal interference detection based on the spectrum analysis optimization data to obtain abnormal interference detection data;

[0146] The detection categories of the abnormal interference detection data specifically include sudden interference, continuous narrowband interference, broadband noise interference, transient frequency drift interference and nonlinear abnormal distortion interference.

[0147] By performing the above operations, among the existing abnormal interference detection methods, there is a relatively commonly used technical method for abnormal interference detection, which is the support vector machine method. However, the traditional support vector machine is a basic model of two classifications, and the classification accuracy, feature extraction depth and classification effect are difficult to meet the complex abnormal interference type detection needs in spectrum analysis. The method of stacking multiple support vector machines is more likely to increase the computational difficulty in a complex integration process, and it is difficult to effectively obtain the technical problem of improving the spectrum analysis performance. This scheme creatively adopts a joint improved multi-classification support vector machine method combined with feature enhancement to perform abnormal interference detection. Through feature enhancement and multi-dimensional joint improvement, the classifiable categories and overall classification performance of the support vector machine are improved, and the training ideas of weight decomposition and projection parameter optimization are used to improve the degree of performance optimization, providing better data guarantee for abnormal interference detection.

[0148] Example 5, see Figure 1 , Figure 2 and Figure 5 This embodiment is based on the above embodiment. In step S4, the spectrum trend prediction is used to predict and construct the change trend of the spectrum data. Specifically, based on the spectrum analysis optimization data and the abnormal interference detection data, a time series integration network combined with one-dimensional convolution feature extraction is used to perform spectrum trend prediction to obtain spectrum trend prediction data. Specifically, the following steps are included:

[0149] Step S41: Convolution feature extraction, specifically constructing a one-dimensional convolution feature extraction layer, extracting local trend features from the spectrum analysis optimization data and abnormal interference detection data, and constructing average pooling to perform feature dimensionality reduction to obtain convolution feature data;

[0150] Step S42: constructing a bidirectional long short-term memory subnet, specifically constructing a standard bidirectional long short-term memory subnet, extracting long-term trend features based on the convolution feature data through a bidirectional structure, and obtaining bidirectional temporal network feature data;

[0151] The calculation formula of the standard bidirectional long short-term memory subnet is:

[0152] ;

[0153] In the formula, is the bidirectional temporal network characteristic data, is the forward output hidden state, is the backward output hidden state, t is the time index;

[0154] Step S43: constructing a gated recurrent subnet, specifically constructing a standard gated recurrent unit including a reset gate and an update gate as a gated recurrent subnet, and extracting short-term change trend features according to the convolution feature data through the gated recurrent subnet to obtain short-term temporal network feature data;

[0155] Step S44: time series network integration, specifically, integrating the bidirectional time series network feature data and the short-term time series network feature data through feature splicing to obtain a time series integrated network feature output, and performing trend prediction according to the time series integrated network feature output to obtain trend prediction output data;

[0156] The calculation formula of the feature splicing is:

[0157] ;

[0158] Where H is the characteristic output of the time series integration network, is the bidirectional temporal network characteristic data, It is the short-term time series network characteristic data;

[0159] Step S45: Spectrum trend prediction model training, specifically, performing spectrum trend prediction model training through the convolution feature extraction, the bidirectional long short-term memory subnet, the gated recurrent subnet and the timing network integration to obtain the spectrum trend prediction model Model SP ;

[0160] Step S46: Spectrum trend prediction, specifically using the spectrum trend prediction model Model SP , performing spectrum trend prediction based on the spectrum analysis optimization data and abnormal interference detection data to obtain spectrum trend prediction data;

[0161] The spectrum trend prediction data specifically includes spectrum signal time period prediction values, spectrum signal trend characteristics and abnormal trend identifiers.

[0162] By performing the above operations, in order to solve the technical problem that in the existing spectrum trend prediction methods, the spectrum trend prediction is processed by the time series model in the existing technical scheme, but the existing bidirectional time series network has better long-term feature capture capability, but has poor capture capability of the trend change of the spectrum signal at the micro level, this scheme creatively adopts the time series integration network combined with one-dimensional convolution feature extraction to perform spectrum trend prediction, and integrates the bidirectional long short-term memory network and the unidirectional gated recurrent network, and the two model structures complement each other through time series integration, which not only optimizes the prediction effect of the macro spectrum trend characteristics, but also improves the capture capability of the short-term trend change, and comprehensively improves the effect of the spectrum trend prediction.

[0163] Example 6, see Figure 1 and Figure 2 This embodiment is based on the above embodiment. In step S5, the spectrum scheduling management is used to perform spectrum dynamic optimization and spectrum resource allocation optimization in combination with the results of abnormal interference detection and spectrum trend prediction. Specifically, based on the spectrum analysis optimization data, the abnormal interference detection data and the spectrum trend prediction data, a reinforcement learning method enhanced by a proximal strategy optimization algorithm is used to perform spectrum scheduling management to obtain wide bandwidth frequency analysis scheduling reference data. Specifically, the following steps are included:

[0164] Step S51: reinforcement learning environment construction, specifically, constructing a reinforcement learning environment through reinforcement learning modeling, and performing reinforcement learning training according to the reinforcement learning environment, wherein the reinforcement learning modeling specifically includes state parameter construction, action parameter construction and reward function construction;

[0165] The state parameters specifically include spectrum signal occupancy parameters, historical spectrum trend parameters, abnormal fault interference parameters and predicted trend parameters;

[0166] The action parameters specifically include bandwidth allocation adjustment actions, spectrum resource time slot scheduling actions, and high-load area spectrum usage planning actions;

[0167] The reward function is specifically calculated by weighted combination of resource utilization maximization target, interference minimization target and system delay minimization target;

[0168] Step S52: construct a proximal policy gradient optimization algorithm, specifically by limiting the policy update amplitude and constructing a clipping objective function to construct a proximal policy gradient optimization algorithm. The calculation formula is:

[0169] ;

[0170] Where, L PPO is the proximal policy gradient optimization function, E t (·) obtains the expected function for the time index t, min(·) is the minimization objective function, is the strategy probability ratio parameter, which is used to indicate the degree of change between the current strategy and the previous iteration strategy. is a reinforcement learning variable, used to represent the parameter set in the reinforcement learning environment, is the advantage function, which is used to represent the comparison between the current strategy and the average strategy. clip(·) is the clipping operation function. It is the discriminant parameter of the clipping operation, which is used to stabilize the training process;

[0171] Step S53: Proximal policy reinforcement learning enhancement, specifically, by applying the proximal policy gradient optimization algorithm, according to the reinforcement learning environment, performing reinforcement learning training enhanced by the proximal policy optimization algorithm until the training converges, and performing reinforcement learning enhancement;

[0172] Step S54: spectrum scheduling management model training, specifically, training the spectrum scheduling management model through the reinforcement learning environment construction, the proximal policy gradient optimization algorithm and the proximal policy reinforcement learning enhancement, to obtain the spectrum scheduling management model Model SM ;

[0173] Step S55: spectrum scheduling management, specifically using the spectrum scheduling management model Model SM , performing spectrum scheduling management according to the spectrum analysis optimization data, the abnormal interference detection data and the spectrum trend prediction data, and obtaining wide bandwidth frequency analysis scheduling reference data;

[0174] The wide bandwidth frequency analysis scheduling reference data specifically includes spectrum resource allocation scheme, spectrum utilization, abnormal area interference information reference, prediction trend optimization results and dynamic scheduling optimization results.

[0175] By performing the above operations, in view of the fact that the existing spectrum scheduling management methods usually rely on traditional rule-based optimization algorithms (such as greedy algorithms or linear programming), these methods perform well in small-scale or simple scenarios, but lack global adaptability and real-time optimization capabilities when facing dynamic changes in spectrum resource distribution, complex multi-dimensional interference factors and nonlinear resource requirements. This solution creatively adopts a reinforcement learning method enhanced by a proximal policy optimization algorithm. Through the adaptive characteristics of reinforcement learning and the stability of the proximal policy optimization algorithm, it can efficiently explore and utilize the optimal strategy for spectrum scheduling, dynamically respond to changes in the spectrum environment, and minimize the instability in the strategy update process. The clipping mechanism of the proximal policy optimization algorithm also effectively balances exploration and utilization, thereby achieving efficient and robust spectrum resource allocation. This method can not only improve spectrum utilization, but also better adapt to complex and dynamic application scenarios, and provide exploration and practical experience for the effective application of wide-frequency spectrum analysis.

[0176] Embodiment 7, see Figure 1 and Figure 2 , this embodiment is based on the above embodiment, and a wide-bandwidth frequency spectrum analysis system based on deep learning provided by the present invention includes a spectrum signal acquisition module, a spectrum preprocessing module, an abnormal interference detection module, a spectrum trend prediction module and a spectrum scheduling management module;

[0177] The spectrum signal acquisition module is used for spectrum signal acquisition, obtains a wide-frequency spectrum analysis original data set through spectrum signal acquisition, and sends the wide-frequency spectrum analysis original data set to the spectrum preprocessing module;

[0178] The spectrum preprocessing module is used for spectrum preprocessing, obtains spectrum analysis optimization data through spectrum preprocessing, and sends the spectrum analysis optimization data to the abnormal interference detection module, the spectrum trend prediction module and the spectrum scheduling management module;

[0179] The abnormal interference detection module is used for abnormal interference detection, obtains abnormal interference detection data through abnormal interference detection, and sends the abnormal interference detection data to the spectrum trend prediction module;

[0180] The spectrum trend prediction module is used for spectrum trend prediction, obtains spectrum trend prediction data through spectrum trend prediction, and sends the spectrum trend prediction data to the spectrum scheduling management module;

[0181] The spectrum scheduling management module is used for spectrum scheduling management, and obtains wide bandwidth frequency analysis scheduling reference data through spectrum scheduling management.

[0182] It should be noted that, in this article, relational terms such as first and second, etc. are only used to distinguish one entity or operation from another entity or operation, and do not necessarily require or imply any such actual relationship or order between these entities or operations. Moreover, the terms "include", "comprise" or any other variants thereof are intended to cover non-exclusive inclusion, so that a process, method, article or device including a series of elements includes not only those elements, but also other elements not explicitly listed, or also includes elements inherent to such process, method, article or device.

[0183] While the embodiments of the present invention have been shown and described, it will be appreciated by those skilled in the art that many changes, modifications, substitutions and variations can be made to the embodiments without departing from the principles and spirit of the invention.

[0184] The present invention and its embodiments are described above, and such description is not restrictive. The drawings show only one embodiment of the present invention, and the actual structure is not limited thereto. In short, if ordinary technicians in the field are inspired by it, without departing from the purpose of the invention, they can design a structure and embodiment similar to the technical solution without creativity, which should belong to the protection scope of the present invention.

Claims

1. A wide bandwidth frequency spectrum analysis method based on deep learning, characterized in that: The method comprises the following steps: Step S1: spectrum signal acquisition, through broadband spectrum data acquisition, broadband frequency spectrum data acquisition is performed to obtain a wide-frequency spectrum analysis original data set; Step S2: spectrum preprocessing, by performing signal denoising, data normalization and data quality optimization operations to obtain spectrum analysis optimization data; Step S3: abnormal interference detection, used to detect abnormal data in the spectrum signal and perform abnormal classification, specifically based on the spectrum analysis optimization data, adopt the joint improved multi-classification support vector machine method combined with feature enhancement to perform abnormal interference detection, and obtain abnormal interference detection data, specifically including the following steps: step S31: construct a basic model of a multi-classification support vector machine; step S32: linear discriminant feature dimensionality reduction optimization; step S33: multi-dimensional linear joint improvement; step S34: improved multi-classification support vector machine iterative training; step S35: abnormal interference detection; Step S4: spectrum trend prediction, which is used to predict and construct the change trend of spectrum data, specifically, based on the spectrum analysis optimization data and abnormal interference detection data, a time series integration network combined with one-dimensional convolution feature extraction is used to perform spectrum trend prediction to obtain spectrum trend prediction data; Step S5: spectrum scheduling management, which is used to perform dynamic spectrum optimization and spectrum resource allocation optimization based on the results of abnormal interference detection and spectrum trend prediction. Specifically, based on the spectrum analysis optimization data, the abnormal interference detection data and the spectrum trend prediction data, a reinforcement learning method enhanced by a proximal strategy optimization algorithm is used to perform spectrum scheduling management to obtain wide-bandwidth frequency analysis scheduling reference data.

2. The method for wide bandwidth frequency spectrum analysis based on deep learning according to claim 1, characterized in that: In step S1, the wide frequency spectrum analyzes the original data set, which specifically includes timestamp data, frequency data, signal amplitude data, signal noise data, signal sampling rate data and sampling device information data.

3. The method for wide bandwidth frequency spectrum analysis based on deep learning according to claim 2, characterized in that: In step S3, the abnormal interference detection is used to detect abnormal data in the spectrum signal and perform abnormal classification. Specifically, based on the spectrum analysis optimization data, a joint improved multi-classification support vector machine method combined with feature enhancement is used to perform abnormal interference detection to obtain abnormal interference detection data, which specifically includes the following steps: Step S31: construct a basic model of a multi-classification support vector machine, specifically by constructing a stacked binary support vector machine objective function, performing multi-classification calculations by calculating the maximum category score of the objective function, and applying a resampling method to optimize the category imbalance problem, thereby constructing a basic model of a multi-classification support vector machine. The calculation formula of the stacked binary support vector machine objective function is: ； In the formula, F TSVM is the stacked binary support vector machine objective function, w is the support vector weight, b is the support vector bias term, is a slack variable used to represent the degree of misclassification, C is the support vector regularization parameter, x is the original data input, used to represent the spectrum analysis optimization data, i is the category index, and w i is the support vector weight corresponding to the i-th category, n is the total number of data samples, j is the data sample index, is the slack variable corresponding to the jth data sample of the ith category, is the maximum selected category score, which is used to represent the final category output detected after stacking, b i is the support vector bias term corresponding to the i-th category; Step S32: linear discriminant feature dimensionality reduction optimization, specifically constructing a linear discriminant analysis operator, amplifying the variance of samples between classes and reducing the variance of samples within classes, and performing feature dimensionality reduction optimization on the frequency domain feature data set in the spectrum analysis optimization data through orthogonal linear discriminant analysis to obtain dimensionality reduction optimized feature data; The linear discriminant analysis operator specifically includes total scatter matrix construction, inter-class matrix construction and orthogonal linear discriminant analysis modeling; The calculation formula for constructing the total scatter matrix is: ； In the formula, S t is the total scatter matrix, which is used to represent the overall variance distribution of all samples, X is the original feature matrix, X T is the transpose of the original feature matrix; The calculation formula for constructing the inter-class matrix is: ； In the formula, S b is the inter-class matrix, which is used to represent the distribution of the mean values ​​of each fault interference class. X is the original feature matrix, X T is the transpose of the original feature matrix, Y is the classification information matrix, which is used to provide the category feature information of signal fault interference, Y T is the transpose of the classification information matrix; The calculation formula of the orthogonal linear discriminant analysis modeling is: ； In the formula, F LDA is the objective function of orthogonal linear discriminant analysis, A is the projection matrix, which is used as the optimal linear discriminant parameter, Tr is the matrix trace parameter, which is used to represent the sum of the matrix diagonal matrix, S t is the total divergence matrix, S b is the between-class matrix; Step S33: multi-dimensional linear joint improvement, specifically, performing three-dimensional iterative linear joint improvement, including abnormal fault multi-classification decision function improvement, iterative objective function improvement and linear discriminant improvement, performing multi-classification support vector machine joint improvement, and obtaining a multi-dimensional linear joint improved multi-classification support vector machine model; Step S34: Improve the iterative training of the multi-classification support vector machine, specifically, through the multi-classification support vector machine basic model, the linear discriminant feature dimensionality reduction optimization and the multi-dimensional linear joint improvement, the improved multi-classification support vector machine iterative training is performed to obtain the abnormal interference detection model Model CS ; Step S35: abnormal interference detection, specifically using the abnormal interference detection model Model CS , performing abnormal interference detection based on the spectrum analysis optimization data to obtain abnormal interference detection data; The detection categories of the abnormal interference detection data specifically include sudden interference, continuous narrowband interference, broadband noise interference, transient frequency drift interference and nonlinear abnormal distortion interference.

4. The method for wide bandwidth frequency spectrum analysis based on deep learning according to claim 3, characterized in that: In step S33, the multi-dimensional linear joint improvement includes the following steps: Step S331: improving the objective function of multi-classification of abnormal faults, specifically, decomposing the weight matrix of the basic model of the multi-classification support vector machine into a first projection matrix and a second projection matrix, constructing the original model of multi-classification of abnormal faults, and constructing the objective function of multi-classification of abnormal faults, and improving the loss optimization objective function by introducing Lagrange multiplier parameters; The calculation formula of the abnormal fault multi-classification objective function is: ； In the formula, F SPQ is the abnormal fault multi-classification objective function, w is the support vector weight, b is the support vector bias term, is a slack variable used to indicate the degree of misclassification, P is the first projection matrix parameter, Q is the second projection matrix parameter, C is the support vector regularization parameter, x is the original data input, i is the category index, and w i is the support vector weight corresponding to the i-th category, n is the total number of data samples, j is the data sample index, is the slack variable corresponding to the jth data sample of the ith category, is the regularization parameter, Y is the classification information matrix, X T is the transpose of the original feature matrix, P i is the first projection matrix of the ith category, Q i is the second projection matrix of the ith category, is the abnormal fault multi-classification improvement to maximize the selected category score, is the transpose of the first projection matrix of the i-th category; The improved calculation formula of the loss optimization objective function is: ； Where, L i is the improved loss optimization objective function corresponding to the i-th data category, is the Lagrange multiplier parameter for the jth data sample of the ith category, is the Lagrange multiplier penalty parameter for the jth data sample corresponding to the i-th category; Step S332: Linear discriminant improvement, specifically, introducing orthogonal constraints in the first projection matrix and the second projection matrix through linear discriminant improvement, and constructing a first projection matrix constraint function and a second projection matrix constraint function; The calculation formula of the first projection matrix constraint function is: ； Where P i is the first projection matrix of the i-th category, where X is the original feature matrix, X is used to represent the first projection matrix after linear discriminant improvement and the introduction of orthogonal constraints. T is the transpose of the original feature matrix, Y is the classification information matrix, which is used to provide the category feature information of signal fault interference, Y T is the transpose of the classification information matrix, is the transpose of the second projection matrix of the ith category, is the regularization parameter, v i is the Lagrange multiplier weight corresponding to the i-th category, is the spatial dimension identity matrix, s is the subspace dimension, which is used to represent the total number of category dimensions; The calculation formula of the second projection matrix constraint function is: ； In the formula, Q i is the second projection matrix of the i-th category, where is specifically used to represent the second projection matrix after linear discriminant improvement and introduction of orthogonal constraints; Step S333: alternating iterative joint optimization, specifically, performing joint optimization of the multi-classification support vector machine by alternatingly updating the Lagrange multiplier parameters, the first projection matrix and the second projection matrix, and calculating the global optimal solution of the first projection matrix and the second projection matrix by fixing the Lagrange multiplier parameters to obtain the optimal multi-classification support vector machine model.

5. The method for wide bandwidth frequency spectrum analysis based on deep learning according to claim 4, characterized in that: In step S4, the spectrum trend prediction is used to predict and construct the change trend of the spectrum data. Specifically, based on the spectrum analysis optimization data and the abnormal interference detection data, a time series integration network combined with one-dimensional convolution feature extraction is used to perform spectrum trend prediction to obtain spectrum trend prediction data, which specifically includes the following steps: Step S41: Convolution feature extraction, specifically constructing a one-dimensional convolution feature extraction layer, extracting local trend features from the spectrum analysis optimization data and abnormal interference detection data, and constructing average pooling to perform feature dimensionality reduction to obtain convolution feature data; Step S42: constructing a bidirectional long short-term memory subnet, specifically constructing a standard bidirectional long short-term memory subnet, extracting long-term trend features based on the convolution feature data through a bidirectional structure, and obtaining bidirectional temporal network feature data; Step S43: constructing a gated recurrent subnet, specifically constructing a standard gated recurrent unit including a reset gate and an update gate as a gated recurrent subnet, and extracting short-term change trend features according to the convolution feature data through the gated recurrent subnet to obtain short-term temporal network feature data; Step S44: time series network integration, specifically, integrating the bidirectional time series network feature data and the short-term time series network feature data through feature splicing to obtain a time series integrated network feature output, and performing trend prediction according to the time series integrated network feature output to obtain trend prediction output data; Step S45: Spectrum trend prediction model training, specifically, performing spectrum trend prediction model training through the convolution feature extraction, the bidirectional long short-term memory subnet, the gated recurrent subnet and the timing network integration to obtain the spectrum trend prediction model Model SP ; Step S46: Spectrum trend prediction, specifically using the spectrum trend prediction model Model SP , performing spectrum trend prediction based on the spectrum analysis optimization data and abnormal interference detection data to obtain spectrum trend prediction data; The spectrum trend prediction data specifically includes spectrum signal time period prediction values, spectrum signal trend characteristics and abnormal trend identifiers.

6. The method for wide bandwidth frequency spectrum analysis based on deep learning according to claim 5, characterized in that: In step S5, the spectrum scheduling management is used to perform spectrum dynamic optimization and spectrum resource allocation optimization in combination with the results of abnormal interference detection and spectrum trend prediction. Specifically, based on the spectrum analysis optimization data, the abnormal interference detection data and the spectrum trend prediction data, a reinforcement learning method enhanced by a proximal strategy optimization algorithm is used to perform spectrum scheduling management to obtain wide-bandwidth frequency analysis scheduling reference data, which specifically includes the following steps: Step S51: reinforcement learning environment construction, specifically, constructing a reinforcement learning environment through reinforcement learning modeling, and performing reinforcement learning training according to the reinforcement learning environment, wherein the reinforcement learning modeling specifically includes state parameter construction, action parameter construction and reward function construction; The state parameters specifically include spectrum signal occupancy parameters, historical spectrum trend parameters, abnormal fault interference parameters and predicted trend parameters; The action parameters specifically include bandwidth allocation adjustment actions, spectrum resource time slot scheduling actions, and high-load area spectrum usage planning actions; The reward function is specifically calculated by weighted combination of resource utilization maximization target, interference minimization target and system delay minimization target; Step S52: constructing a proximal policy gradient optimization algorithm, specifically by limiting the policy update amplitude and constructing a clipping objective function to construct a proximal policy gradient optimization algorithm; Step S53: Proximal policy reinforcement learning enhancement, specifically, by applying the proximal policy gradient optimization algorithm, according to the reinforcement learning environment, performing reinforcement learning training enhanced by the proximal policy optimization algorithm until the training converges, and performing reinforcement learning enhancement; Step S54: spectrum scheduling management model training, specifically, training the spectrum scheduling management model through the reinforcement learning environment construction, the proximal policy gradient optimization algorithm and the proximal policy reinforcement learning enhancement, to obtain the spectrum scheduling management model Model SM ; Step S55: spectrum scheduling management, specifically using the spectrum scheduling management model Model SM , spectrum scheduling management is performed based on the spectrum analysis optimization data, the abnormal interference detection data and the spectrum trend prediction data to obtain wide-bandwidth frequency analysis scheduling reference data.

7. The method for wide bandwidth frequency spectrum analysis based on deep learning according to claim 6, characterized in that: In step S5, the wide bandwidth frequency analysis scheduling reference data specifically includes spectrum resource allocation scheme, spectrum utilization, abnormal area interference information reference, prediction trend optimization result and dynamic scheduling optimization result.

8. A wide-bandwidth frequency spectrum analysis system based on deep learning, used to implement a wide-bandwidth frequency spectrum analysis method based on deep learning as described in any one of claims 1 to 7, characterized in that: It includes a spectrum signal acquisition module, a spectrum preprocessing module, an abnormal interference detection module, a spectrum trend prediction module and a spectrum scheduling management module.

9. A wide bandwidth frequency spectrum analysis system based on deep learning according to claim 8, characterized in that: The spectrum signal acquisition module is used for spectrum signal acquisition, obtains a wide-frequency spectrum analysis original data set through spectrum signal acquisition, and sends the wide-frequency spectrum analysis original data set to the spectrum preprocessing module; The spectrum preprocessing module is used for spectrum preprocessing, obtains spectrum analysis optimization data through spectrum preprocessing, and sends the spectrum analysis optimization data to the abnormal interference detection module, the spectrum trend prediction module and the spectrum scheduling management module; The abnormal interference detection module is used for abnormal interference detection, obtains abnormal interference detection data through abnormal interference detection, and sends the abnormal interference detection data to the spectrum trend prediction module; The spectrum trend prediction module is used for spectrum trend prediction, obtains spectrum trend prediction data through spectrum trend prediction, and sends the spectrum trend prediction data to the spectrum scheduling management module; The spectrum scheduling management module is used for spectrum scheduling management, and obtains wide bandwidth frequency analysis scheduling reference data through spectrum scheduling management.

Images