CN114254713A - Classification system and method based on time-frequency transformation and dynamic mode decomposition - Google Patents

Classification system and method based on time-frequency transformation and dynamic mode decomposition Download PDF

Info

Publication number
CN114254713A
CN114254713A CN202210187798.XA CN202210187798A CN114254713A CN 114254713 A CN114254713 A CN 114254713A CN 202210187798 A CN202210187798 A CN 202210187798A CN 114254713 A CN114254713 A CN 114254713A
Authority
CN
China
Prior art keywords
data
classification
module
time
matrix
Prior art date
Legal status (The legal status is an assumption and is not a legal conclusion. Google has not performed a legal analysis and makes no representation as to the accuracy of the status listed.)
Granted
Application number
CN202210187798.XA
Other languages
Chinese (zh)
Other versions
CN114254713B (en
Inventor
冯立辉
杨景宏
陈子健
卢继华
Current Assignee (The listed assignees may be inaccurate. Google has not performed a legal analysis and makes no representation or warranty as to the accuracy of the list.)
Beijing Institute of Technology BIT
Original Assignee
Beijing Institute of Technology BIT
Priority date (The priority date is an assumption and is not a legal conclusion. Google has not performed a legal analysis and makes no representation as to the accuracy of the date listed.)
Filing date
Publication date
Application filed by Beijing Institute of Technology BIT filed Critical Beijing Institute of Technology BIT
Priority to CN202210187798.XA priority Critical patent/CN114254713B/en
Publication of CN114254713A publication Critical patent/CN114254713A/en
Application granted granted Critical
Publication of CN114254713B publication Critical patent/CN114254713B/en
Active legal-status Critical Current
Anticipated expiration legal-status Critical

Links

Images

Classifications

    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06F—ELECTRIC DIGITAL DATA PROCESSING
    • G06F18/00—Pattern recognition
    • G06F18/20—Analysing
    • G06F18/24—Classification techniques
    • G06F18/241—Classification techniques relating to the classification model, e.g. parametric or non-parametric approaches
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06F—ELECTRIC DIGITAL DATA PROCESSING
    • G06F18/00—Pattern recognition
    • G06F18/20—Analysing
    • G06F18/21—Design or setup of recognition systems or techniques; Extraction of features in feature space; Blind source separation
    • G06F18/213—Feature extraction, e.g. by transforming the feature space; Summarisation; Mappings, e.g. subspace methods
    • G06F18/2135—Feature extraction, e.g. by transforming the feature space; Summarisation; Mappings, e.g. subspace methods based on approximation criteria, e.g. principal component analysis
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06F—ELECTRIC DIGITAL DATA PROCESSING
    • G06F2218/00—Aspects of pattern recognition specially adapted for signal processing
    • G06F2218/12—Classification; Matching

Landscapes

  • Engineering & Computer Science (AREA)
  • Data Mining & Analysis (AREA)
  • Theoretical Computer Science (AREA)
  • Computer Vision & Pattern Recognition (AREA)
  • Bioinformatics & Computational Biology (AREA)
  • Bioinformatics & Cheminformatics (AREA)
  • Artificial Intelligence (AREA)
  • Evolutionary Biology (AREA)
  • Evolutionary Computation (AREA)
  • Physics & Mathematics (AREA)
  • General Engineering & Computer Science (AREA)
  • General Physics & Mathematics (AREA)
  • Life Sciences & Earth Sciences (AREA)
  • Complex Calculations (AREA)

Abstract

本发明涉及一种基于时频变换与动态模式分解的分类系统及方法，属于数据分类、图像分类及信号识别技术领域。所述系统,包括预处理模块、时频分析模块、动态模式分解模块、能量特征提取模块、各阶矩求解模块以及分类模块；所述方法，包括：将采集的数据进行预处理得到有效信号；将有效信号先进行时频分析再动态模式分解，得到多个特征值与动态模态并排序得到从大到小的能量值、对应特征值与动态模态，组合形成能量特征矩阵；提取能量特征矩阵的前S个元素并丢弃后续数据，更新能量特征矩阵，再求该能量特征矩阵的二范数得到变换后的特征矩阵；基于变换后的特征矩阵作为待分类数据进行阈值分类并输出分类结果。所述方法能实现了较高的分类准确率。

Figure 202210187798

The invention relates to a classification system and method based on time-frequency transformation and dynamic pattern decomposition, belonging to the technical fields of data classification, image classification and signal recognition. The system includes a preprocessing module, a time-frequency analysis module, a dynamic mode decomposition module, an energy feature extraction module, a moment solving module and a classification module; the method includes: preprocessing the collected data to obtain valid signals; Perform time-frequency analysis on the effective signal and then decompose the dynamic mode to obtain multiple eigenvalues and dynamic modes and sort them to obtain energy values, corresponding eigenvalues and dynamic modes from large to small, and combine them to form an energy feature matrix; extract energy features The first S elements of the matrix and the subsequent data are discarded, the energy feature matrix is updated, and the second norm of the energy feature matrix is obtained to obtain the transformed feature matrix; based on the transformed feature matrix as the data to be classified, threshold classification is performed and the classification result is output. . The method can achieve higher classification accuracy.

Figure 202210187798

Description

Classification system and method based on time-frequency transformation and dynamic mode decomposition
Technical Field
The invention relates to a classification system and a classification method based on time-frequency transformation and dynamic mode decomposition, and belongs to the technical field of data classification, image classification and signal identification.
Background
Data classification is the grouping together of data having some common attribute or characteristic, and the data are distinguished by the attribute or characteristic of its category. In other words, the information with the same content or the same property is classified into a category, and the different information is distinguished, and then the relationship between the sets is determined, so as to form an organized classification system. For signal processing and classification, for example: the data classification is also widely applied in the related fields of Internet of things identification, space spectrum resource management and control, radio frequency security and the like, and the signal classification and identification have great application value and practical significance.
In practical applications, signals received after the signals sent by the radio frequency signal source pass through a complex channel are often unstable and statistics of the signals are time-varying. In order to classify the radio frequency signals, the radio frequency signal features are often required to be extracted. The methods for extracting features generally include fourier analysis, short-time fourier transform, wavelet transform, power spectrum, cepstrum, and the like. The Fourier analysis method cannot perform local characterization on non-stationary radio frequency signal characteristics, and can only be used for knowing the global characteristics of the signal in a time frequency domain. Although short-time fourier transform can describe frequency information in a certain local time period, the stability of a signal in a short time is difficult to guarantee as the signal time increases. The wavelet transform can characterize local characteristics of signals in a time-frequency domain, but the wavelet transform belongs to a linear time-frequency analysis method as the traditional Fourier analysis method, and cannot describe the instantaneous power spectral density of the signals. Therefore, neither time-frequency analysis nor spectrogram analysis can comprehensively and intuitively reflect all the characteristics of the radio-frequency signal in the time domain and the frequency domain, and the problem of time-frequency resolution compromise becomes the bottleneck of time-frequency characteristic extraction. Furthermore, the frequency information of the signal varies at different times, and it is difficult to find a suitable analysis window to accommodate the entire time period of the signal.
Besides linear time-frequency analysis, a quadratic time-frequency analysis method is also provided. The existing research results show that the time frequency resolution obtained by using quadratic time frequency analysis to process the non-stationary signals is higher, and the quadratic time frequency analysis method is an effective method for processing the non-stationary signals. Such as the Cohen-like time-frequency distribution and the Wigner-Ville distribution (WVD). The Cohen-like time-frequency distribution is proposed by Cohen in the middle of the 60's of the 20 th century, and Cohen finds that many time-frequency distributions are the deformation of WVD and can be uniformly expressed. The Cohen-class time-frequency distribution is two-dimensional Fourier transform of a kernel-weighted fuzzy function, and is also called generalized bilinear time-frequency distribution, WVD has good time-frequency aggregation characteristics, but for multi-component signals, the W-V distribution can generate cross terms, generate false signals and cause interference on the time-frequency characteristics of radio frequency signals, thereby bringing serious influence to signal classification.
The Wigner-Hough transform (WHT) is used for carrying out Hough transform on a signal time-frequency diagram on the basis of WVD, and can inhibit cross term interference. The Hough transform transforms a given curve in the original image space into one point in the parameter space, using the duality of points and lines. The detection problem of a given curve in an original image is converted into a peak value in a parameter space, so that the overall detection characteristic is converted into a local detection characteristic. However, the Hough transform again results in a matrix with too high dimensions after the transform. Dynamic Mode Decomposition (DMD) is physically more significant than conventional dimensionality reduction methods such as Principal Component Analysis (PCA). Therefore, the energy of the matrix can be calculated by the DMD, and the characteristics of the matrix are reflected from the energy perspective. The dynamic information extracted by the DMD is called a dynamic modality, and is used for describing a potential physical mechanism in a dynamic system, or projecting a complex system onto a dynamic system with a very small degree of freedom, so as to extract the dynamic information from the dynamic system. The DMD avoids extracting dynamic information by a model-based method, adopts a data-based process, and is also suitable for experimental and numerical dynamic data. The DMD can reduce the dimensionality of data, and the processed modal matrix can also be used as a characteristic matrix for reflecting matrix information.
Disclosure of Invention
The invention aims to provide a classification system and a classification method based on time-frequency transformation and dynamic mode decomposition aiming at the defects of difficult classification of complex data and low radio frequency signal identification rate transmitted through a complex channel.
In order to achieve the purpose, the following technical scheme is adopted:
the classification system based on time-frequency transformation and dynamic mode decomposition comprises a preprocessing module, a time-frequency analysis module, a dynamic mode decomposition module, an energy feature extraction module, each order moment solving module and a classification module;
the preprocessing module comprises a data truncation unit, a denoising unit and an enhancement unit; the time-frequency analysis module comprises a time-frequency distribution solving unit, a time-frequency relation graph output unit and a Hough transformation unit;
the preprocessing module cuts, denoises and enhances input data and outputs preprocessed signals; the preprocessed signals are output to an H matrix through a time-frequency analysis module, and the H matrix is subjected to dynamic mode decomposition by a dynamic mode decomposition module to obtain a plurality of characteristic values and dynamic modes; the energy characteristic extraction module extracts the first S energy values and the corresponding characteristic values thereof and the dynamic modes according to the sequence from big to small according to the energy value difference between different modes, rearranges the S energy values and the corresponding characteristic values to obtain a transformed characteristic matrix, and each order moment solving module calculates each order moment of the characteristic matrix; the classification module classifies according to each order moment;
the preprocessing module is connected with the video analysis module, the video analysis module is connected with the dynamic mode decomposition module, the dynamic mode decomposition module is connected with the energy feature extraction module, the energy feature extraction module is connected with each order moment solving module, and each order moment solving module is connected with the classification module;
the data truncation unit in the preprocessing module is connected with the denoising unit, and the denoising unit is connected with the enhancement unit;
the classification method based on time-frequency transformation and dynamic mode decomposition comprises the following steps:
step 1, preprocessing acquired data to obtain effective signals;
the preprocessing comprises signal truncation, denoising and enhancement operations;
step 2, performing time-frequency analysis on the effective signals obtained in the step 2 to obtain an H matrix;
the preferable time-frequency analysis is Wigner-Hough transformation, which comprises the steps of solving time-frequency distribution, outputting a time-frequency relation graph and Hough transformation to obtain an H matrix, and the method specifically comprises the following substeps:
step 2.1, solving Wigner-Ville distribution of the effective signals, namely solving Fourier transformation of instantaneous correlation functions of the effective signals, solving time-frequency relations corresponding to the signals, and drawing corresponding time-frequency graphs;
step 2.2, carrying out Hough transformation on the time-frequency diagram to obtain an H matrix reflecting W-H transformation characteristics of the signal;
step 3, performing dynamic mode decomposition on the H matrix obtained in the step 3 to obtain a plurality of characteristic values and dynamic modes;
step 4, arranging the energy values of the plurality of dynamic modes in a descending order to obtain ascending energy values, corresponding characteristic values and rearranged dynamic modes, and combining to form an energy characteristic matrix;
step 5, extracting the first S elements of the energy characteristic matrix, discarding subsequent data, updating the energy characteristic matrix, and solving the two-norm of the energy characteristic matrix to obtain a transformed characteristic matrix;
s is greater than or equal to 2;
step 6, performing threshold classification based on the transformed feature matrix as data to be classified, specifically comprising the following substeps:
step 6.1, whether the classification is finished or not is judged firstly when the classification starts, if only 1 type of data exists in the data to be classified at the moment, the classification is finished, and if not, the step 6.2 is skipped;
step 6.2, comparing various types of data to be classified, if the mean value of the energy characteristic matrix of certain type of data is always larger than the signals of other types, separating the data, removing the data from the data to be classified, skipping to step 6.1 to judge whether the classification is finished, and skipping to step 6.3 if not;
6.3, comparing various energy characteristic matrixes, if the energy characteristic variance of certain data is always larger than that of other data, classifying the data, removing the data from the data to be classified, skipping to the step 6.1 to judge whether the classification is finished, and if not, skipping to the step 6.4;
step 6.4, comparing the third moment of the energy characteristic matrixes of various types, if the third moment of the energy characteristic matrix of certain type of data is always larger than that of other types of data, classifying the data, removing the data from the data to be classified, skipping to step 6.1 to judge whether the classification is finished, and if not, skipping to step 6.5;
step 6.5, continuously comparing the higher-order moments of the various energy characteristic matrixes until one type of data is removed from the data to be classified;
6.6, sequentially circulating the steps 6.1 to 6.5 until all data are separated;
and 7, outputting a classification result.
Advantageous effects
Compared with the existing classification method, the classification system and method based on time-frequency transformation and dynamic mode decomposition have the following beneficial effects:
1. the classification method adopts W-H transformation, namely Hough transformation is carried out on the basis of Wigner-Ville distribution; the W-H conversion can inhibit noise in signals and reduce interference of cross terms generated in Wigner-Ville distribution;
2. the classification method adopts the DMD, which is a dynamic information extraction mode focusing on data, so that the defect of poor algorithm applicability caused by model-based processing of some extraction algorithms is overcome, and the method is more suitable for processing different types of data; the DMD can conveniently extract time domain change corresponding to the space domain component, the change is called as a dynamic mode and is related to dynamic oscillation frequency attenuation and growth rate;
3. the classification method aims at image data and wireless signals passing through complex channels, and achieves high classification accuracy through preprocessing, time-frequency transformation, DMD decomposition and moment calculation of an energy matrix.
Drawings
FIG. 1 is a diagram of the composition and connection relationship of a classification system based on time-frequency transformation and dynamic mode decomposition according to the present invention;
FIG. 2 is a flowchart illustrating the specific implementation of time-frequency transformation in a classification method based on time-frequency transformation and dynamic mode decomposition according to the present invention;
FIG. 3 is a flowchart of a classification method based on time-frequency transformation and dynamic mode decomposition according to the present invention;
FIG. 4 is a block diagram of a signal processing system for a classification system based on time-frequency transformation and dynamic mode decomposition according to an embodiment of the present invention;
FIG. 5 shows the first moment and second moment results of Iris data set classification when the classification method based on time-frequency transformation and dynamic mode decomposition is implemented.
Detailed Description
The classification system and method based on time-frequency transformation and dynamic mode decomposition according to the present invention will be further described and illustrated in detail with reference to the accompanying drawings and embodiments.
Example 1
This embodiment illustrates a process of classifying an Iris data set in a UCI database by using the method of the present invention. The total amount of data in the Iris dataset is 150; the data dimension is 1x4, and the data category is 3 types. Attributes of data in the Iris dataset include: 1. sepal length in cm, 2 sepal width in cm, 3 petal length in cm, 4 petal width in cm, 5 Class Iris Setosa, Iris Versicolour, and Iris Virginica.
FIG. 1 is a diagram illustrating the components and connections of the classification system based on time-frequency transformation and dynamic mode decomposition; as can be seen from fig. 1, the classification system includes a preprocessing module, a time-frequency analysis module, a dynamic mode decomposition module, an energy feature extraction module, an order moment solving module, and a classification module; and the preprocessing module is connected with the video analysis module, the video analysis module is connected with the dynamic mode decomposition module, the dynamic mode decomposition module is connected with the energy feature extraction module, the energy feature extraction module is connected with each order moment solving module, and each order moment solving module is connected with the classification module.
FIG. 2 is a flow chart of the time-frequency transformation implementation of the classification method based on time-frequency transformation and dynamic mode decomposition; the time-frequency analysis is specifically W-H transform in this embodiment. As can be seen from fig. 2, first, the Wigner-Ville distribution of valid data is obtained, that is, the fourier transform of the instantaneous correlation function of the valid data is obtained, a corresponding time-frequency relationship is obtained, a corresponding time-frequency graph is drawn, then Hough transform is performed on the time-frequency graph, and an H matrix reflecting the W-H transform characteristics of the signal is obtained, because the Hough transform is thought to find the peak value problem of a parameter space, that is, the edge of the image is detected and enhanced, low-frequency noise is generally widely and uniformly distributed in the image, and therefore is not easily detected, the purpose of suppressing noise is achieved, and meanwhile, an interference cross term generated by the Wigner-Ville distribution is also difficult to be detected as the peak value after the Hough transform, so that the purpose of suppressing the interference cross term is also achieved; then decomposing the W-H transformation characteristic matrix H through the DMD, and decomposing the W-H transformation characteristic matrix H into a plurality of characteristic values and dynamic modes; and then rearranging the energy values and the corresponding characteristic values thereof with the dynamic modes according to the difference of the energy values among different modes from large to small to obtain a transformed characteristic matrix.
Fig. 4 and fig. 3 are schematic diagrams of the classification system and the classification method based on time-frequency transformation and dynamic mode decomposition, respectively. As can be seen from fig. 3, the method adopts threshold classification, specifically:
s10, preprocessing the three types of data to be classified in the Iris data set to obtain effective data;
the preprocessing comprises truncation, denoising and enhancement in specific implementation;
s20, performing time-frequency analysis on the effective data obtained in the step S10, wherein Wigner-Hough transformation is obtained to obtain an H matrix;
s20 specifically includes the following substeps:
s21, solving Wigner-Ville distribution of effective data, namely solving Fourier transform of an instantaneous correlation function of the effective signal, solving a time-frequency relation corresponding to the data, and drawing a corresponding time-frequency graph;
s22, carrying out Hough transformation on the time-frequency diagram to obtain an H matrix reflecting W-H transformation characteristics of the signal;
s30, performing DMD decomposition on the H matrix which is obtained in S20 and reflects the W-H transformation characteristics of the signals to obtain a plurality of characteristic values and dynamic modes, and specifically comprising the following substeps:
s31 matrix H obtained based on Hough transformation1 NBy H1 N-1=[H1H2…H N-1 ]And H2 N=[H2…H N-1 H N ]Construction matrix H1 N-1And H2 NAnd H is1 N=[ H1H2…H N-1 H N ]；
S32 according to the matrix H1 N-1And H2 NBy A = H2 NH1 N-1+Constructing a relationship matrix A, H1 N-1+Represents a timing matrix H1 N-1The generalized inverse matrix of (2);
s33, general formula H1 N-1=UΣV*To H1 N-1Performing singular value decomposition;
wherein, the decomposed U and V are unitary matrixes, the matrix sigma is an orthogonal matrix of singular value decomposition, the matrix sigma is a diagonal matrix, and V is*Is the adjoint of V;
s34 according to the matrix H2 NAnd the relation matrix A is formed by B = U*AU constructs an approximate relation matrix B;
wherein, U*Is the companion matrix of U;
s35, performing characteristic decomposition on the approximate relation matrix B to obtain a characteristic value lambda, a characteristic vector omega and a matrix Q consisting of the characteristic vectors;
s36 according to the matrix H2 NMatrix U, matrix Σ, and matrix Q pass Φ = H2 NUΣ+Q, calculating the mode phi of dynamic mode decomposition; wherein, sigma+Is the generalized inverse of the matrix Σ;
thereby decomposing the matrix H into a plurality of eigenvalues and dynamic modes;
s40 rearranging the energy values and the corresponding characteristic values and the dynamic modes from large to small according to the energy values among different modes to form an energy characteristic matrix; then extracting the first three energy values of the energy characteristic matrix, discarding subsequent data, taking the logarithm of the energy values, and then solving the two norms to obtain a transformed characteristic matrix;
s50, threshold classification is carried out based on the characteristic matrix, and the threshold classification method specifically comprises the following substeps:
s51, firstly judging whether the classification is finished when the classification is started, finishing the classification if only 1 type of data exists in the data to be classified at the moment, and jumping to S52 if not;
s52, comparing various energy characteristic matrixes, if the energy characteristic mean value of a certain signal is always larger than that of other signals, separating the signals, removing the signals from the signals to be classified, jumping to S51 to judge whether the classification is finished, and otherwise jumping to the step 8.3;
s53, comparing various energy characteristic matrixes, if the energy characteristic variance of a certain signal is always larger than that of other signals, separating the signals, removing the signals from the signals to be classified, jumping to S51 to judge whether the classification is finished, and otherwise, jumping to S54;
s54, comparing the third moment of the energy characteristic matrixes of various types, if the third moment of the energy characteristic matrix of a certain type of signals is always larger than the signals of other types, separating the signals, removing the signals from the signals to be classified, jumping to S51 to judge whether the classification is finished, and otherwise, jumping to S55;
s55, continuously comparing higher-order moments of various energy characteristic matrixes until one signal is removed from the signals to be classified;
s56, sequentially circulating S51 to S55 until all signals are separated;
and repeating the steps S10 to S56 for multiple times, and counting the classification accuracy.
In specific implementation, WHT is performed on each 1 × 4 data, and an energy average is obtained by DMD. The total number of each kind of data is 50, and the energy of 25 data is taken as the mean value and the variance, so that the results of two experiments are obtained. Selecting 75 data as a first experiment result, and taking the first 25 data of each type; the second experiment result selects 75 data, and takes the last 25 data of each type.
The first moment and second moment results of the two experimental energy matrices are shown in tables 1 and 2, respectively.
TABLE 1 mean (first moment) results of two experiments
UCI database Iris data Mean First experiment Second experiment
Iris Setosa 10.3153 10.11495
Iris Versicolour 10.40695 10.30931
Iris Virginica 12.93793 13.12349
TABLE 2 mean variance (second moment) results of two experiments
UCI database Iris data Var First experiment Second experiment
Iris Setosa 10.3153 10.11495
Iris Versicolour 10.40695 10.30931
Iris Virginica 12.93793 13.12349
FIG. 5 shows the first moment and second moment results of Iris data set classification when the classification method based on time-frequency transformation and dynamic mode decomposition is implemented. 5a, the two experiments can be ensured, and the Mean of the Iris Virginica energy matrix is the highest; further based on the second moment value obtained in 5b, Iris Setosa and Iris Versicolor can be distinguished in two experiments.
While the foregoing is directed to the preferred embodiment of the present invention, it is not intended that the invention be limited to the embodiment and the drawings disclosed herein. Equivalents and modifications may be made without departing from the spirit of the disclosure, which is to be considered as within the scope of the invention.

Claims (7)

1.一种基于时频变换与动态模式分解的分类系统，其特征在于：包括预处理模块、时频分析模块、动态模式分解模块、能量特征提取模块、各阶矩求解模块以及分类模块；1. a classification system based on time-frequency transformation and dynamic pattern decomposition, is characterized in that: comprise preprocessing module, time-frequency analysis module, dynamic pattern decomposition module, energy feature extraction module, each order moment solving module and classification module; 所述预处理模块包括数据截断单元、去噪单元以及增强单元；时频分析模块包括求解时频分布单元、时频关系图输出单元以及Hough变换单元；The preprocessing module includes a data truncation unit, a denoising unit and an enhancement unit; the time-frequency analysis module includes a time-frequency distribution unit for solving, a time-frequency relationship graph output unit and a Hough transform unit; 预处理模块与视频分析模块相连，视频分析模块与动态模式分解模块相连，动态模式分解模块与能量特征提取模块相连，能量特征提取模块与各阶矩求解模块相连，各阶矩求解模块与分类模块相连；预处理模块中的数据截断单元与去噪单元相连，去噪单元与增强单元相连；预处理模块将输入数据进行截断、去噪以及增强，输出预处理后信号；预处理后信号经时频分析模块输出H矩阵，动态模式分解模块将H矩阵进行动态模式分解得到多个特征值与动态模态；能量特征提取模块将不同模态之间能量值的差异按照从大到小的顺序提取前S个能量值与其对应的特征值与动态模态重新排列得到变换后的特征矩阵，各阶矩求解模块计算特征矩阵的各阶矩；分类模块按照各阶矩进行分类。The preprocessing module is connected with the video analysis module, the video analysis module is connected with the dynamic mode decomposition module, the dynamic mode decomposition module is connected with the energy feature extraction module, the energy feature extraction module is connected with each order moment solving module, and each order moment solving module is connected with the classification module The data truncation unit in the preprocessing module is connected with the denoising unit, and the denoising unit is connected with the enhancement unit; the preprocessing module truncates, denoises and enhances the input data, and outputs the preprocessed signal; The frequency analysis module outputs the H matrix, and the dynamic mode decomposition module decomposes the dynamic mode of the H matrix to obtain multiple eigenvalues and dynamic modes; the energy feature extraction module extracts the difference in energy values between different modes in descending order. The first S energy values and their corresponding eigenvalues and dynamic modes are rearranged to obtain the transformed eigenmatrix, and each order moment solving module calculates each order moment of the eigenmatrix; the classification module classifies according to each order moment. 2.一种基于时频变换与动态模式分解的分类方法，其特征在于：包括如下步骤：2. a classification method based on time-frequency transformation and dynamic pattern decomposition, is characterized in that: comprise the steps: 步骤1、将采集的数据进行预处理得到有效信号；Step 1. Preprocess the collected data to obtain a valid signal; 步骤2、将步骤2得到的有效信号进行时频分析，得到H矩阵；Step 2. Perform time-frequency analysis on the effective signal obtained in step 2 to obtain an H matrix; 步骤3、将步骤3得到的H矩阵进行动态模式分解，得到多个特征值与动态模态；Step 3. Perform dynamic mode decomposition on the H matrix obtained in step 3 to obtain multiple eigenvalues and dynamic modes; 步骤4、将多个动态模态的能量值按照从大到小的顺序排列，得到从大到小的能量值、对应的特征值与重新排列的动态模态，组合形成能量特征矩阵；Step 4. Arrange the energy values of multiple dynamic modes in descending order, and obtain the energy values, corresponding eigenvalues and rearranged dynamic modes in descending order, and combine them to form an energy characteristic matrix; 步骤5、提取能量特征矩阵的前S个元素并丢弃后续数据，更新能量特征矩阵，再求该能量特征矩阵的二范数得到变换后的特征矩阵；Step 5, extracting the first S elements of the energy characteristic matrix and discarding the subsequent data, updating the energy characteristic matrix, and then finding the second norm of the energy characteristic matrix to obtain the transformed characteristic matrix; 步骤6、基于上述变换后的特征矩阵作为待分类数据进行阈值分类；Step 6, performing threshold classification based on the transformed feature matrix as the data to be classified; 步骤7、输出分类结果。Step 7, output the classification result. 3.根据权利要求2所述的分类方法，其特征在于：所述预处理，包括信号截断、去噪及增强操作。3 . The classification method according to claim 2 , wherein the preprocessing includes signal truncation, denoising and enhancement operations. 4 . 4.根据权利要求2所述的分类方法，其特征在于：时频分析为Wigner-Hough变换，包括求解时频分布、输出时频关系图以及Hough变换，得到H矩阵。4 . The classification method according to claim 2 , wherein the time-frequency analysis is Wigner-Hough transform, which includes solving the time-frequency distribution, outputting a time-frequency relationship diagram and Hough transform to obtain an H matrix. 5 . 5.根据权利要求2所述的分类方法，其特征在于：步骤2、具体包括如下子步骤：5. classification method according to claim 2 is characterized in that: step 2, specifically comprises following sub-step: 步骤2.1、求有效信号的Wigner-Ville分布，即求有效信号瞬时相关函数的傅里叶变换，求解信号对应的时频关系，并画出相应的时频图；Step 2.1, find the Wigner-Ville distribution of the effective signal, that is, find the Fourier transform of the instantaneous correlation function of the effective signal, solve the time-frequency relationship corresponding to the signal, and draw the corresponding time-frequency diagram; 步骤2.2、将上述时频图做Hough变换，得到反映信号W-H变换特征的H矩阵。Step 2.2: Hough transform the above time-frequency diagram to obtain an H matrix reflecting the W-H transform characteristics of the signal. 6.根据权利要求2所述的分类方法，其特征在于：所述S大于等于2。6 . The classification method according to claim 2 , wherein the S is greater than or equal to 2. 7 . 7.根据权利要求2所述的分类方法，其特征在于：步骤6、具体包括如下子步骤：7. classification method according to claim 2 is characterized in that: step 6, specifically comprises the following substeps: 步骤6.1、分类开始时首先判断分类是否完成，如果此时待分类数据中只有1类数据则分类结束，否则跳至步骤6.2；Step 6.1. At the beginning of the classification, first determine whether the classification is completed. If there is only one type of data in the data to be classified at this time, the classification ends, otherwise skip to step 6.2; 步骤6.2、比较各类待分类数据，如果存在某一类数据的能量特征矩阵的均值总是大于其他类别的信号，那么此类数据被分出，同时将此类数据从待分类数据中去除并跳至步骤6.1判断分类是否结束，否则跳至步骤6.3；Step 6.2. Compare various types of data to be classified. If the mean value of the energy characteristic matrix of a certain type of data is always greater than that of other types of signals, then this type of data is separated, and at the same time, this type of data is removed from the data to be classified and added. Go to step 6.1 to judge whether the classification is over, otherwise go to step 6.3; 步骤6.3、比较各类能量特征矩阵，如果存在某一类数据的能量特征方差总是大于其他类别数据，那么此类数据被分出，同时将此类数据从待分类数据中去除并跳至步骤6.1判断分类是否结束，否则跳至步骤6.4；Step 6.3. Compare various energy characteristic matrices. If there is a certain type of data whose energy characteristic variance is always greater than that of other types of data, then this type of data is separated, and at the same time, this type of data is removed from the data to be classified and skips to the step 6.1 Determine whether the classification is over, otherwise skip to step 6.4; 步骤6.4、比较各类能量特征矩阵的三阶矩，如果存在某一类数据的能量特征矩阵的三阶矩总是大于其他类别数据，那么此类数据被分出，同时将此类数据从待分类数据中去除并跳至步骤6.1判断分类是否结束，否则跳至步骤6.5；Step 6.4. Compare the third-order moments of various energy characteristic matrices. If the third-order moment of the energy characteristic matrix of a certain type of data is always greater than that of other types of data, then this type of data is separated, and this type of data is removed from the data to be Remove it from the classified data and skip to step 6.1 to determine whether the classification is over, otherwise skip to step 6.5; 步骤6.5、继续比较各类能量特征矩阵的更高阶矩，直至有一类数据从待分类数据中去除为止；Step 6.5. Continue to compare the higher order moments of various energy characteristic matrices until one type of data is removed from the data to be classified; 步骤6.6、依次循环步骤6.1到步骤6.5直至所有数据均被分出。Step 6.6. Repeat step 6.1 to step 6.5 in turn until all data are separated.
CN202210187798.XA 2022-02-28 2022-02-28 A classification system and method based on time-frequency transform and dynamic pattern decomposition Active CN114254713B (en)

Priority Applications (1)

Application Number Priority Date Filing Date Title
CN202210187798.XA CN114254713B (en) 2022-02-28 2022-02-28 A classification system and method based on time-frequency transform and dynamic pattern decomposition

Applications Claiming Priority (1)

Application Number Priority Date Filing Date Title
CN202210187798.XA CN114254713B (en) 2022-02-28 2022-02-28 A classification system and method based on time-frequency transform and dynamic pattern decomposition

Publications (2)

Publication Number Publication Date
CN114254713A true CN114254713A (en) 2022-03-29
CN114254713B CN114254713B (en) 2022-05-27

Family

ID=80800068

Family Applications (1)

Application Number Title Priority Date Filing Date
CN202210187798.XA Active CN114254713B (en) 2022-02-28 2022-02-28 A classification system and method based on time-frequency transform and dynamic pattern decomposition

Country Status (1)

Country Link
CN (1) CN114254713B (en)

Cited By (1)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN117116289A (en) * 2023-10-24 2023-11-24 吉林大学 Ward medical intercom management system and method thereof

Citations (2)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN110515096A (en) * 2019-08-20 2019-11-29 东南大学 Satellite navigation interference signal identification device and method based on convolutional neural network
US20220060365A1 (en) * 2019-02-28 2022-02-24 Espressif Systems (Shanghai) Co., Ltd. Mimo-ofdm wireless signal detection method and system capable of channel matrix pre-processing during detection

Patent Citations (2)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
US20220060365A1 (en) * 2019-02-28 2022-02-24 Espressif Systems (Shanghai) Co., Ltd. Mimo-ofdm wireless signal detection method and system capable of channel matrix pre-processing during detection
CN110515096A (en) * 2019-08-20 2019-11-29 东南大学 Satellite navigation interference signal identification device and method based on convolutional neural network

Non-Patent Citations (3)

* Cited by examiner, † Cited by third party
Title
DE MARCHI L等: "Warped Wigner-Hough transform for defect reflection enhancement in ultrasonic guided wave monitoring", 《MATHEMATICAL PROBLEMS IN ENGINEERING》 *
SHIJIE R等: "Application of time-frequency analysis to doppler signal", 《2009 INTERNATIONAL FORUM ON COMPUTER SCIENCE-TECHNOLOGY AND APPLICATIONS. IEEE》 *
孟小曼等: "基于Wigner-Hough变换的卫星导航接收机扫频干扰信号检测", 《合肥工业大学学报(自然科》 *

Cited By (2)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN117116289A (en) * 2023-10-24 2023-11-24 吉林大学 Ward medical intercom management system and method thereof
CN117116289B (en) * 2023-10-24 2023-12-26 吉林大学 Ward medical intercom management system and method thereof

Also Published As

Publication number Publication date
CN114254713B (en) 2022-05-27

Similar Documents

Publication Publication Date Title
US10361680B2 (en) Fuzzy entropy based noisy signal processing method and iterative singular spectrum analysis soft threshold de-noising method
Prochazka et al. Wavelet transform use for feature extraction and EEG signal segments classification
US7308378B2 (en) System and method for identifying an object
CN105205112A (en) System and method for excavating abnormal features of time series data
Abdulrahman et al. Face recognition using enhancement discrete wavelet transform based on MATLAB
Hao et al. Data amplification: A unified and competitive approach to property estimation
CN106503733B (en) A useful signal identification method based on NA-MEMD and GMM clustering
CN114897002A (en) Intrapulse Modulation Identification Method of Low Intercept Probability Radar Signal Based on LPINet
CN114254713A (en) Classification system and method based on time-frequency transformation and dynamic mode decomposition
Lu et al. Image-specific prior adaptation for denoising
WO2012050581A1 (en) Dataset compression
Liu et al. Convergence rates of a partition based Bayesian multivariate density estimation method
CN113221736A (en) Electrocardio identity recognition method and system based on multi-feature collaborative non-negative matrix factorization
CN115328661B (en) Computing power balance execution method and chip based on voice and image characteristics
Hao et al. An improved multivariate wavelet denoising method using subspace projection
Xu et al. Unraveling the veil of subspace rip through near-isometry on subspaces
CN1256704C (en) Spectral Feature Extraction and Comparison Method of Remote Sensing Image
CN118115853A (en) A broadband signal detection method based on lightweight fully convolutional neural network
Berthelot et al. Filtering-based analysis comparing the DFA with the CDFA for wide sense stationary processes
CN108256581B (en) Gabor wavelet domain copula model image classification method
Qiao et al. Complex wavelet based texture classification
Jumanov et al. Improving the quality of identification and filtering of micro-object images based on neural networks
CN117951610B (en) Communication signal identification and classification method based on characteristic data analysis
Gao et al. An improved feature-weighted method based on K-NN
CN117095208B (en) Lightweight scene classification method for photoelectric pod reconnaissance image

Legal Events

Date Code Title Description
PB01 Publication
PB01 Publication
SE01 Entry into force of request for substantive examination
SE01 Entry into force of request for substantive examination
GR01 Patent grant
GR01 Patent grant