A hierarchical decomposition-based feature extraction method

Abstract

The present application belongs to the technical field of feature extraction and signal classification, and particularly relates to a hierarchical decomposition feature extraction method. The method comprises: preprocessing a collected two-dimensional time sequence to obtain multi-channel effective data; establishing a hierarchical decomposition feature extraction structure; the hierarchical decomposition feature extraction structure comprises a receiving domain matrix, a modulation domain matrix and a pulse generation domain matrix; initializing loop parameters, feature extraction structure parameters and square matrices; initializing the receiving domain matrix, the modulation domain matrix and the pulse generation domain matrix as all-0 matrices; hierarchically decomposing the multi-channel effective data to obtain hierarchical decomposition feature extraction results; inputting a pulse signal matrix into a classifier for data classification to obtain classification results. The method can better extract signal and data features collected in a complex environment, further improving data classification accuracy; the feature extraction structure of the method is simple, does not need to extract features through machine learning training, and has low computational complexity.

Description

Technical Field

[0001] This invention belongs to the field of feature extraction and signal classification technology, and particularly relates to a feature extraction method based on hierarchical decomposition. Background Technology

[0002] Signal classification typically involves first extracting features using a generator, and then inputting these features into a classifier for classification. The quality of signal feature extraction directly impacts the accuracy of subsequent signal classification. In recent years, with the advancement of deep learning in machine learning, neural networks have been widely applied in signal classification, continuously improving the accuracy and speed of feature extraction. However, many shortcomings still require further optimization. On one hand, most neural network models have complex structures and high computational requirements; on the other hand, these models process specific datasets and require continuous training.

[0003] This invention utilizes a hierarchical decomposition method for signal feature extraction. The aim is to effectively extract features from signals and data collected in complex environments and improve the accuracy of subsequent classification. To achieve this goal, this invention proposes a hierarchical decomposition-based feature extraction method, establishing a feature extraction structure comprising a receiving domain, a modulation domain, and a pulse generation domain, and initializing the parameters and correlation matrix of this structure. Then, effective data is cyclically decomposed into ordered element factors through hierarchical decomposition to obtain sequential details from coarse to fine. In subsequent experiments, combined with existing neural networks for classification, the classification accuracy was significantly improved. Summary of the Invention

[0004] The purpose of this invention is to address the shortcomings of difficult feature extraction from signals and data collected in complex environments, as well as the lack of obvious extracted features, which leads to low classification and recognition accuracy. This invention proposes a feature extraction method based on hierarchical decomposition.

[0005] To achieve the above objectives, the present invention adopts the following technical solution:

[0006] The hierarchical decomposition-based feature extraction method includes the following steps:

[0007] S1. Preprocess the collected two-dimensional time series to obtain multi-channel effective data;

[0008] S2. Establish a hierarchical decomposition feature extraction structure;

[0009] The hierarchical decomposition feature extraction structure includes a receiver domain matrix, a modulation domain matrix, and a pulse generation domain matrix;

[0010] The receiving domain matrix includes a feed channel matrix and a link channel matrix;

[0011] The modulation domain matrix includes an internal activity term matrix;

[0012] The pulse generation domain matrix includes a pulse signal matrix and a threshold matrix;

[0013] S3. Initialize the loop parameters, feature extraction structure parameters, and the square matrix;

[0014] The loop parameters include the number of loops and the total number of loops;

[0015] The feature extraction parameters include link field radius, threshold amplitude scaling term, threshold attenuation term, link attenuation term, link amplitude scaling term, link strength, attenuation time constant, and threshold amplification factor.

[0016] The dimension of the square matrix is ​​2 * link field radius + 1;

[0017] S4. Initialize all matrices in the hierarchical decomposition feature extraction structure to be all zero matrices;

[0018] The matrices in the hierarchical decomposition feature extraction structure include the receiving domain matrix, the modulation domain matrix, and the pulse generation domain matrix;

[0019] S5. Perform hierarchical decomposition on the multi-channel effective data to obtain the hierarchical decomposition feature extraction results;

[0020] S5, specifically:

[0021] S51. Initialize the loop count to 1;

[0022] S52. Convolve the pulse signal matrix with the square matrix to obtain the convolution result;

[0023] S53. Multiply the negative exponential function value of the threshold attenuation term by the feed channel matrix, add the result of multiplying the threshold amplitude scaling term by the convolution result, and add the effective data of the multiple channels to obtain the updated feed channel matrix.

[0024] S54. Multiply the negative exponent of the link decay term by the link channel matrix, and add the result of multiplying the link magnitude scaling term by the convolution result to obtain the updated link channel matrix.

[0025] S55. Add 1 to the link strength, multiply it by the link channel matrix, and then multiply it by the feed channel matrix to obtain the internal activity term matrix;

[0026] S56. Compare the internal activity term matrix with the threshold matrix. If the element value of the internal activity term matrix is ​​greater than the element value of the threshold matrix, set the element value at the corresponding position of the pulse signal matrix to 1, otherwise set it to 0 to obtain the pulse signal matrix.

[0027] S57. Add the product of the negative exponential function value of the decay time constant and the threshold matrix to the product of the threshold amplification factor and the pulse signal matrix, and assign the sum of the two products to the threshold matrix.

[0028] S58. Determine if the number of loops is equal to the total number of loops. If not, increment the number of loops by 1 and jump to S52. If the number of loops is equal to the total number of loops, the loop ends, save the hierarchical decomposition feature extraction results of the pulse signal matrix as valid data in S56, and jump to S6.

[0029] S6. Input the pulse signal matrix into the classifier to classify the data and obtain the classification result.

[0030] The preprocessing described in S1 includes normalization, filtering, signal truncation, denoising, and enhancement.

[0031] The two dimensions in the two-dimensional time series described in S1 refer to the time and channel dimensions, respectively, and the channel dimension is the same as the number of channels in a multi-channel system.

[0032] The dimensions of the feed channel matrix, link channel matrix, internal activity term matrix, pulse signal matrix, and threshold matrix described in S2 are the same as the dimensions of the single-channel effective data.

[0033] The total number of cycles in S3 ranges from 20 to 50; the number of cycles ranges from 1 to the total number of cycles.

[0034] The value at the center of the square matrix in S3 is initialized to 0, and the values ​​at other positions are the reciprocal of the sum of the squares of the distances between that position and the center position.

[0035] The value of the link attenuation term ranges from 1 to 5;

[0036] The threshold attenuation term has a value range of 1 to 5;

[0037] The decay time constant ranges from 1 to 5;

[0038] The value of the link strength ranges from 0 to 1;

[0039] The value range of the link amplitude scaling term is from 0.1 to 0.9;

[0040] The value range of the threshold amplitude scaling term is from 0.1 to 0.9;

[0041] The threshold amplification factor ranges from 15 to 25;

[0042] The radius of the link field ranges from 2.5 to 3.5.

[0043] The value at the center of the matrix is ​​initialized to 0, and the values ​​at other positions are the reciprocal of the sum of the squares of the distances from that position to the center.

[0044] Beneficial effects

[0045] The hierarchical decomposition-based feature extraction method described in this invention has the following advantages compared with existing feature extraction methods:

[0046] 1. The hierarchical decomposition feature extraction method described above can effectively extract the signal and data features collected in complex environments, thereby improving the accuracy of subsequent data classification.

[0047] 2. The hierarchical decomposition feature extraction method has a simple feature extraction structure. Compared with the traditional neural network feature extraction method, it does not require machine learning training to extract features and has low computational complexity. Attached Figure Description

[0048] Figure 1 This is an overall flowchart of a hierarchical decomposition feature extraction method according to the present invention. Detailed Implementation

[0049] The hierarchical decomposition-based feature extraction method of the present invention will be further described and described in detail below with reference to the accompanying drawings and embodiments.

[0050] This invention is based on a hierarchical decomposition feature extraction method, which can be applied to feature extraction and subsequent signal classification in various scenarios, and is especially suitable for feature extraction of signals and data collected in complex environments.

[0051] In specific implementation, the hierarchical decomposition feature extraction method of this invention first preprocesses the collected two-dimensional time series to obtain single-channel effective data; then, a hierarchical decomposition feature extraction structure is established, and the parameters and various correlation matrices of the hierarchical decomposition feature extraction structure are initialized; then, the single-channel effective data is cyclically decomposed to generate ordered element factors, obtaining sequential details from coarse to fine, and the hierarchical decomposition feature extraction result of the effective data is output; finally, the result is sent to a classifier for data classification to obtain the classification result. The specific steps include:

[0052] A feature extraction method based on hierarchical decomposition.

[0053] To achieve the above objectives, the present invention adopts the following technical solution:

[0054] The hierarchical decomposition-based feature extraction method includes the following steps:

[0055] S1. Preprocess the collected two-dimensional time series to obtain multi-channel effective data;

[0056] S2. Establish a hierarchical decomposition feature extraction structure;

[0057] The hierarchical decomposition feature extraction structure includes a receiver domain matrix, a modulation domain matrix, and a pulse generation domain matrix;

[0058] The receiving domain matrix includes a feed channel matrix and a link channel matrix;

[0059] The modulation domain matrix includes an internal activity term matrix;

[0060] The pulse generation domain matrix includes a pulse signal matrix and a threshold matrix;

[0061] S3. Initialize the loop parameters, feature extraction structure parameters, and the square matrix;

[0062] The loop parameters include the number of loops and the total number of loops;

[0063] The feature extraction parameters include link field radius, threshold amplitude scaling term, threshold attenuation term, link attenuation term, link amplitude scaling term, link strength, attenuation time constant, and threshold amplification factor.

[0064] The dimension of the square matrix is ​​2 * link field radius + 1;

[0065] S4. Initialize all matrices in the hierarchical decomposition feature extraction structure to be all zero matrices;

[0066] The matrices in the hierarchical decomposition feature extraction structure include the receiving domain matrix, the modulation domain matrix, and the pulse generation domain matrix;

[0067] S5. Perform hierarchical decomposition on the multi-channel effective data to obtain the hierarchical decomposition feature extraction results;

[0068] S5, specifically:

[0069] S51. Initialize the loop count to 1;

[0070] S52. Convolve the pulse signal matrix with the square matrix to obtain the convolution result;

[0071] S53. Multiply the negative exponential function value of the threshold attenuation term by the feed channel matrix, add the result of multiplying the threshold amplitude scaling term by the convolution result, and add the effective data of the multiple channels to obtain the updated feed channel matrix.

[0072] S54. Multiply the negative exponent of the link decay term by the link channel matrix, and add the result of multiplying the link magnitude scaling term by the convolution result to obtain the updated link channel matrix.

[0073] S55. Add 1 to the link strength, multiply it by the link channel matrix, and then multiply it by the feed channel matrix to obtain the internal activity term matrix;

[0074] S56. Compare the internal activity term matrix with the threshold matrix. If the element value of the internal activity term matrix is ​​greater than the element value of the threshold matrix, set the element value at the corresponding position of the pulse signal matrix to 1, otherwise set it to 0 to obtain the pulse signal matrix.

[0075] S57. Add the product of the negative exponential function value of the decay time constant and the threshold matrix to the product of the threshold amplification factor and the pulse signal matrix, and assign the sum of the two products to the threshold matrix.

[0076] S58. Determine if the number of loops is equal to the total number of loops. If not, increment the number of loops by 1 and jump to S52. If the number of loops is equal to the total number of loops, the loop ends, save the hierarchical decomposition feature extraction results of the pulse signal matrix as valid data in S56, and jump to S6.

[0077] S6. Input the pulse signal matrix into the classifier to classify the data and obtain the classification result.

[0078] This method can easily collect signals in complex environments and extract data features, significantly improving the accuracy of subsequent classification and recognition.

[0079] Example 1

[0080] This invention uses 871 audio data points (6373 characters each) of normal individuals and 671 audio data points (6373 characters each) of patients with depression from the MODMA (Motion of Depression Research Dataset) to perform hierarchical decomposition and feature extraction. The MODMA dataset records audio data from 52 subjects during interviews, reading, and image descriptions; each subject recorded 29 audio data points; the 52 subjects included 29 normal individuals and 23 patients with depression; the interview process involved answering 18 questions with positive, neutral, and negative connotations; the reading process involved reading a short story and three sets of words with positive, neutral, and negative connotations; and the image description process involved freely describing four images.

[0081] The method described in this invention is used for feature extraction from data, and then a classifier is used to obtain classification results. The results in Tables 1 and 2 show that this method has a beneficial effect on feature extraction, enabling the extracted features to achieve higher accuracy in subsequent classification. (See attached table.) Figure 1 Based on the overall flowchart of the hierarchical decomposition feature extraction method, this embodiment illustrates how to implement it.

[0082] like Figure 1 As shown, the hierarchical decomposition-based feature extraction method of the present invention includes the following steps:

[0083] 1) The collected two-dimensional time series is preprocessed to obtain single-channel valid data;

[0084] Preprocessing includes normalization, filtering, signal truncation, denoising, and enhancement. After processing, the length of the effective data S is 6373.

[0085] The effective data of length 6373 is specifically the voiceprint feature data of the extracted speech signal, including energy features, spectral features and prosodic features.

[0086] 2) Establish a hierarchical decomposition feature extraction structure: mainly including the receiving domain, modulation domain, and pulse generation domain. The receiving domain includes the feed channel matrix F and the link channel matrix L; the modulation domain includes the internal activity term matrix U; and the pulse generation domain includes the pulse signal matrix Y and the threshold matrix E.

[0087] Among them, the dimensions of the feed channel matrix F, the link channel matrix L, the internal activity term matrix U, the pulse signal matrix Y, and the threshold matrix E are the same as the dimensions of the single-channel effective data;

[0088] 3) Initialize the parameters of the hierarchical decomposition feature extraction structure; set the total number of iterations M to 30, link attenuation term al to 1.0, threshold attenuation term af to 0.1, attenuation time constant ae to 1.0, link strength β to 0.1, link amplitude scaling term vl to 0.2, threshold amplitude scaling term vf to 0.5, and threshold amplification factor ve to 20.

[0089] 4) Initialize all matrices in the hierarchical decomposition feature extraction structure; the feed channel matrix F, link channel matrix L, internal activity term matrix U, pulse signal matrix Y, and threshold matrix E are all initialized to 0, and the square matrix K is initialized to...

[0090] 5) The single-channel valid data is decomposed into ordered element factors in a cyclic manner through hierarchical decomposition to obtain sequential details from coarse to fine. This includes the following sub-steps:

[0091] 5.1) Calculate the convolution W between the pulse signal matrix Y and the square matrix K;

[0092] W = Y * K

[0093] Where Y is the pulse signal matrix, K is a square matrix, and matrix W is the result of the convolution of Y and K, used to update the feed channel matrix F and the link channel matrix L.

[0094] 5.2) Calculate the feed channel matrix F;

[0095] F = e -af ×F+vf×W+S

[0096] Where F is the feed channel matrix, af is the threshold attenuation term, vf is the threshold amplitude scaling term, and S is the effective data.

[0097] 5.3) Calculate the link channel matrix L;

[0098] L = e -al ×L+vl×W

[0099] Where L is the link channel matrix, al is the link attenuation term, and vl is the link amplitude scaling term.

[0100] 5.4) Calculate the internal activity term matrix U in the modulation domain;

[0101] U = F × (1 + β × L)

[0102] Where U is the internal activity term matrix in the modulation domain.

[0103] 5.5) Calculate the pulse signal matrix Y;

[0104] Y = double(U>E)

[0105] Where Y is the pulse signal matrix, U is the internal activity term matrix, and E is the threshold matrix;

[0106] 5.6) Calculate the threshold matrix E;

[0107] E = e -ae ×E+ve×Y

[0108] Where E is the threshold matrix, ae is the decay time constant, and ve is the threshold amplification factor. 5.7)

[0110] If i ≠ 30, then let i = i + 1, and return to step 5.1) and repeat the process;

[0111] If i = 30, the loop ends and step 6 is executed.

[0112] 6) The pulse signal matrix Y is the result of hierarchical decomposition feature extraction of the output effective data;

[0113] 7) The results of the hierarchical decomposition feature extraction of the effective data are sent to the subsequent classifier for data classification, and the classification results are shown in Table 1 and Table 2.

[0114] The original data is a combination of energy features, spectral features, and prosodic features obtained by performing a single voiceprint feature extraction on the original speech data.

[0115] The hierarchical decomposition feature extraction method described above is a feature extraction method based on the original data.

[0116] Table 1 compares the accuracy of classification of raw data using the patented feature extraction method and without it.

[0117]

[0118] The classification methods and models used in Table 1 are implemented using a classification learner based on MATLAB.

[0119] Fine-grained trees are suitable for handling fine-grained classification problems. Based on a tree structure, they classify samples by progressively dividing the feature space. At each node, branches occur based on certain attributes of the feature until a leaf node is reached and the final class label is assigned.

[0120] K-Nearest Neighbors (KNN) is a classic machine learning algorithm. It is an instance-based learning method that classifies or regresses unknown samples based on the class information of neighboring samples.

[0121] The basic idea of ​​the KNN algorithm is to construct a model in a feature space based on training samples with known categories. When a new sample needs to be classified, the KNN algorithm finds its K nearest neighbors based on the distance between the new sample and the training samples. Then, based on the category information of these K neighbors, a majority vote is used to determine the category of the new sample. Fine-grained KNN, used to solve fine-grained classification problems, is characterized by its flexibility in adjustment and optimization according to the characteristics and needs of the data. Optimizable KNN is an improvement and optimization of the traditional KNN algorithm, aiming to improve the efficiency and performance of the algorithm. It reduces computational complexity, accelerates the neighbor search process, and improves the ability to process large-scale and high-dimensional data by using methods such as KD-trees, nearest neighbor indexing, feature selection, parallel computing, and approximate search.

[0122] Support Vector Machine (SVM) is a supervised learning algorithm used for classification and regression. It achieves classification or regression tasks by mapping data to a high-dimensional space and finding the optimal hyperplane. The goal of SVM is to find a hyperplane that maximizes the margin between different classes (maximum margin), such that the projections of data points onto this hyperplane are as separate as possible. SVM has good generalization ability and can effectively handle high-dimensional and non-linear data.

[0123] Table 2 shows the accuracy of classification of the original data after using the patented feature extraction method.

[0124]

[0125] The classification methods and models used in Table 2 are implemented using MATLAB's classification learner. Linear SVM achieves the classification task by constructing a linear hyperplane. The goal of linear SVM is to find a hyperplane that separates samples of different classes as much as possible, maximizing the margin between the sample points and the hyperplane. It is suitable for linearly separable datasets and has good computational efficiency.

[0126] Quadratic SVM is an extension of linear SVM for handling non-linearly separable data. It constructs a hyperplane by introducing quadratic and cross terms to achieve more accurate classification.

[0127] Cubic SVM is a further extension of quadratic SVM, using cubic terms and cross terms to build a more complex hyperplane to better accommodate the classification of nonlinear data.

[0128] Medium Gaussian SVM is a nonlinear classification method based on the Gaussian kernel function. It uses the Gaussian kernel function to map the feature space to a higher-dimensional space, making the nonlinear relationships of the original features more apparent. By selecting appropriate Gaussian kernel function parameters, it can effectively classify datasets of moderate complexity.

[0129] Coarse Gaussian SVM is an improvement over Medium Gaussian SVM, suitable for more complex datasets. By using a larger Gaussian kernel parameter, Coarse Gaussian SVM can handle more complex nonlinear relationships, thereby improving classification accuracy and robustness.

[0130] As can be seen from the above classification results, the hierarchical decomposition feature extraction method can effectively extract data features and significantly improve the accuracy of data classification.

[0131] The above description is merely a preferred embodiment of the present invention, and the present invention should not be limited to the content disclosed in this embodiment and the accompanying drawings. Any equivalent or modified embodiments made without departing from the spirit of the present invention fall within the scope of protection of the present invention.

Claims

1. A hierarchical decomposition feature extraction method based on hierarchical decomposition feature extraction method, the object of feature extraction of which is speech data, comprising the following steps: S1, pre-processing the collected two-dimensional time sequence to obtain multi-channel effective data; S2, establishing a hierarchical decomposition feature extraction structure; The hierarchical decomposition feature extraction structure comprises a receiving domain matrix, a modulation domain matrix and a pulse generation domain matrix; The receiving domain matrix comprises a feed channel matrix and a link channel matrix; The modulation domain matrix comprises an internal activity item matrix; The pulse generation domain matrix comprises a pulse signal matrix and a threshold matrix; S3, initializing loop parameters, feature extraction structure parameters and square matrix; The loop parameters comprise the number of loops and the total number of loops; The feature extraction parameters comprise the link field radius, the threshold amplitude scaling item, the threshold decay item, the link decay item, the link amplitude scaling item, the link strength, the decay time constant and the threshold amplification coefficient; The dimension of the square matrix is 2*link field radius+1; S4, initializing each matrix in the hierarchical decomposition feature extraction structure as a full 0 matrix; Each matrix in the hierarchical decomposition feature extraction structure comprises a receiving domain matrix, a modulation domain matrix and a pulse generation domain matrix; S5, performing hierarchical decomposition on the multi-channel effective data to obtain hierarchical decomposition feature extraction results; S5, specifically: S51, initializing the number of loops as 1; S52, convolving the pulse signal matrix with the square matrix to obtain a convolution result; S53, multiplying the negative exponential function value of the threshold decay item with the feed channel matrix, then adding the result of multiplying the threshold amplitude scaling item with the convolution result, and finally adding the multi-channel effective data to obtain the updated feed channel matrix; S54, multiplying the negative exponential of the link decay item with the link channel matrix, then adding the result of multiplying the link amplitude scaling item with the convolution result to obtain the updated link channel matrix; S55, multiplying the link strength plus 1 with the link channel matrix, then multiplying the result with the feed channel matrix to obtain the internal activity item matrix; S56, comparing the internal activity item matrix with the threshold matrix, if the element value of the internal activity item matrix is greater than the element value of the threshold matrix, setting the element value of the pulse signal matrix at the corresponding position as 1, otherwise setting it as 0, to obtain the pulse signal matrix; S57, adding the product of the negative exponential function value of the decay time constant and the threshold matrix to the product of the threshold amplification coefficient and the pulse signal matrix, and assigning the sum of the two products to the threshold matrix; S58, judging whether the number of loops is equal to the total number of loops, if not, increasing the number of loops by 1 and jumping to S52; if the number of loops is equal to the total number of loops, the loop ends, the pulse signal matrix of S56 is saved as the hierarchical decomposition feature extraction result of the effective data and jumps to S6; S6, inputting the pulse signal matrix into a classifier for data classification to obtain a classification result.

2. The hierarchical decomposition-based feature extraction method of claim 1, wherein, The pre-processing of S1 comprises normalization, filtering, signal truncation, denoising and enhancement.

3. The hierarchical decomposition-based feature extraction method of claim 1, wherein, The two dimensions in the two-dimensional time sequence of S1 respectively refer to time and channel dimensions, and the dimension of the channel is the same as the number of channels of the multi-channel.

2. The hierarchical decomposition feature extraction method of claim 1, wherein the two-dimensional time sequence is a two-dimensional time sequence of a speech signal.

3. The hierarchical decomposition feature extraction method of claim 1, wherein the two-dimensional time sequence is a two-dimensional time sequence of a speech signal.

4. The hierarchical decomposition feature extraction method of claim 1, wherein the two-dimensional time sequence is a two-dimensional time sequence of a speech signal.

5. The hierarchical decomposition feature extraction method of claim 1, wherein the two-dimensional time sequence is a two-dimensional time sequence of a speech signal.

6. The hierarchical decomposition feature extraction method of claim 1, wherein the two-dimensional time sequence is a two-dimensional time sequence of a speech signal.

7. The hierarchical decomposition feature extraction method of claim 1, wherein the two-dimensional time sequence is a two-dimensional time sequence of a speech signal.

8. The hierarchical decomposition feature extraction method of claim 1, wherein the two-dimensional time sequence is a two-dimensional time sequence of a speech signal.

9. The hierarchical decomposition feature extraction method of claim 1, wherein the two-dimensional time sequence is a two-dimensional time sequence of a speech signal.

10. The hierarchical decomposition feature extraction method of claim 1, wherein the two-dimensional time sequence is a two-dimensional time sequence of a speech signal.

11. The hierarchical decomposition feature extraction method of claim 1, wherein the two-dimensional time sequence is a two-dimensional time sequence of a speech signal.

12. The hierarchical decomposition feature extraction method of claim 1, wherein the two-dimensional time sequence is a two-dimensional time sequence of a speech signal.

13. The hierarchical decomposition feature extraction method of claim 1, wherein the two-dimensional time sequence is a two-dimensional time sequence of a speech signal.

14. The hierarchical decomposition feature extraction method of claim 1, wherein the two-dimensional time sequence is a two-dimensional time sequence of a speech signal.

15. The hierarchical decomposition feature extraction method of claim 1, wherein the two-dimensional time sequence is a two-dimensional time sequence of a speech signal.

16. The hierarchical decomposition feature extraction method of claim 1, wherein the two-dimensional time sequence is a two-dimensional time sequence of a speech signal.

17. The hierarchical decomposition feature extraction method of claim 1, wherein the two-dimensional time sequence is a two-dimensional time sequence of a speech signal.

18. The hierarchical decomposition feature extraction method of claim 1, wherein the two-dimensional time sequence is a two-dimensional time sequence of a speech signal.

19. The hierarchical decomposition feature extraction method of claim 1, wherein the two-dimensional time sequence is a two-dimensional time sequence of a speech signal.

20. The hierarchical decomposition feature extraction method of claim 1, wherein the two-dimensional time sequence is a two-dimensional time sequence of a speech signal.

21. The hierarchical decomposition feature extraction method of claim 1, wherein the two-dimensional time sequence is a two-dimensional time sequence of a speech signal.

22. The hierarchical decomposition feature extraction method of claim 1, wherein the two-dimensional time sequence is a two-dimensional time sequence of a speech signal.

23. The hierarchical decomposition feature extraction method of claim 1, wherein the two-dimensional time sequence is a two-dimensional time sequence of a speech signal.

24. The hierarchical decomposition feature extraction method of claim 1, wherein the two-dimensional time sequence is a two-dimensional time sequence of a speech signal.

25. The hierarchical decomposition feature extraction method of claim 1, wherein the two-dimensional time sequence is a two-dimensional time sequence of a speech signal.

26. The hierarchical decomposition feature extraction method of claim 1, wherein the two-dimensional time sequence is a two-dimensional time sequence of a speech signal.

27. The hierarchical decomposition feature extraction method of claim 1, wherein the two-dimensional time sequence is a two-dimensional time sequence of a speech signal.

28. The hierarchical decomposition feature extraction method of claim 1, wherein the two-dimensional time sequence is a two-dimensional time sequence of a speech signal.

29. The hierarchical decomposition feature extraction method of claim 1, wherein the two-dimensional time sequence is a two-dimensional time sequence of a speech signal.

30. The hierarchical decomposition feature extraction method of claim 1, wherein the two-dimensional time sequence is a two-dimensional time sequence of a speech signal.

31. The hierarchical decomposition feature extraction method of claim 1, wherein the two-dimensional time sequence is a two-dimensional time sequence of a speech signal.

32. The hierarchical decomposition feature extraction method of claim 1, wherein the two-dimensional time sequence is a two-dimensional time sequence of a speech signal.

33. The hierarchical decomposition feature extraction method of claim 1, wherein the two-dimensional time sequence is a two-dimensional time sequence of a speech signal.

34. The hierarchical decomposition feature extraction method of claim 1, wherein the two-dimensional time sequence is a two-dimensional time sequence of a speech signal.

35. The hierarchical decomposition feature extraction method of claim 1, wherein the two-dimensional time sequence is a two-dimensional time sequence of a speech signal.

36. The hierarchical decomposition feature extraction method of claim 1, wherein the two-dimensional time sequence is a two-dimensional time sequence of a speech signal.

37. The hierarchical decomposition feature extraction method of claim 1, wherein the two-dimensional time sequence is a two-dimensional time sequence of a speech signal.

38. The hierarchical decomposition feature extraction method of claim 1, wherein the two-dimensional time sequence is a two-dimensional time sequence of a speech signal.

39. The hierarchical decomposition feature extraction method of claim 1, wherein the two-dimensional time sequence is a two-dimensional time sequence of a speech signal.

40. The hierarchical decomposition feature extraction method of claim 1, wherein the two-dimensional time sequence is a two-dimensional time sequence of a speech signal.

41. The hierarchical decomposition feature extraction method of claim 1, wherein the two-dimensional time sequence is a two-dimensional time sequence of a speech signal.

42. The hierarchical decomposition feature extraction method of claim 1, wherein the two-dimensional time sequence is a two-dimensional time sequence of a speech signal.

43. The hierarchical decomposition feature extraction method of claim 1, wherein the two-dimensional time sequence is a two-dimensional time sequence of a speech signal.

44. The hierarchical decomposition feature extraction method of claim 1, wherein the two-dimensional time sequence is a two-dimensional time sequence of a speech signal.

45. The hierarchical decomposition feature extraction method of claim 1, wherein the two-dimensional time sequence is a two-dimensional time sequence of a speech signal.

46. The hierarchical decomposition feature extraction 4. The hierarchical decomposition-based feature extraction method of claim 1, wherein, The dimensions of the feed channel matrix and the link channel matrix, the internal activity item matrix, the pulse signal matrix and the threshold matrix are the same as the dimensions of the single-channel effective data.

5. The hierarchical decomposition-based feature extraction method of claim 1, wherein, The total cycle number is in the range of 20 to 50, and the cycle number is in the range of 1 to the total cycle number.

6. The hierarchical decomposition-based feature extraction method of claim 1, wherein, The value of the center position of the square matrix is initialized to 0, and the values of other positions are the reciprocal values of the square sum of the distance of the position from the center position. The value of the link attenuation item is in the range of 1 to 5. The value of the threshold attenuation item is in the range of 1 to 5. The value of the decay time constant is in the range of 1 to 5. The value of the link strength is in the range of 0 to 1. The value of the link amplitude scaling item is in the range of 0.1 to 0.

9. The value of the threshold amplitude scaling item is in the range of 0.1 to 0.

9. The value of the threshold amplification coefficient is in the range of 15 to 25. The value of the link field radius is in the range of 2.5 to 3.

5. The value of the center position of the square matrix is initialized to 0, and the values of other positions are the reciprocal values of the square sum of the distance of the position from the center position.

Images