A classification method and apparatus based on feature intersection and fusion feature selection

Abstract

This invention relates to a classification method and apparatus based on feature cross-selection and fusion feature selection. The method includes: constructing a dataset comprising several speech data points and labels; extracting feature data from each speech data point and constructing a concatenated two-dimensional matrix accordingly; performing a first fusion feature selection, adaptive feature cross-selection, and a second fusion feature selection on the concatenated two-dimensional matrix based on the labels to obtain data to be classified, and dividing the data to be classified into a training set and a test set; the adaptive feature cross-selection is performed as follows: if the linear correlation coefficient between features is greater than a threshold, a non-linear transformation is performed before feature cross-selection; otherwise, feature cross-selection is performed directly; the training set and labels are input into a classification learner to obtain a trained classification learner; the test set is input into the trained classification learner to obtain the classification result. The apparatus implements the method. This invention uses two fusion feature selections, improving the dimensionality explosion problem during feature cross-selection and exhibiting better classification robustness.

Description

Technical Field

[0001] This invention belongs to the field of machine learning, specifically relating to a classification method and apparatus based on feature cross-fertilization and feature selection. Background Technology

[0002] Time series represent signals collected over a period of time, exhibiting temporal order and consisting of data points sorted chronologically. Time series analysis encompasses both diagnosing past data and predicting future data. Numerous applications have significantly contributed to the development of time series analysis. Typical applications include medicine, weather, economics, and astronomy. Typical time series analysis methods include statistical methods, machine learning-based methods, and ensemble methods. Machine learning, particularly in classification applications, offers significant advantages when processing time series data. This type of data is widely used in energy, finance, and healthcare, including signals such as electricity load, stock price trends, electrocardiograms, and speech. Through time series classification, future trends can be predicted, anomalies detected, and more informed decisions made. Speech classification, as a form of category learning for the auditory channel, helps individuals quickly identify and interpret sound stimuli, possessing unique survival significance.

[0003] The choice of feature extraction and selection methods is crucial to the final result. Therefore, exploring new feature extraction and selection techniques has long been a research hotspot in machine learning. In feature extraction, to extract deeper features from data, secondary feature extraction methods are often employed, including: first extracting features and then performing feature cross-interaction, or extracting features first and then performing secondary extraction, dimensionality increase, or dimensionality reduction. Feature cross-interaction includes explicit and implicit features. Implicit feature cross-interaction typically uses deep neural networks, which is highly complex. Explicit feature cross-interaction is usually based on linear transformations of existing features, such as inner products and Cartesian products. This type of cross-interaction has poor fitting to nonlinear models, can only use a fixed cross-interaction method on two types of features at a time, and cannot adaptively select the best cross-interaction method based on the characteristics of existing features, nor does it retain the original features after cross-interaction. Regarding feature selection, although various feature selection methods based on different principles have been proposed, current technologies often use a single feature selection method. This cannot always use the most suitable selection for the current feature in different scenarios, thus often failing to achieve optimal results. Summary of the Invention

[0004] The purpose of this invention is to address the shortcomings of existing display feature cross-validation methods, such as difficulty in fitting nonlinear models, limited number of feature types per cross-validation, poor adaptability, failure to retain original features, and low classification accuracy due to single feature selection. This invention proposes a classification method and apparatus based on feature cross-validation and fusion feature selection. The method first collects the subject's speech signal, extracts multiple feature data from the collected speech data, performs a first fusion feature selection after concatenating all the feature data, and then performs adaptive feature cross-validation on the first-selection feature data to obtain cross-validation feature data. The cross-validation feature data undergoes a second fusion feature selection to obtain the data to be classified, which is then used for machine learning. Through machine learning, a classification method capable of classifying the input speech data is finally obtained. The classification apparatus is used to implement classification based on feature cross-validation and fusion feature selection.

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

[0006] As a first aspect of the present invention, a classification device based on feature crossover and fusion feature selection is proposed, including a speech acquisition module, a feature extraction module, a first fusion feature selection module, an adaptive feature crossover module, a second fusion feature selection module, and a model training module;

[0007] The speech acquisition module is connected to the feature extraction module, the feature extraction module is connected to the first fusion feature selection module, the first fusion feature selection module is connected to the adaptive feature cross module, the adaptive feature cross module is connected to the second fusion feature selection module, and the second fusion feature selection module is connected to the model training module.

[0008] The speech acquisition module converts the speech signal into speech data using a digital-to-analog conversion and transmits it to the feature extraction module. The feature extraction module extracts multiple features from the speech data and transmits the resulting feature data to the first fusion feature selection module. The first fusion feature selection module selects a predetermined number of features from the feature data and transmits the first-selection data to the adaptive feature cross module. The adaptive feature cross module performs adaptive feature cross on the first-selection data and transmits the resulting cross-selected feature data to the second fusion feature selection module. The second fusion feature selection module selects a predetermined number of features from the cross-selected feature data and transmits the resulting data to be classified to the model training module. The model training module uses the aforementioned data to be classified to train the model and finally derives the classification model with the highest training accuracy.

[0009] The voice acquisition module includes a microphone and an analog-to-digital converter;

[0010] The microphone in the voice acquisition module is used to acquire voice data. The voice data acquired by the microphone is converted from analog to digital by an analog-to-digital converter and then transmitted to the feature extraction module.

[0011] The feature extraction module includes a wavelet feature extraction unit, a filter feature extraction unit, and an entropy feature extraction unit;

[0012] The number of wavelet feature extraction units is W, and W is greater than or equal to 2.

[0013] The number of the filter feature extraction units is F, where F is greater than or equal to 1;

[0014] The number of entropy feature extraction units is E, where E is greater than or equal to 1.

[0015] The feature extraction module receives the above-mentioned speech data, and the above three feature extraction units run in parallel to extract feature data of W group wavelet features, F group filter features and E group entropy features respectively.

[0016] The feature data of the wavelet features of group W, the filter features of group F, and the entropy features of group E are passed to the first fusion feature selection module.

[0017] The first fusion feature selection module includes a first average MIC calculation unit, a first feature selection unit, and a first feature supplementation unit;

[0018] The number of the first feature selection units is C1, where C1 is greater than or equal to 2;

[0019] The first fusion feature selection module receives the feature data and uses the first average MIC calculation unit to calculate the average MIC of the feature with the label and feature judged by the doctor and other features. Based on this MIC, the above C1 first feature selection units run in parallel to select the corresponding number of features of C1 class. The first feature supplementation unit supplements the features based on the difference between the number of features after taking the union of the above C1 class features and the number of features to be retained, and obtains the data after the first selection and passes it to the adaptive feature cross module.

[0020] The adaptive feature cross module includes a linear correlation coefficient calculation unit, a nonlinear transformation unit, and a Kronecker product unit.

[0021] The adaptive feature cross-multiplication module receives the data after the first selection, calculates the average linear correlation coefficient between features using the linear correlation coefficient calculation unit, and determines whether this linear correlation coefficient is greater than or equal to a threshold S. If not, i.e., less than the threshold S, the data after the first selection is passed as input data to the Kronecker product unit; if yes, i.e., greater than or equal to the threshold S, the data after the first selection is processed by the nonlinear transformation unit and then passed as input data to the Kronecker product unit. The Kronecker product unit calculates the Kronecker product between the input data and the data after the first selection to obtain the cross-multiplication feature data, which is then passed to the second feature selection module.

[0022] The threshold S has a value range of 0 ≤ S ≤ 1;

[0023] The second feature selection module includes a second average MIC calculation unit, a second feature selection unit, and a second feature supplementation unit;

[0024] The number of the first feature selection units is C2, where C2 is greater than or equal to 2;

[0025] The second feature selection module receives the cross-feature data and uses the second average MIC calculation unit to calculate the average MIC of the feature and the label, and the feature and other features. Based on this MIC, the above C2 second feature selection units run in parallel to select the corresponding number of features of class C2. The second feature supplementation unit supplements features based on the difference between the number of features after taking the union of the above C2 features and the number of features to be retained, and obtains the data to be classified and passes it to the model training module.

[0026] The model training module includes several classification models;

[0027] The model training module receives the data to be classified and puts it into several classification models for training. The model with the highest training accuracy is taken as the output of the model training module.

[0028] As a second aspect of the present invention, a classification method based on feature intersection and fusion feature selection is proposed, comprising the following steps:

[0029] S1. The collected speech is converted from analog to digital to obtain speech data, and each speech data is tagged according to the doctor's diagnosis.

[0030] The number of voice data entries is X; the value of X is in the range of 2 ≤ X;

[0031] S2. Extract several types of features from the speech data;

[0032] The number of feature types is N; the value of N is 1 ≤ N;

[0033] The nth type feature of each voice data has Yn There are n features, and the number of types of the features is the same as the number of feature types; where n = 1, 2, ..., N, and N is the number of feature types; the total number of the several types of features is...

[0034] S3. Concatenate the feature data corresponding to each speech data to obtain a concatenated vector, which is then expressed as a concatenated two-dimensional matrix.

[0035] The length of the concatenated vector is The spliced ​​two-dimensional matrix is ​​a two-dimensional matrix with X rows and Y columns, including feature data of all X voice data, with each row corresponding to one voice data and each column corresponding to one feature;

[0036] S4. Perform the first fusion feature selection on all features in the spliced ​​two-dimensional matrix according to the labels to obtain the data after the first selection;

[0037] The first fusion feature selection is specifically as follows:

[0038] S41. Calculate the maximum mutual information coefficient between the feature and the label in the feature data, and the average maximum information coefficient between the feature and other features;

[0039] The feature is the y-th feature, where y = 1, 2, ..., Y; the maximum mutual information coefficient with the label is MIC1. y MIC1 y The value range is 0 ≤ MIC1 y ≤1; the average maximum information coefficient with respect to other features is MIC2. y MIC2 y The value range is 0 ≤ MIC2 y ≤1;

[0040] S42. Calculate the mean of the maximum mutual information coefficients between all features and labels, and the mean of the average maximum information coefficients between the feature and other features.

[0041] The mean of the maximum mutual information coefficient with the tag The range of values ​​is The mean of the average information coefficients with other features The range of values ​​is

[0042] S43, based on S42 and Use the appropriate feature selection method to select and retain the feature data, and obtain the selected features;

[0043] Specifically, S43 is as follows:

[0044] S431. Select the corresponding number of features using different feature selection methods according to the number of feature data to be retained;

[0045] The number of feature data to be retained is M; the value of M is 1 ≤ M ≤ Y; the number of feature selection methods is C1, and it includes at least one method that does not consider the correlation between features and one method that does consider the correlation between features; the value of C1 is 2 ≤ C1; the corresponding number is and indivual;

[0046] S432. Take the union of the features obtained using different feature selection methods in S431 to obtain the selected features to be supplemented.

[0047] The number of selected features to be supplemented is Mm, where m represents the number of features repeatedly selected by different feature selection methods;

[0048] S433. Supplement the features of repeated selection in S432 to obtain the selected features, specifically:

[0049] like If the m features are removed from S431 without considering the correlation between features in S431, then m unique features are selected to supplement the feature selection; otherwise, if the m features are removed from S431 with consideration of the correlation between features in S431, then m unique features are selected to supplement the feature selection.

[0050] S44. Select the feature data in S3 based on the selected features to obtain the data after the first selection;

[0051] The number of rows and columns of the data after the first selection is X, with each row corresponding to one voice data point and each column corresponding to one feature.

[0052] S5. Perform adaptive feature crossing on the data selected in S44 after the first selection to obtain the cross-linked feature data, specifically:

[0053] S51. Calculate the linear correlation coefficient between every two columns of feature data in the first selection data in S44, and then calculate the mean of all linear correlation coefficients.

[0054] The linear correlation coefficient r ij ,i,j=1,2,...,M,j≠i;r ij The range of values ​​for is 0 ≤ r ij ≤1; the mean of the linear correlation coefficient The range of values ​​is

[0055] S52, if in S51 If the value is less than the threshold S, calculate the Kronecker product of each row of data in S34 with itself; otherwise, first calculate the result of each row of data after passing through the nonlinear function, and then calculate the Kronecker product of the result with the row of data.

[0056] The threshold S has a range of 0 ≤ S ≤ 1; the Kronecker product is a square matrix of dimension M in both cases;

[0057] S53. Arrange the elements of the square matrix corresponding to each row of data obtained in S52 into a row vector. Then the feature data corresponding to all physiological data can be represented by a two-dimensional matrix, which is called cross feature data.

[0058] The length of the row vector is M. 2 The number of rows in the cross-feature data is X, and the number of columns is M. 2 Each row corresponds to one speech data point, and each column corresponds to one cross feature.

[0059] S6. Based on the labels, perform a second fusion feature selection on the cross features of the cross-feature data obtained in S53 to obtain the second selected data.

[0060] The second fusion feature selection is specifically as follows:

[0061] S61. Calculate the maximum mutual information coefficient between the feature and the label in the feature data, and the average maximum information coefficient between the feature and other features.

[0062] The feature is the z-th feature, where z = 1, 2, ..., M 2 The maximum mutual information coefficient between the tag and the target is MIC3. z MIC3 z The value range is 0 ≤ MIC3 z ≤1; the average maximum information coefficient with respect to other features is MIC4. z MIC4 z The value range is 0 ≤ MIC4 z ≤1;

[0063] S62. Calculate the mean of the maximum mutual information coefficients between all features and labels, and the mean of the average maximum information coefficients between the feature and other features;

[0064] The mean of the maximum mutual information coefficient with the tag The range of values ​​is The mean of the average information coefficients with other features The range of values ​​is

[0065] S63, based on S62 and Use the appropriate feature selection method to select and retain the feature data, and obtain the selected features;

[0066] Specifically, S63 is as follows:

[0067] S631. Select the corresponding number of features using different feature selection methods according to the number of feature data to be retained;

[0068] The desired number of feature data to be retained is L; the value of L ranges from 1 to L to M²; the number of feature selection methods is C², and it includes at least one method that does not consider the correlation between features and one method that does consider the correlation between features; the value of C² ranges from 2 to C²; the corresponding quantities are respectively and indivual;

[0069] S632. Take the union of the features obtained using different feature selection methods in S631 to obtain the selected features to be supplemented.

[0070] The number of selected features to be supplemented is Ll, ​​where l represents the number of features repeatedly selected by different feature selection methods;

[0071] S633. Supplement the features of repeated selection in S632 to obtain the selected features, specifically:

[0072] like If the feature in S631 is removed without considering the correlation between features, then l unique features are selected for supplementation; otherwise, l unique features are selected for supplementation after removing the features in S631 with consideration of the correlation between features.

[0073] S64. Select the cross feature data from S53 based on the selected features to obtain the data to be classified;

[0074] The number of rows in the data to be classified is X, the number of columns is L, each row corresponds to one speech data point, and each column corresponds to one feature.

[0075] S7. Divide the data to be classified obtained in S64 into training set and test set according to the proportion;

[0076] S8. Input the training set and labels from S7 into the classification learner to train the model, and select the model with the highest training accuracy for export.

[0077] S9. Input the test set obtained in S7 into the model obtained in S8 to get the classification result;

[0078] Thus, through steps S1 to S9, a classification method based on feature crossover and fusion feature selection has been completed.

[0079] Beneficial effects

[0080] The classification method and apparatus based on feature intersection and fusion feature selection proposed in this invention have the following advantages compared with existing classification methods and apparatuses:

[0081] 1. The method uses two fusion feature selections, which improves the situation where the dimension of the cross features is too high due to dimensionality explosion when performing feature cross-cutting;

[0082] 2. The fusion feature selection comprehensively considers the correlation between features and labels, and between features and other features, reducing redundancy among features while ensuring the effectiveness of the selected features;

[0083] 3. The fusion feature selection method can adaptively adjust the feature selection method used based on the above two types of correlation, and has good robustness when facing data with different characteristics;

[0084] 4. The adaptive feature crossing can complete the crossing of more than two types of features in a single calculation;

[0085] 5. The adaptive feature crossing retains the effective features when no crossing occurs, while outputting cross-cross results for different types of features;

[0086] 6. The adaptive feature crossing can adaptively adjust the crossing method according to the linear correlation between the data to be crossed, and use nonlinear functions to calculate features with high linear correlation, thereby enhancing the fit to the nonlinear model;

[0087] 7. The nonlinear function calculation of the adaptive feature crossing is only used on one side of the data to be crossed, which improves the redundancy of "feature 1 crossing feature 2" and "feature 2 crossing feature 1" at the same time. The newly generated "feature 1 crossing feature 2 after nonlinear function" and "feature 2 crossing feature 1 after nonlinear function" enrich the deep features obtained by crossing. Attached Figure Description

[0088] Figure 1 This is a schematic diagram of the structure of a classification device based on feature intersection and fusion feature selection according to the present invention;

[0089] Figure 2 This is a flowchart of the overall process of a classification method based on feature intersection and fusion feature selection according to the present invention;

[0090] Figure 3This is a flowchart of the fusion feature selection process in a classification method based on feature intersection and fusion feature selection according to the present invention;

[0091] Figure 4 This is a flowchart of the adaptive feature crossing process of a classification method based on feature crossing and fusion feature selection according to the present invention;

[0092] Figure 5 This invention relates to a classification device based on feature crossover and fusion feature selection, which presents a classification result using a confusion matrix and human statistics.

[0093] Figure 6 This invention relates to a classification method and apparatus based on feature crossover and fusion feature selection, which presents a classification result using speech statistics represented by a confusion matrix. Detailed Implementation

[0094] The following description, in conjunction with the accompanying drawings and embodiments, provides a further explanation and detailed description of the classification method and apparatus based on feature intersection and fusion feature selection proposed in this invention.

[0095] Example 1

[0096] The present invention provides a classification device based on feature crossover and fusion feature selection, comprising a speech acquisition module, a feature extraction module, a first fusion feature selection module, an adaptive feature crossover module, a second fusion feature selection module, and a model training module;

[0097] In specific implementation, the voice acquisition module is connected to the feature extraction module, the feature extraction module is connected to the first fusion feature selection module, the first fusion feature selection module is connected to the adaptive feature cross module, the adaptive feature cross module is connected to the second fusion feature selection module, and the second fusion feature selection module is connected to the model training module.

[0098] The speech acquisition module converts the speech signal into speech data using a digital-to-analog conversion and transmits it to the feature extraction module. The feature extraction module extracts multiple features from the speech data and transmits the resulting feature data to the first fusion feature selection module. The first fusion feature selection module selects a predetermined number of features from the feature data and transmits the first-selection data to the adaptive feature cross module. The adaptive feature cross module performs adaptive feature cross on the first-selection data and transmits the resulting cross-selected feature data to the second fusion feature selection module. The second fusion feature selection module selects a predetermined number of features from the cross-selected feature data and transmits the resulting data to be classified to the model training module. The model training module uses the data to be classified to train the model and finally derives the classification model with the highest training accuracy.

[0099] The adaptive feature cross-multiplication module receives the data after the first selection, calculates the average linear correlation coefficient between features using the linear correlation coefficient calculation unit, and determines whether this linear correlation coefficient is greater than or equal to a threshold S. If not, i.e., less than the threshold S, the data after the first selection is passed as input data to the Kronecker product unit; if yes, i.e., greater than or equal to the threshold S, the data after the first selection is processed by the nonlinear transformation unit and then passed as input data to the Kronecker product unit. In a specific implementation, S = 0.5.

[0100] The feature extraction module includes a wavelet feature extraction unit, a filter feature extraction unit, and an entropy feature extraction unit; the voice acquisition module includes a microphone and an analog-to-digital converter; the microphone in the voice acquisition module is used to acquire voice, and the voice acquired by the microphone is converted into voice data by the analog-to-digital converter and then transmitted to the feature extraction module.

[0101] The feature extraction module receives the speech data, and the feature extraction units operate in parallel to extract feature data of W groups of wavelet features, F groups of filter features and E groups of entropy features respectively.

[0102] The number of wavelet feature extraction units is W, and W is greater than or equal to 2; in specific implementations, W = 2 or W = 3.

[0103] The number of the filtering feature extraction units is F, where F is greater than or equal to 1; in a specific implementation, F = 2.

[0104] The number of entropy feature extraction units is E, where E is greater than or equal to 1; in a specific implementation, E = 4.

[0105] The feature data of the wavelet features of group W, the filter features of group F, and the entropy features of group E are passed to the first fusion feature selection module.

[0106] The first fusion feature selection module includes a first average MIC calculation unit, a first feature selection unit, and a first feature supplementation unit;

[0107] The number of the first feature selection units is C1; in a specific implementation, C1 = 2.

[0108] The first fusion feature selection module receives the feature data, calculates the average MIC of the feature with the label and other features judged by the doctor using the first average MIC calculation unit, and the C1 first feature selection units run in parallel according to this MIC to select the corresponding number of features of class C1. The first feature supplementation unit supplements the features according to the difference between the number of features after taking the union of the features of class C1 and the number of features to be retained, and obtains the data after the first selection and passes it to the adaptive feature cross module.

[0109] The adaptive feature crossing module includes a Kronecker product unit;

[0110] The Kronecker product unit performs the Kronecker product on the input data and the data after the first selection to obtain the cross-feature data, which is then passed to the second feature selection module.

[0111] The second feature selection module includes a second average MIC calculation unit, a second feature selection unit, and a second feature supplementation unit;

[0112] The number of the second feature selection units is C2; in a specific implementation, C2 = 2.

[0113] The second feature selection module receives the cross-feature data, calculates the average MIC of the feature and label, and the feature and other features using the second average MIC calculation unit, and the C2 second feature selection units run in parallel according to this MIC to select the corresponding number of features of class C2. The second feature supplementation unit supplements features according to the difference between the number of features after taking the union of the features of class C2 and the number of features to be retained, and obtains the data to be classified and passes it to the model training module.

[0114] The model training module includes several classification models;

[0115] The model training module receives the data to be classified and puts it into several classification models for training, and takes the model with the highest training accuracy as the output of the model training module.

[0116] Example 2

[0117] As a second aspect of the present invention, a classification method based on feature intersection and fusion feature selection is proposed, comprising the following steps:

[0118] S1. Perform analog-to-digital conversion on the collected speech to obtain speech data and add tags to each piece of speech data;

[0119] S2. Extract several types of features from the speech data;

[0120] S3. Concatenate the feature data corresponding to each speech data to obtain a concatenated vector, which is then expressed as a concatenated two-dimensional matrix.

[0121] S4. Perform the first fusion feature selection on all feature data in the concatenated two-dimensional matrix according to the labels to obtain the first selected data. Specifically, this involves: selecting the desired number of feature data based on the mean of the maximum mutual information coefficients between all features and labels, and the mean of the average maximum mutual information coefficients between the feature and other features; selecting the corresponding number of features using different feature selection methods; taking the union of the features obtained by different feature selection methods to obtain the selected features to be supplemented; supplementing the repeatedly selected features to obtain the selected features; then using different feature selection methods to select feature data to retain to obtain the selected features; and finally selecting the corresponding feature data based on the selected features to obtain the first selected data, including the following sub-steps:

[0122] S41. Calculate the maximum mutual information coefficient between the feature and the label in the feature data, and the average maximum information coefficient between the feature and other features;

[0123] S42. Calculate the mean of the maximum mutual information coefficients between all features and labels, and the mean of the average maximum information coefficients between the feature and other features.

[0124] S43. Based on the results obtained in S42, use the appropriate feature selection method to select and retain feature data to obtain the selected features;

[0125] S44. Select the feature data in S3 based on the selected features to obtain the data after the first selection;

[0126] S5. Perform adaptive feature crossing on the data after the first selection to obtain the cross-linked feature data;

[0127] The adaptive feature crossing is as follows: if the linear correlation coefficient between features is greater than the threshold, a nonlinear transformation is performed first, followed by feature crossing; otherwise, feature crossing is performed directly.

[0128] S6. Based on the labels, perform a second fusion feature selection on the cross features of the cross-feature data to obtain the data to be classified.

[0129] S7. Divide the data to be classified into training set and test set according to the proportion;

[0130] S8. Input the training set and labels into the classification learner to train the model, and select and export the model with the highest training accuracy.

[0131] S9. Input the test set into the model obtained in S8 to get the classification result.

[0132] The first fusion feature selection in S4 or the second fusion feature selection in S6 are specifically as follows:

[0133] SA1, calculate the maximum mutual information coefficient between the feature and the label in the feature data, and the average maximum information coefficient between the feature and other features;

[0134] SA2. Calculate the mean of the maximum mutual information coefficients between all features and labels, and the mean of the average maximum information coefficients between the feature and other features.

[0135] SA3. Based on the results obtained from SA2, use the corresponding feature selection method to select and retain feature data, and obtain the selected features;

[0136] SA4. Select cross-feature data based on the selected features to obtain the data to be classified.

[0137] The number of voice data items in S1 is X; the value of X is in the range of 2 ≤ X;

[0138] The number of feature types mentioned in S2 is N; the value of N is 1≤N;

[0139] The nth type feature of each voice data has Y n There are n features, and the number of types of the features is the same as the number of feature types; where n = 1, 2, ..., N; the total number of the various types of features is...

[0140] The length of the concatenated vector described in S3 is The concatenated two-dimensional matrix is ​​a two-dimensional matrix with X rows and Y columns, including feature data of all X speech data points, with each row corresponding to one speech data point and each column corresponding to one feature. If this is the first feature selection for fusion, the SA includes the following sub-steps:

[0141] SA1's feature data includes Y features, denoted as y = 1, 2, ..., Y; the maximum mutual information coefficient with the label is MIC1. y MIC1 y The value range is 0 ≤ MIC1 y ≤1; the average maximum information coefficient with respect to other features is MIC2. y MIC2 y The value range is 0 ≤ MIC2 y ≤1;

[0142] The mean of the maximum mutual information coefficients between SA2 and the tag The range of values ​​is The mean of the average information coefficients with other features The range of values ​​is

[0143] SA3 specifically refers to: SA31, selecting a corresponding number of features using different feature selection methods based on the number of feature data to be retained;

[0144] The number of feature data to be retained is M; the value of M is 1 ≤ M ≤ Y; the number of feature selection methods is C1, and it includes at least one method that does not consider the correlation between features and one method that does consider the correlation between features; the value of C1 is 2 ≤ C1; the corresponding quantities are respectively and indivual;

[0145] SA32. Take the union of the features obtained from SA31 using different feature selection methods to obtain the selected features to be supplemented.

[0146] The number of selected features to be supplemented is Mm, where m represents the number of features repeatedly selected by different feature selection methods;

[0147] SA33, supplementing the features of repeated selection in SA32, yields the selected features, specifically: if If m unique features are removed from S431 under the method in SA31 that does not consider the correlation between features, then m unique features are removed from S431 under the method in SA31 that considers the correlation between features.

[0148] SA4 states that after the first selection, the number of rows in the data is X, and the number of columns is M. Each row corresponds to one speech data point, and each column corresponds to one feature.

[0149] S5 specifically refers to: S51, calculating the linear correlation coefficient between every two columns of feature data in the first selection of data, and then calculating the mean of all linear correlation coefficients;

[0150] The linear correlation coefficient r mentioned in S51 ij ,i,j=1,2,...,M,j≠i;r ij The range of values ​​for is 0 ≤ r ij ≤1; the mean of the linear correlation coefficient The range of values ​​is

[0151] S52, if in S51 If the value is less than the threshold S, calculate the Kronecker product of each row of data in S34 with itself; otherwise, first calculate the result of each row of data after passing through the nonlinear function, and then calculate the Kronecker product of the result with the row of data.

[0152] The threshold S has a range of 0 ≤ S ≤ 1; the Kronecker product is a square matrix of dimension M in both cases;

[0153] S53. Arrange the elements of the square matrix corresponding to each row of data obtained in S52 into a row vector. Then the feature data corresponding to all physiological data can be represented by a two-dimensional matrix, which is called cross feature data.

[0154] The length of the row vector is M. 2 The number of rows in the cross-feature data is X, and the number of columns is M. 2 Each row corresponds to one speech data point, and each column corresponds to one cross feature.

[0155] If it is a second fusion feature selection, step S6 includes sub-steps in the order of S61, S62, S63, and S64, specifically as follows:

[0156] The feature described in S61 is the z-th feature, where z = 1, 2, ..., M 2 The maximum mutual information coefficient between the tag and the target is MIC3. z MIC3 z The value range is 0 ≤ MIC3 z ≤1; the average maximum information coefficient with respect to other features is MIC4. z MIC4 z The value range is 0 ≤ MIC4 z ≤1;

[0157] The mean of the maximum mutual information coefficients with the tag as described in S62 The range of values ​​is The mean of the average information coefficients with other features The range of values ​​is

[0158] Specifically, S63 is as follows:

[0159] S631. Select the corresponding number of features using different feature selection methods according to the number of feature data to be retained;

[0160] The desired number of feature data to be retained is L; the value of L ranges from 1 to L to M²; the number of feature selection methods is C², and it includes at least one method that does not consider the correlation between features and one method that does consider the correlation between features; the value of C² ranges from 2 to C²; the corresponding quantities are respectively and indivual;

[0161] S632. Take the union of the features obtained using different feature selection methods in S631 to obtain the selected features to be supplemented.

[0162] The number of selected features to be supplemented is Ll, ​​where l represents the number of features repeatedly selected by different feature selection methods;

[0163] S633. Supplement the features of repeated selection in S632 to obtain the selected features, specifically:

[0164] like If the feature in S631 is removed without considering the correlation between features, then l unique features are selected for supplementation; otherwise, l unique features are selected for supplementation after removing the features in S631 with consideration of the correlation between features.

[0165] The number of rows and columns of the data to be classified in S64 is X, with each row corresponding to one speech data point and each column corresponding to one feature.

[0166] In practice, Figure 1 This is a schematic diagram of a classification device based on feature cross-fertilization and fusion feature selection according to the present invention. The classification device based on feature cross-fertilization and fusion feature selection of this application includes a speech acquisition module, a feature extraction module, a first fusion feature selection module, an adaptive feature cross-fertilization module, a second fusion feature selection module, and a model training module.

[0167] The speech acquisition module is connected to the feature extraction module, the feature extraction module is connected to the first fusion feature selection module, the first fusion feature selection module is connected to the adaptive feature cross module, the adaptive feature cross module is connected to the second fusion feature selection module, and the second fusion feature selection module is connected to the model training module.

[0168] The speech acquisition module converts the speech signal into speech data using a digital-to-analog conversion and transmits it to the feature extraction module. The feature extraction module extracts multiple features from the speech data and transmits the resulting feature data to the first fusion feature selection module. The first fusion feature selection module selects a predetermined number of features from the feature data and transmits the first-selection data to the adaptive feature cross module. The adaptive feature cross module performs adaptive feature cross on the first-selection data and transmits the resulting cross-selected feature data to the second fusion feature selection module. The second fusion feature selection module selects a predetermined number of features from the cross-selected feature data and transmits the resulting data to be classified to the model training module. The model training module uses the aforementioned data to be classified to train the model and finally derives the classification model with the highest training accuracy.

[0169] The adaptive feature cross-multiplication module receives the data after the first selection, calculates the average linear correlation coefficient between features using the linear correlation coefficient calculation unit, and determines whether this linear correlation coefficient is greater than or equal to a threshold S. If not, i.e., less than the threshold S, the data after the first selection is passed as input data to the Kronecker product unit; if yes, i.e., greater than or equal to the threshold S, the data after the first selection is processed by the nonlinear transformation unit and then passed as input data to the Kronecker product unit. S is greater than or equal to 1 / 3 and less than or equal to 1. Specifically, in this embodiment, S is 1 / 3, or it can be 0.5.

[0170] The feature extraction module includes a wavelet feature extraction unit, a filter feature extraction unit, and an entropy feature extraction unit; the voice acquisition module includes a microphone and an analog-to-digital converter; the microphone in the voice acquisition module is used to acquire voice, and the voice acquired by the microphone is converted into voice data by the analog-to-digital converter and then transmitted to the feature extraction module.

[0171] Furthermore, the feature extraction module receives the speech data, and the three feature extraction units mentioned above operate in parallel to extract feature data of W groups of wavelet features, F groups of filter features, and E groups of entropy features, respectively.

[0172] The number of wavelet feature extraction units is W, where W is greater than or equal to 2; specifically in this embodiment, W is 2, or it can be 5.

[0173] The number of the filtering feature extraction units is F, where F is greater than or equal to 1; specifically in this embodiment, F is 1, or it can be 2.

[0174] The number of entropy feature extraction units is E, where E is greater than or equal to 1; specifically in this embodiment, E is 1, or it can be 3.

[0175] The feature data of the wavelet features of group W, the filter features of group F, and the entropy features of group E are passed to the first fusion feature selection module.

[0176] The first fusion feature selection module includes a first average MIC calculation unit, a first feature selection unit, and a first feature supplementation unit;

[0177] Preferably, the number of the first feature selection units is C1, where C1 is greater than or equal to 2; specifically in this embodiment, C1 is 2, but it can also be 4.

[0178] The first fusion feature selection module receives the feature data and uses the first average MIC calculation unit to calculate the average MIC of the feature with the label and feature judged by the doctor and other features. Based on this MIC, the above C1 first feature selection units run in parallel to select the corresponding number of features of C1 class. The first feature supplementation unit supplements the features based on the difference between the number of features after taking the union of the above C1 class features and the number of features to be retained, and obtains the data after the first selection and passes it to the adaptive feature cross module.

[0179] The adaptive feature cross module includes a linear correlation coefficient calculation unit, a nonlinear transformation unit, and a Kronecker product unit.

[0180] The Kronecker product unit calculates the Kronecker product between the input data and the data after the first selection to obtain the cross-feature data and passes it to the second feature selection module.

[0181] The threshold S has a value range of 1 / 3 ≤ S ≤ 1.

[0182] The second feature selection module includes a second average MIC calculation unit, a second feature selection unit, and a second feature supplementation unit;

[0183] The number of the second feature selection units is C2, where C2 is greater than or equal to 2; specifically in this embodiment, C2 is 2, but it can also be 5.

[0184] The second feature selection module receives the cross-feature data and uses the second average MIC calculation unit to calculate the average MIC of the feature and the label, and the feature and other features. Based on this MIC, the above C2 second feature selection units run in parallel to select the corresponding number of features of class C2. The second feature supplementation unit supplements features based on the difference between the number of features after taking the union of the above C2 features and the number of features to be retained, and obtains the data to be classified and passes it to the model training module.

[0185] The model training module includes several classification models;

[0186] The model training module receives the data to be classified and puts it into several classification models for training. The model with the highest training accuracy is taken as the output of the model training module.

[0187] Figure 2 This is a flowchart illustrating the overall process of a classification method based on feature crossover and fusion feature selection according to the present invention; the method specifically includes:

[0188] S1. The collected speech is converted from analog to digital to obtain speech data, and each speech data is tagged according to the doctor's diagnosis.

[0189] The total number of voice data points was 546, collected from 78 subjects, including 24 healthy individuals and 54 patients with depression, with 7 voice data points collected from each subject.

[0190] S2. Extract several types of features from the speech data.

[0191] The number of feature types is 4, namely wavelet features based on Bior 2.4, wavelet features based on Bior 3.7, features based on RASTA filtering, and entropy features.

[0192] Each speech data point has 150 feature data points based on the wavelet features of Bior 2.4; each speech data point has 150 feature data points based on the wavelet features of Bior 3.7; each speech data point has 615 feature data points based on the features of RASTA filtering; each speech data point has 13 feature data points based on the entropy features; the total number of the aforementioned feature types is 928.

[0193] S3. Concatenate the feature data corresponding to each speech data, and then express the concatenated vector as a concatenated two-dimensional matrix.

[0194] The length of the concatenated vector is 928; the concatenated two-dimensional matrix is ​​a two-dimensional matrix with 546 rows and 928 columns, including the feature data of all 546 speech data; each row corresponds to one speech data and each column corresponds to one feature;

[0195] S4. Based on the labels, perform the first fusion feature selection on all features in the spliced ​​two-dimensional matrix to obtain the feature data after the first selection.

[0196] Figure 3 This is a flowchart of the adaptive feature crossing process of a classification method based on feature crossing and fusion feature selection according to the present invention.

[0197] The first fusion feature selection is specifically as follows:

[0198] S41. Calculate the maximum mutual information coefficient between the feature and the label in the feature data, and the average maximum information coefficient between the feature and other features;

[0199] The feature is the y-th feature, where y = 1, 2, ..., 928; the maximum mutual information coefficient with the label is MIC1. y MIC1 y The value range is 0 ≤ MIC1 y ≤1; the average maximum information coefficient with respect to other features is MIC2. y MIC2 y The value range is 0 ≤ MIC2 y ≤1.

[0200] S42. Calculate the mean of the maximum mutual information coefficients between all features and labels, and the mean of the average maximum information coefficients between the feature and other features.

[0201] The mean of the maximum mutual information coefficient with the tag The range of values ​​is The mean of the average information coefficients with other features The range of values ​​is

[0202] S43, based on S42 and Use the appropriate feature selection method to select and retain the feature data, and obtain the selected features;

[0203] Specifically, S43 is as follows:

[0204] S431. Select the corresponding number of features using different feature selection methods according to the number of feature data to be retained;

[0205] The desired number of feature data to be retained is 120; the number of feature selection methods is 2, namely a feature selection method based on Neighbor Component Analysis (NCA) and a feature selection method based on Minimum Redundancy Maximum Relevance (mRMR); the corresponding numbers are respectively... and indivual;

[0206] S432. Take the union of the features obtained using different feature selection methods in S431 to obtain the selected features to be supplemented.

[0207] The number of selected features to be supplemented is 120-m, where m represents the number of features repeatedly selected by different feature selection methods;

[0208] S433. Supplement the features of repeated selection in S432 to obtain the selected features, specifically:

[0209] like Alternatively, select m unique features from S431 after removing the features in NCA-based feature selection method to supplement the selection; otherwise, select m unique features from S431 after removing the features in MRMR-based feature selection method to supplement the selection.

[0210] S44. Select the feature data in S3 based on the selected features to obtain the data after the first selection;

[0211] The data after the first selection has 546 rows and 120 columns. Each row corresponds to one voice data point, and each column corresponds to one feature.

[0212] Figure 4 This is a flowchart of the cross-feature data selection process in a classification method based on feature cross-and fusion feature selection according to the present invention.

[0213] S5. Perform adaptive feature crossing on the data selected in S44 after the first selection to obtain the cross-linked feature data, specifically:

[0214] S51. Calculate the linear correlation coefficient between every two columns of feature data in the first selection data in S44, and then calculate the mean of all linear correlation coefficients.

[0215] The linear correlation coefficient r ij ,i,j=1,2,...,120,j≠i;r ij The range of values ​​for is 0 ≤ r ij ≤1; the mean of the linear correlation coefficient The range of values ​​is

[0216] S52, if in S51 If the value is less than the threshold S, calculate the Kronecker product of each row of data in S34 with itself; otherwise, first calculate the result of each row of data after passing through the nonlinear function, and then calculate the Kronecker product of the result with the row of data.

[0217] The threshold S is 0.5; the Kronecker product is a square matrix of dimension 120 in both cases;

[0218] S53. Arrange the elements of the square matrix corresponding to each row of data obtained in S52 into a row vector. Then the feature data corresponding to all physiological data can be represented by a two-dimensional matrix, which is called cross feature data.

[0219] The length of the row vector is 120. 2 The cross-feature data has 546 rows and 120 columns. 2 Each row corresponds to one speech data point, and each column corresponds to one cross feature.

[0220] S6. Based on the labels, perform a second fusion feature selection on the cross features of the cross-feature data obtained in S53 to obtain the second selected data.

[0221] The second fusion feature selection is specifically as follows:

[0222] S61. Calculate the maximum mutual information coefficient between the feature and the label in the feature data, and the average maximum information coefficient between the feature and other features.

[0223] The feature is the z-th feature, where z = 1, 2, ..., 120.2 The maximum mutual information coefficient between the tag and the target is MIC3. z MIC3 z The value range is 0 ≤ MIC3 z ≤1; the average maximum information coefficient with respect to other features is MIC4. z MIC4 z The value range is 0 ≤ MIC4 z ≤1;

[0224] S62. Calculate the mean of the maximum mutual information coefficients between the feature and the label in all feature data, and the mean of the average maximum information coefficients between the feature and other features.

[0225] The mean of the maximum mutual information coefficient with the tag The range of values ​​is The mean of the average information coefficients with other features The range of values ​​is

[0226] S63, based on S62 and Use the appropriate feature selection method to select and retain the feature data, and obtain the selected features;

[0227] Specifically, S63 is as follows:

[0228] S631. Select the corresponding number of features using different feature selection methods according to the number of feature data to be retained;

[0229] The desired number of feature data to be retained is 120; the number of feature selection methods is 2, namely a feature selection method based on Neighbor Component Analysis (NCA) and a feature selection method based on Minimum Redundancy Maximum Relevance (mRMR); the corresponding numbers are respectively... and indivual;

[0230] S632. Take the union of the features obtained using different feature selection methods in S631 to obtain the selected features to be supplemented.

[0231] The number of selected features to be supplemented is 120-l, where l represents the number of features repeatedly selected by different feature selection methods;

[0232] S633. Supplement the features of repeated selection in S632 to obtain the selected features, specifically:

[0233] like If the NCA-based feature selection method removes the features in S631, then the l non-repeating features are selected for supplementation; otherwise, if the mRMR-based feature selection method removes the features in S631, then the l non-repeating features are selected for supplementation.

[0234] S64. Select the cross feature data from S53 based on the selected features to obtain the data to be classified;

[0235] The data to be classified has 546 rows and 120 columns. Each row corresponds to one speech data point and each column corresponds to one feature.

[0236] S7. Divide the data to be classified obtained in S64 into a training set and a test set in a ratio of 8:2.

[0237] S8. Input the training set and labels from S7 into the classification learner to train the model, and select the model with the highest training accuracy for export.

[0238] S9. Input the test set obtained in S7 into the model obtained in S8 to obtain the classification result of each speech; count the classification results of 9 speech samples of each subject. If at least 5 of them are classified as depression, the subject is considered to have depression.

[0239] The classification results are compared with the labels to generate a confusion matrix. The proportion of people or speech entries with the same classification result as the label is calculated as the accuracy. The confusion matrices based on people and speech entries are shown in the figure. Figure 5 This invention relates to a classification device based on feature crossover and fusion feature selection, which presents a classification result using a confusion matrix and human statistics. Figure 5 In the middle, 0 represents mdd; 1 represents hc; all 11 people correctly judged the mdd category; all 5 people correctly judged the hc category; therefore, the accuracy rate based on the number of people is 100%. Figure 6 The classification result based on speech statistics, represented by a confusion matrix, of the classification method and apparatus of the present invention based on feature cross and fusion feature selection has an accuracy of 96.43% based on speech statistics. Figure 6 In the mdd class, 0 represents mdd; 1 represents hc. 73 voice samples in the mdd class were correctly identified; 4 voice samples were incorrectly identified. In the hc class, all 36 voice samples were correctly identified.

[0240] 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 classification device based on feature intersection and fusion feature selection, characterized in that, It includes a voice acquisition module, a feature extraction module, a first fusion feature selection module, an adaptive feature cross module, a second fusion feature selection module, and a model training module; The voice acquisition module converts the voice signal into voice data using digital-to-analog conversion and then transmits it to the feature extraction module. The feature extraction module extracts multiple features from the speech data and passes the resulting feature data to the first fusion feature selection module. The first fusion feature selection module includes a first average MIC calculation unit, a first feature selection unit, and a first feature supplementation unit. The first fusion feature selection module selects a predetermined number of features from the feature data and passes the first-selection data to the adaptive feature cross module. The adaptive feature cross module performs adaptive feature cross on the first-selection data and passes the resulting cross-linked feature data to the second fusion feature selection module. The second fusion feature selection module includes a second average MIC calculation unit, a second feature selection unit, and a second feature supplementation unit. The second fusion feature selection module selects a predetermined number of features from the cross-linked feature data and passes the resulting data to be classified to the model training module. The model training module uses the data to be classified to train the model and ultimately derives the classification model with the highest training accuracy. The adaptive feature cross module receives the data after the first selection, calculates the average linear correlation coefficient between features using the linear correlation coefficient calculation unit, and determines whether this linear correlation coefficient is greater than or equal to the threshold S. If not, i.e. less than the threshold S, the data after the first selection is passed as input data to the Kronecker product unit; if yes, i.e. greater than or equal to the threshold S, the data after the first selection is passed through the nonlinear transformation unit and then passed as input data to the Kronecker product unit.

2. The classification device based on feature intersection and fusion feature selection according to claim 1, characterized in that, The feature extraction module includes a wavelet feature extraction unit, a filter feature extraction unit, and an entropy feature extraction unit; the voice acquisition module includes a microphone and an analog-to-digital converter; the microphone in the voice acquisition module is used to acquire voice, and the voice acquired by the microphone is converted into voice data by the analog-to-digital converter and then transmitted to the feature extraction module.

3. The classification device based on feature intersection and fusion feature selection according to claim 1, characterized in that, The feature extraction module receives the speech data, and the feature extraction units operate in parallel to extract feature data of W groups of wavelet features, F groups of filter features and E groups of entropy features respectively. The number of wavelet feature extraction units is W, and W is greater than or equal to 2. The number of the filter feature extraction units is F, where F is greater than or equal to 1; The number of entropy feature extraction units is E, where E is greater than or equal to 1. The feature data of the wavelet features of group W, the filter features of group F, and the entropy features of group E are passed to the first fusion feature selection module.

4. A classification device based on feature intersection and fusion feature selection according to claim 1, characterized in that, The number of the first feature selection units is C1; The first fusion feature selection module receives the feature data, calculates the average MIC of the feature with the label and other features judged by the doctor using the first average MIC calculation unit, and the C1 first feature selection units run in parallel according to this MIC to select the corresponding number of features of class C1. The first feature supplementation unit supplements the features according to the difference between the number of features after taking the union of the features of class C1 and the number of features to be retained, and obtains the data after the first selection and passes it to the adaptive feature cross module.

5. A classification device based on feature intersection and fusion feature selection according to claim 1, characterized in that, The adaptive feature crossing module includes a Kronecker product unit; The Kronecker product unit performs the Kronecker product on the input data and the data after the first selection to obtain the cross-feature data, which is then passed to the second feature selection module.

6. A classification device based on feature intersection and fusion feature selection according to claim 1, characterized in that, The number of the second feature selection units is C2; The second fusion feature selection module receives the cross-feature data, calculates the average MIC of the feature and label, and the feature and other features using the second average MIC calculation unit, and the C2 second feature selection units run in parallel according to this MIC to select the corresponding number of features of class C2. The second feature supplementation unit supplements features according to the difference between the number of features after taking the union of the features of class C2 and the number of features to be retained, and obtains the data to be classified and passes it to the model training module. The model training module includes several classification models; The model training module receives the data to be classified and puts it into several classification models for training, and takes the model with the highest training accuracy as the output of the model training module.

7. A classification method based on feature intersection and fusion feature selection, wherein the classification method is the classification method executed by the classification device based on feature intersection and fusion feature selection as described in any one of claims 1 to 6, characterized in that, include: S1. Perform analog-to-digital conversion on the collected speech to obtain speech data and add tags to each piece of speech data; S2. Extract several types of features from the speech data; S3. Concatenate the feature data corresponding to each speech data to obtain a concatenated vector, which is then expressed as a concatenated two-dimensional matrix. S4. Perform the first fusion feature selection on all feature data in the spliced ​​two-dimensional matrix according to the labels to obtain the data after the first selection; S5. Perform adaptive feature crossing on the data after the first selection to obtain the cross-linked feature data; The adaptive feature crossing is as follows: if the linear correlation coefficient between features is greater than the threshold, a nonlinear transformation is performed first, followed by feature crossing; otherwise, feature crossing is performed directly. S6. Based on the labels, perform a second fusion feature selection on the cross features of the cross-feature data to obtain the data to be classified. S7. Divide the data to be classified into training set and test set according to the proportion; S8. Input the training set and labels into the classification learner to train the model, and select and export the model with the highest training accuracy. S9. Input the test set into the model obtained in S8 to get the classification result.

8. The classification method based on feature crossover and fusion feature selection according to claim 7, characterized in that, The first or second fusion feature selection specifically refers to: SA1, calculate the maximum mutual information coefficient between the feature and the label in the feature data, and the average maximum information coefficient between the feature and other features; SA2. Calculate the mean of the maximum mutual information coefficients between all features and labels, and the mean of the average maximum information coefficients between the feature and other features. SA3. Based on the results obtained from SA2, use the corresponding feature selection method to select and retain feature data, and obtain the selected features; SA4. Select cross-feature data based on the selected features to obtain the data to be classified.

9. A classification method based on feature intersection and fusion feature selection according to claim 8, characterized in that, The number of voice data items in S1 is X; the value of X is in the range of 2 ≤ X; The number of feature types mentioned in S2 is N; the value of N is 1≤N; The nth type feature of each voice data has There are 1 feature data, and the number of types of the feature data is the same as the number of feature types; wherein The total number of the aforementioned features is ; The length of the concatenated vector described in S3 is The spliced ​​two-dimensional matrix is ​​a two-dimensional matrix with X rows and Y columns, including feature data of all X speech data, with each row corresponding to one speech data and each column corresponding to one feature. The feature data mentioned in SA1 includes Y Each feature is denoted as . The maximum mutual information coefficient between the tag and the target is: ; The range of values ​​is The average maximum information coefficient with other features is: ; The range of values ​​is ; The mean of the maximum mutual information coefficients between SA2 and the tag ; The range of values ​​is The mean of the average information coefficients of the above and other features. ; The range of values ​​is ; The SA3 mentioned above is specifically: SA31. Select the corresponding number of features using different feature selection methods based on the number of feature data to be retained; The number of feature data to be retained is M; the value of M is in the range of 1≤M≤Y; The number of feature selection methods is It contains at least one method that does not consider the correlation between features and one method that does consider the correlation between features; The range of values ​​is The corresponding quantities are respectively and indivual; SA32. Take the union of the features obtained from SA31 using different feature selection methods to obtain the selected features to be supplemented. The number of selected features to be supplemented is Mm, where m represents the number of features repeatedly selected by different feature selection methods; SA33, supplementing the features of repeated selection in SA32, yields the selected features, specifically: if If m unique features are removed from S431 under the method of SA31 that does not consider the correlation between features, then m unique features are removed from S431 under the method of SA31 that considers the correlation between features. SA4 states that after the first selection, the number of rows in the data is X, and the number of columns is M. Each row corresponds to one speech data point, and each column corresponds to one feature.

10. A classification method based on feature intersection and fusion feature selection according to claim 7, characterized in that, S5 specifically refers to: S51, calculating the linear correlation coefficient between every two columns of feature data in the first selection of data, and then calculating the mean of all linear correlation coefficients; The linear correlation coefficient mentioned in S51 ; The range of values ​​is The mean of the linear correlation coefficients ; The range of values ​​is ; S52, if in S51 If the value is less than the threshold S, calculate the Kronecker product of each row of data in S34 with itself; otherwise, first calculate the result of each row of data after passing through the nonlinear function, and then calculate the Kronecker product of the result with the row of data. The threshold S has a range of 0 ≤ S ≤ 1; the Kronecker product is a square matrix of dimension M in both cases; S53. Arrange the elements of the square matrix corresponding to each row of data obtained in S52 into a row vector. Then the feature data corresponding to all physiological data can be represented by a two-dimensional matrix, which is called cross feature data. The length of the row vector is The number of rows and columns of the cross-feature data is X. Each row corresponds to one speech data point, and each column corresponds to one cross feature.

Images