A feature subset selection method and device, and a storage medium

Abstract

The application discloses a feature subset selection method and device and a storage medium. The method comprises the following steps: acquiring high-dimensional feature data; obtaining a feature importance list by using a random forest; obtaining a feature cross-correlation matrix by using Spearman correlation; obtaining a fitting degree of each feature according to a target GRU model; performing feature relationship type fusion on the feature importance list, the feature cross-correlation matrix and the fitting degree of each feature to obtain an overall evaluation value of each feature; continuously performing feature sorting on the overall evaluation value of each feature, removing the feature with the lowest evaluation value and determining whether the number of the remaining features is less than a required number of features until the number of the remaining features is not greater than the required number of features; and obtaining a feature subset. The feature subset selection method of the application uses Spearman correlation analysis, a random forest and a GRU for joint evaluation, completes diversified feature evaluation, realizes feature data dimension reduction, improves the feature evaluation processing flow and reduces the time complexity and the space complexity for subsequent work.

Description

Technical Field

[0001] This invention relates to the fields of machine learning and pattern recognition technology, and in particular to a method, apparatus and storage medium for selecting a feature subset. Background Technology

[0002] In recent years, multi-sensor data monitoring has been widely used in signal processing, industrial process control, image processing, and fully automated target tracking. The resulting massive amounts of data are crucial for describing the targets to be identified. Since the descriptive attributes (i.e., features) of targets are often high-dimensional and redundant, dimensionality reduction is necessary. Feature selection involves choosing a subset of features from a given dataset that is beneficial to subsequent algorithms, thereby achieving dimensionality reduction.

[0003] Therefore, how to select feature subsets is of great significance for dimensionality reduction of high-dimensional data based on multi-sensor data, and is a technical problem that needs to be solved urgently by those skilled in the art. Summary of the Invention

[0004] This invention provides a method, apparatus, and storage medium for selecting feature subsets. Through specific methods and strategies, Spearman correlation analysis, random forest, and GRU are used for joint evaluation to achieve diversified feature evaluation, thereby reducing the dimensionality of feature data, improving the feature evaluation process, and reducing the time and space complexity for subsequent work.

[0005] To address the aforementioned technical problems, embodiments of the present invention provide a method for selecting a feature subset, comprising:

[0006] Obtain high-dimensional feature data;

[0007] The high-dimensional feature data is processed to obtain a target random forest model, and a feature importance list is obtained based on the analysis and processing of the target random forest model.

[0008] The high-dimensional feature data is analyzed using Spearman correlation to obtain the feature cross-correlation matrix;

[0009] Based on the sliding window method, the columns of each feature in the feature cross-correlation matrix are processed to establish a prediction dataset for each feature;

[0010] The target GRU model is obtained based on the prediction dataset of each feature, and the fitting measure of each feature is obtained based on the analysis and processing of the target GRU model.

[0011] The feature importance list, the feature cross-correlation matrix, and each feature fitting metric are fused using a feature relationship model to obtain the overall evaluation value of each feature.

[0012] Continuously sort the overall evaluation values ​​of each feature, remove the feature with the lowest evaluation value, and determine whether the number of remaining features is less than the required number of features, until the number of remaining features is not greater than the required number of features.

[0013] The corresponding remaining feature data are identified as a feature subset.

[0014] As one preferred embodiment, the high-dimensional feature data is processed to obtain the target random forest model, specifically including:

[0015] The high-dimensional feature data is divided into a first training set and a first test set;

[0016] The first training set is imported into a random forest to train the model, and the random forest parameters are tuned using grid search and cross-validation.

[0017] The random forest model is evaluated using the first test set to obtain the target random forest model and the importance of each corresponding feature.

[0018] As one preferred embodiment, the construction process of the random forest includes:

[0019] The decision tree model is trained by randomly sampling N times with replacement from the first training set and using the corresponding N datasets.

[0020] Features are selected from the M attributes of the sample, and node splitting features are selected using either information gain or Gini coefficient strategies.

[0021] Repeat the above steps until the nodes are inseparable;

[0022] By ensemble N decision trees, a random forest is formed.

[0023] As one preferred embodiment, the analysis and processing based on the target random forest model to obtain a list of feature importance specifically includes:

[0024] The importance value of each feature in the random forest is obtained sequentially to obtain the feature importance list, and the importance value is represented by the following formula:

[0025] RandomForest→max{RF y}

[0026] Among them, RF y Let y be the importance value of the y-th feature in the random forest.

[0027] As one preferred approach, Spearman correlation is used to analyze the high-dimensional feature data to obtain the feature cross-correlation matrix, as shown in the following formula:

[0028] SpearmanRelation→min{R i，j},i≠j

[0029] Among them, R i，j Let be the correlation coefficient between the i-th feature and the j-th feature.

[0030] As one preferred approach, based on the analysis and processing of the target GRU model, the feature fitting metrics are obtained, as shown in the following formula:

[0031] GRU Prediction→max{R 2 -adjusted}

[0032] Among them, R 2 -adjusted is the adjusted coefficient of determination.

[0033] As a preferred embodiment, the feature importance list, the feature cross-correlation matrix, and each feature fitting metric are fused using a feature relationship model to obtain an overall evaluation value for each feature, specifically including:

[0034] Calculate using the following formula:

[0035]

[0036] Among them, F k SR is the overall evaluation value of the k-th feature; n is the number of features; k R for the k-th feature 2 -adjusted value; RF k R represents the importance value of the k-th feature in the random forest. k，j R is the correlation coefficient between the k-th feature and the j-th feature; i，k Let be the correlation coefficient between the i-th feature and the k-th feature;

[0037] Another embodiment of the present invention provides a feature subset selection device, comprising:

[0038] The data acquisition module is used to acquire high-dimensional feature data;

[0039] The feature importance list module is used to process the high-dimensional feature data to obtain the target random forest model, and to obtain the feature importance list based on the analysis and processing of the target random forest model.

[0040] The feature cross-correlation matrix module is used to analyze the high-dimensional feature data using Spearman correlation to obtain the feature cross-correlation matrix;

[0041] The feature prediction dataset module is used to process the column of each feature in the feature cross-correlation matrix according to the sliding window method to establish the feature prediction dataset.

[0042] Each feature fitting measurement module is used to obtain the target GRU model based on the prediction dataset of each feature, and to obtain the fitting measurement of each feature based on the analysis and processing of the target GRU model.

[0043] The overall evaluation value module for each feature is used to perform feature-relational fusion on the feature importance list, the feature cross-correlation matrix and each feature fitting measure to obtain the overall evaluation value of each feature.

[0044] The repeated filtering module is used to continuously sort the overall evaluation values ​​of each feature, remove the feature with the lowest evaluation value, and determine whether the number of remaining features is less than the required number of features, until the number of remaining features is not greater than the required number of features.

[0045] The feature selection module is used to determine the corresponding remaining feature data as a feature subset.

[0046] In another embodiment of the present invention, a computer-readable storage medium is provided, the computer-readable storage medium storing a computer program, wherein when the device containing the computer-readable storage medium executes the computer program, the method for selecting a feature subset as described above is implemented.

[0047] Compared with the prior art, the beneficial effects of the embodiments of the present invention are at least one of the following:

[0048] (1) Compared with the prior art, the present invention applies information fusion technology to feature selection algorithm;

[0049] (2) Existing feature selection algorithms only evaluate features in a relatively simple way, often using only the feature-label as a reference. This invention, however, innovatively combines Spearman correlation analysis and random forest, and subsequently introduces the GRU prediction model to evaluate single feature data (R^2-adjusted), thus further improving feature evaluation.

[0050] (3) This invention introduces a new evaluation mechanism that integrates feature-label information, information between features, and information within each feature, achieving the goal of accurately removing redundant and mixed features, and reducing the time and space complexity for subsequent work.

[0051] (4) The number of iterations can be specified by the user while maintaining a high classification accuracy. This method can be adjusted by users, especially experienced users, according to specific circumstances.

[0052] (5) In the GRU feature evaluation module, the model framework is adjustable to adapt to different datasets and represent sensitive features;

[0053] (6) Before executing the algorithm of this invention, features can be constructed through user-defined methods, which increases the convenience and interpretability for users. Attached Figure Description

[0054] Figure 1 This is a flowchart illustrating a method for selecting a feature subset in one embodiment of the present invention.

[0055] Figure 2 This is a schematic diagram of the internal structure of the GRU in one embodiment of the present invention;

[0056] Figure 3 This is a logic block diagram of a method for selecting a feature subset in one embodiment of the present invention;

[0057] Figure 4 This is a schematic diagram of the structure of a feature subset selection device in one embodiment of the present invention;

[0058] Figure label:

[0059] The modules are as follows: 11. Data acquisition module; 12. Feature importance list module; 13. Feature cross-correlation matrix module; 14. Prediction dataset for each feature module; 15. Fit measurement module for each feature module; 16. Overall evaluation value module for each feature module; 17. Repeated filtering module; 18. Feature selection module. Detailed Implementation

[0060] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. The purpose of providing these embodiments is to make the disclosure of the present invention more thorough and comprehensive. Based on the embodiments of the present invention, all other embodiments obtained by those of ordinary skill in the art without creative effort are within the scope of protection of the present invention.

[0061] In the description of this application, the terms "first," "second," "third," etc., are used for descriptive purposes only and should not be construed as indicating or implying relative importance or implicitly specifying the number of indicated technical features. Therefore, a feature defined with "first," "second," "third," etc., may explicitly or implicitly include one or more of that feature. In the description of this application, unless otherwise stated, "a plurality of" means two or more.

[0062] In the description of this application, it should be noted that, unless otherwise expressly specified and limited, the terms "installation," "connection," and "linking" should be interpreted broadly. For example, they can refer to fixed connections, detachable connections, or integral connections; they can refer to mechanical connections or electrical connections; they can refer to direct connections or indirect connections through an intermediate medium; and they can refer to the internal communication between two components. The terms "vertical," "horizontal," "left," "right," "upper," "lower," and similar expressions used herein are for illustrative purposes only and do not indicate or imply that the device or component referred to must have a specific orientation, or be constructed and operated in a specific orientation, and therefore should not be construed as limiting the invention. The term "and / or" as used herein includes any and all combinations of one or more of the associated listed items. Those skilled in the art can understand the specific meaning of the above terms in this application based on the specific circumstances.

[0063] In the description of this application, it should be noted that, unless otherwise defined, all technical and scientific terms used in this invention have the same meaning as commonly understood by one of ordinary skill in the art to which this invention pertains. The terminology used in this specification is for the purpose of describing specific embodiments only and is not intended to limit the invention. Those skilled in the art can understand the specific meaning of the above terms in this application based on the specific circumstances.

[0064] It should be noted in advance that existing feature selection methods can be categorized into feature selection using a filter evaluation strategy, feature selection using a wrapper evaluation strategy, and embedded feature selection methods, depending on the specific strategy employed. To highlight the beneficial effects of this invention, the existing technologies will be described in detail below:

[0065] (1) Filtering methods. Filtering methods evaluate each feature based on the inherent properties of the data, such as divergence and correlation, and introduce indicators such as thresholds. For example, the variance selection method first calculates the variance of each feature, and then selects features with variances greater than the threshold. Its advantages are simple algorithm and strong scalability.

[0066] (2) Wrapper Method. The Wrapper method detects the interaction between two or more features and uses a specific objective function to assist in feature selection. It selects or removes some features at a time, repeating this process until the desired number of features is obtained. For example, recursive feature elimination uses a base model for multiple rounds of training. After each round, it eliminates some features with lower weights based on the `coef_` or `feature_importances_` returned by the learner, and then performs the next round of training based on the new feature set. The Wrapper method is theoretically more accurate than the Filter method.

[0067] (3) Embedded methods. Embedded methods use machine learning models to obtain the weights of each feature, and then use wrapper methods for feature selection. For example, Recursive Feature Elimination based on Random Forest Estimation (RF-RFE) uses a random forest estimator to evaluate the importance of features, obtains a ranked list, removes tail features, and continues training with the retained features until the optimal feature subset is obtained. Embedded methods embed feature selection into the model construction process, combining the advantages of wrapper methods and machine learning algorithms, as well as the high computational efficiency of filtering methods.

[0068] In addition to the methods mentioned above, other methods such as Laplace score (LS) and principal component analysis (PCA) can also select appropriate features based on their feature evaluation methods.

[0069] In summary, feature selection methods prioritize features based on importance, redundancy, and complementarity. However, some existing methods are computationally complex and may perform poorly under certain conditions, exhibiting limited generalization ability. Therefore, research and practical applications of feature selection for multi-sensor data streams are essential.

[0070] Currently, the superior cloud computing capabilities and information transmission efficiency enable certain online monitoring and decision-making systems (such as bearing fault early warning systems) to provide more timely and accurate warnings. They also support more complex algorithms, yielding accurate feature subsets with a smaller number of elements. However, some feature selection algorithms interpret and evaluate features from a single perspective, which can easily lead to erroneous feature deletion or selection, thus affecting the system's judgment and causing unnecessary losses. The following sections will explain each of these methods in detail.

[0071] (1) The Filter method only explores the data itself and its properties, lacking a reasonable interpretation of certain data indicators. In addition, there is no scientific guidance for the selection of thresholds, resulting in a lack of completeness in the evaluation of various features. Therefore, the Filter method has great limitations.

[0072] (2) The Wrapper method is theoretically more accurate than the Filter method, but it has a high time complexity because it requires multiple rounds of training and iterations. Secondly, the choice of objective function has a significant impact on the model's capabilities, but setting the objective function is not easy and the algorithm is complex. Finally, this method is prone to overfitting when there are insufficient samples.

[0073] (3) The Embedded method has similar disadvantages to the Wrapper method, with high model space and time complexity, especially when using ensemble learning methods as feature importance estimators; secondly, since the sensitivity of labels to features varies across datasets, the estimator needs to be evaluated; finally, a single estimator only evaluates the feature-label correlation, lacking the issues of the features themselves and the relationships between features.

[0074] In summary, relying solely on feature-to-feature correlations or feature-label correlations is insufficient to fully represent the "whole picture" of features, failing to consider the relationship between redundancy and feature importance, and easily leading to misselection. Furthermore, high-dimensional data often brings the "curse of dimensionality." Redundant features also consume memory space, impacting storage and computational efficiency.

[0075] While methods such as Filter, Wrapper, and Embedded, based on their respective perspectives and incorporating correlation theory, may achieve certain results in some situations, their effectiveness can be significantly impacted by redundancy between features or large differences between features themselves. This can lead to poor generalization ability of the trained model and consequently affect the system's early warning capabilities. Therefore, this invention aims to address the incompleteness of the feature evaluation system, more scientifically resolve the "curse of dimensionality" problem in multi-feature scenarios, optimize the diversity of feature evaluation, and improve the fusion methods of various data information to maintain a high accuracy benchmark and rationally select features.

[0076] One embodiment of the present invention provides a method for selecting a feature subset. For details, please refer to [link to relevant documentation]. Figure 1 , Figure 1 The diagram shows a flowchart of a method for selecting a feature subset in one embodiment of the present invention, which includes steps S1 to S8, as follows:

[0077] S1. Obtain high-dimensional feature data;

[0078] S2. Process the high-dimensional feature data to obtain the target random forest model, and obtain the feature importance list based on the analysis and processing of the target random forest model;

[0079] S3. Analyze the high-dimensional feature data using Spearman correlation to obtain the feature cross-correlation matrix;

[0080] S4. Using the sliding window method, process the columns of each feature in the feature cross-correlation matrix to establish a prediction dataset for each feature.

[0081] S5. Obtain the target GRU model based on the predicted dataset of each feature, and obtain the fitting metric for each feature based on the analysis and processing of the target GRU model.

[0082] S6. Perform feature relationship fusion on the feature importance list, the feature cross-correlation matrix and each feature fitting measure to obtain the overall evaluation value of each feature;

[0083] S7. Continuously sort the overall evaluation values ​​of each feature, remove the feature with the lowest evaluation value, and determine whether the number of remaining features is less than the required number of features, until the number of remaining features is not greater than the required number of features.

[0084] S8. Determine the corresponding remaining feature data as a feature subset.

[0085] The feature subset selection method provided by this invention rationally selects features and improves the system's decision-making ability. This method considers the correlation between features and labels, redundancy and correlation between features, and the interpretability of single feature data using random forest, Spearman correlation analysis, and GRU model, respectively. It also defines a new evaluation method to fuse information, weigh feature contributions, and select suitable feature subsets.

[0086] It should be noted that the type of feature data in this invention is determined by the actual application scenario. For example, in the field of bearing monitoring, the feature data can be vibration acceleration signals, which will not be elaborated here.

[0087] The main process of the method of the present invention is divided into three steps: feature correlation calculation (corresponding to steps S1 to S5), information fusion (corresponding to step S6), and iteration (corresponding to step S7).

[0088] Feature correlation calculation mainly consists of three aspects: Random Forest, Spearman correlation, and GRU feature evaluation, which correspond to feature-label correlation, inter-feature correlation, and feature self-evaluation, respectively. These are explained in detail below:

[0089] I. Random Forest

[0090] Feature selection for high-dimensional data often revolves around judging feature inclusion based on feature importance, while random forests can select a suitable subset of features based on the importance of each variable (feature), and are therefore widely used. Its basic principle is as follows:

[0091] Random forests consist of multiple decision tree classifiers. The algorithm randomly samples data and applies it to decision trees, with the final result output based on a 'voting' method.

[0092] The construction process of a random forest is as follows:

[0093] (1) Randomly sample N times with replacement and train a decision tree model using N datasets;

[0094] (2) Randomly select features from the M attributes of the sample (select features), and use strategies such as information gain or Gini coefficient as node splitting features;

[0095] (3) Repeat step (2) until the nodes are inseparable;

[0096] (4) Integrate the N decision trees to form a random forest.

[0097] For a trained random forest model, if the number of variables is n, then the importance of each variable (feature) can be represented by the matrix [RF1, RF2, RF3, ..., RF]. n ]express.

[0098] II. Spearman Correlation

[0099] The Spearman correlation coefficient, also known as the Spearman rank correlation, is an indicator that measures the correlation between two variables using a monotonic equation (with redundancy removed). The algorithm sorts the two variables in ascending order and assigns ranks, representing each rank. Then, it calculates the Spearman correlation coefficient by differencing the ranks and applying the formula. The formula is as follows:

[0100]

[0101] d i =x i -y i #(2)

[0102] In the formula: n is the number of data; x i Let y be the rank of the i-th element in X. i Similarly); d i It is a set of rank differences.

[0103] For n variables (features), the pairwise correlation between any two variables can be represented by matrix R, as shown below:

[0104]

[0105] According to Spearman's definition, matrix R is a symmetric matrix with all its main diagonals equal to 1.

[0106] III. GRU Feature Evaluation

[0107] For details, please see Figure 2 , Figure 2The diagram shown illustrates the internal structure of a GRU in one embodiment of the present invention. A Gated Recurrent Unit (GRU) is a type of neural network used for prediction tasks and is also a type of Recurrent Neural Network (RNN).

[0108] GRU makes memory h t Integrating long and short-term memories (including past information h) t-1 And current information The forgetting and selective memory (weighting) are achieved through the update gate. Figure 1 Chinese r t and z t These represent the reset gate and the update gate, respectively. t Used to determine the state information of the previous moment and write it into the candidate set. The amount of information in time, z t This is used to determine the amount of information imported when importing the current state from the current state. Because r t and z t Both use the sigmoid function for activation, therefore their values ​​are both between 0 and 1. The gate structure principle and forward propagation of the GRU model can be expressed by the following formula:

[0109]

[0110] In the formula: σ(x) is the sigmoid activation function; w r w z W represents the model parameters to be trained, in matrix form; x t σ(x) represents the input at the current time point; tanh(x) is the tanh activation function. The formulas for σ(x) and tanh(x) are:

[0111]

[0112] This paper uses the GRU model to fit the feature data (the column containing the feature) to adjust the coefficient of determination (R²). 2 The adjusted value was evaluated and used as the initial evaluation value for the feature. 2 The adjusted formula is as follows:

[0113]

[0114] In the formula: RSS is the residual sum of squares, which is the sum of the squares of the differences between the observed values ​​and the predicted values; TSS is the population sum of squares, which is the sum of the squares of the differences between the observed values ​​and the observed mean; n is the sample size; k is the sample length. The formulas for RSS and TSS are:

[0115]

[0116]

[0117] In the formula: y i For the i-th observation; This is the i-th predicted value; This represents the mean of all observed samples.

[0118] For ease of expression, R will be used. 2 -adjusted is replaced by SR. For n variables (features), the initial evaluation value of each variable can be obtained from the matrix [SR1, SR2, ..., SR...]. n ]express.

[0119] Of the three components of the method proposed in this paper—random forest, Spearman correlation analysis, and GRU model—the output of each stage is the key, and the fusion of information after the output of each stage is even more crucial, as it can more comprehensively evaluate each feature.

[0120] Under multiple data categories, random forests maintain high accuracy while being less prone to overfitting, and they can provide estimates of the importance of each variable (feature) in the classification, reflecting the degree of importance of each feature to the label. Therefore, the desired result for the feature importance obtained from random forests is:

[0121] RandomForest→max{RF y}#(8)

[0122] Where: RF y Let y be the importance (value) of the y-th feature in the random forest.

[0123] Spearman correlation is a non-parametric statistical indicator of the correlation between two variables. It has no data distribution requirements, avoiding the reliability issues caused by inconsistencies between data distribution characteristics and the usage method, thus offering greater applicability. Its output reflects a certain amount of redundancy between the two variables; therefore, the desired correlation coefficient between the two variables is:

[0124] SpearmanRelation→min{R i，j},i≠j#(9)

[0125] In the formula: R i，j Let be the correlation coefficient between the i-th feature and the j-th feature.

[0126] GRU can be trained on large amounts of data to obtain a prediction model with good generalization and robustness. Compared with RNN, it solves the problem of subsequent nodes having poor perception of information from previous nodes, while also achieving higher accuracy. Based on this, the fitting effect of the GRU model on the feature data (the column containing the feature) is used to adjust the coefficient of determination (R²). 2 The adjusted coefficient of determination (-adjusted) evaluates the features. It reflects the disorder and heterogeneity within the sequence containing the feature. Therefore, the desired result of adjusting the coefficient of determination is:

[0127] GRU Prediction→max{R 2 -adjusted}#(10)

[0128] To ensure that each component meets the above requirements, and to integrate the components while balancing feature importance and reducing feature redundancy and clutter, the following formula is designed:

[0129]

[0130] In the formula: F k SR is the overall model evaluation value for the k-th feature; n is the number of features; k R for the k-th feature 2 -adjusted value.

[0131] Then, the absolute value of the overall evaluation value of each feature is taken, and the process is repeated iteratively by sorting and removing tail features until the number of remaining features equals the required number of features.

[0132] In summary, please refer to Figure 3 , Figure 3 The diagram shown is a logical block diagram of a method for selecting a feature subset in one embodiment of the present invention, which includes the following steps 1-5.

[0133] Step 1: First, split the dataset into training and testing sets. Use the training set to import the random forest to train the model and fine-tune the random forest parameters using grid search and cross-validation.

[0134] Step 2: Further, the random forest model is evaluated using a test set, and the final model and the importance of each feature are obtained.

[0135] Step 3: Simultaneously, on the one hand, Spearman correlation analysis is used to obtain the feature cross-correlation matrix; on the other hand, each feature's column is used to construct a prediction dataset for each feature via a sliding window method, and this dataset is divided into a training set and a test set. The training set is used to import the GRU model to train the model, and then the test set is used to import the trained model and calculate the adjusted coefficient of determination (R²). 2-adjusted), to complete the fitting measurement of each feature.

[0136] Step 4: Use formula (11) to perform feature relationship fusion on the obtained feature importance list, feature cross-correlation matrix and each feature fitting measure to obtain the overall evaluation (value) of each feature.

[0137] Step 5: Then, take the absolute value of the final evaluation value of each feature and sort the features. Remove the feature with the lowest evaluation value and determine whether the number of retained features is less than the required number of features. If it is not greater than the required number of features, the feature selection is complete; otherwise, combine the retained features and perform the next round of feature evaluation and deletion.

[0138] Another embodiment of the present invention provides a feature subset selection device; for details, please refer to [link to relevant documentation]. Figure 4 , Figure 4 The diagram shown illustrates the structure of a feature subset selection device according to one embodiment of the present invention, which includes:

[0139] Data acquisition module 11 is used to acquire high-dimensional feature data;

[0140] The feature importance list module 12 is used to process the high-dimensional feature data to obtain a target random forest model, and to obtain a feature importance list based on the analysis and processing of the target random forest model.

[0141] The feature cross-correlation matrix module 13 is used to analyze the high-dimensional feature data using Spearman correlation to obtain the feature cross-correlation matrix;

[0142] The feature prediction dataset module 14 is used to process the column of each feature in the feature cross-correlation matrix according to the sliding window method to establish the feature prediction dataset.

[0143] Each feature fitting measurement module 15 is used to obtain a target GRU model based on the prediction dataset of each feature, and to obtain a feature fitting measurement based on the analysis and processing of the target GRU model.

[0144] The overall evaluation value module 16 for each feature is used to perform feature relationship fusion on the feature importance list, the feature cross-correlation matrix and each feature fitting measure to obtain the overall evaluation value of each feature.

[0145] The repeated filtering module 17 is used to continuously sort the overall evaluation values ​​of each feature, remove the feature with the lowest evaluation value, and determine whether the number of remaining features is less than the required number of features, until the number of remaining features is not greater than the required number of features.

[0146] Feature selection module 18 is used to determine the corresponding remaining feature data as a feature subset.

[0147] In another embodiment of the present invention, a computer-readable storage medium is provided, the computer-readable storage medium storing a computer program, wherein when the device containing the computer-readable storage medium executes the computer program, the method for selecting a feature subset as described above is implemented.

[0148] Compared with existing technologies, this invention applies information fusion technology to feature selection algorithms. Existing feature selection algorithms only evaluate features in a relatively singular way, often using only feature-label as a reference. This invention, however, innovatively combines Spearman correlation analysis and random forest, and subsequently introduces a GRU prediction model to evaluate single feature data (R^2-adjusted), further improving feature evaluation. This invention introduces a new evaluation mechanism that integrates feature-label information, inter-feature correlation information, and internal information of each feature, achieving accurate removal of redundant and mixed features, reducing time and space complexity for subsequent work. This invention allows for user-specified iteration counts while maintaining high classification accuracy. This method allows users, especially experienced users, to adjust the number of iterations according to specific circumstances. In the GRU feature evaluation module, the model framework is adjustable to adapt to different datasets and represent sensitive features. Before executing the algorithm of this invention, features can be constructed using user-defined methods, increasing user convenience and interpretability.

[0149] The embodiments described above are merely illustrative of several implementations of the present invention, and while the descriptions are specific and detailed, they should not be construed as limiting the scope of the present invention. It should be noted that those skilled in the art can make various modifications and improvements without departing from the concept of the present invention, and these modifications and improvements all fall within the scope of protection of the present invention. Therefore, the scope of protection of this patent should be determined by the appended claims.

Claims

1. A method for selecting a feature subset, characterized in that, include: Acquire high-dimensional feature data for the bearing fault early warning system; The high-dimensional feature data includes vibration acceleration signals; The high-dimensional feature data is processed to obtain a target random forest model, and a feature importance list is obtained based on the analysis and processing of the target random forest model. The high-dimensional feature data is analyzed using Spearman correlation to obtain the feature cross-correlation matrix; Based on the sliding window method, the columns of each feature in the feature cross-correlation matrix are processed to establish a prediction dataset for each feature; The target GRU model is obtained based on the prediction dataset of each feature, and the fitting measure of each feature is obtained based on the analysis and processing of the target GRU model. The feature importance list, the feature cross-correlation matrix, and each feature fitting metric are fused using a feature relationship model to obtain an overall evaluation value for each feature; wherein, the specific calculation method for fusion of the feature importance list, the feature cross-correlation matrix, and each feature fitting metric to obtain an overall evaluation value for each feature includes: in, For the first The overall evaluation value of each feature; The number of features; For the first Features value, To adjust the coefficient of determination; For the first The importance value of each feature in the random forest; For the first The first feature and the second The correlation coefficient between the features; For the first The first feature and the second Correlation coefficients between features; Continuously sort the overall evaluation values ​​of each feature, remove the feature with the lowest evaluation value, and determine whether the number of remaining features is less than the required number of features, until the number of remaining features is not greater than the required number of features. The corresponding remaining feature data are identified as a feature subset.

2. The method for selecting a feature subset as described in claim 1, characterized in that, The high-dimensional feature data is processed to obtain the target random forest model, specifically including: The high-dimensional feature data is divided into a first training set and a first test set; The first training set is imported into a random forest to train the model, and the random forest parameters are tuned using grid search and cross-validation. The random forest model is evaluated using the first test set to obtain the target random forest model and the importance of each corresponding feature.

3. The method for selecting a feature subset as described in claim 2, characterized in that, The construction process of the random forest includes: The decision tree model is trained by randomly sampling N times with replacement from the first training set and using the corresponding N datasets. Features are selected from the M attributes of the sample, and node splitting features are selected using either information gain or Gini coefficient strategies. Repeat the above steps until the nodes are inseparable; By ensemble N decision trees, a random forest is formed.

4. The method for selecting a feature subset as described in claim 1, characterized in that, The analysis and processing based on the target random forest model yields a list of feature importance, specifically including: The importance value of each feature in the random forest is obtained sequentially to obtain the feature importance list, and the importance value is represented by the following formula: in, For the first The importance value of each feature in the random forest.

5. The method for selecting a feature subset as described in claim 4, characterized in that, The high-dimensional feature data is analyzed using Spearman correlation to obtain the feature cross-correlation matrix, as shown in the following formula: in, For the first The first feature and the second The correlation coefficient between the features.

6. The method for selecting a feature subset as described in claim 5, characterized in that, Based on the analysis and processing of the target GRU model, the fitting metrics for each feature are obtained, as shown in the following formula: in, To adjust the coefficient of determination.

7. A device for selecting a feature subset, characterized in that, include: The data acquisition module is used to acquire high-dimensional feature data of the bearing fault early warning system; The high-dimensional feature data includes vibration acceleration signals; The feature importance list module is used to process the high-dimensional feature data to obtain the target random forest model, and to obtain the feature importance list based on the analysis and processing of the target random forest model. The feature cross-correlation matrix module is used to analyze the high-dimensional feature data using Spearman correlation to obtain the feature cross-correlation matrix; The feature prediction dataset module is used to process the column of each feature in the feature cross-correlation matrix according to the sliding window method to establish the feature prediction dataset. Each feature fitting measurement module is used to obtain the target GRU model based on the prediction dataset of each feature, and to obtain the fitting measurement of each feature based on the analysis and processing of the target GRU model. The overall evaluation value module for each feature is used to perform feature-relationship fusion on the feature importance list, the feature cross-correlation matrix, and each feature fitting metric to obtain an overall evaluation value for each feature; wherein, the specific calculation method for performing feature-relationship fusion on the feature importance list, the feature cross-correlation matrix, and each feature fitting metric to obtain an overall evaluation value for each feature includes: in, For the first The overall evaluation value of each feature; The number of features; For the first Features value, To adjust the coefficient of determination; For the first The importance value of each feature in the random forest; For the first The first feature and the second Correlation coefficients between features; For the first The first feature and the second Correlation coefficients between features; The repeated filtering module is used to continuously sort the overall evaluation values ​​of each feature, remove the feature with the lowest evaluation value, and determine whether the number of remaining features is less than the required number of features, until the number of remaining features is not greater than the required number of features. The feature selection module is used to determine the corresponding remaining feature data as a feature subset.

8. A computer-readable storage medium, characterized in that, The computer-readable storage medium stores a computer program, wherein when the device containing the computer-readable storage medium executes the computer program, it implements the method for selecting a feature subset as described in any one of claims 1 to 6.

Images