Iot multi-view feature selection method and system based on similar matrix fusion

Abstract

The application discloses an Internet of Things multi-view feature selection method and system based on similarity matrix fusion, and relates to the technical field of machine learning.The method comprises the following steps: firstly, incomplete data of each view is spliced into a wide table according to columns, and a complete matrix of each view is constructed through sampling and padding; secondly, the data of each view is clustered multiple times, and corresponding similarity matrices are constructed; all the similarity matrices are stacked into a tensor along the third dimension, an adaptive weight fusion is introduced under a low-rank constraint, a unified common similarity matrix is learned for feature value decomposition, a feature vector corresponding to a positive feature value is extracted to form a pseudo label; a high-confidence sample is screened in combination with a pace learning strategy, and each view feature selection coefficient and a self-pace weight are jointly optimized; and importance scores are calculated based on the features of each view to obtain an optimal subset.The application realizes feature selection of incomplete multi-view data of an industrial Internet of Things through similarity matrix tensor fusion, adaptive low-rank constraint and joint optimization of pace learning.

Description

Technical Field

[0001] This invention relates to the field of machine learning technology, and more specifically to a method and system for multi-view feature selection in the Internet of Things based on similarity matrix fusion. Background Technology

[0002] With the rapid development of Industrial Internet of Things (IIoT) technology, a large number of sensors, monitoring devices, and smart terminals have been deployed in industrial production environments, generating massive amounts of multi-source heterogeneous data. This data often comes from different types of acquisition devices (such as temperature sensors, vibration sensors, image acquisition devices, sound monitors, etc.), forming a typical multi-view data structure, providing a rich source of information for industrial equipment condition monitoring, fault diagnosis, and production optimization.

[0003] However, in real-world industrial scenarios, multi-view data often suffers from incompleteness due to sensor malfunctions, network transmission interruptions, and asynchronous data acquisition, meaning that data from some views is missing. This incompleteness poses a significant challenge to traditional multi-view learning methods, directly impacting the accuracy and reliability of feature selection.

[0004] Multi-view feature selection, as an important method for reducing data dimensionality and extracting key features, can effectively remove redundant information and improve the efficiency and performance of subsequent machine learning tasks. However, existing multi-view feature selection methods have the following shortcomings:

[0005] First, most existing methods assume that all view data is complete and available, which cannot effectively handle the data gap problem that is common in the Industrial Internet of Things, resulting in a significant performance degradation in practical applications;

[0006] Second, when integrating multi-view information, traditional methods often simply apply equal weights or independent processing to each view, ignoring the complementarity and consistency between different views and failing to fully explore the potential correlation structure of multi-view data.

[0007] Third, existing similarity matrix construction methods mostly use fixed similarity measurement methods, lack adaptive learning capabilities, and are difficult to accurately characterize the complex relationships between samples in industrial data;

[0008] Fourth, when dealing with incomplete multi-view data, existing methods lack effective learning strategies to balance the learning process of samples with different difficulties, and are easily affected by noisy data and missing data, which affects the robustness and generalization ability of the model.

[0009] Therefore, there is an urgent need for a multi-view feature selection method that can effectively handle incomplete data from the Industrial Internet of Things, fully integrate multi-view similarity information, and has adaptive learning capabilities. Summary of the Invention

[0010] To address the aforementioned issues, this invention proposes a method and system for selecting multi-view features in IoT based on similarity matrix fusion. This invention achieves robust, consistent, and interpretable screening of key features in incomplete industrial IoT data through multi-view similarity fusion, tensor low-rank modeling, pace learning guidance, and joint feature scoring.

[0011] On the one hand, IoT multi-view feature selection methods based on similarity matrix fusion include:

[0012] S1, concatenate each incomplete industrial IoT data column by column to obtain a wide table data matrix, and extract and supplement the wide table data matrix to construct the data matrix of each industrial IoT view;

[0013] S2, perform multiple clustering operations on the data matrix of each Industrial IoT view to construct a similarity matrix for each view;

[0014] S3 stacks the similarity matrices of each view into a third-order tensor and applies tensor learning constraints to capture the global low-rank structure across views. At the same time, adaptive view weights are introduced to perform weighted fusion of the global low-rank structure across views to obtain the common similarity matrix.

[0015] S4. Eigenvalue decomposition is performed on the common similarity matrix to extract the pseudo-label matrix. Step learning guides the construction of a multi-view feature selection model for incomplete industrial IoT data.

[0016] S5 calculates the importance score of each industrial IoT view feature based on the multi-view feature selection model, sorts all features in descending order of score, and selects the top preset number of features as the final feature subset.

[0017] Furthermore, S1, specifically, is as follows:

[0018] Incomplete data from each Industrial IoT Concatenate the columns into a wide table, as follows:

[0019] ;

[0020] ;

[0021] in, Represents a wide table containing multi-view data; Indicates the total dimensions of the view; Indicates the dimension of a single view v; 'v' indicates the overlay of multiple view dimensions; 'v' indicates a single view.

[0022] The data matrix of the multi-view wide table is extracted and supplemented to construct the data matrix for each Industrial IoT view, as follows:

[0023] ;

[0024] in, This represents the data matrix for each Industrial IoT view after extraction and supplementation.

[0025] Furthermore, in S2, the following formula is used to perform multiple clustering calculations on the data matrix of each Industrial IoT view:

[0026] ;

[0027] Multiple clustering operations can yield k clustering results. Then, the similarity matrix is ​​constructed from the results of the k clustering operations, as follows:

[0028] ;

[0029] in, The matrix represents each view; U represents the clustering result; This indicates transpose.

[0030] Furthermore, in S3, the formulas for calculating the similarity matrix are as follows:

[0031] ;

[0032] in, This represents a tensor obtained by stacking the similarity matrices of each view along the third dimension; This represents the shared similarity matrix learned uniformly. Indicates the weighted parameters; This represents the F-norm.

[0033] Furthermore, in S4, eigenvalue decomposition is performed on the shared similarity matrix to extract the pseudo-label matrix. The calculation formula is as follows:

[0034] ;

[0035] ;

[0036] ;

[0037] in, Represents the pseudo-label matrix; This represents the first largest eigenvalue; This represents the c-th largest eigenvalue; This indicates that the elements within the brackets form a diagonal matrix; This indicates taking the maximum value; This represents a matrix composed of the first c eigenvectors; Represents the square root of a diagonal matrix;

[0038] A multi-view feature selection model is constructed by guiding incomplete industrial IoT data through pace learning. The calculation formula is as follows:

[0039] ;

[0040] ;

[0041] in The pace learning ratio parameter for each IoT view; Indicates the initial parameters of self-synchronization; Indicates the maximum self-synchronizing parameter; Represents the self-synchronization matrix; Represents the feature selection matrix; Represents the pseudo-label matrix; Self-synchronization parameters for each data point; This indicates data selected through auto-stepping; This indicates a constraint.

[0042] Furthermore, in S5, all features are sorted in descending order of score, and a predetermined number of features are selected as the final feature subset, specifically including:

[0043] Calculate the importance score of each feature in each IoT view. :

[0044] ;

[0045] in, This represents the j-th row of the feature selection matrix for the v-th IoT view;

[0046] Summarize the feature importance scores of all IoT views:

[0047] ;

[0048] Sort all feature scores in T2 in descending order, and select the top m features as the final feature subset:

[0049] ;

[0050] Where m is the preset number of feature selections.

[0051] On the other hand, an IoT multi-view feature selection system based on similarity matrix fusion includes:

[0052] The extraction and supplementation module is used to concatenate each incomplete industrial IoT data column by column to obtain a wide table data matrix, and to extract and supplement the wide table data matrix to construct the data matrix of each industrial IoT view.

[0053] The clustering module is used to perform multiple clustering operations on the data matrix of each Industrial IoT view to construct a similarity matrix for each view;

[0054] The weighted fusion module is used to stack the similarity matrix of each view into a third-order tensor and apply tensor learning constraints to capture the global low-rank structure across views. At the same time, adaptive view weights are introduced to perform weighted fusion of the global low-rank structure across views to obtain the common similarity matrix.

[0055] The step-guided module is used to extract pseudo-label matrices by eigenvalue decomposition of shared similar matrices and to construct multi-view feature selection models from incomplete industrial IoT data through step-learning.

[0056] The feature selection module is used to calculate the importance score of each industrial IoT view feature based on the multi-view feature selection model, sort all features in descending order of score, and select the top preset number of features as the final feature subset.

[0057] The present invention adopts the above technical solution and has the following beneficial effects:

[0058] (1) This invention uses tensor-quantized similarity matrix modeling and adaptive weighted low-rank fusion to explicitly mine the shared structure and complementary information among multiple views, avoid view bias, generate a more consistent and globally discriminative common similarity representation, and provide reliable guidance for subsequent pseudo-label generation and feature selection;

[0059] (2) This invention combines wide table splicing and sampling mechanism with step learning to dynamically screen high-confidence samples, effectively alleviating the common missing, noise and heterogeneous interference in the industrial Internet of Things, and improving the stability and generalization ability of the model.

[0060] (3) The present invention aggregates the importance score based on the row norm of the feature selection matrix of each view, taking into account both view specificity and overall contribution, and supports the selection of key features preset on demand, which meets the requirements of industrial scenarios for lightweight and deployability. Attached Figure Description

[0061] Figure 1 This is a flowchart of the IoT multi-view feature selection method based on similarity matrix fusion according to an embodiment of the present invention;

[0062] Figure 2 This is a diagram of an IoT multi-view feature selection system based on similarity matrix fusion according to an embodiment of the present invention;

[0063] Figure 3 This is a schematic diagram of the ACC index under different feature selection numbers in an embodiment of the present invention. Detailed Implementation

[0064] The present invention will be further described in detail below with reference to the embodiments and accompanying drawings, but the embodiments of the present invention are not limited thereto.

[0065] like Figure 1 As shown, the present invention provides a method for multi-view feature selection in the Internet of Things based on similarity matrix fusion, comprising:

[0066] S1: For each incomplete industrial IoT data, the data matrix of the wide table is obtained by concatenating the columns. The data matrix of the wide table is then supplemented to construct the data matrix of each industrial IoT view.

[0067] Specifically, each industrial IoT incomplete data Concatenate the columns into a wide table, as follows:

[0068] ;

[0069] ;

[0070] in, Represents a wide table containing multi-view data; Indicates the total dimensions of the view; Indicates the dimension of a single view v; 'v' indicates the overlay of multiple view dimensions; 'v' indicates a single view.

[0071] The data matrix of the multi-view wide table is extracted and supplemented to construct the data matrix for each Industrial IoT view, as follows:

[0072] ;

[0073] in, This represents the data matrix for each Industrial IoT view after extraction and supplementation.

[0074] S2 performs multiple clustering operations on the data matrix of each Industrial IoT view to construct a similarity matrix for each view.

[0075] Specifically, the formula for multiple clustering calculations on the data matrix of each Industrial IoT view is as follows:

[0076] ;

[0077] Multiple clustering operations can yield k clustering results. Then, the similarity matrix is ​​constructed from the results of the k clustering operations, as follows:

[0078] ;

[0079] in, The matrix represents each view; U represents the clustering result; This indicates transpose.

[0080] S3 stacks the similarity matrices of each view into a third-order tensor and applies tensor learning constraints to capture the global low-rank structure across views. At the same time, adaptive view weights are introduced to perform weighted fusion of the global low-rank structure across views to obtain a common similarity matrix.

[0081] Specifically, the formulas for calculating the similarity matrix are as follows:

[0082] ;

[0083] in, This represents a tensor obtained by stacking the similarity matrices of each view along the third dimension; This represents the shared similarity matrix learned uniformly. Indicates the weighted parameters; This represents the F-norm.

[0084] S4 extracts pseudo-label matrices by eigenvalue decomposition of shared similarity matrices, and constructs a multi-view feature selection model by guiding incomplete industrial IoT data through pace learning.

[0085] Specifically, the pseudo-label matrix is ​​extracted by eigenvalue decomposition of the common similarity matrix. The calculation formula is as follows:

[0086] ;

[0087] ;

[0088] ;

[0089] in, Represents the pseudo-label matrix; This represents the first largest eigenvalue; This represents the c-th largest eigenvalue; This indicates that the elements within the brackets form a diagonal matrix; This indicates taking the maximum value; This represents a matrix composed of the first c eigenvectors; Represents the square root of a diagonal matrix;

[0090] A multi-view feature selection model is constructed by guiding incomplete industrial IoT data through pace learning. The calculation formula is as follows:

[0091] ;

[0092] ;

[0093] in The pace learning ratio parameter for each IoT view; Indicates the initial parameters of self-synchronization; Indicates the maximum self-synchronizing parameter; Represents the self-synchronization matrix; Represents the feature selection matrix; Represents the pseudo-label matrix; Self-synchronization parameters for each data point; This indicates data selected through auto-stepping; This indicates a constraint.

[0094] S5 calculates the importance score of each industrial IoT view feature based on the multi-view feature selection model, sorts all features in descending order of score, and selects the top m features as the final feature subset.

[0095] Specifically, all features are sorted in descending order of score, and the top m features are selected as the final feature subset, which includes:

[0096] Calculate the importance score of each feature in each IoT view. :

[0097] ;

[0098] in, This represents the j-th row of the feature selection matrix for the v-th IoT view;

[0099] Summarize the feature importance scores of all IoT views:

[0100] ;

[0101] Sort all feature scores in T2 in descending order, and select the top m features as the final feature subset:

[0102] ;

[0103] Where m is the preset number of features to select.

[0104] In summary, this invention fills incomplete data with imputation to obtain complete data after imputation. Then, it performs clustering based on the imputed data to generate similarity matrices for each view, thereby effectively guiding feature selection from incomplete data. By stacking the similarity matrices of each view into a third-order tensor and applying low-rank constraints, it can mine potential shared structures across views from a global perspective. Compared with the traditional view-by-view independent processing method, it makes fuller use of the consistency information of multi-view data and improves the reliability of similarity measurement. The step-learning strategy guides the feature selection process. Through a progressive training method from simple samples to complex samples, it avoids the model being misled by difficult or noisy samples in the early stage, which significantly improves the stability and accuracy of feature selection results.

[0105] like Figure 2 As shown, this embodiment also discloses an IoT multi-view feature selection system based on similarity matrix fusion, including:

[0106] The extraction and supplementation module 21 is used to concatenate each incomplete industrial IoT data column by column to obtain a wide table data matrix, and to extract and supplement the wide table data matrix to construct the data matrix of each industrial IoT view.

[0107] Clustering module 22 is used to perform multiple clusterings on the data matrix of each industrial IoT view to construct a similarity matrix for each view;

[0108] The weighted fusion module 23 is used to stack the similarity matrix of each view into a third-order tensor and apply tensor learning constraints to capture the global low-rank structure across views. At the same time, adaptive view weights are introduced to perform weighted fusion of the global low-rank structure across views to obtain a common similarity matrix.

[0109] The step-guiding module 24 is used to extract pseudo-label matrices by eigenvalue decomposition of common similarity matrices and to construct multi-view feature selection models by guiding incomplete industrial IoT data through step-learning.

[0110] The feature selection module 25 is used to calculate the importance score of each industrial IoT view feature based on the multi-view feature selection model, sort all features in descending order of score, and select the top m features as the final feature subset.

[0111] Specifically, such as Figure 3 As shown in the figure, the ACC (red line for this method) of the system corresponding to this invention under different feature selection numbers is as follows. The results show that, at different feature selection numbers (m), this method is highly competitive compared to other methods and achieves optimal performance, indicating that it can extract key feature information. The system corresponding to this invention performs feature selection on a multi-view news webpage information dataset. The evaluation metric is clustering accuracy (ACC), which can be used to measure the degree of matching between the feature selection results and the true labels. The value range is [0,1], with a larger value indicating better performance.

[0112] The specific implementation of the IoT multi-view feature selection system based on similar matrix fusion is the same as the IoT multi-view feature selection method based on similar matrix fusion, and will not be described again in this embodiment.

[0113] Although the invention has been specifically shown and described in conjunction with preferred embodiments, those skilled in the art should understand that various changes in form and detail may be made to the invention without departing from the spirit and scope of the invention as defined in the appended claims, all of which shall be within the scope of protection of the invention.

Claims

1. A method for selecting multi-view features in the Internet of Things based on similarity matrix fusion, characterized in that, Includes the following steps: S1, concatenate each incomplete industrial IoT data column by column to obtain a wide table data matrix, and extract and supplement the wide table data matrix to construct the data matrix of each industrial IoT view; S2, perform multiple clustering operations on the data matrix of each Industrial IoT view to construct a similarity matrix for each view; S3 stacks the similarity matrices of each view into a third-order tensor and applies tensor learning constraints to capture the global low-rank structure across views. At the same time, adaptive view weights are introduced to perform weighted fusion of the global low-rank structure across views to obtain the common similarity matrix. S4. Eigenvalue decomposition is performed on the common similarity matrix to extract the pseudo-label matrix. Step learning guides the construction of a multi-view feature selection model for incomplete industrial IoT data. S5. Calculate the importance score of each industrial IoT view feature based on the multi-view feature selection model, sort all features in descending order of score, and select the top-ranked preset number of features as the final feature subset. S1 is specifically as follows: Incomplete data from each Industrial IoT Concatenate the columns into a wide table, as shown below: ； ； in, Represents a wide table containing multi-view data; Indicates the total dimensions of the view; Indicates the dimension of a single view v; 'v' indicates the overlay of multiple view dimensions; 'v' indicates a single view. The data matrix of the multi-view wide table is extracted and supplemented to construct the data matrix for each Industrial IoT view, as follows: ； in, This represents the data matrix for each Industrial IoT view after data extraction and supplementation. The formula for multiple clustering calculations on the data matrix of each Industrial IoT view is as follows: ； Multiple clustering operations can yield k clustering results. Then, the similarity matrix is ​​constructed from the results of the k clustering operations, as follows: ； in, The matrix represents each view; U represents the clustering result; Indicates transpose; The formula for calculating the common similarity matrix is ​​as follows: ； in, This represents a tensor obtained by stacking the similarity matrices of each view along the third dimension. This represents the shared similarity matrix learned uniformly. Indicates the weighted parameters; Denotes the F-norm; The pseudo-label matrix is ​​extracted by eigenvalue decomposition of the common similarity matrix. The calculation formula is as follows: ； ； ； in, Represents the pseudo-label matrix; This represents the first largest eigenvalue; This represents the c-th largest eigenvalue; This indicates that the elements within the brackets form a diagonal matrix; This indicates taking the maximum value; This represents a matrix composed of the first c eigenvectors; Represents the square root of a diagonal matrix; A multi-view feature selection model is constructed by guiding incomplete industrial IoT data through pace learning. The calculation formula is as follows: ； ； in The pace learning ratio parameter for each IoT view; Indicates the initial parameters of self-synchronization; Indicates the maximum self-stepping parameter; Represents the self-synchronization matrix; Represents the feature selection matrix; Represents the pseudo-label matrix; Self-synchronization parameters for each data point; This indicates data selected through auto-stepping; Indicates constraints; All features are sorted in descending order of score, and the top-ranked features (a predetermined number) are selected as the final feature subset, which includes: Calculate the importance score of each feature in each IoT view. : ； in, This represents the j-th row of the feature selection matrix for the v-th IoT view; Summarize the feature importance scores of all IoT views: T2 ； Sort all feature scores in T2 in descending order, and select the top m features as the final feature subset: ； Where m is the preset number of features to select.

2. An IoT multi-view feature selection system based on similarity matrix fusion, characterized in that, include: The extraction and supplementation module is used to concatenate each incomplete industrial IoT data column by column to obtain a wide table data matrix, and to extract and supplement the wide table data matrix to construct the data matrix of each industrial IoT view. The clustering module is used to perform multiple clustering operations on the data matrix of each Industrial IoT view to construct a similarity matrix for each view; The weighted fusion module is used to stack the similarity matrix of each view into a third-order tensor and apply tensor learning constraints to capture the global low-rank structure across views. At the same time, adaptive view weights are introduced to perform weighted fusion of the global low-rank structure across views to obtain the common similarity matrix. The step-guided module is used to extract pseudo-label matrices by eigenvalue decomposition of shared similar matrices and to construct multi-view feature selection models from incomplete industrial IoT data through step-learning. The feature selection module is used to calculate the importance score of each industrial IoT view feature based on the multi-view feature selection model, sort all features in descending order of score, and select the top preset number of features as the final feature subset. S1 is specifically as follows: Incomplete data from each Industrial IoT Concatenate the columns into a wide table, as shown below: ； ； in, Represents a wide table containing multi-view data; Indicates the total dimensions of the view; Indicates the dimension of a single view v; 'v' indicates the overlay of multiple view dimensions; 'v' indicates a single view. The data matrix of the multi-view wide table is extracted and supplemented to construct the data matrix for each Industrial IoT view, as follows: ； in, This represents the data matrix for each industrial IoT view after extraction and supplementation. The formula for multiple clustering calculations on the data matrix of each Industrial IoT view is as follows: ； Multiple clustering operations can yield k clustering results. Then, the similarity matrix is ​​constructed from the results of the k clustering operations, as follows: ； in, The matrix represents each view; U represents the clustering result; Indicates transpose; The formula for calculating the common similarity matrix is ​​as follows: ； in, This represents a tensor obtained by stacking the similarity matrices of each view along the third dimension. This represents the shared similarity matrix learned uniformly. Indicates the weighted parameters; Denotes the F-norm; The pseudo-label matrix is ​​extracted by eigenvalue decomposition of the common similarity matrix. The calculation formula is as follows: ； ； ； in, Represents the pseudo-label matrix; This represents the first largest eigenvalue; This represents the c-th largest eigenvalue; This indicates that the elements within the brackets form a diagonal matrix; This indicates taking the maximum value; This represents a matrix composed of the first c eigenvectors; Represents the square root of a diagonal matrix; A multi-view feature selection model is constructed by guiding incomplete industrial IoT data through pace learning. The calculation formula is as follows: ； ； in The pace learning ratio parameter for each IoT view; Indicates the initial parameters of self-synchronization; Indicates the maximum self-stepping parameter; Represents the self-synchronization matrix; Represents the feature selection matrix; Represents the pseudo-label matrix; Self-synchronization parameters for each data point; This indicates data selected through auto-stepping; Indicates constraints; All features are sorted in descending order of score, and the top-ranked features (a predetermined number) are selected as the final feature subset, which includes: Calculate the importance score of each feature in each IoT view. : ； in, This represents the j-th row of the feature selection matrix for the v-th IoT view; Summarize the feature importance scores of all IoT views: ； Sort all feature scores in T2 in descending order, and select the top m features as the final feature subset: ； Where m is the preset number of features to select.

Images