Multi-view intelligent seal management and control method based on tensor learning

Abstract

The invention discloses a multi-view intelligent seal management and control method based on tensor learning, and belongs to the technical field of tensor calculation and data mining. Comprising the following steps: collecting data and processing the data; decomposing the data by adopting tensor singular value decomposition to obtain a tensor nuclear norm; constructing a multi-view subspace clustering model based on tensor low-rank learning, and optimizing representation of an affinity matrix; performing real-time spectral clustering by using the optimized affinity matrix, and performing early warning and management and control on an abnormal use mode; and generating a seal use report. According to the method, different subspaces in data are analyzed and clustered by constructing a multi-view subspace clustering model; the expression form of the affinity matrix is optimized, so that the relationship between the data can reflect the actual seal use condition more accurately; through the model, the monitoring capability of the seal use behavior can be enhanced, and a more efficient early warning mechanism is provided for the abnormal use behavior.

Description

Technical Field

[0001] The present invention belongs to the technical field of tensor calculation and data mining, and specifically relates to a multi-view intelligent seal management and control method based on tensor learning. Background Art

[0002] With the popularization of digital office, the use of smart seals has become more and more common in enterprises and government agencies. As a tool with legal effect, the frequency of use, scope of authorization, and use process of seals need to be strictly monitored. However, the traditional seal management model relies on manual operation and record keeping, which has many disadvantages, such as improper operation, incomplete records, and lack of transparency in the use process. Especially when multiple people and multiple departments use it together, it is difficult to effectively prevent abuse or abnormal use. In this case, how to achieve intelligent and real-time seal control has become an important issue in corporate compliance management.

[0003] Existing smart seal management and control technologies mostly use data analysis and machine learning methods, but when faced with complex multi-view data (such as usage time, personnel, file type, etc.), conventional methods still have deficiencies in anomaly detection and pattern recognition. On the one hand, traditional multi-view data processing methods are easily affected by noise when faced with high-dimensional data, resulting in inaccurate anomaly detection; on the other hand, the performance of clustering algorithms is limited when faced with complex data structures, making it difficult to efficiently extract key features, affecting the accurate monitoring of seal usage behavior.

[0004] Therefore, current research urgently needs to improve the multi-view subspace clustering algorithm and enhance the ability to process complex data. By introducing advanced technologies such as tensor low-rank learning and tensor singular value decomposition, it is possible to more effectively mine abnormal patterns in the use of seals, optimize the affinity matrix, and enhance the clustering effect, thereby achieving real-time monitoring and early warning of the use of smart seals and providing more scientific decision-making support for management. However, the effectiveness and real-time performance of relevant algorithms in practical applications are still challenging, and further research and optimization are still necessary.

[0005] To sum up, how to solve the problem of complex data processing in seal management solutions can enhance the ability to monitor seal usage behavior. Summary of the invention

[0006] The purpose of the present invention is to provide a multi-view intelligent seal management and control method based on tensor learning to solve the problems raised in the above background technology.

[0007] The object of the present invention is achieved by: a multi-view intelligent seal control method based on tensor learning, characterized in that the method comprises the following steps: Step S1: Collect multi-view data related to seal usage; Step S2: Processing multi-view data using low-rank representation; Step S3: Decompose the data using tensor singular value decomposition, extract key features of seal usage, and obtain tensor nuclear norm; Step S4: construct a multi-view subspace clustering model based on tensor low-rank learning, use the multi-view subspace clustering model to perform cluster analysis on the data, optimize the representation of the affinity matrix, and enable the monitoring and early warning mechanism of seal use; Step S5: Use the optimized affinity matrix to perform real-time spectral clustering to provide early warning and control of abnormal usage patterns; Step S6: Generate a seal usage report to provide decision support for management.

[0008] Preferably, the step S1 collects multi-view data related to the use of the seal, specifically: Multi-view data collection methods include: time dimension data collection, personnel identity data collection, file type data collection, usage frequency and mode data, and device and location data collection. The data of each collection method is regarded as a type, and each type is represented as a view tensor data.

[0009] Preferably, in step S2, the multi-view data is processed using a low-rank representation, specifically: The low rank representation is represented by LRR, and the mathematical formula of LRR is defined as follows: ; ; in, Represents a low-rank tensor The nuclear norm of is the error tensor, Represents the error term The Frobenius norm of , is a regularization parameter used to balance the low-rank approximation and data fitting errors; Use the generated tensor data for the actual stamp, This is the seal data after standardization; is a low-rank tensor.

[0010] Preferably, in step S3, tensor singular value decomposition is used to decompose the data, extract key features of seal use, and obtain the tensor nuclear norm, specifically: The tensor singular value decomposition is used to extract the key behavior features of the stamp. The tensor singular value decomposition is represented by t-SVD. The t-SVD mathematical formula is defined as follows: ; in, Indicates Frontal slice singular values, tensor nuclear norm is defined as the sum of the singular values ​​in all frontal slices; Indicates the size of the first dimension of the tensor, corresponding to the spatial dimension of the data; Indicates the size of the second dimension of the tensor, corresponding to another dimension of the space; Indicates the size of the third dimension of the tensor, which is the time dimension.

[0011] Preferably, in step S4, a multi-view subspace clustering model based on tensor low-rank learning is constructed, and the multi-view subspace clustering model is used to perform cluster analysis on the data, optimize the representation of the affinity matrix, and enable the monitoring and early warning mechanism of seal use, specifically: Step S4-1: construct a multi-view subspace clustering model. The multi-view subspace clustering model is expressed as: (1); in, Representing a tensor The side slice, i.e. ; Indicates The error matrix of each view is To represent a Line and The real number space of columns; Refers to The feature dimension size of each view; Indicates the number of samples; Force each view The columns have a common and consistent size value; is the Laplacian matrix, and is a diagonal matrix with diagonal entries , represents the clustering indicator matrix, where is the number of clusters; is the regularization parameter; To adjust the clustering index matrix in the objective function The relative importance of a term to other terms, emphasizing the consistency of data points in the clustering structure; is the original multi-view tensor data; To express The trace operation of , represents a graph Laplace regularization; is the weight coefficient; Represents the Schatten-p norm of a tensor; To control the power parameter of the norm, it determines the degree of penalty of the tensor core norm; for The transposed matrix of It is the transpose matrix operation; Step S4-2: Use the multi-view subspace clustering model to perform cluster analysis on the data, optimize the representation of the affinity matrix, and enable the monitoring and early warning mechanism of seal use.

[0012] Preferably, the step S4-2: using a multi-view subspace clustering model to perform cluster analysis on the data, optimizing the representation of the affinity matrix, and enabling a monitoring and early warning mechanism for seal use, specifically comprises: By introducing an auxiliary tensor variable , formula (1) is transformed into the following equivalent problem: (2); Among them, the matrix To express constraints Lagrange multipliers of ; matrix Reflects the constraints The degree of violation of To express constraints The Lagrange multiplier of the tensor Reflects the constraints the extent of the violation; For constraints The penalty item, for Penalty items; is the objective function, and the optimization goal is to minimize the value of this function and find an optimal set of variables ; is the introduced auxiliary tensor variable, which represents the shared features of multi-view subspaces; It is a low-rank self-representation form of multi-dimensional tensor; Formula (2) is decomposed into solving the following four sub-problems: Sub-problem (1): Solve ,variable is fixed; (3); in, , ; yes No. OK, we get: (4); At this time, formula (3) becomes: (5); in, To adjust the clustering index matrix in the objective function The relative importance of a term to other terms, emphasizing the consistency of data points in the clustering structure; is the number of views; is the graph Laplacian matrix The transpose of For the Lagrange multiplier for each view; is the auxiliary variable introduced; for Penalty items; The solution of formula (5) is: (6); in, is the identity matrix; It is the Hadamard product, which multiplies the corresponding elements of two matrices or tensors with the same dimension one by one; Indicates that each element is a matrix The symbols of the corresponding elements in ; Sub-problem (2): Solve ,variable is fixed; (7); The optimal solution is: (8); in, express No. List, ; Sub-problem (3): Solve ,variable is fixed; (9); The optimal solution is: (10); Sub-problem (4): Solve ,variable is fixed; (11); Optimal solution Depend on of The eigenvector corresponding to the smallest eigenvalue.

[0013] Preferably, the multi-view subspace clustering model includes an adaptive adjustment mechanism of a weighted Schatten-p norm, and the mathematical formula of the adaptive adjustment mechanism of the weighted Schatten-p norm is defined as follows: ; in, express No. The largest singular value, Represents the weight vector No. elements, is the order of the Schatten-p norm.

[0014] Preferably, in step S5, the optimized affinity matrix is ​​used to perform real-time spectral clustering to provide early warning and control of abnormal usage patterns, specifically: Step S5-1: Setting initial parameters: Set the initial parameters required for the optimization problem, including , convergence condition ; Step S5-2: Loop to update variables: According to formula (6), each view variable Make updates; Calculate and update the error matrix according to formula (8) ; Using the formula Update Lagrange multipliers ; According to formula (10), calculate and update the clustering index matrix ; Step S5-3: Check convergence conditions: Check whether the convergence conditions are met for each view: and ; If the convergence condition is not met, return to step S5-2; if the convergence condition is met, execute step S5-4; Step S5-4: Output result: When the convergence condition is met, the optimized affinity matrix is ​​output, and the final clustering result is obtained through the spectral clustering algorithm.

[0015] Compared with the prior art, the present invention has the following improvements and advantages: 1. By constructing a multi-view subspace clustering model, different subspaces in the data are analyzed and clustered; the representation of the affinity matrix is ​​optimized so that the relationship between the data can more accurately reflect the actual seal usage; through this model, the ability to monitor seal usage behavior can be enhanced, and a more efficient early warning mechanism can be provided for abnormal usage behavior.

[0016] 2. By utilizing the optimized affinity matrix, implementing real-time spectral clustering analysis, and conducting timely warnings and processing, normal usage behaviors can be effectively distinguished from abnormal behaviors, thereby providing more accurate risk control measures to ensure the safe use of seals.

[0017] 3. Based on the data and analysis results processed according to all steps, a detailed seal usage report is automatically generated. The report covers the overall usage of the seal, detected abnormal behaviors and related early warning information, providing an effective decision-making basis for the management; at the same time, the report can help the management better understand the seal usage pattern, optimize the seal management process, and ensure the compliance of the enterprise operation. BRIEF DESCRIPTION OF THE DRAWINGS

[0018] Figure 1 Flow chart of the method of the present invention.

[0019] Figure 2 A flowchart for optimizing the multi-view subspace clustering model of the present invention.

[0020] Figure 3 This is a schematic diagram of the comparison results of the convergence analysis of the Yale and UCI-Digits data sets used in the present invention.

[0021] Figure 4 This is a schematic diagram of the comparison results of convergence analysis of ORL and BBCSport datasets used in the present invention.

[0022] Figure 5 This is a schematic diagram of the comparison results of the convergence analysis of the 3Sources and Scene-15 datasets used in the present invention.

[0023] Figure 6 This is different from the ORL dataset in the present invention. Schematic diagram of value clustering results.

[0024] Figure 7 This is different from the Yale dataset in the present invention. Schematic diagram of value clustering results. DETAILED DESCRIPTION

[0025] The present invention is further summarized below with reference to the accompanying drawings.

[0026] like Figure 1 As shown, a multi-view intelligent seal control method based on tensor learning includes the following steps: Collect multi-view data related to seal usage, specifically: The multi-view data collection methods include: time dimension data collection, personnel identity data collection, file type data collection, usage frequency and mode data collection, and device and location data collection. The data of each collection method is regarded as a type, and each type is represented as a view tensor data; Time dimension data collection: The system log automatically records the precise timestamp of each seal use, including details such as date, time period, and usage duration; time dimension data can help analyze the periodicity of seal use and abnormally high or low frequency usage patterns.

[0027] Personnel identity data collection: Authenticate each seal user (such as through fingerprint, facial recognition, password or smart card), and record the relevant operator identity information, including their position, department and authority level; personnel identity data helps analyze the usage behavior of specific personnel and identify whether there is a risk of unauthorized or unauthorized use.

[0028] Document type data collection: Each time a seal is used, the category of the document being processed is automatically recorded, such as contracts, financial statements, agreements, etc.; this document type information can further help management understand the frequency of seal usage for different types of documents and determine whether there is abnormal signing behavior in specific types of documents.

[0029] Usage frequency and pattern data: Track the frequency of use of each seal within a specific time period and generate usage pattern data for each seal; through this collection method, the regular use of the seal can be monitored to determine whether there is abnormal high or low frequency of use.

[0030] Equipment and location data collection: Through the device management system, the device information (such as computers, mobile devices, etc.) that the seal relies on each time it is used and the geographical location information when it is used are collected; this data can help analyze whether the environment in which the seal is used complies with regulations, such as whether it is used at an authorized location or device, to avoid the risk of illegal movement or remote abuse.

[0031] Step S2: Processing multi-view data using low-rank representation; The collected multi-view tensor data is processed and low-rank representation (LRR) is used. This process identifies the main features in the data and reduces noise interference by reducing dimension and removing noise, ensuring that the normal mode and potential anomalies of seal use are captured more accurately. The specific operations are as follows: Step S2-1: Tensor data preprocessing: First, the collected multi-view tensor data is preprocessed to integrate data of different dimensions into a unified framework; this step ensures that data of each dimension can be processed in the same space and excludes preliminary missing values ​​or outliers. Preprocessing includes operations such as data standardization and normalization, so that different data dimensions have relatively consistent scales.

[0032] Step S2-2: Construct a low-rank representation LRR model based on multi-view tensor data: Low-rank representation is achieved by minimizing the nuclear norm, mining the potential high-order correlations between seal data samples and identifying abnormal patterns of seal use; The low rank representation is represented by LRR, and the mathematical formula of LRR is defined as follows: ; ; in, Represents a low-rank tensor The nuclear norm of is the error tensor, Represents the error term The Frobenius norm of , is a regularization parameter used to balance the low-rank approximation and data fitting errors; Use the generated tensor data for the actual stamp, This is the seal data after standardization; is a low-rank tensor.

[0033] Step S2-3: Perform matrix decomposition operation on tensor data: The low-rank representation LRR model decomposes the original tensor data matrix into a low-rank matrix and a sparse matrix. The low-rank matrix contains the main structures and features in the data and represents the normal usage pattern; while the sparse matrix represents possible abnormal noise or abnormal usage behavior.

[0034] Step S2-4: Noise and anomaly detection: The sparse matrix obtained by decomposition can identify abnormal items in the data; these abnormal items are usually caused by abnormal seal usage behavior, such as abnormally frequent use, unauthorized user operations or non-conventional file types; suspicious usage records are marked based on these abnormal items, providing a basis for subsequent analysis and alarm mechanisms.

[0035] Step S2-5: Feature extraction and pattern recognition: After removing noise and abnormal items, the main features representing normal seal usage behavior are extracted from the low-rank matrix; these features can include the frequency of operation, usage patterns in a specific time period, behavioral habits of specific personnel, etc. The extracted features can form a complete description of the normal usage pattern and provide a reliable basis for subsequent monitoring and analysis.

[0036] Step S2-6: Reconstruct data and store results: After completing the low-rank representation and anomaly detection, the processed data is reconstructed to generate clear, denoised tensor data; this dataset not only contains the main features of normal usage behavior, but also can mark potential abnormal patterns; these results are stored in the database for subsequent anomaly detection and real-time monitoring.

[0037] The multi-view data that has been processed by low-rank representation is further decomposed and analyzed using tensor singular value decomposition (t-SVD) technology; t-SVD can effectively extract key feature information from multi-dimensional data, help reveal the inherent patterns and trends in the use of seals, and calculate the complex relationships hidden in the data through the tensor nuclear norm; In step S3, tensor singular value decomposition is used to decompose the data, extract key features of seal usage, and obtain tensor nuclear norm. The specific operation is as follows: Step S3-1: Tensor structure input: First, receive the multi-view tensor data obtained from the previous step. This data has been processed by noise removal and anomaly detection, but still contains multiple dimensions (such as time, person, file type, etc.); each dimension may reveal different seal usage characteristics, so it is necessary to further extract the core information through t-SVD decomposition technology.

[0038] Step S3-2: Tensor singular value decomposition (t-SVD). t-SVD can extract features from multiple dimensions of a tensor simultaneously by considering a multi-dimensional tensor as an extension of a matrix. This is different from traditional matrix decomposition techniques. The steps are as follows: Fourier transform: First, Fourier transform is performed on the tensor along the time dimension to convert the time domain tensor into a frequency domain representation; Fourier transform helps reduce data complexity and enhance the efficiency of t-SVD decomposition.

[0039] Tensor decomposition: After Fourier transform, the tensor in the frequency domain is gradually decomposed; tensor singular value decomposition can decompose the original tensor into a series of smaller tensor matrices that contain the low-rank structural information, singular values, and eigenvectors of the original data.

[0040] Step S3-3: Tensor nuclear norm calculation: After the decomposition is completed, the nuclear norm of the tensor is calculated, which is a key indicator in t-SVD. The nuclear norm is used to evaluate the complexity of the tensor and the compactness of the data structure, and can reflect the inherent rules in the seal usage pattern; The mathematical formula of the nuclear norm is: ; in, Indicates Frontal slice singular values, tensor nuclear norm is defined as the sum of the singular values ​​in all frontal slices; Indicates the size of the first dimension of the tensor, corresponding to the spatial dimension of the data; Indicates the size of the second dimension of the tensor, corresponding to another dimension of the space; Indicates the size of the third dimension of the tensor, which is the time dimension.

[0041] Through t-SVD decomposition, representative key features can be extracted from multi-view tensors; these features can not only simplify the data structure, but also reveal the inherent trends and patterns of seal usage behavior; the specific results of feature extraction include: Time dimension features: Identify the peak, trough or abnormally frequent usage of seals in a specific time period; this helps monitor whether the seal is abused during non-working hours.

[0042] Personnel dimension features: Through decomposition, the usage patterns of specific operators can be extracted, such as frequent seal use, abnormal permission access, etc.; these features provide support for preventing internal abuse or unauthorized operations.

[0043] File type characteristics: The system can also analyze from the file dimension which file types are frequently stamped; if the seal usage rate of a certain type of file is significantly higher than other types, it may indicate that there are risks in specific business or operating processes.

[0044] Step S3-5: Data simplification and reconstruction: After feature extraction, the decomposition results are used to simplify the data and remove redundant information and complex multi-dimensional structures. The simplified tensor still retains the core features, but its data volume and processing difficulty are greatly reduced, which helps to improve the efficiency of subsequent analysis and monitoring; the data is reconstructed according to the extracted key features to generate a new tensor data set; the reconstructed data is more concise and compact, which can highlight the core patterns of seal use and provide a basis for subsequent clustering analysis.

[0045] Step S3-6: Result storage and output: The results of t-SVD decomposition include low-rank tensors, nuclear norms, key features, etc., and the system stores these results in the database; these decomposition results provide basic data for subsequent monitoring, anomaly detection and decision support; the system also generates a preliminary analysis report, outlining the inherent patterns, key features, and possible abnormal trends in seal use.

[0046] After completing the data decomposition, a multi-view subspace clustering model is constructed based on the results of low-rank representation (LRR) and tensor singular value decomposition (t-SVD). The multi-view subspace clustering model can effectively combine the advantages of the two methods to analyze and cluster different subspaces in the data. In step S4, a multi-view subspace clustering model based on tensor low-rank learning is constructed, and the multi-view subspace clustering model is used to perform cluster analysis on the data, optimize the representation of the affinity matrix, and enable the monitoring and early warning mechanism of seal use, specifically: Step S4-1: construct a multi-view subspace clustering model. The multi-view subspace clustering model is expressed as: (1); in, Representing a tensor The side slice, i.e. ; Indicates The error matrix of each view is To represent a Line and The real number space of columns; Refers to The feature dimension size of each view; Indicates the number of samples; Force each view The columns have a common and consistent size value; is the Laplacian matrix, and is a diagonal matrix with diagonal entries , represents the clustering indicator matrix, where is the number of clusters; is the regularization parameter; To adjust the clustering index matrix in the objective function The relative importance of a term to other terms, emphasizing the consistency of data points in the clustering structure; is the original multi-view tensor data; To express The trace operation of , represents a graph Laplace regularization; is the weight coefficient; Represents the Schatten-p norm of a tensor; To control the power parameter of the norm, it determines the degree of penalty of the tensor core norm; for The transposed matrix of It is the transpose matrix operation; Step S4-2: Use the multi-view subspace clustering model to perform cluster analysis on the data, optimize the representation of the affinity matrix, and enable the monitoring and early warning mechanism of seal use: By introducing an auxiliary tensor variable , formula (1) is transformed into the following equivalent problem: (2); Among them, the matrix To express constraints Lagrange multipliers of ; matrix Reflects the constraints If the constraint is not strictly satisfied, the matrix A penalty term will be introduced to affect the optimization direction; in the actual optimization process, the matrix It will be dynamically updated to gradually reduce the degree of constraint violation so that the final optimization result satisfies the constraint; the tensor To express constraints The Lagrange multiplier of the tensor Reflects the constraints If the constraint is not satisfied, the tensor The optimization direction will be adjusted through the penalty term; tensor The introduction of helps to decompose complex optimization problems into tractable sub-problems, because and Can be optimized separately; For constraints The penalty item, for The penalty parameter determines the strictness of the constraint and the convergence speed and accuracy of the algorithm. It penalizes the solution that does not meet the constraint by amplifying the constraint deviation to reduce the error, or slows down the convergence speed to prevent numerical instability or convergence to a local suboptimal solution caused by over-fast convergence. and , a balance can be found between the accuracy of the solution and the convergence speed; is the objective function, and the optimization goal is to minimize the value of this function and find an optimal set of variables ; is the introduced auxiliary tensor variable, which represents the shared features of multi-view subspaces; It is a low-rank self-representation form of multi-dimensional tensor; Formula (2) is decomposed into solving the following four sub-problems: Sub-problem (1): Solve ,variable is fixed; (3); in, , ; yes No. OK, we get: (4); At this time, formula (3) becomes: (5); in, To adjust the clustering index matrix in the objective function The relative importance of a term to other terms, emphasizing the consistency of data points in the clustering structure; is the number of views; is the graph Laplacian matrix The transpose of For the Lagrange multiplier for each view; is the auxiliary variable introduced; for Penalty items; The solution of formula (5) is: (6); in, is the identity matrix; It is the Hadamard product, which multiplies the corresponding elements of two matrices or tensors with the same dimension one by one; Indicates that each element is a matrix The symbols of the corresponding elements in ; Sub-problem (2): Solve ,variable is fixed; (7); The optimal solution is: (8); in, express No. List, ; Sub-problem (3): Solve ,variable is fixed; (9); The optimal solution is: (10); Sub-problem (4): Solve ,variable is fixed; (11); Optimal solution Depend on of The eigenvector corresponding to the smallest eigenvalue.

[0047] Step S5: Use the optimized affinity matrix to perform real-time spectral clustering to warn and control abnormal usage patterns, specifically: like Figure 2 As shown, step S5-1: setting initial parameters: Set the initial parameters required for the optimization problem, including , convergence condition ; Step S5-2: Loop to update variables: According to formula (6), each view variable Make updates; Calculate and update the error matrix according to formula (8) ; Using the formula Update Lagrange multipliers ; According to formula (10), calculate and update the clustering index matrix ; Step S5-3: Check convergence conditions: Check whether the convergence conditions are met for each view: and ; If the convergence condition is not met, return to step S5-2; if the convergence condition is met, execute step S5-4; Step S5-4: Output result: When the convergence condition is met, the optimized affinity matrix is ​​output, and the final clustering result is obtained through the spectral clustering algorithm.

[0048] Step S6: Generate a seal usage report to provide decision support for management.

[0049] Based on the data and analysis results processed in the previous steps S1-S5, a detailed seal usage report is automatically generated. The seal usage report not only summarizes the overall seal usage, but also highlights the detected abnormal behaviors and related warning information, providing effective decision-making basis for management. The specific operations are as follows: Step S6-1: Data integration and summary: First, the key data obtained from multiple steps are summarized and integrated, including multi-view data, feature information of low-rank representation, results of tensor singular value decomposition, and anomaly detection data of cluster analysis; this step ensures that the report can comprehensively display the use of seals from all dimensions.

[0050] Step S6-2: Abnormal behavior analysis: Organize the abnormal behavior detection results, focus on analyzing and displaying the seal usage records marked as abnormal after low-rank representation and t-SVD decomposition; this part of data helps management understand potential risks and take targeted measures. The specific steps include: Abnormal usage records: Lists detailed information of each abnormal usage record, including operator, time, file type, device and location, etc., to help management clearly understand the specific abnormal behavior.

[0051] Summary of abnormal patterns: The system clusters multiple abnormal behaviors and extracts common abnormal patterns; for example, abnormally high-frequency use, unauthorized operation, use at irregular time periods, etc., which can help management more easily discover potential loopholes in seal use.

[0052] Step S6-3: Warning information generation: Based on the detected abnormal behavior, the system automatically generates relevant warning information. These warning information can be graded according to the severity of the abnormal behavior (such as low, medium, and high risk), and provide specific suggestions and measures for management. The generation process of warning information includes: Risk assessment and classification: The system conducts risk assessment based on the type, frequency, and associated dimensions of abnormal behavior; high-risk abnormal behaviors (such as unauthorized use or multiple illegal operations) will be highlighted to attract the attention of management.

[0053] Suggestions: Depending on the nature of the abnormal behavior, the system provides a series of response suggestions. For example, if unauthorized personnel are found to use the seal, the system may recommend that management strengthen the identity verification mechanism; if abnormally frequent use is found, it may recommend a process review or restrict the seal usage rights.

[0054] Step S6-4: Seal usage trend analysis: Generate a seal usage trend analysis report to reveal the changing trend of seal usage by analyzing the usage data of different time periods, personnel and document types; this analysis helps management understand the long-term usage and optimize the seal management strategy; the steps include: Time trend analysis: Displays the usage trend of seals in different time periods (such as days, weeks, and months), helping management understand the peak and off-peak usage periods of seals.

[0055] Personnel usage trends: Based on the operator's usage data, the system analyzes the usage of different personnel or departments, identifies whether there are some personnel or departments with abnormal usage frequency, and then discovers potential risks.

[0056] Document type analysis: Analyze the signature frequency of different document types to reveal the categories of documents with high-frequency signatures, helping management focus on the use of seals in specific business processes.

[0057] Step S6-5: Decision support and optimization suggestions: Based on the data collected and the analysis results from steps S6-1 to S6-4, a series of decision support information is provided to help management make more scientific seal management decisions, including the following: Suggestions for optimizing the seal management process: The system may suggest optimizing the seal usage process, such as improving the approval process, strengthening personnel authority management, introducing more intelligent management and control measures, etc.

[0058] Suggestions for improving compliance: If certain operations are found to be inconsistent with corporate compliance requirements, the system will remind management to make improvements and provide corresponding compliance management suggestions.

[0059] In order to verify the feasibility and effect of the present invention, a large number of experiments are carried out on multiple data sets to verify the effectiveness of the multi-view subspace clustering algorithm improved based on tensor low-rank learning.

[0060] The main datasets are: Yale: A total of 165 grayscale images provided by 15 participants are used as the dataset. These images show rich diversity by introducing different lighting conditions, facial expressions and occlusion factors.

[0061] ORL database: Developed by the Olivetti Laboratory in Cambridge, UK, this database contains 400 facial images of 40 individuals at different time points, lighting conditions, facial expression states (including eyes open and closed, smiling and not smiling), and whether glasses are worn. This database and the Yale database both belong to the field of facial images, and by extracting intensity, local binary pattern (LBP) and Gaussian wavelet packet (Gabor) features, a multivariate representation of multi-view data is achieved.

[0062] 3Sources: This dataset is a multi-view text dataset collected from three online news sources: BBC, Reuters, and The Guardian, covering 416 different news reports. The dataset comes from six categories of topic tags: business, entertainment, health, politics, sports, and technology. It is a multi-view data with three characteristics.

[0063] BBCSport: A news story dataset containing sports news articles from five subject areas on the BBC Sport website from 2004-2005, often used as a benchmark for machine learning research.

[0064] UCI-Digits: This dataset consists of handwritten digits (ranging from 0 to 9) from the Dutch public utility map library, with a total of ten categories of Arabic numerals, each category containing 200 images.

[0065] Scene-15: This database includes 15 scene categories and 4485 images. All images come from a wide range of indoor and outdoor environments, such as industry, bedroom, kitchen, office, store, etc.

[0066] In the experiment, three types of image features are extracted, namely 1800-dimensional PHOW features, 1180-dimensional PRI-CoLBP, and 1240-dimensional CENTRIST features, and they are regarded as different views.

[0067] The convergence of the model is verified through experiments, and its convergence performance is evaluated on six data sets: like Figure 3 , Figure 4 , Figure 5 As shown, the variable errors are listed The data of the number of iterations show that all iterations are stable after reaching 20, which proves that the method converges effectively. The experiment shows that the algorithm of the present invention exhibits excellent convergence performance.

[0068] Also discussed are the parameters The impact on data, Figure 6 , Figure 7 List the different Value clustering results; From the figure, we can see that the clustering performance increases with The value of the variable has certain fluctuations. When other parameters are fixed, The best clustering effect is achieved on the Yale dataset when When , the best clustering effect is achieved on the ORL dataset. This is because The change of the value affects the difference between the singular values. By selecting a suitable value, the method of the present invention can better utilize complementary information and reduce the influence of other noises such as illumination.

[0069] Table 1 describes the clustering effect of the method on several data sets: ACC NMI Purity F-score Recall AR Yale 0.981 0.977 0.981 0.963 0.964 0.960 UCI-Digits 0.995 0.986 0.995 0.990 0.990 0.988 ORL 0.986 0.991 0.967 0.990 0.990 0.988 BBC Sport 0.867 0.865 0.904 0.841 0.845 0.869 3source 0.884 0.938 0.865 0.837 0.834 0.825 Scene-15 0.971 0.9906 0.975 0.987 0.987 0.986

[0070] The above description is only an embodiment of the present invention and is not intended to limit the present invention. For those skilled in the art, the present invention may have various modifications and variations. Any modification, equivalent substitution, improvement, etc. made within the spirit and principle of the present invention should be included in the scope of the claims of the present invention.

Claims

1. A multi-view intelligent seal management and control method based on tensor learning, characterized in that: The method comprises the following steps: Step S1: Collect multi-view data related to seal usage; Step S2: Processing multi-view data using low-rank representation; Step S3: Decompose the data using tensor singular value decomposition, extract key features of seal usage, and obtain tensor nuclear norm; Step S4: construct a multi-view subspace clustering model based on tensor low-rank learning, use the multi-view subspace clustering model to perform cluster analysis on the data, optimize the representation of the affinity matrix, and enable the monitoring and early warning mechanism of seal use; Step S5: Use the optimized affinity matrix to perform real-time spectral clustering to provide early warning and control of abnormal usage patterns; Step S6: Generate a seal usage report to provide decision support for management.

2. According to a multi-view smart seal management and control method based on tensor learning according to claim 1, it is characterized in that: The step S1 collects multi-view data related to the use of the seal, specifically: Multi-view data collection methods include: time dimension data collection, personnel identity data collection, file type data collection, usage frequency and mode data, and device and location data collection. The data of each collection method is regarded as a type, and each type is represented as a view tensor data.

3. According to a multi-view smart seal management and control method based on tensor learning according to claim 1, it is characterized in that: In step S2, the multi-view data is processed using low-rank representation, specifically: Low-rank representation is achieved by minimizing the nuclear norm, mining the potential high-order correlations between seal data samples and identifying abnormal patterns of seal use; The low rank representation is represented by LRR, and the mathematical formula of LRR is defined as follows: ； ； in, Represents a low-rank tensor The nuclear norm of is the error tensor, Represents the error term The Frobenius norm of , is a regularization parameter used to balance the low-rank approximation and data fitting errors; Use the generated tensor data for the actual stamp, This is the seal data after standardization; is a low-rank tensor.

4. The multi-view smart seal management and control method based on tensor learning according to claim 1, characterized in that: In step S3, the data is decomposed by using tensor singular value decomposition to extract key features of seal usage and obtain tensor nuclear norm, specifically: The tensor singular value decomposition is used to extract the key behavior features of the stamp. The tensor singular value decomposition is represented by t-SVD. The t-SVD mathematical formula is defined as follows: ； in, Indicates Frontal slice singular values, tensor nuclear norm is defined as the sum of the singular values ​​in all frontal slices; Indicates the size of the first dimension of the tensor, corresponding to the spatial dimension of the data; Indicates the size of the second dimension of the tensor, corresponding to another dimension of the space; Indicates the size of the third dimension of the tensor, which is the time dimension.

5. The multi-view smart seal management and control method based on tensor learning according to claim 1, characterized in that: In step S4, a multi-view subspace clustering model based on tensor low-rank learning is constructed, and the multi-view subspace clustering model is used to perform cluster analysis on the data, optimize the representation of the affinity matrix, and enable the monitoring and early warning mechanism of seal use, specifically: Step S4-1: construct a multi-view subspace clustering model. The multi-view subspace clustering model is expressed as: （1）； in, Representing a tensor The side slice, i.e. ; Indicates The error matrix of each view is To represent a Line and The real number space of columns; Refers to The feature dimension size of each view; Indicates the number of samples; Force each view The columns have a common and consistent size value; is the Laplacian matrix, and is a diagonal matrix with diagonal entries , represents the clustering indicator matrix, where is the number of clusters; is the regularization parameter; To adjust the clustering index matrix in the objective function The relative importance of a term to other terms, emphasizing the consistency of data points in the clustering structure; is the original multi-view tensor data; To express The trace operation of , represents a graph Laplace regularization; is the weight coefficient; Represents the Schatten-p norm of a tensor; To control the power parameter of the norm, it determines the degree of penalty of the tensor core norm; for The transposed matrix of It is the transpose matrix operation; Step S4-2: Use the multi-view subspace clustering model to perform cluster analysis on the data, optimize the representation of the affinity matrix, and enable the monitoring and early warning mechanism of seal use.

6. A multi-view smart seal management and control method based on tensor learning according to claim 5, characterized in that: The step S4-2: using the multi-view subspace clustering model to perform cluster analysis on the data, optimizing the representation of the affinity matrix, and enabling the monitoring and early warning mechanism of seal use, specifically: By introducing an auxiliary tensor variable , formula (1) is transformed into the following equivalent problem: （2）； Among them, the matrix To express constraints Lagrange multipliers of ; matrix Reflects the constraints The degree of violation of To express constraints The Lagrange multiplier of the tensor Reflects the constraints the extent of the violation; For constraints The penalty item, for Penalty items; is the objective function, and the optimization goal is to minimize the value of this function and find an optimal set of variables ; is the introduced auxiliary tensor variable, which represents the shared features of multi-view subspaces; It is a low-rank self-representation form of multi-dimensional tensor; Formula (2) is decomposed into solving the following four sub-problems: Sub-problem (1): Solve ,variable is fixed; （3）； in, , ; yes No. OK, we get: （4）； At this time, formula (3) becomes: （5）； in, To adjust the clustering index matrix in the objective function The relative importance of a term to other terms, emphasizing the consistency of data points in the clustering structure; is the number of views; is the graph Laplacian matrix The transpose of For the Lagrange multiplier for each view; is the auxiliary variable introduced; for Penalty items; The solution of formula (5) is: （6）； in, is the identity matrix; The Hadamard product multiplies the corresponding elements of two matrices or tensors of the same dimension one by one; Indicates that each element is a matrix The symbols of the corresponding elements in ; Sub-problem (2): Solve ,variable is fixed; （7）； The optimal solution is: （8）； in, express No. List, ; Sub-problem (3): Solve ,variable is fixed; （9）； The optimal solution is: （10）； Sub-problem (4): Solve ,variable is fixed; （11）； Optimal solution Depend on of The eigenvector corresponding to the smallest eigenvalue.

7. The multi-view smart seal management and control method based on tensor learning according to claim 5, characterized in that: The multi-view subspace clustering model includes an adaptive adjustment mechanism of a weighted Schatten-p norm. The mathematical formula of the adaptive adjustment mechanism of the weighted Schatten-p norm is defined as follows: ； in, express No. The largest singular value, Represents the weight vector No. elements, is the order of the Schatten-p norm.

8. The multi-view smart seal management and control method based on tensor learning according to claim 6, characterized in that: In step S5, the optimized affinity matrix is ​​used to perform real-time spectral clustering to provide early warning and control of abnormal usage patterns, specifically: Step S5-1: Setting initial parameters: Set the initial parameters required for the optimization problem, including , convergence condition ; Step S5-2: Loop to update variables: According to formula (6), each view variable Make updates; Calculate and update the error matrix according to formula (8) ; Using the formula Update Lagrange multipliers ; According to formula (10), calculate and update the clustering index matrix ; Step S5-3: Check convergence conditions: Check whether the convergence conditions are met for each view: and ; If the convergence condition is not met, return to step S5-2; if the convergence condition is met, execute step S5-4; Step S5-4: Output result: When the convergence condition is met, the optimized affinity matrix is ​​output, and the final clustering result is obtained through the spectral clustering algorithm.

Images