An evolution pattern recognition method and system based on time-series tensor segmentation

Abstract

The application provides an evolution pattern recognition method and system based on time series tensor segmentation, the method comprising: collecting third-order time series tensor data as input data; dividing the third-order time series tensor into multiple tensor segments through a sliding window; extracting core tensors of the tensor segments based on Tucker decomposition, and clustering based on a time series weighted clustering method to obtain a coarse-grained segmentation result; optimizing the coarse-grained segmentation result based on information gain to obtain a fine-grained segmentation result; extracting correlation features of each segment based on a maximum information coefficient; introducing a nearest neighbor rule, and using an unsupervised three-way clustering algorithm to recognize the evolution pattern corresponding to each segment. The application also constructs an evolution pattern recognition system based on time series tensor segmentation, which is formed by three parts of data acquisition, time series tensor segmentation and evolution pattern mining, and monitoring APP, and forms a complete system, so that the segmentation of time series tensors and the effective recognition of evolution patterns are realized.

Description

Technical Field

[0001] This invention relates to the field of temporal data pattern recognition technology, and in particular to an evolutionary pattern recognition method and system based on temporal tensor segmentation. Background Technology

[0002] In today's digital age, the generation and collection of data are experiencing explosive growth, with time series data being particularly important due to its wide application across multiple fields. Third-order time series tensors, as complex data structures, typically contain multiple time series collected from various locations (such as users and devices) using various types of sensors. These user or location-collected sequences contain rich latent patterns that evolve and change over time, exhibiting complex and dynamic characteristics. For example, in industrial production, equipment operating status data helps predict failures and optimize production processes. Therefore, accurately extracting the pattern information contained in time series tensors is crucial for understanding the underlying patterns in the data, providing effective decision support, and achieving intelligent system management.

[0003] To delve deeper into the pattern information within time series tensors, preprocessing of the data by dividing it into time sub-intervals is typically required. Existing time series segmentation methods mainly fall into three categories: morphological methods, clustering-based methods, and statistical optimization-based methods. These methods have made significant progress in time series segmentation. However, most existing time series tensor segmentation methods focus on univariate time series, with few algorithms specifically designed for multivariate time series tensors. Furthermore, most segmentation methods require manual pre-setting of the number of segments, and the parameter tuning process increases computational complexity and reduces algorithm portability. In addition, the segmentation points obtained by sliding window-based segmentation methods are numerically integer multiples of the window size, causing deviations from the actual results. Moreover, current tensor segmentation algorithms rarely start from the tensor structure itself, neglecting the interaction between the three different dimensions: time, variables, and location.

[0004] In the field of time series data analysis, clustering is a method based on historical data that can be used to determine the potential similarity between individuals and divide them into clusters, effectively reflecting the differences and similarities among large amounts of data. In current clustering methods for time series, the computational complexity and performance are profoundly affected by the number of attribute variables in the sequence. When the number of attribute variables is large and the time span is long, the computational efficiency will decrease significantly. Furthermore, time series tensors simultaneously contain multiple dimensions, including time, geography, and variables. Traditional two-way distance methods cannot effectively solve the problem of segment pattern recognition in tensors, necessitating three-way clustering.

[0005] Another concern in clustering algorithms is determining the number of clusters. Typically, the number of clusters is unknown during clustering, a problem known as the cluster numbering problem. This requires prior knowledge of changes occurring in the study area, or it can be defined by human experts through a supervised procedure. When prior knowledge is unclear or expert experience is lacking, an effective method is to determine the final number of clusters based on the clustering results with different numbers of clusters. This method is simple and effective, but requires repeated experiments and calculations to find a good score, consuming significant computational resources. Summary of the Invention

[0006] To address the aforementioned problems, the present invention aims to provide a method and system for evolutionary pattern recognition based on time series tensor segmentation. This method utilizes an adaptive segmentation algorithm for time series tensor parameters based on time series weighted clustering and an unsupervised three-way clustering algorithm for time series based on correlation feature extraction to achieve effective segmentation and evolutionary pattern recognition of time series tensors.

[0007] To solve the above-mentioned technical problems, the present invention provides the following technical solution:

[0008] On the one hand, an evolutionary pattern recognition method based on temporal tensor segmentation is provided, the method comprising the following steps:

[0009] S1. Collect third-order temporal tensor data as input data;

[0010] S2. Divide the third-order temporal tensor into multiple tensor segments using a sliding window;

[0011] S3. Extract the core tensors of the tensor segments based on Tucker decomposition, and perform clustering based on temporal weighted clustering method to obtain coarse-grained segmentation results;

[0012] S4. Optimize the coarse-grained segmentation result based on information gain to obtain a fine-grained segmentation result;

[0013] S5. Extract the correlation features of each segment based on the maximum information coefficient;

[0014] S6. Introduce the nearest neighbor rule and use an unsupervised three-way clustering algorithm to identify the evolutionary pattern corresponding to each segment.

[0015] On the other hand, an evolutionary pattern recognition system based on temporal tensor segmentation is provided to implement the method described in any of the above embodiments, the system comprising:

[0016] The data acquisition module is used to acquire third-order time-series tensor data as input data;

[0017] The temporal tensor segmentation module is used to divide the third-order temporal tensor into multiple tensor segments using a sliding window; extract the core tensors of the tensor segments based on Tucker decomposition, and perform clustering based on a temporal weighted clustering method to obtain coarse-grained segmentation results; optimize the coarse-grained segmentation results based on information gain to obtain fine-grained segmentation results.

[0018] The evolution pattern recognition module is used to extract the correlation of each segment based on the maximum information coefficient; it introduces the nearest neighbor rule and uses an unsupervised three-way clustering algorithm to identify the evolution pattern corresponding to each segment.

[0019] In addition, the system also includes a monitoring app.

[0020] The beneficial effects of the technical solution provided by this invention include at least the following:

[0021] In this embodiment of the invention, the Temporal Tensor Parameter Adaptive Segmentation Algorithm (PASTWC) based on temporal weighted clustering first uses a sliding window to cluster the tensor segments extracted by Tucker decomposition, obtaining coarse-grained classification results. Then, fine-grained segmentation is achieved through information gain optimization. Next, the evolutionary pattern recognition of the segmented fragments is achieved using unsupervised three-way clustering based on correlation feature representation. This invention's method not only performs accurate three-way clustering on multivariate time series data but also effectively identifies potential evolutionary patterns. Experimental results on multiple artificial and real datasets demonstrate that the proposed algorithm exhibits superior performance and can effectively complete the segmentation of time series tensors and the identification of evolutionary patterns.

[0022] This invention utilizes a data acquisition module to better obtain the monitoring data that needs to be mined, and constructs a monitoring APP for convenient real-time monitoring. The data acquisition module, the evolutionary pattern recognition method based on time-series tensor segmentation, and the monitoring APP are interconnected to form a complete system. Attached Figure Description

[0023] To more clearly illustrate the technical solutions in the embodiments of the present invention, the accompanying drawings used in the description of the embodiments will be briefly introduced below. Obviously, the accompanying drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0024] Figure 1 This is a flowchart of an evolution pattern recognition method based on temporal tensor segmentation provided by an embodiment of the present invention;

[0025] Figure 2 This is a schematic diagram of the coarse-grained segmentation process provided in an embodiment of the present invention;

[0026] Figure 3 This is a schematic diagram of information gain and objective function provided in an embodiment of the present invention;

[0027] Figure 4 This is a schematic diagram of the time-series weighted clustering process provided in an embodiment of the present invention;

[0028] Figure 5 This is a schematic diagram of fine-grained segmentation provided in an embodiment of the present invention;

[0029] Figure 6 This is a schematic diagram of temporal segment pattern recognition based on unsupervised three-way clustering provided in an embodiment of the present invention. Detailed Implementation

[0030] To make the objectives, technical solutions, and advantages of the embodiments of the present invention clearer, 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, not all, of the embodiments of the present invention. All other embodiments obtained by those skilled in the art based on the described embodiments of the present invention without creative effort are within the scope of protection of the present invention.

[0031] First, the basic theory used in the embodiments of this invention will be explained as follows:

[0032] (1) Loss function based on information entropy

[0033] In information theory, Shannon entropy (H) quantifies the uncertainty in a probability distribution, reflecting how much information can be obtained from a given quantity. Higher entropy indicates greater uncertainty, meaning less information can be inferred about the quantity of interest.

[0034]

[0035] Where p(n) represents the probability of n-1 outcomes, and H(·) represents the entropy of the probability distribution. Given the segmentation of a time series, a loss function l based on information entropy is introduced, which measures the reduction in entropy when the time series is segmented according to a specific segmentation method. More specifically, l reflects the expected reduction in entropy when the time series X is further segmented, and is defined as follows:

[0036]

[0037] Where K+1 is the number of segments, |X k | represents the length of the k-th segment. H(X) is a segment of the entire time series and represents the entropy of the entire time series.

[0038] (2) Hierarchical clustering algorithm TW-FINCH

[0039] TW-FINCH is a hierarchical clustering algorithm used to group semantically consistent graphical frames in a video. It aims to infer the boundaries of semantically consistent visual concepts in a video to achieve action segmentation. It constructs a graph where nodes represent members, p and q in the formula, and each edge weight W(p,q) represents the semantic proximity between two members.

[0040]

[0041] W(p,q)=Ω f (p,q)·Ω t (p,q) (4)

[0042] Where Ω f (p,q) is the spatial adjacency matrix, Ω t (p,q) is the temporal adjacency matrix. TW-FINCH merges semantically close members and repeats this process until a hierarchical structure of partitions is completed, where different partitions are different action segments in the video.

[0043] (3) Maximum information coefficient

[0044] Mutual information refers to the degree to which the information content of one random variable decreases when the information content of another random variable is known. The greater the correlation between the two variables, the greater the mutual information value; when they are independent, the mutual information value is 0. The main idea of ​​the Maximum Information Coefficient (MIC) is: assuming there is some correlation between the two variables, a certain method is used to divide the scatter plot of the joint samples of the two variables into a grid, so that most points fall into a few cells. This solves the problem of calculating the joint probability in mutual information.

[0045] Definition 1: Given a finite set of ordered pairs Given positive integers a and b, and the sample size num, let the dataset D be distributed in a coordinate system. Divide the dataset D into an a*b grid using a and b. The variable values ​​in D fall into the grid, resulting in the corresponding probability distribution D|. G By changing the mesh partitioning, different mutual information values ​​can be obtained, with the largest mutual information value shown below:

[0046] I * (D,a,b)=maxI(D| G (5)

[0047] Definition 2: For ease of comparison, the maximum mutual information value is normalized so that its value is between [0,1]:

[0048]

[0049] Definition 3: Suppose that the number of ordered pairs in dataset D is num, then the formula for the MIC of two variables A and B in the set is defined as follows:

[0050]

[0051] Where B(num) = num ε This represents the upper limit of the mesh division. The constant ε is set based on empirical values ​​or scale. A value that is too large will lead to false correlations, while a value that is too small will only detect simple correlations. Typically, when B(num) = num 0.6 The effect is best at that time.

[0052] Unlike traditional multivariate time series analysis, the purpose of this invention is to achieve segmentation and evolution pattern recognition of time series tensors, which can also be viewed as the collaborative segmentation and pattern recognition of multiple multivariate time series. A time series tensor is essentially a multidimensional array that can meaningfully represent a wide range of spatiotemporal datasets. Table 1 lists the symbols used in this invention and their descriptions.

[0053] Table 1 Symbols and Definitions

[0054]

[0055]

[0056] This invention provides an evolutionary pattern recognition method based on temporal tensor segmentation, such as... Figure 1 As shown, the processing flow of this method includes the following steps:

[0057] S1. Collect third-order temporal tensor data as input data.

[0058] The third-order time series tensor data refers to multivariate time series data collected by sensors, including data from fields such as information retrieval data, environmental monitoring data, human activity data, financial analysis data, and building energy consumption data.

[0059] S2. Divide the third-order temporal tensor into multiple tensor segments using a sliding window.

[0060] The collected third-order temporal tensor X∈R M×N×T Given the sampling results of N variables for M users / locations at T time points; based on the changing trend of the third-order time series tensor, find K segmentation points through a sliding window, and divide X into K+1 tensor segments along the time dimension.

[0061] Specifically, X can be viewed as a stack of multiple multivariate time series data from M users / locations, denoted as X = {X} d(1) ,X d (2),...,X d(m) ,...,X d(M)},in:

[0062]

[0063] Let d(m) represent the measured value of variable v(n) at time t(i), where m = 1,...,M; n = 1,...,N; i = 1,...,T. The purpose of this invention is to adaptively find K segmentation points {b1,b2,...,b...} based on the trend of time-series tensor changes. k ,...,b K Divide X into K+1 segments along the time dimension, denoted as X = {X1, X2, ..., X...} k ,...,X K ,X K+1}, where b k It is the kth segmentation point, and X k This is the k-th segment, where k = 1, ..., K. After segmentation, segment X... k The corresponding time interval is s k , and s k =[b k-1 ,b k ), with a length of |s k |=|b k -b k-1 |

[0064] The set of time intervals obtained by segmentation is:

[0065] S = {s1, s2, ..., s} k ,...,s K ,s K+1}={[b0,b1),[b1,b2),…,[b k-1 ,b k ),…,[b K-1 ,b K ),[b K ,b K+1 ]},in

[0066] 1 = b0 < b1 < ... < b k <...<b K <b K+1 =T. Segment X k It can be represented as in:

[0067]

[0068] It is a multivariate time series of user / location d(m) in the kth segment. It is its nth variable v(n) at time t(b) k-1 The measured values ​​of (+h), m=1,2,...,M; n=1,2,...,N; h=1,2,...,(|s k |-1).

[0069] After the temporal tensor data has been collaboratively divided into K+1 time sub-intervals using the segmentation algorithm, three-way clustering is still needed to obtain complete pattern information from the temporal tensor. Clustering these subsequences from different users and locations requires dividing them into C spatial clusters under the dual constraints of spatial continuity and the discovery of discontinuous spatial repetitive geographic patterns, ensuring that the data within each cluster is as similar as possible, while the data between clusters is as different as possible.

[0070] S3. Extract the core tensors of the tensor segments based on Tucker decomposition, and perform clustering based on temporal weighted clustering method to obtain coarse-grained segmentation results.

[0071] like Figure 2 As shown, Figure 2 In the diagram, (a) represents the feature extraction process, (b) represents the segmentation based on temporal weighted clustering, and (c) represents the coarse-grained segmentation result of the temporal tensor. Tensor segments obtained through a sliding window are decomposed using Tucker decomposition to extract the corresponding core tensors. These core tensors are then clustered using a parameter-free temporal weighted method. During the clustering process, each core tensor is initially treated as a separate cluster. To facilitate fast clustering, a spatiotemporal graph representation matrix is ​​constructed, where the result of each clustering iteration effectively represents a segmentation result. The information gain of the current segmentation result is then calculated as follows: Figure 3 As shown in (a) and (b), after several iterations, all individuals are merged into a single cluster. The segment with the highest objective function value (indicated by an asterisk) is selected as the final segment. Figure 2 In this process, the fourth iteration was selected as the final coarse-grained segmentation result, such as... Figure 2 As shown in (c). This clustering method is very efficient, typically requiring only 3 to 6 iterations to obtain the final result.

[0072] Specifically, assume that the third-order temporal tensor X follows a sliding window... It is divided into J windows along the time dimension, represented as X = {X1, X2, ..., X...} j ,...,X J The j-th window is X. j ∈ M×N×W ,j∈1,...,J,have It is the time series of user / location d(m) in the j-th window:

[0073]

[0074] in Let v(n) represent the value of the user / location variable d(m) at time u in the j-th window, where n = 1, 2, ..., N; m = 1, 2, ..., M; u = 0, 1, 2, ..., (W-1); j = 1, 2, ..., J.

[0075] To reduce computational overhead and capture the geometric integrity of tensor data, this invention utilizes Tucker decomposition to extract features from the resulting tensor segments. A third-order tensor X... j ∈ M×N×W It can be approximated as:

[0076]

[0077] A in formula (11) j ∈ M×P B j ∈ N×Q C j ∈ I×R It is tensor X j Factor matrices of different dimensions and P≤M, Q≤N, R≤I. This corresponds to the core tensor, which can be considered as a dimensionality-reduced feature of the original tensor, and each factor matrix can be considered as a principal component in different dimensions. The optimization objective of this invention is to minimize the Frobenius norm error between the original tensor segment and the reconstructed tensor segment, given a predefined rank of the core tensor.

[0078]

[0079] HOSVD is used as an initialization step to obtain coarse estimates of the factor matrices and core tensors. Subsequently, HOOI iterates over these factor matrices to refine them, resulting in a more accurate and efficient decomposition. This combination leverages the global decomposition capability of HOSVD and the local optimization capability of HOOI, thereby improving decomposition accuracy while reducing computational complexity. Furthermore, to better capture the relationships between multivariate time series at different locations, orthogonality constraints are not imposed on the time dimension, but rather on the variable dimension and the user / location dimension.

[0080] Furthermore, the pre-order rank of the core tensor (i.e. The values ​​of P, Q, and R in the equation affect the decomposition results and computational complexity. An ideal rank should ensure that the resulting core tensor is of a suitable size and can accurately reconstruct the original tensor. Therefore, to adaptively determine the j-th tensor segment g...j The ideal rank of the core tensor, based on the optimization objective of formula (12), is:

[0081]

[0082] In this embodiment of the invention, a Bayesian optimization algorithm is used to determine the optimal tensor rank. When determining the rank of the core tensor, the weight α controls the trade-off between error and rank, and is set to 0.0005.

[0083] Based on formula (13), the tensor segment X can be obtained. j The ideal rank of the core tensor of all tensor segments is fixed, but the ideal rank of other tensor segments may differ. To determine the final ideal rank of the core tensor of all tensor segments, in each A window is randomly selected within the window, and the optimal rank of its core tensor is calculated. Then, in... The rank with the highest frequency among all tensor segments is selected as the final core tensor rank for decomposition.

[0084] For ease of description, we will refer to the tensor segments {X1,X2,....,X...} j ,...,X J The resulting core tensor is denoted as {g1, g2, ..., g}. j ,...,g J}, where when the time dimension length t of the last interval divided by the sliding window is less than the rank R of the time dimension of the determined core tensor g, the last interval tensor X is... J With X J-1 After concatenation in the time dimension, tensor decomposition is performed to obtain the core tensors. At this point, there are a total of J-1 core tensors. Let J be the core tensors. w =J-1; when the interval tensor X J With X J-1 When merging is not required to solve the core tensor, there are a total of J core tensors, denoted as J. w =J. Based on this, the resulting core tensor is denoted as J.

[0085] Furthermore, inspired by TW-FINCH, this invention proposes using a spatiotemporal graph representation matrix G to encode feature space and temporal proximity, with the temporal series as the modulation factor. This matrix is ​​based on the similarity matrices Ω in the feature space and spatiotemporal space, respectively. f and Ω tConstruction. Using G, connect all core tensors to their nearest neighbors and group them into clusters. The initial result divides the original tensors into segments. Then, similar clusters are repeatedly merged to obtain further consecutive segments. Each step forms a group of previously obtained clusters, and the recursion terminates when only one cluster remains. During the iteration, the result with the largest rate of change in information gain is selected as the final segmentation result.

[0086] Specifically, in the segmentation process, a matrix is ​​first defined. Node V is the extracted core tensor Edge E describes the proximity of the core tensor in the feature space and time dimension. Here, assume G(p,q) represents g. p and g q The corresponding edges between them, where g p g q belong And 1≤p≤J W , 1≤q≤J W .

[0087] To obtain the spatiotemporal graph representation matrix G, we first construct similarity matrices in the feature space. The edge weights are calculated as follows:

[0088]

[0089] Where dis(g) p ,g q ) is g p and g q The Euclidean distance between them. To accommodate the subsequent time-weighted matrix, Normalized within the range [0,1], it is represented as:

[0090]

[0091] Then the matrix Transform into the final spatial similarity matrix Ω f :

[0092]

[0093] Secondly, construct another time similarity matrix. The edge weights are calculated as follows:

[0094]

[0095] Ω t (p,q) represents g p and g q The time difference between them. Time series matrix Ω tThe modulation factor used as the feature space map, where Ω t (p,q) provides a weighting mechanism. Finally, Ω f With Ω t Combined, we obtain the graph matrix of timing modulation.

[0096] Ω(p,q)=Ω f (p,q)·Ω t (p,q) (18)

[0097] The final spatiotemporal graphical representation matrix is ​​obtained based on Ω. The feature space and temporal distance are encoded. For any node in Ω, its shortest edge is set to 1, while all other edges are set to 0.

[0098]

[0099] Clustering is achieved by iteratively connecting each node to its nearest neighbor using a spatiotemporal graph representation matrix G. For one iteration, a simple clustering result is as follows: Figure 4 As shown. The resulting spatiotemporal representation matrix G is as follows. Figure 4 As shown in (a), each row contains only a few elements with a value of 1, while the other elements are set to 0. This means that the shortest edge of each node is set to 1, and all other edges are set to 0. Figure 4 (b) illustrates multiple clusters by connecting all nodes to their nearest neighbors. For example, nodes 10, 11, and 12 are grouped into one cluster because they are closest to each other. Each node represents a core tensor extracted from a tensor segment, so the clustering results can be mapped back to the original tensors for segmentation. Specifically, each tensor segment has a cluster label, so segmentation can be achieved by concatenating tensor segments with the same label, such as... Figure 4 As shown in (c).

[0100] After the first iteration, a refined TW-FINCH clustering method is applied to the mean of all obtained clusters. In fact, each iteration of clustering corresponds to a segmentation result. The process terminates when only one cluster remains. Finally, the segmentation result with the largest rate of change in information gain is selected as the final coarse-grained segmentation.

[0101]

[0102]

[0103] Among them l (e) Let k be the information gain of the e-th iteration. The k-th segment of the e-th iteration can be represented as: The corresponding time interval is s (e) (k). yes User / Location d (m) The probability of the nth variable occurring. Based on formula (20), the coarse-grained set of segmentation points can be obtained.

[0104] S4. Optimize the coarse-grained segmentation results based on information gain (information entropy) to obtain fine-grained segmentation results.

[0105] Since the segmentation results after clustering are based on a sliding window, the position of each segment point is a multiple of the window size, and optimization is needed within its left and right windows.

[0106] To obtain the fine-grained segmentation results, the coarse-grained segmentation point set is needed. Further optimization was performed for the nth segment point. Its two adjacent windows are merged into one interval. The tensor segments contained within this time interval are represented as follows: Optimized segment point b k This is obtained by maximizing the information gain l(k):

[0107]

[0108] Where l(k) is the segmentation point b k Information gain at time, b k tensor It is divided into two parts, the left side is The one on the right is They are X kl and X kr User / Location d (m) The probability of the nth variable occurring. Based on formula (24), the set of fine-grained segmentation points {b1,b2,...,b} can be obtained. k ,...,b K}

[0109] Figure 5 To optimize the process of segmentation points. In Figure 5 In (a), to optimize the coarse-grained segmentation point b1, the time interval of the two windows surrounding point b1 is selected, with a total length of 2W. Then, a greedy search is employed. Figure 5In (b), assuming the interval contains 12 time points, the information gain is calculated for each time point as a segmentation point. Since the number of segmentation points is fixed, the point with the highest information gain is the optimal choice. Initially, the segmentation point is point b1, i.e., the sixth point is selected. However, after fine-grained segmentation, the third point is selected as the new segmentation point because it produces the highest information gain. Finally, the fine-grained segmentation result is as follows: Figure 5 As shown in (c).

[0110] The overall process of the time series tensor parameter adaptive segmentation algorithm (PASTWC) based on time series weighted clustering provided by this invention is shown in Table 2.

[0111] Table 2 PASTWC Algorithm Flow

[0112]

[0113] Next, the unsupervised three-way clustering algorithm based on correlation feature extraction (UTCFTS) is used to effectively identify the evolution patterns of the segmented fragments. The overall process is as follows: Figure 6 As shown, features are first extracted from each segment based on correlation, and then unsupervised clustering is performed on the extracted features to identify the pattern category corresponding to each segment.

[0114] S5. Extract the correlation of each segment based on the maximum information coefficient.

[0115] Among them, correlation includes variable correlation, time correlation, and user / location correlation.

[0116] (1) Variable correlation extraction

[0117] For a location d(m), there are K+1 segments, and (K+1)*M subsequence segments need to be clustered.

[0118] The k-th segment of location d(m) can be represented as:

[0119]

[0120] Where b k-1 and b k For two adjacent segmentation points, A multivariate time series segment containing N distinct variables can be viewed as a subsequence segment containing multiple univariate variables, encompassing the interaction information between the time and variable dimensions. Considering that the covariance coefficient is only sensitive to linear relationships and struggles to describe the numerous nonlinear relationships among variables, the maximum mutual information (MIC) matrix is ​​used to measure the correlation between different variables. The maximum mutual information coefficient matrix of this multivariate time series is... It can be calculated using the following formula:

[0121]

[0122] yes In the middle variable n and variable Maximum mutual information coefficient between It is a symmetric matrix, that is The elements of the upper triangular matrix, excluding the diagonal elements, completely contain the relationship between any two distinct variables. Therefore, the intrinsic correlation between any two distinct variables can be represented by a maximum mutual information sequence. To describe, L = (N-1)N / 2. We can know that... The length of the cross-information coefficient depends on the number of variables N, not the length of the subsequence segments. That is, two multivariate time series of different lengths but with the same number of variables, after variable correlation extraction, are sequences with the maximum mutual information coefficient of equal length. Therefore, fast metrics such as Euclidean distance can be used to measure the similarity between different subsequence segments.

[0123] (2) Time correlation extraction

[0124] According to equations (5), (6), (7), (29), and (30), multivariate time series of different lengths can also be converted into maximum mutual information coefficient sequences of the same length in this way, allowing them to be compared and analyzed under the same standard. After constructing the covariance sequence of each multivariate time series (MTS), principal component analysis (PCA) can be used to analyze the composition of these maximum mutual information coefficient sequences, thereby extracting features that can represent the original multivariate time series.

[0125] For the previously obtained multivariate time series dataset Each subsequence It can be converted into the corresponding maximum mutual information coefficient sequence. In this way, the entire dataset can be represented as a multivariate time series with maximum mutual information coefficients.

[0126] Next, PCA analysis is performed on the maximum mutual information coefficient to obtain the eigenvector matrix V. C and eigenvalues ​​Λ C and principal component Y C Eigenvalue Λ C This represents the contribution of each eigenvector to the variance in the dataset, while the eigenvector V... C This indicates the main direction of change in the data.

[0127] In PCA, the eigenvectors corresponding to the top λ largest eigenvalues ​​are typically selected to construct a new feature space. This preserves the main information of the data. For ease of representation, assume there is a time series X∈R. n*m PCA can be used to transform the original multivariate time series X∈R n*m Transform into a new multivariate time series Y∈R n*λ Its dimension is reduced from n*m to n*λ, where λ is the number of principal components selected and retained. The transformed sequence Y can be represented by the following formula:

[0128] Y n×λ =X n×m V m×λ (31)

[0129] In the algorithm of this invention, after feature extraction of the multivariate time series with maximum mutual information coefficients, each original multivariate time series... It is transformed into a feature vector, which consists of the first λ rows of the V matrix, representing the features in the dataset D. The features are then extracted. Ultimately, the entire dataset D is transformed into a feature matrix F with dimensions U*λ, where U is the number of sequences in the dataset, U = M*(K+1). In the feature extraction process of this invention, the covariance matrix Σ... C Principal component analysis is performed to obtain the eigenvector matrix V. C and eigenvalues ​​Λ C Eigenvector matrix V C Each column represents an eigenvector, and these eigenvectors define a new feature space. Eigenvalues ​​Λ C This represents the variance contribution of the data in each eigenvector direction, which is the amount of information.

[0130] To select the most important eigenvectors, they are typically sorted by their eigenvalues, and the eigenvectors corresponding to the top λ largest eigenvalues ​​are chosen. These most important eigenvectors best represent the information in the original data. For each multivariate time series in the dataset... Its corresponding maximum mutual information coefficient sequence It can be projected onto a new feature space formed by these λ most important feature vectors.

[0131]

[0132] Here V C (:,1:nk) represents a matrix formed by selecting the first λ columns from VC. Using this formula, each multivariate time series... It is transformed into a λ-dimensional eigenvector Ultimately, a dataset D containing U subsequences can be represented by a feature matrix F through dimensionality reduction, which contains the features of each multivariate time series in the dataset. eigenvectors

[0133]

[0134] In this process, each multivariate time series It can be effectively represented as a λ-dimensional feature vector, with the dimension ranging from (b k -b k-1 The 1×N is reduced to 1×λ, which is typically much smaller than the original dimension.

[0135] (3) User / Location Relevance

[0136] Since multiple time series originate from different users or locations, the influence of time series data across different locations must be considered. Data typically shows a positive correlation only within a certain spatial / locational range; therefore, sequences geographically close will have a stronger correlation. For different users, sequences obtained from information collected from users with similar behaviors will be more similar. Therefore, when geographical coordinates are missing, the similarity of sequences can be used to determine the proximity of users and the spatial dimension. Specifically, for a complete time series X collected from different locations or users in a time series tensor... d(m) and For variable n, the Maximum Information Coefficient (MIC) is used as a measure. This coefficient allows us to discover the degree of correlation between two time series on variable n without assuming a linear relationship, where n = 1,...,N.

[0137]

[0138] because For multivariate time series, we can define the overall user or location m. The degree of correlation between them is:

[0139]

[0140] According to equation (35), for the time series X at the m-th location... d(m) An M-dimensional feature can be obtained. This reflects the correlation between different time periods in terms of location or user.

[0141] S6. Introduce the nearest neighbor rule and use an unsupervised three-way clustering algorithm to identify the evolutionary pattern corresponding to each segment.

[0142] Traditionally, cluster-based pattern recognition has been influenced by initialization and requires a given number of clusters prior to the algorithm. To construct a clustering algorithm that requires no initialization and automatically identifies the number of pattern types, this invention utilizes the concept of entropy. First, let the scaling factor α... c Let U be the probability that a sample belongs to class c (the input here is the feature of the sample, hereinafter referred to as the sample; as mentioned above, there are a total of U = (K+1)*M samples, which will be referred to as U samples for ease of representation), and c = 1,...,C. Then... This becomes the average value of the information, also known as the proportionality coefficient α. c The corresponding entropy value.

[0143] When α c When α = 1 / C, the entropy reaches its maximum value. Therefore, this term is added to the objective function J(z,A) of the k-means algorithm as a penalty. Then, a pattern is constructed to estimate α by minimizing the entropy. c to obtain α c The most information. Minimize. Equivalent to maximizing Therefore, a new objective function is proposed as follows:

[0144]

[0145] To determine the number of clusters, let another entropy term be considered, with membership z as an example. uc Let z represent the membership degree of the u-th sample to the c-th class. uc and proportionality coefficient α c Combining them, we get term z uc lnα c The overall objective function is as follows:

[0146]

[0147] The Lagrange of the equation is:

[0148]

[0149] Membership degree z uc The update equation is as follows:

[0150]

[0151] Cluster center a c The update equation is as follows:

[0152]

[0153] As shown in (39), the membership degree can be either 1 or 0. Therefore, z can be... uc xu Let represent the sample in the c-th cluster. Equation (40) updates the cluster center not simply by selecting the average value of the cluster, but by selecting the sample that is closest to other samples from the average value of the cluster as the new cluster center.

[0154] By comparing the proportionality coefficient α c Taking the partial derivatives of the Lagrange function, we can obtain the following update equation for the (t+1)th iteration:

[0155]

[0156] For the c-th mixing ratio that is currently being updated All categories need to be considered To mitigate the impact, a weighted average of all proportionality coefficients is introduced into equation (41). This ensures that the algorithm is updated. This allows us to consider the distribution of all categories in the entire dataset, thus more accurately reflecting the probability of each data point belonging to each category. For the c-th mixing ratio... if Less than the sum of the weighted averages of all proportionality coefficients The new mixing ratio In other words, in the next iteration, the smaller proportion will decrease, the larger proportion will increase, and then competition will occur.

[0157] For certain iterative processes, cluster 1 < k ≤ C (t) In other words, if α c <0 or α c If the ratio is ≤1 / U, it is considered an illegal proportionality coefficient. The clusters corresponding to these coefficients are discarded, and the number of clusters C in the current iteration is then calculated. (t) Updated to

[0158]

[0159] Where |{}| represents the cardinality of the set {}. After updating the cluster number c, the remaining mixing ratio and corresponding It needs to be renormalized to:

[0160]

[0161] Next, we will examine equation (37) and The γ and β parameters of the two terms are updated. The parameter update equations for γ and β are as follows:

[0162]

[0163]

[0164] The algorithm can automatically reduce the number of clusters based on the parameter γ, while simultaneously obtaining estimates of the corresponding parameters. Furthermore, if... and If the difference between them is small, then β must be increased to enhance its competitiveness. If and If the differences between the clusters are large, β will decrease to maintain stability. Since β can jump at any time, when the cluster number C is stable, setting β = 0 means that C will no longer decrease.

[0165] To accelerate the clustering process and improve its accuracy, a nearest neighbor rule is introduced to speed up the convergence of clusters. In the initial iteration, all sample points are treated as a single cluster center. During the merging process, based on the calculation results of equation (35), for each sample x... u ,u=1,...,U, that is choose The sequence X of locations m', m'∈[1,K] whose maximum mutual information coefficient is greater than 2 / 3 of the values ​​of other locations. d(m') The nearest neighbor set is formed by the τ segments smaller than a given threshold.

[0166] Based on the number of nearest neighbors for each sample, spatial points are sorted in descending order, and a list O is created. The point with the most neighbors in list O is selected as the first cluster center, and all its neighbors are assigned to that cluster. For subsequent points in list O, if a point does not belong to any cluster, and more than 50% of its spatiotemporal neighbors also do not belong to any cluster, then that point is considered the center of a new cluster.

[0167]

[0168] This method accelerates the first iteration. Furthermore, after each iteration, for each sample x... u ,Right now Find the original location or user time series X corresponding to the MIC coefficients obtained from formula (35). d(m) Nearest location and user time sequence The closest fragment Make a judgment for i,j∈1,...,K+1. Assume... In the current iteration, it belongs to cluster c i The cluster center is a i , In the current iteration, it belongs to cluster c j The cluster center is a j :

[0169]

[0170] This is equivalent to optimizing the cluster after each iteration based on the correspondence between the closest segments of different users or locations, taking into account the association information of different locations or users.

[0171] The overall process of the Time Series Unsupervised Three-Way Clustering Algorithm (UTCFTS) based on correlation feature extraction is shown in Table 3.

[0172] Table 3 UTCFTS Algorithm Flow

[0173]

[0174]

[0175] In summary, this invention addresses the challenges of existing technologies by proposing a parameter adaptive segmentation algorithm based on temporal-weighted clustering for time series tensors (PASTWC). This algorithm first divides a third-order time series tensor into multiple tensor segments using a sliding window. Then, combining Bayesian search and Tucker decomposition, the tensor segments obtained by the sliding window are decomposed into core tensors as input. This feature extraction method based on data compression can capture the underlying structure of the tensor in a low-dimensional form. Coarse-grained segmentation is then achieved based on an improved temporal-weighted clustering algorithm (TW-FINCH), eliminating the need for manually setting the number of segments. Considering that the segmentation result is only an integer multiple of the sliding window size, PASTWC uses the idea of ​​information gain to achieve fine-grained optimization within the tensor segment intervals of the neighborhood on both sides of the coarse-grained segmentation point.

[0176] To address the shortcomings of existing clustering methods, this invention proposes an unsupervised tri-clustering algorithm for timeseries based on correlation feature representation (UTCFTS) to identify latent temporal patterns in time series tensors. First, correlation features along the variable dimension and the time dimension are extracted from different multivariate time series. To measure the nonlinear relationship between different variables, the maximum information criterion (MIC) is introduced to replace the original covariance coefficient. The extracted features are combined with the UK-means algorithm, and an improved MIC is used to set spatial rules to assist in achieving tri-clustering. This enables accurate tri-clustering of multivariate time series data and effectively identifies potential evolutionary patterns.

[0177] The main contributions of this invention are as follows: (1) An adaptive temporal tensor segmentation algorithm is proposed based on the tensor structure itself. This algorithm realizes the collaborative segmentation of temporal tensors and does not require any manual parameter setting during the segmentation process; (2) Considering that the segmentation point position based on the sliding window segmentation is only an integer multiple of the sliding window size, the segmentation point position is finely optimized based on information entropy, realizing fine-grained segmentation of temporal tensors; (3) The feature extraction method based on correlation is used to comprehensively extract the variable dimension and time dimension features in the sub-segments obtained by segmentation, transforming the high-dimensional and redundant multivariate time series into a simple vector representation, reducing the computational cost of clustering and improving execution efficiency; (4) An improved MIC is used to judge the correlation of different users or different locations, which can effectively solve the spatial dimension information measurement under the lack of geographical information, and based on this, spatial rules are constructed to assist in the realization of three-way clustering; (5) The existing advantages of the Uk-means algorithm are retained, the extracted features are used as input, irrelevant features are eliminated, and the optimal number of clusters is automatically found to realize the recognition of evolution patterns.

[0178] Accordingly, embodiments of the present invention also provide an evolutionary pattern recognition system based on temporal tensor segmentation, the system comprising:

[0179] The data acquisition module is used to acquire third-order time-series tensor data as input data;

[0180] The temporal tensor segmentation module is used to divide the third-order temporal tensor into multiple tensor segments using a sliding window; extract the core tensors of the tensor segments based on Tucker decomposition, and perform clustering based on a temporal weighted clustering method to obtain coarse-grained segmentation results; optimize the coarse-grained segmentation results based on information gain to obtain fine-grained segmentation results.

[0181] The evolution pattern recognition module is used to extract the correlation of each segment based on the maximum information coefficient; it introduces the nearest neighbor rule and uses an unsupervised three-way clustering algorithm to identify the evolution pattern corresponding to each segment.

[0182] The system in this embodiment can be used to execute Figure 1 The technical solutions of the method embodiments shown are similar in principle and in effect, and will not be described again here.

[0183] Furthermore, the data acquisition module includes sensor components, a serial port server, a management workstation, an interface server, an information center storage server, and a platform core switch.

[0184] The sensor components are connected to the management workstation via the serial port server. The management workstation is connected to the interface server. The interface server is connected to the information center storage server via a unidirectional isolation gateway. The information center storage server is connected to the platform core switch via optical fiber. The platform core switch is connected to multiple cloud computing nodes and web servers.

[0185] Furthermore, the system also includes a monitoring app, which comprises a client, a server, and a system management backend.

[0186] The client is used for user registration and login, online query, modification, and logout; the server is used for registration and login verification, as well as data transmission, addition, modification, and deletion functions; the system management backend is used for database management.

[0187] The evolutionary pattern recognition system based on temporal tensor segmentation constructed in this invention consists of three main interconnected parts: data acquisition, temporal tensor segmentation and evolutionary pattern mining, and a monitoring APP, forming a complete system that provides a useful reference for research and development in related fields.

[0188] In the experimental verification section, this invention first introduces the evaluation index of segmentation accuracy, then uses multiple artificially synthesized datasets with real segmentation information (ground truth) and real temporal tensor datasets to verify the segmentation accuracy of the proposed method, and finally uses real multivariate spatiotemporal datasets to evaluate the effectiveness of the proposed method in collaborative segmentation.

[0189] In this embodiment of the invention, multiple time-series tensor datasets are used for verification. The experimental data include multivariate time-series datasets from various industries such as information retrieval, environmental monitoring, human activity identification, financial analysis, and building energy consumption. Specific information about the experimental data is shown in Table 4.

[0190] Table 4 Experimental Dataset

[0191]

[0192]

[0193] Choosing appropriate evaluation metrics is crucial in assessing the accuracy of segmentation results. Two metrics will be used here to evaluate the segmentation algorithm.

[0194] (1) F1. Assume the set of segmentation points obtained from the experiment is {b} k |k=1,2,...,K}, the set of true segmentation points is: When the dataset used to evaluate the segmentation method itself has real segmentation information, the widely used evaluation metric is F1:

[0195]

[0196] Using F1 directly as the evaluation metric tends to penalize reasonable segmentation schemes that are close to the true segmentation points. To address this issue, this invention defines a set of true segmentation point intervals. For each segment position, where The center is 0.025T, the radius is 0.025T, and T is the length of the time series. If the set of segment points obtained from the experiment is {b} k |k=1,2,...,K} falls within the range of |k=1,2,...,K} Within the specified range, a result can be marked as a true positive (TP); otherwise, it is marked as a false positive (FP). If the true segmentation interval does not contain any segmentation positions obtained by the algorithm, it is marked as a false negative (FN).

[0197] (2) Average p-lag. In some special cases, such as when the experimental segmentation points obtained by two segmentation methods to be evaluated both fall within a reasonable segmentation interval, F1 will be unable to further evaluate their merits. To solve this problem, we mainly evaluate the accuracy of the segmentation results by calculating the average absolute distance between the experimental segmentation point and its corresponding nearest real segmentation point. Furthermore, to gain an intuitive understanding of the algorithm's performance on different datasets, we divide the average absolute distance by the effective radius to obtain the average p-lag:

[0198]

[0199] GK represents the number of segments marked as true positives (TPs). The lower the average p-lag, the more accurate the segmentation results. In this way, the average p-lag can complement F1, thereby evaluating the accuracy and effectiveness of the segmentation results.

[0200] To verify the effectiveness of the proposed algorithm, it was compared with several time-series segmentation algorithms, including:

[0201] (1)DMM: A time-series segmentation algorithm based on minimum coding loss.

[0202] (2) TTW-FINCH: Directly extend TW-FINCH to the temporal tensor level, where the input is the original data of each time frame.

[0203] (3) GGS: A classic segmentation method based on Gaussian fitting, which requires manual input of the number of segments. Here, when processing time series tensors, the tensor is directly unfolded as input and the actual number of segments is directly input.

[0204] (4) IGTTS: Extend the IGTS algorithm to the temporal tensor level.

[0205] The comparison results of different algorithms are shown in Tables 5 and 6:

[0206] Table 5 shows the F1 scores of PASTWC and other algorithms on different datasets.

[0207]

[0208] Table 6 shows the average p-lag of PASTWC and other algorithms on different datasets.

[0209]

[0210] Tables 5 and 6 show the segmentation performance evaluation of different segmentation methods. Obviously, compared with other comparison methods, PASTWC achieved the highest F1 score and the lowest average number of p-lags. Furthermore, PASTWC's average segmentation performance on multiple datasets is better than other comparison methods, achieving the best overall performance and demonstrating high segmentation effectiveness and accuracy.

[0211] Furthermore, this invention evaluated the proposed pattern recognition algorithm based on 12 public MTS datasets, 10 of which came from the UEA dataset and 2 were constructed from human activity datasets. Details of the datasets are shown in Table 7.

[0212] Table 7. Description of Dataset Information

[0213]

[0214] The experimental evaluation indicators are as follows:

[0215] (1) RAND Index (RI);

[0216]

[0217] The Rand Index (RI) measures the similarity between two clustering results. It evaluates the quality of clustering by comparing the relationship between sample pairs with true labels and predicted labels.

[0218] True positive (TP) indicates that two elements belong to the same category in the true classification and are also assigned to the same cluster in the clustering result; false positive (FP) indicates that two elements do not belong to the same category in the true classification but are assigned to the same cluster in the clustering result; true negative (TN) indicates that two elements do not belong to the same category in the true classification and are assigned to different clusters in the clustering result; false negative (FN) indicates that two elements belong to the same category in the true classification but are assigned to different clusters in the clustering result. The value of RI ranges from [0,1], where 1 represents a perfect match and 0 represents a complete mismatch.

[0219] (2) Normalized mutual information (NMI);

[0220] Normalized Mutual Information (NMI) is an information theory-based metric used to measure the degree of information sharing between real and predicted labels. It quantifies the similarity between clustering results and real labels by calculating mutual information; NMI is a normalization of mutual information. Mutual information measures the amount of information shared between two random variables (e.g., real and predicted labels).

[0221]

[0222] Where t is the real label, and The experimental labels obtained from the experiment, It is t and The joint probability, p(t) and They are t and The marginal probability of H(T) and They are T and The entropy of NMI represents the amount of information shared by a single variable. The value range of NMI is [0,1], where 1 represents complete agreement and 0 represents no information sharing.

[0223] The comparison algorithm is as follows:

[0224] MTSC-FCM-CD: Constructs complex networks based on the similarity between FCMs, and leverages these networks to build long-distance global information. It applies community detection to capture clusters of arbitrary shapes within the network, enabling clustering of multivariate temporal sequences.

[0225] Corclust: Calculates empirical correlations (such as Pearson sample correlation coefficients) between pairs of spatial points to identify spatial and spatiotemporal neighborhoods with similar characteristics and performs iterative clustering based on ranking and selection to achieve spatiotemporal clustering.

[0226] TICC: The Toeplitz Inverse Covariance Clustering Algorithm (TICC) clusters time series data, defining each cluster as a dependency network or Markov random field and describing the interdependencies across different variables and time.

[0227] Mc2PCA: Assuming a common projection axis as the prototype for each cluster in K-Means clustering, and using the reconstruction error of the MTS data to redistribute the clustering results and monitor the convergence of the transformation process.

[0228] U-kmeans-DTW: The original algorithm of Uk-means, where the input is still the original data, and the distance is measured using the DTW algorithm to measure the effectiveness of feature extraction.

[0229] UTCFTS-A: A variant of UTCFTS A, which directly takes the CSPCA algorithm as input and does not have MIC-assisted clustering. It is used to measure the effectiveness of feature extraction and nearest neighbor rules based on MIC.

[0230] The comparison results of different algorithms are shown in Table 8:

[0231] Table 8 Comparison of RandIndex and NormalizedMutual Information using different algorithms

[0232]

[0233]

[0234] As can be seen, the UTCFTS algorithm proposed in this invention has achieved good results on multiple datasets, and the evaluation metrics also show that the algorithm can identify different types of patterns well, which has obvious advantages over other comparative algorithms.

[0235] The above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the protection scope of the present invention.

Claims

1. An evolutionary pattern recognition method based on temporal tensor segmentation, characterized in that, Includes the following steps: S1. Collect third-order temporal tensor data as input data; The third-order time-series tensor data refers to multivariate time-series data collected by sensors, including information retrieval data, environmental monitoring data, human activity data, financial analysis data, and building energy consumption data; the collected third-order time-series tensor... X ∈ R M×N×T for M Users / locations in T Within a certain time period N The sampling results of each variable; S2. Divide the third-order temporal tensor into multiple tensor segments using a sliding window; Based on the changing trend of the third-order temporal tensor, K segmentation points are found by using a sliding window, and X is divided into K+1 tensor segments along the time dimension; S3. Extract the core tensors of the tensor segments based on Tucker decomposition, and perform clustering based on temporal weighted clustering to obtain coarse-grained segmentation results, specifically including: For tensor segments obtained through a sliding window, the corresponding core tensors are extracted using Tucker decomposition. During the clustering process, each core tensor is iteratively connected to its nearest neighbor using a spatiotemporal graph representation matrix to form a cluster. Each iteration of clustering corresponds to a segmentation result. After multiple iterations, the segmentation result with the largest rate of change in information gain is selected as the final coarse-grained segmentation result. S4. Optimize the coarse-grained segmentation result based on information gain to obtain a fine-grained segmentation result, specifically including: The time interval of two windows around the coarse-grained segmentation point is selected. A greedy search is used to calculate the information gain of each time point in the interval as a segmentation point. The point with the largest information gain is selected as the optimized segmentation point, thus obtaining the fine-grained segmentation result. S5. Extract the correlation features of each segment based on the maximum information coefficient; Among them, correlation includes variable correlation, time correlation, and user / location correlation; S6. Introducing the nearest neighbor rule, an unsupervised three-way clustering algorithm is used to identify the evolutionary patterns corresponding to each segment, specifically including: Set the entropy coefficient α c For a sample to belong to the first c Class probability, membership degree z uc For the first u For the nth sample c The degree of membership in a class; Given the initial number of clusters Initial cluster centers Initial entropy coefficient ,in U The number of samples; Clustering is performed using the nearest neighbor rule: Based on the number of nearest neighbors of each sample, the spatiotemporal points are sorted in descending order to create a list; the point with the most neighbor points in the list is selected as the first cluster center, and all its neighbor points are assigned to this cluster; for subsequent points in the list, if a point does not belong to any cluster and more than 50% of its neighbor points do not belong to any cluster, then the point is regarded as the center of the new cluster. The initial clustering results obtained based on the nearest neighbor rule are iterated multiple times according to the set objective function to update the entropy coefficient and membership degree, and further update the cluster centers and the members of each cluster; among them, by discarding clusters to adjust and ; when Stop iterating and output the evolution pattern corresponding to each segment.

2. An evolutionary pattern recognition system based on temporal tensor segmentation, the system being used to implement the method as described in claim 1, characterized in that, The system includes: The data acquisition module is used to acquire third-order time-series tensor data as input data; The temporal tensor segmentation module is used to divide the third-order temporal tensor into multiple tensor segments using a sliding window; extract the core tensors of the tensor segments based on Tucker decomposition, and perform clustering based on a temporal weighted clustering method to obtain coarse-grained segmentation results; optimize the coarse-grained segmentation results based on information gain to obtain fine-grained segmentation results. The evolution pattern recognition module is used to extract the correlation of each segment based on the maximum information coefficient; it introduces the nearest neighbor rule and uses an unsupervised three-way clustering algorithm to identify the evolution pattern corresponding to each segment.

3. The system according to claim 2, characterized in that, The data acquisition module includes sensor components, a serial port server, a management workstation, an interface server, an information center storage server, and a platform core switch; The sensor components are connected to the management workstation via the serial port server. The management workstation is connected to the interface server. The interface server is connected to the information center storage server via a unidirectional isolation gateway. The information center storage server is connected to the platform core switch via optical fiber. The platform core switch is connected to multiple cloud computing nodes and web servers.

4. The system according to claim 2, characterized in that, The system also includes a monitoring app, which includes a client, a server, and a system management backend. The client is used for user registration and login, online query, modification, and logout; the server is used for registration and login verification, as well as data transmission, addition, modification, and deletion functions; the system management backend is used for database management.

Images