A method, device and medium for clustering variable-length multivariable time series
By extracting local shapelets from multi-source sensor data of industrial equipment and constructing a natural neighbor graph for community detection, the problem of unified feature representation of variable-length multivariable time series and collaborative optimization of local features and global clustering structure is solved, achieving more accurate clustering results.
Patent Information
- Application Number
- CN202611218019.2
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2026-08-12
- Publication Date
- 2026-11-10
AI Technical Summary
Existing technologies struggle to simultaneously address the unified feature representation of variable-length, multivariate time series, the extraction of local dynamic features, and the collaborative optimization of local features and global clustering structures. In particular, in multi-source sensor data from industrial equipment, existing methods are easily affected by signal noise, and the local feature selection process is independent of the global clustering structure, leading to inaccurate clustering results.
By extracting local shapelets from variable-length, multivariable time series generated by multi-source sensors in industrial equipment, and converting the distance between the shapelet and the time series sliding window into a unified-dimensional feature vector, a natural neighbor graph is constructed and community detection is performed. The modularity is then fed back to the shapelet genetic evolution process to optimize local features and form closed-loop associations.
It achieves unified feature representation and local dynamic feature extraction for variable-length multivariable time series, and the clustering results correspond to specific operating segments of industrial equipment and local sensor dynamics, thereby improving the accuracy and consistency of clustering.
Smart Images

Figure CN123069002A_ABST
Abstract
Description
Technical Field
[0001] This invention relates to the field of industrial equipment condition monitoring and multi-source sensor signal processing technology, specifically to a variable-length multivariable time series clustering method, device, and medium. Background Technology
[0002] With the development of industrial equipment operation monitoring and equipment status analysis technologies, industrial robots, CNC machine tools, generator sets, and other industrial equipment are typically equipped with multiple sensors during operation to collect physical signals such as vibration, temperature, current, pressure, rotational speed, acoustic emission, cutting force, torque, or displacement. Because industrial equipment operates through different cycles, operational events, processing stages, start-up and shutdown processes, or load changes, the duration of these different processes is usually inconsistent. Therefore, segmenting the signals simultaneously collected by multiple sensors according to their respective operating processes results in multivariate time series of varying lengths.
[0003] In the actual operation of industrial equipment, there are coupling relationships between signals collected by different sensors. For example, when an industrial robot performs grasping, handling, or placement actions, its joint current, end effector force, vibration, position, or velocity signals change with the execution process; when a CNC machine tool is running, its spindle current, vibration, acoustic emission, rotational speed, and cutting force signals change with the machining process. The resulting multivariate time series not only have inconsistent lengths but also exhibit correlations between different variable dimensions. For such variable-length multivariate time series, analysis using a single sensor channel or a single threshold is insufficient to fully characterize the state differences of industrial equipment during different operating processes.
[0004] For time series of different lengths, existing technologies can convert them to the same length using methods such as zero-padding and linear interpolation, and then perform feature extraction or clustering on the converted time series. However, since the padding or interpolated data are not the original physical signals actually acquired by the sensor, the above processing may change the data distribution of the original time series and may introduce bias or invalid data into the time series, thereby affecting the subsequent characterization of the dynamic changes of the original sensor.
[0005] For similarity analysis between variable-length time series, existing technologies also include distance-based methods, such as Dynamic Time Warping (DTW), its differentiable variants, and Shape-based Distance (SBD). These methods typically measure the similarity between different time series using elastic distance or shape distance. However, in multi-source sensor scenarios, as the length and dimensionality of the time series increase, distance calculations are required for a large number of time series, making the calculation process quite complex. Furthermore, these methods are susceptible to signal noise, and the distance results obtained from the complete time series are difficult to directly correlate with specific sensor channels or local dynamic segments within the time series.
[0006] Existing technologies can also represent variable-length multivariate time series using model-based or feature-based methods. Model-based methods, such as autoencoders and recurrent neural networks, learn the latent representation of the time series; however, the latent features obtained by such models are often difficult to establish a direct correspondence with specific sensor channels and specific operating stages in industrial equipment. Feature-based methods can use symbolic aggregation approximation (SAX) and principal component analysis (PCA) to reduce the dimensionality of the time series, thereby transforming time series of different lengths into a unified feature space. However, when performing global feature extraction on a complete time series, local dynamic information such as local shocks, plateau changes, drift, or periodic anomalies may be lost.
[0007] To characterize local dynamic features in time series, existing technologies have proposed time series processing methods based on shapelets. Shapelets can represent distinctive local subsequences within a time series; in industrial equipment sensor signals, they can correspond to local signal changes such as gripping impacts, sudden changes in spindle vibration, acoustic emission peaks, or temperature drift. Compared to the global features of a complete time series, shapelets can describe the morphological changes of multivariate time series from a local time region.
[0008] However, determining suitable shapelets for clustering from a large number of local subsequences remains challenging when known state labels are lacking. Existing shapelet clustering methods can select shapelets based on inter-cluster distances or employ pseudo-labels or self-supervised learning to process shapelets. When the duration of a multivariate time series is long, each time series can generate a large number of local subsequences, correspondingly increasing the number of candidate shapelets. Therefore, further screening of shapelets to represent the time series is necessary from a large number of candidate local subsequences. Although existing techniques can reduce the number of candidates using random sampling or candidate pruning, an evaluation relationship still needs to be established between the shapelet selection process and the final clustering structure.
[0009] After obtaining the feature representation of a time series, existing technologies typically employ clustering algorithms such as DBSCAN and fuzzy clustering to further divide the time series. They can also convert the time series into nodes in a complex network, establishing connections between nodes based on the distances between time series, and then determining the corresponding time series clusters through the network structure. For multi-source sensor data from industrial equipment, nodes in a complex network can correspond to time series segments formed by different operating cycles, operational events, or processing stages, while different communities within the network can correspond to different operating state categories.
[0010] However, existing network clustering methods typically establish network connections based on the distances between complete time series. The network structure is easily influenced by the distance function used, and the local dynamic features within the time series are not fully considered during network construction. On the other hand, if local features representing the time series are predetermined, and clustering is performed directly based on these fixed local features, the selection process of these local features is independent of the overall network structure formed by subsequent clustering. This makes it difficult to adjust the local features representing the time series according to the global relational structure formed during the clustering process. Summary of the Invention
[0011] This invention provides a method, device, and medium for clustering variable-length multivariable time series, aiming to solve the problem that existing technologies are unable to simultaneously achieve unified feature representation of variable-length multivariable time series, extraction of local dynamic features, and collaborative optimization of local features and global clustering structure.
[0012] To achieve the above objectives, the first aspect of the present invention provides a variable-length multivariate time series clustering method, comprising the following steps: Collect physical signals generated by multiple sensors during the operation of industrial equipment to obtain multi-source sensor signal sequences; The multi-source sensor signal sequence is segmented according to the operation process of the industrial equipment to obtain a set of variable-length multivariable time series. Extract the initial shapelet group containing the source sensor channels from the variable-length multivariable time series set; Based on the initial shapelet group, the variable-length multivariate time series set is converted into a feature vector set; Construct a natural neighbor graph based on the set of feature vectors; Community detection is performed on the natural neighbor graph to obtain clustered communities and the modularity corresponding to the clustered communities; Based on the modularity, the initial shapelet group is subjected to genetic evolution optimization to obtain an optimized shapelet group; Based on the optimized shapelet group, the variable-length multivariate time series set is clustered to obtain the time series clustering results; The operational status category of the corresponding operational segment of the industrial equipment is determined based on the time series clustering results.
[0013] Furthermore, the method for segmenting the multi-source sensor signal sequence according to the operation process of the industrial equipment to obtain a variable-length multivariable time series set includes: The segmentation position of the multi-source sensor signal sequence is determined based on the operating cycle, operation events, processing stages, start-up and shutdown processes, or load changes of the industrial equipment. The multi-source sensor signal sequence is segmented according to the segmentation position to obtain multiple running segments; The signal sequences corresponding to multiple sensor channels in each of the aforementioned running segments are determined as multivariable time series, thereby obtaining the variable-length multivariable time series set.
[0014] Furthermore, the method for extracting an initial shapelet set recording source sensor channels from the variable-length multivariate time series set includes: Multiple multivariate time series are sampled from the variable-length multivariate time series set according to a preset sampling rate; Determine the local farthest deviation point in each variable dimension of each of the multivariate time series; Count the number of local farthest deviation points at each time point across all the variable dimensions; Candidate key points are obtained by filtering each time point based on the number of the local farthest deviation points. The candidate key points are filtered based on the time step distance between adjacent candidate key points to obtain key points; The corresponding multivariate time series is segmented using each of the key points as the start point, midpoint, and end point, respectively, to obtain multivariate subsequences with a preset shapelet length; Cluster the multivariate subsequences to obtain a candidate shapelet set; Multiple shapelet groups are determined based on the differences between candidate shapelets in the candidate shapelet set; The multiple shapelet groups are determined as the initial shapelet group, and the source sensor channel corresponding to each shapelet is recorded.
[0015] Furthermore, the method for converting the variable-length multivariate time series set into a feature vector set based on the initial shapelet group includes: Multiple shapelets are determined from the initial shapelet group; Slide each shapelet on each multivariate time series in the variable-length multivariate time series to obtain multiple sliding windows corresponding to each shapelet; Calculate the Euclidean distance between each shapelet and the corresponding plurality of sliding windows; The minimum Euclidean distance is determined from multiple Euclidean distances to obtain the shapelet distance between the multivariate time series and the corresponding shapelet; The feature vector corresponding to the multivariable time series is obtained by arranging multiple shapelets in the initial shapelet group according to the multiple shapelet distances corresponding to the multiple shapelet arrangements; The feature vector set is obtained by collecting the feature vectors corresponding to each multivariate time series in the set of variable-length multivariate time series.
[0016] Furthermore, the method for constructing a natural neighbor graph based on the set of feature vectors includes: Each feature vector in the feature vector set is identified as a node; Calculate the Euclidean distance between each of the nodes; The nearest neighbor nodes of each node are determined iteratively based on the Euclidean distance between each node. The neighborhood parameters are determined based on the nearest neighbor nodes corresponding to each node. The set of natural neighbors corresponding to each node is determined based on the neighborhood parameters. The natural neighbor graph is obtained by establishing edges between nodes based on the set of natural neighbors corresponding to each node.
[0017] Furthermore, the method for performing community detection on the natural neighbor graph to obtain clustered communities and the modularity corresponding to the clustered communities includes: Each node in the natural neighbor graph is initialized as a community; Based on the modularity increment generated by moving nodes to adjacent communities, the nodes are moved to obtain the first community division result; Each community in the first community division result is aggregated into new nodes to obtain an aggregated network; Repeatedly perform node movement and community aggregation on the aggregated network to obtain the second community partitioning result; The clustering community is determined when the modularity corresponding to the second community division result stops increasing; The modularity of the cluster is calculated based on the edges between nodes in the cluster and the degree of each node.
[0018] Furthermore, the method for performing genetic evolution optimization on the initial shapelet group based on the modularity to obtain an optimized shapelet group includes: Each of the initial shapelet groups is identified as a chromosome individual to obtain an initial population; The modularity of each initial shapelet group is determined as the fitness of the corresponding chromosome individual; Multiple parent chromosome individuals are selected from the initial population based on the fitness; Perform a crossover operation on the shapelets of multiple parent chromosome individuals to obtain offspring chromosome individuals; Perform mutation operations on the shapelets in the offspring chromosome individuals to obtain mutated chromosome individuals; Based on the mutant chromosome individuals, a natural neighbor graph is reconstructed and community detection is performed to obtain the modularity corresponding to the mutant chromosome individuals; Based on the modularity corresponding to the individual mutant chromosome, the individuals mutant chromosome and the parent chromosome individuals are screened to obtain an updated population; The selection, crossover, mutation, and screening processes are repeated on the updated population to obtain the final population; The optimized shapelet group is obtained by identifying the shapelet group with the highest modularity from the final population.
[0019] Furthermore, the method for clustering the variable-length multivariate time series set based on the optimized shapelet group to obtain time series clustering results, and determining the operating state category of the corresponding operating segment of the industrial equipment based on the time series clustering results, includes: Based on the optimized shapelet group, the variable-length multivariate time series set is converted into an optimized feature vector set; Construct an optimized natural neighbor graph based on the optimized feature vector set; The optimized natural neighbor graph is subjected to community detection to obtain the final clustered communities; The time series clustering result corresponding to the variable-length multivariable time series set is determined based on the final clustering community. Based on the time series clustering results, running segments belonging to the same final clustering community are identified as belonging to the same running state category.
[0020] To achieve the above objectives, a second aspect of the present invention provides an electronic device including a memory and a processor, the memory being used to store a program that supports the processor in executing the variable-length multivariate time series clustering method, and the processor being configured to execute the program stored in the memory.
[0021] To achieve the above objectives, a third aspect of the present invention provides a computer-readable storage medium storing a computer program, wherein the computer program, when executed by a processor, performs the steps of the variable-length multivariate time series clustering method.
[0022] The beneficial effects of this invention are: Compared with existing technologies, the present invention provides a variable-length multivariable time series clustering method, device, and medium. This method extracts local shapelets from variable-length multivariable time series formed by multi-source sensors of industrial equipment and uses the distance between the shapelets and the sliding windows of each time series to convert time series of different lengths into feature vectors of a unified dimension, without forcibly unifying the time series length through padding or interpolation. Furthermore, it constructs a natural neighbor graph based on the feature vectors and performs community detection to obtain clustered communities reflecting the overall association structure of the time series. Simultaneously, the modularity corresponding to the community division is used as fitness feedback to the genetic evolution process of the shapelet group, iteratively optimizing the shapelet group used to characterize local dynamic features. This creates a closed-loop association between local feature extraction and the global clustering structure, thus balancing the unified feature representation of variable-length multivariable time series, local dynamic feature extraction, and clustering structure optimization. Moreover, by preserving the sensor channels and time positions corresponding to the shapelets, the clustering results can correspond to specific operational segments of industrial equipment and their local sensor dynamics. 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.
[0024] Figure 1 This is a schematic diagram illustrating a state recognition scenario and data challenge for multi-source sensor variable-length time-series signals in industrial equipment, as disclosed in an embodiment of the present invention. Figure 1 (a) in the diagram is a schematic diagram of an industrial equipment and multi-source sensor signal acquisition scenario; Figure 1 (b) in the diagram is a visualization of multi-source sensor signals; Figure 1 (c) in the figure is a schematic diagram of the data challenge of variable-length multivariate time series.
[0025] Figure 2 This is a schematic diagram of the overall framework of a variable-length multivariate time series clustering method disclosed in an embodiment of the present invention.
[0026] Figure 3 This is a schematic diagram of a shapelet representation process for a variable-length multivariable time series disclosed in an embodiment of the present invention.
[0027] Figure 4This is a schematic diagram of a variable-length multivariate time series clustering process based on natural neighbor graphs and community detection, as disclosed in an embodiment of the present invention.
[0028] Figure 5 This is a schematic diagram of a shapelet group genetic evolution optimization process disclosed in an embodiment of the present invention.
[0029] Figure 6 This is a pairwise comparison graph of the NMI method of the present invention and different multivariate time series clustering methods on 12 publicly available multivariate time series datasets, as disclosed in an embodiment of the present invention. Figure 6 (a) in the figure is a pairwise comparison diagram of the method of the present invention and MUSLA's NMI. Figure 6 (b) in the figure is a pairwise comparison diagram of the method of the present invention and the NMI of MFC-DTW. Figure 6 (c) in the figure is a pairwise comparison diagram of the method of the present invention and the NMI of CSMVC. Figure 6 (d) in the figure is a pairwise comparison diagram of the method of the present invention and the NMI of DTW. Figure 6 (e) in the figure is a pairwise comparison diagram of the method of the present invention and the NMI of DeTSEC. Figure 6 (f) in the figure is a pairwise comparison diagram of the NMI of the method of the present invention and CD. Figure 6 (g) in the figure is a pairwise comparison diagram of the NMI of the method of the present invention and SDTW. Figure 6 (h) in the figure is a pairwise comparison diagram of the NMI of the method of the present invention and TCK. Figure 6 In the diagram (i), there is a pairwise comparison of the NMI of the method of the present invention and Mc2PCA. Figure 6 (j) in the figure is a pairwise comparison diagram of the method of the present invention and the NMI of PDC.
[0030] Figure 7 This is an example diagram of a majority class multivariate time series segment of an LP5 industrial robot execution fault dataset disclosed in an embodiment of the present invention.
[0031] Figure 8 This is a robustness comparison chart of the clustering results of the present invention and MUSLA under different proportions of available unlabeled samples, as disclosed in an embodiment of the present invention. Figure 8 (a) in the figure is a comparison of the proportion of unlabeled samples and NMI in the ArabicDigits dataset. Figure 8 (b) in the figure is a comparison of the proportion of unlabeled samples and NMI in the JapaneseVowels dataset. Figure 8 (c) in the figure is a comparison of the proportion of unlabeled samples in the ECG dataset with NMI. Figure 8 (d) in the figure is a comparison of the proportion of unlabeled samples in the CharacterTrajectories dataset with NMI. Figure 8 (e) in the figure is a comparison of the proportion of unlabeled samples and NMI in the CMUsubject16 dataset. Figure 8 (f) in the figure is a comparison of the proportion of unlabeled samples and NMI in the WalkvsRun dataset. Figure 8 In the figure, (g) is a comparison plot of the relationship between the proportion of unlabeled samples and NMI in the LP1 dataset. Figure 8 (h) in the figure is a comparison of the proportion of unlabeled samples and NMI in the LP2 dataset. Figure 8 In the figure (i), there is a comparison of the proportion of unlabeled samples and NMI in the LP3 dataset. Figure 8 In the figure, (j) is a comparison plot of the relationship between the proportion of unlabeled samples and NMI in the LP4 dataset. Figure 8 In the figure, (k) is a comparison plot of the relationship between the proportion of unlabeled samples and NMI in the LP5 dataset. Figure 8 (l) in the figure is a comparison of the proportion of unlabeled samples in the Libras dataset with NMI.
[0032] Figure 9 This is a schematic diagram illustrating the variation of clustering parameters under different shapeslet numbers and lengths as disclosed in an embodiment of the present invention. Figure 9 (a) in the figure shows the relationship between the number of shapelets, shapelet length, and NMI in the ArabicDigits dataset. Figure 9 (b) in the figure shows the relationship between the number of shapelets, shapelet length, and NMI in the JapaneseVowels dataset. Figure 9 (c) in the figure shows the relationship between the number of shapelets, shapelet length, and NMI in the ECG dataset. Figure 9 In the diagram (d), the relationship between the number of shapelets, shapelet length, and NMI in the CharacterTrajectories dataset is shown. Figure 9 In the graph (e), the relationship between the number of shapelets, shapelet length, and NMI in the CMUsubject16 dataset is shown. Figure 9 In the graph (f), the relationship between the number of shapelets, shapelet length, and NMI in the WalkvsRun dataset is shown. Figure 9 In the graph (g), the relationship between the number of shapelets, shapelet length, and NMI in the LP1 dataset is shown. Figure 9 In the diagram (h), the relationship between the number of shapelets, shapelet length, and NMI in the LP2 dataset is shown. Figure 9 In the diagram, (i) represents the relationship between the number of shapelets, shapelet length, and NMI in the LP3 dataset. Figure 9 In the diagram, (j) represents the relationship between the number of shapelets, shapelet length, and NMI in the LP4 dataset. Figure 9 In the diagram, (k) represents the relationship between the number of shapelets, shapelet length, and NMI in the LP5 dataset. Figure 9 (l) in the figure represents the relationship between the number of shapelets, shapelet length and NMI in the Libras dataset.
[0033] Figure 10 This is a schematic diagram of a variable-length multivariate time series clustering process based on the CharacterTrajectories dataset disclosed in an embodiment of the present invention; Figure 10 (a) in the figure is a schematic diagram of the original variable-length multivariate time series and candidate shapelets corresponding to the four motion modes in the CharacterTrajectories dataset. Figure 10 (b) in the diagram is a schematic diagram of the optimized shapelet set obtained through genetic evolution. Figure 10 (c) in the figure is a schematic diagram of the clustering results of variable-length multivariate time series based on complex network representation. Detailed Implementation
[0034] The technical solution of the present invention will be clearly and completely described below with reference to the accompanying drawings and specific embodiments. It should be understood that the following embodiments are only used to illustrate the present invention and are not intended to limit the scope of protection of the present invention. In the absence of conflict, the following embodiments and the technical features in the embodiments can be combined with each other.
[0035] This embodiment provides a variable-length multivariate time series clustering method. This method can be applied to multi-source sensor signals generated during the operation of industrial equipment such as industrial robots, CNC machine tools, and generator sets. Taking the actual physical signals generated during the operation of industrial equipment as the processing object, the method forms a clustering result of variable-length multivariate time series through local shapelet feature extraction, feature vector transformation, natural neighbor graph construction, community detection, and shapelet genetic evolution. Based on the clustering result, the operating state category of the corresponding operating segment of the industrial equipment is determined.
[0036] like Figure 1 As shown, industrial robots, CNC machine tools, and generator sets can collect physical signals such as vibration, temperature, current, pressure, rotational speed, acoustic emission, cutting force, torque, displacement, position, or velocity through multiple sensors installed at corresponding locations during operation. Since the duration of different operating cycles, operational events, processing stages, start-up and shutdown processes, or load changes in industrial equipment varies, continuous sensor signals can be segmented according to the actual operating process to form multivariate time series of varying lengths.
[0037] As shown in Table 1, the processing in this embodiment takes the multi-source sensor signals of industrial equipment as input, and forms the corresponding time series clustering results through variable length segmentation, candidate shapelet extraction, shapelet group optimization, complex network construction and community detection. Each shapelet can further record its corresponding sensor channel, time position and operation stage, each complex network node corresponds to an actual operation segment, and different network communities correspond to different operation state categories.
[0038] Table 1. Summary of Input-Processing-Output for Multi-Source Sensor Status Identification in Industrial Equipment
[0039] like Figure 2 As shown, the method in this embodiment generally includes three interrelated processing parts: variable-length multivariate time series representation, complex network-based clustering, and shapelet feature learning optimization. Variable-length multivariate time series representation is used to map multivariate time series of different lengths to a shapelet distance feature space of consistent dimension; complex network clustering is used to construct a relationship graph between running segments and form cluster communities based on the feature space; shapelet feature learning optimization performs genetic evolution on shapelet groups according to the modularity corresponding to the community division, and reuses the optimized shapelet groups for time series representation and clustering, thus forming a closed-loop processing process.
[0040] Specifically, a variable-length multivariate time series clustering method includes the following steps: Step S100: Collect physical signals generated by multiple sensors during the operation of industrial equipment to obtain a multi-source sensor signal sequence.
[0041] In practice, based on the structure and operating status of the industrial equipment to be monitored, multiple sensors are set at the corresponding locations of the industrial equipment, and the corresponding physical quantities are continuously collected according to the preset sampling frequency.
[0042] For example, sensors can be installed on industrial robots to collect joint current, end effector force, vibration, position, and speed; sensors can be installed on CNC machine tools to collect spindle current, three-axis vibration, acoustic emission, spindle speed, cutting force, and temperature; and sensors can be installed on generator sets to collect vibration, temperature, current, pressure, and speed.
[0043] Each sensor generates a continuous sampling signal according to a unified time base. If the industrial equipment includes… Each sensor variable involved in the state analysis forms a result at any sampling time point from... Multivariate observations consisting of multiple variables are formed by multiple consecutive sampling time points, which correspond to a multi-source sensor signal sequence.
[0044] in, This indicates the dimension of the variable participating in the current time series clustering process. It can correspond to the number of sensors or multiple effective signal channels obtained from multiple sensors.
[0045] For industrial robots, during processes such as approaching the gripping position, performing gripping, transferring parts, carrying parts, and approaching the placement position, the current, force, vibration, position, and speed signals generated by various sensors change with the robot's actions. For CNC machine tools, in different machining stages, the spindle current, vibration, acoustic emission, speed, cutting force, and temperature change with the interaction between the tool and the workpiece. For generator sets, vibration, temperature, current, pressure, and speed can all change accordingly during normal operation, load changes, and equipment degradation.
[0046] This yields a multi-source sensor signal sequence corresponding to specific industrial equipment and its actual operation.
[0047] Step S200: The multi-source sensor signal sequence is segmented according to the operation process of the industrial equipment to obtain a variable-length multivariable time series set.
[0048] After obtaining a continuous sequence of multi-source sensor signals, the time segmentation position is determined according to the corresponding operating cycle, operation event, processing stage, start-up and shutdown process or load change process of the industrial equipment, and the signals of different sensor channels are synchronously segmented according to the time segmentation position.
[0049] For example, industrial robots can be divided into one grasping process, one handling process, or one placement process; CNC machine tools can be divided into one machining cycle, different toolpath stages, or different cutting stages; generator sets can be divided into one operating cycle, one load change process, or one start-stop process.
[0050] Because different operational processes last for varying durations, the resulting multiple operational segments have different time lengths. Within the same operational segment, data from multiple sensor channels are combined according to their time positions to form a multivariate time series.
[0051] Suppose the obtained set of variable-length multivariate time series is:
[0052] in, Represents a set of time series of varying lengths and multiple variables; Indicates the number of time series in the set; Indicates the first The multivariate time series corresponding to each running segment; different They can have different durations.
[0053] For any 1 A multivariate time series It can be represented as having Dimensions of variables and A data sequence at several time points, where... Indicates the dimension of the variable. This indicates the length of the current multivariate time series. For different execution segments, Different values can be selected.
[0054] Using the above method, it is not necessary to pre-transform each running segment to the same length using zero padding or linear interpolation, but rather to retain the original variable-length structure formed by the actual operation of industrial equipment.
[0055] In the subsequent candidate shapelet extraction process, it can be based on the sampling rate. From the set of variable-length multivariable time series The mid-sampled time series is used for subsequent shapelet candidate generation. Sampling rate. The preferred α is selected from 10%, 20%, 50%, and 100% to control the proportion of time series involved in the initial local feature extraction.
[0056] Step S300: Extract the initial shapelet group containing source sensor channels from the variable-length multivariable time series set.
[0057] This step is used to extract shapelets that can describe local dynamic changes from variable-length multivariable time series and form multiple initial shapelet groups.
[0058] In this context, a shapelet is a local subsequence within a multivariate time series, consisting of data from multiple variable channels within a continuous time interval. A shapelet not only includes the corresponding numerical sequence but can also be associated with its source sensor channel, its temporal position within the original runtime segment, and its corresponding runtime stage.
[0059] Specifically, the process includes the following steps: First, according to the sampling rate The extracted multivariate time series is subjected to Local Farthest Deviation Point (LFDP) detection.
[0060] For length of For each multivariate time series, the LFDP is determined separately for each variable dimension. In one specific implementation, the number of LFDPs initially determined for each variable dimension is:
[0061] in, This indicates the duration of the current multivariate time series.
[0062] Subsequently, for each time point in this multivariate time series, the time point's position in the entire series is statistically analyzed. The number of times each variable dimension is identified as an LFDP.
[0063] Furthermore, based on the dimension of the variables Set a key point filtering threshold. In one implementation, the filtering threshold is set as follows:
[0064] Time points with LFDP statistics below the filtering threshold are filtered out, and the remaining time points are identified as candidate key points.
[0065] Considering that adjacent candidate keypoints may generate local subsequences with a high degree of overlap, the shapelet length can also be used as a reference. Further filtering of candidate keypoints is performed. When the time step distance between two adjacent candidate keypoints is less than the shapelet length... When half of the candidate key points are obtained, either merge the adjacent candidate key points or delete one of the candidate key points to obtain the final key points.
[0066] Subsequently, each key point was used as the start, mid, and end point of a local subsequence, and a length of [length missing] was extracted from the corresponding multivariate time series. Multivariable local subsequences.
[0067] Therefore, a key point can generate multiple candidate local subsequences from different relative positions to cover cases where the key change occurs at the beginning, middle or end of the subsequence.
[0068] like Figure 3 As shown, multiple original multivariate time series of different lengths are first sampled and keypoint filtered, and then multiple candidate local subsequences are extracted based on the keypoints. After obtaining significant local subsequences, K-means clustering is performed on the local subsequences to reduce the duplication of highly similar candidate subsequences.
[0069] Let the number of candidate local subsequences obtained after keypoint processing be... Candidate control parameters are used. Then the local subsequence can be divided into: Clusters, among which, This indicates the number of keypoints obtained through keypoint detection. This represents a positive integer used to control the number of candidate shapelets. Each cluster center obtained from K-means clustering is designated as a candidate shapelet, thus forming a set of candidate shapelets.
[0070] Furthermore, in order to form multiple initial shapelet groups from the candidate shapelet set, the pairwise Euclidean distances between each candidate shapelet are calculated to form a candidate shapelet difference matrix.
[0071] First, select a candidate shapelet from the candidate shapelet set as the first shapelet in the current shapelet group. Then, among the unselected candidate shapelets, calculate the average difference between each candidate shapelet and the currently selected shapelet, and prioritize adding the candidate shapelet with the larger average difference to the current shapelet group. Repeat the above process until the current shapelet group contains... There are shapelets. Among them, This indicates the number of shapelets contained in a shapelet group. By changing the initially selected candidate shapelets and repeating the above diversification process, multiple initial shapelet groups can be formed. These multiple initial shapelet groups together constitute the initial population for subsequent genetic evolution, with each shapelet group corresponding to one chromosome individual.
[0072] During the shapelet formation process, the source sensor channel, time position in the original running segment, and corresponding running stage of the shapelet are recorded simultaneously, so that the features formed by the shapelet can be traced back to the local physical dynamics actually generated by the industrial equipment.
[0073] like Figure 3 As shown, after key point screening, candidate shapelet generation and candidate compression, multiple initial shapelet groups are formed. Each initial shapelet group is further used for unified feature representation of variable-length multivariable time series.
[0074] In one parameter setting method, the number of shapelets in the shapelet group From Select; shapelet length It can be based on the minimum length of all original multivariate time series. To determine, for example, from 0.1 0.2 0.3 0.4 and 0.5 Choose from. When When this is the case, the shapelet length can be set to 2. .
[0075] Candidate control parameters It can be selected from positive integers from 1 to 10, and matched with the sampling rate. Number of shapelets The shapelet length ratio and genetic algorithm parameters are used together as candidate hyperparameters for unsupervised selection.
[0076] The above parameter values can be used as a specific implementation method. In actual use, they can also be set according to the sensor sampling frequency, the time scale of the running segment, and computing resources.
[0077] Step S400: Based on the initial shapelet group, convert the variable-length multivariate time series set into a feature vector set.
[0078] For any initial shapelet group, calculate the distance between each multivariate time series in the variable-length multivariate time series set and each shapelet in the shapelet group, thereby converting multivariate time series of different lengths into feature vectors of the same dimension.
[0079] Let the first The multivariate time series are Its length is The dimension of the variable is ;No. Each shapelet is Its length is Shapelet Along multivariate time series Slide along the time direction so that the shapelet sequentially interacts with... The time windows of equal duration are compared.
[0080] No. The multivariate time series and the first Distance between shapelets for: (1) in, Indicates the first A multivariate time series With the a shapelet The shapelet distance between them; Represents a time series index; Indicates shapelet index; Representing multivariate time series Length; Indicates the shapelet length; Indicates the number of dimensions of the variable; Indicates the starting position of the shapelet's sliding over the time series; Indicates the variable dimension index; This represents the time position index inside the shapelet; Indicates the first In the multivariate time series, the first The dimension of the variable is in the first... The value at each time position; Indicates the first The first shapelet The dimension of the variable is in the first... The value at each time position.
[0081] By traversing all feasible sliding positions The minimum value of the obtained Euclidean distance is taken as the distance between the current multivariate time series and the shapelet.
[0082] When a shapelet group contains When the shapelet is n, the nth shapelet is n. The multivariate time series are respectively compared with the above Perform the above distance calculation on each shapelet to obtain:
[0083] Further to the first A multivariate time series is converted into: (2) in, Indicates the first The feature vectors after transformation of a multivariate time series; to These represent the time series and the first to second elements in the shapelet group, respectively. The distance between shapelets; This represents the number of shapelets in the shapelet group, and is equal to the dimension of the transformed feature vector.
[0084] Therefore, regardless of whether the original lengths of the input multivariate time series are the same, as long as the same shapelet group is used for transformation, the dimension can be obtained. eigenvectors.
[0085] like Figure 3 As shown, the original variable-length multivariate time series, after candidate shapelet extraction and shapelet grouping, is mapped to a feature vector of uniform dimension through shapelet transform. Each vector dimension corresponds to a specific shapelet, thus allowing further determination of its source sensor channel and local time location through the shapelet.
[0086] Step S500: Construct a natural neighbor graph based on the set of feature vectors.
[0087] In this step, each feature vector obtained in step S400 is identified as a node in the complex network, and connections between nodes are established based on the Euclidean distance between them. Let the set of feature vectors include... There are nodes, denoted as follows: Each node corresponds to a running cycle segment, operation event segment, processing stage segment, start / stop process segment, or other running segment formed in step S200.
[0088] First, the Euclidean distance between the feature vectors of any two nodes is calculated, and the nearest neighbor relationship of each node is determined based on the Euclidean distance. To reduce the impact of fixed neighborhood parameters on different data distributions, this embodiment uses the NaturalNeighbor mechanism to determine the neighborhood parameters. Neighborhood parameters It can be determined according to the following formula: (3) in, This represents the final determined neighborhood parameters; Indicates the number of iterations in the neighbor search; Indicates the number of nodes in the network; Indicates the first One node; Indicates the first During the round-robin neighbor search process, the node This is considered as the number of the corresponding nearest neighbor nodes; This represents a function used to determine whether the number of corresponding reverse neighbors is zero.
[0089] function Represented as: (4) in, Indicates input to function The number of neighbors. The function value is 1 when the number of input neighbors is zero; otherwise, the function value is 0.
[0090] As the neighbor search rounds The system continuously updates the nearest neighbor set and the number of reverse nearest neighbors for each node. When there are no nodes with zero reverse nearest neighbors, or when the corresponding statistics do not change in two consecutive search rounds, the current neighbor search round is determined as the neighborhood parameter. .
[0091] Determine neighborhood parameters Next, connections are established between the nodes. For any node... With nodes Corresponding edge It can be represented as: (5) in, Represents a node With nodes The connection relationship between them; This indicates that an edge is established between two nodes; This indicates that no edge is established between the two nodes; Represents a node In neighborhood parameters The corresponding set of natural neighbors; Represents a node The corresponding set of natural neighbors.
[0092] A natural neighbor graph is formed by traversing all nodes. Thus, the relationships between different running segments are represented by similarity relations in a unified shapelet feature space, and one running segment corresponds to one network node.
[0093] like Figure 4 As shown, each variable-length multivariate time series is represented by a shapelet to form a network node. The connection relationship between nodes is used to represent the proximity relationship of different running segments in a unified shapelet feature space.
[0094] Step S600: Perform community detection on the natural neighbor graph to obtain clustered communities and the modularity corresponding to the clustered communities.
[0095] The Louvain algorithm is used to detect communities in the natural neighbor graph formed in step S500. Specifically, Louvain community detection includes initialization, local node movement, and community aggregation. During initialization, each node in the natural neighbor graph is treated as an independent community.
[0096] During the local movement of nodes, for the current node to be processed, the modularity increment generated when moving the node to each adjacent community is calculated, and the node is moved to the adjacent community with the largest modularity increment. After all nodes have completed one round of local movement, multiple nodes belonging to the same community are aggregated, and each community is treated as a new node, forming a new aggregated network.
[0097] Furthermore, the node local movement and community aggregation process are re-executed on the aggregated network until the modularity no longer increases, thus obtaining the final community division.
[0098] Modularity corresponding to network partitioning This can be represented as: (6) in, This indicates the modularity of the current community division; Let f(x) represent the number of edges in the network, satisfying the following condition:
[0099] in, Represents a node With nodes The connection relationship between them, that is, whether there are edges; Represents a node The degree; Represents a node The degree; This represents the community consensus indicator function, when a node... With nodes When they belong to the same community, ,otherwise, .
[0100] When the node When joining a community, the corresponding module degree increment This can be represented as: (7) in, This represents the increment in module degree corresponding to node movement; This represents the number of edges within the current community; Represents a node The number of edges connected to nodes within the community; This represents the total number of edges connecting all nodes in the community; Represents a node The degree; This represents the number of edges in the entire network.
[0101] During a round of local node movement, the node is moved to the corresponding The largest neighboring community; if there is no community that increases the modularity, then keep the current community to which the node belongs unchanged.
[0102] like Figure 4 As shown, after Louvain community detection, the natural neighbor graph is divided into multiple communities. Nodes within the same community correspond to multivariate time series that are mutually adjacent in the shapelet feature space. Step S600 not only obtains the clustering communities corresponding to the current shapelet group, but also obtains the modularity corresponding to that clustering community. The modularity is further used as an evaluation criterion when performing shapelet group optimization in step S700.
[0103] Step S700: Perform genetic evolution optimization on the initial shapelet group based on the modularity to obtain an optimized shapelet group.
[0104] The multiple initial shapelet groups formed in step S300 serve as chromosome individuals in the genetic algorithm, and the multiple chromosome individuals together constitute the initial population.
[0105] For each initial shapelet group, perform the shapelet feature transformation in step S400, the natural neighbor graph construction in step S500, and the Louvain community detection in step S600 to obtain the modularity corresponding to that shapelet group. The modularity As the fitness of the corresponding shapelet group.
[0106] like Figure 5 As shown, genetic evolution processing includes initialization, evaluation, selection, crossover, mutation, and filtering processes.
[0107] During the initialization phase, multiple initial shapelet groups are formed into an initial population, with each shapelet group corresponding to a chromosome individual.
[0108] During the evaluation phase, the modularity of the natural neighbor graph constructed by each shapelet group is used as the fitness of that shapelet group.
[0109] During the selection phase, several individuals can be randomly selected from multiple shapelet groups, and the shapelet group with the higher modularity can be retained as the parent individual.
[0110] During the crossover phase, the two parent shapelet groups are crossed according to the preset crossover probability.
[0111] The crossing can include shapelet-level crossing and shapelet fragment-level crossing. In shapelet-level crossing, complete shapelets from the first parent shapelet group are swapped with complete shapelets from the second parent shapelet group. In shapelet fragment-level crossing, fragments are swapped between the local time intervals corresponding to the two parent shapelets, while maintaining the multivariable structure of the resulting shapelet.
[0112] During the mutation phase, the crossover shapelet undergoes random mutation. This mutation can change the corresponding multivariate values at one or more time points of the selected shapelet, while maintaining the shapelet's variable dimension and overall time length. Although the genetically evolved shapelet is no longer limited to local subsequences directly extracted from the original time series, it should maintain morphological similarity to the original local sensor data. This expands the shapelet search range while preserving the interpretable correspondence between the optimized shapelet and the original sensor's local dynamics.
[0113] After selection, crossover, and mutation, new offspring shapelet groups are formed. For each offspring shapelet group, step S400 to S600 are executed sequentially again to perform shapelet transform, natural neighbor graph construction, Louvain community detection, and modularity calculation to obtain the fitness of the offspring shapelet group.
[0114] In the filtering phase, the offspring population is compared with the parent population, and shapelet groups with higher fitness are retained for the next generation based on their modularity. Selection, crossover, mutation, evaluation, and filtering are repeated until the set number of genetic algorithm iterations is reached. Finally, the shapelet group with the highest modularity is selected from the last generation of shapelet groups and designated as the optimized shapelet group. Therefore, the shapelet group is not determined solely by the local candidate selection process in step S300, but rather by using the network community structure formed by the overall time series clustering to perform reverse evaluation of local shapelets and iterative updates through the genetic algorithm.
[0115] In one specific implementation, this method does not use the actual labels of the runtime state during the hyperparameter selection process. Let the candidate hyperparameter space be... This includes the sampling rate Candidate control parameters Number of shapelets Shapelet length ratio Genetic algorithm iteration count, offspring ratio, shapelet-level crossover probability, shapelet fragment-level crossover probability, and mutation probability.
[0116] For any candidate parameter combination The clustering process is repeated using multiple different random seeds, and the hyperparameter combination is determined based on modularity, cluster partitioning stability, and non-degeneracy. The corresponding unsupervised evaluation value can be determined according to the following formula: (8) in, Hyperparameter combination The corresponding unsupervised comprehensive evaluation value; Indicates the current candidate hyperparameter combinations; Represent the candidate hyperparameter space; This represents the average graph modularity obtained under this combination of hyperparameters; This indicates the stability among predicted cluster partitions formed by running the program multiple times with the same parameter combination. This indicates a non-degradability evaluation term, used to constrain single cluster communities or highly imbalanced cluster partitions; Represents the weighting coefficients of the stability term; Represents the weighting coefficient of the non-degenerate term; and Maintain consistency across different datasets.
[0117] By comparing different combinations of candidate parameters Then, determine the final combination of parameters to be used. The entire parameter selection process does not require the use of actual runtime status labels.
[0118] Step S800: Cluster the variable-length multivariable time series set based on the optimized shapelet group to obtain the time series clustering results.
[0119] After obtaining the optimized shapelet group, the optimized shapelet group is applied again to the variable-length multivariate time series set obtained in step S200.
[0120] First, the shapelet distances between each variable-length multivariate time series and each shapelet in the optimized shapelet group are calculated according to formula (1). For each multivariate time series, its shapelet distances with all shapelets in the optimized shapelet group are arranged according to formula (2) to obtain optimized feature vectors. The optimized feature vectors corresponding to all multivariate time series are combined into an optimized feature vector set. Then, each optimized feature vector in the optimized feature vector set is used as a node, and the neighborhood parameters are determined according to formulas (3) and (4), and an optimized natural neighbor graph is constructed according to formula (5). The Louvain community detection process in step S600 is further used to divide the optimized natural neighbor graph into communities to form the final clustered communities.
[0121] Therefore, this embodiment forms the following closed-loop processing procedure: original variable-length multivariate time series; local key point detection; candidate shapelet extraction; initial shapelet group construction; shapelet feature transformation; natural neighbor graph construction; Louvain community detection; modularity evaluation; shapelet group genetic evolution; optimized shapelet feature transformation; optimized natural neighbor graph construction; final community detection; time series clustering results.
[0122] The overall processing flow described above can also be executed according to the process shown in Algorithm 1: Input a variable-length multivariate time series set Sampling rate Shapelet length Number of shapelets Candidate control parameters In addition to the genetic algorithm parameters, an initial shapelet group is first generated, and then the modularity of each shapelet group is evaluated. During the iteration of the genetic algorithm, new shapelet groups are generated through selection, shapelet-level crossover, fragment-level crossover and mutation. Shapelet transformation, graph construction, community detection and modularity evaluation are repeatedly performed to finally obtain the optimized shapelet group with the maximum modularity and the final cluster community.
[0123]
[0124] Step S900: Determine the operating status category of the corresponding operating segment of the industrial equipment based on the time series clustering results.
[0125] Since each network node in step S500 corresponds to a specific industrial equipment operation segment formed in step S200, the final clustering community obtained in step S800 can be mapped back to the specific industrial equipment operation segment. Multiple operation segments belonging to the same final clustering community are determined to be in the same operation state category.
[0126] For example, for industrial robots, the different communities formed by clustering can be further mapped to normal execution, grasping failure, handling abnormality, or placement abnormality. For CNC machine tools, different clusters can correspond to normal machining, tool wear, spindle abnormality, or feed abnormality. For generator sets, different clusters can correspond to normal operation, load fluctuation, bearing degradation, or unknown abnormality. For independent clusters that cannot be mapped to existing operating status categories, they can be retained as unknown operating status categories and subsequently marked based on equipment maintenance information or manual analysis results.
[0127] As shown in Table 2, different sensor inputs and operating states to be identified can be selected according to the type of equipment in different industrial equipment application scenarios. For example, industrial robots can use joint current, end effector force, vibration, position, and speed; CNC machine tools can use spindle current, three-axis vibration, acoustic emission, rotational speed, cutting force, and temperature; and generator sets can use vibration, temperature, current, pressure, and rotational speed.
[0128] Table 2 Mapping Table of Industrial Equipment Implementation Examples
[0129] Meanwhile, since shapelets retain the source sensor channels, time locations, and operating stages, when a certain cluster community has a clear correspondence with certain shapelets, local sensor dynamic information corresponding to that operating state category can be further obtained.
[0130] For example, if a cluster community mainly corresponds to a local vibration mutation shapelet at a specific time location of a vibration channel, then the sensor channel and its corresponding local time interval can be associated with that operating state category. Thus, the clustering results can maintain a correspondence with actual industrial equipment, actual operating segments, and specific sensor dynamics.
[0131] To further illustrate the above methods, the following provides corresponding implementation methods in conjunction with different industrial equipment.
[0132] Example 1: Clustering of Industrial Robot Operation Segments In this embodiment, the industrial equipment to be processed is an industrial robot. During the operation of the industrial robot, multi-source sensor signals such as joint current, end effector force, vibration, position, and velocity are collected, and the continuous sensor signals are divided into multiple operation segments according to the robot's operation process. The operation segments can correspond to operation stages such as approaching the grasping position, transferring parts, moving the carried parts, and approaching the placement position.
[0133] Step S300: Extract local key points and candidate shapelets from each running segment, and then construct multiple initial shapelet groups; Step S400: Convert each running segment into shapelet distance feature vectors; Step S500: Construct a natural neighbor graph; Step S600: Use the Louvain algorithm to form cluster communities; Step S700: Perform genetic evolution on the shapelet groups based on modularity; and Step S800: Form the final clustering result.
[0134] As shown in Table 3, LP1 to LP5 can be used as auxiliary data for clustering the operating states of industrial robots. Specifically: LP1 corresponds to the differences in execution states during the robot's approach to the grasping position; LP2 corresponds to the differences in execution states during the robot's part transfer process; LP3 corresponds to the differences in part position or operational state after a failed transfer; LP4 corresponds to the differences in execution states during the robot's approach to the placement position; and LP5 corresponds to the differences in execution states during the robot's movement carrying parts. Each of LP1 to LP5 includes six variable dimensions, with sample sizes of 50, 30, 30, 75, and 100, respectively.
[0135] Table 3 Usage Boundaries of LP1–LP5
[0136] It should be noted that LP1 to LP5 are used for auxiliary verification of the operating status or execution failure of industrial robots, and are not used as experimental data for CNC machine tools or generator sets.
[0137] Example 2: Clustering of CNC Machine Tool Operation Segments In this embodiment, the industrial equipment is a CNC machine tool. During the operation of the CNC machine tool, multi-source sensor signals such as spindle current, three-axis vibration, acoustic emission, rotational speed, cutting force, and temperature are collected. The continuous sensor signals are divided into multiple variable-length multivariable time series according to the actual machining cycle, toolpath stage, or cutting stage. Steps S300 to S800 are executed sequentially for the time series to form the final cluster community.
[0138] Based on the actual operation records of the CNC machine tool and the corresponding relationships of the operation segments, different communities can be further correlated with normal machining, tool wear, spindle abnormalities, or feed abnormalities. By using the source sensor channels and time positions recorded by the shapelet, the corresponding communities can also be correlated with local dynamic segments such as spindle vibration, acoustic emission, spindle current, or cutting force.
[0139] Example 3: Clustering of Generator Set Operation Segments In this embodiment, the industrial equipment is a generator set. Vibration, temperature, current, pressure, and speed signals are collected during the generator set's operation. Continuous multi-source sensor signals are segmented according to normal operating cycles, load change processes, start-up and shutdown processes, or other operating phases, forming multiple variable-length multivariable time series. Following steps S300–S800, shapelet extraction, unified feature representation, natural neighbor graph construction, Louvain community detection, and shapelet genetic evolution are performed to obtain the final clustered communities.
[0140] Different clustering communities can correspond to normal operation, load fluctuation, bearing degradation, or unknown anomalies based on actual operation records. For operation segments without existing status labels, this method can still form clustering communities by utilizing the local morphological features of multivariable sensor signals and the overall relationship between operation segments.
[0141] To illustrate how the above method handles variable-length multivariate time series, multiple publicly available multivariate time series datasets can be used for auxiliary verification.
[0142] As shown in Table 4, auxiliary verification data may include ArabicDigits, JapaneseVowels, ECG, CharacterTrajectories, CMUsubject16, WalkvsRun, LP1-LP5, and Libras.
[0143] Table 4. Statistical information for 12 publicly available multivariate time series datasets.
[0144] After all hyperparameters are determined, Normalized Mutual Information (NMI) can be used to perform external auxiliary evaluation of the clustering results and known categories.
[0145] In one auxiliary verification implementation, the experiments can be conducted on a workstation equipped with an Intel i7-1260U CPU, 16GB of RAM, and a GeForce RTX 3090 24GB GPU. Except for DeTSEC, which uses GPU acceleration, the other comparative experiments run on a CPU platform.
[0146] NMI can be represented as: (9) in, Represents normalized mutual information; Indicates the number of clusters; This represents the number of samples in the entire multivariate time series. Indicates the first partition result of the first partition. The number of multivariate time series included in each category; Indicating the second partition result, the first... The number of multivariate time series included in each category; This indicates that the corresponding number in the two partitioning results is... The category and the The number of time series contained in the intersection of the categories.
[0147] Adjusted Rand Index, Purity, cluster accuracy, or Rand Index can also be used as external auxiliary evaluation indicators.
[0148] It should be noted that the above-mentioned real category information is only used for the final auxiliary evaluation after the hyperparameters are fixed, and is not used for shapelet extraction in step S300, community division in step S600, or shapelet genetic evolution in step S700.
[0149] The methods can be compared using distance-based, feature-based, and model- or graph-structure-based approaches respectively.
[0150] Distance-based methods can include DTW, Soft-DTW, and MFC-DTW; feature-based methods can include PDC, Mc2PCA, TCK, MUSLA, and CSMVC; and model- or graph-based methods can include DeTSEC and Community Detection.
[0151] As shown in Table 5, after 10 independent runs on 12 auxiliary datasets, the average NMI for each method and the number of datasets that yielded the best results can be calculated.
[0152] Table 5 Performance Comparison by Methodological Paradigm
[0153] This implementation achieves an average NMI of 0.6431 across 12 datasets, and obtains the highest NMI in the corresponding comparisons on WalkvsRun, LP1, LP2, LP3, LP4, and Libras.
[0154] like Figure 6 As shown, this method can be further compared in pairs with MUSLA, MFC-DTW, CSMVC, DTW, DeTSEC, CD, SDTW, TCK, Mc2PCA, and PDC. For some very short time series or samples with insignificant local morphological changes, the processing results based on local shapelets may be limited by the number of effective local subsequences.
[0155] For example, such as Figure 7 As shown, some segments in LP5 exhibit subtle local changes. In such cases, additional sensor channels can be added for processing, or local shapelet extraction can be performed by combining information from different stages of the process.
[0156] set up: Number of segments to run; For the dimension of the variable; The average length of a multivariate time series; The number of shapelets in a shapelet group; The length of the shapelet; The population size for the genetic algorithm; This represents the number of iterations in the genetic algorithm. Let be the number of edges in the natural neighbor graph.
[0157] LFDP-based candidate local subsequence extraction requires scanning according to the sampling rate. The computational complexity of the extracted multivariate time series can be expressed as follows:
[0158] The computational cost of shapelet transform for a single shapelet group can be expressed as:
[0159] In the process of constructing the natural neighbor graph, the computational cost of calculating the pairwise distance between feature vectors can be expressed as:
[0160] The processing capacity of the Louvain community detection is approximately equal to the number of edges in the graph. Related. Therefore, the overall computational cost of evaluating the genetic evolution process can be expressed as:
[0161] In practice, the number of candidates for subsequent processing can be reduced by sampling, key point screening, K-means candidate compression, and controlling the number of iterations of the genetic algorithm.
[0162] As shown in Table 6, the running time of different clustering methods on each auxiliary dataset can be recorded to illustrate the computational overhead corresponding to different processing procedures.
[0163] Table 6. Running time (seconds) of different algorithms on the benchmark dataset
[0164] The runtime results show that this method is not the shortest in absolute runtime among all the compared methods; its average runtime ranks 4th in the comparisons shown in Table 6. Its computational overhead mainly comes from candidate shapelet generation, shapelet transform, natural neighbor graph construction, and genetic evolution evaluation. Subsequent computational costs can be reduced by sampling, LFDP keypoint selection, K-means candidate compression, and controlling the number of iterations in the genetic algorithm. Therefore, this method represents a trade-off between some additional computational overhead and shapelet-level interpretability and competitive clustering performance.
[0165] To verify how well the processing performs as the number of unlabeled running fragments available for feature learning decreases, the proportion of samples used for shapelet learning can be gradually reduced.
[0166] For example, the proportion of data available for feature learning can be gradually reduced from 100% to 90%, 80%, 70%, 60%, and 50%.
[0167] like Figure 8 As shown, the NMI of this method can be compared with that of another shapelet clustering method under different data proportions. In ArabicDigits, when the proportion of data used for feature learning decreases from 100% to 50%, the NMI of this method changes from 0.7245 to 0.6651. In WalkvsRun, when only 50% of the data is used, the corresponding NMI of this method is 0.5132. Under the small sample condition of LP1, the corresponding NMI is 0.1736. The above results can be used as auxiliary validation under the condition of varying unlabeled sample size.
[0168] Furthermore, MUSLA was compared as another method for learning interpretable shapelet features: In ArabicDigits, when the amount of data available for feature learning decreased from 100% to 50%, MUSLA's NMI decreased from 0.2481 to 0.0907, a decrease of approximately 63%, while our method decreased from 0.7245 to 0.6651, a decrease of approximately 8%. In WalkvsRun, using only 50% of the data, our method's NMI was 0.5132, while MUSLA's was 0.2311. Under the small sample condition of LP1, our method's NMI was 0.1736, while MUSLA's was 0.0423. These results indicate that using modularity feedback to genetically evolve shapelets can obtain relatively stable feature representations when the number of available unlabeled samples decreases.
[0169] To examine the keypoint extraction, natural neighbor graph construction, and genetic evolution processes separately, three ablation methods can be constructed: keypoint removal mechanism; natural neighbor removal mechanism; and genetic algorithm optimization removal. The corresponding results are shown in Table 7.
[0170] Table 7 Ablation experiments of ESCN
[0171] For example, on LP1, the NMI of the full method is 0.7181, while the NMI after removing the keypoint mechanism is 0.6190; on Libras, the NMI of the full method is 0.6563, while the NMI after removing the natural neighbor mechanism is 0.6316. These ablation methods allow for the examination of the correspondence between LFDP keypoint extraction, Natural Neighbor graph construction, shapelet genetic evolution, and the final time-series clustering results. The shapelet length can be further adjusted. and number of shapelets This allows us to obtain clustering results under different parameter combinations.
[0172] like Figure 9 As shown, the number of shapelets These can be set to: 1, 2, 5, 10, 20, and 50 respectively. The shapelet length and the minimum time series length are also relevant. The ratios between these ratios can be set to: 0.1, 0.2, 0.3, 0.4, and 0.5, respectively. For cases where the minimum time series length is short, the following can be used: The shapelet length is used as the basis for determining the number and length of shapelets suitable for the current industrial equipment and the current sensor sampling data, since the local dynamic time scales of different data are different.
[0173] The parameter variation results also indicate the number of shapelets The impact on clustering results is usually greater than that of shapelet length. For example, in ArabicDigits, As the number of shapelets increases from 1 to 50, the NMI improves from 0.1628 to 0.7432; in Japanese Vowels, the NMI improves from 0.2231 to 0.8197. CMUsubject16 and WalkvsRun are relatively stable with these parameter changes, while ECG, LP2, LP3, and LP5 are more sensitive to changes in shapelet length. When the number of shapelets is too large and redundant local patterns are introduced, the clustering performance may tend to stabilize, for example, LP3... The case when =50.
[0174] To further illustrate the complete processing from the original variable-length multivariate time series to the final community clustering, we can use CharacterTrajectories as an example. This dataset contains 2558 three-dimensional time series samples across 20 categories. Each time series was obtained by sampling at 200Hz and includes three variables: x-axis motion velocity, y-axis motion velocity, and pen tip pressure. Four motion patterns will be selected for illustration.
[0175] like Figure 10 As shown in (a), the original multivariate time series corresponding to the four motion modes have different lengths and contain different local dynamic changes. Candidate shapelets are extracted from the time series according to step S300, and an optimized shapelet set is obtained through genetic evolution in step S700.
[0176] like Figure 10 As shown in (b) in the example, the optimized shapelet group consists of 20 shapelets, namely: Different shapelets correspond to different local motion modes, such as local high-frequency changes, rapid directional changes, or local velocity plateaus. For each original time series, the distance between it and the 20 shapelets is calculated according to formula (1), thereby forming a 20-dimensional feature vector according to formula (2).
[0177] Furthermore, the feature vectors are used as network nodes, and a complex network is constructed using the natural neighbor approach.
[0178] In this example, the four selected motion patterns include 88, 64, 48, and 75 instances respectively, totaling 275 nodes. The adaptively determined neighborhood parameters are: The resulting complex network consists of 1896 edges. The Louvain algorithm is then used for community detection.
[0179] like Figure 10 As shown in (c), nodes with similar local motion characteristics form densely connected communities, while there are relatively few connections between different communities, thus forming the corresponding time series clustering results.
[0180] thus, Figure 10 The complete diagram illustrates the processing relationships between: original variable-length multivariate time series; candidate local shapelets; optimized shapelet groups; shapelet distance features; complex networks; community detection; and time series clustering results.
[0181] In summary, this implementation method takes the physical signals actually collected by multiple sensors during the operation of industrial equipment as input, and uses multivariate time series of different lengths formed by segmenting the actual operation process as the processing object; it obtains local dynamic representations through local key point extraction and shapelet group construction, and then converts multivariate time series of different lengths into shapelet distance feature vectors of uniform dimension; it constructs a natural neighbor graph using the feature vectors and performs Louvain community detection, while using the modularity corresponding to the community division as the fitness of the genetic algorithm to iteratively optimize the shapelet group; finally, it uses the optimized shapelet group to regenerate the time series clustering results, and maps the clustering results to the actual operation segments and operation status categories of the industrial equipment.
[0182] According to another aspect of the embodiments of this application, an electronic device is also provided, including a processor and a memory, wherein the processor is configured to implement the steps of the method when executing a computer program stored in the memory.
[0183] In the above embodiments of the present invention, the descriptions of each embodiment have different focuses. For parts not described in detail in a certain embodiment, please refer to the relevant descriptions of other embodiments.
[0184] In the several embodiments provided in this application, it should be understood that the disclosed technical content can be implemented in other ways. The device embodiments described above are merely illustrative; for example, the division of units can be a logical functional division, and in actual implementation, there may be other division methods. For instance, multiple units or components may be combined or integrated into another system, or some features may be ignored or not executed. Furthermore, the displayed or discussed mutual coupling, direct coupling, or communication connection may be through some interfaces; the indirect coupling or communication connection between units or modules may be electrical or other forms.
[0185] Furthermore, the functional units in the various embodiments of the present invention can be integrated into one processing unit, or each unit can exist physically separately, or two or more units can be integrated into one unit. The integrated unit can be implemented in hardware or as a software functional unit.
[0186] If the integrated unit is implemented as a software functional unit and sold or used as an independent product, it can be stored in a computer-readable storage medium. Based on this understanding, the technical solution of the present invention, in essence, or the part that contributes to the prior art, or all or part of the technical solution, can be embodied in the form of a software product. This computer software product is stored in a storage medium and includes several instructions to cause a computer device (which may be a personal computer, server, or network device, etc.) to execute all or part of the steps of the methods described in the various embodiments of the present invention. The aforementioned storage medium includes various media capable of storing program code, such as USB flash drives, read-only memory (ROM), random access memory (RAM), portable hard drives, magnetic disks, or optical disks.
[0187] The above description is only a preferred embodiment of the present invention. It should be noted that for those skilled in the art, several improvements and modifications can be made without departing from the principle of the present invention, and these improvements and modifications should also be considered within the scope of protection of the present invention.
Claims
1. A variable-length multivariate time series clustering method, characterized in that, Includes the following steps: Collect physical signals generated by multiple sensors during the operation of industrial equipment to obtain multi-source sensor signal sequences; The multi-source sensor signal sequence is segmented according to the operation process of the industrial equipment to obtain a set of variable-length multivariable time series. Extract the initial shapelet group containing the source sensor channels from the variable-length multivariable time series set; Based on the initial shapelet group, the variable-length multivariate time series set is converted into a feature vector set; Construct a natural neighbor graph based on the set of feature vectors; Community detection is performed on the natural neighbor graph to obtain clustered communities and the modularity corresponding to the clustered communities; Based on the modularity, the initial shapelet group is subjected to genetic evolution optimization to obtain an optimized shapelet group; Based on the optimized shapelet group, the variable-length multivariate time series set is clustered to obtain the time series clustering results; The operational status category of the corresponding operational segment of the industrial equipment is determined based on the time series clustering results.
2. The variable-length multivariate time series clustering method as described in claim 1, characterized in that, The method for segmenting the multi-source sensor signal sequence according to the operation process of the industrial equipment to obtain a variable-length multivariable time series set includes: The segmentation position of the multi-source sensor signal sequence is determined based on the operating cycle, operation events, processing stages, start-up and shutdown processes, or load changes of the industrial equipment. The multi-source sensor signal sequence is segmented according to the segmentation position to obtain multiple running segments; The signal sequences corresponding to multiple sensor channels in each of the aforementioned running segments are determined as multivariable time series, thereby obtaining the variable-length multivariable time series set.
3. The variable-length multivariate time series clustering method as described in claim 1, characterized in that, The method for extracting an initial shapelet set recording source sensor channels from the variable-length multivariate time series set includes: Multiple multivariate time series are sampled from the variable-length multivariate time series set according to a preset sampling rate; Determine the local farthest deviation point in each variable dimension of each of the multivariate time series; Count the number of local farthest deviation points at each time point across all the variable dimensions; Candidate key points are obtained by filtering each time point based on the number of the farthest local deviation points. The candidate key points are filtered based on the time step distance between adjacent candidate key points to obtain key points; The corresponding multivariate time series is segmented using each of the key points as the start point, midpoint, and end point, respectively, to obtain multivariate subsequences with a preset shapelet length; Cluster the multivariate subsequences to obtain a candidate shapelet set; Multiple shapelet groups are determined based on the differences between candidate shapelets in the candidate shapelet set; The multiple shapelet groups are determined as the initial shapelet group, and the source sensor channel corresponding to each shapelet is recorded.
4. The variable-length multivariate time series clustering method as described in claim 3, characterized in that, The method for converting the variable-length multivariate time series set into a feature vector set based on the initial shapelet group includes: Multiple shapelets are determined from the initial shapelet group; Slide each shapelet on each multivariate time series in the variable-length multivariate time series to obtain multiple sliding windows corresponding to each shapelet; Calculate the Euclidean distance between each shapelet and the corresponding plurality of sliding windows; The minimum Euclidean distance is determined from multiple Euclidean distances to obtain the shapelet distance between the multivariate time series and the corresponding shapelet; The feature vector corresponding to the multivariable time series is obtained by arranging multiple shapelets in the initial shapelet group according to the multiple shapelet distances corresponding to the multiple shapelet arrangements; The feature vector set is obtained by collecting the feature vectors corresponding to each multivariate time series in the set of variable-length multivariate time series.
5. The variable-length multivariate time series clustering method as described in claim 3, characterized in that, Methods for constructing a natural neighbor graph based on the aforementioned feature vector set include: Each feature vector in the feature vector set is identified as a node; Calculate the Euclidean distance between each of the nodes; The nearest neighbor nodes of each node are determined iteratively based on the Euclidean distance between each node. The neighborhood parameters are determined based on the nearest neighbor nodes corresponding to each node. The set of natural neighbors corresponding to each node is determined based on the neighborhood parameters. The natural neighbor graph is obtained by establishing edges between nodes based on the set of natural neighbors corresponding to each node.
6. The variable-length multivariate time series clustering method as described in claim 3, characterized in that, The method for performing community detection on the natural neighbor graph to obtain clustered communities and the modularity corresponding to the clustered communities includes: Each node in the natural neighbor graph is initialized as a community; Based on the modularity increment generated by moving nodes to adjacent communities, the nodes are moved to obtain the first community division result; Each community in the first community division result is aggregated into new nodes to obtain an aggregated network; Repeatedly perform node movement and community aggregation on the aggregated network to obtain the second community partitioning result; The clustering community is determined when the modularity corresponding to the second community division result stops increasing; The modularity of the cluster is calculated based on the edges between nodes in the cluster and the degree of each node.
7. The variable-length multivariate time series clustering method according to any one of claims 1-6, characterized in that, The method for performing genetic evolution optimization on the initial shapelet group based on the modularity to obtain an optimized shapelet group includes: Each of the initial shapelet groups is identified as a chromosome individual to obtain an initial population; The modularity of each initial shapelet group is determined as the fitness of the corresponding chromosome individual; Multiple parent chromosome individuals are selected from the initial population based on the fitness; Perform a crossover operation on the shapelets of multiple parent chromosome individuals to obtain offspring chromosome individuals; Perform mutation operations on the shapelets in the offspring chromosome individuals to obtain mutated chromosome individuals; Based on the individual with the mutated chromosome, a natural neighbor graph is reconstructed and community detection is performed to obtain the modularity corresponding to the individual with the mutated chromosome. Based on the modularity corresponding to the individual mutant chromosome, the individuals mutant chromosome and the parent chromosome individuals are screened to obtain an updated population; The selection, crossover, mutation, and screening processes are repeated on the updated population to obtain the final population; The optimized shapelet group is obtained by determining the shapelet group with the highest modularity from the final population.
8. The variable-length multivariate time series clustering method as described in claim 1, characterized in that, The method for clustering the variable-length multivariate time series set based on the optimized shapelet group to obtain time series clustering results, and determining the operating state category of the corresponding operating segment of the industrial equipment based on the time series clustering results, includes: Based on the optimized shapelet group, the variable-length multivariate time series set is converted into an optimized feature vector set; Construct an optimized natural neighbor graph based on the optimized feature vector set; The optimized natural neighbor graph is subjected to community detection to obtain the final clustered communities; The time series clustering result corresponding to the variable-length multivariable time series set is determined based on the final clustering community. Based on the time series clustering results, running segments belonging to the same final clustering community are identified as belonging to the same running state category.
9. An electronic device, comprising a memory and a processor, characterized in that, The memory is used to store programs that support the processor in executing any of the variable-length multivariate time series clustering methods of claims 1-8, the processor being configured to execute the programs stored in the memory.
10. A computer-readable storage medium storing a computer program thereon, characterized in that, When the computer program is run by the processor, it performs the steps of any of the variable-length multivariate time series clustering methods described in claims 1-8.
