Industrial process one-dimensional time sequence signal self-supervision change point detection method and system
By using twin neural networks and one-dimensional convolutional layer in industrial processes, combined with sliding windows and data enhancement technology, the problem of changing point detection in complex industrial scenarios is solved, and the label-free data is achieved High accuracy and generalization change point detection are suitable for a variety of industrial processes.
Patent Information
- Application Number
- CN202410574521.1
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2024-05-10
- Publication Date
- 2024-11-15
- Estimated Expiration
- 2044-05-10
AI Technical Summary
The prior art is difficult to effectively detect variable points in timing signals in complex industrial scenarios, especially due to changes in data distribution states and continuous variable points. The supervised learning method requires a large number of labels and is costly, and the existing methods cannot detect simultaneously. Mutational and continuous change points.
The self-supervised variable point detection method of downstream classifier composed of a time-series characterization encoding module based on twin neural networks and one-dimensional convolutional layer is used to train the model without labels, and generate anchor points and data augmentation technology using sliding windows and data augmentation technology. Negative samples, comparative losses and discriminative losses are calculated for model training, and online detection is realized.
In the absence of label data and unknown properties, effective point change detection is realized, which improves the accuracy and generalization ability of detection, and can simultaneously monitor the mutation and continuous point change of time series, which is suitable for different industrial processes.
Smart Images

Figure CN118965206A_ABST
Abstract
Description
Technical Field
[0001] The present invention relates to the field of information technology, and in particular to a method and system for self-supervising change point detection of one-dimensional time series signals in an industrial process. Background Art
[0002] A change point refers to a situation in a sequence or process where the observed value follows different properties before and after a certain position or time. This position or time point is called a change point. In industrial processes, the occurrence of abnormal conditions is generally manifested as a switch in production status, and the properties of related signals will change. Identifying the time position where the production status switches in the time series signal related to the industrial process is called change point detection. Detecting change points helps understand the behavioral changes of industrial systems and provide timely alerts for potential failures.
[0003] According to the existing literature, the current time series change point detection methods can be divided into the following two categories: The first category is the detection method based on statistical analysis, which designs statistical indicators to measure the difference in properties between different intervals of the time series. When two adjacent sequence intervals have different statistical properties, the time point between the two intervals is judged to be a change point. The second category is a supervised detection method based on data-driven, which assigns a pattern label to each data point or fragment in the sequence, and then learns the mapping relationship from sequence features to pattern labels through a classification model. When the sequences before and after a certain time point have different labels, the time point is judged to be a change point.
[0004] In the prior art, the first type of change point detection method based on statistical analysis requires accurate modeling of the time series distribution state, continuity and other properties before detection. However, the time series signals collected in complex industrial scenarios have dynamic and coupled characteristics, which are difficult to model using ordinary time series models. The second type of supervised change point detection method based on data-driven requires a large amount of data with high-quality labels for model training. However, the cost of comprehensively collecting data labels in complex industrial scenarios is high, and the labels are task-specific and cannot be applied to different industrial processes. In addition, in actual industrial time series signals, there are both sudden change points due to changes in the data distribution state, and more potential continuous change points that do not conform to the sequence consistency. Most of these two types of methods only focus on the single-class properties of the sequence and ignore the detection of continuous change points. Summary of the invention
[0005] In view of the defects in the prior art, the object of the present invention is to provide a method and system for self-supervising change point detection of one-dimensional time series signals in industrial processes.
[0006] According to one aspect of the present invention, a method for self-supervised change point detection of one-dimensional time series signals in an industrial process is provided, comprising:
[0007] Obtain a training set based on an unlabeled historical time series that does not contain change points;
[0008] Establishing a change point detection model, the change point detection model includes a time series representation encoding module based on a twin neural network and a downstream classifier consisting of a one-dimensional convolutional layer; the time series representation encoding module maps the two time series segments in the training set to the same feature space and outputs the time series representations of the two time series segments; the downstream classifier determines whether the two time series representations are similar;
[0009] Based on the training set, calculating the contrast loss of the temporal representation encoding module and the change point discrimination loss of the downstream classifier, and training the change point detection model;
[0010] Use two adjacent sliding windows to continuously extract two adjacent time segments in the test time series samples, input the trained change point detection model to carry out online detection, and determine the change point position.
[0011] Preferably, the training set includes: anchor point samples obtained through a sliding window, negative samples established through data enhancement, negative sample pairs consisting of negative samples and corresponding anchor point samples, and positive sample pairs consisting of different anchor point samples.
[0012] Preferably, the anchor point samples obtained by the sliding window include:
[0013] Using a sliding window of length w, we continuously extract the time series X = {x0, x1, x2, ..., x T}, forming a set of anchor point samples;
[0014] The starting point i of the sliding window will continue to grow as the window moves. The segment from time t = i to time t = i + w-1 is the i-th anchor point sample.
[0015] Preferably, the negative sample s corresponding to the anchor sample is constructed by data enhancement neg , the data enhancement methods include scaling, rotation, and translation; among them,
[0016] The scaling data enhancement method amplifies the time series value by a ratio of k to N (0, 1) to change the distribution state of the original time series: x' i =kx i ;
[0017] Rotational data enhancement method adds an upward or downward trend g~N(0,1) to the time series, changing the continuity of the original time series: x' i =x i ±gi;
[0018] The translation data enhancement method randomly selects subsequences of any length in the time series, adds an increasing or decreasing mutation value s~N(0,1) to its sequence value, and changes the distribution state of the original time domain sequence: x' i =x i ±s.
[0019] Preferably, the input of the time series representation encoding module based on the twin neural network is two time series segments, and the output is two time series representations, and its structure includes:
[0020] Two network branches, whose structures are exactly the same and whose network weights are shared, each network branch includes a one-dimensional convolution block, and each one-dimensional convolution block includes two convolution layers and a maximum pooling layer;
[0021] The fully connected layer fuses the outputs of the two network branches and outputs a time series representation.
[0022] Preferably, the contrast loss of the temporal representation encoding module is
[0023] in is the temporal representation of the i-th anchor point sample, is the temporal representation of the i-th negative sample, is the temporal representation of the jth anchor sample, D(·) is the similarity measurement function between the temporal representations using the cosine distance, and T1 and T2 are the temperature coefficients for adjusting the similarity values of the sample pairs.
[0024] Preferably, the downstream classifier is trained by a change point discrimination loss based on a binary cross entropy loss. The downstream classifier has two inputs. When the input is an anchor sample representation and negative sample representation When , the downstream classifier is trained to output 1, representing a negative sample Different from the anchor point sample properties; if the input is two anchor point samples and Train the downstream classifier to output 0.
[0025] According to a second aspect of the present invention, there is provided a self-supervisory change point detection system for one-dimensional time series signals of an industrial process, comprising:
[0026] Data collection module: obtain training sets based on unlabeled historical time series without change points;
[0027] A time series representation encoding module based on a twin neural network maps the two time series segments in the training set to the same feature space and outputs the time series representations of the two time series segments;
[0028] The downstream classifier composed of one-dimensional convolutional layers determines whether two temporal representations are similar;
[0029] Training module: Based on the training set, the contrast loss of the temporal representation encoding module and the change point discrimination loss of the downstream classifier are calculated to train the change point detection model;
[0030] Application module: Use two adjacent sliding windows to continuously extract two adjacent time segments in the test time series samples, input the trained change point detection model to carry out online detection, and determine the change point position.
[0031] According to a third aspect of the present invention, there is provided a terminal, comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor can be used to execute the method described, or to run the system described, when executing the program.
[0032] According to a fourth aspect of the present invention, there is provided a computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, can be used to execute the method described, or to run the system described.
[0033] Compared with the prior art, the embodiments of the present invention have at least one of the following beneficial effects:
[0034] A self-supervised change point detection method for one-dimensional time series signals of an industrial process in an embodiment of the present invention uses contrastive learning to achieve self-supervised, non-parametric detection. It can learn effective and generalized data representations to train models on data that has not been labeled and has unknown properties, thereby ensuring model accuracy and generalization for different tasks, and effectively solving problems such as the difficulty in obtaining industrial process data labels and the easy failure of labels.
[0035] A self-supervised change point detection method for one-dimensional time series signals in an industrial process in an embodiment of the present invention designs a data enhancement method for time series, provides a reference object for detecting two types of change points: mutation and continuity, so that the model comprehensively monitors multiple properties such as the distribution state of time series data, time series consistency / continuity, etc., thereby improving the accuracy of change point identification. BRIEF DESCRIPTION OF THE DRAWINGS
[0036] Other features, objects and advantages of the present invention will become more apparent from the detailed description of non-limiting embodiments made with reference to the following drawings:
[0037] Figure 1 It is a flow chart of a method for self-supervisory change point detection of one-dimensional time series signals in an industrial process according to an embodiment of the present invention;
[0038] Figure 2 It is a schematic diagram of an offline training process of a self-supervised change point detection method for a one-dimensional time series signal of an industrial process in a preferred embodiment of the present invention;
[0039] Figure 3 It is a schematic diagram of the change point online detection process of the self-supervised change point detection method of the industrial process one-dimensional time series signal in a preferred embodiment of the present invention;
[0040] Figure 4 1 is a diagram of detection results in a specific embodiment of the present invention, wherein (a) represents the change point detection from the normal mode to the periodic abnormal mode; (b) represents the change point detection from the normal mode to the rising trend abnormal mode; (c) represents the change point detection from the normal mode to the falling trend abnormal mode; (d) represents the change point detection from the normal mode to the upward step abnormal mode; (e) represents the change point detection from the normal mode to the downward step abnormal mode;
[0041] Figure 5 It is a comparison result diagram in a specific embodiment of the present invention. Specific embodiments
[0042] The present invention is described in detail below in conjunction with specific embodiments. The following embodiments will help those skilled in the art to further understand the present invention, but are not intended to limit the present invention in any form. It should be noted that, for those of ordinary skill in the art, several variations and improvements may be made without departing from the concept of the present invention. These all belong to the protection scope of the present invention.
[0043] In one embodiment of the present invention, a self-supervised change point detection method for a one-dimensional time series signal of an industrial process is provided. Figure 1 As shown, including:
[0044] S100, obtains the training set based on the unlabeled historical time series without change points;
[0045] S200, establishing a change point detection model, the change point detection model includes a time series representation encoding module based on a twin neural network and a downstream classifier composed of a one-dimensional convolutional layer; the time series representation encoding module maps two time series segments in a training set to the same feature space and outputs the time series representations of the two time series segments; the downstream classifier determines whether the two time series representations are similar;
[0046] S300, based on the training set, calculates the contrast loss of the temporal representation encoding module and the change point discrimination loss of the downstream classifier, and trains the change point detection model;
[0047] S400 uses two adjacent sliding windows to continuously extract two adjacent time segments in the test time series samples, inputs the change point detection model trained in S300 to carry out online detection, and determines the change point position.
[0048] The above embodiments can effectively improve the accuracy of online change point identification when industrial process data is unlabeled and the time series properties are not quantified. By introducing a comparative learning method for model training, the model is driven to obtain valuable and generalizable information from the original data itself, so that it can accurately identify the locations of mutational and continuous change points, and has excellent generalization ability. It is a universal change point detection method that can be applied to different industrial process monitoring.
[0049] In a preferred embodiment, S100 is implemented to construct a model training set. The unlabeled historical time series without change points is processed. Anchor samples are obtained through a sliding window. Negative samples are established through data enhancement, and negative sample pairs with the corresponding anchor sample, and different anchor samples constitute positive sample pairs. Specifically, the steps are as follows:
[0050] S101, using a sliding window of length w, continuously extracts historical time series data X = {x0x1, x2, ..., x T The segment from time t = i to time t = i + w-1 is the i-th anchor sample. Among them, the starting point i of the sliding window will continue to grow as the window moves.
[0051] S102, establish anchor samples through data enhancement The corresponding negative sample Data augmentation is to perform time domain transformation on data to expand samples. In order to distinguish the distribution state, continuity and other temporal properties of negative samples from anchor samples, the data augmentation methods used to construct negative samples include scaling, rotation and translation:
[0052] The scaling data enhancement method amplifies the sequence values of random time steps according to a certain ratio k~N(0,1), changing the distribution state of the original time domain sequence:
[0053] x′ i =kx i
[0054] The rotational data enhancement method adds an upward or downward trend g~N(0,1) to the time series, changing the continuity of the original time domain series:
[0055] x′ i =x i ±gi
[0056] The translational data enhancement method randomly selects subsequences of any length in the time series, adds increasing or decreasing mutation values s~N(0,1) to its sequence values, and changes the distribution state of the original time domain sequence:
[0057] x′ i =xi ±s
[0058] S103: Construct positive sample pairs and negative sample pairs. The i-th anchor point sample and its negative sample constitute a negative sample pair It forms a positive sample pair with other samples in the anchor sample set
[0059] The above-mentioned embodiment designs a data enhancement method for time series, provides a reference object for detecting both mutational and continuity types of change points, and enables the model to comprehensively monitor multiple properties such as the distribution state of time series data and temporal consistency / continuity, thereby improving the accuracy of change point identification.
[0060] In a preferred embodiment, S200 is implemented to establish a change point detection model. This embodiment is divided into two steps, as follows:
[0061] S201, establish a temporal representation encoding module based on a twin neural network. Map two input temporal segments of the temporal representation encoding module to the same feature space through a weight-sharing dual-channel network, and output temporal representations of the two temporal segments.
[0062] In some specific examples, S201 is divided into two steps, S201.1, two network branches are established respectively. The structures of the two network branches are exactly the same and the network weights are shared. Each network branch is composed of a one-dimensional convolution block. Each one-dimensional convolution block contains two convolution layers with a convolution kernel size of 3 and a maximum pooling layer with a pooling size of 2.
[0063] S201.2, establish a fully connected layer, fuse the outputs of the two network branches, and form a complete temporal representation encoding module. The input of the temporal representation encoding module is two temporal fragments in the time domain, and the output is two temporal representations.
[0064] S202, establish a downstream classifier composed of a one-dimensional convolutional layer. The input of the downstream classifier is the two time series representations output by the time series representation encoding module, and judge whether the two time series representations are similar.
[0065] In this embodiment, a time series representation module based on a twin neural network performs nonlinear feature transformation on the original data of the input time series segment to reduce redundant information irrelevant to the change point detection task, and converts the original data into a time series representation.
[0066] In a preferred embodiment, S300 is implemented to train the change point detection model offline by contrast learning training strategy. The training process is as follows: Figure 2 As shown. Using the training set obtained by S100, the contrast loss of the temporal representation encoding module and the change point discrimination loss of the downstream classifier are calculated to train the entire model. Specifically, Figure 2 As shown:
[0067] S301: Use the data set obtained in S100 to train the temporal representation encoding module. The anchor point sample obtains its corresponding temporal representation h through the temporal representation encoding module. anc , the negative sample is calculated through the temporal representation encoding module to obtain the temporal representation h neg The temporal properties of anchor samples at different times are similar, while the negative samples s neg The temporal nature of is different from that of the anchor point samples, and the contrast loss function is designed as follows:
[0068]
[0069] in is the temporal representation of the i-th anchor point sample, is the temporal representation of the i-th negative sample, is the temporal representation of the jth anchor sample. D(·) is the similarity measurement function between the temporal representations, which is calculated using the cosine distance. T1 and T2 are the temperature coefficients for adjusting the similarity values of the sample pairs. By training the temporal representation encoding module with this loss function, the distance between the temporal representations of different anchor samples with similar properties can be made as close as possible, while the negative sample representations are as far away from them as possible, thereby improving the distinguishability of the temporal representation.
[0070] S302: Using the obtained time series representation, train the downstream classifier through the change point discrimination loss based on the binary cross entropy loss. The downstream classifier has two inputs. When the input is the anchor sample representation and negative sample representation When , the downstream classifier is trained to output 1, representing a negative sample Different from the anchor point samples. If the input is two anchor point samples and The downstream classifier is trained to output 0. Specifically, the calculation formula is:
[0071]
[0072] Among them, y i is the true label representing whether the input data of the downstream classifier is similar. is the discrimination result given by the downstream classifier
[0073] In a preferred embodiment, S400 is implemented, using the trained change point detection model, using two adjacent sliding windows to continuously extract two adjacent time segments in the test time series sample, inputting the change point detection model to perform online detection, and determining the change point position. Specifically, Figure 3 As shown:
[0074] S401: Use two sliding windows with a width of w to extract time series segments from the test time series stream to be detected, which are historical segments s and history ={x k-w , x k+1-w , ..., x k-1} and future fragments future ={x k+1 x w , x k+2-w , ..., x k}. The future segment contains the latest data point x that arrived in the time series stream k , while the historical fragment does not include it. Online detection of change points requires checking the latest data point x k Whether it is a change point.
[0075] S402: The historical fragment S history and future fragments future Input the trained time series representation encoding module to obtain the historical time series representation h history and future temporal representation h future .
[0076] S403: Input the two time series representations obtained in S402 into the trained downstream classifier. If the output of the downstream classifier is 1, it means that the nature of the future segment is different from the historical segment, that is, the latest time point sequence value x in the future segment is k The addition of changes the nature of the data. At this time, the latest data point x k If the output of the downstream classifier is 0, it means that the latest data point is not a change point. At this time, the sliding window of S401 is further moved backward to extract the new data continuously generated in the test time series flow, and steps S401-S403 are repeated for testing until a change point is detected.
[0077] Based on the same inventive concept, in another embodiment of the present invention, a self-supervised change point detection system for one-dimensional time series signals of an industrial process is provided, comprising:
[0078] Data collection module: obtain training sets based on unlabeled historical time series without change points;
[0079] A time series representation encoding module based on a twin neural network maps the two time series segments in the training set to the same feature space and outputs the time series representations of the two time series segments;
[0080] The downstream classifier composed of one-dimensional convolutional layers determines whether two temporal representations are similar;
[0081] Training module: Based on the training set, the contrast loss of the temporal representation encoding module and the change point discrimination loss of the downstream classifier are calculated to train the change point detection model;
[0082] Application module: Use two adjacent sliding windows to continuously extract two adjacent time segments in the test time series samples, input the trained change point detection model to carry out online detection, and determine the change point position.
[0083] The modules / units in the above examples of the present invention may specifically refer to the implementation techniques of the corresponding steps of the method for self-supervised change point detection of one-dimensional time series signals of industrial processes in the above embodiments, which will not be described in detail here.
[0084] Based on the same inventive concept, in other embodiments of the present invention, a terminal is provided, comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor can be used to execute the method described, or to run the system described, when executing the program.
[0085] Based on the same inventive concept, in other embodiments of the present invention, a computer-readable storage medium is provided, on which a computer program is stored, and when the program is executed by a processor, it can be used to execute the method described, or to run the system described.
[0086] In order to verify the feasibility and technical effect of the above-mentioned embodiment, in a specific implementation of the present invention, the online detection signal of the white body processing deviation is collected to form a white body data set. The processing deviations detected by multiple white bodies continuously produced on the production line constitute a sequence, which can reflect the production status of the processing process. A total of 113 detection point deviation sequences were collected. Each detection point corresponds to a one-dimensional sequence of processing deviations. The time point when the sequence switches from normal mode to abnormal mode is the change point that needs to be identified. Figure 4 The five figures (a), (b), (c), (d), and (e) are respectively result figures of the change point detection method using the embodiment of the present invention, which can effectively identify the change point positions.
[0087] Figure 5 It is the result of the change point detection method of the embodiment of the present invention and the comparison object method, and the method effect is measured using three indicators: precision, recall and mean absolute error. The method of the embodiment of the present invention has obvious advantages in all three indicators.
[0088] The method in the embodiment of the present invention weakens the dependence of the method on training labels through contrastive learning. Therefore, when industrial process data is unlabeled and the data properties are not quantified, the change point detection accuracy and the generalization ability of the method can be improved through self-supervision and non-parametric forms, and can cope with different application scenarios. The method in the embodiment of the present invention constructs reference objects for detecting two types of change points, namely mutation and continuity, through data enhancement, so that the change point detection model can comprehensively monitor multiple properties such as the distribution state of time series data, time series consistency / continuity, thereby improving the accuracy of the detection results.
[0089] The above describes the specific embodiments of the present invention. It should be understood that the present invention is not limited to the above specific embodiments, and those skilled in the art can make various modifications or variations within the scope of the claims, which does not affect the essence of the present invention. The above preferred features can be used in any combination without conflicting with each other.
Claims
1. A self-supervised change point detection method for one-dimensional time series signals in industrial processes, characterized in that: include: Obtain a training set based on an unlabeled historical time series that does not contain change points; Establishing a change point detection model, the change point detection model includes a time series representation encoding module based on a twin neural network and a downstream classifier consisting of a one-dimensional convolutional layer; the time series representation encoding module maps the two time series segments in the training set to the same feature space and outputs the time series representations of the two time series segments; the downstream classifier determines whether the two time series representations are similar; Based on the training set, calculating the contrast loss of the temporal representation encoding module and the change point discrimination loss of the downstream classifier, and training the change point detection model; Use two adjacent sliding windows to continuously extract two adjacent time segments in the test time series samples, input the trained change point detection model to carry out online detection, and determine the change point position.
2. The method for self-supervised change point detection of one-dimensional time series signals in an industrial process according to claim 1, characterized in that: The training set includes: anchor point samples obtained through sliding windows, negative samples established through data enhancement, negative sample pairs consisting of negative samples and corresponding anchor point samples, and positive sample pairs consisting of different anchor point samples.
3. The method for self-supervised change point detection of one-dimensional time series signals in an industrial process according to claim 2, characterized in that: The anchor point samples obtained by sliding window include: Using a sliding window of length w, we continuously extract the time series X = {x0, x1, x2, ..., x T }, forming a set of anchor point samples; The starting point i of the sliding window will continue to grow as the window moves. The segment from time t = i to time t = i + w-1 is the i-th anchor point sample.
4. The method for self-supervised change point detection of one-dimensional time series signals in industrial processes according to claim 3, characterized in that: Construct the negative sample s corresponding to the anchor sample through data enhancement neg , the data enhancement methods include scaling, rotation, and translation; among them, The scaling data enhancement method amplifies the time series value according to the ratio k~N(0,1) to change the distribution state of the original time series: x′ i =kx i ; Rotational data enhancement method adds an upward or downward trend g~N(0,1) to the time series, changing the continuity property of the original time series: x′ i =x i ±gi; The translation data enhancement method randomly selects subsequences of any length in the time series, adds rising or falling mutation values s~N(0,1) to the subsequence values, and changes the distribution state of the original time domain sequence: x′ i =x i ±s.
5. The method for self-supervised change point detection of one-dimensional time series signals in industrial processes according to claim 1, characterized in that: The input of the temporal representation encoding module based on the twin neural network is two temporal segments, and the output is two temporal representations; The structure of the timing characterization encoding module includes: Two network branches, both of which have exactly the same structure and share network weights. Each network branch includes a one-dimensional convolution block, and each one-dimensional convolution block includes two convolution layers and a maximum pooling layer; The fully connected layer fuses the outputs of the two network branches and outputs a time series representation.
6. The method for self-supervised change point detection of one-dimensional time series signals in industrial processes according to claim 1, characterized in that: The contrast loss of the temporal representation encoding module is in is the temporal representation of the i-th anchor point sample, is the temporal representation of the i-th negative sample, is the temporal representation of the jth anchor sample, D(·) is the similarity measurement function between the temporal representations using the cosine distance, and T1 and T2 are the temperature coefficients for adjusting the similarity values of the sample pairs.
7. The method for self-supervised change point detection of one-dimensional time series signals in industrial processes according to claim 6, characterized in that: The downstream classifier is trained by the change point discriminant loss based on the binary cross entropy loss. The downstream classifier has two inputs. When the input is the anchor sample representation and negative sample representation When , the downstream classifier is trained to output 1, representing a negative sample Different from the anchor point sample properties; if the input is two anchor point samples and Train the downstream classifier to output 0.
8. A self-supervised change point detection system for one-dimensional time series signals in industrial processes, characterized in that: include: Data collection module: obtain training sets based on unlabeled historical time series without change points; A time series representation encoding module based on a twin neural network maps the two time series segments in the training set to the same feature space and outputs the time series representations of the two time series segments; The downstream classifier composed of one-dimensional convolutional layers determines whether two temporal representations are similar; Training module: Based on the training set, the contrast loss of the temporal representation encoding module and the change point discrimination loss of the downstream classifier are calculated to train the change point detection model; Application module: Use two adjacent sliding windows to continuously extract two adjacent time segments in the test time series samples, input the trained change point detection model to carry out online detection, and determine the change point position.
9. A terminal comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, characterized in that: When the processor executes the program, it can be used to execute the method described in any one of claims 1 to 7, or to run the system described in claim 8.
10. A computer-readable storage medium having a computer program stored thereon, characterized in that: When the program is executed by a processor, it can be used to execute the method described in any one of claims 1 to 7, or to run the system described in claim 8.
Citation Information
Patent Citations
Time series data anomaly detection method based on self-supervised learning multi-head attention network
CN115309736A
Self-supervised action recognition method based on cross-modal time sequence contrast learning
CN116721458A
