Context awareness-based contrast coding semi-supervised time sequence classification system

By using a context-aware contrastive coding semi-supervised time series classification system, which employs a Crossformer encoder and mask contrastive learning method, the problem of scarce time series data labels in the process industry is solved, and efficient classification with limited labeled data is achieved.

CN118898022APending Publication Date: 2024-11-05ZHEJIANG SCI-TECH UNIV
View PDF 0 Cites 1 Cited by

Patent Information

Application Number
CN202410777001.0
Authority / Receiving Office
CN · China
Patent Type
Applications(China)
Current Assignee / Owner
Filing Date
2024-06-17
Publication Date
2024-11-05

AI Technical Summary

Technical Problem

In process industries, labeled time-series data is scarce and costly, and existing technologies struggle to effectively utilize large amounts of unlabeled data for classification tasks.

Method used

A context-aware contrastive coding semi-supervised time series classification system is adopted, which includes a two-stage time-context-aware data augmentation module, a cross-dimensional spatial coding module, and a multi-task contrastive learning module. It uses the Crossformer encoder and mask contrastive learning method to capture temporal and cross-dimensional dependencies.

Benefits of technology

It improves model performance in scenarios with limited labeled data by enhancing classification accuracy through rich feature representations and cross-dimensional dependency learning.

✦ Generated by Eureka AI based on patent content.

Smart Images

  • Figure CN118898022A_ABST
    Figure CN118898022A_ABST
Patent Text Reader

Abstract

The invention discloses a context awareness-based contrast coding semi-supervised time sequence classification system, which comprises a two-stage time sequence-context awareness data enhancement module, a cross-dimension space coding module and a multi-task contrast learning module, the two-stage context-aware data enhancement module is used for performing data enhancement on an original data level and a data potential space representation level, learning data time sequence dependence and cross-dimension dependence, and generating stable enhanced view codes; the cross-dimension space coding module realizes refined representation of dimension levels by using a Crossform encoder, and captures a complex dependency relationship between dimensions; the multi-task comparative learning module captures time sequence dependence and cross-dimension dependence through mask comparative learning and double-view prediction tasks. According to the method, robust time sequence feature representation can be effectively obtained, the effectiveness and universality of the method are verified on a plurality of public data sets, and the method is particularly suitable for semi-supervised time sequence classification and has great significance in processing the situation that a large amount of marked data is lacked.
Need to check novelty before this filing date? Find Prior Art

Description

Technical Field

[0001] The present invention belongs to the technical field of time series analysis and deep learning, and specifically relates to a context-aware contrast coding semi-supervised time series classification system. Background Art

[0002] Semi-supervised learning is a special learning method that combines the characteristics of supervised learning and unsupervised learning. In semi-supervised learning, the training data set consists of two parts: one part is data with clear labels, which can be used directly to train the model; the other part is unlabeled data, which cannot be used directly for supervised learning, but can indirectly help improve the performance of the model in some way. In actual situations, labeled data is often limited, especially in some fields, such as process industries, where manual labeling of large amounts of data is not only time-consuming but also costly. Therefore, those skilled in the art have proposed a semi-supervised learning method to solve the situation where there is less labeled data but more unlabeled data.

[0003] Semi-supervised classification is an application form of semi-supervised learning. It is a classification method for when there is less labeled data and a large amount of unlabeled data. It uses both labeled data and unlabeled data for training, thereby obtaining a model with better performance than the classifier trained with labeled data alone. Specifically, semi-supervised classification mainly uses a labeled data set and a semi-supervised data set for training and obtains a classifier that can accurately predict the label of a new data point, where the labeled data set contains multiple data samples and the label corresponding to each sample; the unlabeled data set only contains data samples without corresponding label information.

[0004] When semi-supervised learning is applied to time series data, it forms semi-supervised time series learning. Semi-supervised time series learning is the application of semi-supervised learning in the field of time series data. It considers that the time series data set contains both labeled and unlabeled time series samples, and aims to improve the model's performance in classification, prediction or other tasks of time series data by utilizing the information of unlabeled time series data.

[0005] The semi-supervised time series classification task mainly discusses how to use a large amount of unlabeled time series data to classify related time series tasks in scenarios where the amount of time series label data is extremely scarce or the labeling cost is expensive. Common application scenarios are mainly concentrated in process industries. In these industries, production data is mostly in the form of time series data, and because of the assembly line production characteristics of process industries, the amount of data generated by the entire production process is very large. For product quality inspection, due to the limitations of actual operations, sampling inspection is often used, and it is impossible to perform quality inspection on each product. The semi-supervised time series classification task can effectively solve this problem. Based on this, the present invention proposes a context-aware contrast coding semi-supervised time series classification system, which uses learning cross-dimensional dependencies to improve the classification performance of the model in scenarios with a small amount of time series labeled data. Summary of the invention

[0006] Based on the above status of the prior art, the present invention proposes a context-aware contrast coding semi-supervised time series classification system.

[0007] In order to achieve the above purpose, the present invention adopts the following technical solutions:

[0008] A context-aware contrastive coding semi-supervised time series classification system includes a two-stage time series-context-aware data enhancement module, a cross-dimensional spatial encoding module, and a multi-task contrastive learning module; wherein:

[0009] The two-stage temporal-context-aware data enhancement module is used to enhance data at the raw data level and the data latent space representation level, learn data temporal dependencies and cross-dimensional dependencies, and generate stable enhanced view encodings;

[0010] The cross-dimensional spatial encoding module uses a Crossformer encoder to extract features, achieves refined representation at the dimensional level, and captures complex dependencies between dimensions. The Crossformer encoder is introduced for feature extraction to overcome the limitations of the traditional Transformer architecture, and can learn to distinguish potential interdependencies between dimensions and refine the representation at the dimensional level.

[0011] The multi-task contrastive learning module captures temporal dependency and cross-dimensional dependency by performing mask contrastive learning and dual-view mask prediction tasks.

[0012] In the present invention, the two-stage time series-context-aware data enhancement module enhances the time series data through two consecutive stages, thereby enhancing the model's learning ability for data time series dependency and cross-dimensional dependency.

[0013] The multi-task contrastive learning module strengthens the model's perception of temporal and cross-dimensional information by designing specific learning tasks.

[0014] Preferably, the two-stage time series-context-aware data enhancement module includes two stages: time series data enhancement and dimensional space data enhancement.

[0015] Preferably, the dimensional space data enhancement stage is used to enhance the dimensional space of time series data and enhance the model's understanding of time and dimensional context information.

[0016] Further preferably, in the time series data enhancement stage, two discrete data enhancement operations are used along the time axis, introducing a combination of overturning and jittering, as well as a combination of shifting and scaling, which can deeply study the time context and promote the encoder to extract the essential characteristics of the data.

[0017] Specifically, Overturning involves segment-level operations across time. Given a sample x containing T time steps, the present invention divides x into N segments along the time axis and randomly selects r×N segments from all segments, where r∈[0,1] represents the selection ratio, and randomly flips the time order within each segment. On the one hand, flipping and dithering are combined, where the dithering adopts Gaussian noise N(μ,σ 2 ) to get the enhanced view x1 + On the other hand, by combining the shift and the scaling, it is possible to obtain another temporally enhanced view x2 + Formally, the proposed combined operation of time-context data augmentation is shown in the following formula, namely:

[0018] x1 + =overturning(x)+N(μ,σ 2 )

[0019] x2 + =scaling(shifting(x))

[0020] Among them, σ is the standard deviation of Gaussian distribution, μ is the mean of Gaussian distribution, x is the original data, x1 + With x2 + are the corresponding enhanced data respectively.

[0021] Preferably, the dimensional space data enhancement stage is used to enhance the dimensional space of the time series data by introducing random noise into the latent representation of the encoder to enrich the feature representation of the data, thereby enriching the temporal and dimensional context information contained in the latent feature representation.

[0022] Specifically, in the second stage of data augmentation, first, the data augmentation view x1 is + and x2 + Input encoder to obtain the potential representation h1 in the encoding space + ∈T×D and h2 + ∈T×D, where D represents the dimension of each time step. Then, in h1 + and h2 + Random noise in the dimension is introduced in , as an implicit attribute-context enhancement. One of the reasons for performing random masking in the encoding space is that considering the one-dimensional time series data, it is impractical to perform dimensionality enhancement on the original data. Therefore, a cascaded enhanced view encoding is obtained, denoted as z1 + and z2 + , where z1 + =Dropout(h1 + )∈T×D,z2 + =

[0023] Dropout(h2 + )∈T×D, this encoding method enriches the temporal and dimensional context information contained in the latent feature representation.

[0024] Preferably, the mask contrast learning task adopts a dual-view method to construct positive sample pairs and negative sample pairs, and performs learning by distinguishing contrast.

[0025] The dual-view mask prediction task enables the encoder to effectively learn cross-dimensional dependencies by masking at the dimensional level in the encoder spatial representation.

[0026] Preferably, the mask contrast learning task uses a dual-view method to construct positive sample pairs and negative sample pairs, and learns by distinguishing contrast. The core of the mask contrast learning method is to generate two different but related data views and construct positive sample pairs and negative sample pairs, thereby generating multiple samples. Subsequently, these samples are sequentially passed through the encoder, dimensional mask, and Crossformer-based encoder for feature extraction.

[0027] Specifically, starting from an initial queue of N input data, these data are subjected to a two-stage data augmentation process to obtain two different data views, resulting in 2N samples. Subsequently, these samples are sequentially passed through an encoder, a dimension mask, and a Crossformer-based encoder for feature extraction. The above process ultimately produces the corresponding context vectors ct1 and ct2. In the proposed framework, the context vectors obtained from the same sample are designated as positive pairs, formally represented as (ct1, ct2). These vectors form negative pairs with the remaining 2N-2 sample pairs. Regarding the overall contrast loss for this task The corresponding equation is shown as follows:

[0028]

[0029] Among them, similarity(u,v) represents the cosine similarity between vectors u and v, τ is the temperature parameter, and N represents the amount of data.

[0030] Preferably, the dual-view mask prediction task allows the encoder to effectively learn cross-dimensional dependencies by masking at the dimensional level in the encoder spatial representation. In traditional methods, masks are mainly used for data normalization. Unlike traditional methods, the dual-view prediction task proposed in the present invention mainly emphasizes masking at the dimensional level in the encoder spatial representation, thereby improving the encoder's ability to learn interdependencies between data dimensions.

[0031] Specifically, the enhanced data is encoded using an encoder to obtain spatial variables z1 and z2, and the original data is also encoded to obtain the encoder spatial variable h3. The goal is to use the context vectors extracted from z1 and z2 based on the Crossformer, namely ct1 and ct2, and gradually reconstruct the original data, namely h3, by implementing simple logarithmic linear transformations on these context vectors to effectively promote the encoder to learn the interdependencies between data dimensions, thereby generating a robust data representation. The detailed process of this task is as follows: After receiving N input samples and applying dual-view enhancement, 2N context vectors are generated, denoted as ct, and there are N context vectors corresponding to the original data. By applying appropriate logarithmic linear transformations to the two enhanced views, their predicted values ​​for each time step of the original data can be obtained. These values ​​are associated with the latent variables of the corresponding original data as positive samples, and are formed as negative samples with the remaining N-1 instances, and the two are trained with contrast loss. Mask-based prediction loss The specific expression is as follows:

[0032]

[0033] Among them, h t,k represents the representation of the kth time step in the tth original data in the latent space, T represents the length of the encoder space representation in terms of time steps, and N represents the number of samples. k represents a linear function that maps ct back to h t ,k has the same dimension. The total loss of the dual-view mask prediction task is expressed as is a combination of the above two prediction task losses, as shown below:

[0034]

[0035] Among them, λ1 and λ2 are hyperparameters of the loss weight.

[0036] This paper proposes a context-aware contrastive coding semi-supervised time series classification system, which mainly adopts attribute mask enhancement for latent views and mask prediction and contrastive learning task loss based on Crossformer encoder. This paper learns a wider range of time series data feature representations by extracting dimensional masks of latent space representation to capture cross-dimensional dependencies, thereby improving the model performance in scenarios with a small amount of labeled data. BRIEF DESCRIPTION OF THE DRAWINGS

[0037] In order to more clearly illustrate the embodiments of the present invention or the technical solutions in the prior art, the drawings in the embodiments are briefly introduced below.

[0038] Figure 1 This is an overall framework diagram of a context-aware contrast coding semi-supervised time series classification system according to a preferred embodiment of the present invention.

[0039] Figure 2 This is a time-series dimension data enhancement diagram proposed in a preferred embodiment of the present invention.

[0040] Figure 3 A schematic diagram of the Crossformer cross-dimensional dependency proposed in a preferred embodiment of the present invention.

[0041] Figure 4 Figure 2 is a dual-view mask prediction task diagram proposed for a preferred embodiment of the present invention. DETAILED DESCRIPTION

[0042] The following describes the embodiments of the present invention through specific embodiments, and those skilled in the art can easily understand other advantages and effects of the present invention from the contents disclosed in this specification. The present invention can also be implemented or applied through other different specific embodiments, and the details in this specification can also be modified or changed in various ways based on different viewpoints and applications without departing from the spirit of the present invention. It should be noted that the following embodiments and features in the embodiments can be combined with each other without conflict.

[0043] The present invention mainly focuses on the performance optimization of semi-supervised time series classification problems, and mainly adopts attribute mask enhancement for potential views and mask prediction and contrast learning task loss based on Crossformer encoder. By extracting the dimensional mask of the latent space representation to capture cross-dimensional dependencies, a wider range of time series data feature representations can be learned, thereby improving the model performance in scenarios with a small amount of labeled data. The specific description is given in the form of an embodiment.

[0044] A preferred embodiment of the present invention provides a context-aware contrastive coding semi-supervised time series classification system for improving model performance in scenarios with a small amount of labeled data.

[0045] This embodiment provides a method to improve the model performance in scenarios with a small amount of labeled data, such as Figure 1 As shown in the figure, it includes a two-stage temporal-context-aware data enhancement module, a cross-dimensional spatial encoding module, and a multi-task contrastive learning module, as follows:

[0046] The two-stage time-series-context-aware data enhancement module is used to enhance data at the raw data level and the data latent space representation level, learn data temporal dependencies and cross-dimensional dependencies, and generate stable enhanced view encodings. The two-stage time-series-context-aware data enhancement module enhances time series data through two consecutive stages, thereby enhancing the learning ability of the machine learning model for data temporal dependencies and cross-dimensional dependencies.

[0047] The cross-dimensional spatial encoding module uses the Crossformer encoder to achieve refined representation at the dimensional level and capture the complex dependencies between dimensions. The cross-dimensional spatial encoding module introduces the Crossformer encoder for feature extraction, overcoming the limitations of the traditional Transformer architecture, and can effectively distinguish the potential interdependencies between dimensions and refine the representation at the dimensional level.

[0048] The multi-task contrastive learning module uses mask contrastive learning and dual-view prediction tasks to capture temporal dependencies and cross-dimensional dependencies. The multi-task contrastive learning module strengthens the model's perception of temporal and cross-dimensional information by designing specific learning tasks.

[0049] like Figure 2 As shown, in this embodiment, the time series-context-aware data enhancement module deploys the following two stages.

[0050] (1) Time series data enhancement stage

[0051] The temporal data enhancement stage uses two discrete data enhancement operations along the time axis, introducing a combination of overturning and jittering, as well as a combination of shifting and scaling, which can deeply study the temporal context and promote the encoder to extract the essential features of the data.

[0052] Specifically, Overturning involves segment-level operations across time. Given a sample x containing T time steps, this embodiment divides x into N segments along the time axis and randomly selects r×N segments from all segments, where r∈[0,1] represents the selection ratio, and randomly flips the time order within each segment. On the one hand, flipping and jittering are combined, where jittering uses Gaussian noise N(μ,σ2) to obtain an enhanced view x1 + On the other hand, by combining the shift and the scaling, it is possible to obtain another temporally enhanced view x2 + Formally, the proposed combined operation of time-context data augmentation is shown in the following formula, namely:

[0053] x1 + =overturning(x)+N(μ,σ2)

[0054] x2 + =scaling(shifting(x))

[0055] Among them, σ is the standard deviation of Gaussian distribution, μ is the mean of Gaussian distribution, x is the original data, x1 + With x2 + are the corresponding enhanced data respectively.

[0056] (2) Dimensional space data enhancement stage

[0057] The dimensional space data enhancement stage is used to enhance the dimensional space of time series data by introducing random noise into the encoder's latent representation to enrich the feature representation of the data, thereby enriching the temporal and dimensional contextual information contained in the latent feature representation.

[0058] Specifically, in the second stage of data augmentation, first, the data augmentation view x1 is + and x2 + Input encoder to obtain the potential representation h1 in the encoding space + ∈T×D and h2 + ∈T×D, where D represents the dimension of each time step. Then, in h1 + and h2 +Random noise in the dimension is introduced in , as an implicit attribute-context enhancement. One of the reasons for performing random masking in the encoding space is that considering the one-dimensional time series data, it is impractical to perform dimensionality enhancement on the original data. Therefore, a cascaded enhanced view encoding is obtained, denoted as z1 + and z2 + ,in

[0059] z1 + =Dropout(h1 + )∈T×D,z2 + = Dropout(h2 + )∈T×D, this encoding method enriches the temporal and dimensional context information contained in the latent feature representation.

[0060] In this embodiment, the cross-dimensional spatial encoding module introduces a Crossformer encoder for feature extraction, which is significantly different from the traditional Transformer architecture. The Crossformer encoder uses discrete encoding of different dimensions, so it explicitly targets cross-dimensional correlations, is particularly suitable for dual-view mask prediction tasks, and has shown significant effectiveness in practical applications.

[0061] Specifically, this embodiment uses an encoder to encode and mask two data views to generate latent representations h1 and h2. Then, dimension segment encoding is introduced into these representations through an embedding layer for dimension segments. A two-stage time window attention mechanism is then used to enhance the extraction of temporal dependencies and cross-dimensional dependencies across different time window levels. This process produces the corresponding context vectors ct1 and ct2. Figure 3 The process of feature extraction using Crossformer is visually described. The dimension segmentation encoding process is as follows:

[0062]

[0063] Among them, h 1:T represents the latent variable h after dimension segment encoding, L seg is the length of the corresponding segment, represents the i-th segment with dimension d, T represents the total number of time steps, and D represents the total dimension of the features.

[0064] In this embodiment, the multi-task contrastive learning module deploys the following two tasks, where the detailed information of the dual-view mask prediction task is as follows: Figure 4 shown.

[0065] (1) Mask contrast learning task

[0066] Mask contrastive learning uses a dual-view method to construct positive and negative sample pairs, and learns by distinguishing contrast. The core of the mask contrastive learning method is to generate multiple samples by generating two different but related data views and constructing positive and negative sample pairs. Subsequently, these samples are sequentially passed through the encoder, dimensional mask, and Crossformer-based encoder for feature extraction.

[0067] Specifically, starting from an initial queue containing N input data, these data are subjected to a two-stage data augmentation process to obtain two different data views, resulting in 2N samples. Subsequently, these samples are passed through an encoder, a dimension mask, and a Crossformer-based encoder for feature extraction. The above process ultimately generates the corresponding context vectors ct1 and ct2. Within the framework proposed in the present invention, the context vectors obtained from the same sample are designated as positive pairs, formally expressed as (ct1, ct2). These vectors form negative pairs with the remaining 2N-2 sample pairs. Regarding the overall contrast loss for this task The corresponding equation is shown as follows:

[0068]

[0069] Among them, similarity(u,v) represents the cosine similarity between vectors u and v, τ is the temperature parameter, and N represents the amount of data.

[0070] (2) Dual-view mask prediction task

[0071] The dual-view prediction task allows the encoder to effectively learn cross-dimensional dependencies by masking at the dimensional level in the encoder spatial representation. Unlike traditional methods, where masks are mainly used for data normalization, the dual-view prediction task proposed in the present invention mainly emphasizes masking at the dimensional level in the encoder spatial representation, thereby improving the encoder's ability to learn inter-dimensional dependencies between data dimensions.

[0072] Specifically, the enhanced data is encoded using an encoder to obtain spatial variables z1 and z2, and the original data is also encoded to obtain the encoder spatial variable h3. The goal is to use the context vectors extracted from z1 and z2 based on the Crossformer, namely ct1 and ct2, and gradually reconstruct the original data, namely h3, by implementing simple logarithmic linear transformations on these context vectors to effectively promote the encoder to learn the interdependencies between data dimensions, thereby generating a robust data representation. The detailed process of this task is as follows: After receiving N input samples and applying dual-view enhancement, 2N context vectors are generated, denoted as ct, and there are N context vectors corresponding to the original data. By applying appropriate logarithmic linear transformations to the two enhanced views, their predicted values ​​for each time step of the original data can be obtained. These values ​​are associated with the latent variables of the corresponding original data as positive samples, and are formed as negative samples with the remaining N-1 instances, and the two are trained with contrast loss. Mask-based prediction loss The specific expression is as follows:

[0073]

[0074] Among them, h t ,k represents the representation of the kth time step in the tth original data in the latent space, T represents the length of the encoder space representation in terms of time steps, and N represents the number of samples. k represents a linear function that maps ct back to h t ,k has the same dimension. The total loss of the dual-view mask prediction task is expressed as is a combination of the above two prediction task losses, as shown below:

[0075]

[0076] Among them, λ1 and λ2 are hyperparameters of the loss weight.

[0077] In summary, the present invention proposes a context-aware contrastive coding semi-supervised time series classification system. The present invention uses a two-stage time-context-aware data enhancement module to effectively learn the time dependency and cross-dimensional dependency of data and generate stable enhanced view coding; uses a cross-dimensional spatial coding module to perform refined representation at the dimensional level and capture complex dependencies between dimensions; and uses a multi-task contrastive learning module to further enhance the ability to capture time dependency and cross-dimensional dependency.

[0078] The embodiments described above are merely descriptions of preferred implementation modes of the present invention, and are not intended to limit the scope of the present invention. Without departing from the design spirit of the present invention, various modifications and improvements made to the technical solutions of the present invention by ordinary technicians in this field should all fall within the protection scope of the present invention.

Claims

1. A semi-supervised time series classification system based on context-aware contrastive coding, Its characteristics are It includes a two-stage temporal-context-aware data enhancement module, a cross-dimensional spatial encoding module, and a multi-task contrastive learning module; The two-stage temporal-context-aware data enhancement module is used to enhance data at the raw data level and the data latent space representation level, learn data temporal dependencies and cross-dimensional dependencies, and generate enhanced view encodings; The cross-dimensional space encoding module uses a Crossformer encoder to extract features, achieve dimensional-level quantitative representation, and capture dependencies between dimensions; The multi-task contrastive learning module performs mask contrastive learning and dual-view mask prediction tasks, capturing temporal dependencies and cross-dimensional dependencies.

2. The context-aware contrastive coding semi-supervised time series classification system according to claim 1, Its characteristics are: The two-stage time series-context-aware data enhancement module enhances the time series data through two consecutive stages, and the two consecutive stages specifically include a time series data enhancement stage and a dimensional space data enhancement stage.

3. The context-aware contrastive coding semi-supervised time series classification system according to claim 2, Its characteristics are: In the temporal data enhancement stage, two discrete data enhancement operations are used along the time axis, namely, a combination of overturning and jittering, and a combination of shifting and scaling, so that the encoder can extract the essential features of the data.

4. The context-aware contrastive coding semi-supervised time series classification system according to claim 3, Its characteristics are: The Overturning involves segment-level operations across time. Given a sample x containing T time steps, it is divided into N segments along the time axis, and r×N segments are randomly selected from all segments, where r∈[0,1] represents the selection ratio, and the time order within each segment is randomly flipped. The flipping and jittering are combined, where the jittering adopts Gaussian noise N(μ,σ 2 ) for an enhanced view Combine swiping and zooming to get another time-enhanced view The combined operation of time-context data enhancement is shown in the following formula: Among them, σ is the standard deviation of the Gaussian distribution, and μ is the mean of the Gaussian distribution.

5. The context-aware contrastive coding semi-supervised time series classification system according to claim 4, Its characteristics are: Dimensional Space The data enhancement stage is used to enhance the dimensional space of time series data by introducing random noise into the encoder’s latent representation to expand the feature representation of the data.

6. The context-aware contrastive coding semi-supervised time series classification system according to claim 5, Its characteristics are: dimensional space In the data enhancement phase, the data enhancement view is first and Input encoder to obtain the latent representation in the encoding space and where D represents the dimension of each time step; then, and Introducing random noise in the dimension as implicit attribute-context enhancement; The cascade enhanced view encoding is obtained, which is represented as and in 7. The context-aware contrastive coding semi-supervised time series classification system according to claim 1, Its characteristics are: The mask contrast learning task adopts a dual-view method to construct positive sample pairs and negative sample pairs, and performs learning through distinguishing contrast.

8. The context-aware contrastive coding semi-supervised time series classification system according to claim 7, Its characteristics are: mask Contrastive loss for contrastive learning tasks The equation is shown in the following formula: Among them, similarity(u,v) represents the cosine similarity between vectors u and v, τ is the temperature parameter, and N represents the amount of data.

9. The context-aware contrastive coding semi-supervised time series classification system according to claim 8, Its characteristics are: The dual-view mask prediction task enables the encoder to learn cross-dimensional dependencies by masking at the dimensional level in the Crossformer encoder spatial representation.

10. The context-aware contrastive coding semi-supervised time series classification system according to claim 9, Its characteristics are: Mask-based prediction loss in dual-view mask prediction task The specific expression is as follows: Among them, h t , k represents the representation of the kth time step in the tth original data in the latent space, T represents the length of the encoder space representation in terms of time steps, and N represents the number of samples; W k represents a linear function that maps ct back to h t ,k has the same dimension; the total loss of the dual-view mask prediction task is expressed as As shown below: Among them, λ 1 and λ 2 is a hyperparameter of the loss weight.

Citation Information

Cited By

  • Time series data classification system based on mixed contrast learning network

    CN120105190A