Self-supervision-based small sample time sequence identification method and model construction method

Abstract

The invention discloses a self-supervision-based small sample time sequence identification method and a model construction method, and the method comprises the steps: constructing a network model for self-supervision pre-training, which comprises a feature extraction module and mask reconstruction; random sampling and data enhancement are carried out from unlabeled data, and a pseudo-label small sample time sequence task is constructed; a self-supervised contrast mask method is used in a pseudo-label small sample time sequence task to help a model to learn global information and time dependence in samples and consistency and discriminative information between the samples so as to enhance the feature extraction capability of the model; and finely tuning the model on the downstream small sample time sequence identification task and evaluating the model performance. According to the small sample time sequence identification method and system based on self-supervision, effective time sequence feature extraction can be achieved for small sample time sequence data, dependence on a large amount of labeled data is reduced, and the identification accuracy and generalization ability of a time sequence identification model under the small sample condition are improved.

Description

Technical Field

[0001] The present invention belongs to the field of multivariate time series analysis and application, and in particular relates to a small sample time series recognition method and a model building method based on self-supervision. Background Art

[0002] Time series data plays an important role in various practical applications, including financial market analysis, weather forecasting, medical diagnosis, industrial equipment monitoring and many other fields. These fields rely on the dynamic changes of time series data to make important decisions. However, time series data often have complex situations such as non-stationarity and noise interference, and may contain multiple dimensions. In addition, time series data has a strong temporal dependency, that is, the data at one point in time usually depends on the previous data. These problems make traditional statistical modeling and machine learning methods based on feature engineering show great limitations when processing time series data.

[0003] Unlike traditional time series analysis methods, deep learning does not require complex feature engineering, and can automatically learn time series representations in a data-driven manner and achieve excellent recognition results. Although deep learning has achieved great success in the field of time series recognition, the success of deep learning often relies on the support of a large amount of labeled data. In the fields of medical data, equipment monitoring, etc., labeled data is scarce due to the difficulty of collecting time series data and the high cost of labeling. In the case of small sample time series with limited labeled data, deep learning models are usually difficult to extract effective time series features from the data, and overfitting is prone to occur during model training, resulting in poor recognition performance. In the prior art, Chinese patent CN116070128A discloses a time series equipment fault diagnosis method based on contrastive self-supervised learning. This method ignores the time dependency within the time series and is difficult to fully capture the complex information of the time series data. Moreover, the self-supervised learning task of this method is not specially designed for small sample scenarios, which limits its application effect in small sample time series recognition tasks. Therefore, for small sample multivariate time series data, how to design a time series feature learning and recognition algorithm that does not require a large amount of labeled data based on the data characteristics of the time series itself is still an urgent problem to be solved. Summary of the invention

[0004] In view of the actual situation and data characteristics of small sample time series recognition tasks, the present invention proposes a self-supervised small sample time series recognition model construction method and recognition method, which solves the technical problem that the existing time series recognition methods based on deep learning are difficult to extract effective time series features from small sample time series data, and are prone to overfitting during model training, resulting in poor recognition performance; the present invention focuses on using a large amount of unlabeled time series data under small sample conditions to help the model learn effective time series feature representations, and the model solves the problems of low recognition accuracy and poor generalization caused by unstable training and overfitting through self-supervised contrast mask reconstruction pre-training. Through the method proposed by the present invention, the model can effectively extract time series features and recognize patterns for small sample time series data, while reducing dependence on large-scale labeled time series data and improving the recognition performance of the model.

[0005] The present invention is implemented by the following technical solutions: In the first aspect, the present invention provides a method for constructing a small sample time series recognition model based on self-supervision, comprising the following steps: S1, obtain unlabeled multivariate time series data for self-supervised pre-training; S2, constructing a network model for self-supervised pre-training, the model includes a backbone network for extracting time series features and a mask reconstruction module for reconstructing time series data from masked time series features; S3, construct pseudo-label small sample time series task: Randomly sampled from unlabeled multivariate time series data Samples are initialized to get the support queue; each pseudo-label small sample time series task randomly samples from unlabeled multivariate time series data to get different pseudo-labels query samples; the process of selecting support samples from the support queue is modeled as an entropy regularized optimal transmission problem, and by solving the optimal transmission problem, it is possible to find a corresponding support samples; generate a pseudo-label matching matrix based on the correspondence between pseudo-label query samples and support samples; update the support queue using pseudo-label query samples; S4, based on the self-supervised contrastive mask reconstruction method, the model is pre-trained. The model learns from the contrastive learning of the relationship between samples and the mask reconstruction task of the relationship within the sample obtained from the constructed pseudo-label small sample time series task, by minimizing the contrast loss and the mask reconstruction loss; S5, repeat S3 and S4 to optimize the pre-trained network model on the pseudo-label small sample time series task; transfer the pre-trained network model to the labeled small sample time series recognition task and adjust the pre-training weights to obtain a small sample time series recognition model based on self-supervision.

[0006] Furthermore, the mask reconstruction module consists of a linear layer and a ReLU activation function.

[0007] Furthermore, constructing semantically meaningful small sample tasks satisfies two conditions: first, the support samples and query samples with the same label should be semantically similar, and second, all samples are different.

[0008] Furthermore, when constructing the pseudo-label small sample time series task, S3 also includes mask enhancement and Gaussian noise data enhancement on the pseudo-label query samples to obtain mask query samples and enhanced query samples respectively. Mask enhancement randomly masks time periods with frame lengths of 1 to 5 in different dimensions of the multivariate time series to destroy the temporal dependency of the original time series. Gaussian noise data enhancement adds Gaussian noise with an average intensity of 0.01 to the time series to simulate real-world disturbances.

[0009] Furthermore, using pseudo-label query samples to update the support queue includes: after each pseudo-label small sample task is completed, query the sample Insert support queue The head of the queue and the tail of the queue are removed from the supporting queue samples.

[0010] Furthermore, S4 includes: Contrastive learning of the relationship between samples, dividing the positive and negative sample pairs based on the correspondence between the enhanced query features and the supporting features; in the pre-training process, the meta-contrastive loss is calculated based on the supporting features and the enhanced query features to optimize the model, so that the model learns the correlation and discriminative information between samples; Mask reconstruction of intra-sample relations, sending the mask query features to the mask reconstruction module to obtain reconstructed query samples, and calculating the mask reconstruction loss based on the time series values ​​of the reconstructed query samples and the original query samples to optimize the model, so that the model can learn the global information and time-dependent relations within the samples.

[0011] Furthermore, the contrast loss is:

[0012] in and They are enhanced query features and supporting features respectively. is a pseudo-label matching matrix. and supporting features When the corresponding sample is a positive sample , otherwise it is a negative sample . is a temperature parameter that adjusts the smoothness of the similarity measure. Represents the sample characteristics and Similarity in feature space; The mask reconstruction loss is:

[0013] Right now Each masked element is calculated for the masked position of all query samples The reconstruction estimate of and the true value The squared error.

[0014] In the second aspect, the present invention provides a small sample time series recognition method based on self-supervision, which migrates the pre-trained model to a labeled small sample time series recognition task and adjusts the pre-trained weights to obtain a small sample time series recognition model based on self-supervision, and uses the time series to be identified as the input of the small sample time series recognition model based on self-supervision. The backbone network outputs time series features, and the classifier outputs the predicted small sample time series category.

[0015] The present invention can also provide a computer device, including a processor and a memory, the memory is used to store a computer executable program, the processor reads the computer executable program from the memory and executes it, and when the processor executes the computer executable program, it can implement the above-mentioned method for constructing a small sample time series recognition model based on self-supervision.

[0016] At the same time, a computer-readable storage medium is provided, in which a computer program is stored. When the computer program is executed by a processor, the above-mentioned method for constructing a small sample time series recognition model based on self-supervision can be implemented.

[0017] Compared with the prior art, the present invention has at least the following beneficial technical effects: This paper proposes a contrastive mask reconstruction pre-training method based on self-supervised learning. Through contrastive learning tasks on the constructed pseudo-labeled small sample time series tasks, the model learns the consistency information and discriminative information between samples, and through mask reconstruction tasks, the model is forced to capture the global information and time dependency within the sample. Based on this method, the time series features learned on the unlabeled sequence data of the model have strong discriminative and generalization capabilities, and can be effectively transferred to the small sample time series recognition task to improve the recognition performance; The self-supervised small sample time series recognition model provided by the present invention can be used in multiple practical application scenarios, including but not limited to medical diagnosis, industrial equipment monitoring, financial market forecasting, etc. In these scenarios, the data collection and annotation costs are high, but the self-supervised small sample time series recognition model constructed by the method of the present invention can learn effective time series feature representation from unlabeled data, thereby achieving high-accuracy time series recognition with only a small amount of labeled data, thereby improving the intelligence level and application value of the recognition system. BRIEF DESCRIPTION OF THE DRAWINGS

[0018] Figure 1 It is a flow chart of the present invention.

[0019] Figure 2 It is an algorithm flow chart of the self-supervised contrastive mask reconstruction method. DETAILED DESCRIPTION

[0020] In order to make the purpose, technical solution and advantages of the present invention clearer, the present invention is further described in detail below with reference to the accompanying drawings and specific examples. These examples are merely illustrative and not limiting of the present invention.

[0021] like Figure 1 As shown, the present invention provides a small sample time series recognition method based on self-supervision, comprising: Step 1: Obtain multivariate time series data for self-supervised pre-training.

[0022] Taking the electrocardiogram signal in the medical field as an example, specifically, the present invention selects the public Heartbeat data set as self-supervised pre-training data. The electrocardiogram signal provided by this data set is recorded by placing multiple electrodes (called leads) at different positions on the human body surface, capturing the electrical activity of the heart in a specific time period from different angles, and belongs to a multivariate time series data. The present invention aims to enable the model to learn effective time series feature representation from unlabeled time series data, so the training data of the Heartbeat data set is directly used to pre-train the model, and the training data used in the pre-training process does not require labeling information.

[0023] Step 2: Build a network model for self-supervised pre-training.

[0024] For multivariate time series data of electrocardiogram type, the present invention constructs a deep learning model suitable for multivariate time series analysis. The model includes a time series feature extraction module using transformer as the backbone network. And a mask reconstruction module consisting of a linear layer and a ReLU activation function During the pre-training process, the temporal feature extraction module Used to extract temporal features from unlabeled time series samples, mask reconstruction module It is used to predict the original time series values ​​at the masked positions based on the time series features of the time series data where some inputs are masked (i.e., some values ​​of the time series are set to 0).

[0025] Step 3: Build a pseudo-label small sample time series task.

[0026] To construct a semantically meaningful small-shot task, two conditions must be met: (1) support samples and query samples with the same label should be semantically similar, and (2) all samples should be different. Based on these requirements, the pseudo-label small-shot time series recognition task is constructed in the following way: First, we obtain an unlabeled multivariate time series training set from the Heartbeat dataset. Each pseudo-labeled small sample time series task is randomly sampled from the unlabeled multivariate time series data. Samples , and the The sample assumption is query samples with different pseudo labels. All small sample tasks share the same random sampling from unlabeled multivariate time series data. The support queue is initialized with samples Then, the process of selecting support samples from the support queue is modeled as an entropy regularized optimal transmission problem. By solving this optimal transmission problem, the support samples from the support queue are selected. For each pseudo-label query sample turn up Corresponding supporting samples , and then query samples based on pseudo labels and supporting samples The correspondence between them generates a pseudo label matching matrix ;Finally, use the pseudo-label query sample Update support queue .

[0027] In addition, in order to construct more diverse small sample tasks to help the model learn self-supervised feature representation, the present invention uses a data enhancement method designed for multivariate time series to help search for support samples. This data enhancement method is used for pseudo-label query samples. Use mask enhancement and Gaussian noise data enhancement to obtain mask query samples and enhanced query samples , and further expand the diversity of training data under small sample conditions. Mask enhancement randomly masks time periods with frame lengths of 1 to 5 in different dimensions of multivariate time series to destroy the temporal dependence of the original time series, and Gaussian noise data enhancement adds Gaussian noise with an average intensity of 0.01 to the time series to simulate real-world disturbances.

[0028] For the search process of support samples, specifically, first query the sample for the mask and support queues Using the transformer backbone network Extract the corresponding features and perform normalization to obtain the mask query features ( ) and support queue features ( ) ,in represents the L2 normalization process; then the selection process of support samples is modeled as an entropy regularized optimal transmission problem:

[0029] in Indicates that the value range is 0~1 from Mask query features To Support queue features The transmission matrix, represents the regularization coefficient, = Represents the transmission matrix The Sinkhorn-Knopp algorithm is used to quickly obtain the approximate optimal solution to the entropy regularized optimal transmission problem. *, and can be considered as the current mask query feature and support queue features Therefore, the transmission matrix can be selected * before each line Elements to support queue features The corresponding supporting features are obtained in , that is, each mask query feature has a corresponding Support features. From the supporting sample The corresponding features extracted from .

[0030] For the generation process of the pseudo-label matching matrix, the pseudo-label matching matrix Query features by mask and supporting features The corresponding relationship between them is obtained. Query features Corresponding Support features include Supported features When (i.e. query feature and supporting features When the corresponding sample is a positive sample) On the contrary, when Query features Corresponding Support features do not include Supported features When (i.e. query feature and supporting features When the corresponding sample is a negative sample) .

[0031] For updates to the support queue, after each small sample build task is completed, the sample will be queried Insert support queue The head of the queue is removed from the queue at the same time. samples to dynamically update the support queue to increase the diversity of small sample tasks.

[0032] Step 4: Self-supervised contrast mask reconstruction of pre-trained model.

[0033] In the medical field, it is difficult to obtain a large amount of labeled data due to the high cost of labeling, patient privacy and data confidentiality. It is often difficult to effectively extract features by directly training deep learning models, and overfitting is prone to occur during the training process, resulting in poor recognition performance. In order to solve the above problems in this task, the present invention proposes a self-supervised contrast mask reconstruction pre-training method, see Figure 2 The algorithm flow chart in , in the constructed pseudo-label small sample time series task, the positive and negative samples are distinguished by contrastive learning, which prompts the model to learn the consistency and discriminative information between samples; through the mask reconstruction task, the model's understanding of the global information and time dependency of the samples is enhanced. The combination of the two enables the model to learn effective time series feature representations with generalization in unlabeled multivariate time series data, thereby achieving better recognition performance in small sample time series tasks.

[0034] For the contrastive learning task of the relationship between samples, the goal of contrastive learning is to encourage the model to distinguish between positive and negative samples by maximizing the similarity between positive sample pairs and minimizing the similarity between negative sample pairs, thereby learning discriminative feature representations. Different from the common contrastive learning method, in this example, the anchor sample is Enhanced query samples after adding Gaussian noise Send to transformer backbone network The enhanced query features obtained ( ) ,represent The corresponding categories of pseudo labels are Indicates L2 normalization. In support of features In the example, each anchor sample is selected from the support queue The most similar support features are used as their corresponding positive samples. Relatively speaking, similar samples selected from the support queue by other anchor samples are their negative samples. Then, for positive and negative samples, the contrast loss is calculated:

[0035] in and They are enhanced query features and supporting features respectively. is a pseudo-label matching matrix. and supporting features When the corresponding sample is a positive sample , otherwise it is a negative sample . is a temperature parameter that adjusts the smoothness of the similarity measure. Represents the sample characteristics and Similarity in feature space.

[0036] For the mask reconstruction task of in-sample relations, in each input query sample middle, w represents the length in the time dimension, h The number of variables representing the multivariate time series. As part of the self-supervised pre-training, the mask reconstruction task generates a Generate binary mask , and then for the query sample Perform masking operation so that part of the sample input is masked (part of the input value is set to 0) to obtain the corresponding mask query sample ;in Represents the dot multiplication operation. During the mask reconstruction process, the transformer backbone network of the model Calculate the mask query feature ( ) ,in represents L2 normalization. Then, for the query sample The binary mask The mask position set to 0, the mask reconstruction module of the model Query features based on masks The context information contained in the complete query sample The reconstruction estimate of , thus predicting the mask position in the original query sample By reconstructing the masked time series fragments, the model can further learn to extract the global information and time dependency relationships within the time series.

[0037] Defines the index set of the masked elements ,in represents the time dimension, Represents the vector dimension. When the mask When querying samples The elements at the corresponding positions in will be reconstructed and considered in the loss function, and the number of reconstructed elements is .

[0038] Specifically, for the mask-enhanced query sample Corresponding features ( ) , the network model is reconstructed through a mask module Features Perform linear transformation and ReLU activation function activation to output the complete query sample Estimates Then in the small sample task Query samples and the corresponding reconstructed query samples The mask reconstruction loss is calculated element-wise at the masked position:

[0039] Right now Each masked element is calculated for the masked position of all query samples The reconstruction estimate of and the true value The squared error.

[0040] By minimizing the contrast loss and mask reconstruction loss , the pre-trained network model can be optimized on the pseudo-label small sample time series task, enhancing the feature extraction ability and generalization of the network model for electrocardiogram type time series data.

[0041] Step 5: Verify the effectiveness of the model on the small sample time series recognition task The temporal feature representation learned by the model through self-supervised pre-training can be effectively transferred to the small-sample time series recognition task. By fine-tuning the pre-trained weights of the model on the labeled small-sample time series recognition task, the model can achieve excellent recognition performance.

[0042] Specifically, first, a linear classifier is added to the last layer of the pre-trained network model to form a recognition model for ECG time series data. Then, the model is fine-tuned using the labeled training data in the Heartbeat dataset to obtain the final ECG time series recognition model. Finally, the test sample of the Heartbeat dataset is sent to the ECG time series recognition model, the transformer backbone network outputs the time series features, and then the classifier outputs the predicted ECG recognition category. The recognition accuracy of the model is obtained by comparing the recognition category of the test sample with the labeled category of the sample.

[0043] Specifically, the present invention verifies the model validity of the recognition performance of the proposed method through the accuracy index, and compares the proposed method with a variety of existing time series recognition algorithms in the case of small samples on the Heartbeat time series dataset of the medical electrocardiogram type. The selected comparison benchmark methods include the classic machine learning model XGBoost, the deep learning model LSTM, and the transformer model that has not been pre-trained by self-supervised contrast mask reconstruction. The experimental results in Table 1 fully verify the effectiveness of the self-supervised small sample time series recognition method of the present invention. Compared with the classic time series recognition method, the self-supervised small sample time series recognition method enables the model to extract more effective and robust time series features from the electrocardiogram small sample time series data to achieve better recognition accuracy.

[0044] Table 1 Accuracy of time series recognition methods on the Heartbeat test set

[0045] In addition, the present invention can also provide a computer device, including a processor and a memory, the memory is used to store a computer executable program, the processor reads part or all of the computer executable program from the memory and executes it, and when the processor executes part or all of the computer executable program, it can implement the method for constructing a small sample time series recognition model based on self-supervision described in the present invention, and can also execute the small sample time series recognition method based on self-supervision.

[0046] On the other hand, the present invention provides a computer-readable storage medium, which stores a computer program. When the computer program is executed by a processor, it can implement the method for constructing a small sample time series recognition model based on self-supervision described in the present invention, and can also execute the small sample time series recognition method based on self-supervision.

[0047] The computer device may be a laptop computer, a desktop computer or a workstation.

[0048] The processor of the present invention may be a central processing unit (CPU), a digital signal processor (DSP), an application specific integrated circuit (ASIC) or a readily available field programmable gate array (FPGA).

[0049] The memory of the present invention may be an internal storage unit of a laptop computer, a desktop computer or a workstation, such as a memory or a hard disk; or an external storage unit, such as a mobile hard disk or a flash memory card.

[0050] Computer-readable storage media may include computer storage media and communication media. Computer storage media include volatile and non-volatile, removable and non-removable media implemented by any method or technology for storing information such as computer-readable instructions, data structures, program modules or other data. Computer-readable storage media may include: read-only memory (ROM), random access memory (RAM), solid-state drive (SSD) or optical disk, etc. Among them, random access memory may include resistance random access memory (ReRAM) and dynamic random access memory (DRAM).

[0051] Although the present invention has been described in detail with general description and specific embodiments, it is obvious to those skilled in the art that some modifications or improvements may be made thereto based on the present invention. Therefore, these modifications or improvements made on the basis of not departing from the spirit of the present invention all belong to the scope of protection claimed in the present invention.

Claims

1. A method for constructing a small sample time series recognition model based on self-supervision, characterized in that: The following steps are involved: S1, obtain unlabeled multivariate time series data for self-supervised pre-training; S2, constructing a network model for self-supervised pre-training, the model includes a backbone network for extracting time series features and a mask reconstruction module for reconstructing time series data from masked time series features; S3, construct pseudo-label small sample time series task: Randomly sampled from unlabeled multivariate time series data Samples are initialized to get the support queue; each pseudo-label small sample time series task randomly samples from unlabeled multivariate time series data to get different pseudo-labels query samples; the process of selecting support samples from the support queue is modeled as an entropy regularized optimal transmission problem, and by solving the optimal transmission problem, it is possible to find a corresponding support samples; generate a pseudo-label matching matrix based on the correspondence between pseudo-label query samples and support samples; update the support queue using pseudo-label query samples; S4, based on the self-supervised contrastive mask reconstruction method, the model is pre-trained. The model learns from the contrastive learning of the relationship between samples and the mask reconstruction task of the relationship within the sample obtained from the constructed pseudo-label small sample time series task, by minimizing the contrast loss and the mask reconstruction loss; S5, repeat S3 and S4 to optimize the pre-trained network model on the pseudo-label small sample time series task; transfer the pre-trained network model to the labeled small sample time series recognition task and adjust the pre-training weights to obtain a small sample time series recognition model based on self-supervision.

2. The method for constructing a small sample time series recognition model based on self-supervision according to claim 1, characterized in that: The mask reconstruction module consists of a linear layer and a ReLU activation function.

3. The method for constructing a small sample time series recognition model based on self-supervision according to claim 1, characterized in that: Constructing semantically meaningful small sample tasks satisfies two conditions: first, support samples and query samples with the same label should be semantically similar, and second, all samples are different.

4. The method for constructing a small sample time series recognition model based on self-supervision according to claim 1, characterized in that: When constructing the pseudo-label small sample time series task, S3 also includes mask enhancement and Gaussian noise data enhancement on the pseudo-label query samples to obtain mask query samples and enhanced query samples respectively. Mask enhancement randomly masks time periods with frame lengths of 1 to 5 in different dimensions of the multivariate time series to destroy the temporal dependency of the original time series. Gaussian noise data enhancement adds Gaussian noise with an average intensity of 0.01 to the time series to simulate real-world disturbances.

5. The method for constructing a small sample time series recognition model based on self-supervision according to claim 1, characterized in that: Using pseudo-label query samples to update the support queue includes: each time the pseudo-label small sample task is completed, the query sample Insert support queue The head of the queue and the tail of the queue are removed from the supporting queue samples.

6. The method for constructing a small sample time series recognition model based on self-supervision according to claim 1, characterized in that S4 include: Contrastive learning of the relationship between samples, dividing the positive and negative sample pairs based on the correspondence between the enhanced query features and the supporting features; in the pre-training process, the meta-contrastive loss is calculated based on the supporting features and the enhanced query features to optimize the model, so that the model learns the correlation and discriminative information between samples; Mask reconstruction of intra-sample relations, sending the mask query features to the mask reconstruction module to obtain reconstructed query samples, and calculating the mask reconstruction loss based on the time series values ​​of the reconstructed query samples and the original query samples to optimize the model, so that the model can learn the global information and time-dependent relations within the samples.

7. The method for constructing a small sample time series recognition model based on self-supervision according to claim 1, characterized in that: The contrast loss is: in and They are enhanced query features and support features. is a pseudo-label matching matrix. and supporting features When the corresponding sample is a positive sample , otherwise it is a negative sample , is the temperature parameter used to adjust the smoothness of the similarity measure, Represents the sample characteristics and Similarity in feature space; The mask reconstruction loss is: Right now Each masked element is calculated for the masked position of all query samples The reconstruction estimate of and the true value The squared error.

8. A small sample time series recognition method based on self-supervision, characterized in that: The pre-trained model is transferred to the labeled small sample time series recognition task and the pre-trained weights are adjusted to obtain a small sample time series recognition model based on self-supervision. The time series to be identified is used as the input of the small sample time series recognition model based on self-supervision. The backbone network outputs time series features, and the classifier outputs the predicted small sample time series category.

9. A computer device, characterized in that: It includes a processor and a memory, the memory is used to store a computer executable program, the processor reads the computer executable program from the memory and executes it, and when the processor executes the computer executable program, it can implement the method for constructing a small sample time series recognition model based on self-supervision as described in any one of claims 1 to 7 or the method for recognizing a small sample time series based on self-supervision as described in claim 8.

10. A computer-readable storage medium, characterized in that: A computer program is stored in a computer-readable storage medium. When the computer program is executed by a processor, it can implement the method for constructing a small sample time series recognition model based on self-supervision as described in any one of claims 1 to 7 or the method for recognizing a small sample time series based on self-supervision as described in claim 8.

Images