Low-quality time series data representation learning method based on patch alignment auto-encoder

Abstract

The invention discloses a low-quality time series data representation learning method based on a patch alignment auto-encoder, and the method comprises the steps: generating low-quality time series data with Gaussian noise based on original low-quality time series data; constructing a low-quality time series data representation learning model; training a model based on the low-quality time sequence data; and calculating the difference between the model reconstruction time sequence and the original time sequence, inputting to-be-processed time sequence data into the trained representation learning model, and selecting a proper output module according to a specific downstream task to obtain a corresponding downstream task result. The method is used for extracting high-quality data representation of low-quality time series data, time series data with missing values and noise can be processed at the same time, the reliability of data analysis is improved, and high-fidelity data representation is provided for downstream tasks.

Description

Technical Field

[0001] The present invention relates to a low-quality time series data representation learning method based on a patch alignment autoencoder, and belongs to the technical field of time series data analysis. Background Art

[0002] In many practical applications, such as meteorological monitoring, healthcare, and urban transportation, time series data can be subject to interference from equipment failures, environmental disturbances, or human factors, leading to quality issues such as missing information and noise contamination. For example, while meteorological data should theoretically be collected daily, actual data distribution shows significant sampling interval inequality not only between data points within a single time series, but also between different time series, with significant variations in sampling frequency. In addition to high missing rates, low-quality data may be accompanied by noise that obscures true climate patterns, causing the data to deviate from the true value. The presence of low-quality data significantly impacts the reliability of data analysis, and high missing rates can lead to the loss of critical information. Furthermore, the accompanying noise in the data further distorts the underlying data distribution. Therefore, designing robust representation learning methods is crucial.

[0003] Traditional time series representation methods, such as piecewise aggregate approximation, symbolic aggregate approximation, Shapelet extraction, and discrete wavelet transform, often require deep domain background knowledge as support, and the entire processing process is time-consuming and inefficient. Therefore, some studies have integrated neural differential equations into existing networks to achieve continuous expression of the model's hidden state. Contiformer combines the advantages of neural ordinary differential equations in continuous dynamic modeling with the attention mechanism of Transformer to effectively capture the complex relationships and dynamic changes in the data. However, such neural differential equation-based methods generally face the problem of high computational cost in the solution process. In addition, when there is noise in the data, such methods have limited ability to reduce the impact of noise on model performance.

[0004] In contrast, deep learning technology has shown great advantages in time series representation learning, and can automatically identify and extract hidden patterns and features in time series data. TAKE proposes an RNN-based autoencoder architecture that can effectively fuse the features of different variables and time steps in multivariate time series when data is missing, thereby generating a compressed representation of the multivariate time series. mTAN designs a multi-temporal attention mechanism that can assign different attention weights to data points in a time series at different time scales and intervals, thereby highlighting data information that is more critical to the overall time series characteristics and task objectives, and solving the problem of difficulty in focusing on important information due to missing data. Although the above methods have alleviated the negative impact of low-quality data on model performance to a certain extent, they still fail to effectively handle the situation where missing values ​​and noise exist in the data at the same time. Summary of the Invention

[0005] Purpose of the invention: In view of the problems and shortcomings of the existing technology, the present invention provides a low-quality time series data representation learning method based on patch alignment autoencoder with good feature extraction effect and good practicality.

[0006] Technical Solution: A method for learning low-quality time series data representation based on patch alignment autoencoders, including the following steps:

[0007] Step (1): Based on the original low-quality time series data, generate low-quality time series data with Gaussian noise, referred to as low-quality time series data;

[0008] Step (2): Construct a low-quality time series data representation learning model based on patch alignment autoencoder;

[0009] Step (3): Input low-quality time series data and train the low-quality time series data representation learning model;

[0010] Step (4): Calculate the difference between the time series reconstructed by the learning model representing low-quality time series data and the original time series. If the difference is lower than the preset threshold, execute step (5); otherwise, return to step (3).

[0011] Step (5): Input the time series data to be processed into the trained low-quality time series data representation learning model, select the appropriate output module according to the specific downstream task, and obtain the corresponding downstream task results.

[0012] Preferably, the step (1) is specifically as follows:

[0013] The input low-quality time series can be represented as Where N represents the total number of variables, and the nth variable contains L n observations, the i-th observation is represented by the timestamp and the corresponding values composition.

[0014] More preferably, the original low-quality time series data is climate data, medical data, human activity data, server data, satellite data, or equipment operation data. In practical applications, problems such as sensor failure and communication link interruptions can cause significant loss of some collected data, significantly reducing data quality. Although unified data quality standards have not yet been established in the field of time series analysis, extensive experimental research has shown that when the missing rate of time series data exceeds 50% at the same sampling frequency, the processing and analysis performance of various models will be significantly reduced. Based on this empirical evidence, time series data with a missing rate exceeding 50% at the same sampling frequency standard can be defined as low-quality data.

[0015] More preferably, noise injection is used to randomly inject Gaussian noise into each low-quality input time series to simulate data pollution in the real world and generate low-quality time series data with Gaussian noise.

[0016] Preferably, the step (2) is specifically as follows:

[0017] Based on the patch alignment autoencoder, a low-quality time series data representation learning model is constructed, consisting of a patch alignment module, an encoder, and a decoder. The patch alignment module mainly consists of a transformable patch technology and a temporal encoding mechanism. The transformable patch technology is used to segment the noisy time series, and the temporal encoding mechanism is used to preserve the temporal information of the time series. The encoder mainly consists of an adaptively weighted TCN module and a Transformer architecture. The adaptively weighted TCN module is used to capture the local periodic characteristics of each patch, and the Transformer architecture is used to capture long-range dependencies. The decoder is a fully connected layer that converts the feature vector in the latent space back into time series data.

[0018] More preferably, for the transformable patch technology: the input of the technology is the low-quality time series of each unit after the noise is introduced, the low-quality time series is segmented into multiple patches, each patch has a uniform time interval, and then the segmented patches are output.

[0019] More preferably, for the time encoding mechanism: the input of the mechanism is the segmented patches output by the transformable patch technology, the observation data timestamp in each patch is encoded in a time embedding manner, and then the patch embedding obtained by splicing the time embedding and the patch is output; the formula of the time embedding encoding process is as follows:

[0020]

[0021] Among them, t represents the timestamp, d prepresents the embedding dimension of the p-th patch, D t is the total dimension of the embedding, i.e. d p The maximum value of d p = 0, the linear term ω 0p t+α 0p Used to capture the aperiodic pattern of the p-th patch evolving over time, ω 0p and α 0p are two learnable parameters; when 0 <d p <D t When ω dp and α dp Represent the frequency and phase of the sine function, respectively, and are used to capture the periodic characteristics of the time series.

[0022] More preferably, the patch alignment module is composed of a transformable patch technology and a time coding mechanism, and the input of the patch alignment module is the low-quality time series of each unit after the introduction of noise; the transformable patch technology is used to segment the noisy time series, and the time coding mechanism is used to retain the time information of the sequence.

[0023] More preferably, for the adaptive weighted TCN module: the input of this module is the patch embedding output by the temporal encoding mechanism, the temporal features of the input are adaptively input using meta-filter parameters, TCN is applied for convolution operation, and then the potential embedding of each patch is output.

[0024] More preferably, for the Transformer architecture: the input of the architecture is the latent embedding of the patch output by the adaptive weighted TCN module, the latent representation is combined with the positional encoding, and the output is obtained using the multi-head attention mechanism, and then the latent representation of each patch is output.

[0025] More preferably, the encoder consists of an adaptive weighted TCN module and a Transformer architecture, and the input of the encoder is the patch embedding output by the temporal coding mechanism; the adaptive weighted TCN module is used to capture the local periodic features of each patch, and the Transformer architecture is used to capture long-distance dependencies.

[0026] More preferably, for the decoder: the decoder is a fully connected layer, the input of the decoder is the potential representation of the patch output by the Transformer architecture, the feature vector in the latent space is converted back to time series data through a linear transformation, and then the ReLU activation function is applied to process it, and finally the reconstructed time series is output.

[0027] More preferably, the step (3) is specifically as follows:

[0028] The low-quality time series data representation learning model is trained, and the back propagation algorithm is used to optimize the parameters according to the results of the loss function during training.

[0029] More preferably, the loss function is specifically:

[0030]

[0031] Where N represents the total number of variables, L n Indicates the length of the sequence, represents the i-th observation value of the n-th variable in the original time series X, Represents the reconstructed time series X ′ The i-th observation of the n-th variable in .

[0032] A computer device includes a memory, a processor, and a computer program stored in the memory and executable on the processor. When the processor executes the computer program, the low-quality time series data representation learning method based on patch alignment autoencoder as described above is implemented.

[0033] A computer-readable storage medium stores a computer program for executing the low-quality temporal data representation learning method based on patch alignment autoencoder as described above.

[0034] Compared with the prior art, the present invention has the following beneficial effects:

[0035] 1. High Robustness: The proposed method for learning low-quality time series data representations based on patch-aligned autoencoders achieves natural alignment of time series using transformable patching technology, effectively overcoming the feature shift caused by uneven sampling intervals in traditional patching techniques. Combined with a random Gaussian noise injection mechanism to simulate real-world data contamination scenarios, the model maintains stable feature extraction performance despite high missing rates and noise interference. Compared to traditional methods, it achieves lower MSE metrics in prediction tasks and higher F1 scores for anomaly detection, demonstrating its practical application value.

[0036] Second, strong multi-task generalization: Our patch-aligned autoencoder-based representation learning method for low-quality time series data is based on a unified representation learning framework for denoising autoencoders. The resulting high-fidelity latent features seamlessly support downstream tasks such as prediction, classification, and anomaly detection. Experiments on a variety of downstream tasks demonstrate superior performance compared to traditional methods in metrics such as prediction accuracy, anomaly detection F1 score, and classification accuracy, demonstrating the model's cross-domain applicability. BRIEF DESCRIPTION OF THE DRAWINGS

[0037] Figure 1This is a framework diagram of a low-quality temporal data representation learning method (PAMDA) based on patch alignment autoencoder in an embodiment of the present invention.

[0038] Figure 2 This is the specific structure of the encoder in the low-quality time series data representation learning method based on patch alignment autoencoder in an embodiment of the present invention.

[0039] Figure 3 This figure shows the experimental results of the low-quality time series data representation learning method based on patch alignment autoencoder in anomaly detection tasks in an embodiment of the present invention. DETAILED DESCRIPTION

[0040] The present invention is further illustrated below with reference to specific examples. It should be understood that these examples are only used to illustrate the present invention and are not used to limit the scope of the present invention. After reading the present invention, modifications of various equivalent forms of the present invention made by those skilled in the art all fall within the scope defined by the claims attached to this application.

[0041] A method for learning low-quality time series data representation based on patch alignment autoencoders, including:

[0042] Step (1): Generate low-quality time series data with Gaussian noise based on the original low-quality time series data;

[0043] The input low-quality time series can be represented as Where N represents the total number of variables, and the nth variable contains L n observations, the i-th observation is represented by the timestamp and the corresponding values The original low-quality time series data includes climate data, medical data, human activity data, server data, satellite data, or equipment operation data. Specifically, climate data includes snowfall, snowfall depth, precipitation, maximum temperature, and minimum temperature. Medical data includes clinical data. Human activity data includes data on different human activities collected by sensors located on the left ankle, right ankle, waist, and chest. Server data includes CPU, memory, disk, network, and other data. Satellite data includes global soil moisture, large-area vegetation coverage, and other data. Equipment operation data includes data collected by different devices in the same facility.

[0044] For the input low-quality time series X, set the noise rate v (v = 0.1) to calculate the number of elements m to which noise needs to be added. Randomly select the position index of m elements, generate a Gaussian noise value ∈ with a mean of μ (μ = 0) and a standard deviation of σ (σ = 0) for each selected position, and add it to the original data set to achieve random perturbation of the data. The generation of Gaussian noise can be achieved using formula (3):

[0045]

[0046] Among them, ∈ i Represents the noise added to the i-th element. The dataset after the noise is introduced can be expressed as

[0047] Step (2): Construct a low-quality time series data representation learning model based on patch alignment autoencoder;

[0048] The specific structure of the model is as follows Figure 1 The model mainly consists of three parts: patch alignment module, encoder and decoder. The patch alignment module is composed of transformable patch technology and temporal coding mechanism, the encoder is composed of adaptive weighted TCN module and Transformer architecture, and the decoder is a fully connected layer.

[0049] The input of the transformable patch technique is the low-quality time series of each unit after the introduction of noise. The low-quality time series of each unit after the noise is introduced Split into P subsequence patches Where l1=1,r1=L,l p =l p-1 +L,r p =r p-1 +L, L represents the total length of each time series, and then output the segmented patch

[0050] The input of the temporal encoding mechanism is the segmented patch output by the transformable patching technique A time embedding method is introduced to encode the timestamps of the observation data in each patch. The representation of each timestamp can be composed of one non-periodic dimension and multiple periodic dimensions. The encoding process formula is as follows:

[0051]

[0052] Among them, t represents the timestamp, d p represents the embedding dimension of the p-th patch, D t is the total dimension of the embedding, i.e. d p The maximum value is set to 10. p = 0, the linear term ω 0p t+α 0p Used to capture the aperiodic pattern of the p-th patch evolving over time, ω 0p and α 0p are two learnable parameters; when 0 <d p <D t When ω dp and αdp Represent the frequency and phase of the sine function, respectively, and are used to capture the periodic characteristics of the time series.

[0053] Combined with the learned temporal embedding representation, the observation value in each patch can be expressed as follows:

[0054]

[0055] Among them, || represents the splicing operation. Then the output time embedding is spliced ​​with the patch embedding obtained by patching.

[0056] The patch alignment module is composed of a transformable patch technique and a time encoding mechanism. The input of the patch alignment module is the low-quality time series of each unit after the introduction of noise. Can get patch embed

[0057] The input of the adaptive weighted TCN module is the patch embedding output by the temporal encoding mechanism The convolution kernel that matches the input feature dimension is dynamically generated through the meta-filter. The calculation formula of the filter is as follows:

[0058]

[0059] Among them, L p Represents a transformable patch The sequence length, is the filter derived from the d-th feature map, D in Indicates the dimension of the input features. d Represents a meta-filter that can be instantiated by a learnable neural network, through which filter parameters are independently derived for each feature map, ultimately forming D-1 filters.

[0060] Specifically, the meta-filter first represents the input Perform global average pooling to compress the spatial dimension. Then, through two layers of fully connected layers and ReLU activation function, extract the nonlinear features. The last layer of fully connected layer is combined with Softmax function to generate a filter f that can adaptively match the input. d .

[0061] Enter Apply TCN to perform convolution operations with D-1 filters respectively, and finally get the potential embedding of each patch The latent embedding is calculated by formula (6).

[0062]

[0063] Considering that in the case of sparse time series or high temporal resolution of patch segmentation, some patches may have no observations, so this paper incorporates the patch mask into the potential embedding of each patch block. The calculation process is as follows:

[0064]

[0065] Among them, when there are observations in the transformable patch, m p = 1; when there is no observation in the transformable patch, m p = 0. Combining the embedded representations of P patches, the potential representation output by the adaptive weighted TCN module can be defined as h 1:P = Then output the latent embedding of each patch

[0066] The input to the Transformer architecture is the latent embedding h of the patch output by the adaptively weighted TCN module 1:P First, the potential representation h 1:P With positional encoding Combined to preserve the location information of the patch block. The variable n is used in the fusion potential representation and position encoding PE 1:P The input after is expressed as Then, the multi-head attention mechanism is used to convert the input representation into the query matrix Bond Matrix Sum Matrix Transformer output obtained using scaled dot product attention The calculation formula is as follows:

[0067]

[0068] in, is a learnable weight matrix, H is the number of attention heads, which is set to 1. Combining the output representation of the time series, the final output form of the Transformer can be expressed as The latent representation for all variables in the same patch can be identified as Then output the latent representation of each patch

[0069] The specific structure of the encoder is as follows Figure 2 The encoder consists of an adaptive weighted TCN module and a Transformer architecture. The input of the encoder is the patch embedding output by the temporal encoding mechanism. The potential representation of the patch can be obtained

[0070] The input to the decoder is the latent representation of the patch output by the Transformer architecture For each transformable patch, the latent representation First, apply the projection layer to convert its dimensions. The definition of the projection layer is as follows:

[0071]

[0072] Among them, W p and b p are learnable parameters, representing the weight matrix and bias vector of the projection layer, is the reconstructed sequence of the pth patch. To ensure that the reconstructed patch It can conform to the distribution of the original data and apply the ReLU activation function after the projection layer to obtain the reconstructed output of the patch Splice the reconstructed outputs of all patches of N variables to get the final reconstructed time series

[0073] Step (3): Input the processed data in step (1) and train the low-quality time series data representation learning model constructed in step (2);

[0074] For the constructed model, a training set, validation set, and test set were constructed. The model was initialized and trained using the training set data. The model parameters were adjusted using the validation set and evaluation metrics. The model performance was tested using the test set. It is worth mentioning that to prevent overfitting, an early stopping mechanism was introduced. When the loss on the validation set did not decrease significantly within 10 consecutive epochs, the training process would automatically terminate. Each experiment was repeated five times, each using a different random seed, and the average of all experimental results was reported to mitigate random fluctuations in the results.

[0075] In this embodiment, unless otherwise specified, the proportions of the training dataset, validation dataset, and test dataset are 70%, 10%, and 20%, respectively.

[0076] Use the loss function to train and optimize the parameters to complete the model training and finally output the reconstructed time series X ′ .

[0077] The loss function is specifically:

[0078]

[0079] Where N represents the total number of variables, L n Indicates the length of the sequence, represents the i-th observation value of the n-th variable in the original time series X, Represents the reconstructed time series X′ The i-th observation of the n-th variable in .

[0080] Step (4): Calculate the difference between the model reconstructed time series and the original time series. If the difference is lower than the preset threshold, execute step (5); otherwise, return to step (3).

[0081] Step (5): Input the time series data to be processed into the trained representation learning model, select the appropriate output module according to the specific downstream task, and obtain the corresponding downstream task results.

[0082] Obviously, those skilled in the art should understand that the various steps of the low-quality time series data representation learning method based on patch alignment autoencoder of the above-mentioned embodiment of the present invention can be implemented by a general-purpose computing device. They can be concentrated on a single computing device or distributed on a network composed of multiple computing devices. Optionally, they can be implemented with program codes executable by the computing device, so that they can be stored in a storage device and executed by the computing device. In some cases, the steps shown or described can be performed in an order different from that shown here, or they can be made into individual integrated circuit modules, or multiple modules or steps therein can be made into a single integrated circuit module for implementation. Thus, the embodiments of the present invention are not limited to any specific combination of hardware and software.

[0083] In order to verify the effectiveness and superiority of the low-quality time series data representation learning method based on patch alignment autoencoder proposed in this invention in time series data analysis, experiments will be conducted on the mainstream tasks (prediction, classification and anomaly detection) in the three major time series data analysis fields.

[0084] (1) Prediction task

[0085] The benchmark datasets and related configuration item settings used in the prediction task experiments are shown in Table 1:

[0086] Table 1 Benchmark datasets used in prediction task experiments

[0087]

[0088]

[0089] The comparison methods used in the prediction task experiments include the most cutting-edge Transformer-based models (iTransformer, PatchTST, Crossformer), graph neural network-based models (t-PatchGNN, GraphWavenet), multi-layer perceptron-based models (DLinear), neural differential equation-based models (NeuralFlows, CRU), time series general models (TimesNet), and large language model-based models (Time-LLM).

[0090] The metrics used in the prediction task experiments are mean square error (MSE) and mean absolute error (MAE). In the experimental results table, bold indicates the best, and underlined indicates the suboptimal.

[0091] In the prediction task experiment, the benchmark data sets used include USHCN, PhysioNet and Human Activity, which represent the application requirements of three real-world scenarios: meteorology, medicine, and human activities. Each data set contains a continuous time series data, and obtains sample data through a sliding window. It should be noted that USHCN uses the climate data of the previous 24 months to predict the climate conditions of the next month, PhysioNet uses the data of the previous 24 hours as observations to predict the clinical data within the next 24 hours, and Human Activity uses the data of the previous 3000 milliseconds as observations to predict the data of the next 1000 milliseconds. The experimental results are shown in Table 2. The method of the present invention achieved the best results in the prediction task experiment, which is higher than other comparison methods.

[0092] Table 2 Experimental results of prediction tasks

[0093]

[0094] (2) Classification task

[0095] The benchmark datasets and related configuration item settings used in the classification task experiments are shown in Table 3:

[0096] Table 3 Benchmark datasets used in classification task experiments

[0097]

[0098] The comparison methods used in the classification task experiments include the most cutting-edge convolutional neural network-based model (ROCKET), the Transformer-based model (TST), the Shapelet-based model (CSL), and the general time series model (TimesNet).

[0099] The metric used in the classification task experiment is accuracy (ACC). In the experimental results table, bold indicates the best, and underlined indicates the suboptimal.

[0100] The classification task experiments used benchmark datasets including Ethanol Concentration, Face Detection, Handwriting Heartbeat, Japanese Vowels, PEMS-SF, Self-Regulation SCP1, Self-Regulation SCP2, Spoken Arabic Digits, and UWave Gesture Library. These datasets are derived from real-world scenarios such as human motion recognition, electrocardiogram monitoring, and audio. The experimental results are shown in Table 4. Our method achieved third-place average accuracy in the classification task experiments and ranked first or second across all five datasets, demonstrating significant performance.

[0101] Table 4 Experimental results of classification tasks

[0102]

[0103]

[0104] (3) Anomaly detection task

[0105] The benchmark datasets and related configuration item settings used in the anomaly detection task experiment are shown in Table 5:

[0106] Table 5 Benchmark datasets used in anomaly detection task experiments

[0107]

[0108] The comparison methods used in the anomaly detection task experiment include the most cutting-edge OmniAnomaly (Omni), Reformer (Re), Anomaly Transformer (AT), LSSL, DLinear, TimesNet and iTransformer.

[0109] The F1 score is the metric used in the anomaly detection task experiment. In the experimental results table, bold indicates the best performance, while underlined indicates suboptimal performance.

[0110] In the anomaly detection task experiment, the benchmark datasets used include SMD, MSL, SMAP, SWaT and PSM, which come from various real-world scenarios. The F1 score in the experimental results is as follows: Figure 3 As shown in Table 6, the proposed method achieved the best results in the anomaly detection task experiments. The specific experimental results are shown in Table 6. The proposed method achieved the best F1 score in the anomaly detection task experiments.

[0111] Table 6 Experimental results of anomaly detection task

[0112]

Claims

1. A low-quality time series data representation learning method based on patch alignment autoencoder, characterized by: The steps include: Step (1): Generate low-quality time series data with Gaussian noise based on the original low-quality time series data; Step (2): Construct a low-quality time series data representation learning model based on patch alignment autoencoder; Step (3): Input low-quality time series data and train the low-quality time series data representation learning model; Step (4): Calculate the difference between the time series reconstructed by the learning model representing low-quality time series data and the original time series. If the difference is lower than the preset threshold, execute step (5); otherwise, return to step (3). Step (5): Input the time series data to be processed into the trained low-quality time series data representation learning model, select the output module according to the specific downstream task, and obtain the corresponding downstream task results.

2. The low-quality time series data representation learning method based on patch alignment autoencoder according to claim 1 is characterized in that The step (1) is specifically as follows: The input low-quality time series can be represented as Where N represents the total number of variables, and the nth variable contains L n observations, the i-th observation is represented by the timestamp and the corresponding values composition; The original low-quality time series data include climate data, medical data, human activity data, server data, satellite data, or equipment operation data; Noise injection is used to randomly inject Gaussian noise into each low-quality time series input to simulate data pollution in the real world and generate low-quality time series data with Gaussian noise.

3. The low-quality time series data representation learning method based on patch alignment autoencoder according to claim 1 is characterized in that The step (2) is specifically as follows: Based on the patch alignment autoencoder, a low-quality time series data representation learning model is constructed, including a patch alignment module, an encoder and a decoder; the patch alignment module includes a transformable patch technology and a time encoding mechanism. The transformable patch technology is used to segment the noisy time series, and the time encoding mechanism is used to retain the time information of the time series; the encoder includes an adaptive weighted TCN module and a Transformer architecture. The adaptive weighted TCN module is used to capture the local periodic features of each patch, and the Transformer architecture is used to capture long-distance dependencies; the decoder is a fully connected layer used to convert the feature vector in the latent space back to time series data.

4. The low-quality time series data representation learning method based on patch alignment autoencoder according to claim 3 is characterized in that For the transformable patch technique: the input of this technique is the low-quality time series of each unit after the introduction of noise, the low-quality time series is segmented into multiple patches, where each patch has a uniform time interval, and then the segmented patches are output.

5. The low-quality time series data representation learning method based on patch alignment autoencoder according to claim 3 is characterized in that For the temporal encoding mechanism: the input of this mechanism is the segmented patches output by the transformable patch technology. The observation data timestamps in each patch are encoded in a temporal embedding manner, and then the patch embedding obtained by splicing the time embedding and the patch is output. The formula for the temporal embedding encoding process is as follows: Among them, t represents the timestamp, d p represents the embedding dimension of the p-th patch, D t is the total dimension of the embedding, i.e. d p The maximum value; when d p = 0, the linear term ω 0p t+α 0p Used to capture the aperiodic pattern of the p-th patch evolving over time, ω 0p and α 0p are two learnable parameters; when 0 <d p <D t When ω dp and α dp Represent the frequency and phase of the sine function, respectively, and are used to capture the periodic characteristics of the time series.

6. The low-quality time series data representation learning method based on patch alignment autoencoder according to claim 3 is characterized in that The patch alignment module is composed of a transformable patch technology and a time encoding mechanism. The input of the patch alignment module is the low-quality time series of each unit after the introduction of noise; the transformable patch technology is used to segment the noisy time series, and the time encoding mechanism is used to preserve the time information of the sequence; Adaptive Weighted TCN Module: This module takes as input the patch embeddings output by the temporal encoding mechanism, adapts the temporal features of the input using meta-filter parameters, applies TCN for convolution, and then outputs the latent embedding of each patch. Transformer architecture: The input of this architecture is the latent embedding of the patch output by the adaptive weighted TCN module, which combines the latent representation with the positional encoding and obtains the output using a multi-head attention mechanism, and then outputs the latent representation of each patch.

7. The low-quality time series data representation learning method based on patch alignment autoencoder according to claim 3 is characterized in that The encoder consists of an adaptively weighted TCN module and a Transformer architecture. The input of the encoder is the patch embedding output by the temporal encoding mechanism; the adaptively weighted TCN module is used to capture the local periodic characteristics of each patch, and the Transformer architecture is used to capture long-range dependencies. The decoder is a fully connected layer. The input of the decoder is the potential representation of the patch output by the Transformer architecture. The feature vector in the latent space is converted back to time series data through linear transformation, and then the ReLU activation function is applied to process it, and finally the reconstructed time series is output.

8. The low-quality time series data representation learning method based on patch alignment autoencoder according to claim 3 is characterized in that: The step (3) is specifically as follows: Train the learning model for low-quality time series data representation, and use the backpropagation algorithm to optimize the parameters according to the results of the loss function during training; The loss function is specifically: Where N represents the total number of variables, L n Indicates the length of the sequence, represents the i-th observation value of the n-th variable in the original time series X, Represents the reconstructed time series X ′ The i-th observation of the n-th variable in .

9. A computer device, characterized in that: The computer device includes a memory, a processor, and a computer program stored in the memory and executable on the processor. When the processor executes the computer program, the steps of the low-quality time series data representation learning method based on patch alignment autoencoder as described in any one of claims 1 to 8 are implemented.

10. A computer-readable storage medium, characterized in that: The computer-readable storage medium stores a computer program for executing the low-quality temporal data representation learning method based on patch alignment autoencoder according to any one of claims 1 to 8.

Images