Multi-scale decomposition-based multivariate time series classification method, program, equipment and storage medium

Abstract

The invention belongs to the technical field of computers, and particularly relates to a multivariate time sequence classification method based on multi-scale decomposition, a program, equipment and a storage medium. According to the method, a multi-scale time slice method is introduced, an input time sequence is divided into non-overlapping time slices along a time dimension, modeling is performed on subsequences in different layers, and by using different time slice sizes in different PatchMama layers, a model can capture multi-scale features in time sequence data, so that the classification accuracy is improved, and the classification efficiency is improved. And the model can process data of different time scales, and has relatively strong generalization ability. The PatchMama model designed by the invention fuses two perspectives of different time resolutions and time distances, and can more accurately capture complex structures and dynamic changes in time sequence data, so that information in the data is more comprehensively understood, and the classification capability of MTSC problems is remarkably improved.

Description

Technical Field

[0001] The present invention belongs to the field of computer technology, and specifically relates to a multivariate time series classification method, program, device and storage medium based on multi-scale decomposition. Background Art

[0002] The task of time series classification (TSC) has always been one of the hot topics of research. In recent years, with the rise of deep neural networks, the method of extracting time series features using deep neural networks has been widely used in time series classification tasks. This method can effectively extract useful features from time series data, thereby improving the accuracy and efficiency of classification. Multi-channel convolutional networks were first proposed to deal with the problem of multivariate time series classification (MTSC), demonstrating its potential in processing time series data. A series of convolution-based methods have also achieved remarkable success in such tasks. These methods can effectively extract local and global features from time series. Some methods use random convolution kernels to train linear classifiers, and their simplicity and efficiency make them considered as a powerful time series classification method in recent studies. In addition, some scholars have designed hybrid networks to perform time series classification, such as combining LSTM layers and stacked convolutional layers and structures such as squeeze and excitation blocks to extract richer features. In addition, graph convolutional networks have also been introduced into the task of time series classification, which further improves the accuracy and robustness of classification by mining the graph structure between time series for comprehensive analysis.

[0003] Although the above methods have achieved remarkable results in time series classification tasks, they still have certain limitations when processing multi-scale time series data. Real time series data often contains complex information on multiple time scales from short-term to long-term. Many architectures proposed recently extract features by dividing time at different time resolutions, and they have demonstrated their capabilities in multivariate time series classification tasks. However, when these architectures divide time slices according to time resolution, they cannot guarantee that each time slice has the same time interval, which leads to inconsistent time distances between time slices. This inconsistency in time distance may have a certain impact on the accurate extraction of data features and the performance of the model. Summary of the invention

[0004] The purpose of the present invention is to provide a multivariate time series classification method based on multiscale decomposition, which can combine time resolution and time distance for multiscale modeling, and overcome the problem of inconsistency in time distance caused by dividing time periods by different time resolutions.

[0005] A multivariate time series classification method based on multi-scale decomposition includes the following steps:

[0006] Step 1: Stack multiple PatchMamba layers and use different time slice sizes in different PatchMamba layers;

[0007] Step 2: Input the time series X to be classified into k PatchMamba layers for processing. The k PatchMamba layers gradually decompose the time series X into k components {S 1 ,...,S k}, each component is represented by a low-dimensional representation generated by its corresponding layer {E 1 ,...,E k}, representing the characteristics of time series X at different levels;

[0008] Step 3: Place {S 1 ,...,S k} are input into the pre-trained classification model to obtain the corresponding components S i The classification results of all components are combined, and the classification result to which the most components belong is taken as the classification of time series X.

[0009] Furthermore, the PatchMamba layer includes a time series decomposition module, an encoder module, a decoder module, and a time series recovery module; 0 =X, in the i-th PatchMamba layer, the encoder module first performs Z i-1 Process and generate a low-dimensional representation E i ; The decoder module is based on E i Rebuilt as S i , the time series recovery module converts S i Restored to the original dimension, the time series decomposition module performs Z i-1 Perform time series segmentation operation to obtain Z i , Z i =Z i-1 -S i ; Z i As the input of the next layer, the i+1th PatchMamba layer.

[0010] Furthermore, the k PatchMamba layers in step 2 gradually decompose the time series X into k components {S 1 ,...,S k The specific steps of} are:

[0011] Step 2.1: Initialize Z 0 =X, i=1;

[0012] Step 2.2: Z i-1 Input to the i-th PatchMamba layer;

[0013] Step 2.3: The encoder module of the i-th PatchMamba layer is Z i-1 Process and generate a low-dimensional representation E i ;

[0014] Step 2.4: The decoder module of the i-th PatchMamba layer is based on E i Rebuilt as S i ;

[0015] Step 2.5: The time series decomposition module of the i-th PatchMamba layer performs Z i-1 Perform time series segmentation operation to obtain Z i ;

[0016] Z i =Z i-1 -S i

[0017] Step 2.6: If i<k, set i=i+1 and return to step 2.2.

[0018] Furthermore, in the encoder module of the PatchMamba layer, the input Z is firstly transformed into i-1 Perform correlation calculations, and then use the Mamba model to perform correlation learning calculations between patches and within patches, and finally generate E i ; A shape transformation operation is performed after each calculation.

[0019] A computer device / equipment / system comprises a memory, a processor and a computer program stored in the memory, wherein the processor executes the computer program to implement the steps of the multivariate time series classification method based on multi-scale decomposition.

[0020] A computer-readable storage medium stores a computer program / instruction, which, when executed by a processor, implements the steps of the multivariate time series classification method based on multi-scale decomposition.

[0021] A computer program product comprises a computer program / instruction, which, when executed by a processor, implements the steps of the multivariate time series classification method based on multi-scale decomposition.

[0022] The beneficial effects of the present invention are:

[0023] The present invention introduces a multi-scale time slice method, divides the input time series into non-overlapping time slices along the time dimension, and models the subsequences in different layers. By using different time slice sizes in different PatchMamba layers, the model can capture the multi-scale features in the time series data, thereby improving the classification accuracy, and also enables the model to process data of different time scales, with strong generalization ability. The PatchMamba model designed by the present invention integrates two perspectives of different time resolutions and time distances, and can more accurately capture the complex structure and dynamic changes in the time series data, thereby more comprehensively understanding the information in the data, and significantly improving the classification ability of the MTSC problem. BRIEF DESCRIPTION OF THE DRAWINGS

[0024] Figure 1 Execute the flow chart for the PatchMamba model.

[0025] Figure 2 Schematic diagram of encoder and decoder module.

[0026] Figure 3 Data graphs to evaluate the impact of different patch sizes on classification accuracy, including AWR (upper left), AF (upper right), SCP1 (lower left), and UWGL (lower right).

[0027] Figure 4 Illustration of the adjacency matrix learned for the encoder and decoder.

[0028] Figure 5 This is the structure diagram of PatchMamba. DETAILED DESCRIPTION

[0029] The present invention is further described below in conjunction with the accompanying drawings.

[0030] The present invention relates to the field of deep learning models for time series analysis and proposes a new model named PatchMamba. The structure of the PatchMamba module is as follows: Figure 5 As shown in the figure, the model is based on the Mamba model and introduces a multi-scale time slice mechanism. The model combines the two perspectives of different time resolutions and time distances, and can more accurately capture the complex structure and dynamic changes in time series data, thereby more comprehensively understanding the information in the data and significantly improving the classification ability of the MTSC problem.

[0031] Patching: For the input data (X, Y), X∈R C×L , where C is the number of channels and L is the time step. According to the predefined time slice size P of the PatchMamba layer i, the time slices are divided, the original data X is decomposed into multiple time slices of different sizes, and then the divided time slices are unified to obtain a high-dimensional tensor with a shape of C×L'×p i , Each PatchMamba layer i Different parameters can divide the time series into different time resolutions, which allows the model to capture multi-scale features.

[0032] The Mamba model is a simplified SSM (State Space Model) architecture. The SSM transform is derived from the classic state space model, which maps a one-dimensional input signal x(t)∈R to a one-dimensional output signal y(t)∈R through an N-dimensional hidden state h(y)∈R N The state equation describes how the state changes (state matrix A) based on how the input affects the state (input matrix B), and the output equation describes how the state is transformed into the output (output matrix C).

[0033] h'=Ah(t)+Bx(t),y(t)=Ch(t)

[0034] Discretization: To process the discrete input sequence x=(x 0 ,x 1 ,...)∈R L The S4 model uses a step size Δ (which can be regarded as the resolution of the continuous input x(t)) to discretize the continuous parameters A and B into This is achieved through the zero-order hold technique (ZOH). Using the zero-order hold technique, a new learnable parameter, step size Δ, is generated. The formula for the discretization process is as follows:

[0035]

[0036] After A and B are discretized, the formula can be restated as:

[0037]

[0038] SSM can be efficiently calculated by RNN. This recursive process is reformulated and calculated as a convolution operation. The SSM convolution kernel calculation formula is as follows:

[0039]

[0040] Mamba integrates SSM and selection mechanism. The selection mechanism is implemented by setting the input SSM function parameters, and the input-based interaction is realized on the sequence. The parameters B, C, and Δ depend on the input sequence x:

[0041] B, C, Δ=Linear(x)

[0042] Where B∈R B×L×N , C∈R B×L×N ,Δ∈R B×L×D . xInput x∈R B×L×D is the full shape, where B represents the batch size and D represents the number of channels.

[0043] Mamba uses hardware-aware techniques to efficiently compute Selective SSM.

[0044] The present invention stacks multiple PatchMamba layers, uses different time slice sizes in different PatchMamba layers, and k PatchMamba layers gradually decompose the input X into k components {S 1 ,...,S k}, each component is represented by a low-dimensional representation generated by its corresponding layer {E 1 ,...,E k}, which represents the characteristics of the input data at different levels, so that the model can capture the multi-scale features in the time series data, thereby improving the classification accuracy, and also enables the model to process data of different time scales with strong generalization ability.

[0045] The PatchMamba layer contains a time series decomposition module, an encoder, a decoder, and a time series recovery module.

[0046] In the time series decomposition module, the data is first divided into multiple time slices. Different from the usual architecture that uses a uniform time slice size to divide the data, the present invention defines a time slice size set {S 1 ,...,S k}, each time slice size S corresponds to a time series segmentation operation, which simulates the relationship between time series of different sizes.

[0047]

[0048] Where P l , P s Represents the maximum time slice and the i-th time slice.

[0049] In time series analysis, existing deep learning methods usually directly learn the representation of input data X and map it to the output. However, this method may encounter difficulties in processing time series containing multiple complex time patterns. To solve this problem, the present invention decomposes the time series X to more easily capture the correlation between time slices of different sizes. The decomposition of X is:

[0050]

[0051] Among them, S i and R represent the i-th component and residual, respectively.

[0052] For the i-th PatchMamba layer, Z i Refers to the remainder after the first k components are decomposed from the input X:

[0053]

[0054] In other words, the current remainder is equal to the remainder of the previous layer minus the generated component S of the current layer. i :

[0055] Z i =Z i-1 -S i

[0056] Initial definition Z 0 =X.

[0057] Then reorder Z in the time series decomposition module i-1 , and feed it back to the encoder module to generate a low-dimensional representation E of the i-th component i :

[0058] E i =g i (Z i-1 )

[0059] Decoder module from E i Reconstruction S i , the time series recovery module restores the time slice to its original dimension, and then i-1 Subtract S from i , we get Z i , and Z i As input, it is sent to the next PatchMamba layer for further decomposition. i Projected and summed through the linear layer, we get

[0060]

[0061] The brief process of PatchMamba is as follows Figure 1 shown.

[0062] Time series decomposition module: PatchMamba model predefines a time slice p for each layer i, which is a hyperparameter that can be tuned for a specific dataset. To transform an input time series with C channels and L time steps into time slices of size p, the steps are as follows:

[0063] 1) Time series padding: Use zero values ​​to fill the beginning of the time series to ensure that the total length of the time series can be divided by the time slice size p. This step ensures that the subsequent time slice segmentation is carried out smoothly.

[0064] 2) Multi-scale time slice segmentation: Based on a predefined set of time slice sizes, the padded time series is segmented into multiple time slices of different sizes. The largest time slice is further segmented into smaller time slices to capture the time series features at different scales. We believe that smaller time slices already contain enough details and do not need to be further divided. For example, the predefined time slice size set is [16, 8, 4, 2, 1, 1], and a time slice of size 16 will be subdivided into [8, 4, 2, 1, 1] to accommodate different time scales.

[0065] 3) Non-overlapping time slice extraction: In the dimension of the time series, non-overlapping time slices are extracted according to the step size p. This step ensures that each time slice contains a unique and non-overlapping part of the original time series.

[0066] 4) Data rearrangement: The extracted time slices will be rearranged, and time slices of different sizes will be unified to create a new time slice dimension. The model will get a high-dimensional tensor with a shape of C×L'×p, where C is the number of channels of the input time series. p is the size of the time slice.

[0067] Encoder and decoder modules: Figure 2 As shown, both the encoder module and the decoder module are built based on the Mamba framework for feature extraction in different dimensions. Three types of Mamba blocks are used in the present invention, which are used in the encoder module and the decoder module respectively:

[0068] Inter-channel Mamba block: allows communication between different channels to capture inter-channel correlation.

[0069] Inter-patch Mamba blocks: Allow communication between different patches to capture global context information.

[0070] Intra-Patch Mamba Blocks: Allows communication between different time steps within a patch to capture sub-sequence level changes.

[0071] The structure of the encoder module: The encoder module combines these Mamba blocks and a linear layer to convert the input data S iConvert to low-dimensional representation E i .

[0072] Structure of the decoder module: The decoder module has the same number and type of blocks as the encoder module, but in reverse order. It receives the low-dimensional representation E generated by the encoding block. i , and gradually reconstruct it into the original sequence S i .

[0073] For high-dimensional tensors, a channel Mamba block, an inter-patch Mamba block, an intra-patch Mamba block, and a linear layer are sequentially passed to capture the correlation between channels, the global context, and the changes at the subsequence level, respectively. The decomposed time series S i Convert to low-dimensional representation E i , so as to facilitate subsequent calculation and output. Figure 2 The lower dimensional representation E i The module is reconstructed into S by passing through a linear layer, an intra-patch Mamba block, an inter-patch Mamba block, and an inter-channel Mamba block. i .like Figure 2 As shown in the lower part. In the Unpatching module, S i Restored to its original dimensions.

[0074] Embodiment 1:

[0075] A multivariate time series classification method based on multi-scale decomposition includes the following steps:

[0076] Step 1: Stack multiple PatchMamba layers and use different time slice sizes in different PatchMamba layers;

[0077] Step 2: Input the time series X to be classified into k PatchMamba layers for processing. The k PatchMamba layers gradually decompose the time series X into k components {S 1 ,...,S k}, each component is represented by a low-dimensional representation generated by its corresponding layer {E 1 ,...,E k}, representing the characteristics of time series X at different levels;

[0078] Step 2.1: Initialize Z 0 =X, i=1;

[0079] Step 2.2: Z i-1 Input to the i-th PatchMamba layer;

[0080] Step 2.3: The encoder module of the i-th PatchMamba layer is Zi-1 Process and generate a low-dimensional representation E i ;

[0081] In the encoder module of the PatchMamba layer, the input Z is firstly transformed into the channel dimension using the Mamba model. i-1 Perform correlation calculations, and then use the Mamba model to perform correlation learning calculations between patches and within patches, and finally generate E i ;A shape transformation operation is performed after each calculation;

[0082] Step 2.4: The decoder module of the i-th PatchMamba layer is based on E i Rebuilt as S i ;

[0083] Step 2.5: The time series decomposition module of the i-th PatchMamba layer performs Z i-1 Perform time series segmentation operation to obtain Z i ;

[0084] Z i =Z i-1 -S i

[0085] Step 2.6: If i<k, set i=i+1 and return to step 2.2;

[0086] Step 3: Place {S 1 ,...,S k} are input into the pre-trained classification model to obtain the corresponding components S i The classification results of all components are combined, and the classification result to which the most components belong is taken as the classification of time series X.

[0087] The present invention proposes a multivariate time series classification PatchMamba model based on multiscale decomposition, which introduces a multiscale time slice method, divides the input time series into non-overlapping time slices along the time dimension, and models the subsequences in different layers. By using different time slice sizes in different PatchMamba layers, the model can capture the multiscale features in the time series data, thereby improving the classification accuracy, and also enables the model to process data of different time scales, with strong generalization ability. The PatchMamba classification model proposed by the present invention has significant technical effects and can effectively improve the classification performance of time series data.

[0088] In order to fully evaluate the effectiveness of the present invention, the following experimental verification is carried out. The data sets are all selected from the widely used UEA Multivariate Time Series Classification (MTSC) archive, from which 10 multivariate data sets with different characteristics are carefully selected: ArticularyWordRecognition (AWR), Atrial Fibrillation (AF), CharacterTrajectories (CT), Cricket, FaceDetection (FD), FingerMovements (FM), MotorImagery (MI), SelfRegulationSCP1 (SCP1), SelfRegulationSCP2 (SCP2) and UWaveGestureLibrary (UWGL). The main features of these data sets are shown in Table 1.

[0089] Table 1 Main characteristics of the dataset

[0090]

[0091] For comprehensive evaluation, the present invention selects the following popular baseline methods for comparison: PatchTST ("A time series is worth 64 words: Long-term forecasting with transformers"), Crossformer ("Crossformer: Transformer utilizing cross-dimension dependency for multivariate time series forecasting"), TimesNet ("Timesnet: Temporal 2d-variation modeling for general time series analysis"), Dlinear ("Are transformers effective for time series forecasting"), LightTS ("Less is more: Fast multivariate time series forecasting with light sampling-oriented mlp structures"), ETSformer ("Etsformer: Exponential smoothing transformers for time-series forecasting"), NST ("Non-stationary Transformers: Exploring the Stationarity in Time Series Forecasting"), FEDformer ("Fedformer: Frequency enhanced decomposed transformer for long-term series forecasting"), TST ("A transformer-based framework for multivariate time series representation learning"), LS ("Learning time-series shapelets"), Informer ("Informer: Beyond efficient transformer for long sequence time-series forecasting"), GTN ("Gated transformernetworks for multivariate time series classification") and MR ("Minirocket: A very fast (almost) deterministic transform for time series classification").

[0092] All models were implemented in PyTorch and trained using an NVIDIA GeForce RTX 4090 GPU with 24GB of memory.

[0093] These baseline methods cover a variety of different time series classification techniques, providing a rich comparative perspective for the experiment. By comparing with these baseline methods, the performance of the present invention on different data sets can be more comprehensively evaluated. This experiment uses accuracy as the metric, and the results are shown in Table 2. The result of "N / A" means that the corresponding operation cannot be executed and produce results due to insufficient memory.

[0094] Table 2 Performance of different methods on different datasets

[0095]

[0096] Table 2 summarizes the classification accuracy of all compared methods. The present invention is able to surpass the previous classifiers in terms of average accuracy, and for each data set, the classification performance of the present invention is either the most accurate or very close to the most accurate model, showing an improvement in the classification ability of the MTSC problem.

[0097] The results of the ablation experiment are shown in Table 3.

[0098] Table 3 Ablation experiment

[0099]

[0100] From the analysis of the results in Table 3, it can be seen that further decomposition of modules has a significant impact on improving the prediction accuracy of the model. In model design, it is very important to extract multi-scale features by optimizing the combination of time slice sizes. The time series decomposition module also helps the model capture the temporal dynamic characteristics in time series data. In terms of efficiency and local pattern recognition, the intra-patch module shows advantages, while the inter-patch module is better at capturing global patterns.

[0101] In addition, in order to study the effect of the time slice size setting on the classification accuracy of the present invention, parameter analysis is performed on the Articulary Word Recognition (AWR), Action Features (AF), Shape Context Points 1 (SCP1) and Unipen On-line Handwriting Recognition with Geometric Features and Lexicon (UWGL) datasets. The introduction of a maximum time slice size of 16 significantly improves the classification accuracy of PatchMamba. The visualization results are shown in Figure 2. Figure 3 shown.

[0102] Each element of the adjacency matrix reflects the strength of the correlation between each pair of nodes. Visualize the encoder and decoder matrices learned by PatchMamba on the UEA dataset, as shown in Figure 4 shown.

[0103] It can be concluded that the adjacency matrix of the encoder shows that the correlation between most nodes is weak, and only a few node pairs maintain strong correlation, which reflects that the trend sequence is a large-scale sequence dominated by global information, and only a very small amount of local spatial information is retained. The adjacency matrix of the decoder shows that many node pairs have strong correlation, which reflects that the multivariate time series is multi-scale and contains a lot of local and global information.

[0104] Through the above technical solution, the PatchMamba model can realize multi-scale analysis of time series data, thereby more accurately capturing complex patterns and features in the data. In addition, since the size and number of time slices can be adjusted according to specific tasks and data sets, the present invention has strong flexibility and adaptability.

[0105] The above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. For those skilled in the art, the present invention may have various modifications and variations. Any modification, equivalent replacement, improvement, etc. made within the spirit and principle of the present invention shall be included in the protection scope of the present invention.

Claims

1. A multivariate time series classification method based on multi-scale decomposition, characterized in that: The following steps are involved: Step 1: Stack multiple PatchMamba layers and use different time slice sizes in different PatchMamba layers; Step 2: Input the time series X to be classified into k PatchMamba layers for processing. The k PatchMamba layers gradually decompose the time series X into k components {S1,...,S k }, each component is represented by a low-dimensional representation {E1,...,E k }, representing the characteristics of time series X at different levels; Step 3: Replace {S1,...,S k } are input into the pre-trained classification model to obtain the corresponding components S i The classification results of all components are combined, and the classification result to which the most components belong is taken as the classification of time series X.

2. The multivariate time series classification method based on multiscale decomposition according to claim 1, characterized in that: The PatchMamba layer includes a time series decomposition module, an encoder module, a decoder module, and a time series recovery module; Z0=X. In the i-th PatchMamba layer, the encoder module first performs i-1 Process and generate a low-dimensional representation E i ; The decoder module is based on E i Rebuilt as S i , the time series recovery module converts S i Restored to the original dimension, the time series decomposition module performs Z i-1 Perform time series segmentation operation to obtain Z i , Z i =Z i-1 -S i ; Z i As the input of the next layer, the i+1th PatchMamba layer.

3. The multivariate time series classification method based on multiscale decomposition according to claim 2, characterized in that: In step 2, the k PatchMamba layers gradually decompose the time series X into k components {S1,...,S k The specific steps of} are: Step 2.1: Initialize Z0=X, i=1; Step 2.2: Z i-1 Input to the i-th PatchMamba layer; Step 2.3: The encoder module of the i-th PatchMamba layer is Z i-1 Process and generate a low-dimensional representation E i ; Step 2.4: The decoder module of the i-th PatchMamba layer is based on E i Rebuilt as S i ; Step 2.5: The time series decomposition module of the i-th PatchMamba layer performs Z i-1 Perform time series segmentation operation to obtain Z i ; Z i =Z i-1 -S i Step 2.6: If i<k, set i=i+1 and return to step 2.

2.

4. The multivariate time series classification method based on multiscale decomposition according to claim 3 is characterized by: In the encoder module of the PatchMamba layer, the input Z is firstly coded using the Mamba model in the channel dimension. i-1 Perform correlation calculations, and then use the Mamba model to perform correlation learning calculations between patches and within patches, and finally generate E i ; A shape transformation operation is performed after each calculation.

5. A computer device / equipment / system comprising a memory, a processor and a computer program stored in the memory, characterized in that: The processor executes the computer program to implement the steps of the method according to any one of claims 1 to 4.

6. A computer-readable storage medium having a computer program / instruction stored thereon, characterized in that: When the computer program / instructions are executed by a processor, the steps of the method according to any one of claims 1 to 4 are implemented.

7. A computer program product comprising a computer program / instructions, characterized in that: When the computer program / instructions are executed by a processor, the steps of the method according to any one of claims 1 to 4 are implemented.

Images