Time sequence classification method based on multi-scale adaptive graph convolutional neural network

Abstract

The invention provides a multi-scale adaptive Graph Convolutional Neural Network (MS-AGCNN for short), which is used for a classification task of time sequence data, and is characterized in that the time sequence data can be classified by using the MS-AGCNN, and the time sequence data can be classified by using the MS-AGCNN. The neural network is combined with a long short-term memory (LSTM) network and a one-dimensional convolutional layer to effectively extract multi-scale features of a time sequence. Through the adaptive graph structure module, the MS-AGCNN can accurately capture the correlation and difference between the features, and utilizes the graph convolution module to fuse the dependency relationship between different features. Furthermore, the neural network adopts a unique multi-scale feature fusion method to deeply capture complex relationships among cross-scale features, including similarity, difference and interaction, so that the expression ability of the features is remarkably enhanced. According to the method, the LSTM and the one-dimensional convolutional layer are combined to enrich the input information of the graph structure, and the correlation and the difference between the features are extracted through the training process of the neural network, so that the accuracy and the robustness of time sequence classification are improved.

Description

Technical Field

[0001] The present invention relates to the field of time series data classification. Specifically, it relates to a deep learning model that combines multi-scale feature extraction, adaptive graph structure, and graph convolutional neural network for efficiently and accurately classifying time series data. Background Art

[0002] Time series data widely exists in various practical applications, such as financial market prediction, weather forecasting, medical diagnosis, equipment fault detection, and intelligent transportation systems. Accurately classifying time series data is of great significance for decision-making and analysis in these fields. However, time series data usually has complex dynamic patterns, non-linear features, and multi-scale characteristics, which pose challenges to the classification task.

[0003] Traditional time series classification methods include methods based on statistical feature extraction and methods based on dynamic time warping (DTW). However, these methods often perform poorly when dealing with time series data with high complexity and non-linear features. With the rapid development of deep learning technology, models such as long short-term memory networks (LSTM) and convolutional neural networks (CNN) have been widely applied to the feature extraction and classification tasks of time series data. These models can capture complex patterns and dependencies in time series, significantly improving the classification accuracy.

[0004] Recurrent Neural Networks (RNN) and its variants, such as Long Short-Term Memory (LSTM) and Gated Recurrent Unit (GRU), are effective tools for processing time series data. By introducing recurrent connections and gating mechanisms, they can capture long-term dependencies in time series. However, traditional RNNs are prone to the problems of gradient vanishing or gradient explosion when dealing with long sequences, and although LSTM and GRU alleviate these problems to a certain extent, there is still room for improvement in terms of the diversity of feature extraction and the interpretability of the model.

[0005] Convolutional Neural Networks (CNN) were originally mainly used for image processing, but they also perform well in time series classification tasks. Through convolutional operations, CNN can extract local features in time series data and perform dimensionality reduction and feature selection through pooling operations. However, a single CNN model has limitations in capturing global temporal dependencies.

[0006] To overcome the above problems, the present invention proposes a multi-scale adaptive graph convolutional neural network (MS-AGCNN), which realizes multi-scale feature extraction of time series data and captures complex dependencies by combining LSTM, one-dimensional convolutional layer, adaptive graph structure, and graph convolution module. This model can significantly improve the accuracy and robustness of time series classification, providing a more efficient and reliable solution for applications in related fields. Summary of the Invention

[0007] The present invention relates to a deep neural network model for time series classification, aiming to solve the problems of incomplete feature extraction, insufficient classification performance, and poor model interpretability in existing time series classification methods by combining long short-term memory network (LSTM), one-dimensional convolutional layer, adaptive graph structure module, and multi-scale feature fusion technology. The innovation of the present invention lies in proposing a multi-scale adaptive graph convolutional neural network (MS-AGCNN), which realizes efficient and accurate classification of time series data and has broad application prospects and practical value.

[0008] First, the present invention focuses on multi-scale feature extraction of time series data. Traditional time series classification methods often only focus on features at a single scale, resulting in incomplete feature extraction and limited classification performance. To overcome this problem, the present invention introduces LSTM and one-dimensional convolutional layer to perform preliminary feature extraction at fine scale and coarse scale respectively. The bidirectional long short-term memory network (BiLSTM) is used to process time series data. Through its internal gating mechanism and cell state, BiLSTM can effectively capture long-term dependencies in time series data and input them as fine-grained features into subsequent processing modules. The one-dimensional convolutional layer is used to process time series data. Feature extraction is performed by sliding the convolutional kernel over the sequence data with a step size of k. Compared with the traditional method of sampling every k time points, the use of the convolutional layer improves the comprehensiveness of feature extraction, enhances the flexibility and interpretability of the model. The size of the convolutional kernel is M and the step size is k, where M and k are preset parameters.

[0009] Secondly, the present invention proposes an adaptive graph structure module for further capturing complex relationships between features. In time series classification tasks, the relationships between features are often non-linear, complex, and difficult to measure with traditional Euclidean distance. To solve this

[0010] problem, the present invention uses the correlation and difference between the embedded features obtained by preliminary feature extraction to create a nearest neighbor graph by calculating the Pearson correlation coefficient. This adaptive graph structure can better capture complex relationships between features and enhance the model's ability to understand data.

[0011] Meanwhile, to further improve the feature selection ability and non-linear expression ability, the present invention also integrates three one-dimensional convolutional layers to form a temporal convolutional module, and processes the output of the convolutional layer through a gated linear unit (GLU) and a PReLU activation function. Then, the present invention further fuses and transforms the features through a graph convolutional module. The graph convolution operation can fuse the dependency relationships between data features and the features in the time domain, and perform feature transformation using a layer-specific trainable weight matrix and activation function. This step not only enhances the model's ability to model the complex relationships between features, but also improves the model's ability to process time series data. Through the graph convolutional module, the present invention can better capture the dynamic change characteristics and temporal dependency relationships in time series data. Then, the further extraction of time series features is jointly completed by combining the temporal convolutional module and the graph convolutional module.

[0012] Finally, the present invention uses a multi-scale feature fusion technique to connect, calculate, and combine the feature vectors extracted at different scales to generate diverse feature representations. Through multi-scale feature fusion, the present invention can capture the complex relationships between cross-scale features, including similarities, differences, and interactions. This cross-scale feature fusion not only enhances the feature expression ability, but also improves the classification accuracy of the model. Compared with traditional single-scale feature extraction methods, the multi-scale feature fusion technique of the present invention can more comprehensively capture the feature information in time series data, thereby improving the classification performance.

[0013] In summary, the time series classification deep neural network model of the present invention realizes the efficient and accurate classification of time series data by optimizing the model structure and feature extraction method. Compared with the prior art, the present invention has achieved significant improvements in classification performance, comprehensiveness of feature extraction, and model interpretability. The innovation of the present invention lies in the combination of LSTM, one-dimensional convolutional layers, an adaptive graph structure module, and a multi-scale feature fusion technique, which improves the accuracy and efficiency of time series classification by optimizing the model structure and feature extraction method. The time series classification deep neural network model of the present invention has broad application prospects and practical value, and has potential application value in multiple fields such as financial forecasting, meteorological analysis, medical diagnosis, and industrial production monitoring. BRIEF DESCRIPTION OF THE DRAWINGS

[0014] Figure 1 The general architecture of MS-AGCNN mainly includes three processes: (1) preprocessing of multivariate time series and preliminary feature extraction by LSTM and one-dimensional convolutional layers; (2) creation of a graph structure and further extraction of multi-scale features by an adaptive graph structure module; (3) multi-scale feature fusion and final classification process. DETAILED DESCRIPTION OF THE INVENTION

[0015] The preferred embodiments of the present invention will be described below in conjunction with the accompanying drawings. It should be understood that the preferred embodiments described herein are only used to illustrate and explain the present invention and are not intended to limit the present invention.

[0016] It should be noted that the steps shown in the flowchart of the accompanying drawings can be executed in a computer system such as a set of computer-executable instructions. And although the logical order is shown in the flowchart, in some cases, the steps shown or described can be executed in a different order than here.

[0017] As shown in the figure, the present invention provides a time series classification method based on multi-scale feature extraction and adaptive graph neural network, including the following steps:

[0018] Step 1: Data processing

[0019] Process the publicly available dataset into the same standard. Generally speaking, the time series will be processed into a three-dimensional matrix, where each dimension represents the number of samples, the number of features, and the length of the time series respectively. The number of features and the length of the time series for each sample should be the same.

[0020] Step 2: Preliminary feature extraction

[0021] Input the processed time series data into the neural network. The neural network contains three branches, which respectively represent three different scales.

[0022] One of the branches is processed by BiLSTM to represent the feature embedding of the sequence under the original scale. The other two branches are both one-dimensional convolutions. The downsampling process is approximately completed by setting the stride equal to the downsampling rate. In this way, the number of time series features also becomes a fixed 128.

[0023] Step 3: Graph creation

[0024] Based on the feature embeddings extracted from different scales, we create nearest neighbor graphs respectively. That is to say, at this time, the size of the adjacency matrix we create is 128×128, and the standard is the Pearson correlation coefficient.

[0025] Step 4: Further extract features through the adaptive graph structure module

[0026] The adaptive graph structure module mainly consists of two sub - modules, the graph convolution module and the temporal convolution module. The temporal convolution module is composed of three one - dimensional convolutional layers. Specifically, the input data is first processed through three convolutional layers. The output of one convolutional layer is divided into a linear part and a gated part, and the gated linear unit (GLU) operation is used to enhance the feature selection ability. At the same time, the output of another convolutional layer is processed through the PReLU activation function. Finally, the outputs of the three convolutional layers are added together and batch normalization is performed. To fuse the dependencies between data features and the features in the time domain, we construct the graph convolution module. The module can be stacked or extended according to the scale and complexity of specific situations. As shown in the figure, the middle graph convolution layer is a bridge connecting the two temporal convolution modules. The input data is processed in the time dimension through the temporal convolution module, and then the graph convolution operation is used to capture the dependencies between different features in the data. Specifically, the graph convolution operation can be expressed by the following formula: Finally, we complete the dimensionality reduction operation through the average pooling layer to compress the time dimension.

[0027] Step 5: Multi - scale feature fusion

[0028] For three different scales, after we complete the final feature extraction through the adaptive graph structure module, it will become three feature embedding matrices of sample number × 128. At this time, by connecting, calculating, and combining these feature vectors of different scales, we generate diverse feature representations. Specifically

[0029]

[0030] where, o i represents the features extracted from different scales. These operations can capture the complex relationships between cross - scale features, including similarities, differences, and interactions, enhancing the feature expression ability. Finally, we input the fused features into the classifier to obtain the final classification result.

[0031] Step 6: Input to the classifier

[0032] We finally input the feature embeddings that fuse multiple scales into the classifier. The final classifier completes the final classification operation through the linear layer and the softmax activation function

[0033] Through the above steps, the present invention can effectively extract multi - scale features in time - series data and use the adaptive graph neural network for feature extraction and classification, thereby improving the accuracy and robustness of time - series classification.

[0034] The above is only a preferred specific embodiment of the present application, but the protection scope of the present application is not limited thereto. Any changes or substitutions that can be easily conceived by those skilled in the art within the technical scope disclosed in the present application should be covered by the protection scope of the present application. Therefore, the protection scope of the present application should be subject to the protection scope of the claims.

Claims

1. A deep neural network model for time series classification, characterized by The following steps are involved: a. Preliminary extraction of multi-scale features: Use the long short-term memory network (LSTM) to perform preliminary fine-scale feature extraction on time series data. Through its internal gating mechanism and cell state, it can effectively capture the long-term dependencies in time series data and use them as fine-grained features. A one-dimensional convolutional layer is used to perform preliminary feature extraction on a coarse scale for time series data. The convolution kernel is slid on the sequence data with a step size of k to extract features. Compared with the traditional method of sampling every k time points, this method improves the comprehensiveness of feature extraction and enhances the flexibility and interpretability of the model. Among them, the convolution kernel size of the one-dimensional convolution layer is M, the step size is k, and M and k are preset parameters; b. Adaptive graph structure module: Using the correlation and difference between the embedded features obtained by preliminary feature extraction, the nearest neighbor graph is created by calculating the mean square error (MSE), and its negative exponential value is used as the weight of the edge to enhance the model's ability to capture the complex relationship between features; The temporal convolution module is composed of three one-dimensional convolutional layers. The output of the convolutional layer is processed by the gated linear unit (GLU) and PReLU activation functions to capture the temporal dependency of the data and enhance the feature selection and nonlinear expression capabilities. Construct a graph convolution module, fuse the dependencies between data features and time domain features through graph convolution operations, and use layer-specific trainable weight matrices and activation functions to perform feature transformation; c. Multi-scale feature fusion: Connect, calculate and combine feature vectors extracted at different scales to generate diverse feature representations; Capturing the complex relationships between cross-scale features, including similarities, differences, and interactions, to enhance the expressive power of features; d. Category: The fused features are input into the classifier to obtain the final classification result.

2. The deep neural network model for time series classification according to claim 1, characterized in that: The long short-term memory network (LSTM) effectively captures the long-term dependencies in time series data through its internal gating mechanism and cell state, and uses it as fine-grained features, thereby improving the classification accuracy of the model.

3. The deep neural network model for time series classification according to claim 1, characterized in that: The one-dimensional convolution layer extracts features by sliding the convolution kernel on the sequence data. Compared with the traditional method of sampling every k time points, it improves the comprehensiveness of feature extraction and enhances the flexibility and interpretability of the model.

4. The deep neural network model for time series classification according to claim 1, characterized in that: The adaptive graph structure module uses mean square error (MSE) to create a nearest neighbor graph and uses its negative exponential value as the weight of the edge, which enhances the model's ability to capture complex relationships between features and improves the model's classification performance.

5. The deep neural network model for time series classification according to claim 1, characterized in that: The temporal convolution module integrates three one-dimensional convolutional layers and processes the output of the convolutional layer through the gated linear unit (GLU) and PReLU activation functions, which enhances the feature selection ability and nonlinear expression ability and improves the model's ability to process time series data.

6. The deep neural network model for time series classification according to claim 1, characterized in that: The graph convolution module captures the dependencies between different features in the data through graph convolution operations, and uses layer-specific trainable weight matrices and activation functions to transform features, thereby improving the model's ability to model complex relationships between features.

7. The deep neural network model for time series classification according to claim 1, characterized in that: The multi-scale feature fusion generates diversified feature representations by connecting, calculating and combining feature vectors of different scales, and captures the complex relationship between cross-scale features, thereby enhancing the expressiveness of features and improving the classification accuracy of the model.

8. A computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, implements the deep neural network model for time series classification described in any one of claims 1 to 7.

Images