Multivariable time sequence classification method and device for multi-scale period extraction

Abstract

The invention provides a multivariable time sequence classification method and device for multi-scale period extraction, and the method comprises the steps: converting a time domain signal into a frequency domain signal through fast Fourier transform, and extracting periodic features and frequency spectrum information; then, main period characteristics are selected through a self-adaptive period detection module, and the data are divided into a plurality of period segments; thirdly, extracting local features by using a multi-scale convolutional neural network, and capturing local discrimination features of different time scales in combination with an Inception module; modeling a cross-time and cross-variable global dependency relationship through a self-attention mechanism, and extracting long-range dependency features; and finally, carrying out weighted fusion on the local features and the global features, and inputting into a classifier to generate a classification result. According to the method, multi-period and multi-scale features can be better captured, the classification precision is improved, particularly, the method is excellent in performance in biological signals, electroencephalogram signals, action recognition and multi-source data analysis, and higher generalization ability and robustness are achieved.

Description

Technical Field

[0001] This invention relates to the fields of multivariate time series analysis, deep learning and pattern recognition, and in particular to a multivariate time series classification method and apparatus with multi-scale period extraction. Background Technology

[0002] Multivariate time series classification (MTSC) is a key technology for analyzing time series data composed of multiple variables, and it is widely used in fields such as health monitoring, traffic management, and environmental monitoring. Traditional time series classification methods mostly rely on manual feature extraction and classic machine learning algorithms, such as Support Vector Machines (SVM) and Random Forests. When dealing with high-dimensional and complex data, these methods often require extensive feature engineering and struggle to automatically adapt to different time scales and complex relationships between variables, resulting in low processing efficiency and poor generalization ability. With the development of deep learning technology, time series classification methods based on Convolutional Neural Networks (CNN) and Recurrent Neural Networks (RNN) have gradually become a research hotspot. These methods significantly improve classification performance through automated feature learning. In recent years, the Transformer model has become an important tool in time series analysis due to its powerful global dependency modeling capabilities, especially demonstrating advantages in long-term dependency modeling.

[0003] However, existing deep learning methods still face several challenges when processing time-series data with multi-period and multi-scale features. First, traditional models largely rely on global attention mechanisms, which can lead to excessive computational complexity when dealing with high-dimensional data, particularly limiting performance in long-term series or high-dimensional multivariate data. Second, existing methods struggle to balance capturing local features and long-term dependencies, especially in the interaction between short-term local patterns and long-term global dependencies, making it difficult to achieve ideal performance. Furthermore, existing models often fail to effectively integrate time-frequency domain features when processing features across time scales, limiting their adaptability and generalization ability in complex contexts.

[0004] To address these issues, some researchers have attempted to improve model performance on complex time-series data by introducing advanced architectures such as multi-scale convolutions and Transformer modules. While these methods have improved classification accuracy to some extent, they still suffer from shortcomings such as low feature extraction efficiency and poor cross-timescale adaptability when dealing with complex data that is multi-periodic and multivariate. Therefore, there is an urgent need for a novel multivariate time-series classification method that can simultaneously consider local features and global dependencies, and possess stronger cross-domain adaptability and computational efficiency to meet the complex challenges in real-world applications. Summary of the Invention

[0005] To address the above problems, this invention provides a multivariate time series classification method and apparatus with multi-scale periodic extraction.

[0006] According to a first aspect, the present invention provides a multivariate time series classification method based on multi-scale periodic feature extraction, the method comprising the following steps:

[0007] Obtain multivariate time series data to be classified;

[0008] Based on the multivariate time series data, the time domain signal is converted into a frequency domain signal using Fast Fourier Transform (FFT) to extract the periodic features and spectral information in the time series;

[0009] The adaptive period detection module automatically selects the main periodic features from the frequency domain signal and decomposes the data into multiple periodic segments.

[0010] Multi-scale convolutional neural networks (CNNs) are used to extract local features from the periodic segments, capturing local discriminative features at different time scales;

[0011] The extracted local features are modeled using a self-attention mechanism (Transformer encoder) to extract global features across time steps and variables;

[0012] The local and global features are weighted and fused to obtain the classification features;

[0013] The time series category is identified based on the classification features.

[0014] In some embodiments, the adaptive period detection module performs frequency domain transformation on the time series signal by using fast Fourier transform, calculates a weighted combination of spectral amplitude and autocorrelation score, extracts joint features in the frequency and time domains, and automatically detects and selects multiple significant periodic features by weighting the period using the softmax function.

[0015] The multi-scale convolutional neural network includes multiple Inception modules, where each convolutional kernel has a fixed and different size (e.g., 1, 3, 5, 7, 9). It extracts features at different time scales through multi-scale convolutional operations. The weights of the convolutional kernels are dynamically adjusted based on the statistical characteristics of the input signal. Specifically, a weighting mechanism is used to weight the outputs of different convolutional kernels, enhancing the feature representation at important time scales. This effectively captures local discriminative features in time series data, thereby improving the model's generalization ability and classification accuracy.

[0016] The self-attention mechanism is implemented through a Transformer encoder, employing a multi-head self-attention mechanism to model the dependencies between local features. Through this mechanism, the model can effectively capture long-range dependencies between different time steps and learn cross-domain associations between different variables in the time series, thereby extracting global features. The multi-head mechanism further enhances the model's expressive power, making it more robust and accurate when handling complex time-series data.

[0017] According to a second aspect, the present invention provides a multivariate time series classification device, the device comprising:

[0018] The data acquisition module is configured to acquire multivariate time series data to be classified.

[0019] The frequency domain conversion module is configured to use Fast Fourier Transform (FFT) to convert the multivariate time series data from the time domain to the frequency domain signal;

[0020] The period detection module is configured to select the main periodic features from the frequency domain signal through adaptive period detection and decompose the data into multiple periodic segments;

[0021] The local feature extraction module is configured to extract local features from the periodic segment using a multi-scale convolutional neural network (CNN);

[0022] The global feature extraction module is configured to use a self-attention mechanism to model local features, thereby extracting global features across time and variables;

[0023] The feature fusion module is configured to perform a weighted fusion of the local features and the global features to obtain classification features;

[0024] The recognition module is configured to identify the category of a time series based on classification features.

[0025] In some embodiments, the periodic detection module is specifically configured as follows:

[0026] The periodic components in the frequency domain signal are analyzed by an adaptive period detection method, the spectral amplitude and autocorrelation score are calculated, and multiple significant periodic features are automatically selected by a weighted combination strategy to decompose the data into multiple periodic segments.

[0027] The local feature extraction module is specifically configured as follows:

[0028] A multi-kernel convolutional module using the Inception architecture, where each convolutional kernel has a fixed and different size, extracts features from the periodic features through multi-scale convolutional operations, thereby capturing local discriminative features at different time scales.

[0029] The feature fusion module is specifically configured as follows:

[0030] The weighted average method is used to weight and fuse local and global features based on their importance to generate the final classification features.

[0031] The identification module is specifically configured as follows:

[0032] By using fully connected layers and non-linear activation functions, the fused features are mapped to multiple categories, and the final category probabilities are output through the Softmax function for time series classification tasks.

[0033] Compared with existing technologies, the multivariate time series classification method and apparatus provided by this invention have stronger cross-domain adaptability and generalization ability, and can better identify time series with multi-periodic and multi-scale features, thereby improving classification accuracy and precision. It shows significant advantages, especially in complex data processing and cross-domain applications. Attached Figure Description

[0034] To more clearly illustrate the technical solutions of the embodiments of the present invention, the drawings used in the following description of the embodiments will be briefly introduced. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0035] Figure 1 This is a diagram illustrating the architecture of the multivariate time-series classification method provided by the present invention.

[0036] Figure 2 The structure diagram of the frequency domain feature extraction method provided by this invention is shown below;

[0037] Figure 3 This invention provides a visual comparison chart of the results of other multivariate time series classification methods on 22 datasets;

[0038] Figure 4 The flowchart of the multivariate time series classification method provided by the present invention;

[0039] Figure 5 This is a schematic diagram of a multivariate time series classification device provided in an embodiment of the present invention. Detailed Implementation

[0040] To make the objectives, technical solutions, and advantages of the embodiments of this application clearer, the technical solutions in the embodiments of this application will be described below with reference to the accompanying drawings.

[0041] To address the shortcomings of existing multivariate time series classification methods, this invention proposes a method that simultaneously considers local features and global dependencies, possessing strong generalization ability and high computational efficiency. Its basic concept mainly includes: First, designing a periodic feature extraction framework based on Fast Fourier Transform (FFT) to capture periodic components in multivariate time series, thereby improving the model's adaptability to different time scales; Second, constructing a dual-branch feature extraction structure using a multi-scale convolutional neural network (CNN) and a self-attention mechanism. The convolutional branch is used to extract local discriminative features, while the self-attention branch is used to capture long-term dependencies and global dependency information, improving the modeling ability for time series data; Third, through an adaptive cross-scale feature fusion mechanism, dynamically adjusting the fusion weights of features at each scale, thereby improving the model's ability to express multi-period, multi-scale features and enhancing the recognition effect on complex time series data.

[0042] Based on the above concepts and improvements, the multivariate time series classification method provided by this invention can effectively solve the problems of insufficient adaptability and low feature extraction efficiency of existing technologies under cross-time scale and multi-source data conditions, significantly improve classification accuracy and precision, and show significant advantages, especially in the processing of time series data with complex backgrounds and long-term dependencies.

[0043] Figure 1 The overall architecture and technical approach of this invention are given, such as... Figure 1 As shown, this invention proposes a novel multivariate time series classification framework called "Multi-Scale Periodic Feature Extraction Network" (MITNet), which consists of two key components: a multi-scale periodic feature extraction component and a global-local feature fusion component. The multi-scale periodic feature extraction component extracts periodic features from time-domain data using Fast Fourier Transform (FFT) and combines it with a multi-scale convolutional neural network (CNN) to effectively capture local features at different time scales. The global-local feature fusion component models global dependencies across time and variables through a self-attention mechanism and performs weighted fusion with local features, thereby enhancing the multi-scale representation learning of multivariate time series data and improving classification accuracy and precision. The advantage of this framework lies in its ability to simultaneously capture short-term local patterns and long-term global dependency information, and its effective solution to the problem of poor feature adaptability across time scales through an adaptive feature fusion mechanism, thus significantly improving the classification performance of complex time series data.

[0044] The multivariate time series classification method provided by this invention is described below with reference to the accompanying drawings and specific embodiments. Figure 4 The flowchart of the multivariate time series classification method provided by this invention is as follows: Figure 4 As shown, the multivariate time series classification method includes the following steps:

[0045] S1. Obtain the multivariate time series data to be classified;

[0046] The multivariate time series data to be classified can come from different fields and various sensors. This data may contain values ​​of multiple variables at different time periods, such as biosignal data, meteorological data, and traffic flow data. The multivariate time series classification method provided by this invention has good cross-domain adaptability, can effectively handle data from different sources, reduce differences caused by factors such as sensor type, acquisition time, and data preprocessing, and ensure that the model can adapt to different data sources and maintain high classification accuracy.

[0047] After obtaining the multivariate time series data to be classified, the data can be preprocessed, including operations such as denoising, standardization, and missing value imputation, in order to improve data quality and model robustness.

[0048] S2. Based on the multivariate time series data, use Fast Fourier Transform to extract periodic features;

[0049] It is important to understand that the periodicity of multivariate time series data in this invention refers to the periodic variation of each variable in the time series over time, typically determined by the frequency components of the signal. The Fast Fourier Transform (FFT) can effectively convert time-domain signals into frequency-domain signals, thereby revealing periodic components and spectral information. Periodic features at different time scales can be extracted and further analyzed using FFT, providing strong support for subsequent multi-scale feature extraction and classification.

[0050] This step corresponds to the multi-scale periodic feature extraction component in the MITNet framework. For example... Figure 1 As shown, the FFT module is used to extract periodic features from raw multivariate time series data and transform these features into frequency domain information. By converting the data to the frequency domain, the model can capture periodic variations at different frequencies, which is of great significance for further extracting fine-grained local features and modeling global dependencies.

[0051] After extracting periodic features using FFT, the next step combines a multi-scale convolutional neural network (CNN) and a self-attention mechanism to further extract local and global features from periodic signals at different time scales. Through these steps, the model can comprehensively capture the contextual and local patterns of the time series, improving classification accuracy and robustness.

[0052] S3. Use a multi-scale convolutional neural network to extract local features from the frequency domain signal;

[0053] like Figure 1As shown on the right, the model employs a spatial-frequency dual-branch learning structure, where the frequency branch uses a multi-scale convolutional neural network (CNN) to extract local features from the frequency domain signal. The goal of this part is to capture the details of periodic changes in the time series, especially the fine-grained features for multi-period data.

[0054] In this step, the frequency domain features extracted in S2 are input into a multi-scale convolutional neural network model. By using multiple convolutional operations with different kernel sizes, the model can effectively extract local features at multiple time scales. These local features reflect the variation patterns at various periodic scales in the frequency domain signal, accurately capturing short-term fluctuations and micro-features in the time series.

[0055] The advantage of multi-scale convolutional neural networks lies in their ability to adapt to changes at different time scales, capturing feature information from local to global perspectives, making them particularly suitable for processing complex and highly variable time-series data. Through the design of parallel convolutional layers, the model can extract local discriminative features helpful for classification at different scales and ensure that these features are effectively transmitted and fused in subsequent networks. Ultimately, these extracted local features provide a foundation for global dependency modeling, enhancing the model's performance in multivariate time series classification.

[0056] S4. Use self-attention mechanism to model global dependencies across time and variables, and extract global features;

[0057] like Figure 1 As shown, in the multi-scale periodic feature extraction framework of this invention, the global feature extraction part adopts a self-attention mechanism, which aims to model the global dependencies across time and variables in the time series. This mechanism can capture the long-range dependencies between different time steps and between different variables in the time series, and solves the limitations that traditional convolutional networks may face when processing global information.

[0058] This step inputs the local features extracted in S3 into the self-attention module, which automatically learns the most important global features by calculating the relationships between different time steps or variables in the sequence. The self-attention mechanism effectively captures global dependencies by calculating the correlation between each pair of time steps or variables and dynamically adjusting the weight of each feature based on these correlations. This mechanism can not only model long-term dependencies in time series but also flexibly capture complex interactions between different variables, enhancing the model's ability to perceive dependencies across time scales and variables.

[0059] Unlike traditional fully connected or convolutional networks, self-attention mechanisms compute the relationships between variables at each time step in parallel, resulting in higher computational efficiency and stronger modeling capabilities, especially when dealing with high-dimensional and multivariate time series data. In this way, the model can learn important dependencies globally, further enhancing its ability to identify complex time series patterns.

[0060] By combining local feature extraction and global dependency modeling, the framework of this invention can simultaneously capture fine-grained patterns and long-term global dependencies in time series, thereby greatly improving classification accuracy and the model's generalization ability.

[0061] In some embodiments, after converting the local features into global features, an adaptive attention mechanism can be used to further extract global dependencies.

[0062] This mechanism enhances the learning ability of global features by simulating the brain's neural responses, automatically capturing long-range dependencies between different time steps and variables, further improving the model's ability to recognize global patterns. The adaptive attention mechanism strengthens important global features by dynamically adjusting the weights of features at each time step or variable, while improving the model's adaptability and computational efficiency, demonstrating excellent performance, especially when processing multi-period, multi-scale data.

[0063] This mechanism adaptively adjusts the weighted input features based on the relative importance of global information, ensuring that the model can capture key global dependencies in complex time-series data. This adaptive attention mechanism not only enhances the ability to extract global features but also improves the model's stability and computational efficiency when dealing with high-dimensional and multi-source data.

[0064] S5. The extracted local and global features are weighted and fused, and then input into the classifier to generate the classification result;

[0065] like Figure 1 As shown in the middle section, this step corresponds to the operation of the feature fusion module. This module weightedly fuses local features extracted from multi-scale convolutional neural networks (CNNs) and self-attention mechanisms with global features to form a comprehensive feature representation. This fusion operation allows local features and global dependency information to be effectively combined in the same feature space, thereby improving the model's ability to classify time series data.

[0066] In this process, local and global features are weighted to ensure that the model can dynamically adjust the weights of each feature based on important information in the data. Through this weighted fusion, the model can better capture the interaction between short-term local patterns and long-term global dependencies, further improving its ability to identify complex time-series data.

[0067] The weighted and fused features are input into a classifier for final classification prediction. Based on these fused features, the classifier generates classification results and classifies the input time series. By fusing features from different sources, the method of this invention enhances the perception of global patterns while preserving local details, thereby significantly improving classification accuracy and robustness.

[0068] In some embodiments, the extracted local and global features are weighted and fused before being input into a classifier for classification.

[0069] This embodiment achieves weighted fusion of local and global features through an adaptive feature fusion module. The fused feature representation can simultaneously retain short-term local patterns and long-term global dependency information, ensuring that the model can better handle complex time series data.

[0070] In this embodiment, local features are extracted using a multi-scale convolutional neural network (CNN), while global features are modeled using a self-attention mechanism to capture long-range dependencies across time steps and variables. The weighted fusion process is dynamically adjusted based on the contribution of each feature, thereby maximizing the effectiveness of different features in the classification task. The weighted fused features are then fed into a classifier, which uses these features to generate the final classification result, completing the classification task of multivariate time series data.

[0071] This embodiment fully demonstrates the advantages of fusing local and global features, especially when dealing with tasks with complex dependencies and variable data patterns, which can significantly improve classification accuracy and generalization ability.

[0072] The above describes the steps of the multivariate time series classification method provided by this invention.

[0073] To verify the effectiveness of the multivariate time series classification method provided in this invention, the following experimental results are presented:

[0074] We used 21 multivariate datasets from the UEA benchmark dataset for comparative experiments, and also used a PCR standard sample dataset collected from a real-time quantitative PCR instrument. Details of all datasets are shown in Table 1. The datasets are mainly divided into the following categories: Human Activity Recognition (HAR), Motion, Electrocardiogram (ECG), Electroencephalogram (EEG / MEG), Audio Spectrum (AS), and Other.

[0075] Table 1. Twenty-one UEA datasets and one PCR standard sample dataset collected from a real-time quantitative PCR instrument used in the experiment.

[0076]

[0077] Benchmark results for all baseline methods are derived from their respective publications, ensuring consistent training parameters are used in the comparisons. Our proposed model was implemented on a server running Ubuntu 20.04.3 LTS equipped with one NVIDIA GeForce RTX 3090 GPU. All models were trained and evaluated on the same hardware environment to ensure fairness and comparability of experimental results. We used standard deep learning frameworks to implement all models, ensuring code scalability and reproducibility.

[0078] To ensure the stability of model training, we performed hyperparameter tuning on all methods and used cross-validation or early stopping strategies to prevent overfitting. Each model's training process involved multiple experiments, and the best results were reported.

[0079] Performance was evaluated by calculating accuracy, average accuracy, and the number of datasets on which the classifier achieved its best accuracy. Accuracy refers to the correct classification result of the model on each dataset, and average accuracy is the mean of the accuracy across all datasets. Furthermore, we evaluated the generalization ability of each method by comparing its performance across all datasets. Performance metrics for each experiment were calculated on standard dataset splits and a test set to ensure the validity of the results.

[0080] All experiments were conducted with fixed training and test set partitions. The same data preprocessing steps were used. These measures aim to minimize the influence of external factors on the experimental results and ensure the impartiality and scientific rigor of the comparisons.

[0081] This group of experiments uses 21 equal-length datasets provided by UEA and one PCR standard sample dataset collected from a real-time quantitative PCR instrument to evaluate the performance of our proposed model. For comprehensive comparison, we selected several strong baseline methods in the field of multivariate time series classification (MTSC) for comparison, including bag-of-words symbolic feature method (WEASEL+MUSE), prototype-metric attention prototype network (TapNet), full-scale one-dimensional convolutional network (OS-CNN), time-varying pattern modeling TimesNet, multi-feature fusion network (MF-Net), and dense knowledge-aware network (DKN).

[0082] To avoid bias caused by implementation differences, we prioritized the author's open-source implementation or authoritative reproduction version whenever possible, and used a consistent training / validation process and early stopping strategy on all datasets. For distance-nearest neighbor methods (ED / DTW) that do not require training, we standardized them at the channel level and calculated the distance according to the original definition, as is the standard practice. The experimental evaluation results are shown in Table 2.

[0083] Table 2 compares the performance of recent state-of-the-art MTSC-specific models on 21 UEA datasets and 1 dataset of standard PCR samples collected from a real-time quantitative PCR instrument.

[0084]

[0085] like Figure 2 As shown in the table, "AVG" represents the average accuracy of the corresponding model across 22 datasets, and "Win" represents the number of datasets on which the corresponding model achieved its best accuracy. The results in the table demonstrate that our method achieves the best classification accuracy on most datasets. On all 21 UEA and PCR datasets, our method excels in both average accuracy and the number of datasets on which it achieves best accuracy. Specifically, our method performs exceptionally well on several challenging datasets, such as ArticularWordRecognition, Cricke, and HandMovementDirection, demonstrating its strong generalization ability across various tasks.

[0086] Compared to other methods, our method outperforms traditional baseline methods on multiple datasets. For example, DKN performs well on the ArticularyWordRecognition and HandMovementDirection datasets, but still cannot match our method. Meanwhile, methods such as TimesNet and ConvTran perform well on BasicMotions and NaTOPS, but their results on other datasets are mediocre.

[0087] In particular, on some challenging datasets, such as Atrial Fibrillation and HandMovementDirection, our method significantly improves accuracy, achieving 0.467 and 0.676, respectively. This demonstrates that our method can better capture temporal dependencies and multivariate features when dealing with complex time series data, resulting in more representative model representations.

[0088] Furthermore, while the WEASEL+MUSE and TapNet methods achieved an accuracy of 1.000 on the BasicMotions dataset, they exhibited significant fluctuations on other datasets, further indicating that they may not generalize well to different time series tasks. Our method, however, maintained high performance across multiple datasets, demonstrating its strong adaptability and robustness.

[0089] Figure 3 shows the critical difference (CD) plot based on the posterior Nemenyi test, used to compare the performance of eight different implementation methods on 21 UEA datasets and one dataset of standard PCR samples collected from a real-time quantitative PCR instrument. The CD plot reveals the relative performance of different methods by comparing the mean rank and significance of the differences. Short lines in the figure indicate no significant performance differences between methods, while long lines indicate significant differences between methods.

[0090] from Figure 3 As can be seen, our method ranks first, indicating that it performs best overall among all eight methods. The CD plot further demonstrates the superiority of our method on multiple datasets, especially its significant advantages in accuracy and classification precision.

[0091] To further verify the performance differences between Ours and the core baselines, we conducted an analysis using the paired Wilcoxon signed-rank test: using the accuracy of 22 datasets as samples, we examined the performance differences between Ours and DKN, and between Ours and ConvTran. The results showed that the p-value between Ours and DKN was 0.032, and the p-value between Ours and ConvTran was 0.028, both less than the significance level of 0.05, indicating that Ours statistically significantly outperformed these two core baselines. Further analysis of the baseline limitations revealed that while DKN can model multivariate interactions, it lacks the ability to capture frequency domain periodic features, resulting in an accuracy of only 0.962 on the PCR standard sample dataset; ConvTran relies on a fixed convolution scale for feature extraction, making it unsuitable for multi-scale temporal patterns, achieving an accuracy of only 0.805 on the PEMS-SF traffic flow dataset; while Ours, by fusing fine-grained temporal features, frequency domain periodic features, and spatial correlation information, effectively compensates for the shortcomings of single-feature modeling, ultimately achieving superior and stable classification performance.

[0092] The CD plot in Figure 3 not only demonstrates the leading position of our method among all eight methods, but also further validates its superiority in accuracy and classification precision. The CD plot provides strong visual support for the performance of our method, further demonstrating its powerful performance and broad adaptability when processing various time series data.

[0093] Overall, the data in the tables and the CD plot demonstrate the superiority of our proposed model on 21 UEA and PCR datasets, particularly in terms of overall performance, accuracy, and dataset classification precision, where it outperforms other existing models. Our method provides a better understanding and processing capability for variable time series data through more efficient feature extraction and multi-channel information fusion.

[0094] As can be seen from the above, compared with the prior art, the multivariate time series classification method provided by the present invention has the following beneficial effects:

[0095] 1. Enhanced Cross-Domain Adaptability: By introducing multi-scale periodic feature extraction and adaptive periodic detection modules, this invention effectively reduces the differences in periodicity, frequency, and variable dependencies among time series data from different sources and at different time scales, achieving consistent processing of cross-domain data. Compared with existing methods, this invention maintains high classification accuracy when dealing with data from different periods and sources, significantly improving the model's generalization ability and stability.

[0096] 2. Enhanced ability to recognize multi-periodic and small-scale features: This invention designs a dual-branch feature extraction structure with a multi-scale convolutional neural network and a self-attention mechanism, capable of simultaneously extracting local and global features. The multi-scale convolutional branch effectively captures local patterns at different time scales, while the self-attention mechanism models global dependencies. This design enhances the model's ability to recognize multi-periodic, complex patterns, and fine-grained features, especially exhibiting higher accuracy when processing small-scale features, thus overcoming the shortcomings of traditional methods in recognizing small features and details.

[0097] 3. Significantly Improved Detection Performance: By weighted fusion of local and global features, this invention achieves joint modeling of multi-scale periodic features, outperforming existing methods on multiple evaluation metrics. Experimental results show that, compared to existing deep learning models, this invention significantly improves classification accuracy in multi-period time series classification tasks, exhibiting higher robustness and accuracy, especially in complex environments. Particularly in tests involving cross-timescale and multivariate data, the accuracy improvement can reach over 20%, demonstrating higher practical value and stability.

[0098] According to another aspect, the present invention also provides a multivariate time series classification device, such as... Figure 5 As shown, the device includes:

[0099] Data acquisition module 801 is configured to acquire multivariate time series data to be classified;

[0100] The frequency domain conversion module 802 is configured to use Fast Fourier Transform (FFT) to convert the time domain signal into a frequency domain signal and extract periodic features;

[0101] The local feature extraction module 803 is configured to extract local features from the frequency domain signal using a multi-scale convolutional neural network (CNN);

[0102] The global feature extraction module 804 is configured to use a self-attention mechanism to model global dependencies across time and variables and extract global features;

[0103] The feature fusion module 805 is configured to perform weighted fusion of the local features and the global features, and input the fusion into the classifier to generate a classification result;

[0104] The identification module 806 is configured to identify and predict the category of a multivariate time series based on the classification results.

[0105] In some embodiments, the frequency domain feature extraction module is specifically configured to: convert the time domain signal into a frequency domain signal using a Fast Fourier Transform (FFT); and extract features from the frequency domain signal using a multi-scale convolution module to obtain the extracted frequency domain features.

[0106] The feature fusion module is specifically configured to input the local features and the extracted frequency domain features into the classifier to generate a classification result.

[0107] In some embodiments, the network structure of the convolutional neural network includes multiple self-attention modules configured to capture global dependencies through a Transformer mechanism.

[0108] It should be noted that the above-described device can execute the multivariate time series classification method provided by this invention. Furthermore, the functions of each part of the device are described in the foregoing introduction to the method, and will not be repeated here.

[0109] In the description of the embodiments of this application, the words "exemplary," "for example," or "for instance" are used to indicate examples, illustrations, or explanations. Any embodiment or design described as "exemplary," "for example," or "for instance" in the embodiments of this application should not be construed as being more preferred or advantageous than other embodiments or designs. Specifically, the use of the words "exemplary," "for example," or "for instance" is intended to present the relevant concepts in a specific manner.

[0110] In the description of the embodiments of this application, the term "and / or" is merely a description of the relationship between related objects, indicating that three relationships can exist. For example, A and / or B can represent: A existing alone, B existing alone, and A and B existing simultaneously. Furthermore, unless otherwise stated, the term "multiple" means two or more.

[0111] Furthermore, the terms "first" and "second" are used for descriptive purposes only and should not be construed as indicating or implying relative importance or implicitly specifying the indicated technical features. Thus, a feature defined as "first" or "second" may explicitly or implicitly include one or more of that feature. The terms "comprising," "including," "having," and their variations all mean "including but not limited to," unless otherwise specifically emphasized.

[0112] The specific embodiments described above further illustrate the purpose, technical solution, and beneficial effects of the present invention. It should be understood that the above description is only a specific embodiment of the present invention and is not intended to limit the scope of protection of the present invention. Any modifications, equivalent substitutions, improvements, etc., made on the basis of the technical solution of the present invention should be included within the scope of protection of the present invention.

Claims

1. A multi-scale periodicity-extracted multivariate time series classification method, characterized in that, The method comprises the following steps: Obtaining multivariate time series data to be classified; Converting time domain signals to frequency domain signals based on the time series data using fast Fourier transform to extract periodic features and spectral information in the time series; automatically selecting main periodic features from the frequency domain signals through an adaptive period detection module to decompose the data into multiple period segments; extracting local features from the period segments using a multi-scale convolutional neural network to capture local discriminative features at different time scales; and modeling the local features using a self-attention mechanism to extract global features across time and variables. Weighted fusion of the local features and the global features to obtain classification features; and identifying the category of the time series according to the classification features.

2. The method of claim 1, wherein, The adaptive period detection module converts time series signals to the frequency domain by using fast Fourier transform, calculates the weighted combination of spectral amplitude and autocorrelation score, extracts joint features in the frequency domain and the time domain, and weights the periods through a softmax function to automatically detect and select multiple significant periodic features. The multi-scale convolutional neural network comprises multiple conditional convolution modules, each of which uses a fixed-size convolution kernel and dynamically adjusts the weights of the convolution kernels according to the statistical characteristics of the input signal through a weighting mechanism to adapt to the feature extraction requirements at different time scales, thereby effectively capturing the local features of the time series. The self-attention mechanism is implemented through a Transformer encoder and uses a multi-head self-attention module to model the dependency between local features, thereby extracting global features across time steps and variables.

3. A multi-scale periodicity-extracted multivariate time series classification apparatus characterized by, The device comprises: A data acquisition module configured to obtain multivariate time series data to be classified; A frequency domain conversion module configured to convert the multivariate time series data from the time domain to frequency domain signals using fast Fourier transform; A period detection module configured to select main periodic features from the frequency domain signals through adaptive period detection and decompose the data into multiple period segments; A local feature extraction module configured to extract local features from the period segments using a multi-scale convolutional neural network; A global feature extraction module configured to model the local features using a self-attention mechanism to extract global features across time and variables; A feature fusion module configured to weightedly fuse the local features and the global features to obtain classification features; An identification module configured to identify the category of the time series according to the classification features.

4. The apparatus of claim 3, wherein, The period detection module is specifically configured to: Analyze the periodic components in the frequency domain signals through an adaptive period detection method, calculate the spectral amplitude and autocorrelation score, and automatically select multiple significant periodic features through a weighted combination strategy to decompose the data into multiple period segments; The local feature extraction module is specifically configured to: Use a multi-kernel convolution module with an Inception structure, where each convolution kernel has a fixed and different size, and the period features are extracted through multi-scale convolution operations to capture local discriminative features at different time scales; The feature fusion module is specifically configured to: The local features and the global features are fused by a weighted average method according to the importance of the local features and the global features, so as to generate final classification features; The identification module is specifically configured to: The fused features are mapped to multiple categories by using a full connection layer and a nonlinear activation function, and the final category probability is output by a Softmax function, so as to perform a time series classification task.

Images