Multivariable time sequence classification method and system based on spectrum convolutional network
By using a spectral convolutional network to perform time-spectrum transformation and multi-scale feature extraction on multivariate time series data, the problem of capturing frequency domain characteristics and multi-scale features in existing technologies is solved, and more efficient multivariate time series classification is achieved.
Patent Information
- Application Number
- CN202511284416.5
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2025-09-09
- Publication Date
- 2025-11-21
AI Technical Summary
Existing multivariate time series classification methods struggle to effectively capture frequency domain characteristics, periodicity, and multi-scale features, resulting in limited classification accuracy.
A method based on spectral convolutional networks is used to perform time-spectral transformation, multi-scale feature extraction, and deep learning classification on one-dimensional multivariate time series data, including the integration of enhanced discrete Fourier transform, spectral energy priority strategy, and multi-scale convolutional structure.
It significantly improves the accuracy and robustness of multivariate time series classification, enabling it to more comprehensively capture the complex structure and dynamic features of time series and improve classification efficiency.
Smart Images
Figure CN120995292A_ABST
Abstract
Description
Technical Field
[0001] This invention relates to the field of time series data processing technology, and in particular to a multivariate time series classification method and system based on deep learning. Background Technology
[0002] With the development of information technology, multivariate time series data has been widely used in various fields such as financial market analysis, medical and health monitoring, and industrial process control. Multivariate time series data consists of multiple simultaneously recorded correlated or independent time series, each capturing the time variation of a specific variable. Accurately classifying multivariate time series data is crucial for understanding its underlying patterns, but its inherent complex correlations and rich information make multivariate time series data classification a challenging task.
[0003] While existing multivariate time series classification methods have made some progress, they typically rely on one-dimensional time series analysis, which limits their ability to capture complex temporal dynamics and multi-scale features. Many existing models also struggle with handling frequency domain features.
[0004] Some studies have attempted to convert one-dimensional time series data into two-dimensional representations to extract features more effectively. However, these methods may fall short in capturing the periodicity of time series and fully reflecting their underlying spectral characteristics, and may struggle to adapt to the complexity of the data, especially being relatively inefficient in multivariate time series classification tasks.
[0005] Furthermore, although multi-scale convolutional structures and spectral analysis techniques have shown great potential in time series classification, most models tend to use these techniques independently and fail to integrate them effectively, which may lead to limited analytical perspectives or missing features.
[0006] Therefore, how to effectively extract and integrate the time-domain, frequency-domain, and multi-scale features of time series to improve the accuracy and robustness of multivariate time series classification is a technical problem that urgently needs to be solved. Summary of the Invention
[0007] The purpose of this invention is to address the shortcomings of existing technologies in multivariate time series classification, particularly the difficulty in effectively capturing frequency domain characteristics, periodicity, and multi-scale features, which leads to limited classification accuracy. This invention proposes a multivariate time series classification method and system based on spectral convolutional networks. By performing innovative time-spectral transformation, multi-scale feature extraction, and deep learning classification on the input one-dimensional multivariate time series data, this invention aims to significantly improve the accuracy and efficiency of classification.
[0008] This invention is achieved through the following technical solution:
[0009] The first aspect of this invention proposes a multivariate time series classification method based on spectral convolutional networks. This method is used by computing devices (e.g., servers, embedded devices, etc.) to process and classify input one-dimensional multivariate time series data, specifically including the following steps:
[0010] Step S1. Perform time-spectrum processing on the one-dimensional multivariate time series data to convert the one-dimensional multivariate time series data into a comprehensive one-dimensional feature representation containing frequency domain and periodic structure information;
[0011] Step S2. Perform multi-scale feature extraction on the comprehensive one-dimensional feature representation. Process the comprehensive one-dimensional feature representation through parallel convolutional layers with convolutional kernels of different sizes to capture features at different time scales and obtain multi-scale fused features.
[0012] Step S3. Classify the data according to the multi-scale fusion features to obtain the classification result of the one-dimensional multivariate time series data.
[0013] In a preferred embodiment, step S1 specifically includes the following sub-steps:
[0014] Step S11. Perform enhanced discrete Fourier transform to transform the one-dimensional multivariate time series data to the frequency domain.
[0015] Step S12. Execute the spectrum energy priority strategy and construct a two-dimensional time series representation. Execute the spectrum energy priority strategy, select several key frequency components according to the energy magnitude of the frequency domain data, and reshape the one-dimensional multivariate time series data into at least one two-dimensional time series tensor based on the period length corresponding to the key frequency components.
[0016] Step S13. Perform feature extraction and transformation on each two-dimensional time series tensor generated in steps S1-2-4, extract two-dimensional features from the at least one two-dimensional time series tensor, and convert it into at least one one-dimensional feature vector;
[0017] Step S14. Perform feature aggregation and residual connection, and perform weighted aggregation on the at least one-dimensional feature vector to form the comprehensive one-dimensional feature representation.
[0018] In a preferred embodiment, step S11, the enhanced discrete Fourier transform includes standardizing the spectral data, and its calculation formula is as follows:
[0019]
[0020] Among them, X freq The spectrum data is obtained after Fast Fourier Transform, where μ is the mean amplitude, σ is the standard deviation of the amplitude, and X is the mean amplitude. freq_normThis is the standardized spectrum data.
[0021] In a preferred embodiment, in step S12, a spectrum energy priority strategy is implemented and a two-dimensional time series representation is constructed, the calculation formula of which is as follows:
[0022]
[0023] Among them, E ω L represents the energy corresponding to the frequency point ω. i To be based on the selected key frequency ω i The calculated period length, T, is the length of the original one-dimensional time series data.
[0024] In a preferred embodiment, step S14 involves weighted aggregation of at least one-dimensional feature vectors, and the calculation formula is as follows:
[0025]
[0026] in, The aggregated one-dimensional output features These are the normalized weighting coefficients calculated based on frequency energy. This is a one-dimensional feature vector obtained after two-dimensional feature extraction.
[0027] In a preferred embodiment, the multi-scale feature extraction in step S2 is accomplished by at least two parallel convolutional layers configured with kernels of different sizes, and a feature concatenation unit for concatenating the outputs of the parallel convolutional layers, specifically including the following sub-steps:
[0028] Step S21. The aggregated one-dimensional features output from step S1 are first processed by layer normalization, and then input into the multi-scale convolution module.
[0029] Step S22. Used to enable the input features to pass through multiple convolutional layers with different sized kernels in parallel within the multi-scale convolutional module;
[0030] Step S23. Use dynamic masking technology in each convolutional layer;
[0031] Step S24. This step is used to process the output of each parallel convolutional branch through batch normalization and a ReLU activation function.
[0032] Step S25. This step concatenates the features obtained from all parallel convolutional branches along the channel dimension. The concatenated features are then subjected to batch normalization and a ReLU activation function to obtain the final multi-scale fused features.
[0033] In a preferred embodiment, step S3 specifically includes the following sub-steps:
[0034] Step S31. Used for processing the multi-scale fusion features output in step S2 Perform layer normalization to obtain normalized feature X ln ;
[0035] Step S32. Used to normalize feature X ln It passes through a GELU activation function and optionally a Dropout layer in sequence;
[0036] Step S33. This step inputs the processed features into a fully connected linear layer and outputs the final classification prediction's original logical value y′.
[0037] In a preferred embodiment, step S4 specifically includes the following sub-steps:
[0038] Step S41. Used to define the loss function, for example, using the cross-entropy loss function to calculate the difference between the model's predicted output and the true target category label;
[0039] Step S42. Used to optimize model parameters, for example, by selecting the Adam optimizer and iteratively adjusting the learnable parameters in the network model through the backpropagation algorithm based on the calculated loss function value;
[0040] Step S43. Used to employ specific training strategies to improve the final performance and generalization ability of the model, such as setting the number of training epochs, learning rate, and implementing an early stopping mechanism.
[0041] The second aspect of this invention proposes a multivariate time series classification system based on a spectral convolutional network. This system is used to execute the method described in the first aspect of this invention. The system includes a data receiving and initialization module, a time-spectral processing module, a multi-scale feature extraction module, a classification output module, and a model training module.
[0042] The data receiving and initialization module is used to receive externally input one-dimensional multivariate time series data and can perform preliminary data formatting or preparation work.
[0043] The time-spectrum processing module is responsible for performing enhanced discrete Fourier transform on the input one-dimensional multivariate time series data to convert it to the frequency domain. Next, it implements a spectrum energy priority strategy, including calculating the energy of each frequency, selecting the top-k key frequencies, calculating the corresponding period length, and finally reshaping the original one-dimensional time series data into k two-dimensional time series tensors. Subsequently, the two-dimensional convolutional neural network submodule within this time-spectrum processing module extracts features from each two-dimensional time series tensor and converts them into one-dimensional feature vectors through pooling and reshaping operations. Finally, these one-dimensional feature vectors are weighted and aggregated according to the weights calculated for each frequency energy, and can be combined with the module input through residual connections.
[0044] The multi-scale feature extraction module receives one-dimensional features output by the time-spectrum processing module (or its last cascaded unit) and processed by layer normalization. It contains multiple parallel convolutional layers with different sized kernels to capture the temporal dependencies of different scale ranges from the input features. These convolutional layers use dynamic masking technology. The output of each parallel branch is batch normalized and ReLU activated. Then the features of all branches are concatenated and batch normalized and ReLU activated again to form multi-scale fused features.
[0045] The classification output module first performs layer normalization on the features output by the multi-scale feature extraction module; then it performs a nonlinear transformation using the GELU activation function; finally, it maps the processed features to a predefined category space through a fully connected linear classifier and outputs the classification result.
[0046] The model training module includes a loss function calculation unit and a parameter optimization unit, and can integrate a training strategy control unit. This module uses the training dataset to adjust and optimize the learnable parameters in the system.
[0047] The time-spectrum processing module is used to execute step S1 as described in claim 1;
[0048] The multi-scale feature extraction module is used to perform step S2 as described in claim 1;
[0049] The classification output module is used to perform step S3 as described in claim 1.
[0050] In a preferred embodiment, the time-spectrum processing module includes an enhanced discrete Fourier transform unit, a two-dimensional representation construction unit, a two-dimensional feature extraction and transformation unit, and a feature aggregation unit.
[0051] The enhanced discrete Fourier transform unit is used to apply the enhanced discrete Fourier transform to the one-dimensional multivariate time series data and transform it to the frequency domain.
[0052] The two-dimensional representation construction unit is used to execute a spectrum energy priority strategy, select several key frequency components according to the energy magnitude of the frequency domain data, and reshape the one-dimensional multivariate time series data into at least one two-dimensional time series tensor based on the period length corresponding to the key frequency components.
[0053] The two-dimensional feature extraction and conversion unit is used to extract two-dimensional features from the at least one two-dimensional temporal tensor and convert it into at least one one-dimensional feature vector.
[0054] The feature aggregation unit is used to perform weighted aggregation on the at least one-dimensional feature vector to form the comprehensive one-dimensional feature representation.
[0055] The two-dimensional feature extraction and multi-scale feature extraction in the time-spectrum processing module and the multi-scale feature extraction module are accomplished by a convolutional neural network module containing depthwise separable convolutions.
[0056] Beneficial Effects: This invention aims to provide a more insightful and discriminative classification solution for multivariate time series by employing specific spectral transformation, energy-priority two-dimensional representation construction, and deep fusion of multi-scale convolutional structures. Compared with existing technologies, the multivariate time series classification method and system based on spectral convolutional networks provided by this invention have significant beneficial effects, specifically in the following three aspects;
[0057] First, through innovative time-spectrum processing mechanisms, including enhanced discrete Fourier transform and spectral energy priority strategies, it is possible to effectively reveal and utilize the hidden periodicity and key frequency components in time series data, which is crucial for understanding the intrinsic structure of time series.
[0058] Secondly, it cleverly transforms one-dimensional time series into two-dimensional time series tensors and combines them with advanced two-dimensional convolutional networks for feature extraction, thereby capturing richer structured information and complex dependencies, such as intra-cycle patterns and cross-cycle patterns.
[0059] Third, the introduced multi-scale convolution module can process features at different time scales in parallel, effectively capturing a variety of dynamics from short-term subtle fluctuations to long-term evolution trends, making the model's understanding of time series more comprehensive.
[0060] Fourth, the overall network architecture design emphasizes the effective integration and transmission of features, which helps improve learning efficiency and model performance. Ultimately, these design features work together to enable this invention to achieve classification accuracy superior to existing state-of-the-art models on multiple publicly available multivariate time series classification datasets, demonstrating stronger feature extraction and classification capabilities, and providing a more powerful and efficient technical means for complex multivariate time series analysis and applications. Attached Figure Description
[0061] Figure 1 This is a schematic diagram of the overall process of the spectral convolutional network method proposed in this invention.
[0062] Figure 2 This is a schematic diagram of the overall architecture of the spectral convolutional network of the present invention, including the composition and connection relationship of the main processing modules such as the time-spectral processing module and the multi-scale convolution module.
[0063] Figure 3 This is a detailed flowchart illustrating the two-dimensional feature extraction and conversion process within the time-spectrum processing module of this invention.
[0064] Figure 4 This is a detailed flowchart of the multi-scale convolution module in this invention. Detailed Implementation
[0065] The embodiments of the present invention will be described in detail below with reference to the accompanying drawings: These embodiments are implemented based on the technical solution of the present invention, and provide detailed implementation methods and specific operation processes, but the protection scope of the present invention is not limited to the following embodiments.
[0066] like Figure 1 As shown, a multivariate time series classification method based on spectral convolutional networks includes the following steps:
[0067] Step S1: Perform time-spectral processing on the input one-dimensional multivariate time series data. This includes the following sub-steps:
[0068] Step S1-1: Perform the enhanced discrete Fourier transform. This step first processes the one-dimensional time series data X. 1D The real-number fast Fourier transform is applied to convert it to the spectral domain, yielding the spectral data X. freq The result is then normalized to eliminate the influence of sequence length T. The calculation formula is as follows:
[0069]
[0070] Next, the obtained spectrum data X is calculated. freq Amplitude |X freq，i The mean μ and standard deviation σ of | (where i is the frequency index and n is the total number of frequency points) are calculated using the following formulas:
[0071]
[0072] Finally, the calculated mean μ and standard deviation σ are used to analyze the spectral data X. freq Standardization processing is performed to obtain standardized spectrum data X. freq_normThis provides a standardized feature scale, helping the model to effectively extract information and avoid interference from outliers or extreme values in the original amplitude. The calculation formula is as follows:
[0073]
[0074] Step S1-2 involves implementing a spectrum energy priority strategy and constructing a two-dimensional time series representation. This step begins with the unnormalized spectrum data X from step S1-1. freq The amplitude is calculated, and the energy E corresponding to each frequency point ω is calculated. ω To preserve the original proportions and ensure accurate identification of the frequencies that have the greatest impact on time series variations, the calculation formula is as follows:
[0075]
[0076] Then, based on the calculated energy E at each frequency ω Sort the data and select the top data with the highest energy. k Key frequencies This allows us to focus on the frequency components that have the greatest impact on changes in time series.
[0077] In this embodiment, the value of k can be set according to the specific application scenario and data characteristics. Then, for each selected key frequency ω... i (where i∈{1,...,k}) Calculate its corresponding period length L i The calculation formula is as follows: Where T is the length of the original one-dimensional time series data. Finally, based on the calculated key frequencies ω... i The corresponding period length L i The original one-dimensional time series data X 1D Remodeled into k two-dimensional temporal tensors
[0078] In the embodiment, the two-dimensional temporal tensor The dimension can be set to (ω) i L i ) or (L i ω i Such a two-dimensional structure can reveal the change patterns of time series within a period and the change patterns across different periods.
[0079] Steps S1-3 involve feature extraction and transformation of the two-dimensional temporal tensors generated in step S1-2. This process can be referenced... Figure 3 The diagram shows the internal flow of the time-spectrum processing module. This process processes each two-dimensional time-series tensor obtained in steps S1-2. Each input is fed into a two-dimensional convolutional neural network module for two-dimensional feature extraction, resulting in a two-dimensional feature output. like Figure 3 As shown.
[0080] In this embodiment, the ConvNeXt module can be configured to include an initial convolutional layer to expand the number of input channels to the number of intermediate channel layers, followed by a GELU activation function to enhance feature processing capabilities, and finally another convolutional layer to compress the number of channels from the intermediate layer back to the original or target dimension, thereby completing the integration and refinement of information. The ConvNeXt module utilizes depthwise separable convolution technology to process spatial relationships (depthwise convolution) and inter-channel relationships (pointwise convolution) separately. Then, the extracted two-dimensional features are processed... An adaptive average pooling 2D operation is applied to adjust the feature map dimension, ensuring that the output dimension is aligned with the target length, and then it is reshaped into the corresponding 1D feature vector. The calculation formula can be expressed as:
[0081]
[0082] Step S1-4 involves feature aggregation and residual connection. This step first utilizes the energy corresponding to each selected key frequency calculated in step S1-2. The weight coefficients are generated by processing using the Softmax function. These weights reflect the relative importance of each selected frequency (and its corresponding period), and the calculation formula is shown below:
[0083]
[0084] Here, l-1 represents the energy value from the previous layer or the initial calculation. Next, the k one-dimensional feature vectors obtained in steps S1-3 are... Based on the corresponding weighting coefficients obtained in this step We perform a weighted summation to obtain the one-dimensional output feature of the current time-frequency spectrum processing module at layer l after aggregation. The calculation formula is as follows:
[0085]
[0086] Finally, if multiple time-spectrum processing modules are connected in series, the output features obtained by the current module in the l-th layer will be processed. Through residual connection and its input features The output (i.e., the output or original input of the previous layer module) is added together to form the final output of the l-th layer module, in order to enhance information flow and gradient propagation. The calculation formula is as follows:
[0087]
[0088] in, The aggregated features obtained after processing from step S1-1 to this step are as follows: Figure 4 As shown.
[0089] Step S2: Perform multi-scale feature extraction on the one-dimensional features output from Step S1. This process is mainly accomplished through a multi-scale convolution module; for details, please refer to [link / reference]. Figure 4 This step first performs a layer normalization process on the aggregated one-dimensional features output from step S1, and then inputs them into the multi-scale convolutional module for deeper feature learning. Inside the multi-scale convolutional module, the input features are passed in parallel through multiple convolutional layers, each configured with a different sized kernel to capture features across different scales, such as from fine-grained rapid fluctuations to long-term trend changes. Figure 4 As shown.
[0090] In this embodiment, three parallel convolutional layers can be configured, with kernel sizes of 3, 5, and 7, respectively. Optionally, in each convolutional layer, a dynamic masking technique is employed to generate a specific mask for each layer. These masks are determined by the kernel size and a predefined maximum kernel length, ensuring that the convolutional kernel is computed only within its designated effective region, thereby reducing unnecessary computation and potentially lowering noise interference. Then, the output of each parallel convolutional branch is subjected to batch normalization and a ReLU activation function. Finally, the features obtained from processing all parallel convolutional branches are concatenated along the channel dimension to form a wider feature representation. The concatenated features are then subjected to batch normalization and a ReLU activation function to obtain the final multi-scale fused features.
[0091] Step S3: Classify and output the multi-scale fusion features obtained in Step S2. This step first processes the multi-scale fusion features output in Step S2. Perform layer normalization to obtain normalized feature X ln This is done by adjusting the statistical properties of each layer of features to accelerate training and improve the model's generalization performance. The calculation formula is shown below:
[0092]
[0093] Then, the obtained normalized feature X ln The model is then passed through a GELU activation function to introduce non-linearity, enhancing its ability to handle complex patterns, and optionally through a Dropout layer to prevent overfitting. Finally, the processed features are input into a fully connected linear layer, which maps the high-dimensional features to a predefined number of categories, outputting the final raw logistic value y′ of the classification prediction. The calculation formula is shown below:
[0094] y′=Linear(GELU(X ln )).
[0095] Step S4, Model Training. To enable the constructed network model to effectively perform the classification task, it needs to be trained. First, a loss function is defined. In this example, the cross-entropy loss function is used to measure the difference between the probability distribution of the model's predicted output and the true target class label. For N samples and C classes, the calculation formula can be expressed as:
[0096]
[0097] in, It is the model's true class y for the i-th sample. i The predicted logical value.
[0098] Then, the Adam optimizer is selected, and based on the calculated loss function value, all learnable parameters in the network model are iteratively adjusted using the backpropagation algorithm to minimize the loss function. Optionally, specific training strategies are employed to improve the model's final performance and generalization ability.
[0099] In the embodiments, a maximum number of training cycles can be set, such as 30 cycles; a learning rate can be set, such as 1e-3; and an early stopping mechanism can be implemented. For example, if the loss on the validation set fails to decrease significantly within a preset number of consecutive cycles, the training process is terminated early to prevent the model from overfitting on the training data and to preserve the model state with the best performance.
[0100] The descriptive terms "example," "implementation," or "sample" used in this specification are intended to illustrate the structure, function, or feature of embodiments of the present invention. The use of these terms is exemplary and not intended to limit the scope of the embodiments of the present invention. Furthermore, the described structures, functions, or features may be combined in a practical manner in multiple examples.
[0101] The foregoing has described in detail the embodiments of the present invention. It is worth noting that these embodiments are merely intended to aid those skilled in the art in understanding the principles and knowledge of the present invention and should not be construed as limiting the application scenarios of the present invention. Modifications, variations, and transformations made to the above embodiments based on the principles and spirit of the present invention still fall within the scope of protection of the present invention. Obviously, it is unnecessary to exhaustively describe all implementation methods here.
Claims
1. A multivariate time series classification method based on spectral convolutional networks, characterized in that, The method is used by computing devices (such as servers, embedded devices, etc.) to process and classify input one-dimensional multivariate time series data, and specifically includes the following steps: Step S1. Perform time-spectrum processing on the one-dimensional multivariate time series data to convert the one-dimensional multivariate time series data into a comprehensive one-dimensional feature representation containing frequency domain and periodic structure information; Step S2. Perform multi-scale feature extraction on the comprehensive one-dimensional feature representation. Process the comprehensive one-dimensional feature representation through parallel convolutional layers with convolutional kernels of different sizes to capture features at different time scales and obtain multi-scale fused features. Step S3. Classify the data according to the multi-scale fusion features to obtain the classification result of the one-dimensional multivariate time series data.
2. The multivariate time series classification method based on spectral convolutional networks according to claim 1, characterized in that, Step S1 specifically includes the following sub-steps: Step S11. Perform enhanced discrete Fourier transform to transform the one-dimensional multivariate time series data to the frequency domain. Step S12. Execute the spectrum energy priority strategy and construct a two-dimensional time series representation. Execute the spectrum energy priority strategy, select several key frequency components according to the energy magnitude of the frequency domain data, and reshape the one-dimensional multivariate time series data into at least one two-dimensional time series tensor based on the period length corresponding to the key frequency components. Step S13. Perform feature extraction and transformation on each two-dimensional time series tensor generated in steps S1-2-4, extract two-dimensional features from the at least one two-dimensional time series tensor, and convert it into at least one one-dimensional feature vector; Step S14. Perform feature aggregation and residual connection, and perform weighted aggregation on the at least one-dimensional feature vector to form the comprehensive one-dimensional feature representation.
3. The multivariate time series classification method based on spectral convolutional networks according to claim 1, characterized in that, In step S11, the enhanced discrete Fourier transform includes standardizing the spectral data, and its calculation formula is shown below: Among them, X freq The spectrum data is obtained after Fast Fourier Transform, where μ is the mean amplitude, σ is the standard deviation of the amplitude, and X is the mean amplitude. freq_norm This is the standardized spectrum data.
4. The multivariate time series classification method based on spectral convolutional networks according to claim 1, characterized in that, In step S12, a spectrum energy priority strategy is implemented and a two-dimensional time series representation is constructed. The calculation formula is shown below: Among them, E ω L represents the energy corresponding to the frequency point ω. i To be based on the selected key frequency ω i The calculated period length, T, is the length of the original one-dimensional time series data.
5. The multivariate time series classification method based on spectral convolutional networks according to claim 1, characterized in that, In step S14, at least one-dimensional feature vectors are weighted and aggregated, and the calculation formula is as follows: in, The aggregated one-dimensional output features These are the normalized weighting coefficients calculated based on frequency energy. This is a one-dimensional feature vector obtained after two-dimensional feature extraction.
6. The multivariate time series classification method based on spectral convolutional networks according to claim 1, characterized in that, The multi-scale feature extraction in step S2 is accomplished through at least two parallel convolutional layers configured with kernels of different sizes, and a feature concatenation unit for concatenating the outputs of the parallel convolutional layers. Specifically, it includes the following sub-steps: Step S21. The aggregated one-dimensional features output from step S1 are first processed by layer normalization, and then input into the multi-scale convolution module. Step S22. Used to enable the input features to pass through multiple convolutional layers with different sized kernels in parallel within the multi-scale convolutional module; Step S23. Use dynamic masking technology in each convolutional layer; Step S24. This step is used to process the output of each parallel convolutional branch through batch normalization and a ReLU activation function. Step S25. This step concatenates the features obtained from all parallel convolutional branches along the channel dimension. The concatenated features are then subjected to batch normalization and a ReLU activation function to obtain the final multi-scale fused features.
7. The multivariate time series classification method based on spectral convolutional networks according to claim 1, characterized in that, Step S3 specifically includes the following sub-steps: Step S31. Used for processing the multi-scale fusion features output in step S2 Perform layer normalization to obtain normalized feature X ln ; Step S32. Used to normalize feature X ln It passes through a GELU activation function and optionally a Dropout layer in sequence; Step S33. This step inputs the processed features into a fully connected linear layer and outputs the final classification prediction's original logical value y′.
8. The multivariate time series classification method based on spectral convolutional networks according to claim 1, characterized in that, Step S4 specifically includes the following sub-steps: Step S41. Used to define the loss function, for example, using the cross-entropy loss function to calculate the difference between the model's predicted output and the true target category label; Step S42. Used to optimize model parameters, for example, by selecting the Adam optimizer and iteratively adjusting the learnable parameters in the network model through the backpropagation algorithm based on the calculated loss function value; Step S43. Used to employ specific training strategies to improve the final performance and generalization ability of the model, such as setting the number of training epochs, learning rate, and implementing an early stopping mechanism.
9. A multivariate time series classification system based on a spectral convolutional network, the system being used to perform the method according to any one of claims 1-8, characterized in that, The system includes a data receiving and initialization module, a time-spectrum processing module, a multi-scale feature extraction module, a classification output module, and a model training module. The data receiving and initialization module is used to receive externally input one-dimensional multivariate time series data and can perform preliminary data formatting or preparation work. The time-spectrum processing module is responsible for performing enhanced discrete Fourier transform on the input one-dimensional multivariate time series data to convert it to the frequency domain. Next, it implements a spectrum energy priority strategy, including calculating the energy of each frequency, selecting the top-k key frequencies, calculating the corresponding period length, and finally reshaping the original one-dimensional time series data into k two-dimensional time series tensors. Subsequently, the two-dimensional convolutional neural network submodule within this time-spectrum processing module extracts features from each two-dimensional time series tensor and converts them into one-dimensional feature vectors through pooling and reshaping operations. Finally, these one-dimensional feature vectors are weighted and aggregated according to the weights calculated for each frequency energy, and can be combined with the module input through residual connections. The multi-scale feature extraction module receives one-dimensional features output by the time-spectrum processing module (or its last cascaded unit) and processed by layer normalization. It contains multiple parallel convolutional layers with different sized kernels to capture the temporal dependencies of different scale ranges from the input features. These convolutional layers use dynamic masking technology. The output of each parallel branch is batch normalized and ReLU activated. Then the features of all branches are concatenated and batch normalized and ReLU activated again to form multi-scale fused features. The classification output module first performs layer normalization on the features output by the multi-scale feature extraction module; then it performs a nonlinear transformation using the GELU activation function; finally, it maps the processed features to a predefined category space through a fully connected linear classifier and outputs the classification result. The model training module includes a loss function calculation unit and a parameter optimization unit, and can integrate a training strategy control unit. This module uses the training dataset to adjust and optimize the learnable parameters in the system. The time-spectrum processing module is used to execute step S1 as described in claim 1; The multi-scale feature extraction module is used to perform step S2 as described in claim 1; The classification output module is used to perform step S3 as described in claim 1.
10. A multivariate time series classification system based on a spectral convolutional network according to claim 9, characterized in that, The time-spectrum processing module includes an enhanced discrete Fourier transform unit, a two-dimensional representation construction unit, a two-dimensional feature extraction and transformation unit, and a feature aggregation unit. The enhanced discrete Fourier transform unit is used to apply the enhanced discrete Fourier transform to the one-dimensional multivariate time series data and transform it to the frequency domain. The two-dimensional representation construction unit is used to execute a spectrum energy priority strategy, select several key frequency components according to the energy magnitude of the frequency domain data, and reshape the one-dimensional multivariate time series data into at least one two-dimensional time series tensor based on the period length corresponding to the key frequency components. The two-dimensional feature extraction and conversion unit is used to extract two-dimensional features from the at least one two-dimensional temporal tensor and convert it into at least one one-dimensional feature vector. The feature aggregation unit is used to perform weighted aggregation on the at least one-dimensional feature vector to form the comprehensive one-dimensional feature representation. The two-dimensional feature extraction and multi-scale feature extraction in the time-spectrum processing module and the multi-scale feature extraction module are accomplished by a convolutional neural network module containing depthwise separable convolutions.