Deep learning-based time series data analysis system and method
By combining a sliding window recurrent neural network and a dynamically updated recurrent neural network with a deep learning-based time series data analysis system, and optimizing model weights using adaptive update rate and self-supervised loss, the problems of gradient vanishing and computational overhead in long sequence processing of existing models are solved, and efficient and real-time time series analysis is achieved.
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- UNIV OF CHINESE ACAD OF SOCIAL SCI
- Filing Date
- 2026-01-27
- Publication Date
- 2026-05-08
AI Technical Summary
Existing time series analysis models suffer from gradient vanishing/exploding problems when processing long sequence data, resulting in huge computational overhead and difficulty in meeting real-time requirements. Furthermore, the adaptive mechanism lacks a theoretical basis and cannot dynamically adjust parameters according to sequence characteristics, leading to large performance fluctuations.
A deep learning-based time series data analysis system is adopted, including input preprocessing, sliding window recurrent neural network, dynamically updated recurrent neural network and meta-learning optimization module. The dynamic multilayer perceptron layer weights are optimized by adaptive update rate and self-supervised loss. Combined with the fusion of short-term and long-term features, the model behavior is dynamically adjusted to adapt to different sequence characteristics.
It significantly improves the model's performance in long sequence tasks, reduces computational complexity, maintains high accuracy and real-time performance, and enhances the model's stability and generalization ability, making it suitable for time series data analysis scenarios with high real-time requirements.
Smart Images
Figure CN121997263A_ABST
Abstract
Description
Technical Field
[0001] This invention relates to the field of machine learning technology, and more specifically to a time-series data analysis system and method based on deep learning. Background Technology
[0002] Time series data analysis, as a key application direction in the field of artificial intelligence, has wide application value in fields such as financial forecasting, medical and health monitoring, and industrial equipment monitoring. With the rapid development of the Internet of Things and sensor technology, the scale and complexity of time series data collection are growing exponentially, which puts forward higher requirements for the performance of time series analysis models. At present, deep learning-based time series analysis models have become the mainstream technical solution, among which recurrent neural networks (RNNs) and their variants, such as LSTM, GRU and Transformer architectures, are widely used in various time series analysis tasks.
[0003] However, existing time series analysis models face significant challenges when processing long sequence data. Traditional RNN models suffer from gradient vanishing / exploding problems when dealing with long sequences, making it difficult to effectively capture long-term dependencies. While Transformers can handle long sequences, their computational overhead is enormous when dealing with extremely long sequences, making it difficult to meet real-time requirements. Furthermore, existing models typically use fixed parameters during inference, failing to dynamically adjust their behavior based on the characteristics of the input sequence. This results in significant performance fluctuations when processing sequences with different characteristics, making it impossible to maintain both high accuracy and computational efficiency.
[0004] In existing technical solutions, although some adaptive mechanisms have been introduced to improve model performance, they generally suffer from two key drawbacks: first, the adaptive mechanisms lack an effective theoretical basis and cannot accurately adjust the parameter update intensity according to the sequence characteristics; second, the computational efficiency is low and cannot meet the real-time requirements while ensuring accuracy. Summary of the Invention
[0005] In view of the above-mentioned shortcomings of the existing technology, the present invention provides a time series data analysis system and method based on deep learning, which can effectively solve the problems mentioned in the existing technology.
[0006] To achieve the above objectives, the present invention provides the following technical solution:
[0007] This invention provides a time-series data analysis system based on deep learning, comprising:
[0008] The input preprocessing module is used to perform standardization and segmentation on the input time series data. The standardization process uses the mean and standard deviation calculated based on the training dataset to normalize the input time series data. The segmentation process divides the normalized time series data into segments according to a fixed window of eight thousand time steps.
[0009] The sliding window recurrent neural network module is used to process the input time series data within each sliding window of eight thousand time steps to generate short-term time series feature vectors;
[0010] A dynamic recurrent neural network module is used to process long-term time-series data. It includes a static multilayer perceptron layer, a dynamic multilayer perceptron layer, and a weight update unit. The weights of the static multilayer perceptron layer remain unchanged during inference, while the weights of the dynamic multilayer perceptron layer are dynamically updated according to an adaptive update rate during inference. The weight update unit only updates the weights of the dynamic multilayer perceptron layer in the last quarter layer of the recurrent neural network.
[0011] The meta-learning optimization module is used to dynamically update the initialization parameters of the recurrent neural network module through the outer loop optimization, so that the dynamic update recurrent neural network module can dynamically adjust the adaptive update rate according to the characteristics of the input sequence during the inference process;
[0012] The output fusion module is used to fuse the short-term time-series feature vector generated by the sliding window recurrent neural network module with the long-term time-series feature vector generated by the dynamically updated recurrent neural network module to generate the final time-series analysis result.
[0013] The adaptive update rate is dynamically calculated based on the complexity score and sequence length of the current time-series data, and the calculation formula is as follows:
[0014] ;
[0015] in, To achieve adaptive update rate, , and For learnable hyperparameters optimized through meta-learning, For complexity score, The sequence length is given.
[0016] Furthermore, the complexity score is the entropy value of the current sequence, calculated using the following formula:
[0017] ;
[0018] in, The relative frequency of element i in the current sequence. It is the natural logarithm.
[0019] Furthermore, the static multilayer perceptron layer and the dynamic multilayer perceptron layer have the same structure, both including a first fully connected layer, a second fully connected layer, layer normalization units, and residual connection units, wherein the input dimension of the first fully connected layer is... The number of hidden units is The activation function is GELU, and the input dimension of the second fully connected layer is... The output dimension is The layer normalization unit normalizes the output of the first fully connected layer, and the residual connection unit adds the input and output of the first fully connected layer.
[0020] Furthermore, the weight update unit performs the following operations during inference to adapt to sequence changes:
[0021] pass The matrix performs a linear transformation on the input time-series data to generate a corrupted input, where This is the weight matrix of the dynamic multilayer perceptron layer, used to perturb the input time-series data to obtain the corrupted input.
[0022] pass The matrix performs a linear transformation on the input time-series data to generate the target, where This is the target weight matrix for the dynamic multilayer perceptron layer, used to generate targets from the input time-series data to obtain the target.
[0023] The self-supervised loss is calculated using the following formula:
[0024] ;
[0025] in, For self-monitoring losses, Input time series data for the current time step. For dynamic multilayer perceptron layers;
[0026] Based on adaptive update rate and self-monitoring loss The gradient updates the weights of the dynamic multilayer perceptron layer, and the calculation formula is as follows:
[0027] ;
[0028] in, For the weights of the dynamic multilayer perceptron layer, For self-monitoring loss The gradient with respect to the weight w.
[0029] Furthermore, the output fusion module calculates the adaptive fusion weights using the following formula:
[0030] ;
[0031] ;
[0032] in, For static multilayer perceptron fusion weights, The fusion weights are for the dynamic multi-layer perceptron layer.
[0033] A deep learning-based time series data analysis method includes the following steps:
[0034] S100. Standardize and segment the input time series data, wherein the standardization uses the mean and standard deviation calculated based on the training dataset, and the segmentation is performed according to eight thousand time steps.
[0035] S200: Process the input time series data within each sliding window of eight thousand time steps to generate short-term time series feature vectors;
[0036] S300. Use a dynamically updated recurrent neural network module to process long-term time-series data. The dynamically updated recurrent neural network module includes a static multilayer perceptron layer, a dynamic multilayer perceptron layer, and a weight update unit. The weights of the static multilayer perceptron layer remain unchanged during inference. The weights of the dynamic multilayer perceptron layer are dynamically updated according to an adaptive update rate during inference. The weight update unit only updates the weights of the dynamic multilayer perceptron layer in the last quarter layer of the recurrent neural network.
[0037] S400. Use the meta-learning optimization module to optimize the initialization parameters of the dynamic update recurrent neural network module, so that the dynamic update recurrent neural network module can dynamically adjust the adaptive update rate according to the characteristics of the input sequence during the inference process.
[0038] S500: The output fusion module is used to fuse the short-term time-series feature vector generated by the sliding window recurrent neural network module with the long-term time-series feature vector generated by the dynamically updated recurrent neural network module to generate the final time-series analysis result.
[0039] Furthermore, the complexity score is calculated as follows: the entropy value of the current sequence is calculated using the formula:
[0040] ;
[0041] in, The relative frequency of element i in the current sequence. It is the natural logarithm.
[0042] Furthermore, the weight update operation of the dynamic multilayer perceptron layer includes:
[0043] pass The matrix performs a linear transformation on the input time-series data to generate a corrupted input, where This is the weight matrix of the dynamic multilayer perceptron layer, used to perturb the input time-series data to obtain the corrupted input.
[0044] pass The matrix performs a linear transformation on the input time-series data to generate the target, where This is the target weight matrix for the dynamic multilayer perceptron layer, used to generate targets from the input time-series data to obtain the target.
[0045] The self-supervised loss is calculated using the following formula:
[0046] ;
[0047] in, For self-monitoring losses, Input time series data for the current time step. For dynamic multilayer perceptron layers;
[0048] Based on adaptive update rate and self-monitoring loss The gradient updates the weights of the dynamic multilayer perceptron layer, and the calculation formula is as follows:
[0049] ;
[0050] in, For the weights of the dynamic multilayer perceptron layer, For self-monitoring loss The gradient with respect to the weight w.
[0051] Furthermore, the fusion operation of the output fusion module includes:
[0052] The fusion weights of a static multilayer perceptron are calculated using the following formula:
[0053]
[0054] The formula for calculating the fusion weights of a dynamic multilayer perceptron layer is as follows:
[0055]
[0056] Multiply the output of the static multilayer perceptron layer by The output of the dynamic multilayer perceptron layer multiplied by Then they are added together to obtain the fused temporal feature vector.
[0057] The technical solution provided by this invention has the following advantages compared with the known prior art:
[0058] This invention uses an adaptive update rate mechanism to dynamically adjust the weight update intensity based on the complexity and length of the input sequence. This increases the update intensity on complex sequences to fully adapt to specific patterns, and decreases the update intensity on simple sequences to avoid over-adjustment. This significantly improves the model's performance under various sequence characteristics, enabling the model to achieve a performance level comparable to Transformer in long sequence tasks.
[0059] This invention significantly reduces the computational complexity of the inference process by updating only the dynamic multilayer perceptron layer weights of the last quarter layer of the recurrent neural network, thereby optimizing computational efficiency while maintaining high-precision analysis results. It is particularly suitable for time-series data analysis scenarios with high real-time requirements.
[0060] This invention, through a meta-learning optimization module and an adaptive fusion mechanism, enables the model to retain long-term general knowledge learned from training data during inference while dynamically adjusting according to the characteristics of the current sequence. This significantly improves the model's stability and generalization ability, allowing it to achieve excellent analysis results on time series data with different domains and characteristics. It effectively solves the problem of performance degradation in traditional time series analysis models when processing long sequences. Attached Figure Description
[0061] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the accompanying drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are merely some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without any creative effort.
[0062] Figure 1 This is a schematic diagram of the system structure of the present invention;
[0063] Figure 2 This is a schematic diagram of the method flow of the present invention. Detailed Implementation
[0064] To make the objectives, technical solutions, and advantages of the embodiments of the present invention clearer, the technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some, not all, of the embodiments of the present invention. All other embodiments obtained by those skilled in the art based on the embodiments of the present invention without creative effort are within the scope of protection of the present invention.
[0065] The present invention will be further described below with reference to embodiments.
[0066] Example:
[0067] Reference Figure 1 A time-series data analysis system based on deep learning includes:
[0068] The input preprocessing module is used to standardize and segment the input time-series data. The specific steps are as follows: First, the system loads the global mean calculated on the training set during the pre-training phase. with standard deviation The input raw time series data is processed according to the formula. Standardization processes are performed, among which To normalize the original time-series data, the distribution is standardized to eliminate biases caused by differences in units and numerical ranges, thereby improving the stability and convergence speed of model training. Subsequently, the standardized continuous sequence is segmented into segments with a fixed length of 8000 time steps using a sliding window. If the total sequence length is less than 8000 steps, zeros are padded at the end of the sequence to the full window length. If the sequence exceeds 8000 steps, it is sequentially segmented into multiple non-overlapping or partially overlapping segments. The overlap step size can be set according to task requirements, such as 10% of the window length to enhance temporal continuity. This segmentation strategy transforms long sequences into multiple fixed-length subsequences that can be processed in parallel, satisfying the model's input length requirements and facilitating efficient batch computation on hardware such as GPUs.
[0069] The sliding window recurrent neural network module is used to process the input time series data within each sliding window of eight thousand time steps to generate short-term time series feature vectors;
[0070] In one specific embodiment, the sliding window recurrent neural network module employs a gated recurrent unit (GRU) or long short-term memory (LSTM) network structure to extract temporal features for each 8000-step window. The specific process is as follows: data within the window is input into the RNN unit sequentially according to time steps; the hidden state at each time step is updated based on the current input and the hidden state of the previous time step; after the entire window is processed, the hidden state of the last time step is taken as the condensed representation of the window, and then mapped to a fixed-dimensional short-term temporal feature vector through a fully connected layer. This module focuses on capturing local dependencies and short-term fluctuation patterns within the window. Its advantages include high computational efficiency, small parameter count, and suitability for scenarios with high real-time requirements. By using fixed window partitioning, the model can process ultra-long sequences in a streaming manner, avoiding the memory pressure caused by loading the entire sequence at once.
[0071] The dynamically updated recurrent neural network module, used to process long-term time-series data, consists of static multilayer perceptron layers, dynamic multilayer perceptron layers, and a weight update unit. The static layer weights remain fixed after training to maintain the general time-series knowledge learned from the training data; the dynamic layer weights, on the other hand, are dynamically adjusted during inference based on the current input sequence to adapt to changes in data distribution. The weight update unit only updates the last quarter of the dynamic layers in the network. This design significantly reduces online computational overhead while ensuring that the model's deep features can flexibly adapt to new sequences.
[0072] The specific update mechanism is as follows: at each time step, the weight matrix of the dynamic layer is used. For input Perform a linear transformation to generate the perturbed "destructed input"; simultaneously use another learnable matrix. Generate the target vector; then calculate the mean squared error between the two as the self-supervised loss, and update it according to the adaptive update rate. Gradient descent is used to update the weights of dynamic layers. This mechanism enables the model to learn online and continuously optimize during inference, making it particularly suitable for scenarios where the data distribution changes gradually over time.
[0073] The meta-learning optimization module is used to dynamically update the initialization parameters of the recurrent neural network module through the outer loop optimization, so that the dynamic update recurrent neural network module can dynamically adjust the adaptive update rate according to the characteristics of the input sequence during the inference process;
[0074] Specifically, the meta-learning optimization module runs during the offline training phase, employing a model-independent meta-learning framework to optimize the initialization parameters and update rate hyperparameters of the dynamically updated recurrent neural network module. The specific process is as follows: several task sets are sampled from multiple tasks with different distributions, each task containing a support set and a query set; in the inner loop, a small amount of gradient updates are performed on the support set for each task, simulating the dynamic adjustment process during inference; in the outer loop, the meta-loss is calculated based on the performance of each task on the query set, and the initialization parameters and hyperparameters are updated using gradient descent. After meta-learning optimization, the model possesses the ability to quickly adapt to new tasks, requiring only a small number of new sequences to adjust the dynamic layer weights, greatly enhancing its cross-domain and cross-scenario generalization performance.
[0075] The output fusion module is used to fuse the short-term time-series feature vector generated by the sliding window recurrent neural network module with the long-term time-series feature vector generated by the dynamically updated recurrent neural network module to generate the final time-series analysis result.
[0076] The adaptive update rate is dynamically calculated based on the complexity score and sequence length of the current time series data. The calculation formula is as follows:
[0077] ;
[0078] in, To achieve adaptive update rate, , and For learnable hyperparameters optimized through meta-learning, For complexity score, The sequence length is given.
[0079] Furthermore, the complexity score is the entropy value of the current sequence, calculated using the following formula:
[0080] ;
[0081] in, The relative frequency of element i in the current sequence. It is the natural logarithm.
[0082] Furthermore, the static and dynamic multilayer perceptron layers have the same structure, both including a first fully connected layer, a second fully connected layer, layer normalization units, and residual connection units, wherein the input dimension of the first fully connected layer is... The number of hidden units is The activation function is GELU, and the input dimension of the second fully connected layer is... The output dimension is The layer normalization unit normalizes the output of the first fully connected layer, and the residual connection unit adds the input and output of the first fully connected layer.
[0083] Furthermore, the weight update unit performs the following operations during inference to adapt to sequence changes:
[0084] pass The matrix performs a linear transformation on the input time-series data to generate a corrupted input, where This is the weight matrix of the dynamic multilayer perceptron layer, used to perturb the input time-series data to obtain the corrupted input.
[0085] pass The matrix performs a linear transformation on the input time-series data to generate the target, where This is the target weight matrix for the dynamic multilayer perceptron layer, used to generate targets from the input time-series data to obtain the target.
[0086] The self-supervised loss is calculated using the following formula:
[0087] ;
[0088] in, For self-monitoring losses, Input time series data for the current time step. For dynamic multilayer perceptron layers;
[0089] Based on adaptive update rate and self-monitoring loss The gradient updates the weights of the dynamic multilayer perceptron layer, and the calculation formula is as follows:
[0090] ;
[0091] in, For the weights of the dynamic multilayer perceptron layer, For self-monitoring loss The gradient with respect to the weight w.
[0092] Furthermore, the output fusion module calculates the adaptive fusion weights using the following formula:
[0093] ;
[0094] ;
[0095] in, For static multilayer perceptron fusion weights, The fusion weights are for the dynamic multi-layer perceptron layer.
[0096] Specifically, this weight design allows the model to rely more on the stable knowledge provided by the static layer when the sequence is simple or the update rate is low; while when the sequence is complex and updates are frequent, it focuses more on the new patterns captured by the dynamic layer. The weighted summed fusion vector is input to a task-specific output layer, such as a Softmax classification layer or a linear regression layer, to generate the final time series analysis results, such as anomaly scores, class labels, or predicted values.
[0097] A deep learning-based time series data analysis method includes the following steps:
[0098] S100. Standardize and segment the input time series data. Standardization uses the mean and standard deviation calculated based on the training dataset. Segmentation is performed according to eight thousand time steps.
[0099] S200: Process the input time series data within each sliding window of eight thousand time steps to generate short-term time series feature vectors;
[0100] S300. Use a dynamically updated recurrent neural network module to process long-term time-series data. The dynamically updated recurrent neural network module includes a static multilayer perceptron layer, a dynamic multilayer perceptron layer, and a weight update unit. The weights of the static multilayer perceptron layer remain unchanged during inference, while the weights of the dynamic multilayer perceptron layer are dynamically updated according to an adaptive update rate during inference. The weight update unit only updates the weights of the dynamic multilayer perceptron layer in the last quarter layer of the recurrent neural network.
[0101] S400. Use the meta-learning optimization module to optimize the initialization parameters of the dynamic update recurrent neural network module, so that the dynamic update recurrent neural network module can dynamically adjust the adaptive update rate according to the characteristics of the input sequence during the inference process.
[0102] S500: The output fusion module is used to fuse the short-term time-series feature vector generated by the sliding window recurrent neural network module with the long-term time-series feature vector generated by the dynamically updated recurrent neural network module to generate the final time-series analysis result.
[0103] Furthermore, the complexity score is calculated as follows: calculate the entropy value of the current sequence, using the formula:
[0104] ;
[0105] in, The relative frequency of element i in the current sequence. It is the natural logarithm.
[0106] Furthermore, the weight update operation of the dynamic multilayer perceptron layer includes:
[0107] pass The matrix performs a linear transformation on the input time-series data to generate a corrupted input, where This is the weight matrix of the dynamic multilayer perceptron layer, used to perturb the input time-series data to obtain the corrupted input.
[0108] pass The matrix performs a linear transformation on the input time-series data to generate the target, where This is the target weight matrix for the dynamic multilayer perceptron layer, used to generate targets from the input time-series data to obtain the target.
[0109] The self-supervised loss is calculated using the following formula:
[0110] ;
[0111] in, For self-monitoring losses, Input time series data for the current time step. For dynamic multilayer perceptron layers;
[0112] Based on adaptive update rate and self-monitoring loss The gradient updates the weights of the dynamic multilayer perceptron layer, and the calculation formula is as follows:
[0113] ;
[0114] in, For the weights of the dynamic multilayer perceptron layer, For self-monitoring loss The gradient with respect to the weight w.
[0115] Furthermore, the fusion operation of the output fusion module includes:
[0116] The fusion weights of a static multilayer perceptron are calculated using the following formula:
[0117]
[0118] The formula for calculating the fusion weights of a dynamic multilayer perceptron layer is as follows:
[0119]
[0120] Multiply the output of the static multilayer perceptron layer by The output of the dynamic multilayer perceptron layer multiplied by Then they are added together to obtain the fused temporal feature vector.
[0121] The above embodiments are only used to illustrate the technical solutions of the present invention, and are not intended to limit it. Although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some of the technical features. Such modifications or substitutions will not cause the essence of the corresponding technical solutions to deviate from the protection scope of the technical solutions of the embodiments of the present invention.
Claims
1. A time-series data analysis system based on deep learning, characterized in that, include: The input preprocessing module is used to perform standardization and segmentation on the input time series data. The standardization process uses the mean and standard deviation calculated based on the training dataset to normalize the input time series data. The segmentation process divides the normalized time series data into segments according to a fixed window of eight thousand time steps. The sliding window recurrent neural network module is used to process the input time series data within each sliding window of eight thousand time steps to generate short-term time series feature vectors; A dynamic recurrent neural network module is used to process long-term time-series data. It includes a static multilayer perceptron layer, a dynamic multilayer perceptron layer, and a weight update unit. The weights of the static multilayer perceptron layer remain unchanged during inference, while the weights of the dynamic multilayer perceptron layer are dynamically updated according to an adaptive update rate during inference. The weight update unit only updates the weights of the dynamic multilayer perceptron layer in the last quarter layer of the recurrent neural network. The meta-learning optimization module is used to dynamically update the initialization parameters of the recurrent neural network module through the outer loop optimization, so that the dynamic update recurrent neural network module can dynamically adjust the adaptive update rate according to the characteristics of the input sequence during the inference process; The output fusion module is used to fuse the short-term time-series feature vector generated by the sliding window recurrent neural network module with the long-term time-series feature vector generated by the dynamically updated recurrent neural network module to generate the final time-series analysis result. The adaptive update rate is dynamically calculated based on the complexity score and sequence length of the current time-series data, and the calculation formula is as follows: ； in, To achieve adaptive update rate, , and For learnable hyperparameters optimized through meta-learning, For complexity score, The sequence length is given.
2. The time-series data analysis system based on deep learning according to claim 1, characterized in that, The complexity score is the entropy value of the current sequence, calculated using the following formula: ； in, The relative frequency of element i in the current sequence. It is the natural logarithm.
3. The time-series data analysis system based on deep learning according to claim 1, characterized in that, The static and dynamic multilayer perceptron layers have the same structure, both including a first fully connected layer, a second fully connected layer, layer normalization units, and residual connection units, wherein the input dimension of the first fully connected layer is... The number of hidden units is The activation function is GELU, and the input dimension of the second fully connected layer is... The output dimension is The layer normalization unit normalizes the output of the first fully connected layer, and the residual connection unit adds the input and output of the first fully connected layer.
4. The time-series data analysis system based on deep learning according to claim 1, characterized in that, The weight update unit performs the following operations during inference to adapt to sequence changes: pass The matrix performs a linear transformation on the input time-series data to generate a corrupted input, where This is the weight matrix of the dynamic multilayer perceptron layer, used to perturb the input time-series data to obtain the corrupted input. pass The matrix performs a linear transformation on the input time-series data to generate the target, where This is the target weight matrix for the dynamic multilayer perceptron layer, used to generate targets from the input time-series data to obtain the target. The self-supervised loss is calculated using the following formula: ； in, For self-monitoring losses, Input time series data for the current time step. For dynamic multilayer perceptron layers; Based on adaptive update rate and self-monitoring loss The gradient updates the weights of the dynamic multilayer perceptron layer, and the calculation formula is as follows: ； in, For the weights of the dynamic multilayer perceptron layer, For self-monitoring loss The gradient with respect to the weight w.
5. The time-series data analysis system based on deep learning according to claim 1, characterized in that, The output fusion module calculates the adaptive fusion weights using the following formula: ； ； in, For static multilayer perceptron fusion weights, The fusion weights are for the dynamic multi-layer perceptron layer.
6. A time-series data analysis method based on deep learning, characterized by the following steps: include: S100. Standardize and segment the input time series data, wherein the standardization uses the mean and standard deviation calculated based on the training dataset, and the segmentation is performed according to eight thousand time steps. S200: Process the input time series data within each sliding window of eight thousand time steps to generate short-term time series feature vectors; S300. Use a dynamically updated recurrent neural network module to process long-term time-series data. The dynamically updated recurrent neural network module includes a static multilayer perceptron layer, a dynamic multilayer perceptron layer, and a weight update unit. The weights of the static multilayer perceptron layer remain unchanged during inference. The weights of the dynamic multilayer perceptron layer are dynamically updated according to an adaptive update rate during inference. The weight update unit only updates the weights of the dynamic multilayer perceptron layer in the last quarter layer of the recurrent neural network. S400. Use the meta-learning optimization module to optimize the initialization parameters of the dynamic update recurrent neural network module, so that the dynamic update recurrent neural network module can dynamically adjust the adaptive update rate according to the characteristics of the input sequence during the inference process. S500: The output fusion module is used to fuse the short-term time-series feature vector generated by the sliding window recurrent neural network module with the long-term time-series feature vector generated by the dynamically updated recurrent neural network module to generate the final time-series analysis result.
7. The time series data analysis method based on deep learning according to claim 6, characterized in that, The complexity score is calculated as follows: Calculate the entropy value of the current sequence using the formula: ； in, The relative frequency of element i in the current sequence. It is the natural logarithm.
8. The time series data analysis method based on deep learning according to claim 6, characterized in that, The weight update operation of the dynamic multilayer perceptron layer includes: pass The matrix performs a linear transformation on the input time-series data to generate a corrupted input, where This is the weight matrix of the dynamic multilayer perceptron layer, used to perturb the input time-series data to obtain the corrupted input. pass The matrix performs a linear transformation on the input time-series data to generate the target, where This is the target weight matrix for the dynamic multilayer perceptron layer, used to generate targets from the input time-series data to obtain the target. The self-supervised loss is calculated using the following formula: ； in, For self-monitoring losses, Input time series data for the current time step. For dynamic multilayer perceptron layers; Based on adaptive update rate and self-monitoring loss The gradient updates the weights of the dynamic multilayer perceptron layer, and the calculation formula is as follows: ； in, For the weights of the dynamic multilayer perceptron layer, For self-monitoring loss The gradient with respect to the weight w.
9. The time series data analysis method based on deep learning according to claim 6, characterized in that, The fusion operation of the output fusion module includes: The fusion weights of a static multilayer perceptron are calculated using the following formula: The formula for calculating the fusion weights of a dynamic multilayer perceptron layer is as follows: Multiply the output of the static multilayer perceptron layer by The output of the dynamic multilayer perceptron layer multiplied by Then they are added together to obtain the fused temporal feature vector.