CN-121997263-A - Time sequence data analysis system and method based on deep learning

CN121997263ACN 121997263 ACN121997263 ACN 121997263ACN-121997263-A

Abstract

The invention relates to the technical field of machine learning, in particular to a time sequence data analysis system and method based on deep learning, wherein the system comprises an input preprocessing module for carrying out standardized processing and segmentation processing on input time sequence data; the device comprises a sliding window cyclic neural network module, a dynamic updating cyclic neural network module, a meta-learning optimization module and an output fusion module, wherein the sliding window cyclic neural network module is used for processing input time sequence data in a sliding window of each eight thousand time steps to generate a short-term time sequence feature vector, the dynamic updating cyclic neural network module is used for processing long-term time sequence data, the meta-learning optimization module is used for dynamically updating initialization parameters of the cyclic neural network module through outer layer cyclic optimization, and the output fusion module is used for fusing the feature vectors to generate a final time sequence analysis result. The invention innovatively designs a weight updating mechanism for updating only the last quarter layer of the cyclic neural network, thereby remarkably reducing the computational complexity, and effectively solving the contradiction between the computational efficiency and the model adaptability in the long-sequence time sequence analysis while maintaining high precision.

Inventors

HE YUANJING
CHEN HONGBO

Assignees

中国社会科学院大学

Dates

Publication Date: 20260508
Application Date: 20260127

Claims (9)

1. A time series data analysis system based on deep learning, comprising: The input preprocessing module is used for carrying out standardization processing and segmentation processing on the input time sequence data, wherein the standardization processing is used for carrying out normalization on the input time sequence data by using a mean value and a standard deviation calculated based on a training data set, and the segmentation processing is used for segmenting the normalized time sequence data according to a fixed window of eight thousand time steps; The sliding window cyclic neural network module is used for processing the input time sequence data in each eight thousand time steps in the sliding window to generate a short-term time sequence feature vector; the dynamic updating cyclic neural network module is used for processing long-term time sequence data and comprises a static multilayer perceptron layer, a dynamic multilayer perceptron layer and a weight updating unit, wherein the weight of the static multilayer perceptron layer is kept unchanged in the reasoning process, the weight of the dynamic multilayer perceptron layer is dynamically updated according to the self-adaptive updating rate in the reasoning process, and the weight updating unit only updates the weight of the dynamic multilayer perceptron layer of the last quarter layer of the cyclic neural network; The element learning optimization module is used for dynamically updating the initialization parameters of the cyclic neural network module through outer cyclic optimization, so that the dynamic cyclic neural network module can dynamically adjust the self-adaptive update rate according to the characteristics of the input sequence in the reasoning process; The output fusion module is used for fusing the short-term time sequence feature vector generated by the sliding window cyclic neural network module with the long-term time sequence feature vector generated by the dynamic updating cyclic neural network module to generate a final time sequence analysis result; The self-adaptive update rate is dynamically calculated according to the complexity fraction and the sequence length of the current time sequence data, and the calculation formula is as follows: ; Wherein, the In order to adapt the update rate to the situation, 、 And To learn the optimized learnable superparameters through meta-learning, As a fraction of the complexity of the model, Is the sequence length.
2. The deep learning based time series data analysis system of claim 1, wherein the complexity score is an entropy value of a current sequence, and the calculation formula is: ; Wherein, the For the relative frequency of occurrence of element i in the current sequence, Is natural logarithm.
3. The deep learning based time series data analysis system of claim 1, wherein the static multilayer perceptron layer and the dynamic multilayer perceptron layer have the same structure and each comprise a first fully connected layer, a second fully connected layer, a layer normalization unit and a residual connection unit, 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 full connection layer is The output dimension is The layer normalizing unit normalizes the output of the first full-connection layer, and the residual error connecting unit adds the input and the output of the first full-connection layer.
4. The deep learning based time series data analysis system as claimed in claim 1, wherein the weight updating unit performs the following operations in the reasoning process to accommodate the sequence variation: By passing through The matrix performs linear transformation on the input time sequence data to generate destroyed input, wherein The weight matrix is a dynamic multi-layer perceptron layer and is used for carrying out disturbance processing on input time sequence data to obtain damaged input; By passing through Linearly transforming the input time sequence data by matrix to generate target, wherein The target weight matrix is a target weight matrix of the dynamic multi-layer perceptron layer and is used for generating targets for input time sequence data to obtain targets; And calculating self-supervision loss, wherein the calculation formula is as follows: ; Wherein, the In order to self-monitor the loss of the device, For the input of time series data for the current time step, Is a dynamic multi-layer perceptron layer; according to adaptive update rate And self-monitoring losses The weight of the dynamic multi-layer perceptron layer is updated by the gradient of (1), and the calculation formula is as follows: ; Wherein, the For the weights of the dynamic multi-layer perceptron layer, To self-supervise losses Gradient to weight w.
5. The deep learning-based time series data analysis system of claim 1, wherein the output fusion module calculates an adaptive fusion weight according to the following calculation formula: ; ; Wherein, the Is the fusion weight of the static multilayer perceptron layer, And the fusion weight of the dynamic multi-layer perceptron layer is obtained.
6. The time sequence data analysis method based on deep learning is characterized by comprising the following steps: s100, carrying out standardization and segmentation processing on input time sequence data, wherein the standardization is carried out according to eight thousand time steps by using a mean value and a standard deviation calculated based on a training data set; S200, processing input time sequence data in each sliding window of eight thousand time steps to generate a short-term time sequence feature vector; S300, processing long-term time sequence data by using a dynamic updating cyclic neural network module, wherein the dynamic updating cyclic neural network module comprises a static multilayer perceptron layer, a dynamic multilayer perceptron layer and a weight updating unit, wherein the weight of the static multilayer perceptron layer is kept unchanged in the reasoning process, the weight of the dynamic multilayer perceptron layer is dynamically updated according to an adaptive updating rate in the reasoning process, and the weight updating unit only updates the weight of the dynamic multilayer perceptron layer of the last quarter layer of the cyclic neural network; S400, optimizing initialization parameters of the dynamic updating cyclic neural network module by using a meta-learning optimization module, so that the dynamic updating cyclic neural network module can dynamically adjust the self-adaptive updating rate according to the characteristics of the input sequence in the reasoning process; S500, using an output fusion module to fuse the short-term time sequence feature vector generated by the sliding window cyclic neural network module with the long-term time sequence feature vector generated by the dynamic updating cyclic neural network module, and generating a final time sequence analysis result.
7. The method for deep learning based time series data analysis of claim 6, wherein the complexity score is calculated by calculating the entropy of the current sequence according to the formula: ; Wherein, the For the relative frequency of occurrence of element i in the current sequence, Is natural logarithm.
8. The deep learning based time series data analysis method as claimed in claim 6, wherein the weight update operation of the dynamic multi-layer perceptron layer comprises: By passing through The matrix performs linear transformation on the input time sequence data to generate destroyed input, wherein The weight matrix is a dynamic multi-layer perceptron layer and is used for carrying out disturbance processing on input time sequence data to obtain damaged input; By passing through Linearly transforming the input time sequence data by matrix to generate target, wherein The target weight matrix is a target weight matrix of the dynamic multi-layer perceptron layer and is used for generating targets for input time sequence data to obtain targets; And calculating self-supervision loss, wherein the calculation formula is as follows: ; Wherein, the In order to self-monitor the loss of the device, For the input of time series data for the current time step, Is a dynamic multi-layer perceptron layer; according to adaptive update rate And self-monitoring losses The weight of the dynamic multi-layer perceptron layer is updated by the gradient of (1), and the calculation formula is as follows: ; Wherein, the For the weights of the dynamic multi-layer perceptron layer, To self-supervise losses Gradient to weight w.
9. The deep learning based time series data analysis method according to claim 6, wherein the fusion operation of the output fusion module includes: And calculating the fusion weight of the static multilayer perceptron layer, wherein the formula is as follows: the fusion weight formula for calculating the dynamic multi-layer perceptron layer is as follows: Multiplying the output of the static multilayer perceptron layer by The output of the dynamic multi-layer perceptron layer is multiplied by And then adding to obtain the fused time sequence feature vector.

Description

Time sequence data analysis system and method based on deep learning Technical Field The invention relates to the technical field of machine learning, in particular to a time sequence data analysis system and method based on deep learning. Background The time sequence data analysis is used as a key application direction in the artificial intelligence field, has wide application value in the fields of financial prediction, medical health monitoring, industrial equipment monitoring and the like, along with the rapid development of the Internet of things and sensor technology, the acquisition scale and complexity of the time sequence data are exponentially increased, higher requirements are put forward on the performance of a time sequence analysis model, and at present, the time sequence analysis model based on deep learning becomes a mainstream technical scheme, wherein a cyclic neural network RNN and variants thereof, such as LSTM, GRU and a Transformer architecture, are widely applied to various time sequence analysis tasks. However, the conventional time sequence analysis model faces significant challenges when processing long sequence data, the conventional RNN model has the problems of gradient disappearance/explosion when processing long sequences, and is difficult to effectively capture long-term dependency, while the Transformer can process long sequences, but has huge calculation cost when processing ultra-long sequences, and is difficult to meet real-time requirements, in addition, the conventional model usually adopts fixed parameters in the reasoning process, and cannot dynamically adjust own behaviors according to the characteristics of input sequences, so that performance fluctuation is large when processing sequences with different characteristics, and calculation efficiency cannot be simultaneously considered while high precision is maintained. In the prior art, although some self-adaptive mechanisms are introduced to improve the model performance, two key defects generally exist, namely, the self-adaptive mechanism lacks an effective theoretical basis and cannot accurately adjust the parameter updating strength according to the sequence characteristics, and the calculation efficiency is low and cannot meet the real-time requirement while ensuring the accuracy. Disclosure of Invention Aiming at the defects existing in the prior art, the invention provides a time sequence data analysis system and a time sequence data analysis method based on deep learning, which can effectively solve the problems in the prior art. In order to achieve the above purpose, the invention is realized by the following technical scheme: The invention provides a time sequence data analysis system based on deep learning, which comprises the following steps: The input preprocessing module is used for carrying out standardization processing and segmentation processing on the input time sequence data, wherein the standardization processing is used for carrying out normalization on the input time sequence data by using a mean value and a standard deviation calculated based on a training data set, and the segmentation processing is used for segmenting the normalized time sequence data according to a fixed window of eight thousand time steps; The sliding window cyclic neural network module is used for processing the input time sequence data in each eight thousand time steps in the sliding window to generate a short-term time sequence feature vector; the dynamic updating cyclic neural network module is used for processing long-term time sequence data and comprises a static multilayer perceptron layer, a dynamic multilayer perceptron layer and a weight updating unit, wherein the weight of the static multilayer perceptron layer is kept unchanged in the reasoning process, the weight of the dynamic multilayer perceptron layer is dynamically updated according to the self-adaptive updating rate in the reasoning process, and the weight updating unit only updates the weight of the dynamic multilayer perceptron layer of the last quarter layer of the cyclic neural network; The element learning optimization module is used for dynamically updating the initialization parameters of the cyclic neural network module through outer cyclic optimization, so that the dynamic cyclic neural network module can dynamically adjust the self-adaptive update rate according to the characteristics of the input sequence in the reasoning process; The output fusion module is used for fusing the short-term time sequence feature vector generated by the sliding window cyclic neural network module with the long-term time sequence feature vector generated by the dynamic updating cyclic neural network module to generate a final time sequence analysis result; The self-adaptive update rate is dynamically calculated according to the complexity fraction and the sequence length of the current time sequence data, and the calculation formula is as follows: ; Wherein, the