CN-122020139-A - Feature prediction method and device for non-stationary time sequence data

CN122020139ACN 122020139 ACN122020139 ACN 122020139ACN-122020139-A

Abstract

The invention provides a feature prediction method and a device for non-stationary time sequence data, which relate to the technical field of time sequence prediction and comprise the steps of obtaining a multi-source monitoring sequence; the method comprises the steps of calculating according to a time sequence and a wavelet base sequence in a multi-source monitoring sequence to obtain a waveform coefficient, carrying out optimal wavelet decomposition on the multi-source monitoring sequence under different decomposition layers through the waveform coefficient to obtain a plurality of low-frequency components and a plurality of high-frequency components, inputting the plurality of low-frequency components into a non-stationary attention mechanism to carry out sequence stable processing and prediction to obtain a low-frequency prediction result, predicting the plurality of high-frequency components based on an attention mechanism of a preset sequence model to obtain a high-frequency prediction result, wherein the attention mechanism comprises a characteristic attention mechanism and a time sequence attention mechanism, and carrying out reconstruction based on the low-frequency prediction result and the high-frequency prediction result to obtain a time sequence characteristic prediction result. The invention solves the problem of reduced prediction accuracy.

Inventors

YANG LIU
LI MINGHUI
MA ZHENG
WANG GUO
LIU HENG

Assignees

西南交通大学

Dates

Publication Date: 20260512
Application Date: 20260129

Claims (10)

1. The characteristic prediction method for the non-stationary time sequence data is characterized by comprising the following steps of: Acquiring a multi-source monitoring sequence, wherein the multi-source monitoring sequence comprises a structural response monitoring sequence and an environmental load sequence; Calculating according to a time sequence in the multi-source monitoring sequence and a preset wavelet base sequence to obtain a waveform coefficient, and carrying out optimal wavelet decomposition on the multi-source monitoring sequence under different decomposition layers through the waveform coefficient to obtain a plurality of low-frequency components and a plurality of high-frequency components; inputting a plurality of low-frequency components into a preset non-stationary attention mechanism to perform sequence stationary processing and prediction to obtain a low-frequency prediction result; Predicting a plurality of high-frequency components based on an attention mechanism of a preset sequence model to obtain a high-frequency prediction result, wherein the attention mechanism comprises a characteristic attention mechanism and a time sequence attention mechanism; and reconstructing based on the low-frequency prediction result and the high-frequency prediction result to obtain a time sequence characteristic prediction result.
2. The feature prediction method for non-stationary time series data according to claim 1, wherein calculating to obtain a waveform coefficient according to a time sequence in the multi-source monitoring sequence and a preset wavelet base sequence, and performing optimal wavelet decomposition on the multi-source monitoring sequence under different decomposition layers by using the waveform coefficient to obtain a plurality of low-frequency components and a plurality of high-frequency components, includes: Obtaining a plurality of decomposition layers; calculating according to the time sequence and the wavelet base sequence to obtain a waveform coefficient; Performing discrete processing on the waveform coefficients, performing waveform similarity analysis on the time sequence and the wavelet base sequence through the discrete waveform coefficients, and performing optimal wavelet base selection according to analysis results to obtain optimal wavelet bases; Performing wavelet decomposition on the multi-source monitoring sequence according to the optimal wavelet base to obtain a plurality of initial low-frequency components and a plurality of initial high-frequency components; And processing the initial high-frequency components according to the structural response monitoring sequences under different decomposition layer numbers to obtain periodic components, and respectively carrying out optimization processing on the initial low-frequency components and the initial high-frequency components through the periodic components to obtain a plurality of low-frequency components and a plurality of high-frequency components.
3. The feature prediction method for non-stationary time series data according to claim 2, wherein processing the plurality of initial high frequency components according to the structural response monitoring sequences under different decomposition layers to obtain periodic components, and respectively performing optimization processing on the plurality of initial low frequency components and the plurality of initial high frequency components by the periodic components to obtain a plurality of low frequency components and a plurality of high frequency components, comprises: Superposing a plurality of initial high-frequency components according to the structural response monitoring sequences under different decomposition layer numbers to obtain periodic components; calculating according to the periodic component and the environmental load sequence to obtain a dynamic time bending distance; weighting the preset minimum decomposition layer number according to the dynamic time bending distance and a preset entropy value method to obtain an optimal decomposition layer number; and respectively carrying out optimization selection on the plurality of initial low-frequency components and the plurality of initial high-frequency components according to the optimal decomposition layer number to obtain a plurality of low-frequency components and a plurality of high-frequency components.
4. The non-stationary time series data oriented feature prediction method according to claim 1, wherein the sequence model includes an encoder and a decoder, the high frequency prediction result is obtained by predicting a plurality of the high frequency components based on an attention mechanism of a preset sequence model, the attention mechanism includes a feature attention mechanism and a time series attention mechanism, and the method includes: Integrating the high-frequency components to obtain a high-frequency component set; calculating weighted high-frequency components according to the convolution long short-time memory unit in the characteristic attention mechanism, and inputting the weighted high-frequency components into the convolution long-time memory unit for time sequence information iteration to obtain an encoder hiding state; calculating according to the last time step hidden state in the decoder and the unit state in the decoder, and carrying out weighted summation on the hidden states in the encoder through the time sequence attention mechanism to obtain an intermediate semantic vector; updating the hidden state in the decoder according to the output result of the intermediate semantic vector and the sequence model of the last time step, and obtaining a high-frequency prediction result by connecting the updated hidden state with the intermediate semantic vector and inputting the connection to a full-connection layer in the time sequence attention mechanism for prediction.
5. The feature prediction method for non-stationary time series data according to claim 4, wherein the calculating the weighted high-frequency component according to the convolution long short-time memory unit in the feature attention mechanism, inputting the weighted high-frequency component into the convolution long-short time memory unit for time series information iteration, and obtaining the encoder hiding state comprises: Calculating the hiding state of the last time step in the high-frequency component set and the encoder according to the convolution long short-time memory unit to obtain the hiding state of the current time step; performing convolution operation based on the last time step hiding state in the encoder and the first component in the high-frequency component set to obtain attention weight at the current moment; Carrying out normalized exponential function calculation on the attention weight at the current moment to obtain the importance degree of the structural response monitoring sequence; constructing based on the importance degree of the structural response monitoring sequence and the preset structural response characteristic at the current moment to obtain a weighted high-frequency component; And inputting the weighted high-frequency component into the convolution long-short time memory unit, and performing time sequence information iteration on the hidden state of the current time step to obtain the hidden state of the encoder.
6. The feature prediction method for non-stationary time series data according to claim 4, wherein the calculating according to the last time step hidden state in the decoder and the unit state in the decoder, and the weighting and summing the hidden states in the encoder through the time series attention mechanism, to obtain the intermediate semantic vector, comprises: Calculating according to the hidden state of the last time step in the decoder and the unit state in the decoder to obtain the attention weight of the hidden state; Carrying out normalized exponential function calculation on the attention weight of the hidden state to obtain the influence degree of the hidden state; And carrying out weighted summation on the influence degree of the hidden state and the hidden state in the encoder based on the time sequence attention mechanism to obtain an intermediate semantic vector.
7. The characteristic prediction device for non-stationary time series data is characterized by comprising the following components: The acquisition module is used for acquiring a multi-source monitoring sequence, wherein the multi-source monitoring sequence comprises a structural response monitoring sequence and an environmental load sequence; the selection decomposition module is used for calculating according to the time sequence in the multi-source monitoring sequence and a preset wavelet base sequence to obtain a waveform coefficient, and performing optimal wavelet decomposition on the multi-source monitoring sequence under different decomposition layers through the waveform coefficient to obtain a plurality of low-frequency components and a plurality of high-frequency components; The first prediction module is used for inputting a plurality of low-frequency components into a preset non-stationary attention mechanism to perform sequence stationary processing and prediction to obtain a low-frequency prediction result; The second prediction module is used for predicting a plurality of high-frequency components based on an attention mechanism of a preset sequence model to obtain a high-frequency prediction result, wherein the attention mechanism comprises a characteristic attention mechanism and a time sequence attention mechanism; and the construction module is used for reconstructing based on the low-frequency prediction result and the high-frequency prediction result to obtain a time sequence characteristic prediction result.
8. The non-stationary time series data oriented feature prediction apparatus of claim 7, wherein the selection decomposition module comprises: An acquisition unit configured to acquire a plurality of decomposition levels; The first calculation unit is used for calculating according to the time sequence and the wavelet base sequence to obtain a waveform coefficient; the selection unit is used for carrying out discrete processing on the waveform coefficients, carrying out waveform similarity analysis on the time sequence and the wavelet base sequence through the discrete waveform coefficients, and carrying out optimal wavelet base selection according to an analysis result to obtain an optimal wavelet base; The decomposition unit is used for carrying out wavelet decomposition on the multi-source monitoring sequence according to the optimal wavelet base to obtain a plurality of initial low-frequency components and a plurality of initial high-frequency components; The first processing unit is used for processing the initial high-frequency components according to the structural response monitoring sequences under different decomposition layer numbers to obtain periodic components, and optimizing the initial low-frequency components and the initial high-frequency components through the periodic components to obtain the low-frequency components and the high-frequency components.
9. The non-stationary time series data oriented feature prediction apparatus of claim 8, wherein the first processing unit comprises: the superposition subunit is used for superposing a plurality of initial high-frequency components according to the structural response monitoring sequences under different decomposition layer numbers to obtain periodic components; the second calculating subunit is used for calculating according to the periodic component and the environmental load sequence to obtain a dynamic time bending distance; the first processing subunit is used for carrying out weighting processing on the preset minimum decomposition layer number according to the dynamic time bending distance and a preset entropy value method to obtain an optimal decomposition layer number; and the selection subunit is used for respectively carrying out optimization selection on the plurality of initial low-frequency components and the plurality of initial high-frequency components according to the optimal decomposition layer number to obtain a plurality of low-frequency components and a plurality of high-frequency components.
10. The non-stationary time series data oriented feature prediction apparatus of claim 7, wherein the second prediction module comprises: The second processing unit is used for carrying out integration processing on the high-frequency components to obtain a high-frequency component set; The iteration unit is used for calculating weighted high-frequency components according to the convolution long-short-time memory unit in the characteristic attention mechanism, and inputting the weighted high-frequency components into the convolution long-short-time memory unit for time sequence information iteration to obtain the hidden state of the encoder; The second calculating unit is used for calculating according to the last time step hidden state in the decoder and the unit state in the decoder, and carrying out weighted summation on the hidden states in the encoder through the time sequence attention mechanism to obtain an intermediate semantic vector; And the updating unit is used for updating the hidden state in the decoder according to the output result of the intermediate semantic vector and the sequence model in the last time step, and obtaining a high-frequency prediction result by connecting the updated hidden state with the intermediate semantic vector and inputting the connection result into a full-connection layer in the time sequence attention mechanism for prediction.

Description

Feature prediction method and device for non-stationary time sequence data Technical Field The invention relates to the technical field of time sequence prediction, in particular to a feature prediction method and device for non-stationary time sequence data. Background In the technical field of time sequence prediction, due to the structural characteristics of an engineering structure, the diversity of external loads, and the continuous dynamic changes of environmental factors such as weather, traffic and the like, monitoring data such as vibration, displacement and the like of a structure are acted together, so that the engineering structure always shows a complex unstable state characteristic. Disclosure of Invention The present invention aims to provide a feature prediction method for non-stationary time series data, so as to improve the above problems. In order to achieve the above purpose, the technical scheme adopted by the invention is as follows: in a first aspect, the present application provides a feature prediction method for non-stationary time series data, including: Acquiring a multi-source monitoring sequence, wherein the multi-source monitoring sequence comprises a structural response monitoring sequence and an environmental load sequence; Calculating according to a time sequence in the multi-source monitoring sequence and a preset wavelet base sequence to obtain a waveform coefficient, and carrying out optimal wavelet decomposition on the multi-source monitoring sequence under different decomposition layers through the waveform coefficient to obtain a plurality of low-frequency components and a plurality of high-frequency components; inputting a plurality of low-frequency components into a preset non-stationary attention mechanism to perform sequence stationary processing and prediction to obtain a low-frequency prediction result; Predicting a plurality of high-frequency components based on an attention mechanism of a preset sequence model to obtain a high-frequency prediction result, wherein the attention mechanism comprises a characteristic attention mechanism and a time sequence attention mechanism; and reconstructing based on the low-frequency prediction result and the high-frequency prediction result to obtain a time sequence characteristic prediction result. In a second aspect, the present application further provides a feature prediction apparatus for non-stationary time series data, including: The acquisition module is used for acquiring a multi-source monitoring sequence, wherein the multi-source monitoring sequence comprises a structural response monitoring sequence and an environmental load sequence; the selection decomposition module is used for calculating according to the time sequence in the multi-source monitoring sequence and a preset wavelet base sequence to obtain a waveform coefficient, and performing optimal wavelet decomposition on the multi-source monitoring sequence under different decomposition layers through the waveform coefficient to obtain a plurality of low-frequency components and a plurality of high-frequency components; The first prediction module is used for inputting a plurality of low-frequency components into a preset non-stationary attention mechanism to perform sequence stationary processing and prediction to obtain a low-frequency prediction result; The second prediction module is used for predicting a plurality of high-frequency components based on an attention mechanism of a preset sequence model to obtain a high-frequency prediction result, wherein the attention mechanism comprises a characteristic attention mechanism and a time sequence attention mechanism; and the construction module is used for reconstructing based on the low-frequency prediction result and the high-frequency prediction result to obtain a time sequence characteristic prediction result. The beneficial effects of the invention are as follows: According to the method, the optimal wavelet basis is selected through the multisource monitoring sequences under different decomposition layers, the multisource monitoring sequences are decomposed through the optimal wavelet basis to obtain the low-frequency components and the high-frequency components, the low-frequency components are predicted through the non-stationary attention mechanism, key information in the sequences can be dynamically focused, the low-frequency components are subjected to stationary processing, future trends are predicted based on historical data, so that a low-frequency prediction result is obtained, the high-frequency components are predicted through the attention mechanism of the sequence model, key features and time sequence relations in the data can be focused respectively, so that the prediction accuracy and the robustness are improved, the reconstructed prediction result can comprehensively reflect the influence of dynamic behaviors and environmental loads of engineering structures, scientific basis is provi