CN-121980376-A - Residual life prediction method based on depth residual error shrinkage TCN

CN121980376ACN 121980376 ACN121980376 ACN 121980376ACN-121980376-A

Abstract

The invention relates to a residual life prediction method based on depth residual error shrinkage TCN. The method comprises the following steps of firstly, carrying out adaptive noise total-set empirical mode decomposition on vibration data of a vibration component, extracting multi-domain statistical characteristics, constructing health indexes based on monotonicity, trending and predictability weighted fusion, and dividing the health indexes into a health stage and a degradation stage by using a 3 sigma method. And secondly, constructing a depth residual error shrinkage TCN, and inputting the health index in the first step into the depth residual error shrinkage TCN to obtain the intermediate characteristic of life prediction. And finally, further screening key information in the intermediate features by combining a multi-scale attention mechanism and obtaining a prediction result of the final residual service life. Experiments prove that the method can effectively improve the residual life prediction precision of the vibration part, and compared with the current advanced method, the root mean square error is reduced by at least 31%, and the average absolute error is reduced by at least 24%.

Inventors

WANG PENG
WANG HAO
LI MINGYUAN
WANG CHU
LI XIAOYAN
LV ZHIGANG
DONG MIANMIAN
GAO HUI
WANG BIN
WANG GUANQUN

Assignees

西安工业大学

Dates

Publication Date: 20260505
Application Date: 20251211

Claims (5)

1. The residual life prediction method based on depth residual shrinkage TCN is characterized by comprising the following steps: Firstly, extracting multi-domain statistical characteristics by carrying out self-adaptive noise total-set empirical mode decomposition on vibration data of a vibration component, constructing health indexes based on monotonicity, trending and predictability weighted fusion, and dividing the health indexes into a health stage and a degradation stage by using a 3 sigma method; Step two, constructing a depth residual error shrinkage TCN, and inputting the health index of the step one into the depth residual error shrinkage TCN to obtain an intermediate characteristic of life prediction; and thirdly, screening key information in the intermediate features by combining a multi-scale attention mechanism and obtaining a prediction result of the final RUL.
2. The method for predicting residual life based on depth residual shrinkage TCN according to claim 1, wherein the monotonicity Mon (X) in the step one is calculated as follows: Wherein K represents the total number of data points, no. d/dx represents the derivative of two adjacent data points, and is greater than zero for the number of rising trends and less than zero for the number of falling trends; Trend of The calculation method of (2) is shown in a formula (6); where K represents the total number of data points, Normalized data representing the characteristic value of the object, Normalized data representing time values; the method for calculating the predictability Pro is shown in a formula (7); Wherein x l and x k respectively represent the characteristic values of any two different time points, mean|x l -x k | represents the total average value after the difference value is calculated by all the data points pair by pair, and the overall change amplitude of the characteristic values is measured; the calculation method of the weighted fusion HI is as follows: w i ＝Mon i +Tre i +Pro i (9) where W is the sum of monotonicity, trending and predictability of all statistical characteristics, M is the total number of characteristic curves, Mon i 、Tre i 、Pro i represents monotonicity, trend and predictability of the ith statistical characteristic curve respectively; The weight of the ith statistical feature is obtained by calculating the ratio of W i to W, so that the important statistical feature is enhanced, and the final health index contains more degradation information.
3. The method for predicting residual life based on depth residual shrinkage TCN according to claim 2, wherein in the first step, the 3 sigma phase division is calculated as follows: Wherein HI t represents the health index at time t and σ represents the root mean square value.
4. The method for predicting residual life based on depth residual shrinkage TCN according to claim 3, wherein in the second step, the depth residual shrinkage TCN is composed of a residual structure, a soft threshold function and a channel attention mechanism.
5. The method for predicting residual life based on depth residual shrinkage TCN according to claim 4, wherein the middle soft threshold function and its derivative in the second step are shown in the formulas (13) and (14); Wherein τ is a preset threshold parameter; in the formula (13), when x is greater than a threshold value, only the part exceeding the threshold value is reserved, when- τ is less than or equal to x and less than or equal to τ, the input is in the threshold value range, the input is directly set to 0, and when x is less than or equal to- τ, the input is less than a negative threshold value, and only the part lower than the negative threshold value is reserved.

Description

Residual life prediction method based on depth residual error shrinkage TCN Technical Field The invention relates to the technical field of fault prediction and health management, in particular to a residual life prediction method based on depth residual error shrinkage TCN. Background The PHM technology can realize fault early warning, life prediction and maintenance strategy optimization by monitoring and analyzing the state data of the whole life cycle of the equipment. The core idea is to replace the traditional passive maintenance with active prevention, and reduce the fault risk from the source. For example, for an air compressor in a fuel cell, bearings in its vibrating parts are critical components affecting the reliability of the system. The accurate prediction of the bearing RUL can identify potential faults in advance and trigger early warning, effectively avoid shutdown and production stoppage caused by sudden failure, and has important practical value in engineering practice. At present, the TCN network can accurately capture the long-term degradation information of the bearing through a unique structure of causal convolution and expansion convolution, obtain higher residual life prediction precision, and draw a great deal of attention in the stage of predicting the residual life of the bearing. The patent with application number of 202411262483.2 is a bearing residual life prediction method based on MA and MSTCN, the thought is to combine a multi-head attention mechanism and multi-scale convolution to improve the prediction precision of a TCN network, but the problem is that degradation information contained in health indexes of the TCN network is not abundant enough, and when the TCN network is deepened, gradient paths are longer, the related convolution kernels and weights are more, and gradient explosion phenomenon can occur due to the fact that expansion convolution kernels are continuously enlarged. The patent with application number of 202111181738.9 is an image super-resolution reconstruction method based on a novel depth residual error shrinkage network, and the idea is that a depth residual error shrinkage module is introduced to improve the network structure, so that the definition of a reconstructed picture is effectively enhanced. However, the problem is that the gradient explosion problem caused by the deepening of the layer number of the neural network is not solved. The patent with application number 202510097179.5 is a rolling bearing Health state assessment and prediction method based on improved TCN, and the idea is to screen different time-frequency characteristics as Health Indicators (HI) through an improved algorithm and predict the RUL of the bearing through the TCN with a self-attention mechanism. However, the method has the problems that the simple dimension reduction process is directly carried out, the degradation information contained in the HI is not abundant, and the gradient explosion phenomenon can occur when the number of layers of the TCN network is increased. The problem that the residual life prediction accuracy of the vibration part is not high can not be solved in the above documents, so that the problem that the gradient explosion phenomenon occurs when the degradation information is not rich enough and the network layer number is deepened is to be solved in the prior art. Disclosure of Invention The invention provides a residual life prediction method based on depth residual error shrinkage TCN, which aims to solve the problems of insufficient degradation information and gradient explosion phenomenon when the number of network layers is deepened in the prior art. In order to achieve the purpose, the technical scheme of the invention is as follows, the residual life prediction method based on depth residual error shrinkage TCN comprises the following steps: Firstly, extracting multi-domain statistical characteristics by carrying out self-adaptive noise total-set empirical mode decomposition on vibration data of a vibration component, constructing health indexes based on monotonicity, trending and predictability weighted fusion, and dividing the health indexes into a health stage and a degradation stage by using a 3 sigma method; Step two, constructing a depth residual error shrinkage TCN, and inputting the health index of the step one into the depth residual error shrinkage TCN to obtain an intermediate characteristic of life prediction; and thirdly, screening key information in the intermediate features by combining a multi-scale attention mechanism and obtaining a prediction result of the final RUL. Further, the method for calculating monotonicity Mon (X) in the first step is as follows: Wherein K represents the total number of data points, no. d/dx represents the derivative of two adjacent data points, and is greater than zero for the number of rising trends and less than zero for the number of falling trends; Trend of The calculation method of (2) is