CN-116796258-B - Health state prediction method based on self-adaptive Bayesian deep learning

CN116796258BCN 116796258 BCN116796258 BCN 116796258BCN-116796258-B

Abstract

The invention relates to a health state prediction method based on self-adaptive Bayesian deep learning under multi-source uncertainty. An adaptive model-free technique is introduced in a Bayesian deep learning framework to fully exploit the predictive model capabilities. Firstly, a model-free sag (dropout) method is adopted to quantify cognitive uncertainty, the dropout rate and distribution type can be automatically learned, highly nonlinear degradation characteristics can be better captured, secondly, an arbitrary polynomial chaotic expansion (arbitrary polynomial chaos expansion, aPc) method is adopted to quantify random uncertainty, the method can avoid introducing extra subjectivity into assumed distribution from limited samples or sparse information, finally, a health state prediction framework based on adaptive Bayesian deep learning is provided, a network loss function is constructed and training is carried out by utilizing a variation reasoning method, and the cognitive uncertainty and the random uncertainty are quantified uniformly in a model-free mode, so that average prediction and interval prediction of health states are carried out simultaneously.

Inventors

SUN BO
PAN JUNLIN
WU ZEYU
WANG ZILI
FENG QIANG
REN YI
YANG DEZHEN

Assignees

北京航空航天大学

Dates

Publication Date: 20260505
Application Date: 20230625

Claims (10)

1. A health state prediction method based on adaptive Bayesian deep learning is characterized by comprising the following steps: step 1, aiming at a data set of multidimensional input characteristics and one-dimensional health state characterization quantity For the lithium battery health state prediction, the multidimensional input characteristic is constant current charging time and isobaric dropping time characteristic data, the one-dimensional health state representation quantity is the lithium battery health state, Dividing the time sequence sample length into a training set, a testing set and a verification set, and carrying out normalization processing on a data set; Step2 for one Deep layer neural network, definition of Layer [ (layer ] ) Has the following components Neurons, the first Individual neurons [ ] ) Corresponding to the temporary mask dropout mask parameter Related intermediate hidden variables and mapping functions; Step 3, the dropout mask parameter is mapped through the intermediate hidden variable and the mapping function Modeling is carried out, so that the distribution type of the data is adaptively changed along with the data; step 4, parameter prior distribution of dropout mask Modeling, and carrying out parameter estimation by using a maximum likelihood estimation method; Step 5, calculating the network weight parameters after the dropout mask is added Constructing a variation distribution in a variation inference ; Step 6, according to the data set Constructing a multivariable polynomial output model; step 7, for polynomial coefficients in the output model Solving by Galerkin projection method or random response surface method, and obtaining multidimensional orthogonal basis function in output model , Is that Random variable dimension, and random distributed orthogonal basis function constructed by random polynomial chaos expansion method , Is the order; Step 8, training aPc a model by using a training set, and establishing a mapping relation between multidimensional input features and health state characterization quantities under the condition of considering random uncertainty; step 9, data augmentation is carried out on the test set, the augmented data is input into a trained aPc model, and probability information characteristics of the health state characterization quantity are predicted; Step 10, constructing a prediction distribution model under Bayes deep learning, and constructing a network loss function by using a variable-fraction reasoning method according to variable-fraction distribution in cognitive uncertainty quantization and distribution information characteristics in random uncertainty quantization; Step 11, setting network super parameters, training a Bayesian deep learning model according to a network loss function by using a dropout technology and L2 regularization, and using a random gradient descent algorithm to set parameters in the training process Optimizing to obtain optimal model weight, wherein the strain distribution is used for approximating posterior distribution, dropout is not closed in a test stage, and probability characteristics of health state are reserved; and step 12, outputting the results of the healthy state mean value prediction and the interval prediction by using a Monte Carlo sampling method, decomposing the two types of uncertainty, and representing the total uncertainty, the random uncertainty and the cognitive uncertainty.
2. The method for predicting health status based on adaptive Bayesian deep learning according to claim 1, wherein the step 2 is characterized in that The intermediate hidden variable of a layer is defined as Which obeys the intermediate distribution Intermediate hidden variable And (3) with Through a mapping function And (3) associating: (1) In the formula, Is satisfied with 1) the value range is [0,1], 2) the single and tiny arbitrary function within the definition range, Is satisfied 1) with the ability to convert to a standard distribution of microparameters and 2) with an arbitrary distribution that is easy to sample.
3. The method for predicting health status based on adaptive Bayesian deep learning according to claim 1, wherein in said step 3 Is satisfied with Any form of (c): (2) In the formula, Is the inverse of the mapping function when Compliance with , In the case of a Sigmoid function, The distribution of (2) is expressed as: (3) Distribution of (3) By adjusting And The adaptation approximates all types of dropout distributions.
4. The method for predicting health status based on adaptive bayesian deep learning according to claim 1, wherein the prior distribution of parameters in step 4 is defined as: (4) In the formula, Is the first The multi-variable gaussian hidden state parameters of the layer, And Respectively follow a gaussian distribution and an inverse gamma distribution, And Is approximated with a multi-layer perceptron of gaussian output: (5) (6) In the formula, And Is the first Parameters of layers In response to the weight and the bias, And Is the first Parameters of layers In response to the weight and the bias, By maximising Reasoning: (7) In the formula, Is a multi-variable gaussian hidden state parameter.
5. The health state prediction method based on adaptive Bayesian deep learning as set forth in claim 1, wherein the network weight parameters after dropout mask is added in the step 5 Expressed as: (8) In the formula, Is the first Layer original network parameters, definition of the first Parameter set of layer The method comprises the following steps: (9) Then in the variational derivative The method comprises the following steps: (10) In the formula, Is a variation distribution.
6. The method for predicting the health status based on adaptive bayesian deep learning according to claim 1, wherein the multivariate polynomial model of step 6 is expressed as: (11) In the formula, Is the coefficient of the polynomial, Is the number of coefficients, and the truncated order And feature dimension In relation to the use of a liquid crystal display device, Is a multidimensional orthogonal basis function, under the condition of independent input characteristics, Expressed as the product of one-dimensional orthogonal basis functions : (12) In the formula, Is the index of the multiple variables, Random variable corresponding to order The polynomials are: (13) In the formula, Is that Medium coefficient.
7. The method for predicting health status based on adaptive Bayesian deep learning as set forth in claim 1, wherein the distribution model of the health status characterizer in step 9 is represented by a dual-parameter function, namely 。
8. The method for predicting health status based on adaptive bayesian deep learning according to claim 1, wherein the distribution model predicted in step 10 is expressed as: (14) posterior distribution according to Bayes' formula Expressed as: (15) Using a variation distribution To approximate posterior distribution Calculation of And (3) with KL divergence between: (16) In the formula, And model parameters Irrespective of the reason And Defining a lower bound of evidence: (17) In the formula, Approximated by a monte carlo integration, L2 canonical terms considered as model parameters, minimization And (3) with The objective of the KL divergence between is to maximize ELBO, then the loss function of the network is expressed as: (18) In the formula, Is the model attenuation coefficient and variation distribution Determined according to the formula (10), Based on aPc random uncertainty quantitative determination, the health state characterization quantity is subjected to two parameters For the normal distribution, the loss function is expressed as (19)。
9. The method for predicting health status based on adaptive bayesian deep learning according to claim 1, wherein in the step 11 The prediction distribution under sub-sampling is expressed as: (20) In the formula, , And Is the number of samplings.
10. The method for predicting health status based on adaptive bayesian deep learning according to claim 1, wherein in the step 12 The mean of the subsampled downsampled predictions is expressed as: (21) Expressed as: (22) Expressed as: (23) Expressed as: (24) In the formula, Is the variance that characterizes the total uncertainty, Is a variance that characterizes the random uncertainty, Is a variance that characterizes cognitive uncertainty.

Description

Health state prediction method based on self-adaptive Bayesian deep learning Technical Field The invention provides a health state prediction method based on a self-adaptive Bayes deep learning framework, which is a health state prediction method considering uncertainty quantification, can obtain point estimation and interval estimation results of a system health state, and belongs to the technical field of fault Prediction and Health Management (PHM). Background The multiple increase in hardware, software, personnel, organizations, etc. of modern complex industrial systems increases the complexity of the system, which in turn increases the likelihood of system failure, which often results in unacceptable personnel and property loss during long-term operation. Fault prediction and health management (Prognostics AND HEALTH MANAGEMENT, PHM) techniques can ensure safe and stable operation of complex industrial systems. In PHM, state of health (SOH) prediction is the basis of health management, and it is necessary to provide accurate and reliable input for health management decisions. With the development of sensors and computer technologies, modern industrial systems often increasingly use network-connected devices, and various sensor packages generate a large amount of multidimensional data, so that data driving methods represented by deep learning are widely applied to health state prediction. Models developed using deep learning, however, require assessment of their reliability and effectiveness due to the characteristics of the black box problem, as predictions of these models are subject to data noise and errors of the model itself. Most of the existing deep learning-based health state prediction models obtain point estimation results, but the uncertainty influence is ignored, but the prediction results are often unreliable for reasoning decisions. Deep learning based state of health predictions require quantification of two types of uncertainties—cognitive uncertainties due to model self reasoning and random uncertainties due to data noise. Cognitive uncertainty is a neural network architecture, model parameters, and hyper-parameter cognitive deficit due to data information depletion, which can be eliminated in cases where sufficient amounts of information-rich data are included. Random uncertainty is an inherent property of data that cannot be eliminated even if more data is collected. Two kinds of uncertainties in deep learning require a unified architecture to perform quantization and comprehensive processing to improve the credibility of the prediction result. Almost all current approaches quantify uncertainty in deep learning by introducing model-based hypothesis priors, which limit the modeling and predictive capabilities of the approach. In view of this, the invention proposes a model-free adaptive Bayes deep learning framework to comprehensively consider random uncertainty and cognitive uncertainty for health state prediction, so as to realize health state probability prediction with multi-source uncertainty fusion, and provide reliable decision support for health management. Disclosure of Invention The invention discloses a health state prediction method based on self-adaptive Bayes deep learning, which introduces a self-adaptive model-free technology into a Bayes deep learning framework to fully exert the capability of a prediction model. Firstly, a model-free sag (dropout) method is adopted to quantify the cognitive uncertainty, the dropout rate and the distribution type can be automatically learned, the highly nonlinear degradation characteristic can be better captured, secondly, an arbitrary polynomial chaotic expansion (arbitrary polynomial chaos expansion, aPc) method is adopted to quantify the random uncertainty, the method can avoid introducing extra subjectivity into the assumed distribution from limited samples or sparse information, finally, a health state prediction framework based on adaptive Bayesian deep learning is provided, a network loss function is constructed and training is carried out by utilizing a variational reasoning method, and the cognitive uncertainty and the random uncertainty are uniformly quantified in a model-free mode, and the method flow is shown in a figure 1 and concretely comprises the following steps: step 1, aiming at a data set of multidimensional input characteristics and one-dimensional health state characterization quantity ,Dividing the time sequence sample length into a training set, a testing set and a verification set, and carrying out pretreatment such as normalization on a data set; Step2 for one Deep layer neural network, definition ofLayer [ (layer ]) Has the following componentsNeurons, the firstIndividual neurons [ ]) Corresponding to the temporary mask dropout mask parameterRelated intermediate hidden variables and mapping functions; Step 3, the dropout mask parameter is mapped through the intermediate hidden variable and the mapping functio