CN-116028810-B - Data and model fusion small sample degradation prediction method

Abstract

The invention discloses a data and model fusion small sample degradation prediction method, which comprises the steps of firstly obtaining a degradation sample D of key equipment in the operation process through a sensor or an equipment recorder, processing the degradation sample through a normalization method, dividing the sample into a degradation training sample and a degradation test sample, learning the degradation amount of the training sample corresponding to each time of the key equipment through a neural network to obtain a relation function h i of the encoded degradation amount and different times, integrating the relation function h i of the different times through an aggregator, calculating the mean value and variance of the integrated function, sampling the mean value and variance to obtain a hidden variable z, and finally learning the function g of the hidden variable z and the degradation test sample through the neural network to obtain a key equipment degradation amount prediction value. The method solves the problem that the degradation modeling and predicting effect of a single traditional random process and a neural network method in the prior art is poor.

Inventors

• XIE GUO
• SHANGGUAN ANQI
• MU LINGXIA
• LI YANKAI
• JIN YONGZE
• ZHANG CHUNLI
• YANG YANXI

Assignees

• 西安理工大学

Dates

Publication Date

20260505

Application Date

20221229

Claims (1)

1. 1. The method for predicting the degradation amount of the small sample by fusing the data and the model is characterized by comprising the following steps of: step 1, acquiring a degradation sample D of key equipment in the operation process through a sensor or an equipment recorder; Step 2, processing the degradation sample by a normalization method, and dividing the sample into a degradation training sample and a degradation testing sample; The step 2 specifically comprises the following steps: step 2.1, set the original degradation sample D 1:m ={T 1:m ,X 1:m , , , Wherein, the An overall degradation sample representing a critical device, Representing the j-th critical device degradation sample, Representing the total number of degraded samples, Indicating the i-th monitoring instant of time, For the entire duration of the monitoring that is recorded, Representing the degradation amount of the jth critical device at the ith moment; Representing the length of time for which the entire sample was observed, Indicating the ith observation time; Step 2.2, normalizing the degradation sample X 1:m of the key equipment into the same dimension by a maximum and minimum normalization method, wherein the process is shown in a formula (1): (1) Wherein, the Indicating the amount of degradation corresponding to the jth critical device at time T i , Is the maximum degradation of the jth critical device over the entire monitoring time, Is the minimum degradation of the jth critical device over the entire monitoring time, Is the normalized degradation amount of the j-th key equipment at the moment T i ; Thus obtaining a normalized degraded sample ={ }, , , Wherein, the Representing a normalized integral degradation sample of the key equipment; step 2.3, assuming the duration of the degradation training sample is n, the duration of the test sample is m-n-1, so as to obtain a normalized degradation training sample as = { Normalized degradation test sample is ={ }; Step 3, learning the degradation amount of training samples corresponding to each moment of key equipment through a neural network to obtain a relation function h i of the encoded degradation amount and different moments; the step 3 specifically comprises the following steps: Obtaining a normalized training sample through the step 2 = { The corresponding degradation amount at each moment is% ),...,( ),...,( ) , wherein, That is, each time T i corresponds to one key device training sample degradation amount, and then learning each time T i and training sample degradation amount through the neural network Relation function between That is, n relation functions are obtained , wherein, The degradation amount of the training sample of the equipment in the T i is the degradation amount training sample duration, NN represents a neural network, NN is a back propagation neural network BPNN or a cycle time network RNN; step 4, integrating the relation functions h i at different moments by using an aggregator, calculating the mean value and the variance of the integrated functions, and sampling the mean value and the variance to obtain a hidden variable z; the step 4 specifically comprises the following steps: step 4.1, obtaining the degradation amount of the training sample and each time T i after training by the neural network through the step 3 Relation function between Using an average aggregator for n relation functions Polymerization is performed as shown in formula (2): (2) Where r is the aggregate result and n is the degradation amount training sample duration; step 4.2, calculating the average value of the aggregation result And variance of Sampling the mean value and the variance to obtain hidden variables obeying normal distribution ; Step5, finally combining the degradation test sample obtained in the step2 with the hidden variable z obtained in the step4, learning a function g of the hidden variable z and the degradation test sample through a neural network to obtain a degradation quantity predicted value of key equipment, The step 5 specifically comprises the following steps: step 5.1, assume a normalized degraded sample Obeying gaussian process When the degradation sample based on the Gaussian process is obtained according to the Bayesian formula Distribution as shown in formula (3): (3) Wherein, the Is a normalized overall degradation sample and, Representing normalized degraded samples Is a function of the probability density of (c) in the (c), Is the overall degradation sample monitoring time and, Is the distribution of the degradation sample monitoring time due to Subject to a gaussian process, Is subject to gaussian distribution and, therefore, As shown in formula (4): (4) bringing equation (4) into equation (3) yields a degraded sample distribution, as shown in equation (5): (5) thus, degenerate samples based on Gaussian process Distribution of Also obeys gaussian distribution; step 5.2, learning each monitoring time in the test sample by the neural network NN Relationship function with hidden variable z The random process f (T) is utilized Description, i.e. to describe a relational function Instead of f (T), the combined test degradation samples D n+1:m = { Equation (5) becomes: (6) Wherein, the Is a joint distribution of hidden variables and the amount of degradation of the test sample, Is a priori distribution of hidden variables, Representing the v-th monitoring time in the test sample A relationship function with the hidden variable z; Step 5.3, according to the integral sum relationship Monitoring time based on overall degradation samples Degradation amount of training sample Testing the degradation amount of the sample The distribution is shown in formula (7): (7) Wherein, the Is based on the degradation of the test sample of the hidden variable z The distribution integration value, Is the overall degradation sample monitoring time Degradation amount of training sample Under the condition of (1) testing the degradation amount of the sample Distribution to obtain the degradation of the test sample Associated with the hidden variable z; Obtaining hidden variables obeying normal distribution through the step 4 Combining logarithm and KL divergence algorithm to obtain the degradation of the test sample The distributed evidence lower bound function is calculated as shown in formula (8): (8) Wherein, the Is a distribution of hidden variables z following a normal distribution, Is based on an approximation of the distribution of the hidden variable z subject to a normal distribution under the condition of the overall degenerate sample, Is based on the approximation of the distribution of hidden variable z subject to normal distribution Is not limited to the desired one; By applying a function to the lower bound of evidence Obtaining the extreme value to obtain the degradation amount of the test sample Distribution of results, then the amount of degradation to the test sample Sampling the distribution to obtain the degradation amount of the test sample Predicted values.

Description

Data and model fusion small sample degradation prediction method Technical Field The invention belongs to the technical field of degradation analysis of key equipment, and particularly relates to a data and model fusion small sample degradation prediction method. Background Along with the rapid development of industrial systems, key equipment with high reliability and long service life becomes an important direction for stably operating the system and reducing the cost. However, since analysis of the degradation process of the device is the basis of reliability analysis and residual life prediction, it is generally necessary to predict the future degradation amount by establishing a degradation process model, and further obtain a correlation model between the degradation amount and the device failure. Therefore, establishing an accurate degradation process model has extremely important significance for improving the safe operation of the system. In predicting the degradation amount of the key device, since the degradation amount belongs to time series samples, it is first necessary to establish a correlation model between the degradation amount of each device and time. Then, the degradation amount at the future time is predicted based on the correlation model between the degradation amount and time. In general, the degradation process of a certain device mainly consists of different degradation amounts of samples, which easily generates uncertainty factors, namely heterogeneity among samples, due to deviations existing in the factory of the device. When the degradation process model is built, since the conventional random process model needs to assume distributions of certain parameters in advance, the distributions deviate from the actual situation to a certain extent, and further influence the credibility of the later analysis results. The machine learning method is mainly based on data analysis to obtain a relation model of degradation amount and monitoring time, no prior assumption is needed for model parameters, but uncertainty of a degradation sample may not be learned. If the advantages of the neural network and the random process method can be fused, the accuracy of the degradation process model can be improved, and the reliability analysis result of the key equipment can be further improved. Disclosure of Invention The invention aims to provide a data and model fusion small sample degradation prediction method, which solves the problems of poor degradation modeling and prediction effects of a single traditional random process and a neural network method in the prior art. The technical scheme adopted by the invention is that the method for predicting the degradation of the small sample by fusing the data and the model is implemented according to the following steps: step 1, acquiring a degradation sample D of key equipment in the operation process through a sensor or an equipment recorder; Step 2, processing the degradation sample by a normalization method, and dividing the sample into a degradation training sample and a degradation testing sample; step 3, learning the degradation amount of training samples corresponding to each moment of key equipment through a neural network to obtain a relation function h i of the encoded degradation amount and different moments; step 4, integrating the relation functions h i at different moments by using an aggregator, calculating the mean value and the variance of the integrated functions, and sampling the mean value and the variance to obtain a hidden variable z; and 5, finally, combining the degradation test sample obtained in the step2 with the hidden variable z obtained in the step4, and learning a function g of the hidden variable z and the degradation test sample through a neural network to obtain a degradation quantity predicted value of the key equipment. The present invention is also characterized in that, The step2 is specifically as follows: step 2.1, set the original degradation sample D 1:m＝{T1:m,X1:m, Where X 1:m represents the overall degradation samples of the critical device, j=1,..i represents the j-th critical device degradation sample, l represents the total number of degradation samples, i=1, m represents the i-th monitoring time, m is the entire monitoring duration recorded, X j,i represents the degradation amount of the j-th key device at the i-th time; T 1:m represents the whole sample observation time period, and T i represents the ith observation time; Step 2.2, normalizing the degradation sample X 1:m of the key equipment into the same dimension by a maximum and minimum normalization method, wherein the process is shown in a formula (1): wherein, X j,i represents the degradation amount of the jth key device corresponding to the time T i, max (X j,：) is the maximum degradation amount of the jth key device in the whole monitoring time, max (X j,：) is the minimum degradation amount of the jth key device in the whole monitoring time, and