CN-122021227-A - KL feedback-based complex response surface confidence assessment system and method

CN122021227ACN 122021227 ACN122021227 ACN 122021227ACN-122021227-A

Abstract

A complex response surface confidence assessment system and method based on KL feedback comprises an experiment design module, a confidence assessment module and a response surface design module, wherein the experiment design module obtains initial simulation input control variable design points through a uniform and orthogonal experiment design method, industrial data are obtained through loading control variable points in a corresponding industrial simulation system to simulate an industrial operation process, the data are input into an industrial fault identification model to be trained, a model training statistical condition is obtained, the confidence assessment module selects a corresponding confidence assessment method to carry out confidence assessment model fitting treatment according to the training statistical condition to obtain a fitted assessment model, and the response surface design module carries out KL deviation calculation according to response information of the initial control variable points in the industrial simulation system to obtain a surface design point result. The invention can scientifically and reasonably carry out statistical inference and decision on the testing level of the initial use stage of the fault identification device, and can effectively improve the design level of the response curved surface of the design method.

Inventors

LI JIANXUN
JIANG NAN

Assignees

上海交通大学

Dates

Publication Date: 20260512
Application Date: 20241111

Claims (8)

1. A complex response surface confidence evaluation system based on KL feedback is characterized by comprising an experiment design module, a confidence evaluation module and a response surface design module, wherein the experiment design module obtains initial simulation input control variable design points through a uniform and orthogonal experiment design method, industrial data are obtained through loading control variable points in a corresponding industrial simulation system to simulate an industrial operation process, the data are input into an industrial fault recognition model to be trained, a model training statistical condition is obtained, the confidence evaluation module selects a corresponding confidence evaluation method to carry out confidence evaluation model fitting treatment according to the training statistical condition to obtain a fitted evaluation model, and the response surface design module carries out KL deviation calculation according to response information of the initial control variable points in the industrial simulation system to obtain a surface design point result.
2. The KL feedback-based complex response surface confidence evaluation system as claimed in claim 1, wherein the method for designing the uniform and orthogonal experiment is to select a combination of columns with the best uniformity according to the number of factors to design a test scheme, and uniformly distribute all test points in the whole sample space, and specifically comprises: Wherein P n is an experiment design point set, x i is an i-th experiment design point in the point set, n is the initial set number of experiments, k is the number of factors required to be considered in the experiment design, and C k is a factor sample space constructed by experimental factors.
3. The KL feedback-based complex response surface confidence assessment system is characterized in that the industrial fault identification model comprises a data input layer, a feature extraction layer and a fault detection layer, wherein the data input layer is used for carrying out 0-1 standardization processing on industrial data from an industrial simulation system to obtain industrial data with consistent scales, the feature extraction layer is used for processing the processed industrial data by using a cyclic neural network to obtain depth features of the industrial data, and the fault detection layer is used for carrying out nonlinear mapping between features and faults by using a Softmax classifier on the extracted depth features to obtain a fault identification result.
4. The complex response surface confidence assessment system based on KL feedback according to claim 1, characterized in that the confidence assessment model is a 0-1 distribution-based confidence assessment method or a Wald statistic-based confidence assessment method; The confidence evaluation method based on 0-1 distribution evaluates the confidence of the related event by recognizing the fault as a binomial distribution event and adopting 0-1 distribution, and obtains a confidence interval (p 1 ,p 2 ) meeting a confidence of 1-alpha in binomial distribution under a obeying parameter p through the model, wherein: n is the sample data capacity, z α/2 is the threshold value corresponding to the standard normal distribution for a given confidence level 1-alpha, Is the sample mean value; The confidence evaluation method based on Wald statistics constructs a confidence interval by considering that the fault recognition accuracy and the fault have a similar linear relation, obtaining parameters and a variance-covariance matrix thereof through maximum likelihood estimation, obtaining a response variable for the value X of each prediction variable X under the condition that the confidence is 1-alpha through the model The confidence interval of (2) is: Wherein: Standard error of (2) Is to estimate the estimated values of the linear model parameters beta 0 and beta 1 using a maximum likelihood estimation method, And The standard error of the parameter is extracted from the variance-covariance matrix.
5. The complex response surface confidence assessment method based on the system of any one of claims 1-4 is characterized in that initial industrial simulation system input control variable design points are obtained by means of a uniform and orthogonal experimental design method, sensor variables are obtained through a corresponding industrial simulation system, data are input into an industrial fault recognition model for training, the obtained training statistics are input into a confidence assessment module, the confidence assessment model is selected and fitted according to fault recognition characteristics, and then the experimental design module is guided to simulate the design of input control variables through KL point feedback in the response surface design module, so that statistical inference and decision on the testability level of an industrial fault recognition device are better carried out.
6. The complex response surface confidence evaluation method according to claim 5, which is characterized by comprising the following steps: firstly, obtaining initial simulation input control variable design points by using a uniform and orthogonal experimental design method, obtaining industrial simulation data through a corresponding simulation system, and inputting the data into an industrial fault recognition model for training; secondly, inputting training statistics obtained by the industrial fault recognition model into a confidence evaluation module, and selecting and fitting the confidence evaluation model based on 0-1 distribution and the confidence evaluation model based on Wald statistics based on the fault recognition characteristics; the third step, the response surface design module designs a proper response surface based on KL point design and combines the control input variable characteristics, and feeds back and guides the design of the simulation input control variable and the training of the fault recognition model, and specifically comprises the following steps: 3.1 obtaining the prior knowledge by obtaining the sampling points and the corresponding response values through the conventional sampling method when designing the response curved surface, specifically, the kernel density estimation based on the data { x i }, namely, the kernel density estimation function Wherein the bandwidth is N is the number of samples, H n is the bandwidth, K is the kernel function, d is the dimension of the data, |k|| 2 is the second order norm of the kernel function, μ 2 (K) is the second moment of the kernel function, and H f (x) is the second derivative matrix of the probability density function; 3.2 calculating the KL deviation between the kernel Density estimation function f n (x) and the target Density function g (x) When the KL deviation is minimized, the KL deviation is optimized and discretized for avoiding the calculation disaster caused by high-dimensional integration, and specifically comprises the following steps: Wherein D (f n I f) represents the KL deviation between the kernel density estimation function f n (x) and the target density function f (x), and x i is the generated ith KL point; 3.3 generating KL points by adopting a sequential algorithm, wherein the first point is taken as the global maximum value of f (x), and the mth KL point is taken as Meanwhile, a random Nelder-Mead optimization algorithm is adopted to utilize the pseudo model Instead of f (x), the mth KL point after optimization is: thereby avoiding trapping in the local area; And fourthly, inputting actual industrial fault data into the trained fault recognition model to realize confidence evaluation test in the test stage.
7. The complex surface confidence evaluation method according to claim 6, wherein the orthogonal experimental design method divides the experimental parameter space into uniform regions, and selects representative experimental conditions in each region; the experimental conditions are preferably a set of sample points with good distribution characteristics to cover as much as possible all aspects of the parameter space; The orthogonal experiment design method is based on LHS sampling, and the random variable is assumed to be generated by a known distribution function, so that the input probability distribution is layered, and the obtained sampling experiment points can effectively represent the whole sample space.
8. The method for evaluating the confidence coefficient of the complex response surface according to claim 6, wherein the selection of the method for evaluating the confidence coefficient is performed by judging whether the success rate of the fault recognition is related to the characteristics of the fault itself or not based on the characteristics of the fault recognition, selecting a confidence coefficient evaluating model based on 0-1 distribution when the success rate of the fault recognition is not related to the fault itself, and selecting a confidence coefficient evaluating model based on Wald statistics when the success rate of the fault recognition is related to the fault itself.

Description

KL feedback-based complex response surface confidence assessment system and method Technical Field The invention relates to a technology in the field of equipment testing, in particular to a complex response surface confidence assessment system and method based on feedback of Kullback-Leibler (KL). Background The experimental design of the existing equipment test mainly adopts a space filling design, so that design points are uniformly distributed, but the trend of a response curved surface is difficult to capture, and the design points are wasted. Meanwhile, the existing confidence assessment depends on machine learning, but industrial fault data are scarce and unbalanced, and model under-fitting and generalization capability are easy to be low. Disclosure of Invention Aiming at the problems that the existing experimental design method in the prior art cannot fully fill a sample space and has poor generalization capability, the invention provides a complex response surface confidence evaluation system and method based on KL feedback, which can scientifically and reasonably carry out statistical inference and decision on the testing level of the fault identification device in the initial stage of use and can effectively improve the response surface design level of the design method. The invention is realized by the following technical scheme: The invention relates to a complex response surface confidence assessment system based on KL feedback, which comprises an experiment design module, a confidence assessment module and a response surface design module, wherein the experiment design module obtains initial simulation input control variable design points through a uniform and orthogonal experiment design method, industrial data are obtained through loading control variable points in a corresponding industrial simulation system to simulate an industrial operation process, the data are input into an industrial fault identification model to be trained, a model training statistical condition is obtained, the confidence assessment module selects a corresponding confidence assessment method to carry out confidence assessment model fitting treatment according to the training statistical condition, a fitted assessment model is obtained, and the response surface design module carries out KL deviation calculation according to response information of the initial control variable points in the industrial simulation system, so that a surface design point result is obtained. The method for designing the uniform and orthogonal experiments comprises the steps of selecting the combination of columns with the best uniformity according to the number of factors to design an experiment scheme, and uniformly distributing all the experiment points in the whole sample space, wherein the method comprises the following steps: Wherein P n is an experiment design point set, x i is an i-th experiment design point in the point set, n is the initial set number of experiments, k is the number of factors required to be considered in the experiment design, and C k is a factor sample space constructed by experimental factors. The industrial simulation system is a TE simulation system (TENNESSEE EASTMAN Process Simulation System). The simulation system is a classical simulation model for simulating complex chemical processes. The simulation platform is proposed by tennessee chemical Company (TENNESSEE EASTMAN Company) and used for researching and verifying aspects such as fault detection. The TE simulation system simulates a complex chemical process through 12 input control variables, so as to output process variable data in the industrial process. The industrial fault identification model comprises a data input layer, a feature extraction layer and a fault detection layer, wherein the data input layer is used for carrying out 0-1 standardization processing on industrial data from an industrial simulation system to obtain industrial data with consistent dimensions, the feature extraction layer is used for processing the processed industrial data by using a cyclic neural network to obtain depth features of the industrial data, and the fault detection layer is used for carrying out nonlinear mapping between the features and faults by using a Softmax classifier on the extracted depth features to obtain a fault identification result. The confidence evaluation model is a confidence evaluation method based on 0-1 distribution or a confidence evaluation method based on Wald statistics. The confidence evaluation method based on the 0-1 distribution evaluates the confidence of related events by recognizing faults as binomial distribution events and adopting the 0-1 distribution. Obtaining a confidence interval (p 1,p2) with a confidence level of 1-alpha in the binomial distribution under the obeying parameter p through the model, wherein: n is the sample data capacity, z α/2 is the threshold value corresponding to the standard normal distribution for