Research On Fault Detection Using Generative Adversarial Network

In the modern industrial production process,how to detect faults immediately and precisely is of great importance to the process safety and continuity.With the development of production process intelligence,automation and industrial big data,a large number of sensor signals in industrial processes have been recorded,and data-driven fault detection methods have become mainstream.Data-driven models can be divided into statistical analysis models,machine learning models,deep learning models,etc.according to the modeling method.As one of the deep learning methods,generative adversarial networks are used for unsupervised training through the idea of game confrontation,and the resulting model can be very good.to solve problems such as prediction and classification.Therefore,this paper adopts the generative adversarial network as the fault detection model and conducts research and improvement from data preprocessing,network model,statistics calculation,detection of specific types of faults,etc.The main research contents are as follows:(1)In the traditional fault detection method based on generative adversarial network,the generator input uses random noise and the network training effect is not good.A fault detection strategy of generative adversarial network using encoded input is proposed.An autoencoder is introduced and the hidden variable information after dimensionality reduction is used as the input of the generator.In addition,considering the high computational cost and sensitivity to outliers of generator-based statistics in fault detection methods,a new statistical calculation method is proposed based on latent variables extracted from autoencoders.The effectiveness and improvement of modified method is verified by Tennessee Eastman process and actual coal mill process simulation.(2)In order to improve the detection effect of traditional fault detection methods based on generative adversarial networks for small offset and pulse oscillation type faults,a fault detection method based on multi-block information and generative adversarial network is proposed.This method defines the extraction method of cumulative information and rate of change information,and obtains multiple sub-blocks through information extraction.Then,generative adversarial network is used to establish a detection model for each sub-block and new statistics are obtained through Bayesian fusion.Through numerical cases,Tennessee Eastman and blast furnace ironmaking process simulation experiments,it is verified that multiple pieces of information can extract hidden information in the process,thereby improving the model’s detection results for micro-offset and pulse oscillation type faults.(3)In the fault detection method based on generative adversarial network,most models only consider the different characteristics of a single sample for statistical calculation and ignore the time series characteristics.Therefore,a fault detection method based on the time series correlation between samples and generative adversarial network is proposed.In this method,the long short-term memory networks is introduced into the discriminator model in the generative adversarial network and the remaining time samples are considered when calculating the statistics of the samples to be tested to extract their time series features.The difference between the sample and the reconstructed output of the noise reduction autoencoder is further used as the network input,which reduces the flooding of abnormal information caused by faults caused by normal information in the sample and improves the detection effect of small deviations and other difficult-to-detect faults.Through the simulation of numerical case,Tennessee Eastman process and blast furnace ironmaking process,it is verified that the proposed method improve the detection performance indeed.