Fault Diagnosis Decision For Complex Industrial Systems Based On Deep Learning

With the proposal of “Made in China 2025”,intelligent manufacturing has become an innovative project combining a new generation of information technology and production technology and equipment.Industrial systems such as industrial robots and3 D printers have the characteristics of complex structure and high precision.Complex industrial systems are gradually developing towards high-order,nonlinear and multiagent deep coupling.Therefore,fault diagnosis is very important in the health management of industrial systems.Because the traditional fault diagnosis methods have the limitations of insufficient deep feature extraction,relying on manual experience and label guidance,and poor diagnosis performance of mechanical equipment with unclear fault mechanism,this paper proposes the complex industrial system fault diagnosis methods from supervised fault classification to unsupervised clustering based on deep learning algorithm.First,in the supervised learning scenario of complex industrial systems,a basic deep learning model is used to make fault diagnosis decisions.Noticed that the learning purpose of a typical deep learning algorithm is to extract features and obtain a more effective feature representation,but the essence of diagnosis classification is to minimize the intra-class distance and maximize the inter-class distance.Therefore,this paper further proposes to improve the framework of supervised deep learning by fusing deep feature representation and feature enhancement as a starting point.Since the above methods are all based on labeled scenarios with known fault types,however,there are a large number of non-faulty labeled scenarios in complex industrial systems.In view of this situation,this paper also proposes an unsupervised deep clustering method for fault diagnosis.In conclusion,the main research work in this paper for complex system fault diagnosis is as follows:（1）Aiming at the problem that multiple fault signals of industrial systems are difficult to accurately identify under complex working conditions,a fault diagnosis method based on Deep Boltzmann Machine（DBM）is used for analysis.First,Wavelet Packet Transform（WPT）is used to extract statistical features of the original vibration signal under each fault state.Then,DBM is used to deeply mine the statistical features,so as to extract its abstract intrinsic features.Finally,use the classifier to output the fault diagnosis results,and compared with the current mainstream decision classification methods to analyze the performance and parameter sensitivity of the fault diagnosis model based on the Deep Boltzmann Machine.（2）A fault diagnosis enhancement method based on Joint Interclass and Intraclass Mappings（JIIM）is proposed.The experiment mainly involves feature enhancement mapping for sensor signals and feature extraction and industrial system diagnosis classification based on Echo State Network（ESN）.Firstly,the aggregation effect of intraclass distance on similar attributes and the quantitative representation effect of inter-class distance on the distance between different categories are analyzed,so as to construct an ensemble activation function with class distance and intra-class distance as indicators.Then,the ensemble activation function mapping method is used between the original data and the class center to reduce the intra-class decision distance of the deep features in each class structure respectively,and form a new feature matrix.At the same time,each type of fault is uniformly distributed in the feature space,thereby increasing the decisionmaking distance between different categories.Finally,the feature-enhanced dataset is put into the ESN,and a high diagnosis decision category separability can be achieved with a small computational complexity.（3）An unsupervised modeling method for fault diagnosis of complex industrial systems based on Improved Deep Encoder Clustering（IDEC）is proposed.For the scene with unknown labels in the complex industrial system,IDEC is used for the application test of unsupervised deep clustering fault diagnosis.First,the management information system database is formed by collecting the state variable data of the industrial system under different state categories and resampling at equal intervals.Then the fault diagnosis model based on IDEC is applied to the collected data for deep clustering.IDEC fuses the reconstruction loss and clustering loss of data features to guide the model to fine tune and obtain the optimal clustering results.Compared with the traditional shallow clustering algorithm,it effectively improves the fault diagnosis performance of the model in the complex industrial system of unlabeled scenes.