| Manufacturers are integrating new technologies,including Internet of Things(Io T),and AI and machine learning into their facilities and operations.These factories are equipped with advanced sensors,embedded software and robotics that collect and analyze data.Even higher value data is created when it is combined with machine learning methods to get new levels of insight.Also,benefitting from this is the fault classification task,which is crucial for safe operation of modern factories.A fault classification system monitors the operating state of equipment and uses intelligent algorithms to identify the operating state of the whole process.For intelligent fault classification,data from the industrial processes needs to be carefully analyzed and mined out off important fault-related features.Because industrial process data generally contains strong correlations,high complexity,and nonlinear patterns,fault classification requires a proficient deep learning model.Deep learning demonstrates its great power in processing complicated data.Recent research has shown that Stacked Autoencoder(SAE)based deep learning models can learn deep abstract features from complicated process data.However,a typical SAE cannot extract deep fault-relevant features from input data due to its unsupervised self-reconstruction pre-training.Therefore,to overcome the shortcomings of a typical SAE and other traditional deep learning models,this research focuses on the development of improved deepfeature extraction methods for industrial process fault classification.The three main research contributions of this thesis are as follows:First of all,this thesis creates an understanding of the correlations present in the process data using the autocorrelation study on process variables.This study exposes that data samples over a short time period exhibit strong correlations,which contain valuable information of the process’ s time-series dynamics and this also gives an idea for selection of length of the active time frame used in the dynamic data augmentation process.Therefore,to accommodate the time-series dynamic characteristics of industrial process data,an active time-frame technique is designed for an SAE model to process and augment the incoming process data.The recorded process data is passed through a time-frame of a certain length and the frame captures data values of that period of time and treats it as one data sample to study the flow of data over the specified time period.The inclusion of dynamic process data characteristics in this way leads to the extraction of enhanced and more informative data features that help the deep learning model to understand the process data better.Experimental results using the Tennessee–Eastman(TE)benchmark process and a real-world industrial hydrocracking process show improvements in fault classification performance.Secondly,as the performance of any model majorly depends on the quality of features it can extract from the input data,a new approach to feature extraction is formulated for an SAE model.Thus,to further improve the feature extraction method of an SAE,an Inter-Relational Mahalanobis(IRM)loss function is developed to create the IRM-SAE model.The aim of this model is to analyze the process data with a multi-objective approach.The multi-objective unsupervised loss function binds Mahalanobis loss,an Inter-Relational(IR)loss,and the SAE’s original mean squared reconstruction loss.This way,the extracted features contain more information and meaningful patterns of the incoming process data.Moreover,this model also makes use of the abovementioned dynamic data augmentation technique to obtain dynamic data information as well.These welllearned comprehensive features now provide a better foundation for the classifier to work on and better classify faults in a process.Experimental results compare the performance of the original SAE,the SAE with IR loss,the SAE with Mahalanobis loss,and the SAE with IRM loss.Among these models,the SAE with IRM loss produces the best results in fault classification.Finally,considering the problem that a classification-specific task needs features that contain maximum information regarding that particular classification task,a semi-supervised pre-training strategy is developed for the IRM-SAE model.This strategy plays a vital role in the extraction of fault-related features that contain maximum information related to fault patterns in process data so that the model can easily categorize data samples from different fault types.The semi-supervised pre-training uses both supervised and unsupervised losses in a special pre-training model architecture and guides the feature extraction process towards fault-relevant information in process data.This creates the Inter-Relational Mahalanobis SAE(IRM-SAE)semi-supervised model.Combining this model with the active time frame technique,the model finds unique dynamic features in input data.This complete method improves fault classification by learning faultfeatures that concurrently encode data distribution and internal relationships among data samples.Performance of this model is validated by data from the two processes considered in this research.Elaborate experimentation and comparative results authenticate the effectiveness of each developed method towards enhancing the fault classification performance of an SAE.Figures: 39,Tables: 15,References: 81... |
PDF Full Text Request