Research On Deep-Learning Models For Online Condition Monitoring And Diagnostics In Industrial Systems

Equipment and instrumentation in industrial processes such as manufacturing units,chemical processing plants,power generation stations etc.,are vulnerable to failures.Mechanical degradation by prolonged operations under harsh-environment,dynamic load changes,and inadvertent changes during maintenance operations are a few eminent causes for such failures.Online condition monitoring helps reduce risk to catastrophic failures and ensures safe,reliable and economical operations.It further supports condition based maintenance decision that helps reduce the cost incurred by unnecessary scheduled maintenance approach.'Defense-in-depth' safety concept requires mission critical systems to be redundant and diverse in implementation to avoid common-mode-failures.Early fault identification through online monitoring can be seen as a preventive approach that has potential to reduce the cost associated with redundant implementation of mission critical systems under defense-in-depth safety concept.Data-driven modeling of physical process states and instrumentation-conditions is a prevalent approach to generate an online monitoring system.The real time performance for fault detection and ability to learn system model from history data are the tempting features of such data-driven models.However,development of a reliable model that is robust to the changes in instrumentation performance over its operational time is a challenging task.The increasingly implementation of IOT(Internet of Things)enabled systems has resulted in big-data related to instrumentation and equipment performance.So effective learning strategies are needed that can exploit such big-data for data-driven modeling.Recently,deep learning strategies has resulted in state of the art performance improvements on machine learning tasks such as object recognition,face recognition,speech processing,machine translation etc.Several deep learning architectures have been developed in context of afore-mentioned application domains.Here,in this thesis we explored deep learning architectures to model process and instrumentation conditions for fault detection,identification and diagnosis in an online setting.Sensors are one of the critical infrastructures in any industrial process and may severely affect plant economy.In particular,the focus is to devise a sensor validation model that is robust to large perturbations and significantly improves auto-sensitivity,cross-sensitivity and fault detect-ability performance metrics.Regularization properties of a deep auto-encoder structure are explored to improve afore-mentioned performance metrics.We proposed a Denoising-based Auto-associative Sensor Model(DAASM)for online sensor validation and demonstrated its superiority in simulated as well as real sensor-fault scenarios recorded from a Nuclear Power plant.The performance and uncertainty is compared with relevant models in literature.Similarly,mechanical failures are eminent in rotating machinery e.g.pumps,turbines,motor drives etc.Vibration based fault identification is a common practice for rotary type machinery.However,vibration based fault identification and classification models are sensitive to the features extracted from raw vibration data.We evaluated the performance of traditional vibration features for fault identification and proposed a hybrid deep-model based on Convolution Neural Network(CNN)and Stacked Denoising Auto-Encoder(SDAE)to extract abstract level features from vibration data.The performance is demonstrated on benchmark data collected from an experimental test-rig.Results showed that CNN-based features significantly improved classifier performance and out-performed the traditional features in many cases.