Study on Vibration Based Rotor System Fault Detection and Diagnosis Using Deep Learning Approaches

Abstract

This research demonstrates the implementation of AI-based techniques for rotor machinery's fault detection and diagnosis through a comprehensive experimental and theoretical study applying various Machine Learning (ML), Deep Learning, and Domain Adaptation techniques. The work addresses the significant research gaps, including the domain adaptation task, which is the immediate Industrial requirement. Rotating machines play a vital role in various industries, and the failure of these systems poses a significant threat to asset, production, environment, and human life. The rotor may fail in different ways due to various types of faults; however, in most cases, the fault arises due to common reasons such as manufacturing errors or severe operational conditions. Most failures are due to the severe operational condition for getting higher yield, which is implemented to cope with ever-increasing global market growth and competitive environment. In general, catastrophic failures are observed due to the initiation of a sequence of faults. Initially, the presence of rotor mass imbalance induces other critical faults that eventually lead to catastrophic failures. As an important measure, industries spend millions of funds on the maintenance of machines. In particular, fault prediction in the rotor system has always been a paramount activity in the industry. Out of different maintenance strategies, predictive maintenance gains importance as it enhances the equipment life and reliability of the rotor system. It includes various techniques of condition monitoring for rotor systems, out of which the vibration-based methods are quite popular, accurate, and relatively inexpensive. It is evident that technological enhancement in the last two decades has helped engineers to apply automation in maintenance, and Artificial intelligence based strategies have gained a lot of attention. In industries, using Machine Learning and Deep Learning in predictive maintenance for rotor systems has gained significant momentum in preventing untimely failures. However, there are challenges in digitization of plant maintenance, and ample efforts are underway globally to achieve a robust methodology for continuous health monitoring of rotor systems. Hence, the literature review has been conducted based on Preferred Reporting Items for Systematic Review and Meta-Analyses (PRISMA) protocol guidelines, and research gaps were identified. Research gaps such as automatic feature extraction, overfitting, hyperparameters tuning, domain shift, biased prediction due to imbalance class size etc. are important and systematically analyzed in the present study. For this purpose, different techniques such as SMOTE, SMOTEBoost, ensembling, 1D-CNN, Genetic Algorithm (GA) based optimization, effective dropout layer positioning, Adversarial Discriminative Domain Adaptation (ADDA) etc. are implemented. An additional case of investigation of noise-based signals using the variational mode decomposition (VMD) approach is also studied to check the performance of the proposed methodology. An experimental rig was designed and developed in the laboratory to generate faulty signals of a rotor system. Primarily, two types of faults, rotor mass imbalance and crack of different severity levels are considered for the study. Vibration signals are collected using accelerometers and NI data acquisition system. Various case studies have been conducted to benchmark the performance of the proposed methodology against the traditional ML, DL, and transfer learning approaches. Rotor mass imbalance signal is utilized to demonstrate the performance of the proposed methodology and finally tested for the rotor crack fault prediction on the rotor rig. The results demonstrate that rotor mass imbalance fault prediction using machine learning with balanced class data achieved an accuracy of 95.63%. However, in the presence of noise, the accuracy decreased to 69.45%. Furthermore, the application of VMD to mitigate the effects of noise enhanced the prediction accuracy to 88.89%. However, the available data's class imbalance, which leads to overfitting, has decreased the prediction capabilities to 75.33 percent. After addressing the class imbalance with SMOTEBoost and noise with VMD, the prediction accuracy increased to 81.33%. The overfitting issue is further mitigated by the ensemble technique, and the maximal prediction accuracy of VMD-based Ensembled SMOTEBoost is as high as 86%.

The DL-based method, which uses 1D-CNN, achieved an 84.16% prediction performance with the help of 7 convolutional layers and hyperparameters manually tuned. The hyperparameter optimization problem of CNNs has been resolved with the proposed GA-optimized 1D-CNN. The strategic placement of the dropout layer in 1D-CNN's architecture addresses the inherent overfitting problems of CNNs, while GA optimizes hyperparameters, resulting in enhanced prediction accuracy of up to 97.47% with three convolutional layers. The results of the tests show that while GA optimization decreases the depth of CNN architecture, 1D-CNN based deep learning does away with the need for human intervention for feature extraction. Efficient dropout placement minimizes computational burden and duration by decreasing the learnable parameters of the CNN (network weights) while maintaining optimal prediction accuracy.

Domain adaptation test results show that ML couldn't achieve a better prediction accuracy of more than 41%, while DL could reach 52.3% with 3 convolutional layers and manually tuned hyperparameters. Hence, the domain shift issue has been addressed through ADDA's domain invariant feature learning algorithm. The efficacy of ADDA has been explored by introducing 1D-CNN as a source and a target encoder inside ADDA's architecture to take advantage of CNN's feature extraction capability. Further, the data class imbalance issue has been addressed through SMOTE. The domain adaptation approach of ADDA without SMOTE and the manually tuned hyperparameter could achieve a prediction performance of 60.1%. Hence, to improve the prediction capabilities of ADDA, the dropout layer, and GA-optimized hyperparameters have been used with 1D CNN to tackle CNN's overfitting and hyperparameter issues within ADDA's architecture. Thus, the proposed methodology, which uses SMOTE and a genetically tuned hyperparameter, could increase the domain adaptation prediction performance to a maximum of 86.87%. Thus, the proposed methodology successfully addresses the deep learning based domain adaptation challenges, and it can be scalable in the industrial environment for online continuous monitoring. Further, the present research boosts the possibility of Industry 4.0 implementation at a faster speed.