Development of effective approaches for prognostics of industrial systems

Abstract

Machine availability and reliability are two of the most essential concerns for an industry. Increased availability is required by industries to stay competitive in today’s global market competition. To achieve increased availability, a good maintenance strategy is required that reduces the losses due to unplanned shutdowns and keep the preventive maintenance at minimum. Among all available maintenance strategies; Condition Based Maintenance (CBM) is the most effective strategy to achieve these goals. Effectiveness of Condition Based Maintenance (CBM) strategy depends on accuracy in prediction of Remaining Useful Life (RUL). Prognostic is the technology used to predict the RUL based on monitored parameters. Prognostic approaches can be broadly classified into two categories: physics based prognostic approaches and data driven prognostic approaches. Current research focus is only on data driven prognostic approaches.Presence of noise in the data reduces the accuracy of RUL prediction with data driven prognostic approaches. Presence of unknown initial wear and presence of multiple failure behaviour in data may act as sources of data noise. If these sources of data noise are not handled appropriately, then it may give poor prediction of the RUL. Another major issue found with the data driven prognostics approaches is the larger variation present in the historically observed Condition Monitored (CM) data obtained from the fleet. This leads to poor model performance. Updating the model parameters based on new information for a unit can help in reducing the variation and improve accuracy of prediction. Third issue is handling of multidimensional features (i.e. RMS, kurtosis, skewness, mean, median etc.). Features are generally extracted from the raw data to represent the degradation of the component. Number of features can be extracted from the raw data to represent the degradation of the component. If all these features have been taken as input to the model, then it will over fit the model. Over fitting is the situation where model performs well during training, but shows significantly poor performance during testing.The objective of this thesis is to increase the prediction accuracy of prognostic model while considering the presence of the noise, effect of multidimensional condition monitoring features and continuously updating the model parameters. Different industrial systems have been considered for this study such as aircraft engine, gas turbine, and roller ball bearings. Also, life prediction for a smart material components (SMA springs) is also demonstrated.Remaining Useful Life Prediction of an Aircraft Engine under Unknown Initial Wear is presented. Two Artificial Neural Network (ANN) models were developed. First model is developed by neglecting the effect of the presence of noise in the data (i.e. sample with abnormal initial wear); while the second model is developed after removing the noise associated with the data. Another model is developed by considering another type of noise in the data. Inner race failure, outer race failure and cage failure are the major failure modes associated with ball bearing. Presence of multiple failure modes and their interaction may cause noise in the data set. Clustering and Change Point Detection Algorithm (CPDA) is used for identification of presence of multiple failure behaviour due to multiple failure modes in the data. Combined output of Clustering and CPDA is used for developing RUL prediction model. Separate models for single failure behaviour and multiple failure behaviour are constructed. General Log- Linear Weibull (GLL- Weibull) model is used for the same. Effects of multidimensional features are also considered here.A PCA-ANN based algorithm is used to predict the RUL of ball bearings while considering the effect of the multidimensional features. Instead of giving all features directly to the model, the best three principal component values obtained from PCA are used as input parameters to the model. A risk based maintenance strategy to optimize forecast of a gas turbine failures is also presented. The algorithm does not completely rely on historically observed condition monitored data but also updates the model parameters as and when newinformation is available. Bayesian approach is used to update the model parameters. Second part of thesis focuses on reliability estimation of Shape Memory Alloy (SMA) springs. The reliability of the SMA springs was estimated by using life test data of the springs. The spring has undergone thermo mechanical fatigue and it was observed that recovering to original shape is disappearing with number of cycles due to inelastic deformation. The life prediction model was developed here using GLL- Weibull. Bayesian approach is used to update the parameters of the model. As experiments were performed on accelerated condition; an accelerated life testing model was also developed to extrapolate the Probability Density Function (PDF) at normal use condition.In essence, present thesis contributes towards the development of accurate approaches for prognostic of various components. Thus, the outcome is of high importance in effective planning of Condition Based Maintenance of asset intensive systems and reducing unplanned down time losses to the industries. Also, the novel attempt is made to first time study the life prediction approaches for shape memory allow springs. The results are encouraging and open up further scope prognostics for systems with such components.