PROGNOSTICS OF HIGH-SPEED CYLINDRICAL ROLLER BEARING

Abstract

Rolling Bearings are considered one of critical mechanical components of almost all rotating machines. As rolling element bearings are mainly responsible for failure of the rotating machinery, hence their condition monitoring becomes a crucial task for safe, reliable, and continuous operation. The CBM (condition-based monitoring) is the decision-making tool that utilizes condition-monitoring information to provide optimal maintenance plans. Thus becoming popular among both industrialists and scholars in recent years as it maximizes the availability, reliability, safety, and productivity of engineering systems. In CBM, prognostics and health management (PHM) involves practices of evaluating health conditions under tasks such as diagnostics, prognostics, and decision support systems. The PHM applications mainly determine whether the system has deviated from its normal operating mode or will fail after running a certain operational life. Thus its implementation enables us to determine whether a particular engineering system continues to perform its intended function or, if not, then how much more time it will remain operational. The steps that are generally adopted in the CBM process for bearing are 1) data acquisition through proper instrumentation of the component under consideration, 2) features extraction and selection, 3) bearing health state assessment or fault detection, and 4) decision making for diagnosis and prognosis. Different approaches are employed to detect, isolate and identify the faults in the machinery using statistical approaches, AI-based (artificial intelligence) approaches, and model-based approaches. The data-driven-based prognostic program can be broadly divided into three categories, namely data acquisition, construction of health indicator (HI), health states division, and estimation of RUL (remaining useful life). The condition monitoring data involves the vibration signals, acoustic emission signals, current, and temperature which are acquired through different sensors. In this work, vibration acceleration responses are acquired to estimate the remnant life of the bearing operating under different working conditions. For accurate RUL estimation, deep learning models are utilized, and their performance under varying operating conditions has been extensively investigated. Different deep learning models, such as CNN (convolutional neural network), LSTM (long short-term memory), BiLSTM (bi-directional LSTM), stacked LSTM (SLSTM), and stacked BiLSTM (SBiLSTM), have been utilized for bearing condition monitoring. To further improve the RUL estimation, a comprehensive feature selection method is adopted in this work. Features from different domains namely time, frequency, and time-frequency are extracted to acquire the bearing degradation information. In this work, a comprehensive feature selection strategy is implemented to discard irrelevant and redundant features. Different feature ranking metrics, such as monotonicity, trendability, prognosability, and robustness, are utilized. The selected feature set is then fused together using dimensionality reduction techniques such as PCA (principal component analysis), LDA (linear discriminant analysis), and ISOMAP (isometric mapping). In addition to this, multi-dimensional features are fused using LSTM, SLSTM, BiLSTM, and SBiLSTM models to obtain the degradation trend of bearing. Furthermore, to selectively capture the bearing degradation influential information, an attention mechanism (AM) is implemented. this integration of AM further enhances the prediction accuracy of the RUL model. In bearing health monitoring tasks, problems of cross-domain model performance generally evolve. To effectively mitigate this issue, both domain adaptation and data augmentation techniques are utilized in this work. For this, maximum mean discrepancy (MMD) and domain adaptation layer are implemented. Data augmentation is carried out using generating models such as generative adversarial networks (GAN), variational autoencoder (VAE), and VAEGAN, which utilizes the merits of both GAN and VAE. In this work, input data to deep learning models are fed in both formats of one-dimensional vibration signal (1D) and two-dimensional image form (2D). Therefore, an extensive investigation of the AI techniques has been presented in this work for bearing prognostics tasks. The accurate RUL predictions rely heavily on the quality of training data which contains the bearing degradation information over time. Therefore, in this work, a comprehensive feature selection strategy is adopted to select prognostic-sensitive feature set. Initially, the vibration signals are processed for unwanted fluctuations and noise removal using smoothing filters and wavelet denoising methods. Subsequent to this, features from different domains are extracted to ensure retrieval of maximum degradation information over time. For this purpose, the time, frequency, and time-frequency domain features are utilized to track bearing degradation information. Also, different features represent bearing degradation with different signatures and scales. To ensure the equal contribution of extracted features for RUL prediction, feature transformation need to be performed. A weighted sum of feature ranking metrics with their corresponding weight factors is constructed. Then fitness analysis is carried out for each extracted feature. The optimized set of parameters is obtained by implementing the PSO technique. The features with fitness function value greater than a predefined threshold are retained for health index construction. The feature set thus obtained is utilized as input to the deep learning model to learn the underlying relation of bearing degradation information over time.The accuracy of deep neural networks suffers significantly when tested on different domain data. The presence of environmental noise, different failure modes, and various speeds and loads lead to domain shift problems. To address these issues, this work investigated the cross-domain performance of transfer learning-based bi-directional Long Short-Term Memory (TBiLSTM) network. the domain adaptation is realized using multi-kernel maximum mean discrepancy (MK-MMD). The discrepancy differences between the two domains are computed by minimizing the maximum mean distribution discrepancy of feature values. The proposed methodology is validated using both experimental and IEEE PHM datasets. For both datasets, the proposed method achieves low values of RMSE and MAE, indicating excellent generalized performance for different operating conditions. In the context of scarce faulty data conditions of bearing, data generation methods are generally implemented to mitigate the imbalanced data situation. Data augmentation is implemented using variational auto-encoder generative adversarial network (VAEGAN). For this purpose, phase space trajectories from vibration signals are obtained for conditions of chaos (strong and weak attractor), and quasi-periodic. The characterization operation is implemented using residual deep convolutional neural network (RDCNN). Thus the VAEGAN-RDCNN model is proposed to effectively solve the problems of scarce faulty data conditions. To generate samples of high-quality, latent representations from trained VAE are utilized as enhanced input information vectors to the GAN. Various metrics such as peak signal-t-noise ratio (PSNR), structural similarity index measure (SSIM), Kullback-Leibler divergence (KLD), and histogram analysis (HA) are utilized to assess the quality of images generated through VAEGAN. The comparative performance assessment of VAEGAN-RDCNN is presented using VAE, and WGAN models based on metrics, namely average accuracy, precision, recall, and F1-score.