A deep learning model for prognostics and system health management

Through health condition monitoring, Prognostics and system Health Management (PHM) techniques facilitate the prediction of the occurrence of current and future failures and the estimation of the future state of engineering systems. These predictions enable machine operators to objectively manage and take timely maintenance decisions. PHM techniques analyze data generated from machines by extracting insights that facilitate the prognostics and diagnostics process in engineering systems. However, due to the nature of these engineering systems, it is extremely difficult to generate and acquire run-to-failure data from real engineering systems because of critical risks that might emerge during the data collection process. Recurrent Neural Networks and the variants have been used extensively in PHM. Nevertheless, because of the impediments known in such models, non-convergence during training occurs frequently. As a result, models were built shallow notwithstanding the knowledge that the depth of a model significantly contributes to the models' optimization and accuracy of results. It is upon this that this study proposes a deep residual sequence-to-sequence model with LSTM neurons intending to overcome the drawbacks faced by deep models. The model was trained and validated by PHMs' publicly available vibration signal data. The data was preprocessed such that it is compatible with the proposed model. Time to Start Prediction was detected objectively by setting a threshold on Pearson's Correlation Coefficients then the remaining useful life was estimated. Lastly, the model was evaluated with Cumulative Relative Accuracy and found to be 17.87% more accurate than the other models in comparison