Novel Prognostic Methods for System Degradation Using LSTM

Abstract

Accurate prognosis of degradation trend is considered very important for the maintenance of critical industrial equipment and assets. This leads to an increase in service availability and life expectancy. Accurate and timely degradation prognosis enables maintenance managers to efficiently plan maintenance regimes and reduce failure occurrences. Recently, Artificial Intelligence (AI) based prognosis and prediction techniques have been in the limelight and attracted the interest of the research community. One such popular AI-based technique is the Long Short-Term Memory (LSTM), which is very efficient in making predictions using time series and sequential data. This paper proposes a novel prognostic technique based on LSTM to predict the degradation trend. The proposed LSTM technique uses a dynamic training window approach with a fixed look-back window for forecasting future steps. The size of the training window is iteratively increased for each prediction as more data is available. This enables the model to utilize the complete sequence trends while making future degradation state predictions. To mitigate over-fitting during model training, the dropout technique and L2 regularization are also incorporated into the proposed generic LSTM model. The performance of the proposed LSTM-based technique is evaluated using experimental results on a real-world application and data. As a case study, the degradation trend of Aerial Bundled Cables (ABCs) using actual thermal degradation data acquired from in-service cables (ABC) is predicted. Moreover, the proposed LSTM-based technique is further compared with a particle filter-based statistical prognosis technique. Promising results validate the efficacy of the proposed LSTM-based approach for degradation prognosis.