Bayesian approach and time series dimensionality reduction to LSTM-based model-building for fault diagnosis of a reciprocating compressor

Reciprocating compression machinery is the primary source of compressed air in the industry. Undiagnosed faults in the machinery's components produce a high rate of unplanned stoppage of production processes that can even result in catastrophic consequences. Fault diagnosis in reciprocating compressors requires complex and time-consuming feature-extraction processes because typical fault diagnosers cannot deal directly with raw signals. In this paper, we streamline the deep learning and optimization algorithms for effective fault diagnosis on these machines. The proposed approach iteratively trains a group of long short-term memory (LSTM) models from a time-series representation of the vibration signals collected from a compressor. The hyperparameter search is guided by a Bayesian approach bounding the search space in each iteration. Our approach is applied to diagnose failures in intake/discharge valves on double-stage machinery. The fault-recognition accuracy of the best model reaches 93% after statistical selection between a group of candidate models. Additionally, a comparison with classical approaches, state-of-the-art deep learning-based fault-diagnosis approaches, and the LSTM-based model shows a remarkable improvement in performance by using the proposed approach.