Performance evaluation of LSTM and Bi-LSTM using non-convolutional features for blockage detection in centrifugal pump

Blockages in the suction or discharge side of the pump adversely affect the pump’s performance by reducing the flow rate and head, increasing vibration, noise, and overheating. Thus, diagnosing pump blockages accurately and quickly is vital. In this study, a test rig is developed to detect blockages in the pump (suction, discharge, and simultaneous) at three levels of severity. Initially, a non-convolutional feature layer is developed to extract sets of statistical features (SF), entropy-based parameters, and Holder’s exponent (HE). Next, two sequential learners, LSTM and Bi-LSTM, are trained with statistical parameters and tuned using grid-search optimization. Model with least validation loss is chosen as the base model. Subsequently, the process is repeated with only non-linear features, combination of SF with HE, and combination of all three sets of features. Thus, eight different models are trained (four each for LSTM and Bi-LSTM). To eradicate the impact of random variations, these models are trained ten times and the average values of performance metrics (precision, recall, and F1score) are noted. It is found that LSTM trained with all sets of features is most accurate at 99.01% with F1 score of 98.95%. Although LSTM is superior, it is marginally ahead of Bi-LSTM, but LSTM consumes almost 40% lesser time for training. Toward the end, the ablation study of LSTM is performed to analyze the effects of the most influencing parameters. Thus, the present study emphasizes an automated and robust method for maintenance engineers to diagnose the blockages in the pump with severities of blockages.