Blockage detection in centrifugal pump using semi-supervised machine learning based on SVM and LSTM

Blockages in centrifugal pumps affect their flow rate and performance and can even interrupt their continuous operation. Health monitoring of the pump helps to avoid unwanted stoppages that can further lead to the failure of the whole system. Various supervised machine learning models have been developed in the past to monitor pumps for these faults. These models perform well on large amounts of labelled data, but a shortage of labelled data is a common problem in industrial applications. However, unlabelled data acquired from real-time operation of the pump are easily available but not utilised for training these models. Therefore, this study presents a semi-supervised methodology to detect blockages in pumps along with their severity. First, the support vector machine (SVM) model and the state-of-the-art long short-term memory (LSTM) model are individually trained with only labelled data using the statistical features acquired from the discharge pressure signal. The hyperparameters of both these models are optimised using the grid search optimisation method. Next, pseudo labels are generated for the unlabelled data through a trained SVM model. Pseudo labels defined by the SVM with a confidence greater than 0.9 are further selected to be combined with labelled data to train the LSTM model. The results show that the proposed approach effectively identifies blockage faults with a validation accuracy, test accuracy and F1 score of 97.51 %, 97.39% and 97.9%, respectively.