Fault prediction from horizontal national water pump sound in Diyarbakir drinking water network using machine learning methods

Machines generate noise as a result of the linear or rotary motion of their mechanical components. During operation, the sound changes as a result of axial misalignment, over-stressing and wear of the components that make up the machine. In addition to using sequences or images obtained directly from sounds to train classification models, multiple sequences are created for the models by extracting features from time and frequency values. In this study, 15 1D (one-dimensional) sequences of MFCCs (Mel-Frequency Cepstral Coefficients) and 15 2D (two-dimensional) images were used for each voice to create data to be used in deep learning models such as CNN (Convolutional Neural Network). In machine learning models, a dataset was created by extracting 18 features from Amplitude-time, Mel-spectrogram, MFCCs, ZCRs (Zero Crossing Ratios) and RMS (Root Mean Square Energy) features from each sound. SVM (Support Vector Machine), KNN (K-Nearest Neighbours) and Ensemble Learning models were used in the machine learning models. Ensemble learning model combines SVM, KNN and RF (Random Forest) models. The highest accuracy for the pump fan is 98.83% in the ensemble model and the lowest accuracy is 63.38% in the SVM model, and the highest accuracy for the pump front bearing is 99.68% in the ensemble model and the lowest accuracy is 84.08% in the SVM model. In addition, the highest accuracy for the pump rear bearing is 99.66% in the ensemble model and the lowest accuracy is 89.53% in the SVM model, and the highest accuracy for the pump motor fan and bearing is 98.19% in the KNN model and the lowest accuracy is 75.60% in the SVM model. The highest accuracy for the complete system is calculated as 93.57% in the ensemble model and the lowest accuracy was calculated as 65.11% in the SVM model. These results show that early fault diagnosis of a centrifugal pump can be achieved and faults can be prevented. The implementation of early fault diagnosis and predictive maintenance planning to prevent failures caused by broken or malfunctioning machines during operation is an important economic and energy saving.