A Deep Learning Approach for Fault Diagnosis in Centrifugal Pumps through Wavelet Coherent Analysis and S-Transform Scalograms with CNN-KAN

Centrifugal Pumps (CPs) are critical machine components in many industries, and their efficient operation and reliable Fault Diagnosis (FD) are essential for minimizing downtime and maintenance costs. This paper introduces a novel FD method to improve both the accuracy and reliability of detecting potential faults in such pumps. The proposed method combines Wavelet Coherent Analysis (WCA) and Stockwell Transform (S-transform) scalograms with Sobel and non-local means filters, effectively capturing complex fault signatures from vibration signals. Using Convolutional Neural Network (CNN) for feature extraction, the method transforms these scalograms into image inputs, enabling the recognition of patterns that span both time and frequency domains. The CNN extracts essential discriminative features, which are then merged and passed into a Kolmogorov-Arnold Network (KAN) classifier, ensuring precise fault identification. The proposed approach was experimentally validated on diverse datasets collected under varying conditions, demonstrating its robustness and generalizability. Achieving classification accuracy of 100%, 99.86%, and 99.92% across the datasets, this method significantly outperforms traditional fault detection approaches. These results underscore the potential to enhance CP FD, providing an effective solution for predictive maintenance and improving overall system reliability.