Improvements in Stacking Deep Learning Models for Current and Vibration Signature Analysis in Rotating Machines

Given the difficulty of obtaining representative and quality industrial datasets, such as electrical currents, we propose a multi-model stacking approach. Our method combines the strengths of three distinct neural network architectures in a multi-domain approach, tuned to process data from the frequency, time, and time-frequency domains. We evaluate the proposed architecture on three datasets (CWRU, PU, and PECCE). The stacking model achieved the highest test accuracy of 99% on the PECCE dataset, which represents an 8% improvement over using just one architecture. On the noisy CWRU dataset, we achieved accuracies above 98% for noise levels up to 10 dB. On the PU dataset, we observe that the stacking generalizes effectively to different operating conditions. Therefore, the multi-domain stacking approach has proven itself, confirming its versatility and potential for applicability in diverse industrial scenarios.