Modelagem computacional dirigida por dados para diagnosticar falhas de rolamentos e cavitação em bombas centrífugas

Abstract

Computational modeling based on machine learning has stood out in engineering applications, specifically in monitoring and classifying the degradation of critical components. In this research, machine learning models were developed to monitor the occurrence of cavitation in centrifugal pumps and bearing defects through time signals recorded by sensors. The experimental data originate from the Case Western Reserve University (CWRU) and the UWA System Health Lab Prognostics Data Library (SPDL), involving types and intensities of faults in ball bearings and the presence of the cavitation phenomenon detected via accelerometers. The focus of this research is on constructing a computational solution for data preparation and feature extraction, using the hit counting technique and the Short-Time Fourier Transform (STFT). The first technique involves calculating several parameters that define the waveform in the time domain, such as hit duration, time to reach peak amplitude, and hit interval energy. The second transforms the time-domain signal into a time-frequency spectrogram and subsequently into images. Two machine learning models based on neural networks were used for pattern identification and classification: a Multilayer Perceptron Neural Network (MLP), which employs parameters from the hit counting technique, and a Convolutional Neural Network (CNN), which uses images generated by the STFT. The results obtained with the MLP model using vibration experimental data from the Case Western Reserve University showed an accuracy of 69.50%, precision of 67.07%, recall of 73.50%, and an F1-score of 68.63%. The CNN model, on the other hand, achieved an accuracy of 93.90%, precision of 93.41%, recall of 93.77%, and an F1-score of 92.98%. For the UWA System Health Lab Prognostics Data Library, the results obtained with the MLP model showed an accuracy of 77.90%, precision of 76.25%, recall of 75.68%, and an F1-score of 76.25%. The CNN model achieved an accuracy of 89.17%, precision of 87.57%, recall of 88.05These results were analyzed using K-fold cross-validation, confusion matrices, and performance comparison with machine learning models from the reference study. Additionally, a statistical analysis was conducted to define parameters such as signal thresholding and windowing, as well as to create labels.