Multiclass classification of faulty industrial machinery using sound samples

The intelligent diagnosis of faults in industrial assets helps avoid unexpected interruptions of critical services. Thus, machine learning systems for monitoring machinery have a fundamental role in smart factories. The goal of this paper is to promote generalization by combining binary models into multiclass ones. This approach avoids using customized models, which do not perform well if the required conditions do not match during inference. Hence, instead of proposing multiple models for each particular condition, e.g., a model for each machine type, we evaluate different multiclass models where this information becomes an additional outcome along with the operating status. In our experiments, we use the Synthetic Minority Over-sampling Technique (SMOTE) to deal with imbalanced datasets and the PyCaret library to pre-select the most promising algorithms using binary classification. After selecting the binary models for combination, we evaluate two multiclass approaches, called partially- and totally-mixed, using the Malfunctioning Industrial Machine Investigation and Inspection (MIMII) dataset of machinery audio. The number of classifiers needed is reduced up to 83% with our proposal compared with binary classification used as a baseline approach. Furthermore, in the partially-mixed multiclass proposal, it is possible to obtain the correct classification of anomalous states in at least 88% of the samples for most devices. We show that there is a cost on computational resources for model training, which is compensated by the generalization gains verified using the F1-Score metric two times higher than the binary models. We follow the same idea of building more generic models with another existing dataset, ToyADMOS, which similarly contains machine operating sounds. The experiments confirm the results obtained with the MIMII dataset.