Evaluating Hyperparameter Optimization for Machinery Anomalous Sound Detection

Machinery fault detection plays an important role in industrial applications. In this study, we evaluate the hyperparameter optimization for machinery anomalous sound detection (MASD) using two datasets, IDMT-ISA and MIMII Pump. We used two features, mel and reassigned spectrogram, and two autoencoder models to adapt the features. The hyperparameters are optimized using the Optuna toolkit. The results show that the optimal hyperparameters are different for each dataset, partially for acoustic features and loss functions. The top three importance parameters for both datasets are the same, i.e., learning rate, patience, and loss with different orders of importance. The model could predict the large test data (1399 samples) on the IDMT-ISA dataset with the highest AUC score of 0.90 and on the MIMII Pump dataset with the highest AUC score of 0.89.