ANOMALOUS MACHINE SOUND DETECTION USING TIME DOMAIN GAMMATONE SPECTROGRAM FEATURE AND BASED ON IDNN AND UNET MODEL

• Main Author: Adyatma Hafiz, Primanda
• Format: Final Project
• Language: Indonesia
• Subjects: Teknik (Rekayasa, enjinering dan kegiatan berkaitan)
• Online Access: https://digilib.itb.ac.id/gdl/view/82480
• Institution: Institut Teknologi Bandung

Anomalous machine sound detection system aims to identify the condition of machinery based on its sound. During the development of anomalous machine sound detection system, unsupervised learning methods have advantages over supervised learning methods because they do not require anomalous sound data for model training, which is harder to obtain compared to normal sound data. However, the use of certain unsupervised models such as autoencoder models has not yet been able to provide good accuracy performance in the system due to the model's limitations in handling certain characteristics of machine sounds, especially non-stationary sounds. On the other hand, in terms of features, the use of log Mel spectrogram features is the most popular method, but the classification accuracy achieved is not yet optimal because these features are extracted in the frequency domain using STFT, which has several weaknesses. Therefore, in this final project, anomalous machine sound detection system is proposed using time domain Gammatone spectrogram features extracted in the time domain to improve feature precision. The extracted features are modeled using the IDNN model architecture with 2 variations, namely the AE-IDNN and UNet-IDNN models. In this case, the IDNN model is used because it has the potential to improve accuracy performance, especially for machines with non-stationary sound characteristics, and its combination with the UNet model has the potential to produce a complementary effect because both have their respective advantages. The evaluation results show the advantage of the time domain Gammatone spectrogram features over the log Mel spectrogram features, with a difference in AUC score of 5.9% in the UNet-IDNN model. Additionally, overall, the IDNN model successfully improved the performance of the autoencoder model by 9% and the UNet model by 6.6%. In the IDNN model, the UNet-IDNN model provided better performance compared to the AE-IDNN model by 0.2%. Hyperparameter tuning of the model resulted in an overall AUC score of 92.5%.