RESEARCH
>Anomaly Detection
Anomaly Detection
Anomalous sound detection enhances maintenance efficiency in industries by early detecting equipment malfunctions, ensuring swift responses to security incidents, aiding in accident prevention in transportation, and contributing to noise management and emergency response in urban environments.
In the conventional anomalous sound detection method, a model is trained using a limited number of normal data. Accordingly, the mingling of normal and anomalies unseen during the training can be induced at the test stage, which calls for the need to train with more versatile data. To overcome these limitations, we synthesize samples near the normal data distribution with our proposed Noisy-ArcMix. Noisy-ArcMix significantly improves the compactness of intra-class distribution through the training with virtually synthesized samples near the normal data distribution. More importantly, we observed that the mingling effect between normal and anomalous samples can be reduced further by Noisy-ArcMix, which gains generalization ability through the use of inconsistent angular margins for the corrupted label prediction
In addition to Noisy-ArcMix, we introduce a new input feature, a temporally attended log-Mel spectrogram (TAgram), derived from a temporal attention block. TAgram includes the temporal attention weights broadcasted to spectrogram features, which helps a model to focus on the important temporal regions for capturing crucial features.
Proposed model (TASTgram-Noisy-ArcMix)
Performace comparison with the state-of-the-art models