Resumen
Unsupervised anomalous sound detection, especially self-supervised methods, plays a crucial role in differentiating unknown abnormal sounds of machines from normal sounds. Self-supervised learning can be divided into two main categories: Generative and Contrastive methods. While Generative methods mainly focus on reconstructing data, Contrastive learning methods refine data representations by leveraging the contrast between each sample and its augmented version. However, existing Contrastive learning methods for anomalous sound detection often have two main problems. The first one is that they mostly rely on simple augmentation techniques, such as time or frequency masking, which may introduce biases due to the limited diversity of real-world sounds and noises encountered in practical scenarios (e.g., factory noises combined with machine sounds). The second issue is dimension collapsing, which leads to learning a feature space with limited representation. To address the first shortcoming, we suggest a diffusion-based data augmentation method that employs ChatGPT and AudioLDM. Also, to address the second concern, we put forward a two-stage self-supervised model. In the first stage, we introduce a novel approach that combines Contrastive learning and masked autoencoders to pre-train on the MIMII and ToyADMOS2 datasets. This combination allows our model to capture both global and local features, leading to a more comprehensive representation of the data. In the second stage, we refine the audio representations for each machine ID by employing supervised Contrastive learning to fine-tune the pre-trained model. This process enhances the relationship between audio features originating from the same machine ID. Experiments show that our method outperforms most of the state-of-the-art self-supervised learning methods. Our suggested model achieves an average AUC and pAUC of 94.39% and 87.93% on the DCASE 2020 Challenge Task2 dataset, respectively.