Weakly-Supervised Diagnosis with Attention Models

Sammanfattning: With the rapid development of data-driven solutions, a high-quality dataset with fine labels can well tackle fault diagnosis problems. However, in industry assets, the quality of the dataset depends much on the experience and capability of the annotation engineer. Therefore, strong labels are usually hard and expensive to acquire. The attention model is currently a novel method to support weak labels on strong tasks. This thesis project investigates the application of attention models on the transition from weak tasks to strong tasks. We explore the attention mechanism on estimating the location of the faults, by employing an attention layer and generating the distribution of the attention weight. After verifying the effectiveness of attention models for weak labels on strong tasks, we apply the attention models on a Sound Event Detection (SED) task. To be more specific, more attempts on different types of attention models and parameters of each type are made to improve the results. The result indicates that attention models are feasible and promising on weakly-supervised tasks, to improve the diagnostic performance.