Attention-Boosted Path Exploration Neural Architecture Search for Fault Diagnosis of Rotating Machinery

With the ongoing advancement of artificial intelligence, data-driven fault diagnosis methods have become widely adopted. In practical industrial applications, environmental noise is, however, inevitable, posing significant challenges in diagnosing faults based on subtle vibration signals. The manual process of constructing models is time-consuming and labor-intensive, hindering the efficient development of optimal diagnostic models. This article proposes an attention-boosted path exploration neural architecture search (ABPE-NAS) method designed explicitly for fault diagnosis in rotating machinery to address these challenges. The ABPE-NAS method seamlessly integrates the depth and width of network layers into the model's architecture while incorporating an attention mechanism within the search space to dynamically adjust the weights of different network components, thereby enhancing overall network performance. Additionally, a sequential structure matrix is designed to represent the model architecture, which is iteratively generated through path tracking. By leveraging a recurrent reinforcement learning network to generate this matrix, experimental results demonstrate that the ABPE-NAS method produces an optimal diagnostic model with exceptional performance across various noise levels, validating the effectiveness of the attention-boosted path exploration approach.