Impact monitoring based on domain adversarial transfer learning strategies

Abstract Impact events may cause some damage to aerospace composite structures that are difficult to inspect on the surface, thus threatening the operational safety of the aircraft. Therefore, estimating the impact location and the original impact force is necessary. This paper proposes a deep-learning model for impact monitoring based on feature extraction. The first step employs a Convolutional Neural Network (CNN) to localize the impact region, initially narrowing it to a specific area and then determining a precise location using a weighted center algorithm. In the second part, the Temporal Convolutional Network (TCN) is first utilized for feature extraction, and then the Gated Recurrent Unit (GRU) is used for impact force estimation. During the training of the impact monitoring model, a domain-adversarial transfer learning strategy is employed to extract domain-invariant features between the source and target domains, reducing the data needed for training. This method can monitor impacts on large, complex composite structures using sparse sensor arrays.