Damage Presence, Localization, and Quantification of Aircraft Structure Based on End-to-End Deep Learning Framework

Structural health monitoring (SHM) plays a vital role in promptly identifying structural damage in aircraft, optimizing maintenance, and reducing costs; however, it faces significant challenges in practical applications, in that it mainly needs to process a large number of continuously collected sensor data, which are inevitably contaminated by random noise. This study, therefore, maps the relationship between Lamb wave signal data with noise and the health condition of aircraft structures using an end-to-end approach to construct a deep learning (DL) framework. The framework integrates deep residual convolutional networks (DRSNs) for feature extraction, efficient channel attention (ECA) for feature enhancement, and long short-term memory (LSTM) in analyzing time series data. Lamb wave signal datasets considering different damage locations and severity are obtained by lead zirconate titanate (PZT) sensors on the aircraft structure, and the datasets are destroyed by using multiple levels of Gaussian random noise to approximate the noise disturbances and unavoidable unpredictability of the industrial environment. The experimental data confirms that the performance metrics of the proposed framework for damage presence, localization, and quantification tasks are all above 97% in a noise-free environment. When dealing with high-noise scenes, the framework provides stronger antinoise robustness and higher accuracy compared to existing state-of-the-art (SOTA) methods. Quantitative analysis of ablation experiments and visualization of the t-distributed stochastic neighbor embedding (t-SNE) algorithm is applied to reveal the contribution of each component of the designed framework toward feature extraction and damage identification of Lamb wave signals in noisy environments.