This paper examines differences in remaining useful life (RUL) prognosis models for single-lap joint (SLJ) bonded composite structures under various bending conditions. By analysing crack growth behaviours under different bending stress modes, the study explores the relationship between crack propagation patterns and the fatigue life of SLJ structures, offering valuable reference data for RUL prediction. Two commonly used RUL prediction models are presented: the traditional RUL model (neural networks) and the long short-term memory (LSTM) network model. These models are used to forecast the remaining useful life of SLJ structures under different bending conditions.

The traditional RUL model estimates remaining life by directly analysing crack growth curves and applying fatigue damage theory, making it suitable for quick evaluation and real-time monitoring. The LSTM model, based on deep learning, captures long-term dependencies in time series data and can learn complex nonlinear crack propagation trends, delivering improved prediction accuracy under variable bending conditions.

This study further emphasises how bending conditions affect the performance of RUL models, particularly with respect to crack growth behaviour and model generalisation. While the LSTM model shows strong adaptability and accuracy when sufficient data is available, its computational complexity and training requirements may limit its real-time use. Conversely, although less adaptable, the traditional RUL model remains practical for engineering applications due to its simplicity and efficiency.

Crack detection on pressed panels is essential for maintaining quality in automotive manufacturing. However, achieving stable inspection is challenging due to strong reflections on metallic surfaces, geometric curvature, and varying illumination conditions. To address these issues, this study proposes a hybrid approach that combines shape-based analysis with unsupervised deep learning. Shape-based analysis leverages the panel’s edge geometry and therefore remains relatively robust to illumination changes, while unsupervised models learn the normal appearance of panels and perform anomaly detection without requiring defect labels. Despite these advantages, both methods commonly suffer from over-detection, where structural edges or reflection patterns are misinterpreted as cracks. To mitigate this problem, we introduce a baseline-panel–driven correction map that utilizes prior information extracted from healthy panels. By capturing recurring geometric features and reflection patterns, this correction step effectively suppresses systematic false positives that repeatedly occur in the same regions. When integrated with shape analysis and unsupervised learning, the proposed correction mechanism enhances the stability and accuracy of crack detection, enabling more reliable inspection under the challenging lighting conditions of real automotive press lines.

Structural Health Monitoring (SHM) aims to monitor in real-time the health state of engineering structures. For thin structures, Lamb Waves (LW) are particularly effective for SHM applications. A bonded piezoelectric transducer (PZT) generates LW in the form of a short tone burst, creating an initial wave packet (IWP) that propagates through the structure while interacting with boundaries, defects, and other features. These interactions leave precise yet complex signatures in the signals recorded by the sensors. In practice, however, extracting the specific signature associated with damage is often challenging for traditional SHM signal-processing methods, making reliable damage detection difficult.

To address these limitations, we employ a Deep Learning–based approach known as the Rank Reduction Autoencoder (RRAE). The RRAE is an autoencoder whose latent space is constrained to follow a low-rank SVD structure, ensuring that it captures only the most salient features of the measured signals. In this work, we extend the original concept by enforcing the SVD modes to be predictive of both damage location and severity. This is achieved through an additional neural network that takes the reduced latent representation produced by the RRAE and directly estimates the damage characteristics. The joint training of the RRAE and the feature-extraction MLP therefore yields, at convergence, a robust and powerful tool for damage detection.

The proposed technique is demonstrated on two case studies. The first, a more academic example, involves damage detection on a thin plate. The second is based on an experimental campaign conducted by CETIM and focuses on a 2-meter pipe segment, where the proposed architecture performs damage detection using measurements of acoustic wave emissions.

Guided-wave structural health monitoring is highly sensitive to small defects, but its reliability can degrade severely under coupled time-varying service conditions such as temperature and load changes. To address this challenge, this work proposes an unsupervised adaptive feature enhancement network (UAFEN) for reliable damage alarm without using labeled damage data. The method is built on a one-dimensional convolutional autoencoder trained only with healthy guided-wave signals. A modified multi-kernel maximum mean discrepancy term is introduced to align feature distributions across mixed temperature-load conditions, while a sequence attention module is incorporated to enhance sensitivity to localized damage-related anomalies. During inference, incoming signals are reconstructed by the trained model, and reconstruction errors are used for anomaly judgment and region-level alarm generation. The method is validated on a real metallic wing-box skin instrumented with an integrated piezoelectric sensor network under coupled temperature-load variations. The dataset contains healthy baseline data as well as crack and pit damage data collected under multiple environmental conditions. To improve regional interpretability and monitoring efficiency, the monitored area is partitioned into transverse and longitudinal regions, and their alarm results are fused to obtain the final damage-alarm map. Experimental results show that the proposed method maintains a low false-alarm level under healthy time-varying conditions, with alarm rates below 4.0% in transverse regions and below 2.0% in longitudinal regions. For actual damage regions, the alarm rates reach 74.6% and 70.5% for a 2 mm crack, 92.5% and 95.2% for a 5 mm crack, and 94.6% and 100% for pit damage in transverse and longitudinal regions, respectively. These results demonstrate that the proposed UAFEN can effectively suppress environmental interference while preserving high sensitivity to local structural damage, providing a practical damage-alarm solution for guided-wave monitoring of aircraft structures under complex service conditions.