A hybrid one-dimensional convolutional neural network and long short-term memory framework for vibration-based damage localisation and severity estimation in reinforced concrete bridges and results support the use of the proposed framework for screening-level assessment of bridge damage under controlled simulated conditions.
Abstract
Structural health monitoring (SHM) is essential for assessing the safety and serviceability of bridge structures. Identifying progressive and concurrent damage remains challenging due to the complex and continuous nature of structural deterioration. This study proposes a hybrid one-dimensional convolutional neural network and long short-term memory (1D-CNN–LSTM) framework for vibration-based damage localisation and severity estimation in reinforced concrete bridges. Operational modal analysis is applied to field-measured vibration data from an in-service bridge. A finite element model is updated using particle swarm optimisation, reducing frequency discrepancies from 7–17% to within ±3%. Progressive single-, double-, and triple-element damage scenarios are simulated through systematic stiffness degradation. The resulting modal frequency data are used to train 1D-CNN–LSTM models using Pareto front optimisation. The proposed framework achieves coefficients of determination above 0.80 with low prediction errors (MSE and MAE < 2) for single- and double-element damage scenarios. The results support the use of the proposed framework for screening-level assessment of bridge damage under controlled simulated conditions.
The structural integrity and operational safety of High-Speed Railway (HSR) bridges are increasingly threatened by low-frequency industrial vibrations. As traditional numerical methods are inadequate for real-time monitoring, this study proposes a novel FE-Augmented intelligent prediction framework to evaluate and forecast the structural response of HSR bridges subjected to ambient industrial excitations. Field measurements conducted on an HSR bridge near a stone-processing plant revealed a severe 1.5 Hz global resonance, with lateral displacements exceeding safety limits by a factor of 4.4 even without train loads. To overcome data scarcity for machine learning, a 3D coupled bridge-soil finite element model was developed to augment the vibration dataset. Subsequently, a hybrid deep learning framework integrating Long Short-Term Memory (LSTM) networks and Random Forest (RF) was established. By aligning the LSTM input time-window with the physical wave propagation delay, the predictive model gains physical interpretability. Furthermore, ablation studies confirmed that integrating spatial waveguide information with hybrid features minimizes prediction errors. Validation demonstrates that by fusing LSTM-extracted deep temporal features with handcrafted statistical indicators, the proposed LSTM-RF framework overcomes the amplitude clipping effect common in standard neural networks. It accurately captures complex spatial interference and beat phenomena under multi-machine operations, achieving a high Coefficient of Determination (R
2
= 0.965) and a low Normalized RMSE of 2.43% under extreme load scenarios. The trained hybrid model accelerates prediction inference to the millisecond scale, representing over six orders of magnitude improvement in computational efficiency compared to conventional 3D FEM. This framework provides an efficient, real-time early-warning methodology for safeguarding HSR infrastructure against complex environmental vibrations.
Ran Bi, Chen Wen, Ming Wang et al.· International Journal of Str...· 0 citations
Given the widespread application of carbon fiber reinforced polymer (CFRP) structures in aerospace and automotive manufacturing, the critical role of structural health monitoring (SHM) is increasingly evident. Traditional damage quantification methods exhibit poor transferability and reusability in different structural areas, which are difficult to adapt to the complex and diverse damage characteristics of CFRP structures. To address this challenge, a multi-scale adaptive cross-area damage quantification method based on transfer learning (MSACADQ-TL) is proposed in this paper. Firstly, a multi-scale feature extraction framework is presented, which decomposes signals from various paths into multiple frequency domain sub-signals and analyzes the dynamic trends of these signals across different signal scales, providing richer damage feature information. Moreover, the difference between the source domain and the target domain is minimized to enhance transferability of features. Furthermore, this paper applies a damage quantification based on the Bidirectional Gated Recurrent Unit (BiGRU) model to achieve cross-area damage quantification. To validate the proposed method’s effectiveness, this paper designed three different cross-area damage transfer scenarios. Experimental results demonstrate that the MSACADQ-TL model outperforms other comparison methods in damage quantification tasks, which is effective for quantification damage in CFRP structures.
Zhi-Jun Liu, Haijun Zhang, Lijing Wang· Italian National Conference...· 0 citations
Structural health monitoring based on vibration-induced damage detection has become increasingly important for ensuring the safety and integrity of frame structures. Although deep learning techniques have significantly improved detection performance, many existing models rely on single-domain inputs and therefore fail to fully capture the complex characteristics of structural vibration responses under real-world conditions. To address this limitation, this study proposes a novel temporal-modal parallel convolutional neural network (TM-PCNN) framework that integrates temporal information with modal characteristics for enhanced structural damage detection. The proposed framework adopts a dual-stream parallel feature-extraction architecture, consisting of a temporal convolutional network (TCN) branch for learning long-range temporal dependencies and a two-dimensional convolutional neural network (2D-CNN) branch for extracting latent modal features from inner-product matrix representations. By fusing temporal and modal features, the TM-PCNN framework enables more accurate and robust damage identification. To validate the proposed method, experiments were conducted on a five-story steel frame structure. The TM-PCNN model was compared with several representative baseline methods, including inner product matrix-2D-CNN, TCN, extreme gradient boosting (XGBoost), and light gradient boosting machine (LightGBM). The experimental results show that the proposed model achieved an accuracy of 98.57%. In addition, feature visualization using principal component analysis and t-distributed stochastic neighbor embedding demonstrates that TM-PCNN learns compact and highly separable feature representations. These results confirm that the proposed framework provides an effective and promising solution for structural health monitoring applications.
Dongxue Li, Ying-ni He, Likai Zhang et al.· Structural Health Monitoring· 0 citations
Abstract. High-performance mechanical component structural health monitoring (SHM) is a vital issue in contemporary engineering, especially in the aerospace, automotive, and industrial turbomachinery sectors where component failure may be disastrous. This article introduces a new AI-aided SHM framework with multimodal sensor fusion and ensemble deep learning architecture with one-dimensional convolutional neural networks (1D-CNN) and bidirectional long short-term memory (Bi-LSTM) networks, which can be used to detect fault, classify fault, and predict remaining useful life (RUL) in real-time. The proposed system takes in time-series streams of vibration, acoustic emission, and strain gauge data, runs them through an adaptive signal preprocessing pipeline, and derives hierarchical features of fault relevance without using manually specified features. The tests are done on two benchmark datasets, which include the CWRU bearing fault dataset and a custom gas turbine blade fatigue dataset that were obtained under controlled laboratory settings and a real gas turbine compressor testbed. The CNN-BiLSTM ensemble suggested has an accuracy of fault classification of 98.7 and a mean absolute percentage error (MAPE) of 3.14 to predict RUL with average inference latency of 12.3 ms, which is appropriate to be integrated into embedded systems in real-time. These findings constitute a statistically significant step forward compared to the current state-of-the-art baselines and they generalize and scale to provide an AI-SHM paradigm of safety-critical mechanical systems.
Jasjeet Singh· Materials Research Proceedin...· 0 citations
The accumulation of fatigue damage in heavy-haul locomotive couplers significantly degrades their load-bearing capacity and accelerates structural failure, making accurate damage estimation imperative for operational safety. Although the rainflow counting method (RCM) has long served as the standard for quantifying cumulative fatigue damage under variable amplitude loading, existing RCM-based frameworks are hindered by computationally exhaustive load history traversals and the cumbersome process of repeatedly mapping massive load cycles to physical priors. To address these issues, a dual-order synergistic complementary attention network (DSCANet) surrogate model for cumulative fatigue damage estimation of couplers is proposed in this article. First, operating on meticulously extracted multidomain inputs, an encoder-decoder architecture with skip connections is employed to perform high-dimensional mapping and reconstruction of inputs, effectively preserving critical damage-inducing features. To comprehensively capture fatigue degradation characteristics, the model divides damage assessment into two complementary subtasks. Specifically, the local branch mines cross-domain coupling relationships among damage-contributing features through multiscale dilated convolutions, while the global branch models damage accumulation dependencies and dynamically weights feature importance via a dual-path complementary attention mechanism with asymmetric head designs. The adaptive fusion of these two representations enables precise evaluation of cumulative fatigue damage. Experimental results demonstrate that the proposed DSCANet accurately predicts cumulative fatigue damage while outperforming traditional methods in computational efficiency.
Zirui Song, Shiqian Chen, Maoyong Dong et al.· IEEE Transactions on Instrum...· 0 citations
The development of concrete cracks will affect the durability, serviceability, and safety of civil infrastructure. This study proposes a leakage-controlled convolutional neural network long/short-term memory (LSTM) framework for static crack detection and short-term prediction of crack severity from time-ordered image sequences. The temporal task is defined as a next-step prediction of the crack severity index (CSI) derived from the image, rather than claiming complete physical crack geometry prediction from bounding box annotations. The CSI is calculated based on the ratio of crack-box and region-of-interest areas. The ResNet34 encoder extracts frame-wise spatial features, and a two-layer LSTM models the temporal dependency of structure-level sequences. Sequence construction, augmentation usage, split rules, optimisation settings, and evaluation metrics are selected to improve reproducibility and avoid temporal leakage. On the dataset, the framework achieves 93.0% classification accuracy, 88.3% mean average precision (mAP)@0.5, 0.76 mean intersection over union, 0.015 normalised mean squared error, and 89.2% F1 score. It also clarifies the practical interpretation of CSI growth, discusses detector-level benchmarking, domain shift and uncertainty estimation, and takes this method as an interpretable short-horizon monitoring framework, rather than a fully calibrated crack metrology system.
Jun Tang, Guo-Lin Guo, Ting Wang et al.· Proceedings of the Instituti...· 0 citations
