A Systematic Review of Structural Health Monitoring Using Artificial Neural Networks: From Traditional Neural Networks to Deep Learning Algorithms.

With recent advances in neural network models and computational performance, structural health monitoring studies have shifted their methods from feature engineering or manual feature extraction to the use of machine learning models, especially neural networks and deep learning. The neural networks have been used for damage classification, feature extraction, dimensionality reduction, and image and 3D video-based inspection. Deep learning neural networks have shown a higher performance and accurate prediction and classification of the damage severity and location. This paper presents a detailed literature review of existing neural network-based applications in structural health monitoring and their main types, mathematical models, definitions, applications and feature extraction methods. This paper also surveyed the current literature in the context of implementation methods, quantification levels, neural network types, and mechanisms. A comparison of the performance of different deep learning neural networks and traditional neural networks is presented to help determine the neural network techniques for specific applications. It highlights the performance of deep learning algorithms and hybrid neural networks in terms of accuracy, data storage, and time complexity. The challenges facing structural health monitoring-based neural network applications can be summarized in main points, including but not limited to fully automated approaches, generalization of new data sets for population-based damage detection, model-associated problems, interpretability of the ANNs model, and real-time deployment.