Real-time damage detection in structural health monitoring via reinforcement learning

A novel real-time automated damage detection method leveraging reinforcement learning has been unveiled in a recent study published in Engineering. The research, spearheaded by Chengwen Zhang, Qing Chun, and Yijie Lin from Southeast University, China, introduces an innovative approach that significantly enhances the efficiency and accuracy of damage detection in structural health monitoring (SHM) systems, particularly for architectural heritage.

Traditional damage detection methods, such as those based on finite element model (FEM) updating and sensitivity Jacobian matrices, often fall short in terms of computational efficiency and convergence ratio for online calculations. To address these limitations, the researchers developed the theory-assisted adaptive multi-agent twin delayed deep deterministic policy gradient algorithm (TA2-MATD3). This method integrates reinforcement learning with theoretical frameworks to improve damage detection capabilities.

The study establishes a theoretical framework for reinforcement-learning-driven damage detection and introduces a theory-assisted mechanism and an adaptive experience replay mechanism to enhance the traditional multi-agent twin delayed deep deterministic (MATD3) algorithm. The effectiveness of TA2-MATD3 was demonstrated using 12-month SHM data from a historical residential house built in 1889. The results indicated that TA2-MATD3 achieved a computational efficiency approximately 117 to 160 times higher than traditional methods. The convergence ratio of damage detection on the training set was around 97%, and on the test set, it ranged from 86.2% to 91.9%. Moreover, the method successfully identified main apparent damages found in field surveys.

The TA2-MATD3 algorithm incorporates a teacher actor generator (TAG) to generate virtual damage distributions, which are used to train the student actor network. The training process is divided into four stages: initial training with TAG, updating the student actor network to minimize action errors, gradually increasing the participation of the student actor network, and finally, training the network to maximize long-term benefits while reducing the weight of minimizing action errors. The adaptive experience replay mechanism ensures efficient sampling and learning from the experience replay pool.

The research highlights the potential of reinforcement learning in SHM by significantly improving online computing efficiency and damage detection accuracy. The proposed method not only addresses the challenges of high computational costs and ill-conditioned matrix solutions in traditional FEM updating but also offers a robust solution for real-time damage assessment in architectural heritage. Future work may focus on scaling up the method for larger civil engineering structures and enhancing its adaptability to complex environmental conditions.

The paper “An Automatic Damage Detection Method Based on Adaptive Theory-Assisted Reinforcement Learning,” is authored by Chengwen Zhang, Qing Chun, Yijie Lin. Full text of the open access paper: https://doi.org/10.1016/j.eng.2025.03.026. For more information about Engineering, visit the website at https://www.sciencedirect.com/journal/engineering.