Machine-learning-based methods for output-only structural modal identification

Nonlinearity and randomness are two intrinsic characteristics of the mechanical behavior of concrete material. The structural response under large excitation can barely be predicted without considering these two characteristics. Brilliant works have been done for decades in the material science and computational stochastic mechanics. However, the existed numerical methods are usually parameter dependent and the key mechanical properties of concrete material are determined by empirical recognition. Therefore, in this paper, a data-driven multi-scale constitutive model is proposed for representing the mechanical behavior of concrete material based on the polynomial chaos expansion and stochastic damage model. Several groups of compressive stress–strain data of concrete material are applied to train the proposed model. By cross validation of the prediction and the concrete stress–strain experimental data, the proposed model is firstly verified to have a robust performance to
gain accurate prediction results. Afterwards, the proposed method is compared with a neural network method, the results shows that the proposed method is more robust and accurate than the neural network method.

Structural health monitoring systems continuously monitor the operational state of structures, generating a large amount of monitoring data during the process. The structural responses of extreme events, such as earthquakes, ship collisions, or typhoons, could be captured and further analyzed. However, it is challenging to identify these extreme events due to the interference of faulty data. Real-world monitoring systems suffer from frequent misidentification and false alarms. Unfortunately, it is difficult to improve the system’s built-in algorithms, especially the deep neural networks, partly because the current neural networks only output results and do not provide an interpretable decision-making basis. In this study, a deep learning-based method with visual interpretability is proposed to identify seismic data under sensor faults interference. The transfer learning technique is employed to learn the features of seismic data and faulty data with efficiency. A post-hoc interpretation algorithm termed Gradient-weighted Class Activation Mapping (Grad-CAM) is embedded into the neural networks to uncover the interest regions that support the output decision. The in-situ seismic responses of a cable-stayed long-span bridge are used for method verification. The results show that the proposed method can effectively identify seismic data mixed with various types of faulty data while providing good interpretability.

Structural health monitoring (SHM) is a multi-discipline field that involves the automatic sensing of structural loads and response by means of a large number of sensors and instruments, followed by a diagnosis of the structural health based on the collected data. Because an SHM system implemented into a structure automatically senses, evaluates, and warns about structural conditions in real time, massive data are a significant feature of SHM. The techniques related to massive data are referred to as data science and engineering, and include acquisition techniques, transition techniques, management techniques, and processing and mining algorithms for massive data. This paper provides a brief review of the state of the art of data science and engineering in SHM as investigated by these authors, and covers the compressive sampling-based data-acquisition algorithm, the anomaly data diagnosis approach using …

Surrogate model methods are widely used in structural reliability assessment, but conventional sampling methods require a large number of experimental points to construct a surrogate model. Inspired by the learning process of the AlphaGo, which is essentially optimization of sampling, we proposed a deep reinforcement learning (DRL)-based sampling method for structural reliability assessment. First, the sampling space and the existing samples are transformed into an array that is treated as the state in DRL. Second, a deep neural network is designed as the agent to observe the sampling space and select new experimental points, which are treated as actions. Finally, a reward function is proposed to guide the deep neural network to select experimental points along the limit state surface. Two numerical examples including a benchmark problem are employed to illustrate the sampling ability of the proposed …