Identification of anomaly of tunnel segment strain using an adaptive machine learning model

Anomaly identification is a crucial issue for preventing diseases in tunnel structures. However, distinguishing whether the structure is abnormal from outliers is challenging, as the monitoring dataset composed of time series may contain non-structural abnormal data caused by operational environmental pollution. Therefore, this study aims to propose a structural anomaly identification model to distinguish the structural anomaly from the polluted dataset based on an improved autoencoder model with adaptive loss function modification (AAE). To solve the problem of insufficient abnormal samples in practice, the proposed model is applied to a time series dataset obtained from numerical simulation, where extreme conditions are introduced to generate structural abnormal data, and Gaussian noises are overlaid to pollute the raw data and generate non-structural abnormal data. Then, the anomaly identification capability of the AAE model under different intensities of noise pollution conditions is discussed, along with its reliability compared to other common methods. Experimental results demonstrate that the AAE model can effectively identify structural anomaly information from heavily polluted datasets. The accuracy of the AAE model improved by at least 4% to other models even on a serious polluted dataset. Therefore, the structural anomaly identification model presented in this study is reliable.