شناسایی الگوهای ناهنجاری داده‌های شتاب‌سنج در پل‌های کابلی خود ایستا به روش یادگیری ماشین غیرنظارت شده

Nowadays, Structural Health Monitoring (SHM) has become an important field in civil engineering, aiming to evaluate and monitor the condition of structures and detect any potential anomalies and damage. One of the key tools in SHM is the use of accelerometers to collect structural data for damage detection. However, these sensors can experience anomalies due to noise, sensor failure, or environmental condition changes, among other reasons, unrelated to structural damage. If these anomalies are not identified, they can compromise the entire SHM system's performance, and the system might mistakenly report sensor failure as structural member damage. These anomalies can lead to incorrect data interpretation and, consequently, wrong decision-making. Therefore, developing efficient methods to detect and correct sensor anomalies is of high importance.
In this paper, various sensor anomalies in real bridge benchmark data are evaluated, and an unsupervised machine learning method using a Convolutional Neural Network (CNN) with auto encoder architecture is introduced to identify anomalous data from healthy data. Two metrics, reconstruction error and probability estimation function, were used for anomaly detection. The results show that the use of the probability estimation function significantly increases accuracy and that the proposed method can effectively identify anomalies, as the accuracy improves by about 10% when using the probability estimation function. The proposed method can be applied as an efficient unsupervised method.