Study of the abnormal vibration detection method for in-service structure using semi-supervised Learning by autoencoder

Abstract

This study concerns a method for detecting the structural anomaly of F type support information board from acceleration measurements. The evaluation of structural anomaly is performed by the autoencoder method．By the proposed method, training of the autoencoder is conducted from data of normal conditions, and diagnosis of the structure is conducted from reconstruction error of the autoencoder. To validate the effectiveness of the method, the method is applied to long-term measurement data from actual equipment. To evaluate the detection accuracy and reliability of the method, the method is applied to sites with fluctuating natural frequencies and sites with stable natural frequencies. As a result, the proposed method with the autoencoder is able to correctly evaluate anomalies of the structure with more than 80% detection and a low false positive rate of 5%, and is able to evaluate the normality of the structure stably over a long period of time.

Keywords

• structural diagnosis
• neural network
• anomaly detection
• structural health monitoring
• mode analysis
• vibration analysis
• autoencoder
• structure in service