Vibration method for tendon force loss detection using large kernel convolution and semi-supervised training

As the use of prestressed concrete (PSC) becomes more prevalent in critical infrastructure, it’s increasingly important to accurately monitor the force in the tendons to prevent problems such as cracking and deflection. While various methods have been considered adequate for real-world applications, the use of vibration-based methods has been limited due to the inconsistent relationship between natural frequency and tendon force. This research aims to address this challenge by implementing a deep learning approach. Our proposed method employs a convolutional encoder with large kernel sizes and a semi-supervised loss function. When tested on a laboratory-scale PSC beam, the method demonstrated an impressive ability to differentiate between normal conditions and loss of tendon force, achieving 92.85% accuracy. It effectively categorized and mapped acceleration data in the feature space, including data not used during training. Comparative studies with other machine learning methods confirm the superiority of the proposed approach.