Stress Monitoring Based on Lamb Waves Measurements and a Convolutional Neural Network

Abstract

Monitoring of static and dynamic stresses is crucial for structural performance evaluation and prevention of unexpected structural failure due to overloading, stress concentration and fatigue. For quasi-static stress monitoring, ultrasonic measurement is an attractive method because of its non-destructive nature and readiness for online monitoring. The well-established acoustoelastic effect shows that the ultrasonic wave velocity exhibits a linear relationship with the stress within the linear elasticity region of a material when the wave propagates through the material. In this study, an online stress monitoring technique is proposed for metallic plate-like structures based on Lamb-wave measurements and a Convolution Neural Network (CNN) under dynamic as well as static loadings. First, an aluminum (6061-T6) plate specimen is fabricated, and piezoelectric (PZT) transducers are installed on the specimen. Then, different levels of static loadings are applied to the specimen, and the ultrasonic responses are obtained at each loading level. For Lamb wave generation and sensing, a conventional data acquisition system is used, and the applied loads (ground truths) are measured using the load cell built in the loading machine. Next, a CNN is designed and trained by defining the ultrasonic response in the time domain as the input and the measured stress level as the output. For the input, one-dimensional time domain ultrasonic responses are rearranged to two-dimensional images, and the images constitute the inputs of the CNN without feature extraction. Finally, the performance of the trained CNN is examined using blind test data obtained from various static and constant amplitude cyclic loading conditions. For the monitoring of dynamic loadings, the data acquisition parameters such as the signal duration and the repetition rate are controlled and optimized. The uniqueness of this technique lies in (1) automated stress estimation without conventional feature extraction based on human intervention such as acoustoelastic effects, and (2) stress monitoring under constant amplitude cyclic loading loadings up to 5 Hz.