This study utilizes vibration-based signal analysis as a non-destructive testing technique that involves analyzing the vibration signals produced by a structure to detect possible defects or damage. The study aims to employ deep learning models to identify defects in a 3D-printed cantilever beam by analyzing the beam’s tip displacement given a random input signal generated by an electromagnetic shaker. This study is focused on the output signal without any information of the random input, which is common for structural health monitoring applications in practice. Additionally, the study has revealed that the number of times the test set is applied to the trained model significantly impacts the accuracy of the model’s consistent predictions.

Structural health monitoring provides significant and obvious potential in enhancing our life safety and extending the service life of systems. In practice, structures are often exposed to random excitations without knowing the exact characteristics of its source. This paper proposes a novel method of the implementation of LSTM and GRU models for characterizing anomalies in output-only random vibration signals. In the proposed approach, the time response of a 3D printed PLA beam is measured when subjected to a random excitation and used to train LSTM and GRU models. Healthy time response and four additional cases that contain a small mass at varied locations along the beam are used as model inputs. These inputs represent normal and abnormal signals which are then classified to diagnose the health state of the structure. In this study, the random vibration time responses with large amplitude (so-called signal caricature) were selectively employed for creating the output-only models. The results illustrate the signal caricature data set leading to both more accurate and efficient characterization of the structural health state, compared to utilizing the entire time response as an input to the models. Modal properties obtained by traditional vibration analysis support the effectiveness of utilizing the signal caricature with LSTM/GRU models, demonstrating great potential in identifying defects in practical applications.

Structural health monitoring is an area of growing interest and is worthy of new and innovative approaches. Since the automatic diagnosis of structures is very complex and challenging, recent research to apply deep learning techniques has been actively conducted. In this study, we assumed that a PLA beam copied by 3D printing is the smallest unit constituting a complex structure and applied GRU to detect defects. To set the defect point of the beam, a total of four holes were drilled at regular intervals, and then a mass was attached. Signals at different locations were collected through a vibrator and trained through GRU, and the results were compared in terms of RMSE value. As a result of this experiment, we checked the defect by inputting test data into the trained model and were able to measure the defect degree of the PLA beam with a weighted average F1 score of 84%.

Vibration-based methods have received considerable attention in structural condition monitoring applications. We have proposed a model to detect damaged points of a target structure using the GRU model and classify the 0.84 overall accuracy. To increase the model's accuracy in this research, we propose an ensemble deep learning model using LSTM and bi-directional LSTM incorporated with GRU. Each model predicted its RMSE trend and combined the damage estimation results from both models, which are mostly close to the true damage locations. As a result of synthesizing the three algorithms, the damage point of the cantilever beam was found with an accuracy of 0.88 and a misclassification rate of 0.12. The results indicate that the proposed combined approach provides enhanced reliability than a single algorithm.