Bayesian-Based Hyperparameter Optimization of 1D-CNN for Structural Anomaly Detection

• Published in: Sensors (Basel, Switzerland) Vol. 23; no. 11; p. 5058
• Main Authors: Li, Xiaofei, Guo, Hainan, Xu, Langxing, Xing, Zezheng
• Format: Journal Article
• Language: English
• Published: Switzerland MDPI AG 25.05.2023; MDPI
• Subjects: 1-D convolutional neural network; Accuracy; Analysis; Bayesian optimization algorithm; Big data; decision-level fusion; Deep learning; Identification; Measuring instruments; Neural networks; Optimization algorithms; Sensors; structural anomaly detection; vibration signals; Wavelet transforms; structural anomaly detection; 1-D convolutional neural network; Bayesian optimization algorithm; decision-level fusion; vibration signals
• DOI: 10.3390/s23115058
• ISSN: 1424-8220

With the rapid development of sensor technology, structural health monitoring data have tended to become more massive. Deep learning has advantages when handling big data, and has therefore been widely researched for diagnosing structural anomalies. However, for the diagnosis of different structural abnormalities, the model hyperparameters need to be adjusted according to different application scenarios, which is a complicated process. In this paper, a new strategy for building and optimizing 1D-CNN models is proposed that is suitable for diagnosing damage to different types of structure. This strategy involves optimizing hyperparameters with the Bayesian algorithm and improving model recognition accuracy using data fusion technology. Under the condition of sparse sensor measurement points, the entire structure is monitored, and the high-precision diagnosis of structural damage is performed. This method improves the applicability of the model to different structure detection scenarios, and avoids the shortcomings of traditional hyperparameter adjustment methods based on experience and subjectivity. In preliminary research on the simply supported beam test case, the efficient and accurate identification of parameter changes in small local elements was achieved. Furthermore, publicly available structural datasets were utilized to verify the robustness of the method, and a high identification accuracy rate of 99.85% was achieved. Compared with other methods described in the literature, this strategy shows significant advantages in terms of sensor occupancy rate, computational cost, and identification accuracy.