Inspection Cover Damage Warning System Using Deep Learning Based on Data Fusion and Channel Attention

• Published in: Electronics (Basel) Vol. 14; no. 12; p. 2383
• Main Authors: Zhang, Kaiyu, Wang, Baohua, Chen, Hongyan, Peng, Huaijun, Xue, Lei, Han, Baojiang, Tang, Zhili, Liu, Yuzhang
• Format: Journal Article
• Language: English
• Published: Basel MDPI AG 11.06.2025
• Subjects: Accuracy; Adaptation; Adaptive algorithms; Algorithms; Analysis; Artificial intelligence; Broadband; Curve fitting; Data acquisition; Data integration; Deep learning; Edge computing; Efficiency; Fault diagnosis; Fourier transforms; Impact loads; Infrastructure; Inspection; Machine learning; Manhole covers; Manholes; Neural networks; Nondestructive testing; Optimization; Real time; Safety and security measures; Sensors; Signal processing; Structural vibration; Vibration; Vibration analysis; Warning systems; Wavelet transforms; Working conditions; China
• ISSN: 2079-9292; 2079-9292
• DOI: 10.3390/electronics14122383

This paper explores the application of artificial intelligence in urban energy infrastructure construction and enhances the operation and maintenance safety of infrastructure through edge computing and advanced sensors. At present, urban manhole covers cover a large number of roads, but there is a lack of effective real-time monitoring methods. In order to effectively solve these problems, this study proposes a domain adaptive network algorithm (EDDNet) based on data fusion. By optimizing the loss function, the attention mechanism is used to make the model pay more attention to the deep features related to the abnormal state of the inspection cover. The algorithm solves the problem of broadband vibration analysis and reduces the misclassification rate in various behavioral scenarios, including pedestrian traffic, slow-moving vehicles, and intentional surface collisions. A data acquisition sensor network is established, and a six-degree-of-freedom coupled vibration model and a structural vibration model of the inspection cover are established. The vibration peak under high load conditions is modeled and simulated using impact load data, and a fitting curve is generated to achieve deep optimization of the model and enhance robustness. The experimental results show that the classification accuracy of the network reaches 95.23%, which is at least 10.2% higher than the baseline model.