|
基于计算机视觉的桥梁健康监测数据异常检测 |
Anomaly Detection of Bridge Health Monitoring Data Based on Computer Vision |
投稿时间:2024-01-29 修订日期:2024-04-08 |
DOI: |
中文关键词: 数据异常 时频分析 计算机视觉 桥梁健康监测 |
英文关键词: data anomalies time-frequency analysis computer vision bridge health monitoring |
基金项目:国家自然科学基金资助项目(51878234) |
|
摘要点击次数: 611 |
全文下载次数: 0 |
中文摘要: |
针对桥梁健康监测系统中因设备故障、环境影响等导致部分监测数据产生异常,将时频分析和计算机视觉相结合,用于监测数据的异常检测。首先根据监测数据时域图像对其进行类型划分和标记,采用时频分析法实现监测数据样本可视化,制备图像数据库用于构建和训练深度神经网络模型。然后利用深度学习框架搭建ResNet18神经网络模型并对其进行训练和验证,通过反向传播机制和Adam优化算法更新和优化模型权重参数,使用批标准化、数据增强等方法提高模型准确率和泛化能力。最后使用完成训练的模型对其他桥梁监测数据样本进行异常检测,进一步验证模型性能。结果表明,提出的方法对监测数据异常检测准确率为95.62%,对其他桥梁监测数据样本检测准确率为95.28%,具有良好的稳定性和识别性能。 |
英文摘要: |
In order to address the issue of abnormal monitoring data caused by equipment failures and environmental influences in the bridge health monitoring system, a combined approach of time-frequency analysis and computer vision is proposed for anomaly detection. Firstly, the monitoring data is categorized and labelled based on its time-domain images, and the time-frequency analysis method is utilized to visualize the monitoring data samples, creating an image database for constructing and training a deep neural network model. Subsequently, a ResNet18 neural network model is built using a deep learning framework, it is trained and validated. The model's weight parameters are updated and optimized through the backpropagation mechanism and the Adam optimization algorithm, using methods such as batch normalization and data augmentation to improve the model's accuracy and generalization ability. Finally, the trained model is used to detect anomalies in other bridge monitoring data samples, further validating the model's performance. The results show that the proposed method achieves an accuracy of 95.62% in detecting abnormal monitoring data and 95.28% in detecting other bridge monitoring data samples, demonstrating good stability and recognition performance. |
View Fulltext
查看/发表评论 下载PDF阅读器 |
关闭 |
|
|
|