作者投稿和查稿 主编审稿 专家审稿 编委审稿 远程编辑

计算机工程 ›› 2024, Vol. 50 ›› Issue (10): 370-380. doi: 10.19678/j.issn.1000-3428.0068533

• 开发研究与工程应用 • 上一篇    下一篇

基于电机数据图像化的多时序变量间接卡车误吊起检测

刘嘉杰1, 刘国平2,*(), 胡文山1   

  1. 1. 武汉大学电气与自动化学院, 湖北 武汉 430072
    2. 南方科技大学控制科学技术中心, 广东 深圳 518055
  • 收稿日期:2023-10-10 出版日期:2024-10-15 发布日期:2024-01-12
  • 通讯作者: 刘国平
  • 基金资助:
    国家自然科学基金(62173255); 国家自然科学基金(62073247); 国家自然科学基金(62103308)

Multi-Time-Series Variable Indirect Truck Lifting Detection Based on Motor Data Visualization

LIU Jiajie1, LIU Guoping2,*(), HU Wenshan1   

  1. 1. School of Electrical Engineering and Automation, Wuhan University, Wuhan 430072, Hubei, China
    2. Center for Control Science and Technology, Southern University of Science and Technology, Shenzhen 518055, Guangdong, China
  • Received:2023-10-10 Online:2024-10-15 Published:2024-01-12
  • Contact: LIU Guoping

摘要:

自动化集装箱码头的装卸作业中经常发生集装箱与卡车同时被吊起的安全事故, 导致人员伤亡及货品、车辆的损坏。为解决该问题, 提出一种基于电机数据图像化处理的多时序变量间接卡车误吊起检测方法(MEIN)。该方法通过神经网络分析异步电机在吊起集装箱和卡车的过程中产生的电流和电压异常, 从而判断是否发生了误吊起事故。采集吊机的三相电流和电压数据, 并基于物理公式进行特征工程计算出多个相关时序物理量, 采用滑动窗口、SMOTE-Tomek综合采样的方式扩大样本总数并平衡类别数量, 最后将多时序变量转换为图像的形式以EfficientNet进行分类。实验结果表明, 该方法能在复杂的环境下(例如雨雾天气或轮胎被遮挡)保持稳定的检测性能, 各测试地区的AUC均在0.997以上。相较于传统的基于激光雷达和计算机视觉的检测方法, MEIN方法具有成本低、精度高、计算量小并且抗环境干扰能力强等优点。该方法已在武汉、青岛、钦州、梅山等多地部署, 为提高自动化集装箱码头的作业安全提供一种有效的解决方案。

关键词: 时间序列分类, 卷积神经网络, 合成少数类样本的过采样技术, Tomek Links欠采样技术, 卡车误吊起检测

Abstract:

In automated container terminal operations, simultaneous lifting of containers and trucks has led to significant safety issues, including damage to goods and vehicles and, in severe cases, human fatalities. To mitigate these issues, a novel approach, Motor-data Engineering and image processing-based Indirect truck lifting detection Network (MEIN), is introduced. MEIN collects three-phase current and voltage data from cranes and employs feature engineering based on physical formulas to calculate numerous related time-series physical quantities. It combines a Sliding window with Tomek-Synthetic Minority Oversampling Technique(SMOTE) for sampling, which enhances sample size and balances category quantities. This process converts multi-time-series variables into an image format for classification via EfficientNet. In tests, MEIN demonstrated robust detection capabilities, even in challenging conditions such as adverse weather or obscured tires, achieving an Area Under the receiver operator characteristic Curve (AUC) above 0.997 across various test regions. Compared to traditional detection methodologies such as lidar and computer vision, MEIN offers notable advantages including reduced cost, heightened accuracy, reduced computational demands, and enhanced resilience to environmental disturbances. Its successful implementation in Wuhan, Qingdao, Qinzhou, and Meishan underscores its its effectiveness in improving safety at automated container terminals.

Key words: time-series classification, Convolutional Neural Network(CNN), Synthetic Minority Oversampling Technique(SMOTE), Tomek Links downsampling technique, truck lifting detection