工业机器人夹爪失效检测中深度学习与传统机器学习方法的对比研究

Deep Learning Versus Traditional Machine Learning for Grip-Loss Detection in Industrial Robots: A Systematic Comparative Study

摘要: 工业机器人夹爪失效可能导致严重的生产事故与经济损失，对其进行实时准确检测具有重要工程价值。本文基于包含7355条传感器记录、正负样本比约为29比1的高度失衡机器人关节数据集，系统比较了三种传统机器学习方法，即决策树、支持向量机和K近邻算法，与四种深度学习及特征增强方法，即前馈神经网络、一维卷积神经网络、长短期记忆网络和注意力神经网络，在夹爪失效二分类任务上的性能表现。评估指标涵盖平衡准确率、精确率、召回率、F1分数及曲线下面积，并结合五折交叉验证、学习曲线分析和实际计算耗时进行全面考察。实验结果表明，K近邻算法的平衡准确率达0.8604，综合效率评分为0.8000，训练时间仅0.01秒，均优于其他方法，展现出在工业在线检测场景中的显著优势。本研究为工程师在实际部署中选择合适检测算法提供了量化依据和可复现的评估框架。

Abstract: Grip-loss events in industrial robots can trigger severe production accidents and economic losses, making their real-time detection a problem of considerable engineering relevance. This paper presents a systematic comparison of three traditional machine learning methods, namely Decision Tree, Support Vector Machine, and K-Nearest Neighbors, against four deep learning and feature-augmentation approaches, namely Feedforward Neural Network, 1D-Convolutional Neural Network, Long Short-Term Memory Network, and Attention-based Neural Network, on a grip-loss binary classification task. The dataset comprises 7355 real sensor records collected from a six-joint industrial robot, with a severe class imbalance ratio of approximately 29:1 between normal and grip-loss samples. To address this imbalance, cost-sensitive matrices, sample weighting, and minority-class oversampling are applied consistently across methods. All seven approaches are evaluated under a unified framework using balanced accuracy, precision, recall, F1-score, and area under the ROC curve, supplemented by five-fold cross-validation, learning curve analysis, and measured training and inference times. Experimental results show that K-Nearest Neighbors achieves the highest balanced accuracy of 0.8604, an overall recommendation score of 0.8000, and a training time of only 0.01 seconds, outperforming all other methods on every primary criterion. The findings demonstrate that lightweight instance-based methods can match or surpass more complex deep learning approximations on structured, low-dimensional sensor data with extreme class imbalance, and that high AUC does not guarantee practical utility at a fixed decision threshold. Feature importance analysis consistently identifies Current_J2 as the most discriminative signal for grip-loss detection. The proposed evaluation framework is directly transferable to other highly imbalanced industrial fault detection tasks.