面向预测性维护的CNC刀具磨损状态识别——机器学习与深度学习对比、多轴信号表征及SHAP可解释分析
CNC Tool Wear State Identification for Predictive Maintenance—Comparison of Machine Learning and Deep Learning, Multi-Axis Signal Representation and SHAP Interpretable Analysis
DOI: 10.12677/csa.2026.1611334, PDF,    科研立项经费支持
作者: 黄扬庭, 颜轲越*:广东金融学院大数据与人工智能学院,广东 清远;岳子焕, 李 莹:北京理工大学(珠海)国际教育与培训学院,广东 珠海
关键词: 刀具磨损;多轴信号;分组交叉验证;AdaBoost;SHAP;预测性维护;Tool Wear; Multi-Axis Signals; Grouped Cross-Validation; AdaBoost; SHAP; Predictive Maintenance
摘要: 针对机床时序数据中独立加工实验数量少、相邻窗口相关性强及模型判别依据难以解释的问题,本文构建了结合多轴统计变量、加工实验分组验证与SHAP解释的刀具磨损状态识别方法。本文基于18次计算机数字控制(CNC)蜡铣削实验,从电流、主轴功率、实际速度和速度误差等10个信号通道提取40项统计变量,并采用外层留一加工实验与内层分层四折验证,比较11种机器学习分类器和6种时序深度学习模型;预处理和变量筛选均在训练折内完成。模型比较结果表明,使用全部40项变量的AdaBoost、逻辑回归与GBDT准确率均为77.78%,其中AdaBoost的平衡准确率为77.50%,磨损与未磨损召回率分别为80.00%和75.00%。在6种时序深度学习模型中,1D-CNN的准确率和平衡准确率最高,分别为66.67%和67.50%。变量筛选结果表明,SHAP筛选提高了部分分类器的准确率;在由内层验证从5种分类器中选择模型的实验中,保留SHAP排名前15项变量与使用全部40项变量的准确率均为72.22%。进一步对全变量AdaBoost进行SHAP解释发现,X、Y轴电流水平、主轴电流与速度波动以及Y轴速度误差是模型识别刀具磨损状态的主要信号依据。本文将识别性能比较与模型判别依据相结合,为小样本刀具磨损识别和维护人员复核预测结果提供了方法参考。
Abstract: To address the small number of independent machining experiments, strong correlations between adjacent windows, and difficulties in interpreting model predictions in machine tool time-series data, this study develops a tool wear state identification method combining multi-axis statistical features, validation grouped by machining experiment, and SHAP interpretable analysis. Based on 18 Computer Numerical Control (CNC) wax-milling experiments, this study extracts 40 statistical features from ten signal channels, including current, spindle power, actual velocity, and velocity error, and compares 11 machine learning classifiers and six time-series deep learning models using outer leave-one-experiment-out validation and inner stratified four-fold validation; preprocessing and feature selection are performed within the training folds. The model comparison results show that AdaBoost, logistic regression, and GBDT using all 40 features each achieve an accuracy of 77.78%. AdaBoost achieves a balanced accuracy of 77.50%, with recalls of 80.00% and 75.00% for the worn and unworn classes, respectively. Among the six time series deep learning models, 1D-CNN achieves the highest accuracy and balanced accuracy, at 66.67% and 67.50%, respectively. The feature selection results show that SHAP based selection improves the accuracy of some classifiers; in the experiment where inner validation selects a model from five classifiers, retaining the top 15 SHAP ranked features and using all 40 features both yield an accuracy of 72.22%. Further SHAP analysis of AdaBoost using all features identifies X- and Y-axis current levels, fluctuations in spindle current and velocity, and Y-axis velocity error as the main signal evidence used by the model to identify tool wear states. By combining identification performance comparisons with explanations of model predictions, this study provides a methodological reference for tool wear identification with small samples and for maintenance personnel to review prediction results.
文章引用:黄扬庭, 颜轲越, 岳子焕, 李莹. 面向预测性维护的CNC刀具磨损状态识别——机器学习与深度学习对比、多轴信号表征及SHAP可解释分析[J]. 计算机科学与应用, 2026, 16(11): 169-183. https://doi.org/10.12677/csa.2026.1611334

参考文献

[1] 张媛媛, 杨奕东, 陈晓丽. 清远传统制造业数字化转型问题与对策研究[J]. 商场现代化, 2026(1): 5-8.
[2] Vial, G. (2019) Understanding Digital Transformation: A Review and a Research Agenda. The Journal of Strategic Information Systems, 28, 118-144.
https://doi.org/10.1016/j.jsis.2019.01.003
[3] 田佩彬, 王宝金, 沈锦桃. 刀具磨损状态智能监测技术研究现状与展望[J]. 林业机械与木工设备, 2021, 49(12): 4-8, 20.
[4] Munaro, R., Attanasio, A. and Del Prete, A. (2023) Tool Wear Monitoring with Artificial Intelligence Methods: A Review. Journal of Manufacturing and Materials Processing, 7, Article 129.
https://doi.org/10.3390/jmmp7040129
[5] Mohamed, A., Hassan, M., M’Saoubi, R. and Attia, H. (2022) Tool Condition Monitoring for High-Performance Machining Systems—A Review. Sensors, 22, Article 2206.
https://doi.org/10.3390/s22062206
[6] 王民, 车昌家, 高相胜, 昝涛, 高鹏, 张云飞. 复杂工况条件下刀具磨损状态监测与剩余寿命预测研究进展[J]. 机械工程学报, 2025, 61(19): 299-326.
[7] 钟荣锋. 基于时频融合与故障频率特征提取的智能诊断技术应用[J]. 传感器技术与应用, 2025, 13(6): 951-962.
[8] 杨好, 李军成, 杨磊. 基于小波分析与神经网络的电子设备故障诊断[J]. 仪器与设备, 2024, 12(2): 230-238.
[9] 张嵘, 朱寅, 李帅. 基于机器学习的齿轮箱故障诊断[J]. 建模与仿真, 2023, 12(2): 1415-1427.
[10] 海天恒. 基于二维卷积神经网络的轴承故障智能诊断[J]. 建模与仿真, 2024, 13(2): 1333-1345.
[11] 吴英豪, 姚丹妍, 李蓉, 曹正豪, 肖思凯. 基于多传感器融合与深度学习的工业轴承源域故障诊断[J]. 图像与信号处理, 2026, 15(2): 221-234.
[12] 高程远. 基于融合注意力的卷积神经网络滚动轴承故障诊断模型[J]. 建模与仿真, 2024, 13(5): 5503-5512.
[13] 范慧鹏, 李瑞华, 李福林, 房哲续, 彭六保. 深度学习算法在工业设备故障诊断应用研究[J]. 传感器技术与应用, 2021, 9(4): 193-200.
[14] 胡惠媛, 赵祖玲. 基于深度学习的电气自动化设备故障诊断与预测技术研究[J]. 传感器技术与应用, 2025, 13(3): 447-458.
[15] 成佳闻, 赛希亚拉图, 张超勇, 罗敏. 基于注意力机制堆叠LSTM的多传感器信息融合刀具磨损预测[J]. 工业工程, 2024, 27(3): 64-77, 86.
[16] Roberts, D.R., Bahn, V., Ciuti, S., Boyce, M.S., Elith, J., Guillera‐Arroita, G., et al. (2017) Cross‐Validation Strategies for Data with Temporal, Spatial, Hierarchical, or Phylogenetic Structure. Ecography, 40, 913-929.
https://doi.org/10.1111/ecog.02881
[17] Varma, S. and Simon, R. (2006) Bias in Error Estimation When Using Cross-Validation for Model Selection. BMC Bioinformatics, 7, Article No. 91.
https://doi.org/10.1186/1471-2105-7-91
[18] Kaggle (2026) CNC Mill Tool Wear.
https://www.kaggle.com/datasets/shasun/tool-wear-detection-in-cnc-mill
[19] Lundberg, S.M. and Lee, S.-I. (2017) A Unified Approach to Interpreting Model Predictions. Proceedings of the 31st International Conference on Neural Information Processing Systems, Long Beach, 4-9 December 2017, 4768-4777.
[20] Lundberg, S.M., Erion, G., Chen, H., DeGrave, A., Prutkin, J.M., Nair, B., et al. (2020) From Local Explanations to Global Understanding with Explainable AI for Trees. Nature Machine Intelligence, 2, 56-67.
https://doi.org/10.1038/s42256-019-0138-9
[21] Brodersen, K.H., Ong, C.S., Stephan, K.E. and Buhmann, J.M. (2010) The Balanced Accuracy and Its Posterior Distribution. 2010 20th International Conference on Pattern Recognition, Istanbul, 23-26 August 2010, 3121-3124.
https://doi.org/10.1109/icpr.2010.764
[22] 李思崎, 关博. 基于K均值算法的设备故障分析与应对策略[J]. 应用数学进展, 2023, 12(12): 5050-5056.
[23] 李涛, 陈丽, 高松. 基于组合分类器的MSF发动机综合故障诊断研究[J]. 运筹与模糊学, 2019, 9(2): 170-176.
[24] 王强, 许超, 许春梅. 电力工程中电力设备状态监测与故障诊断技术的最新进展[J]. 传感器技术与应用, 2025, 13(4): 615-631.
[25] 许志鹏, 陈虎. 基于CNN-SVM融合算法的矿用充电机功率模块早期故障识别[J]. 建模与仿真, 2026, 15(8): 51-60.
[26] 尹红珊, 庞进, 李静, 刘安, 王景初, 刘孟奇, 许伟民. 基于机器学习集成模型的配电网预测性维护方法研究[J]. 智能电网, 2026, 16(4): 113-121.
[27] 李嘉钰. AI技术工业生产中的应用研究[J]. 软件工程与应用, 2020, 9(4): 288-293.