Abstract: Accurate monitoring of tool condition is crucial for improving machining quality and efficiency.In the current widely used indirect methods for tool wear monitoring,the single-step or short-term predictions are predominant,without achieving multi-step prediction and suffering from significant cumulative errors.Gaussian process is a machine learning method commonly applied in indirect methods.However,traditional Gaussian process regression has limited accuracy in tool wear prediction due to model structure and algorithm constraints.To address these shortcomings,an improved autoregressive recursive Gaussian process model was proposed for multi-step prediction of tool wear.To reduce cumulative prediction errors,the improved model updating methods and the composite kernel functions were applied to set forgetting factor for samples during model training.Additionally,a bias correction method was incorporated in the prediction process.The effects of each improvement factor on the model were studied,and the accurate multi-step prediction of tool wear state was achieved by combining all favorable factors.The prediction errors reduced by 85.68%,20.67% and 63.32% on three test sets respectively.

Key words: tool condition monitoring, multi-step prediction, Gaussian process, recurrent

摘要： 刀具状态的准确监测对于提高切削加工质量和加工效率至关重要。在当前广泛用于刀具磨损状态监测的间接法中,多以单步或短期预测为主,没有实现多步预测,且累积误差较大。高斯过程是间接法中应用较多的一种机器学习方法,然而传统的高斯过程回归由于模型结构和算法的限制,对刀具磨损预测的精度不高。针对上述不足,提出了改进的自回归递归高斯过程模型对刀具磨损进行多步预测。为了减小预测累积误差,在模型训练中应用了改进的模型更新方式、组合核函数,对样本设置了遗忘因子,在预测中加入了偏差校正方法。研究了各个改进因素对模型的影响并综合所有有利因素,实现了较准确的刀具磨损状态多步预测,在3个测试集上预测误差分别降低了85.68%,20.67%和63.32%。

关键词: 刀具状态监测, 多步预测, 高斯过程, 递归