The prediction of tool wear helps minimize costs and enhance product quality in manufacturing. While existing data-driven models using machine learning and deep learning have contributed to the accurate prediction of tool wear, they often lack generality and require substantial training data for high accuracy. In this paper, we propose a new data-driven model that uses Bayesian Regularized Artificial Neural Networks (BRANNs) to precisely predict milling tool wear. BRANNs combine the strengths and leverage the benefits of artificial neural networks (ANNs) and Bayesian regularization, whereby ANNs learn complex patterns and Bayesian regularization handles uncertainty and prevents overfitting, resulting in a more generalized model. We treat both process parameters and monitoring sensor signals as BRANN input parameters. We conducted an extensive experimental study featuring four different experimental data sets, including the NASA Ames milling dataset, the 2010 PHM Data Challenge dataset, the NUAA Ideahouse tool wear dataset, and an in-house performed end-milling of the Ti6Al4V dataset. We inspect the impact of input features, training data size, hidden units, training algorithms, and transfer functions on the performance of the proposed BRANN model and demonstrate that it outperforms existing state-of-the-art models in terms of accuracy and reliability.

翻译：刀具磨损的预测有助于降低制造成本并提升产品质量。尽管现有基于机器学习和深度学习的数据驱动模型已能够准确预测刀具磨损，但这些模型通常缺乏泛化性，且需要大量训练数据才能实现高精度。本文提出了一种新的数据驱动模型，采用贝叶斯正则化人工神经网络（BRANNs）精确预测铣削刀具磨损。BRANNs融合了人工神经网络（ANNs）与贝叶斯正则化的优势：ANNs负责学习复杂模式，而贝叶斯正则化则处理不确定性并防止过拟合，从而生成更具泛化能力的模型。我们将工艺参数和监测传感器信号均作为BRANN的输入参数。我们开展了广泛的实验研究，涵盖四组不同的实验数据集，包括NASA Ames铣削数据集、2010年PHM数据挑战赛数据集、NUAA Ideahouse刀具磨损数据集以及自建的Ti6Al4V端铣实验数据集。通过分析输入特征、训练数据规模、隐含层单元数、训练算法及传递函数对BRANN模型性能的影响，我们证明了该模型在精度与可靠性方面均优于现有最先进的模型。