Abstract: To achieve effective transfer of tool wear prediction values based on neural network methods under multiple working conditions,a method combining Model Agnostic Meta-Learning (MAML) and Multi-Head Attention (MHA) was proposed to realize tool wear prediction under multiple working conditions.The multi-dimensional information in the cutting process of the tool was collected by multiple sensors,and the time-frequency feature matrix was constructed by extracting time-domain,frequency-domain and time-frequency domain features.The MHA model was utilized as the base model,and the time-frequency feature samples constructed from historical working condition information were leveraged to train this model via the MAML approach to acquire the optimal initialization parameters for the MHA model.In the new working condition,the optimal initialization MHA model was iteratively trained with a small number of initial wear samples several times to adapt to the new condition,thereby predicting the tool wear value for the new working conditions.Finally,relevant experiments demonstrated that the proposed method could achieve effective model transfer in tool wear prediction under multiple working conditions.

Key words: tool wear, meta-learning, attention mechanism, multi-sensor

摘要： 为了采用基于神经网络方法预测刀具磨损值时在多工况条件下进行有效迁移,提出一种结合模型无关元学习（MAML）与多头注意力（MHA）模型的方法,预测多工况下刀具的磨损值。首先通过多传感器收集刀具切削过程中的多维信息,从中提取时域频域和时频域的特征信息,以构建时频特征矩阵;其次将MHA模型作为基模型,利用历史工况信息构建的时频特征样本,通过MAML方法训练该模型,获得最优的MHA模型初始化参数;然后最优的初始化MHA模型通过少量的初始磨损样本迭代训练几次来适应新工况,从而预测新工况的刀具磨损值;最后通过相关实验验证,该方法在多工况下的刀具磨损值预测中能够有效进行模型迁移。

关键词: 刀具磨损, 元学习, 注意力机制, 多传感器