基于注意力机制与深度学习算法的机床主轴系统故障辨识

Fault Identification of Machine Tool Spindle System Based on Attention Mechanism and Deep Learning Algorithm

针对具有复杂非线性特点的数控机床主轴系统整体动态退化故障较难辨识及故障研究难度大的问题,从数据分析入手,提出一种基于注意力机制与深度学习算法的智能化故障辨识方法,研究机床主轴系统的整体故障辨识问题。该方法设计了注意力机制的研究框架,将研究问题分为全局纵向大分类区间和局部横向细粒度区间两个维度:采用训练并调优后推理平均绝对误差达到0.028 7的门控循环单元模型,辨识出大分类区间的全局性退化故障;采用鲁棒性强且辨识准确率达99.7%的残差网络模型,在sym8小波基自适应软阈值降噪的基础上对局部细粒度区间故障进行准确细节辨识。结果表明:该方法可量化地辨识出主轴系统的整体故障;所提注意力机制可使大分类区间无法准确辨识的故障在细粒度区间得到有效区分,类内数据增长梯度由6.6%增加到43.8%;通过对机床主轴系统实际使用中在空载状态下遇到的不对中和局部共振等典型故障,以及在负载加工状态下故障的辨识研究,验证了所提方法的有效性与准确性。

The overall dynamic degradation fault of the numerically-controlled machine tool spindle system with complex nonlinear characteristics is difficultly identified and investigated.An intelligent fault identification method based on attention mechanism and depth learning algorithm is proposed to stud the overall fault identification of spindle system by starting with data analysis.The proposed method is used to design the research framework of attention mechanism,and divide the research problems into global vertical large classification interval dimension and local horizontal fine-grained interval dimension.The gated recurrent unit model with reasoning average absolute error of 0.028 7 after training and tuning is used to identify the global degradation faults in large classification interval.The residual network model with strong robustness and identification accuracy of 99.7% is used to accurately identify the local fine-grained interval faults, based on sym8 wavelet basis adaptive soft threshold noise reduction.The results show that the proposed method is used to quantitatively identify the overall fault of spindle system. The proposed attention mechanism is used to effectively distinguish the faults that cannot be accurately identified in the large classification interval in the fine-grained interval,and the data growth gradient in the category increases from 6.6% to 43.8%. The effectiveness and accuracy of the proposed method are verified by studying the typical faults,such as misalignment and local resonance encountered in the actual use of the machine tool spindle system under no-load,and the fault identification under loading.