Abstract: For industrial data acquisition,there are many problems such as large amount of data and complex interference signals,which lead to complex identification methods and low accuracy of tool wear state.A recognition method based on Ensemble Empirical Mode Decomposition Fast Kurtogram (EEMDFK)component selection and Attention Mechanism Convolutional Neural Network (ACNN)was proposed.Aiming at the difficulty in selecting mode components of ensemble empirical mode,a fast spectrum kurtosis graph selection strategy was used to select the optimal component.The tool vibration signals collected under different conditions were decomposed into fault signal features by integrating empirical mode decomposition.Then,the Intrinsic Mode Function (IMF)was selected by the fast spectral kurtosis graph selection strategy,and the time-frequency analysis of Hilbert Huang Transform (HHT)was carried out to generate time-frequency graphs.The time-frequency diagram was input into the recognition model designed for learning,and the efficiency of feature extraction was improved through the attention mechanism.The wear state of the saved model was recognized in the test set,and the recognition results were output,so as to realize the recognition of different tool wear state.Experimental results showed that the recognition rate of the proposed method could reach 99.7% under different tool states,and the method could realize intelligent recognition under different tool wear states,and had good generalization ability and robustness.

Key words: tool wear, ensemble empirical mode decomposition, fast-Kurtogram, deep learning, attention mechanism

摘要： 针对加工数据采集存在数据量大且干扰信号复杂,导致刀具磨损状态识别方式复杂、识别精度低等问题,提出一种基于快速谱峭度图的集合经验模态分量选取(EEMDFK)与注意力机制的卷积神经网络(ACNN)相结合的识别方法。针对集合经验模态存在选取模态分量困难的情况,引用快速谱峭度图选择策略选取最优分量。通过集合经验模式分解从所采集的不同状况下的刀具振动信号分解出故障信号特征;通过快速谱峭度图选择策略选取内在模函数并进行HHT时频分析,生成时频图;将时频图输入所设计的识别模型进行学习,通过注意力机制提高特征提取效率,并使保存的模型在测试集中对不同刀具磨损状态进行了识别。实验结果表明,该方法对刀具不同状态下的识别率可达99.7%,实现了不同磨损状态下刀具的智能识别,并具有较好的泛化能力和鲁棒性。

关键词: 刀具磨损, 集合经验模式分解, 快速谱峭度, 深度学习, 注意力机制