基于整周期数据和卷积神经网络的谐波减速器健康状态评估

Health condition assessment of harmonic reducer based on integer-period data and convolutional neural network

针对工业机器人谐波减速器循环往复运动、工作节拍不一和转速瞬变而导致其运行状态难以刻画和健康状态不易评估的问题,提出了基于整周期数据和卷积神经网络的谐波减速器健康状态评估方法。首先,运用相位差频谱校正—互相关法对振动信号分割构造整周期数据样本以准确刻画谐波减速器的运行状态信息;其次,应用连续小波变换对整周期数据进行分解以充分展现谐波减速器运转周期内的瞬变特征;最后,利用卷积神经网络在时间和空间上对输入信号的平移、缩放具有高度不变性的特点,充分学习谐波减速器运转周期内的瞬变特征,从而实现对谐波减速器健康状态评估。实验结果显示,所提方法识别准确率达到了90%以上,证明了该方法能够有效评估谐波减速器健康状态,并具有较好的泛化能力和稳健性。

Due to the cyclic motion, different working beats and transient speed, it is difficult to effectively describe the running state and assess the health state of industrial robot harmonic reducer. In this study, a method based on integer-period data and convolutional neural network is proposed to achieve health state assessment of harmonic reducer. Firstly, the phase difference spectrum correction-cross correlation method is used to adaptively segment the original vibration signal and construct the integer-period data samples to accurately describe the running state information of the harmonic reducer. Secondly, continuous wavelet transform is applied to decompose the integer-period data sample to obtain the time-frequency map to fully show the transient characteristics of harmonic reducer in the operation cycle. Finally, convolution neural network is utilized to translate and scale the input signals in time and space with high invariability to fully learn the transient characteristics of the harmonic reducer in each operating cycle. In this way, the health state of the harmonic reducer can be evaluated. Experimental results show that the identification accuracy of the proposed method is over 90%. The effectiveness of the proposed method is verified, which has good generalization ability and robustness.