In order to improve the accuracy and efficiency of bearing condition monitoring,a new framework based on convolution neural network (CNN) and GPU computation is proposed.Vibration signals of beatings are first acquired to obtain the condition information,and then the wavelet coefficients of vibration sig nals are calculated through continuous wavelet transform to extract the time frequency domain features.At last,the wavelet features are input into the CNN to learn high level features,and convolution and pooling operations enable the learned features to be rotational and size invariance.Additionally,the classification function in the top layer of CNN recognizes the condition of bearings.Moreover,the Computer Unified Device Architecture (CUDA) based CPU + GPU heterogeneous computing framework accelerates the computation efficiency.To demonstrate the effectiveness of the proposed method,a run-to -fail dataset of bearings are used.The comparison experiment between CPU + GPU based method and CPU based method are conducted to highlight the computation effectiveness.The results show that the computational efficiency of the proposed method is five times than the CPU based method.%为提高轴承状态监测的准确性和实时性,研究了基于卷积神经网络和GPU运算的轴承状态识别模型.利用振动信号监测轴承性能状态,应用连续小波变换算法对振动信号进行时频变换得到小波系数云图,通过基于卷积神经网络的深度学习方法进行数据驱动的特征学习,卷积和子采样计算提取具有旋转和尺寸不变性的特征向量,最后全连接层对特征向量进行状态识别.采用基于CUDA(Computer Unified Device Architecture)框架的CPU+ GPU异构并行运算对计算模型加速,提高系统的实时性.为验证提出算法的有效性,采集轴承全寿命周期振动信号,运用提出的CPU+GPU计算方法和CPU计算方法分别对轴承运行状态进行识别实验.实验结果表明,所提出的方法,计算速度是CPU计算速度的5倍以上.
展开▼