【摘要】 首先，采用经验模态分解法与小波阈值相结合的去噪方法对采集到的变速箱振动信号进行消噪；其次，基于消噪后的振动信号提取时域、频域及熵特征构成多域原始特征集，并利用Fisher score评价准则进一步从原始特征集中筛选出能够反映健康运行状态的特征子集；最后，建立卷积神经网络模型，并将原始特征集及特征选择处理后的特征子集输入到网络模型中进行分类，考察特征选择对分类准确率的影响。研究表明：所建立的卷积神经网络模型对健康状态的识别准确率能够达到93.55%，较原始特征集直接构建的网络模型识别精度更高，训练速度更快。该研究对维持变速箱的正常运行并抑制机械设备故障的产生具有重大意义。更多还原

【Abstract】 Firstly, the collected transmission vibration signals was de-noised by the de-noising method combining Empirical Mode Decomposition(EMD)and wavelet threshold. Secondly, the time domain, frequency domain and entropy features were extracted to form a multi-domain original feature set based on the de-noised vibration signal, and then the Fisher score evaluation criterion was used to further filter out the feature subset from the original feature set which can reflect the healthy operation state. Finally, a Convolutional Neural Network model(CNN) was built.The original feature set and feature subset processed by feature selection were input into the network model for classification, and investigate the effect of feature selection on the classification accuracy. The research results show that the recognition accuracy of the convolutional neural network model built in the present work can reach 93.55%, which has higher recognition accuracy and faster training speed compared with the network model directly constructed from the original feature set. These research results are of great significant to maintain the normal operation of gearbox and restrain the failure of mechanical equipment.更多还原