Nowadays, most intelligent diagnosis methods focus on fault classification and the discriminative knowledge is unknown due to the ‘black box’ characteristic. However, impulse responses in vibration signals, which is important sign to determine whether mechanical equipment is faulty, are rarely studied under intelligent methods since their recognition is both difficult and time-consuming, especially mixed with noise. Aiming at these problems, a novel impulse recognition method was proposed to capture them from raw mechanical data. Firstly, a single-kernel convolutional neural network is proposed as weak classifier to learn discriminative information from raw data. Then, a coarse-to-fine search is proposed to locate position of impulse response. Finally, the boosting algorithm is used to ensemble several proposed weak classifiers for final output. Vibration signals of bearings with two different faults are utilized to validate the proposed model. The results prove that the proposed approach obtain higher accuracy compared with traditional Laplace wavelet method. Moreover, the extracted kernel functions reveal new knowledge about characteristics of impulse responses, which significantly differs from traditional hypothesis and sheds a light on improvement of relevant approaches.

如今，大多数智能诊断方法都集中在故障分类上，由于“黑匣子”的特性，区分性知识尚不明确。然而，振动信号中的脉冲响应是确定机械设备是否有故障的重要标志，因此很少采用智能方法进行研究，因为它们的识别既困难又费时，尤其是与噪声混合。针对这些问题，提出了一种新颖的脉冲识别方法，可以从原始机械数据中捕获它们。首先，提出了一种单核卷积神经网络作为弱分类器，以从原始数据中学习判别信息。然后，提出了从粗到精的搜索来定位脉冲响应的位置。最后，使用boosting算法将几个建议的弱分类器集成为最终输出。利用具有两个不同故障的轴承的振动信号来验证所提出的模型。结果证明，与传统的拉普拉斯小波方法相比，该方法具有更高的精度。此外，提取的核函数揭示了关于冲激响应特征的新知识，这与传统的假设有很大的不同，并为相关方法的改进提供了启示。