Due to the complicacy of mechanical instruments and the noise interference in the working environment, the equipment status information contained in a single sensor is insufficient, and multi-source information contains more complete status information. In order to effectively fuse multi-sensor information and improve the reliability of diagnosis, a multi-level fusion dual convolution neural network (MFDCNN) for fault diagnosis of rotating machinery is proposed in this paper. This approach realizes multi-level fusion of fault information by utilizing the flexibility of the structure of the convolutional neural network. During the training process, the two subnets automatically extract representative features from the multi-sensor timedomain signal and its frequency spectrum in parallel, and then fuse the extracted features for pattern recognition to achieve end-to-end fault diagnosis. Compared with the single sensor diagnosis method and single level information fusion method, this approach has better diagnosis performance.

由于机械仪表的复杂性和工作环境中的噪声干扰，单个传感器包含的设备状态信息不足，多源信息包含更完整的状态信息。为了有效融合多传感器信息，提高诊断的可靠性，本文提出了一种用于旋转机械故障诊断的多级融合双卷积神经网络（MFDCNN）。该方法利用卷积神经网络结构的灵活性，实现了故障信息的多级融合。在训练过程中，两个子网自动并行地从多传感器时域信号及其频谱中提取代表性特征，然后融合提取的特征进行模式识别，实现端到端的故障诊断。与单传感器诊断方法和单层信息融合方法相比，该方法具有更好的诊断性能。