Deep learning is characterized by strong self-learning and fault classification ability without manually feature extraction stage of traditional algorithms. Deep belief network (DBN) is one of the most classic models of deep learning. However, traditional DBN is mainly restricted to learn automatically single scale features from raw vibration signal while identify the fault type, which implies some important information inherent in other scales of vibration data are neglected, thus causing easily unsatisfactory diagnosis result. To alleviate the problem, this paper presents a novel architecture named multiscale cascading deep belief network (MCDBN) for automatic fault identification of rotating machinery, which is aimed at learning the broader feature representation and improving the recognition precision. Firstly, a sliding window with data overlap is adopted to split the collected raw vibration signal to a group of equal-sized sub-signal, and then the improved multiscale coarse-grained procedure of each sub-signal is conducted to obtain the coarse-grained time series at different scales. Meanwhile, Fourier spectrum at different scales is calculated to capture multiscale characteristics. Finally, multiple DBN architecture with three hidden layers are designed to learn high-level feature representation directly from multiscale characteristics in a parallel manner and accomplish fault identification automatically through cascading way and softmax classifier without artificial expertise. Results of two experimental cases with respect to mechanical fault identification under different working conditions have well indicated that the proposed method is provided with preferable diagnostic performance compared with standard DBN and traditional multiscale feature extractors.

深度学习的特点是强大的自学习和故障分类能力，而无需传统算法的手动特征提取阶段。深度信仰网络（DBN）是深度学习的最经典模型之一。然而，传统的DBN主要局限于在识别故障类型的同时从原始振动信号中自动学习单尺度特征，这意味着忽略了其他尺度的振动数据中固有的一些重要信息，从而容易导致无法令人满意的诊断结果。为了缓解这一问题，本文提出了一种新颖的架构，用于旋转机械故障自动识别的多尺度级联深度置信网络（MCDBN），旨在学习更广泛的特征表示并提高识别精度。首先，采用数据重叠的滑动窗口，将采集到的原始振动信号分解为一组大小相等的子信号，然后对每个子信号进行改进的多尺度粗粒度处理，得到粗粒度时间序列。在不同的规模。同时，计算不同尺度的傅立叶光谱以捕获多尺度特征。最后，设计了具有三个隐藏层的多个DBN架构，以并行方式直接从多尺度特征中学习高级特征表示，并通过级联方式和softmax分类器自动完成故障识别，而无需人工专业知识。