中文摘要：

特征提取是故障智能诊断的关键步骤,然而不同的特征提取方法所得到的特征不同,导致诊断结果也可能有所差异,增加了人工特征选择的难度和不确定性。深度信念网络（Deep Belief Network,DBN）是一种典型的深度学习（Deep Learning）方法,可以通过组合低层特征形成更加抽象的高层表示,发现数据的分布式特征。DBN可直接从低层原始信号出发,通过逐层智能学习得到更好的特征表示,避免特征提取与选择的人工操作,增强识别过程的智能性。将DBN直接应用于轴承振动原始信号的处理,实现轴承故障的分类识别。试验结果表明,DBN可以直接通过原始数据对轴承故障进行分类识别,优先调节时间复杂度偏导数较大的参数,可有效控制DBN的计算成本。

英文摘要：

Feature extraction plays an important role in machine fault intelligent diagnosis. However, different methods, which extract different features from the raw signal, may lead to different diagnostic results. It increases the difficulty in feature selection as well as the diagnosis uncertainty. Deep belief network is one of the typical deep learning methods, which can obtain an abstract representation through the combination of low-level features to discover the data structure distribution. By directly learning from the raw signal and layer-wise training, the deep belief network can obtain a better feature representation, eliminate manual influence on feature extraction, and enhance the intelligence of fault diagnostic process. To verify the effectiveness of DBN, bearing fault experiments were conducted. The experimental results demonstrate that DBN can classify different faults directly with the raw data, and the computational cost of DBN can be depressed by tuning the parameters sensitive to the time complexity.