An end-to-end fault diagnostics method based on convolutional neural network for rotating machinery withmultiple case studies

The fault diagnostics of rotating components are crucial for most mechanical systems since the rotating components faults.are the main form of failures of many mechanical systems. In traditional diagnostics approaches, extracting features from raw.input is an important prerequisite and normally requiresmanual extraction based on signal processing techniques. This suffers.of some drawbacks such as the strong dependence on domain expertise, the high sensitivity to different mechanical systems,.the poor flexibility and generalization ability, and the limitations of mining new features, etc. In this paper, we proposed an.end-to-end fault diagnostics model based on a convolutional neural network for rotating machinery using vibration signals..The model learns features directly from the one-dimensional raw vibration signals without any manual feature extraction..To fully validate its effectiveness and robustness, the proposed model is tested on four datasets, including two public ones.and two datasets of our own, covering the applications of ball screw, bearing and gearbox. The method of manual, signal.processing based feature extraction combined with a classifier is also explored for comparison. The results show that the.manually extracted features are sensitive to the various applications, thus needing fine-tuning, while the proposed framework.has a good robustness for rotating machinery fault diagnostics with high accuracies for all the four applications, without any.application-specific manual fine-tuning.

旋转部件的故障诊断对于大多数机械系统至关重要，因为旋转部件故障是许多机械系统故障的主要形式。在传统诊断方法中，从原始输入中提取特征是一个重要前提，通常需要基于信号处理技术进行手动提取。这存在一些缺陷，例如对领域专业知识的强烈依赖、对不同机械系统的高度敏感性、灵活性和泛化能力差以及挖掘新特征的局限性等。在本文中，我们提出了一种基于卷积神经网络的旋转机械端到端故障诊断模型，该模型使用振动信号。该模型直接从一维原始振动信号中学习特征，无需任何手动特征提取。为了充分验证其有效性和鲁棒性，所提出的模型在四个数据集上进行了测试，包括两个公开数据集和我们自己的两个数据集，涵盖了滚珠丝杠、轴承和变速箱的应用。我们还探索了基于手动信号处理的特征提取与分类器相结合的方法用于比较。结果表明，手动提取的特征对各种应用较为敏感，因此需要微调，而所提出的框架对于旋转机械故障诊断具有良好的鲁棒性，在所有四个应用中都具有较高的准确率，无需任何针对特定应用的手动微调。