Intelligent fault diagnostic systems are fast becoming key instruments in industrial applications. This paper presents a recognition system for diagnosing bearing defects in induction motors. The proposed scheme is comprised of five steps, namely signal segmentation, feature extraction and reduction, fault classification and the decision fusion. First, the vibration signal is segmented into successive equal-length intervals, which are considered as patterns in a recognition problem. The objective is to predict the defect mode (class) for each pattern. Then, the time- and the frequency-domain features are extracted from each interval. At the next step, a small set of distinctive and informative features is found by resorting to different feature reduction techniques to guarantee well-organized learning and immediate and accurate classification. Then, in the fourth step, various classifiers are trained to learn to distinguish between the faulty and healthy states. To make the final decision, different combinations of classifiers are considered using the voting and stacking techniques to enhance the overall performance of the recognition system. Evaluation of the proposed diagnostic scheme on the standard CWRU bearing defect database demonstrates that this system attains reasonable performance measures, validating the ideas put forward in this paper.

中文翻译：

---

感应电机轴承缺陷诊断决策融合方案

智能故障诊断系统正迅速成为工业应用中的关键工具。本文提出了一种用于诊断感应电机轴承缺陷的识别系统。所提出的方案由五个步骤组成，即信号分割、特征提取和减少、故障分类和决策融合。首先，振动信号被分割成连续的等长区间，这些区间被认为是识别问题中的模式。目标是预测每个模式的缺陷模式（类别）。然后，从每个间隔中提取时域和频域特征。在下一步中，通过采用不同的特征减少技术来找到一小组独特且信息丰富的特征，以保证组织良好的学习和即时准确的分类。然后，在第四步中，训练各种分类器以学习区分故障状态和健康状态。为了做出最终决定，使用投票和堆叠技术来考虑分类器的不同组合，以提高识别系统的整体性能。对标准 CWRU 轴承缺陷数据库所提出的诊断方案的评估表明，该系统达到了合理的性能指标，验证了本文提出的想法。