This paper describes two algorithms for feature extraction from the Poincaré plot which is constructed with the vibration signals measured in roller bearings and gearboxes. The extracted features are used for classifying 10 types of fault conditions in a gearbox and 7 types of fault conditions a roller bearings. Both vibration signal datasets were acquired at different loads and speeds. The feature extraction using Algorithm 1 performs the feature calculation from the Poincaré plot constructed with the raw vibration signals. In contrast, the Algorithm 2 requires an additional stage where the vibration signal is pre-processed for identifying the peaks of the signal. This peak sequence is equivalent to a non-uniform sub-sampling of the vibration signal that retains relevant information useful for fault classification. The fault classification is attained by using a multi-class Support Vector Machine. The proposed method is tested using the tenfold cross-validation. Results show that both algorithms could attain classification accuracies as high as 99.3% for the gearbox dataset and 100% for the roller bearings. The results are compared to other classification approaches performed on the same datasets by using other different features. The comparison shows that the approach in this paper has a performance as good as obtained by using well-known statistical features.

本文介绍了两种从庞加莱图中提取特征的算法，该图是根据在滚动轴承和齿轮箱中测得的振动信号构建的。提取的特征用于对变速箱中的10种故障条件进行分类，对滚动轴承进行7种故障条件分类。两种振动信号数据集都是在不同的负载和速度下获取的。使用算法1进行特征提取是根据由原始振动信号构成的庞加莱图进行特征计算。相比之下，算法2需要额外的阶段，在该阶段中，对振动信号进行预处理以识别信号的峰值。该峰值序列等效于振动信号的非均匀子采样，该采样保留了可用于故障分类的相关信息。通过使用多类支持向量机可以实现故障分类。使用十倍交叉验证对提出的方法进行了测试。结果表明，这两种算法对于齿轮箱数据集和滚动轴承的分类准确率均高达99.3％。通过使用其他不同功能，将结果与对相同数据集执行的其他分类方法进行比较。比较表明，该方法具有与使用众所周知的统计特征所获得的性能相同的性能。通过使用其他不同功能，将结果与对相同数据集执行的其他分类方法进行比较。比较表明，该方法具有与使用众所周知的统计特征所获得的性能相同的性能。通过使用其他不同功能，将结果与对相同数据集执行的其他分类方法进行比较。比较表明，该方法具有与使用众所周知的统计特征所获得的性能相同的性能。