基于融合数据增强的柴油机故障诊断方法

Diesel Engine Fault Diagnosis Method Based on Fusion Data Augmentation

针对基于深度学习的柴油机故障诊断方法在训练样本数量匮乏时易发生过拟合并导致诊断准确率下降的问题,提出一种基于融合数据增强的柴油机故障诊断方法。利用融合数据增强方法扩充训练集中的故障样本:先将合成少数过采样技术(synthetic minority oversampling technology,SMOTE)和改进辅助分类器生成对抗网络(auxiliary classifier generative adversarial network,ACGAN)相结合,分两阶段进行样本生成;再通过K近邻算法(K-nearest neighbors,KNN)去除噪声生成样本。然后使用扩充后的训练集训练深度学习故障诊断模型,用于识别未知振动信号。经柴油机故障模拟试验实测信号验证,在仅用10个样本进行训练时,一维卷积神经网络(one-dimensional convolutional neural network,1DCNN)能达到90.21%的故障诊断准确率,且所提融合数据增强方法在不同深度学习模型上均能提高故障诊断准确率。

Aiming at the problem that the diesel engine fault diagnosis method based on deep learning overfitting occured and led to a decrease in diagnosis accuracy when the number of training samples was scarce,an diesel engine fault diagnosis method based on fusion data augmentation was proposed.The fault samples in the training set were expanded using fusion data augmentation methods.The synthetic minority over-sampling technology(SMOTE)and auxiliary classifier generative adversarial network(ACGAN)were combined to generate samples in two stages.K-nearest neighbors(KNN)was used to remove the noise generating samples.Then the expanded training set was used to train a deep learning fault diagnosis model for recognizing unknown vibration signals.The measured signals of the diesel engine fault simulation experiment show that the one-dimensional convolutional neural network(1DCNN)can achieve a fault diagnosis accuracy of 90.21%with only 10 samples used for model training and the proposed fusion data augmentation method can improve the fault diagnosis accuracies of different deep learning models.