基于改进判别字典学习的故障诊断方法

近年来，基于稀疏表示的分类技术在模式识别中取得一定的成功。该框架中，字典的学习和分类器的训练通常是两个独立的模块，降低了方法的识别精度。针对以上问题，提出了一种特征提取和模式识别相融合的改进判别字典学习模型，将重构误差项、稀疏编码判别项及分类误差项进行了整合，并用K奇异值分解算法对目标函数进行优化，实现了字典和分类器的同步学习。该方法先对原始信号进行经验模态分解，并从分解的本征模态函数中提取时、频特征，形成故障样本；然后将训练样本输入改进模型用K奇异值分解优化；最后用习得字典及分类器权重对测试样本进行识别。实验结果表明：该算法不但适用于小样本故障问题，而且鲁棒性和分类性能都明显高于其它算法。    

Fault diagnosis method based on improved discriminative dictionary learning

In recent years, sparse representation based classification method has been successfully employed in pattern recognition. The learning of dictionary and the training of classifier in this method are usually independent in existing approaches, so it reduces the identification accuracy. In this paper, we propose a novel fault diagnosis method based on improved dictionary learning model, which integrates discriminative sparse coding error and classification performance criterion with the reconstruction error. And this model is solved by K-singular value decomposition (K-SVD) algorithm that realizes the synchronization learning of dictionary and classifier. For our method, original signal is decomposed firstly by empirical mode decomposition, and the features of time domain and frequency domain are extracted from the decomposed intrinsic mode functions; then training samples are input into the improved model that is optimized by K-SVD; finally testing samples are identified by using learned dictionary and classification weights. Experimental results show that the algorithm not only can be applied in the small sample faults diagnosis, and the robustness and classification performance are significantly higher than other algorithms.