The performance of traditional vibration based fault diagnosis methods greatly depends on those handcrafted features extracted using signal processing algorithms,which require significant amounts of domain knowledge and human labor,and do not generalize well to new diagnosis domains.Recently,unsupervised representation leing provides an altative promising solution to feature extraction in traditional fault diagnosis due to its superior leing ability from unlabeled data.Given that vibration signals usually contain multiple temporal structures,this paper proposes a multiscale representation leing (MSRL) framework to le useful features directly from raw vibration signals,with the aim to capture rich and complementary fault patt information at different scales.In our proposed approach,a coarse-grained procedure is first employed to obtain multiple scale signals from an original vibration signal.Then,sparse filtering,a newly developed unsupervised leing algorithm,is applied to automatically le useful features from each scale signal,respectively,and then the leed features at each scale to be concatenated one by one to obtain multiscale representations.Finally,the multiscale representations are fed into a supervised classifier to achieve diagnosis results.Our proposed approach is evaluated using two different case studies:motor bearing and wind turbine gearbox fault diagnosis.Experimental results show that the proposed MSRL approach can take full advantages of the availability of unlabeled data to le discriminative features and achieved better performance with higher accuracy and stability compared to the traditional approaches.