In this study， a novel semi-supervised fault diagnosis of planetary gearboxes based on manifold regularized support high er-order tensor machines （MRSHTM） is proposed. In the MRSHTM， CANDECOMP/PARAFAC （CP） decomposition is intro duced to exploit the intrinsic structural information of tensor data， and tensor-based inverse multiquadric kernel function （Tensor IMKF） is defined to construct a Laplacian operator. The constructed graph matrix can better describe the manifold structure be tween tensor data. Besides， the one-versus-rest （OVR） strategy is introduced into the MRSHTM model for multi-class fault diag nosis of planetary gearboxes. Hierarchical multiscale permutation entropy （HMPE） is adopted to extract the three-order tensor fea tures “channel×hierarchical layer×scale”， and then the extracted HMPE values are fed into OVR-MRSHTM for automatic fault identification. The results suggest that the proposed method can achieve semi-supervised fault diagnosis of planetary gearboxes in tensor space.