Abstract: Accurate and timely diagnosis of rotating machinery faults can avoid economic losses and even casualties. Based on this, a SPGAP-ResLSTMnet fault diagnosis method for rotating machinery is proposed in this paper. The method combined multi-manifold LSTM elements, which can remove repeated data in time and save the required data. The combined residual structure can effectively alleviate the gradient dispersion problem caused by the increase of network layers, so as to fully extract fault features. GAP and ELM are used to achieve efficient and accurate fault classification. Case Western Reserve University laboratory data set is selected to complete the comparative experiment between the proposed method and the methods3,6. The results show that the proposed method has higher recognition accuracy for normal signals, rolling body,inner and outer ring fault signals under various loads and different pits, as well as noisy signals in comparison with methods3,6. In addition, higher accuracy and lower loss values can be achieved with fewer training sessions by our method.