Abstract: Due to the interference of noise, equipment eccentricity and other factors, it is hard to accurately extract the steady-type fault parameters of fixed-shaft gears. Additionally, the features extracted by intelligent diagnosis methods are often abstract and difficult to explain. To solve this problem, firstly, an interpretable sparse representation Auto-Encoder network is designed based on the steady-type fault response mechanism and sparse representation theory. The encoding and decoding layers of Auto-Encoder network are equivalent to the solution of the sparse vector and learning of the over complete dictionary respectively. Based on the characteristics of steady-type fault signal parameters, an adaptive optimization algorithm is then designed to realize the fast-global optimization of characteristic parameters. Combining the designed sparse representation Auto-Encoder network and the steady-type fault signal features of fixed-shaft gears, a deep neural network is built to achieve a high-precision feature reconstruction of steady-type fault signals. Finally, the simulation shows that the proposed method can directly extract steady-type fault feature parameters with a clear physical meaning, and has high feature extraction accuracy and good anti-noise performance, which further verifies the effectiveness of the proposed method.