In order to solve the problems of poor feedback accuracy of channel state information (CSI) and inability to adaptively adjust multi-scale features in massive multiple input multiple output (MIMO) system under frequency division duplex (FDD), an adaptive multi feature fusion CSI feedback algorithm for massive MIMO systems was proposed. Firstly, the CSI in the space-frequency domain was moved to the angular-delay domain by using the discrete Fourier transform (DFT), and truncated to perform preliminary compression on the CSI; Then, based on the principle of autoencoder, a CSI feedback network consisting of encoder and decoder was constructed, and a selective kernel network was used to allocate different weights to CSI features of different scales, adaptively adjusting the CSI features; Finally, simulation tests were conducted under the COST 2100 channel model to compare and analyze the proposed algorithm with four CSI intelligent feedback algorithms. The results show that compared to the benchmark algorithm CsiNet, the proposed algorithm has a normalized mean square error (NMSE) improvement of 1.7~9.3 dB and 0.55~2.64 dB under indoor and outdoor conditions, respectively; Compared to the three CSI feedback algorithms with simple fusion of multiple features, the proposed algorithm can better adapt to changes in compression rate and environment. Under the condition of high compression loss in indoor 1/64 compression rate, NMSE also has an improvement of more than 1 dB. The algorithm proposed in this paper introduces a selective kernel network into the autoencoder to realize the adaptive adjustment of multi-scale features and provides a new way of CSI feedback for massive MIMO systems.