Deep learning-based health indicator construction has become a new research and application hotspot in the field of machinery fault diagnostics. The performance of deep learning-based health indicators is largely depending on hand-craft feature extraction and selection. Moreover, the correlation of multi-channel sensor signals is not enough considered. In response to the above problems, a method for constructing health indicators based on multi-channel information fusion based on Deep Fusion Neural Network (DFNN) is designed. First, a multi-channel feature extractor (MFE) is proposed to extract bearing degradation features from the raw vibration signals. Then an adaptive feature selector (AFS) is designed to select useful features automatically. After MFE and AFS, we utilized a bidirectional long-short-term memory (BiLSTM) network to construct bearing health indicator. The proposed method is experimentally verified on the bearing life data set. The result shows that compared with some state-of-the art methods, the health indicator by DFNN is up to 98. 4%, and the monotonic indicator increases by 44%. Therefore, it is able to map the bearing degradation process effectively.