<p>Detecting gait intention based on single-leg information is more challenging than using data from both legs. This paper proposes a user-specific single-leg locomotion mode detection (LMD) method based on a neural network that fuses eight-channel electromyography (EMG) signals and three-dimensional divergent component of motion (DCM). To capture the spatial and temporal correlations in the fused input, a convolutional neural network (CNN)-based classification model was designed. The proposed method achieves reliable detection of three locomotion modes (stair-ascending, stair-descending, and level-walking) even with a small amount of training data. In addition, to improve model robustness, we applied data augmentation and defined exception classes. To quantitatively validate the suitability of the CNN architecture and exception classes, we conducted cross-validation and compared the CNN with multilayer perceptron (MLP) and bidirectional long short-term memory (BiLSTM) models. The CNN achieved higher classification performance, and EMG and DCM fusion demonstrated its effectiveness by improving performance over single-modality inputs. For experimental evaluation, we tested the proposed method under both trained and untrained locomotion environments. Under the trained locomotion environment, the proposed method achieved an LMD success rate of 96.7–97.3% for three users, consistently outperforming the DCM-only and EMG-only methods. Additional single-user evaluations were conducted under untrained locomotion environments, including two-step stair locomotion, speed variations, speed transitions, and obstacle crossing with locomotion mode transitions.</p>