User identification through smartphones and wearable sensors holds promise but faces challenges from similarity and variability in user activities. Visualization of smartphone acceleration signals revealed users’ signals exhibit high similarity, as activities share a common underlying structure. For example, walking elicits a repeated general pattern. Therefore, user identification relies on subtle distinguishing factors in fine activity details. At times, patterns are near-indistinguishable between users. To address this, we developed a method leveraging the assumption that prediction uncertainty increases for nonseparable samples. The input data is divided into subsequences, each independently predicted by a convolutional neural network. Predictions are fused through a weighted averaging scheme, where weights quantify prediction uncertainty using the Monte Carlo dropout method. Through experiments on five real-world data sets, the study demonstrates improved performance in identifying users across a range of activities compared to existing methods. It was also directly compared to state-of-the-art methods using two well-known data sets, improving accuracy by 1.29% in one case and 7.98% in the other. These findings validate the effectiveness of the new approach for continuous user identification, even when faced with unpredictable user behavior.