In complex systems, condition monitoring data consist of multi-source spatiotemporal information collected from multiple sensors. To effectively capture temporal degradation patterns and spatial correlations among measured variables, we propose an adaptive temporal-spatial feature fusion neural network (TSTFNN) based on a dual-stream structure. This framework incorporates parallel temporal and spatial streams to extract temporal dependencies and spatial correlations separately. To overcome the limitations of traditional dot-product self-attention, which often neglects time-series continuity, a convolutional self-attention mechanism is implemented, enhancing the capacity of model to capture sequential continuity and subtle temporal variations. A multiscale convolutional neural network further extracts spatial correlation features across variables, improving global perception capabilities. During feature fusion, an adaptive weighting mechanism enables dynamic integration of temporal and spatial features. To optimize predictive performance, a joint loss function, combining constrained mean squared error (MSE) and feature balance loss, is introduced, facilitating the collaborative learning of temporal and spatial features. Finally, experimental results based on NASA′s C-MAPSS benchmark dataset demonstrate that the proposed method outperforms various state-of-the-art (SOTA) models in terms of multi-source data RUL prediction accuracy.