To address the significant decline in movement intention recognition accuracy caused by the regular physiological feature drift of surface electromyography (sEMG) signals under varying loads in lower limb exoskeleton applications, a recognition method based on muscle physiological characteristics and a bi-branch cross-scale multi-head attention residual network (BCM-ResNet) is proposed. First, the regular pattern of sEMG signal changes under load, which can be termed "physiological reconstruction", is quantitatively revealed through time-frequency analysis. Agglomerative hierarchical clustering confirms that the sEMG feature space exhibits distinct "low, medium, and high" cluster distributions within the 0~14 kg range. This finding demonstrates that load is a key variable driving the regular displacement of the feature space rather than a random disturbance. Secondly, a differentiated feature extraction strategy is proposed by considering the functional heterogeneity of the thigh and calf muscles. A bidirectional long short-term memory (BiLSTM) network is employed to capture the strong temporal dynamics of the thigh muscles (the primary power source), while a multi-scale convolutional neural network (Multi-scale CNN) is utilized to extract local fine-grained features of the calf muscles (the precision regulators). A bi-branch fusion architecture with a multi-head attention (MHA) mechanism is constructed to achieve dynamic decoupling and weighted fusion of intention and load features, utilizing a residual network (ResNet) to enhance the abstraction of high-level features. Additionally, the sensor configuration is optimized through Pareto analysis, determining a 12-channel unilateral layout on the right leg as the optimal setup. Experiments show that the proposed method achieves an average recognition accuracy of 94.78% on a self-built dataset and 92.85% on an international open-source dataset containing 22 subjects, validating the model′s generalization performance. Finally, migration tests on an edge computing terminal verify the algorithm′s execution stability under sub-millisecond inference latency. This research provides reliable algorithmic support for the precise intention recognition of exoskeletons in complex loading environments.