 Cognitive workload recognition is essential for maintaining operator safety and avoiding accidents in human-robot interactions. However, existing models are limited to a single task, making it difficult to recognize workloads across multiple tasks. Additionally, when applying these models to new conditions, the differences in EEG signals between tasks can limit their generalizability. We propose a task transfer framework to address this issue by adapting existing models from one task to another. This approach outperformed other methods by up to 8%, demonstrating its effectiveness in recognizing workload across tasks. This article was authored by Yuying Zhou, Ziming Su, Yifan Niu, and others.