A cross-task brain-controlled robot control method, system, device and medium based on deep transfer learning
By decomposing and training an asymmetric directed transfer weight matrix based on cortical body localization maps and a two-stream encoder, the problem of decoding failure in cross-task brain-controlled robot operation was solved, achieving stable decoding accuracy and user-specific amplitude correction across users, and improving the decoding accuracy of cross-task transfer.
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- NANJING XINYING SMART INFORMATION TECHNOLOGY CO LTD
- Filing Date
- 2026-06-05
- Publication Date
- 2026-07-17
AI Technical Summary
Existing deep transfer learning methods cannot effectively distinguish between the anatomical topology of joint motion control areas and user-specific temporal difference information in cross-task brain-controlled robot operation, leading to decoding failure and requiring a large amount of target user data to retrain the overall model.
By determining the effective set of electrode pairs based on the cortical somatic localization map, an asymmetric directed transfer weight matrix is constructed, and the event-related desynchronization temporal features are decomposed into sequence components and amplitude components. Then, a two-stream encoder is used for directed asymmetric constraint training to achieve separate modeling of cross-user stable joint activation temporal sequence information and user-specific temporal difference amplitude information.
It achieves the goal of maintaining initial decoding accuracy without requiring a large amount of target user data, and implements targeted correction without destroying user-specific amplitude representation during online operation, thereby improving the decoding accuracy and stability of cross-task migration.
Smart Images
Figure CN122401425A_ABST