Despite recent progress in decoupling algorithms for multi-axis force sensors, the high cost of high-fidelity (HF) calibration data severely limits dataset size, resulting in poor generalization in complex multi-axis loading scenarios. To address these challenges, a physics-informed transfer learning-based decoupling algorithm is proposed to reduce the dependence on HF multi-axis coupled calibration data. The proposed method comprises two primary parts. First, HF data are obtained from multi-axis coupled calibration experiments, while low-fidelity (LF) data are generated using a physics-based surrogate model derived via the Least Squares (LS) method. The LF dataset captures the first-order coupling characteristics of the sensor and serves as the source domain, whereas the HF dataset represents the target domain. Second, Bayesian Optimization (BO) is employed to identify optimal hyperparameters that ensure the network structure is commensurate with the nonlinear coupling complexity of the sensor. A fully connected neural network is pre-trained on the LF dataset to encode low-order coupling mechanisms and subsequently fine-tuned using the HF dataset to compensate for higher-order nonlinear effects. Compared with LS, Extreme Learning Machine (ELM), and Artificial Neural Network (ANN), the proposed method achieves superior accuracy, reaching an RRMSE of 0.009 with error reductions of 60.9 %, 30.7 % and 40 % under identical calibration data conditions. Moreover, the proposed method substantially reduces the reliance on HF calibration data, achieving accuracy comparable to ELM and ANN trained with approximately 200 and 250 samples using only 100 HF samples.
扫码关注我们
求助内容:
应助结果提醒方式:
