An Improved IHBA-BP Neural Network for Temperature Compensation of Load Cells

Temperature variations degrade load cell accuracy. To address this problem, an Improved Honey Badger Algorithm (IHBA) was developed to optimize the weights and biases of a BP neural network. IHBA incorporates uniform initialization, a nonlinear weight factor, and Lévy flight with stagnation-aware triggering to overcome the uneven initialization, poor exploration–exploitation balance, and weak local optima escape capability of the standard HBA. To validate the proposed method, a dedicated calibration experimental system was constructed. A 7075 T6 aluminum load cell (50 kN) was tested under 0–50 kN loading–unloading cycles over a temperature range of 0–60 °C. To eliminate random errors, three identical elastomers were fabricated, each tested three times, and the measured values were averaged. The results show that after IHBA-BP compensation, the zero-temperature drift coefficient of the load cell was reduced from 374.8 ppm/°C to 35.09 ppm/°C, and the sensitivity-temperature coefficient was reduced from 936.94 ppm/°C to 45.75 ppm/°C. On the unseen test set, the relative error after compensation was 0.01207, the mean square error was 2.84 × 10−5, and the root mean square error was 0.00533. Compared with IMA-BP, PSO-BP, BP, and polynomial fitting methods, IHBA-BP achieved the lowest error. The proposed method shows strong potential for industrial load cell temperature compensation.