Precision compensation technology of industrial robot based on OOA-DNN and transfer learning

Addressing the issue of increased positioning errors in industrial robots during grasping and handling tasks, we propose an accuracy compensation technique optimized through the Osprey Optimization Algorithm (OOA) − Deep Neural Networks (DNN). This method leverages the efficient equilibrium between global and local searches characteristic of OOA to optimize the initial weights and biases of the DNN model, thereby enhancing training efficiency and model performance. Experimental results demonstrate significant improvement: based on a model trained with data points under a 60 kg load, the average, maximum, and standard deviation of positioning errors were reduced by 92.1%, 91.5%, and 91.0% respectively, outperforming DNN, Particle Swarm Optimized DNN, and Extreme Learning Machine methods. Furthermore, addressing varying degrees of positioning errors caused by different mass loads, we introduce an accuracy compensation method through transfer learning based on the OOA-DNN algorithm model. Experimental results reveal that adopting a freeze-thaw training strategy for transfer parameters achieves the best accuracy compensation effect; under a 120 kg load with 150 training data points, the mean error, maximum error, and standard deviation were reduced by 91.9%, 89.9%, and 88.4%, respectively. These improvements suggest that our method, requiring significantly fewer data, is especially valuable in data-constrained application scenarios.