Thermal error prediction model for long-term operating of machine tool using transfer learning techniques

Thermal errors of a feed-drive system could influence machining inaccuracies significantly. Furthermore, the prediction model’s accuracy may decrease over time due to changes in thermal characteristics after long-term operation of the machine tool. This paper employs transfer learning methods to refine a backpropagation neural network (BPNN) prediction model, enabling it to effectively adapt to data collected 1 year later. First, linear encoder and rotary encoder is proposed to measure ball-screw thermal expansion, which accelerates measurements and enhances model accuracy by minimizing the influence of heat transfer. Validation using a laser interferometer confirmed a maximum error of 1 μm. Secondly, complex experimental conditions were designed to enhance the robustness and generalizability of thermal error prediction model. Unlike previous studies focused solely on varying feedrate, this research also considers reciprocating motion area variability, revealing its significant impact on thermal error distribution. Then a BPNN model was established, achieving a prediction accuracy of 89%. However, conducting experiments a year later showed a decline in the original model’s prediction accuracy, dropping from an initial 89.11 to 67.4% due to evolving thermal characteristics. To overcome the concept-shifting effect, three transfer learning techniques were applied. This paper presents a comparative analysis of three transfer learning methods such as regular transfer learning (RTL), importance weighting network (IWN), and domain-adversarial training of neural networks (DANN). The results demonstrated that while DANN improved average prediction accuracy to 86.64% with a maximum error of 7.06 μm, RTL achieved competitive results with an average prediction accuracy of 85.91% and a maximum error of 5.36 μm. This residual error suggests the RTL’s potential advantage over the other two methods because the RTL method only fine-tuned the parameters from the previous model using limited data. The profiles of the thermal error only change slightly under thermal characteristics drifting. The RTL method could also prevent the imbalance in data quantity between the source and target domains.