Modeling and compensation of small-sample thermal error in precision machine tool spindles using spatial–temporal feature interaction fusion network

Precision machine tools play a pivotal role in high-accuracy machining across diverse sectors, from aerospace to medical devices. Among the factors affecting machining accuracy, thermal error control is crucial. Traditional deep-learning models sequentially connect temporal and spatial models or use parallel architectures to independently analyze thermal information data. However, these approaches struggle to capture long-term relationships and global features within the large-sample thermal information data, as spatiotemporal features are not independent. Then these spatiotemporal models are often ineffective at capturing the long-terms relationships and global features of small-sample thermal information data, leading to reduced prediction accuracy and poor robustness. To address these challenges, a novel spatial–temporal feature interaction fusion network is proposed. This network comprises spatiotemporal feature interaction blocks, a spatiotemporal feature gated fusion layer, and a residual structure. The gated temporal convolution is combined with multi-head self-attention to capture temporal information and integrate temporal features, capturing both short- and long-term relationships. For spatial information, it uses a graph convolution network and multi-head self-attention, integrated through spatial features fusion. This approach facilitates the simultaneous utilization of spatiotemporal features. The thermal information data is then processed through a spatial–temporal feature gated fusion layer, blending information using a gating mechanism. Results indicate that the proposed spatial–temporal feature interaction fusion network significantly surpasses other models, including least square support vector machine, long short-term memory, convolutional neural network-long short-term memory, temporal convolutional network, spatial–temporal graph convolutional network, graph multi-attention network, and spatial–temporal synchronous graph convolutional networks in small-sample thermal information prediction and thermal error compensation. Specifically, a 97.91% reduction in thermal error during compensation is achieved by the proposed spatial–temporal feature interaction fusion network.