An atrous spatial pyramid pooling–fully connected network-based machining time prediction method for intelligent manufacturing

Accurate machining time prediction is essential for production planning, cost control and quotation in the small-batch manufacture of custom-made parts for special vehicles. However, traditional approaches based on empirical formulas or tabulated standards often suffer from low efficiency, large prediction errors and limited generalization capability. To address these limitations, this study proposes a machining time prediction model integrating Atrous Spatial Pyramid Pooling with a fully connected network (ASPP-FCN). First, key time-related features are extracted from Machining Process Specification Cards stored in the enterprise’s Teamcenter system and reconstructed into a multi-channel pseudo-image using a spatial encoding strategy. An improved ASPP module with multiple atrous convolution branches is then employed to capture multi-scale feature representations, while a Squeeze-and-Excitation (SE) attention mechanism adaptively enhances critical channel features. To reduce redundancy in high-dimensional feature representations, a fully connected network is designed to compress and map the fused features while preserving essential information, thereby generating the final machining time prediction. Finally, the proposed model is validated on a real-world production dataset and compared with multiple baseline models. The results demonstrate that ASPP-FCN achieves high prediction accuracy and strong stability in machining time prediction.