An approach for mining typical machining process plans integrating representation learning and spectral clustering

A typical machining process plan (TMPP) refers to the standardized process plan for a specific category of parts or products solidified through long-term manufacturing practices. It embeds substantial reusable process information, allowing for high-quality and rapid process planning for new parts. Consequently, TMPP is generally regarded as process knowledge and stored in knowledge base. With the frequent implementation of computer-aided process planning (CAPP) systems in manufacturing enterprises, a large number of planed process documents have been produced and archived in the database, from which TMPPs can be identified and extracted through knowledge discovery (KD) approaches. Due to complex characteristics of machining process plan (MPP) data including multi-level hierarchies, multi-dimensional attributes and semantic dependencies, TMPPs mining from MPs becomes highly challenging. To address this issue, the paper proposes an approach for mining TMPPs based on representation learning and spectral clustering. In this approach, a representation learning model is developed based on TransD to capture the deep-level features of machining processes comprehensively for the following data analysis. Subsequently, a similarity calculation model is constructed for MPPs data in the form of graph model, and TMPPs mining algorithm is developed using spectral clustering. Finally, an index V_tp is defined to quantify and evaluate the typicality of a TMPP in view of the application of TMPPs in CAPP systems. The proposed method is performed on the dataset obtained from planed process documents of shaft, gear, plate, and box parts, and comparisons are provided with several existing clustering algorithms. The results show that the proposed approach achieves the V_tp value around 0.8. It attains robust clustering results, as evidenced by the mean value of purity (82%), ARI (0.78), and NMI (0.71), outperforming K-means, Ward and DBSCAN clustering algorithm, thus verifying its correctness and effectiveness.