Organizational strategic decision support model in textile manufacturing enterprises under deep learning and knowledge graph

Equipment status data, quality records, maintenance experience and production rules generated during textile manufacturing are scattered across heterogeneous information systems, resulting in the absence of an integrated decision chain linking production status identification and maintenance resource allocation. This study constructs an optimized production decision framework integrating Knowledge Graph (KG), Multi-scale Temporal Convolutional Network (MS-TCN), Relation-aware Graph Attention Mechanism (R-GAM), and Knowledge-Enhanced Deep Reinforcement Learning (K-DRL). The MS-TCN performs multi-scale encoding on speed, temperature, humidity, vibration and energy consumption within successive decision steps. The Textile Manufacturing Domain Knowledge Graph (TM-DKG) uniformly formalizes equipment conditions, fault events, maintenance operations and production performance indicators. The R-GAM aggregates fault nodes and strategy nodes relevant to real-time working conditions according to relation categories. The K-DRL leverages the KG to generate feasible action sets and conducts joint optimization targeting output yield, energy consumption, defect risks, action costs and state switching costs. A public textile production dataset consisting of 1,000 records with 11 fields is adopted to build the state library in experiments. Since this dataset excludes equipment IDs, timestamps and physical sampling frequencies, the experiment is defined as a record-driven and rule-constrained proof-of-concept verification. The TM-DKG-RGAT achieves a Mean Reciprocal Rank (MRR) of 44.1% in the link prediction task. In the simulation environment, the K-DRL obtains an average training reward of 688.7 with a convergence epoch of 420. When 20% Gaussian noise is superimposed, the decision success rate remains at 88.2%. Experimental results demonstrate that relational type encoding and knowledge-based action constraints can improve decision-making effectiveness, training stability and the consistency of interpretable reasoning paths. The proposed method is applicable to the methodological verification of textile production decisions, while its practical on-site performance needs further validation using continuous equipment-level data and maintenance work orders.