Adaptive Garment Manufacturing: A Multi-Modal AI Pipeline for Trend Prediction and Digital Twin Simulation
Mozaffari, Somayeh (2026)
2026
Tietotekniikan DI-ohjelma - Master's Programme in Information Technology
Informaatioteknologian ja viestinnän tiedekunta - Faculty of Information Technology and Communication Sciences
Hyväksymispäivämäärä
2026-06-22
Julkaisun pysyvä osoite on
https://urn.fi/URN:NBN:fi:tuni-202606197771
https://urn.fi/URN:NBN:fi:tuni-202606197771
Tiivistelmä
The fashion industry faces severe sustainability challenges, with overproduction and rigid supply chains generating massive amounts of waste. This thesis presents an end-to-end AI-driven framework for adaptive garment manufacturing that connects real-time social media trend detection with production planning and robotic execution. A multi-modal pipeline combines detections from a YOLOS (You Only Look at One Sequence) vision transformer applied to Instagram and Pinterest images, zero-shot textual analysis of captions using the BART (Bidirectional and Auto-Regressive Transformer) language model, and Google Trends search interest to construct a daily multi-modal trend dataset. These signals are used to train a Temporal Fusion Transformer (TFT) for probabilistic demand forecasting over a 30-day horizon.
The forecast is utilized in a discrete-event digital twin of a garment factory, which models cutting, sewing, and inspection stages, inventory management, and waste tracking. A Proximal Policy Optimization (PPO) agent learns optimal reorder points to maximize profit under stochastic demand, while a Soft Actor-Critic (SAC) agent controls a simulated robotic cutting cell (Gazebo / Robot Operating System 2, ROS 2) to minimize fabric waste.
The integrated system demonstrates that incorporating natural language processing (NLP)-based textual trend signals significantly improves demand forecasts. Without NLP, the forecast is nearly flat at approximately 780 units/day; with NLP, it shows a realistic declining trend from 1249 to 845 units/day. The SAC-based cutting agent reduces average fabric waste by 47% compared to a heuristic baseline (from 15% to 7.91%). During a 30-day evaluation period, the digital twin maintains a throughput of 28,116 garments, while the reinforcement learning (RL)-based inventory policy achieves profit comparable to an optimized fixed-rule baseline. The closed-loop architecture demonstrates that multi-modal AI, cognitive digital twins, and reinforcement learning can jointly reduce overproduction, inventory surplus, and cutting waste, offering a blueprint for sustainable and adaptive manufacturing. The source code and simulation environments developed for this thesis are available at https://github.com/Somayeh5656/adaptive-garment-manufacturing.
The forecast is utilized in a discrete-event digital twin of a garment factory, which models cutting, sewing, and inspection stages, inventory management, and waste tracking. A Proximal Policy Optimization (PPO) agent learns optimal reorder points to maximize profit under stochastic demand, while a Soft Actor-Critic (SAC) agent controls a simulated robotic cutting cell (Gazebo / Robot Operating System 2, ROS 2) to minimize fabric waste.
The integrated system demonstrates that incorporating natural language processing (NLP)-based textual trend signals significantly improves demand forecasts. Without NLP, the forecast is nearly flat at approximately 780 units/day; with NLP, it shows a realistic declining trend from 1249 to 845 units/day. The SAC-based cutting agent reduces average fabric waste by 47% compared to a heuristic baseline (from 15% to 7.91%). During a 30-day evaluation period, the digital twin maintains a throughput of 28,116 garments, while the reinforcement learning (RL)-based inventory policy achieves profit comparable to an optimized fixed-rule baseline. The closed-loop architecture demonstrates that multi-modal AI, cognitive digital twins, and reinforcement learning can jointly reduce overproduction, inventory surplus, and cutting waste, offering a blueprint for sustainable and adaptive manufacturing. The source code and simulation environments developed for this thesis are available at https://github.com/Somayeh5656/adaptive-garment-manufacturing.