A Neuro-Intelligent Framework for Real-Time Production Scheduling Optimization

Thesis Abstract

In contemporary manufacturing environments, real-time production scheduling faces volatile demand, stochastic processing times, and dynamic machine availability, leading to suboptimal utilization, increased cycle times, and elevated work-in-process levels. This study addresses the gap by developing a neuro-intelligent framework that integrates deep learning-based prediction with reinforcement learning-driven scheduling to optimize real-time decision-making in mixed-model production lines. The aim is to design a scalable framework that autonomously adapts to short-term disturbances while maintaining global objectives such as throughput, tardiness, and resource utilization. Specific objectives are (1) to formulate a hybrid neural predictive model that forecasts short-term job arrivals, processing times, and machine states; (2) to develop a reinforcement learning (RL) scheduler that optimizes dispatching, sequencing, and preventive maintenance within a rolling-horizon control scheme; (3) to evaluate the framework against traditional heuristic and metaheuristic schedulers under varying shop-floor scenarios; (4) to assess robustness under data noise and concept drift; and (5) to establish guidelines for deployment in industrial environments. The research adopts a mixed-methods design combining quantitative simulation with qualitative expert evaluation. The population comprises production lines in a mid-to-large-scale discrete-part manufacturing facility in the automotive components sector. A stratified sample of 12 lines, each with 20–40 interchangeable jobs per batch, will be observed over six months, yielding a dataset of approximately 50,000 scheduling events. Data collection instruments include enterprise resource planning (ERP) extracts, shop-floor sensors (machine status, utilization, and downtime logs), and operator-reported disruption events, complemented by structured expert elicitation sessions with shift supervisors. The neural predictive component will employ a sequence-to-sequence LSTM with attention to forecast next-hour arrival rates and processing-time distributions, calibrated via backpropagation through time. The scheduling component will leverage a Deep Q-Network (DQN) with prioritized experience replay to determine dispatching and sequencing actions under a rolling horizon of up to 60 minutes, constrained by machine eligibility and setup costs. The model specification also integrates a Bayesian linear regression module to quantify uncertainty in predictions and a Markov decision process (MDP) framework to formalize state transitions. Validity and reliability will be ensured through cross-validation for the predictive models, back-testing against historical outages, and sensitivity analyses for RL hyperparameters. Ethical considerations include data anonymization and adherence to industry data governance policies. Data analysis will proceed in three phases (i) descriptive and prognostic analysis of historical data to establish baseline performance metrics; (ii) comparative evaluation of the neuro-intelligent framework against baseline schedulers using metrics such as average tardiness, total makespan, machine utilization, and WIP levels; and (iii) ablation studies to isolate the contribution of each component (neural predictor vs. RL scheduler). Statistical tests will include repeated-measures ANOVA to compare performance across scenarios, with effect sizes reported. The RL component’s policy performance will be analyzed through cumulative reward, stability metrics, and convergence diagnostics, while the predictive models will be assessed via RMSE, MAE, and calibration curves. The study will draw on theories of dynamic capability and incremental learning, as well as established scheduling theory and the stochastic Petri net framework to interpret results. Expected findings indicate that the hybrid neuro-intelligent framework reduces average tardiness by 18–25%, increases on-time delivery rates by 10–15%, and lowers WIP by 12–20% under volatile demand and machine downtime scenarios, outperforming heuristic baselines by substantial margins. The framework is expected to demonstrate resilience to noise and occasional concept drift through continuous learning and uncertainty quantification. The study contributes to knowledge by operationalizing a scalable, data-driven, real-time scheduling paradigm that blends predictive analytics with autonomous decision-making, bridging theories of dynamic capability with contemporary deep reinforcement learning for manufacturing systems. Recommendations include integration protocols for ERP and shop-floor data streams, guidelines for selecting RL hyperparameters in production contexts, and a roadmap for pilot deployment in mixed-model lines, with emphasis on data governance, human-in-the-loop oversight, and incremental roll-out to ensure safe and reliable performance.