Abstract
During the past decades, the production performance analysis and optimization have played a significant role in queueing network system production planning and operation design for the manufacturing plants, but how to model the optimized production system performance and solve it is facing great challenges. Since the more and more wide use of queueing network, the non-conforming product scenarios such as rework and departure, cause mass random interference. These problems when modelled in the conventional analysis and optimization models are extremely difficult to solve as they are coupled and NP-hard problems. This paper considers a modified M/M/n/m-first come first service (M/M/n/m-FCFS) queueing network approximation model that is developed to maximize utilization and profit by analysing the queue time and queue length of the products under the constraint conditions. A hybrid meta-heuristics is applied to solve the approximation model, both the service rate and buffer allocation of the queueing network system can be determined. Finally, to assess the effectiveness of the proposed method, extensive numerical experimental results from the approximation model are compared to the ARENA simulation, and sensitivity and validation analysis are also conducted, which prove it accurate, believable, and effective.
Introduction
In modern manufacturing, production lines are rarely linear or deterministic. Real-world scenarios involve products failing inspections, leading to rework loops or total scrap (departure). These non-conforming product scenarios introduce significant random interference, making performance analysis and optimization highly complex. Traditional queueing models often oversimplify these systems by assuming infinite buffers or ignoring rework loops. When realistic constraints like finite capacities and rework rates are introduced, the resulting mathematical models become coupled and NP-hard, posing a significant challenge for optimization.
Methodology
The researchers proposed a three-pronged approach to address these complexities:
1. Modified Queueing Model
A modified $M/M/n/m$-FCFS (First-Come-First-Served) approximation model was developed. This model specifically accounts for how rework and departures influence queue lengths and waiting times across multiple workstations with finite buffer capacities.
2. Hybrid Optimization Framework (GA + ANN)
To solve the optimization problem (maximizing profit and utilization), the study employs a hybrid meta-heuristic approach:
- Genetic Algorithm (GA): Explores the global search space to find optimal service rates and buffer sizes.
- Artificial Neural Network (ANN): Acts as a high-speed proxy for the complex mathematical model. Once trained on system inputs and performance outputs, the ANN allows the GA to evaluate thousands of potential configurations in seconds, bypassing computationally expensive iterations.
3. Experimental Setup and Validation
The framework was tested on a numerical benchmark representing a multi-product system with 11 workstations, 8 finite buffers, and 3 distinct product classes. The model's accuracy was validated against ARENA discrete-event simulation software.
Key Insights and Results
The study revealed several critical findings for production management:
- The 0.6 Service Intensity Threshold: When a station's service intensity exceeds 0.6, queue times and Work-in-Process (WIP) levels tend to grow exponentially, marking a critical operational "danger zone."
- Buffer Optimization: Increasing buffer capacity reduces blocking but yields diminishing returns beyond a certain point (approximately 15 units in this model).
- Performance Gains: The hybrid GA-ANN approach reduced system-wide WIP by 15.99% and average sojourn time by 13.64%.
- Accuracy: Predictions closely matched simulations, with errors as low as 3.53% for WIP levels and a maximum of 12.44% for sojourn time.
- Profit vs. Utilization: The results suggest that maximizing machine utilization does not always equate to maximum profit, as maintenance and capital costs must be balanced against throughput.
Conclusion
This research demonstrates that integrating realistic rework and departure scenarios into queueing models is essential for accurate factory performance prediction. The hybrid GA-ANN framework provides a computationally efficient way to optimize complex, finite-capacity production lines. Future work will focus on relaxing Markovian assumptions to include general distributions and incorporating machine reliability factors like failure and repair cycles.
Summary updated