Robust Scheduling Methodology to Reduce Risk in Aerospace Production Systems
Siedlak, Dennis Jacob Lee
MetadataShow full item record
The knowledge and value gained from collecting data and being able to monitor vehicles’ performance, safety, reliability, etc. have resulted in a sharp increase in the number of sensors being installed on modern aerospace vehicles. Sensor installations, which are commonly performed manually, lead to increased risk for installation errors and quality issues. These disruptions, in turn, contribute to the program cost overruns, increased schedule risk, and production delays seen throughout the industry. As such, reducing the risk and impact of manual installation tasks on aerospace production flows is becoming increasingly important for such highly schedule- and cost-constrained vehicles. Robust scheduling methodologies, which aim to build schedules with reduced risk of cost or time overruns by minimizing the impact of disruptions, have the potential to meet the requirements of these scheduling problems. Despite the benefit to be gained by implementing robust, detailed project scheduling methodologies, traditional, deterministic strategies still tend to dominate the industry. Two research challenges must be overcome to support the implementation of such robust scheduling techniques in an industrial setting: First, the project scheduling methodologies in use today struggle to model and optimize real-world systems. The increasing complexity of modern aerospace vehicles is only going to exacerbate these difficulties. Second, the transition of new planning and scheduling practices from academia to an industrial setting is commonly challenging. Moreover, this transition is not generally discussed alongside the development of new methods. To address these challenges, this dissertation focuses on the development, implementation, and evaluation of a new planning methodology, named PORRTSS: Production Optimization to Reduce Risk Through Simulation-based Scheduling. A representative case study is used to test the methodology’s capability to model a real production environment and search for improved scheduling options. The case study involves planning sensor installation processes within a provided production schedule to reduce the risk of production delays. Traditional scheduling techniques provide a strong framework to plan and optimize, at a medium level of detail, the completion of primary production processes (e.g. structural assembly, system integration, etc.). However, fully defining the interactions and logic required to evaluate the impact resulting from the sensor installations in this scheduling framework is challenging. The discrete-event simulation paradigm simplifies the definition of these production rules and constraints; however, DES models commonly require too much detail, modeling effort, and optimization time/resources to be useful during pre-production planning. The developed methodology addresses this gap by integrating the process optimization strengths of scheduling with the modeling flexibility of simulation. This enables the fast generation of a limited fidelity simulation that can evaluate the impact of sensor installations to support simulation-based schedule optimization. Even with an optimization framework in place, the deployment of scheduling methodologies developed in academia to an industrial setting remains challenging. A primary barrier that limits the implementation of developed scheduling practices is poor interactions between the system and the human planners. The developed methodology works to overcome these challenges by: 1) increasing the transparency of the planning process, 2) improving collaboration among the stakeholders, and 3) enabling the stakeholders to directly modify the sensor installation plan. Increased transparency and improved collaboration is achieved by developing a decision-support tool that provides both system- and detailed-level views of the planning results. Finally, this research does not claim to provide the answer, but instead, recognizes that there may be additional “soft” constraints. As such, it also provides planners with the capability to make manual modifications to the optimized production plans. This ultimately leads to a more implementable and beneficial planning methodology when compared to the many rigid methods developed in academia. The PORRTSS methodology begins by identifying process constraints contained within a provided medium-level production plan. This schedule is accepted as truth, and the identified process constraints are utilized to automatically generate a baseline discrete-event simulation. The DES model contains process logic to control sensor installation processes, and using this logic, the simulation can estimate the impact of a parametrically defined sensor installation plan. With a parametric model in place, a multi-objective, meta-heuristic optimization algorithm (Non-dominated Sorting Genetic Algorithm-II) is linked to the simulation. The optimizer sets the locations within the primary production plan during which each sensor is installed. The flexible nature of the optimization routine enables the inclusion of a variety of objective functions, including process time and heuristic risk metrics. Once convergence is achieved, the resulting non-dominated points are fed into a data analysis and decision support environment. The decision making system is included to support the implementation of the methodology. An initial system-level view and ranking algorithm enable SMEs to quickly identify points of interest. These can then be compared in more detail to identify similarities and differences between the selected plans. A detailed Gantt chart is also utilized to improve transparency and help understand the reasons for potential problems. Finally, the user is able to make manual modifications to a selected plan to include any additional knowledge or understanding. With the PORRTSS methodology in place, the following experiments are performed to test its ability to overcome the aforementioned research challenges. Focusing on the first research challenge, the difficulty in modeling and optimizing the systems of interest, the appropriateness of the simulation logic and model generation strategy is tested. This is accomplished by generating a model from a schedule for a major sub-assembly of a “real-world” aerospace vehicle. The simulation is shown to appropriately match the baseline schedule and estimate the impact of parametrically defined sensor installations. The appropriateness of the optimization integration is then tested by linking the NSGA-II to the simulation model. An initial experiment conducted with deterministic simulation evaluations is shown to improve the objectives of interest in a short time, which demonstrates that the optimization strategy is effective for the problem of interest. This experiment is then expanded to investigate the impact of directly evaluating the risk in a schedule. This is accomplished by allowing for uncertainty in the simulation and running multiple replication per case to estimate the output distribution of the process time. The optimization is then seeded with generation 500 from the deterministic run and evaluated with this additional information. When comparing the optimization runs with and without the robustness information, the optimization considering robustness is shown to make immediate improvements to the population, especially in the process time risk. While this indicates that the additional information is effectively utilized by the optimization routine, a significant number of further runs are needed to make a generalized conclusion. Despite this, these experiments demonstrate that the developed methodology is able to effectively model and efficiently optimize the sensor installation plan. To address the second research challenge, strategies to improve the “implementability” of the developed methodology are investigated. To test the scalability of the methodology, the modeling and optimization strategy are applied to a range of problem sizes and complexities. The results demonstrate that the methodology is capable of handling models representative of the largest size expected to be seen in an industrial setting. Alternative, point-solution optimization algorithms are also investigated to attempt to improve the optimization’s speed. These are shown to perform adequately when optimizing the deterministic model; however, when considering the stochastic model, their performance does not appear to leverage the schedule risk evaluations provided. Therefore, it cannot be shown that the point-solution algorithms expand the applicability of the methodology. A final experiment is conducted to investigate whether the decision-making environment increases the acceptance and deployability of the methodology. Results generated by the NSGA-II for a real-world planning problem are propagated to the decision-making tool. This tool and results are then provided to the industrial engineers, manufacturing engineers, and avionics experts to down-select to a final plan for execution. Following this exercise, the SMEs confirmed the feasibility of the provided plans and leveraged the decision-making system to down-select and compare scenarios of interest. Overall, the real-world implementation demonstrates that the inclusion of the decision-support environment increased transparency and acceptance of the methodology. From these results, the PORRTSS methodology is shown to overcome the two identified research challenges. The modeling and optimization strategy enables the automatic generation and evaluation of feasible alternative installation plans. The inclusion of heuristic robustness metrics and process time risk enables the identification of more robust installation plans. Then, the transparency and freedom provided by the decision-support system is shown to increase the approachability and deployability of the methodology. Ultimately, this methodology enables the replacement of a manual planning process with one that can better estimate and reduce the system-level impact of small installation steps, which can be large contributors to the time to complete a schedule.