Using Neural Networks with Limited Data to Estimate Manufacturing Cost

Neural networks were used to estimate the cost of jet engine components, specifically shafts and cases. The neural network process was compared with results produced by the current conventional cost estimation software and linear regression methods. Due to the complex nature of the parts and the limited amount of information available, data expansion techniques such as doubling-data and data-creation were implemented. Sensitivity analysis was used to gain an understanding of the underlying functions used by the neural network when generating the cost estimate. Even with limited data, the neural network is able produced a superior cost estimate in a fraction of the time required by the current cost estimation process. When compared to linear regression, the neural networks produces a 30% higher R value for shafts and 90% higher R value for cases. Compared to the current cost estimation method, the neural network produces a cost estimate with a 4.7% higher R value for shafts and a 5% higher R value for cases. This significant improvement over linear regression can be attributed to the neural network ability to handle complex data sets with many inputs and few data points.