| Popis: |
This paper presents a new method named Split Training Method (STM) for generating a data driven model of gas turbine engines to be used for anomaly detection and prognosis. The data driven model in this study simulates the relationship among engine parameters, such as sensor outputs and control inputs from pilots. More accurate trend monitoring of the relationship among engine parameters enables earlier anomaly detection with high accuracy. Machine learning techniques such as clustering, nonlinear regression, classification and optimization are used in STM. The input data for generating a simulation model is extracted from the training data by applying clustering algorithm to split into some clusters and outliers. This is why our method is named “Split” Training. Outliers are excluded from training data and each cluster of training data generates a simulation model with nonlinear regression. Data points in the test data are classified to each cluster and used for evaluation of the corresponding model accuracy. In order to generate the simulation model with the highest accuracy, training data extraction is optimized by changing clustering shapes. The shapes of the optimized clusters are unique to individual engines and simulation target parameters. The novelty of STM is that the training data is selected in accordance with the characteristics of the data itself and not selected on the condition determined in advance. The result of comparison between STM and the conventional method shows significant accuracy enhancement. |
