Error Mode Identification in Gas Turbines Through Predictive Maintenance

MAN Energy Solutions SE is a German multinational company that produces diesel engines and turbomachinery for marine and stationary applications. The plant in Oberhausen builds gas turbines for mechanical drive applications (mainly pipeline compressors) and power production. The engines are distributed all over the world, often to remote areas where machine failure can have severe consequences.

MathWorks engagement started in 2020 with a proof of concept using machine learning techniques to detect error conditions in gas turbines. The goal of this consulting project is to automate the time-consuming process of visualizing and manually evaluating measured sensor data, in order to determine gas turbine error conditions at an early stage.

We used GateCycle™ to generate simulated gas turbine data for two different machine types. For each machine type, we generated two data sets: one comprehensive gridded set for training and validation and one random set for testing. We used the training/validation set to build a model in MATLAB^® that predicts the error type and error magnitude with high accuracy using machine learning (classification validation ~ 1%). The model is a combination of a classifier (predicting the error type) and a regression model (predicting the magnitude of the error). Then, we tested the model against the test data (classification test error < 10%).

Since the number of sensors in the engine is usually limited, we then reduced the number of predictors from the full set (21 predictors) down to a minimal set of 6 predictors. Even with a set of only 10 predictors, the accuracy remained in the targeted range (classification test error < 10%).

The next step of the project will be to test the model against real field data. The long-term goal is the integration of the application into the engine control system environment, so engine condition monitoring can be done in real time.