Predicting and Explaining Aircraft Engine Vibration Using Machine Learning and SHAP Values

Vibration is a critical quality parameter in aircraft engines, as excessive levels can propagate through the airframe and be perceptible to passengers in the cabin. During engine acceptance tests, vibrations must remain below strict thresholds. Failures at this stage are costly, often requiring engine disassembly with the consequence of halting production. Traditional statistical correlation analyses between individual variables and vibration outcomes have yielded limited insight due to the high dimensionality and independence of many parameters. To address these limitations, the authors investigate machine learning techniques to both predict and explain engine vibration behaviour prior to acceptance testing. The approach consists of two complementary strategies. The first approach combines build data, containing production records, process configuration changes, engine configuration, and operation data measured during acceptance test to train both Random Forest and Neural Network models for accurate vibration grade prediction. By using SHAP (SHapley Additive exPlanations) key contributing variables are identified enabling deeper insight into root causes. The second approach involves training models solely on engine parts manufacturing data available before the acceptance test, enabling the prediction of vibration levels in advance of testing. The research results indicate that SHAP values offer interpretable insights that align with prior internal findings and help uncover patterns that traditional correlation analysis may overlook. Additionally, early-stage prediction models show promising performance, opening pathways for proactive quality control in engine production.