Input Feature Purification Method for Engine Fault Prediction Based on Physical Prior FMA Analysis

Abstract

To address the issues including redundant measured engine data, low feature interpretability and incompatibility with neural network modeling, this paper proposes an improved PP-FMA feature purification method based on physical priors. Conventional Failure Mode Analysis (FMA) and purely mathematical analysis both have prominent drawbacks: strong subjectivity in feature screening and severe feature redundancy. This paper establishes a coupled PP-FMA analysis system integrating engine physical mechanisms and mathematical models to pre-optimize raw monitoring data and select features with low redundancy and high fault discrimination. Experimental results demonstrate that the proposed method can reduce feature dimensions, strengthen fault-sensitive information and greatly improve the quality of input features. It can provide a stable and reliable data preprocessing scheme for fault prediction of natural gas engines.