Engine health diagnostic apparatus and method

An engine health diagnostic apparatus for analysing health of an engine, the apparatus comprising:

• feature generation circuitry to process vibration sensor data received from a vibration sensor for sensing vibration at a component of said engine and generate a feature vector indicating a plurality of features of the vibration sensor data;
• and
• processing circuitry to process the feature vector generated by the feature generation circuitry using a trained classification model defined by model parameters characterising a boundary of healthy operation learned from a training set of feature vectors corresponding to vibration sensor data captured during healthy operation of the engine, to generate an engine health indication providing a quantitative indication of deviation of the feature vector from the boundary of healthy operation;
• wherein the plurality of features include at least one peak vibration timing feature indicative of a timing of a peak in the vibration sensor data relative to an operating cycle of said engine.

An engine health diagnostic method for analysing health of an engine, the method comprising:

• capturing vibration sensor data using a vibration sensor fitted to sense vibration at a component of said engine;
• processing the vibration sensor data to generate a feature vector indicating a plurality of features of the vibration sensor data;
• and
• processing the feature vector using a trained classification model defined by model parameters characterising a boundary of healthy operation learned from a training set of feature vectors corresponding to vibration sensor data captured during healthy operation of the engine, to generate an engine health indication providing a quantitative indication of deviation of the feature vector from the boundary of healthy operation, where the engine health indication is calculated based on distance between the feature vector and a boundary function representing the boundary of healthy operation as a hyperplane in N-dimensional hyperspace, where N is a number of features of the feature vector;
• wherein the trained classification model is defined by stored model parameters updated by training in dependence on said training set of feature vectors, the training comprising determining the boundary of healthy operation which maximizes a margin between a predetermined fraction of outlying feature vectors of the training set and a remaining fraction of feature vectors of the training set, wherein the predetermined fraction and the remaining fraction are both non-zero.