Self-tuning linear adaptive genetic algorithm for feature selection in machinery fault diagnosis

Abstract

Advanced pattern recognition of a machine learning classifier function, aka the black box allows automated machinery fault diagnosis and outperforms classic decision-making mechanisms. Nonetheless, the black box supervised learning is subject to overfitting when the usefulness of statistical input features is unknown, and results in biased prediction. An established genetic algorithm (GA) feature selection (FS) iteratively searches and extracts quality feature subset as the classifier input targeting fitness function prediction error minimisation. However, static genetic parameters are prone to premature convergence in multi-objective optimisation, while manual parameter tuning is computationally expensive. Thus, this study proposed an optimisation methodology based on adaptive search space strategy with a customised parameter tuning mechanism and stopping criteria revision to improve local convergence and prediction efficacy. By embedding an exploration-exploitation cycle as a function of the iterative fitness, Self-Tuning Linear Adaptive GA (Stella GA) adjusts the standard genetic parameters in parallel. Assuming convergence is detected via unique static fitness evaluation threshold, linear-additive-gain-incrementcontrol equations gradually increase mutation rate and population size in an attempt to enhance population diversity. Alternatively, conservative genetic setting is initiated upon the global best score update to exploit the new search space neighbourhood. Stella GA alters parameters recursively until the hybrid stopping criteria is met to allow the tracking of floating genetic variables in preventing premature termination and computation explosion. As demonstrated in multi-objective optimisation problem with five machinery fault diagnosis benchmarking datasets, Stella GA generated feature subset candidate population capable of deterring premature convergence. A prediction benchmarking against modern classifiers (Deep Learning) and classic FS alternatives (GA, Binary Particle Swarm Optimisation and Neighbourhood Component Analysis) indicates the proposed Stella GA consistently returned classifier with desirable efficacy in accuracy (maximum 4.5% increment in hydraulic system) and confusion matrix statistical indicators (maximum 0.0974 increment in Matthews Correlation Coefficient for pumps), with the optimal feature reduction (maximum 58.59% in rotor fault diagnosis). This result suggests that Stella GA yielded optimal machinery fault diagnosis by further decrement in model overfitting, and the removal of manual tuning and unstable parameter feedback.