Hybrid Explainable AI for Machine Predictive Maintenance: From Symbolic Expressions to Meta-Ensembles

Created by W.Langdon from gp-bibliography.bib Revision:1.8519

@Article{andelic:2025:Processes,

• author = "Nikola Andelic and Sandi {Baressi Segota} and Vedran Mrzljak",
• title = "Hybrid Explainable {AI} for Machine Predictive Maintenance: From Symbolic Expressions to Meta-Ensembles",
• journal = "Processes",
• year = "2025",
• volume = "13",
• number = "7",
• pages = "Article No. 2180",
• keywords = "genetic algorithms, genetic programming",
• ISSN = "2227-9717",
• URL = "https://www.mdpi.com/2227-9717/13/7/2180",
• DOI = "doi:10.3390/pr13072180",
• abstract = "Machine predictive maintenance plays a critical role in reducing unplanned downtime, lowering maintenance costs, and improving operational reliability by enabling the early detection and classification of potential failures. Artificial intelligence (AI) enhances these capabilities through advanced algorithms that can analyse complex sensor data with high accuracy and adaptability. This study introduces an explainable AI framework for failure detection and classification using symbolic expressions (SEs) derived from a genetic programming symbolic classifier (GPSC). Due to the imbalanced nature and wide variable ranges in the original dataset, we applied scaling/normalization and oversampling techniques to generate multiple balanced dataset variations. Each variation was used to train the GPSC with five-fold cross-validation, and optimal hyperparameters were selected using a Random Hyperparameter Value Search (RHVS) method. However, as the initial Threshold-Based Voting Ensembles (TBVEs) built from SEs did not achieve a satisfactory performance for all classes, a meta-dataset was developed from the outputs of the obtained SEs. For each class, a meta-dataset was preprocessed, balanced, and used to train a Random Forest Classifier (RFC) with hyperparameter tuning via RandomizedSearchCV. For each class, a TBVE was then constructed from the saved RFC models. The resulting ensemble demonstrated a near-perfect performance for failure detection and classification in most classes (0, 1, 3, and 5), although Classes 2 and 4 achieved a lower performance, which could be attributed to an extremely low number of samples and a hard-to-detect type of failure. Overall, the proposed method presents a robust and explainable AI solution for predictive maintenance, combining symbolic learning with ensemble-based meta-modelling.",
• notes = "also known as \cite{pr13072180}",

}

Genetic Programming entries for Nikola Andelic Sandi Baressi Segota Vedran Mrzljak

Citations