An integrated physical model and extant data based approach for fault diagnosis and failure prognosis: Application to a photovoltaic module

Abstract

Nowadays, the increasing tendency towards the exploitation of solar energy has yielded many technological advancements. Hybrid approaches are attracting attention worldwide to ensure the comprehensive assessment of photovoltaic modules reliability becoming a crucial issue. The present study is dedicated to the investigation of an innovative approach integrating Bond graph theory, Gaussian mixture models and Similarity-based method for fault detection and remaining useful life prediction. In this context, Bond graphs are exploited first to create a dataset covering diverse operational modes of the system. The identification and evaluation of critical sensors for fault observability is also considered, where the dataset is optimized based on the variance analysis. The Gaussian mixture model with its semi-supervised initialization is then utilized for clustering and fault diagnosis, while remaining useful life estimation is performed using a pairwise similarity technique. Validation results on a photovoltaic panel model demonstrate that the Gaussian mixture model consistently outperforms the classical k-Nearest Neighbors model across all key metrics (accuracy of 0.9396 vs. 0.7577, precision of 0.9192 vs. 0.5570, recall of 0.7849 vs. 0.5628, and F1-score of 0.8666 vs. 0.6707), highlighting its superior performance. The remaining useful lifetime model also achieves high accuracy, with Root Mean Square Error values ranging from 0.0282 to 0.0300, indicating minimal prediction error. Additionally, the R-Squared value of ~0.92 shows that the model explains approximately 92% of the variance in remaining useful lifetime predictions, underscoring its strong predictive capability. The results demonstrate the practical effectiveness of the proposed framework for both single and multiple faults. However, some limitations are noted, such as the exclusion of the transition phase in training data and the reliance on controlled conditions. The outcomes of this work are expected to provide valuable insights into the implementation of efficient hybrid frameworks, contributing to the sustainable development of solar energy.