Deep learning approaches for visual faults diagnosis of photovoltaic systems: State-of-the-Art review

Abstract

PV systems are prone to external environmental conditions that affect PV system operations. Visual inspection of the impacts of faults on PV system is considered a better practice rather than onsite fault detection mechanisms. Faults such as hotspot, dark area, cracks, glass break, wavy lines, snail tracks, corrosion, discoloration, junction box failure and delamination faults have different visual symptoms. EL technology, infrared thermography, and photoluminescence approaches are used to extract and visualize the impact of faults on PV modules. DL based algorithms such as, CNN, ANN, RNN, AE, DBN, TL and hybrid algorithms have shown promising results in domain of visual PV fault detection. This article critically overviews working mechanism of DL algorithms in terms of their limitations, complexity, interpretability, training dataset requirements and capability to work with another DL algorithms. This research article also reviews, critically analyzes, and systematically presents different clustering algorithms based on their clustering mechanism, distance metrics, convergence criteria. Additionally, their performance is also evaluated in terms of DI, CHI, DBI, S-score, and homogeneity. Moreover, this research work explicitly identifies and explains the limitations and contributions of recent and older techniques employed for features extraction, data preprocessing, and decision making by performing SWOT analysis. This research work also recommends future research directions for industry and academia.

Keywords

• Deep learning
• Machine learning
• PV visual faults
• Fault classification