Fuzzy Brain-State-in-a-Box neural network for intelligent classification of solar panel defects

Abstract

The article considers the problem of intelligent classification of solar panel defects in remote monitoring systems based on unmanned aerial vehicles. The relevance of the study is due to the need to increase the accuracy of automated diagnostics of photovoltaic systems in conditions of incomplete data, the presence of noise components and partial overlap of features of different types of defects. The use of traditional computer vision methods and deep neural networks for unmanned aerial vehicle monitoring tasks is often accompanied by significant computational complexity, the need for a large amount of training data and the complexity of implementation on low-power edge-AI platforms. In this regard, the development of compact intelligent classification models capable of operating in conditions of limited computing resources is relevant. The work aims to develop a fuzzy neural model Fuzzy Brain-State-in-a-Box, for intelligent classification of solar panel defects in a compact feature space. The proposed approach combines the mechanisms of Brain-State-in-a-Box associative memory, fuzzy inference and prototypical class representation within a single hybrid architecture. The research methodology is based on the formation of a five-dimensional feature space after preprocessing and dimensionality reduction of experimental data. To increase the stability of the classification, a temperature-controlled fuzzy membership evaluation mechanism was used, which provides adaptive regulation of the level of fuzziness of the classification solution. As part of the study, a software implementation of the Fuzzy Brain-State-in-a-Box model was performed and experimental studies were conducted on a dataset of one thousand samples. The results confirmed the effectiveness of the proposed approach and provided a classification accuracy of ninety-one percent, a macro-F1 measure at the level of eighty-nine hundredths and Cohen's Kappa at the level of eighty-six hundredths. The analysis showed that the use of fuzzy degrees of membership and prototypical representation of classes allows for increasing the model's resistance to noise and partial overlap of features. The practical significance of the work lies in the possibility of using the developed Fuzzy Brain-State-in-a-Box model in systems for monitoring and diagnosing defects of solar panels based on unmanned aerial vehicles and edge-AI platforms, where low computational complexity, compactness of the algorithm and the ability to work in real time are important.