Integrating sound and smoke detection for electrical fault identification within PV inverters: An automated approach

Like many semi-arid countries around the world, Morocco benefits from ample solar radiance throughout the year. Taking advantage of this significant energy source, the local government has deployed vast photovoltaic installations to cover the nation’s growing electricity demand relying mainly on fossil fuels. However, numerous challenges arise along the way while applying this strategy, one of the most critical being maintenance. In this context, and with the goal of improving operational practices and enhancing the reliability of PV systems, this framework introduces a novel approach that aims to identify electrical malfunctions within inverters via sound analysis. Three Convolutional Neural Networks models were developed with various input features: spectrograms, log mel-spectrograms and mel-frequency cepstral coefficients. The latter achieved a precision of 0.9957, a recall of 0.9978, and a loss of 0.0276. Not only does it deliver important performance, but it is also lightweight with a compact size of 4.38MB compared to the spectrogram model and the log mel-spectrogram model which necessitate 520MB and 33.2MB, respectively. The solution is computationally efficient, with only 92849 trainable parameters compared to the spectrogram and log mel-spectrogram models, which have 11356849 and 723633 trainable parameters, respectively, making it well-suited for deployment on edge devices with limited resources. The sound analysis systems were trained to distinguish between electrical problems and normal operating sounds throughout the use of two datasets, labeled “1” for malfunction events and “0” for standard day-to-day circumstances surrounding the inverter. Additionally, due to its enclosed nature, a smoke detection sensor is placed inside the inverter to capture any issue that may have gone undetected by the first method. The entire strategy is designed for real-time performance with alerting mechanisms and the capability to store fault records in straightforward local and remote databases. This approach demonstrates that sound analysis models combined with auxiliary sensors provide a cost-effective and efficient manner to monitor photovoltaic setups, ensuring safety and mitigating downtime through proactive maintenance.