High-Fidelity Soiling Detection for Photovoltaic Panels via Compressive Sensing and Learned Dictionaries

Accurate, high-resolution monitoring of soiling on photovoltaic (PV) panels is critical for optimizing cleaning schedules and maximizing energy yield, yet it presents a significant data challenge. Conventional methods either offer poor spatial resolution or generate massive datasets that are costly to acquire, store, and transmit. This paper introduces a fundamentally different approach to soiling detection by developing a novel framework based on the principles of Compressed Sensing (CS). We propose a stationary, non-contact sensing station that reconstructs a high-resolution soiling map from a small set of non-adaptive, compressive measurements. The system is architecturally based on a single-pixel imaging modality, utilizing a Digital Micromirror Device (DMD) to project a series of structured patterns onto the panel surface while measuring the integrated reflected light. The core novelty of our work lies in moving beyond standard, generic sparsifying bases (e.g., wavelets). We develop a custom, overcomplete dictionary of soiling ”atoms” that is learned directly from a representative set of soiling patterns using the K-SVD algorithm. This tailored dictionary provides a much sparser representation of soiling distributions, enabling significantly more accurate reconstruction from fewer measurements. Furthermore, the framework is extended to multi-spectral compressive sensing by incorporating several filtered photodetectors, allowing for the classification of soiling composition from the same highly-compressed dataset. Theoretical simulations demonstrate that this method can reconstruct soiling maps with high structural similarity to the ground truth using only ۱۵-۲۰% of the data required by conventional Nyquist-rate sampling, establishing a new paradigm for data-efficient, high-fidelity soiling diagnostics.

کلیدواژه ها:

Soiling ، photovoltaic ، compressed sensing ، single-pixel imaging ، dictionary learning ، K-SVD ، sparse representation ، condition monitoring