Photovoltaic (PA) arrays may be affected by a variety of factors during operation, leading to different types of failures. Real-time monitoring, fault diagnosis, and predictive maintenance of PV array data can be realized through machine learning algorithms, an approach that is not limited by geography and can improve system reliability and efficiency. The current-voltage (I-V) curve of a PV array is an important metric that contains a great deal of information about the health of the PV module, which is crucial for timely fault detection and health assessment. However, existing methods only extract part of the information from the I-V curve for diagnostic analysis, without digging deeper into all the information. As a result, the range of detectable PV array faults remains limited. To address the problems, an I-V curve correction algorithm is proposed to correct the effects of irradiance and temperature on the characterization of the same fault type, effectively eliminating the coupling effect of environmental variables on the characterization of fault features. Then, the CatBoost model is used to realize real-time, high-accuracy fault intelligent diagnosis of PV arrays with small samples. The model′s key hyperparameters are optimized using the sparrow search algorithm. Finally, in order to further enhance the optimization ability of the sparrow search algorithm, the sparrow search algorithm is improved by introducing the fusion elite inverse learning strategy and the Cauchy Gaussian variation strategy, so that it achieves the best effect in optimizing the CatBoost model. The results show that when using simulated data for model training and field data for fault diagnosis, only one and two misdiagnosed samples appear in the test set, respectively. The classification accuracy of the deeply optimized integrated learning model CatBoost reaches 99.9% in both cases, demonstrating its exceptional diagnostic performance.