Solar Panel Inspection Employs Drone, AI, Automatic Reporting (2)
Mosaic abnormalities caused by lightning can be detected
In analyzing thermal distribution images shot by drones from the sky in the post-process, AI is not used to identify positions of images and to match them against the solar panel layout drawing of the solar power plant.
Positions of solar panels with abnormal thermal distribution on the layout drawing are identified by using the panel layout drawing of the site obtained in advance from the customer, that is, the power producer or the O&M company, and matching the data on the automatic flight route against positional information of thermal distribution images. This process was automated.
AI is used after this process. Phenomena of abnormal thermal distribution in solar panels are classified into four types and the degrees of urgency for measures are also classified into three levels.
Thermal distribution abnormalities are categorized into "partial overheating/heating (hot spot)," "cluster abnormality," "overheating of entire solar panel" and "overheating of entire string."
"Partial overheating/heating" is caused by microcracks in cells, staining of solar panels or shades on solar panels in many cases. "Cluster abnormality" is caused by disconnection of interconnectors, defective soldering or short circuiting of bypass diodes in many cases (Fig. 2).
"Cluster abnormality" means the stopping of power generation by a cluster (a series circuit of multiple cells that are divided into three one-third sections) due to activation of the bypass diode following an output drop from cells (power generation elements) caused by problems.
"Overheating of entire solar panel" is caused by cracking of cover glass or abnormality in the back sheet in many cases. "Overheating of entire string" is caused by damage to connectors, damage to cables or opening of junction boxes in many cases.
"RAPID machine learning technology," one of the AI technology groups of NEC, is used for automatic analysis.
"RAPID machine learning technology" was developed by incorporating so-called deep learning technology used to teach computers to perform actions similar to human actions including sound recognition, image identification and estimation. Discrimination of normal thermal distribution from abnormal thermal distribution and identification of abnormality trends, which are needed to analyze thermal distribution images shot by drones from the sky, are learned automatically by AI to realize classification of data and detection of abnormalities with high accuracy, according to the company (Fig. 3).
The data needed for automatic learning is called the "teaching data." Normal thermal distribution images and images containing abnormal thermal distribution were collected respectively for solar cells, clusters consisting of cells, solar panels and strings to use as the "teaching data." Teaching of AI was performed using the data to establish a model for automatic analysis by AI.
The analysis accuracy of solar panel thermal distribution images is estimated at 95% in manual analysis for general services. The company made efforts to improve the analysis accuracy by AI to the same or higher level and decided to offer the services because the goal was achieved.