IMPROVING THE METOD OF FUNCTIONING OF THE CYBER-PHYSICAL SYSTEM FOR MONITORING DEFECTS IN PHOTOVOLTAIC MODULES OF A SOLAR POWER STATION

Abstract

The article presents an improved model of the functioning of a cyber-physical system for monitoring defects in photovoltaic modules of a solar power plant. The key feature of the developed system is its integrated architecture, which involves the integration of a surveillance camera model, image processing functions based on a convolutional neural network (CNN), as well as object detection and object tracking algorithms. To ensure the geometric accuracy of image analysis, the surveillance camera is modeled using a pinhole model that allows determining the geometric parameters of images in computer vision tasks, calibrating the camera to determine its internal and external parameters, and correcting lens distortion. Additionally, the developed model provides for automated determination of whether the detected objects belong to predefined classes of defects. The classification is based on the output of a convolutional neural network using the softmax function, which predicts the probability of a defect in each cell of the image grid, providing a quantitative assessment of the confidence in the detected class. An important aspect of the improvement is the integration of Object Detection and Object Tracking technologies, which effectively eliminates the re-detection of already detected defects in the video sequence. This leads to a significant reduction in the number of duplicate and false alarms of the system, increasing its computational efficiency and the reliability of monitoring results. To further improve the tracking accuracy and reliable identification of previously detected defects over time, the model comprehensively uses Deep Simple Online and Realtime Tracking (Deep SORT) algorithms. This approach is based on a combination of two mathematical methods: the Kalman filter to eliminate noise and random outliers in the weighting coefficients of the tracked objects, which ensures more stable and reliable tracking and prediction of the position of objects in subsequent frames, and the Mahalanobis distance to quantify the degree of similarity between the weighting coefficients of already known and newly detected objects, which contributes to more accurate defect identification. In addition, the system integrates the Density-Based Spatial Clustering of Applications with Noise (DBSCAN) clustering algorithm, which classifies detected defect polygons by their spatial location. This allows detecting groups of closely spaced defects, which can be useful for diagnosing system problems or identifying patterns in the distribution of defects on the surface of a solar power plant. The results of the integrated approach demonstrate a significant improvement in the accuracy of defect detection due to the synergistic effect of the combination of CNN for pattern recognition, Softmax for probabilistic classification, DBSCAN for spatial distribution analysis, and Deep SORT for stable tracking. The detection speed is also increased by integrating Object Detection and Object Tracking, which minimizes the need to re-analyze the same image areas. The system's reliability is enhanced by the use of the Kalman filter to reduce the impact of random noise, the Mahalanobis distance for more objective identification, and the DBSCAN algorithm for detecting spatial anomalies.