Abstract
Photovoltaic (PV) installations as an alternative energy source are actively encouraged by governments around the world. However, the surge in deployment of PV systems has exposed energy conversion issues at a larger scale due to damaged PV cells, the cause of which, in many cases can be traced back to the manufacturing process. To address this issue clients commission quality assurance organisations to act as the intermediary quality certifier between the solar manufacturer and the end client. These companies monitor the process of cell manufacturing to guarantee cells leaving the manufacturing complex are defect-free. One of the key tools used for this screening process is Electroluminescence (EL) Imaging, allowing inspectors to essentially ‘X-ray’ PV cells followed by a manual inspection of each cell to unearth and remove defective cells from production. However, the manual inspection of thousands of EL images, brings many limitations such as human error, bias, time consumption and high cost of expert personnel.Succinctly, this research presents a novel framework, Gradient Guided Filter Tuning, for the development of a CNN architecture tasked with the detection of Micro-cracks within PV manufacturing facilities. The resultant architecture performs highly in a wide variety of performance indicators including the conventional accuracy metric but also computational, architectural and post deployment metrics. Inspired by the output filter maps as a result of the proposed framework, the internal framework logic is manipulated and extended to address the issue of domain-specific data scarcity, by generating representative augmentations from within the architectural convolutional blocks, introduced as Filter Fused Data Scaling.
Taking note of the advancements in Convolutional Neural Networks (CNN), the project devised a custom CNN architecture, aimed at the PV manufacturing industry, for automated defect detection in PV cells. Contrary to the convention of requiring a humongous dataset for training CNN’s, the project proposed to work with a small dataset, as a manifestation of accounting for the difficulty in acquiring EL processed PV cell images from the production floor. Focus was directed towards comprehension of variance present within the internal classes of the data, enabling the formation of assumptions to dictate the type of transformations and filter configurations within the proposed architecture. The objective of creating a light-weight architecture, enabled subscription to an empirical approach for filter selection, by propagating network gradients back onto the input image, prior to any optimisation of the weights. This provided a mechanism of interpreting the capacity of the filters with respect to the data. After ensuring the required capacity had been developed, various regularisation techniques were utilized for improving model performance. Our proposed Filter Fused Data Scaling process outperformed generic augmentation strategies enabling the developed architecture to learn from domain-specific, well-representative training dataset outputting an F1-score of 98%. The trained and tuned architecture was compared to various state-of-the-art models in the field of image classification, benchmarked against a wide scope of metrics.
In order to gain the fundamental understanding of PV systems, the research was initiated by focusing on the use of numerical input variables extracted from a deployed PV installation at the University of Huddersfield with a generation capacity of 2.2kW. The aim was to address the issue of data extraction and the associated cost and time by implementing an Artificial Neural Network requiring only the solar irradiance and output power for determining the fault type. As a result of utilizing only two input parameters, we were able to introduce the concept of irradiance-power ‘mapping’ to enhance the training process in a progressive manner. The trained network was validated on a separate PV system maintaining an accuracy of above 95% when trained on a different installation with a significantly higher generation capacity of 4.16kW.
| Date of Award | 19 Aug 2022 |
|---|---|
| Original language | English |
| Supervisor | Tianhua Chen (Main Supervisor) & Peter Mather (Co-Supervisor) |