PSOPruner: PSO-Based Deep Convolutional Neural Network Pruning Method for PV Module Defects Classification

Abstract

Photovoltaic (PV) modules can be damaged by external environment, which can significantly reduce power generation efficiency. Electroluminescent (EL) imaging is an economical and widely applied technique for PV module defects detection. The analysis of EL images, however, is still labor-consuming. To solve this problem, a deep convolutional neural network (DCNN) is proposed to automatically detect and classify defects. The proposed DCNN performs well on a large public EL dataset with more than 2600 images extracted from both monocrystalline and polycrystalline PV modules. Nevertheless, the proposed DCNN cannot be directly applied on portable or embedded devices due to its requirement of large-size memory and intensive computation. To handle this challenge, an evolutionary algorithm-based DCNN pruning method, dubbed PSOPruner, is further developed. The pruning problem of DCNN is formulated as a search problem, which is solved by particle swarm optimization (PSO) algorithm. To improve the quality of the pruning scheme, a tailored trick is considered that the automatic searching process with PSO algorithm is repeated for multiple rounds. To illustrate the effectiveness of the proposed PSOPruner, we compare it with mainstream DCNNs and lightweight CNNs in terms of model complexity and model accuracy. Experimental results demonstrate that the proposed method could efficiently reduce the amount of model parameters with slight drop of accuracy.