Partial Discharge Pattern Recognition via Improved Evolutionary Convolutional Neural Network

Deep learning has surfaced in numerous practical domains, including face recognition, autonomous driving, and fault detection. In recent times, convolutional neural networks (CNNs), emblematic of deep learning methodologies, have been employed extensively for pattern recognition tasks. Nonetheless, the current architecture of CNNs for pattern recognition heavily relies on domain expertise and entails extensive trial and error. To address this challenge, an evolutionary framework for CNNs has been devised to autonomously determine network architectures tailored for partial discharge pattern recognition. First, we transform the partial discharge data into images to eliminate the effect of noise in the data and perform feature extraction on the partial data. Subsequently, we outline a search space grounded on fundamental network components. Finally, we introduce a flexible variable-length encoding scheme along with corresponding reproduction operators. Experimental results confirm the effectiveness of the proposed method, and we demonstrate higher recognition accuracy in two different real-world scenarios compared to existing methods.