Open Access

ARTICLE

An ISSA-Optimized Attention-Enhanced ConvNeXt Model for Partial Discharge Pattern Recognition in Gas-Insulated Switchgear

Rui Huang¹, Ziwei Zhang^2,*, Kari Tusongjiang¹, Bowen Zhang³, Ning Yang³, Xiaowei Li¹, Aimudula Maierdan¹

1 School of Electrical Engineering, Xinjiang University, Urumqi, China
2 Sichuan Energy Internet Research Institute, Tsinghua University, Chengdu, China
3 China Electric Power Research Institute, Beijing, China

* Corresponding Author: Ziwei Zhang. Email:

Computers, Materials & Continua 2026, 89(2), 33 https://doi.org/10.32604/cmc.2026.086232

Received 26 May 2026; Accepted 22 July 2026; Issue published 15 September 2026

Abstract

The accuracy of partial discharge (PD) pattern recognition is essential for assessing the insulation condition of gas-insulated switchgear (GIS). However, in practical recognition tasks, phase-resolved partial discharge (PRPD) patterns often exhibit complex feature distributions, and key discharge characteristics may be weakened during feature extraction. This study proposes an improved sparrow search algorithm (ISSA)-optimized attention-enhanced ConvNeXt model for GIS PD pattern recognition. A multi-criterion grayscale evaluation scheme is first employed to select the most suitable grayscale conversion for PRPD patterns, aiming to preserve informative discharge regions and reduce redundant color interference. Subsequently, an attention-enhanced ConvNeXt model is built, where the convolutional block attention module (CBAM) is embedded after each stage of ConvNeXt to selectively emphasize salient discharge features, and additive angular margin loss (ArcFace) is introduced to enlarge inter-class feature margins and enhance feature discriminability. To automatically determine the optimal hyperparameters and improve optimization stability, the recognition model is then optimized by an ISSA that incorporates a random restart strategy. This strategy triggers population restart upon evolutionary stagnation or diversity collapse, helping the optimizer escape local optima and maintain global search capability. The optimized model is evaluated on five typical GIS PD categories. Experimental results show that the proposed method achieves an accuracy of 97.59%, with precision, recall, and F1-score reaching 97.59%, 97.64%, and 97.61%, respectively. Compared with conventional convolutional neural network and ResNet-series models, the proposed ISSA-optimized ConvNeXt model improves accuracy by 10.95%–13.74%, which provides a practical approach for GIS PD pattern recognition.

---

Keywords

---