现代电子技术

针对气体绝缘开关（GIS）局部放电（PD）在线监测中，深度学习模型在小样本PRPD图谱数据集上易过拟合，且难以部署于资源受限边缘设备的双重挑战，提出一种兼顾高精度与硬件效率的软硬件协同设计方案。在算法层面，设计了一种轻量化卷积神经网络。该网络以跨阶段局部残差块为核心，通过特征重用与分流机制改善梯度传播路径，从而抑制过拟合；再结合上采样与多尺度特征融合策略，增强模型对PRPD图谱细微特征的提取能力。在硬件层面，基于Zynq⁃7z020平台为该网络设计了专用加速器。采用高层次综合（HLS）工具将C++模型代码综合为硬件描述语言，并应用流水线和循环展开指令优化卷积运算。采用双缓冲机制优化数据传输效率，并对权重参数与激活函数进行16 bit定点量化，从而平衡模型精度与存储开销。实验结果表明，所提出的模型在训练平台上的分类准确率达到97.50%，性能优于经过公平适配的主流轻量化网络。将该模型部署到Zynq⁃7z020平台后，准确率仅下降1.03%，存储空间减小约50%。所提出的软硬件协同设计方案能够在保证高识别精度的前提下，显著降低模型的计算与存储开销，可为局部放电等电力设备故障的实时、低功耗在线监测提供参考。

In allusion to the dual challenges in online monitoring of partial discharge (PD) in gas⁃insulated switchgear (GIS), where deep learning models are prone to overfitting on small⁃sample phase resolved partial discharge (PRPD) pattern datasets and are difficult to deploy on resource⁃constrained edge devices, a software⁃hardware co⁃design scheme that balances high accuracy and hardware efficiency is proposed. At the algorithm level, a lightweight CNN is designed. In this network, the cross⁃ stage local residual block is used as the core, and the feature reuse and split mechanisms are used to improve the gradient propagation, thus suppressing overfitting. The upsampling and multi⁃scale feature fusion strategies are combined to enhance the model’s ability to extract the fine⁃grained features from the PRPD pattern. At the hardware level, a dedicated accelerator is designed based on Zynq⁃7z020 platform. The C++ code is compiled into hardware description by means of high⁃level synthesis (HLS), and the instruction ⁃ optimized convolution operation is conducted by means of pipeline and loop unrolling. A double buffering scheme is used to optimize the data transmission efficiency, and 16⁃bit fixed⁃point quantization for weight parameters and activation functions is used to balance model accuracy and storage overhead. The experimental results show that the proposed model can realize a classification accuracy of 97.50% on the training platform, and its performance is better than that of mainstream lightweight networks with fair adaptation. After deploying this model to the Zynq⁃7z020 platform, the accuracy rate is only decreased by 1.03%, and the storage space is reduced by approximately 50%. The proposed hardware⁃software co⁃design scheme can significantly reduce the computational and storage overhead of the model while ensuring high recognition accuracy, and can provide an efficient and reliable solution for real⁃time and low⁃power online monitoring of power equipment faults such as partial discharge.