| 陈睿霖,曹晖,郑晓东,闫大鹏,薛霜思,曲凯,汲胜昌.多尺度门控网络的变电站复杂声源分离研究[J].电子测量与仪器学报,2026,40(3):46-57 |
| 多尺度门控网络的变电站复杂声源分离研究 |
| Multi-scale gated separation network for complex sound source separationin substation acoustic environments |
| |
| DOI： |
| 中文关键词: 声源分离 变电站 无人机 多尺度 门控网络 |
| 英文关键词:sound source separation substation UAV multi-scale gated network |
| 基金项目:国家自然科学基金项目（62306232）资助 |
|
|
| Author | Institution |
| Chen Ruilin | School of Electrical Engineering, Xi′an Jiaotong University, Xi′an 710049,China |
| Cao Hui | School of Electrical Engineering, Xi′an Jiaotong University, Xi′an 710049,China |
| Zheng Xiaodong | School of Electrical Engineering, Xi′an Jiaotong University, Xi′an 710049,China |
| Yan Dapeng | School of Electrical Engineering, Xi′an Jiaotong University, Xi′an 710049,China |
| Xue Shuangsi | School of Electrical Engineering, Xi′an Jiaotong University, Xi′an 710049,China |
| Qu Kai | School of Electrical Engineering, Xi′an Jiaotong University, Xi′an 710049,China |
| Ji Shengchang | School of Electrical Engineering, Xi′an Jiaotong University, Xi′an 710049,China |
|
| 摘要点击次数: 344 |
| 全文下载次数: 203 |
| 中文摘要: |
| 变电站设备安全稳定运行对电力系统至关重要。近年来，无人机巡检因其高效、安全优势，已成为变电站维护重要手段。然而，无人机本体及周围环境噪声与设备运行关键声学信息混叠，严重阻碍基于声学信号的设备状态检测与故障预警。为解决此问题并高效分离变电站设备声源，提出一种多尺度门控声源分离网络（multi-scale gated source separation network, GSN）模型。GSN模型采用编码器 分离器 解码器架构,编码器引入并行多尺度一维深度可分离卷积以捕捉多尺度特征；分离器构建局部建模和全局建模双路径，并通过门控融合机制整合输出；解码器采用逐层一维转置卷积与跳跃连接还原时域信号。在包含变电站设备声、无人机噪声及环境背景声的三源混合数据集上实验验证。结果表明，GSN相较于全卷积时域音频分离网络等主流模型，在尺度不变信号实真比、信号干扰比、皮尔逊相关系数指标上分别提高了0.8~7.1 dB、1.3~9.7 dB、0.032~0.297；训练收敛速度与平稳性上亦具明显优势。GSN模型能有效抑制复杂干扰并还原目标设备声源，为变电站设备声学巡检提供高质量信号。 |
| 英文摘要: |
| The safe and stable operation of substation equipment is paramount for power system reliability. In recent years, UAV inspection has emerged as a crucial maintenance tool in substations due to its efficiency and enhanced safety. However, inherent UAV noise, coupled with ambient environmental sounds, often mixes significantly with the vital acoustic signatures of operational equipment. This severe interference substantially hinders acoustic-based equipment status detection and early fault prognosis. To address this challenge and efficiently isolate substation equipment sounds from such complex mixtures, a multi-scale gated source separation network (GSN) model is proposed. The GSN model adopts an encoder-separator-decoder architecture: its encoder incorporates parallel multi-scale 1D depthwise separable convolutions to capture rich features across various temporal scales; the separator constructs a dual-path structure, comprising a local temporal modeling and a global contextual modeling, integrating their outputs via a gated fusion mechanism; the decoder employs layer-by-layer 1D transposed convolutions with skip connections to reconstruct the time domain signal. Experimental validation was conducted on a tripartite mixed dataset comprising substation equipment sounds, UAV noise, and environmental background noise. Results indicate that GSN has superior performance compared to mainstream models. GSN achieved improvements in SI-SDR by 0.8~7.1 dB, SIR by 1.3~9.7 dB, and PCC by 0.032~0.297. Furthermore, GSN demonstrated notable advantages in training convergence speed and stability. The GSN model effectively suppresses complex interference and faithfully reconstructs target equipment sound sources, thereby providing high-quality signals for acoustic inspection of substation equipment. |
| 查看全文 查看/发表评论 下载PDF阅读器 |
|
|
|
