基于红外图像与声纹信号融合的电厂设备故障深度识别方法

张德有

基于红外图像与声纹信号融合的电厂设备故障深度识别方法

投稿时间：2026-06-12 修订日期：2026-06-29 点此下载全文

引用本文：

摘要点击次数: 33

全文下载次数: 0

作者	单位	邮编
张德有^*	国能山西河曲发电有限公司	036002

中文摘要:针对电厂旋转与电气设备故障识别中单一模态表征不完整、固定权重多模态融合难以适应工况波动与噪声干扰的问题，提出一种红外图像与声纹信号融合的故障深度识别方法：构建双分支特征提取网络，以双向跨模态多头注意力建模热分布与声纹时频特征的关联；以输入自适应的置信度门控在线估计模态可靠度并动态分配融合权重；以跨模态一致性约束正则化决策。在半实物试验平台6类状态、7 200组红外—声纹配对实测样本上，所提方法准确率达97.0%，误报率与漏报率分别为1.7%和0.6%，较特征级拼接与跨模态Transformer融合分别提升3.6和1.9个百分点；声纹信噪比?5 dB时仍保持91.3%。该方法以约8.6 M参数与11.8 ms单样本GPU推理时间兼顾精度、鲁棒性与实时性，满足电厂设备在线监测需求。

中文关键词:故障识别红外图像声纹信号跨模态注意力置信度门控

Deep Fault Recognition Method for Power Plant Equipment Based on Infrared Image and Voiceprint Signal Fusion

Abstract:Aiming at the problems that single-modal information is insufficient for fault recognition of rotating and electrical equipment in power plants and that fixed-weight multimodal fusion can hardly adapt to operating-condition fluctuation and noise, a deep fault recognition method based on infrared image and voiceprint signal fusion is proposed. A dual-branch feature extraction network is built, a bidirectional cross-modal multi-head attention module models the correlation between thermal-distribution and time-frequency voiceprint features, an input-adaptive confidence-gated mechanism estimates modality reliability online to allocate fusion weights dynamically, and a cross-modal consistency constraint regularizes the decisions. On 7 200 paired samples of six states from a semi-physical test bench, the method achieves 97.0% accuracy with a 1.7% false alarm rate and a 0.6% missed detection rate, exceeding feature-concatenation and cross-modal Transformer fusion by 3.6 and 1.9 percentage points, and maintains 91.3% accuracy at a voiceprint SNR of ?5 dB. With about 8.6 M parameters and 11.8 ms GPU inference time per sample, it balances accuracy, robustness and real-time performance for online monitoring of power plant equipment.

keywords:fault recognition infrared image voiceprint signal cross-modal attention confidence gating

查看全文 查看/发表评论 下载pdf阅读器