计算机科学
人工智能
模式识别(心理学)
核(代数)
离散小波变换
鉴定(生物学)
断层(地质)
小波
小波变换
算法
分布(数学)
计算机视觉
平稳小波变换
人工神经网络
卷积神经网络
信号处理
领域(数学分析)
小波包分解
第二代小波变换
多分辨率分析
连续小波变换
故障检测与隔离
系统标识
出处
期刊:
日期:2026-06-01
卷期号:: 575-580
标识
DOI:10.1109/icetis70504.2026.11633396
摘要
To address weak transient feature extraction, low similar-fault accuracy, and severe background noise in distribution networks, this paper proposes a fault diagnosis framework integrating discrete wavelet transform (DWT), an enhanced selective kernel attention (E-SKA) mechanism, and a CNN. The framework first transforms raw sequential signals into a 2D time-frequency feature matrix via DWT. For deep feature refinement, the proposed E-SKA module incorporates a dualpath pooling fusion strategy to capture transient spikes, alongside an independent Sigmoid-based weight selection mechanism to mitigate competitive conflicts among multi-scale components. Residual shortcuts are further embedded to safeguard original transient characteristics against attenuation. Finally, a CNN backbone executes autonomous fault classification. Simulation results indicate that the model yields superior convergence speed, identification accuracy, and noise robustness compared to conventional CNNs and standard attention baselines under intense interference.
科研通智能强力驱动
Strongly Powered by AbleSci AI