阈值
自编码
异常检测
人工智能
太赫兹辐射
模式识别(心理学)
噪音(视频)
鉴定(生物学)
材料科学
干扰(通信)
计算机科学
光谱特征
粒子群优化
异常(物理)
算法
重建算法
电子工程
人工神经网络
离群值
预处理器
滤波器(信号处理)
深度学习
迭代重建
降噪
外推法
生物系统
离散余弦变换
分层(地质)
作者
Xinna Jiang,Hongquan Jiang,Maojie Zhang,Yang Liu,Xiaolu Feng,Yuzheng Zhan
标识
DOI:10.1080/10589759.2026.2704925
摘要
This study proposes a THz-TDS-based framework for spectral optimisation and polyethylene (PE) ageing identification. First, a filtering-pooling and peak attention network (FPAN) is developed to mitigate water vapour interference and system noise under conventional conditions. By learning the mapping from low-SNR and high-SNR spectra, FPAN achieves efficient high-fidelity reconstruction of THz time- and frequency-domain data. Second, PE ageing identification is reformulated as an anomaly detection task to address weak ageing-related spectral features and the limited generalisation of regression-based methods. A cyclic 3D Unet++ autoencoder combined with a multi-statistic reconstruction error thresholding strategy (MSRETR_AE-C3D Unet++) is therefore proposed. Three-dimensional re-encoding and cyclic reconstruction enhance the separability of ageing-related variations, while particle swarm optimisation adaptively determines multiple thresholds. Under a safety-priority principle, no missed detections were observed on the current test split while false alarms were reduced. Experiments demonstrate that FPAN outperforms conventional filtering methods and several deep learning models in reconstruction error, SNR improvement, and peak stability. MSRETR_AE-C3D Unet++ also achieved high precision and no false negatives, demonstrating superior overall performance. The framework provide a promising route towards reliable and automated terahertz non-destructive evaluation of PE ageing.
科研通智能强力驱动
Strongly Powered by AbleSci AI