分析化学2025，Vol.53Issue(12)：2020-2031,12.DOI:10.19756/j.issn.0253-3820.251253
光谱深度学习模型的可解释性研究进展
Research Progress on Interpretability of Spectral Deep Learning Models
摘要
Abstract
Recently,deep learning(DL)has significantly advanced spectral analysis,enhancing precision and generalization through efficient data processing and feature extraction.However,the ″Ablack-box″ nature obscuring the decision process hinders practical application.Consequently,interpretability research has gained attention to improve model transparency,trustworthiness,and scientific reliability.This paper reviews advances in the interpretability of spectral DL models from 2019 to 2025.It outlines feature extraction mechanisms in models like 1D-CNNs and Transformers.The principles,characteristics,and applicability of various interpretability methods are reviewed from perspectives including feature importance evaluation,model reasoning process visualization,and intrinsic interpretability.Feature importance methods encompass gradient-based and perturbation-based approaches.Reasoning process interpretation includes dimensionality reduction visualization and convolutional layer feature maps.Intrinsic interpretability involves incorporating physical constraints and designing interpretable functional modules.Multi-dimensional interpretation combines various methods for a more comprehensive understanding.These approaches improve transparency and reliability while providing a scientific basis for identifying key spectral regions and elucidating mechanisms.Nevertheless,challenges remain,including an incomplete classification framework,a lack of evaluation standards,and balancing transparency with predictive performance.Future directions include developing intrinsically interpretable models with physical constraints,optimizing the performance-interpretability synergy,promoting multi-dimensional applications,establishing evaluation systems,and advancing practical deployment.关键词
深度学习/可解释性/光谱分析/评述Key words
Deep learning/Interpretability/Spectral analysis/Review引用本文复制引用
刘恒钦,孙杰,闵红,安雅睿,刘曙..光谱深度学习模型的可解释性研究进展[J].分析化学,2025,53(12):2020-2031,12.基金项目
中华人民共和国海关总署科技项目(No.2024HK186)资助. Supported by the Scientific Research Project of the General Administration of Customs of the People's Republic of China(No.2024HK186). （No.2024HK186）
