Abstract: Time series classification is a critical technology for intelligent decision-making systems, significantly enhancing system reliability and diagnostic accuracy. However, the task faces two major challenges when dealing with low-value-density time series signals: the inherent sparsity of discriminative features and low signal-to-noise ratio, and the prohibitive computational cost of traditional Neural Architecture Search (NAS) which requires full training of massive candidate models. To address these issues, a low-cost architecture search method was proposed. The discriminative power and robustness of input features were first enhanced through a cross-domain feature enhancement and dynamic selection mechanism. Then, a separable multi-scale search space and a Lightweight Pareto Architecture Search Strategy (LPAS) were designed to automatically balance model accuracy and efficiency. Furthermore, a Multi-dimensional Low-cost Proxy Evaluation Strategy (MLP-ES) was introduced, which rapidly assesses candidate models without complete training by quantifying their feature expressiveness, scalability, and trainability, thereby dramatically reducing computational overhead. Extensive experimental results on the Taiyuan Intelligent Detection and Information Processing (TIDIP) dataset of welding defects and fault diagnosis dataset from the Korea Advanced Institute of Science and Technology (KAIST) demonstrate that high-performance classification models are achieved with an average search time of only 1.5 hours, attaining accuracy rates exceeding 98%. The resulting models also exhibit significantly reduced parameter counts and inference times compared to MnasNet, Quant-NAS, WDD-NAS (Weld Defect Detection Neural Architecture Search) and other comparison methods. The number of parameters is reduced by an average of 19.9%, and the inference time is shortened by an average of 30.8%. These results validate the effectiveness of the approach in providing an efficient and automated architecture search solution for low-value-density time series classification tasks.

Key words: time series signal classification, Neural Architecture Search (NAS), cross-domain feature enhancement, low-cost evaluation, Pareto search

摘要： 时序信号分类是智能决策系统的关键技术，对提升系统可靠性与诊断精度至关重要。针对低价值密度时序信号中判别性特征稀疏、信噪比低的固有难题，以及传统神经架构搜索因需完整训练海量候选模型而导致计算成本高昂的瓶颈，提出廉价架构搜索方法。通过跨域特征增强与动态筛选提升输入特征的判别性与鲁棒性；设计可分离多尺度搜索空间和轻量帕累托架构搜索器(Lightweight Pareto Architecture Search Strategy，LPAS)，实现模型在精度与效率权衡下的自动构建；构建多维廉价代理评估策略(Multi-dimensional Low-cost Proxy Evaluation Strategy，MLP-ES)，通过量化特征表达性、可扩展性与可训练性，在无需完整训练的条件下快速评估候选模型，显著降低计算开销。在太原智能检测与信息处理(Taiyuan Intelligent Detection and Information Processing，TIDIP)焊缝缺陷数据集与韩国科学技术院(Korea Advanced Institute of Science and Technology，KAIST)故障诊断数据集上的实验结果表明，平均仅需约1.5小时搜索即可获得高性能分类模型，准确率均超过98%，且模型参数量与推理时间均显著优于MnasNet、Quant-NAS、WDD-NAS (Weld Defect Detection Neural Architecture Search)等对比方法，参数量平均降低了19.9%，推理时间平均缩短了30.8%，验证了本文方法能够为低价值密度时序信号分类任务提供一个高效、自动化的架构搜索解决方案。

关键词: 时序信号分类, 神经架构搜索, 跨域特征增强, 廉价评估, 帕累托搜索