Scholarlink

<p>Small and medium-sized enterprises (SMEs) face barriers in adopting artificial intelligence, including limited data availability, restricted computing resources, privacy concerns, and limited trust in black-box outputs. An explainable and privacy-aware decision-support framework is proposed for SME operational decision scenarios involving fault classification, anomaly detection, predictive maintenance, tool-wear prediction, and manufacturing-yield prediction. The framework integrates a hybrid convolutional neural network–long short-term memory model,0. federated averaging for distributed training, SHapley Additive exPlanations for interpretability, and a dynamic human-in-the-loop weighting mechanism for expert oversight. Benchmark-based evaluation across five public SME-relevant datasets reports classification F1-scores from 0.88 to 0.92, including 0.92 on Steel Plates, 0.88 on MIMII, and 0.91 on SECOM. Regression results show RMSE reductions of 29.9% on NASA CMAPSS and 34.2% on PHM Milling relative to the centralized CNN–LSTM baseline. Risk-reduction and return-on-investment analyses are presented as scenario-based projections derived from the proposed decision-chain assumptions rather than field-measured outcomes. Differential privacy is treated as an auditable deployment option; the headline predictive results correspond to the non-DP federated setting unless a separate DP-SGD privacy–utility experiment is reported. The results indicate that the framework can support interpretable and governed AI adoption for resource-constrained SMEs, while field validation, seed-level dispersion reporting, and privacy–utility trade-off analysis remain necessary before operational deployment.</p>

错误：搜索内容不能为空，请输入英文关键词
错误：关键词超出字数限制，请精简
高级检索

Privacy-preserving explainable AI framework for SME decision support using federated CNN-LSTM learning

  • 期刊论文

  • 出版模式：

    开放获取

  • 发表日期：

    2026年7月8日

查看全文
  • Wenjin Xu

摘要

Small and medium-sized enterprises (SMEs) face barriers in adopting artificial intelligence, including limited data availability, restricted computing resources, privacy concerns, and limited trust in black-box outputs. An explainable and privacy-aware decision-support framework is proposed for SME operational decision scenarios involving fault classification, anomaly detection, predictive maintenance, tool-wear prediction, and manufacturing-yield prediction. The framework integrates a hybrid convolutional neural network–long short-term memory model,0. federated averaging for distributed training, SHapley Additive exPlanations for interpretability, and a dynamic human-in-the-loop weighting mechanism for expert oversight. Benchmark-based evaluation across five public SME-relevant datasets reports classification F1-scores from 0.88 to 0.92, including 0.92 on Steel Plates, 0.88 on MIMII, and 0.91 on SECOM. Regression results show RMSE reductions of 29.9% on NASA CMAPSS and 34.2% on PHM Milling relative to the centralized CNN–LSTM baseline. Risk-reduction and return-on-investment analyses are presented as scenario-based projections derived from the proposed decision-chain assumptions rather than field-measured outcomes. Differential privacy is treated as an auditable deployment option; the headline predictive results correspond to the non-DP federated setting unless a separate DP-SGD privacy–utility experiment is reported. The results indicate that the framework can support interpretable and governed AI adoption for resource-constrained SMEs, while field validation, seed-level dispersion reporting, and privacy–utility trade-off analysis remain necessary before operational deployment.