Human-in-the-Loop Hybrid Intelligent Decision for Automated Strip Steel Flatness Detection

Intelligent decision-making in high-risk industrial scenarios is critical for maximizing production efficiency and ensuring operational safety. Strip flatness detection is a representative task because missed flattening decisions may lead to downstream quality defects. To address this problem, this paper proposes a Human-in-the-Loop Intelligent Decision Framework (HIL-IDF) and develops a Multimodal Hybrid Intelligent Strip Detection method (MHI-SD) as its task-specific implementation. The proposed framework consists of an expert-rule layer, an intelligent model layer, and a risk-aware decision layer. In MHI-SD, the expert-rule layer rapidly handles samples with deterministic characteristics, while the intelligent model layer combines a time-series branch and an image-based auxiliary branch. The time-series branch extracts segment-wise statistical features and performs weighted ensemble learning using k-nearest neighbors (KNN), random forest (RF), and Light Gradient Boosting Machine (LightGBM). The image branch converts paired bending-leveling curves into two-dimensional morphology images and uses a convolutional neural network (CNN) with a support vector machine (SVM) for auxiliary verification. The risk-aware decision layer is designed from a cost-sensitive perspective and implemented as a threshold-based sequential routing mechanism. Low-confidence samples, high-risk predictions, and samples with model disagreement are routed to auxiliary checking or manual verification. Based on historical production-line data, MHI-SD achieves 96.32% accuracy and 95.19% recall. Together with the ablation and threshold-sensitivity analyses, these results support the effectiveness of the proposed risk-aware decision design and suggest its practical potential for strip flatness decision support under the studied production conditions.