Interpretable Recognition of Fused Magnesium Furnace Working Conditions with Deep Convolutional Stochastic Configuration Networks

Weitao Li, Xinru Zhang, Dianhui Wang, Qianqian Tong, Tianyou ChaiarXiv 2025 · DOI · arXiv

Abstract

Monitoring fused magnesium furnaces is critical for industrial safety and efficiency, yet traditional manual inspection is hazardous and subjective. While deep learning offers automation, standard models suffer from "black-box" opacity and training inefficiencies. This paper presents Deep Convolutional Stochastic Configuration Networks (DCSCNs), a novel framework that employs incremental supervised learning and physically meaningful Gaussian kernels. By integrating a reinforcement learning-based pruning module and an interpretability evaluation component, the proposed model achieves a high recognition accuracy of 92.57% on real-world industrial data while providing human-readable visual explanations. The approach eliminates the need for backpropagation, ensuring fast optimization and a compact model suitable for real-time monitoring.

Introduction

The production of fused magnesia, essential for aerospace and nuclear applications, involves high-temperature furnaces operating under extreme conditions. Identifying abnormal states such as "underburn" or "overheating" is currently a manual task, exposing workers to intense heat and dust. While standard Deep Learning (DL) models can automate this process, they face significant hurdles: the "black box" nature of CNNs hinders operator trust, and traditional backpropagation is computationally expensive and prone to local minima. This research addresses these challenges by developing an interpretable and efficient recognition framework.

Proposed Methodology

The core of the proposed system is the Deep Convolutional Stochastic Configuration Network (DCSCN). Unlike traditional fixed-architecture CNNs, DCSCNs are built incrementally.

Incremental Construction and Gaussian Kernels

The model starts with zero kernels and adds them sequentially, ensuring that recognition error converges step-by-step. Instead of random initialization, the model utilizes Gaussian differential convolution kernels. These kernels are mathematically designed to mimic human perception of textures and edges, making the extracted features inherently more interpretable.

The overall framework of the proposed DCSCN-based recognition system, including interpretability evaluation and adaptive pruning modules.

Optimization and Pruning

A key advantage of this approach is the avoidance of backpropagation. Output weights are calculated using the least squares method, providing significantly faster optimization. To maintain efficiency, a Deep Deterministic Policy Gradient (DDPG) reinforcement learning agent is employed for adaptive pruning. The agent balances a joint reward function considering accuracy, model size, and "interpretable trustworthiness."

Detailed architecture of the DCSCN showing the incremental addition of convolutional layers and feature maps.

Interpretability Evaluation

To quantify interpretability, the authors define an "independent coefficient" for channel feature maps to generate precise Class Activation Maps (CAMs). These maps highlight the specific image regions (e.g., furnace mouth, flame) that triggered the classification, which are then compared against expert-annotated regions of interest using an Intersection over Union (IoU) score.

Experimental Results and Discussion

The model was evaluated on a dataset of 12,000 images from a fused magnesium furnace factory in Liaoning Province, China, covering four conditions: Normal, Underburn, Overheating, and Abnormal Exhaust.

  • Accuracy: The DCSCN achieved a test accuracy of 92.57%, outperforming standard CNNs (90.35%) and Swin Transformers (88.76%).
  • Efficiency: The RL-based pruning significantly reduced parameter size without sacrificing accuracy. The model achieved an inference time of 0.015 seconds, enabling real-time monitoring.
  • Visual Trustworthiness: CAM visualizations confirmed that the model focuses on physically relevant features like the furnace mouth, whereas unpruned models often attend to background noise.

Conclusion

The DCSCN framework provides a robust solution for industrial furnace monitoring by combining high accuracy with inherent interpretability and computational efficiency. By moving beyond traditional backpropagation and incorporating RL-based pruning, the system offers a trustworthy and deployable tool for harsh industrial environments. Future work will explore Stochastic Configuration Machines (SCMs) for even lighter deployment on edge devices and multi-modal data integration.

Summary updated

BibTeX
@misc{li2025interpretable,
  author = {Weitao Li and Xinru Zhang and Dianhui Wang and Qianqian Tong and Tianyou Chai},
  title = {{Interpretable Recognition of Fused Magnesium Furnace Working Conditions with Deep Convolutional Stochastic Configuration Networks}},
  year = {2025},
  eprint = {2501.02740},
  archivePrefix = {arXiv},
  url = {https://arxiv.org/abs/2501.02740},
}
Figure