TLDR: This research paper details a machine vision-based anomaly detection system for real-time failure prediction in steel rolling mills. It uses industrial cameras and deep learning to monitor visual cues like equipment operation and hot bar motion, integrating this with sensor data to identify failure locations and root causes. A six-month deployment showed high accuracy, real-time performance, and significant cost savings by preventing equipment breakdowns and improving operational reliability.
Operating a steel rolling mill is an intricate dance of massive machinery, complex control systems, and countless sensors. Ensuring continuous, high-quality production demands constant, vigilant monitoring. Traditionally, this monitoring has relied on manual inspections or alarms generated by Programmed Logic Controllers (PLCs) based on sensor data. However, these methods often fall short. Manual observation is prone to human error and costly, while PLCs can be slow, require explicit programming, and lack the generalization needed for diverse process monitoring.
The paper, titled Process Integrated Computer Vision for Real-Time Failure Prediction in Steel Rolling Mill, introduces an innovative solution: a Machine Vision-based Anomaly Detection system designed for real-time failure prediction in steel rolling mills. Developed by Vaibhav Kurrey, Sivakalyan Pujari, and Gagan Raj Gupta, this system integrates industrial cameras to visually monitor critical aspects like equipment operation, alignment, and the movement of hot steel bars along the production line.
How the System Works
At its core, the system processes live video streams using deep learning models deployed on a centralized video server. This setup allows for the prediction of equipment failures and potential process interruptions, significantly reducing unplanned breakdown costs. A key advantage is that this computer vision system operates independently, reducing the computational load on existing industrial process control systems (PLCs). This design ensures seamless scalability across the production line with minimal resources and a reduced risk of computational loading-related process interruptions.
Beyond just visual data, the system simultaneously analyzes sensor data from Data Acquisition (DAQ) systems. This fusion of visual and sensor information enables it to identify both the location and potential root causes of failures with greater accuracy. Operators receive actionable insights, facilitating proactive maintenance planning and ultimately improving reliability, productivity, and profitability.
Addressing Traditional Limitations
Traditional anomaly detection systems in steel mills primarily rely on time-series data from sensors measuring torque, RPM, current, and temperature. While somewhat effective, these systems often lack context awareness and cannot visually assess physical issues such as equipment malfunctions, misalignments, surface defects, or mechanical deformations that often precede major failures. Machine vision, with its ability to capture subtle visual changes, fills this critical gap.
The proposed framework is designed as an independent, real-time anomaly detection and monitoring system. It uses high-speed industrial Baumer cameras strategically placed at critical points. These cameras capture visual streams of hot bar motion, roller alignment, and auxiliary equipment operation, which are then transmitted to the server for real-time deep learning analytics.
Key Detection Capabilities
The system employs YOLO-based deep learning models to identify and track objects of interest. It extracts multiple features from each video frame, including:
- Rod detection and vibration analysis: Tracks the steel rod’s center coordinates over time to detect vibrations exceeding predefined thresholds.
- Flapper tracking: Measures flapper displacement against a static baseline, logging deviations as anomalies.
- Diverter shift measurement: Detects diverter positions and converts pixel-level shifts into millimeter displacements.
- Rod presence and billet duration: Automatically detects rod entry and exit, segmenting billet durations to estimate throughput and identify short metal events.
Crucially, the system integrates auxiliary process signals from PLCs, such as mill operational status and material presence data, fetched from IBA-based data acquisition systems. This sensor data fusion allows for conditional activation of the vision pipeline and dynamic suppression of false alerts, enhancing the system’s reliability.
Also Read:
- Accelerating Edge AI Deployment for Industrial Operations
- Unveiling River Pollution: A Deep Learning and Geometric Approach to Debris Monitoring
Real-World Impact and Scalability
All extracted features and statistics are logged into an InfluxDB time-series database, and an alerting module triggers real-time notifications for issues like excessive rod vibration or diverter misalignment. A FastAPI-based web server streams processed video, allowing operators to visually verify detections and alerts, while Grafana dashboards provide long-term trend visualization.
The system was deployed in a fully operational steel bar rolling mill for six continuous months. During this period, it achieved an average precision ([email protected]) of 94.2% for detection models. Vibration anomalies were detected with a recall of 92.5%, and misalignment events with a recall of 90.7%. The system maintained real-time inference with an average end-to-end latency of 280 ms per frame and processed at 42 FPS.
The operational impact was significant: the mill, which previously faced around 60 cobbles (production stoppages) each month, saw a reduction of nearly 10 cobbles per month. This directly translated to a saving of approximately Rs. 1.15 Crore (11.5 million Indian Rupees) monthly, based on a production loss cost of Rs. 21 lakhs per hour of downtime. These results validate the system’s feasibility and effectiveness, demonstrating its ability to improve anomaly detection accuracy, enhance production reliability, and empower operator decision-making in harsh industrial environments.
