Live visualization for greater safety

Close inspection of the tuyeres is particularly critical when fuels are injected at high rates in order to prevent hazardous situations, including tuyere blockages that cause the fuel to burn inside the blowpipe and tuyere. A blockage will result in considerable damage and trigger an emergency stop of the blast furnace, leading to high repair and production loss costs. Traditionally, regular visual inspections through the tuyere peepholes are the means chosen to prevent such hazards, requiring operators to access the tuyere floor, which is a potentially dangerous area around the blast furnace. Frequent inspections are essential to pinpoint emerging hazards in good time.

The Tuyere Phenomena Detection System (TPDS) greatly facilitates these visual inspections by providing a continuous live visualization of the tuyeres on dedicated screens in the control room, and thus at a safe distance from the tuyere floor, as well as on the cast floor as an additional option. The personnel in charge are thus able to detect harmful or potentially dangerous activities at a very early stage and can take appropriate countermeasures in time.

TPDS hardware mounted on tuyere elbow
Tuyere floor with TPDS hardware installed

Beyond the real-time visualization of the injection process, the TPDS provides automatic detection algorithms that enable the system to identify potentially hazardous phenomena related to the injection lances and tuyeres at an early stage and to issue alerts to the relevant operating personnel. Optionally, the TPDS can also interact autonomously with the blast furnace control system to initiate immediate action to protect the integrity of the plant.

Example of dashboard for tuyere blockage detection

Artificial intelligence takes old processes to a new level

Having established the TPDS on the market and seen it used very successfully in practical applications, Paul Wurth identified potential for further improvement, targeting areas that are typically hard to address with conventional software development methods. By using a mixture of artificial intelligence techniques, an array of different possibilities is opened up.

Up to now, the TPDS phenomena detection system applied conventional computer vision techniques to detect emerging problems related to the tuyere. Classic computer vision techniques usually rely on handcrafted algorithms that have been fine-tuned for a small, controlled set of images. The field performance then depends on the availability of images close enough to the set for which those algorithms were optimized. Although the TPDS hardware provides all means of facilitating mechanical adjustments to the camera view as well as clearing visual obstructions, it is not always possible in practice to have images close to such a controlled set, leading to reduced detection performance. While problems of this type are not necessarily intractable with traditional computer vision approaches, recent advances in machine vision have shown that tasks involving high variability are generally better handled by machine learning techniques.

The TPDS interferes with the worlds of process, equipment and maintenance experts and must therefore have expertise from these domains incorporated into its detection logic and decision-making. This expertise is provided by domain experts from the customer and/or Paul Wurth. All industries combined have a proven record of software development projects that have failed due to a high level of friction at the interface of the domain experts and software development teams, leading to either high costs or suboptimal solutions. Paul Wurth is addressing this issue with its DataXpert platform, providing an intuitive, easy-to-use and powerful low-code suite of tools that allow domain experts to digitize their know-how. Using a low code principle enables people who have no or hardly any coding background to use the tools with little training. DataXpert has been leveraged to incorporate additional domain knowledge when detecting tuyere or lance related phenomena.

Machine vision pipeline as a supplement to the TPDS

In order to combine machine learning and additional domain knowledge, Paul Wurth developed an addition to the existing TPDS system – the machine vision pipeline. This processing pipeline applies semantic segmentation to the camera video stream, and a machine learning algorithm to classify images on the pixel level into the different areas of interest: raceway, tuyere, lance, and injection. Additional transformation steps make use of the semantic understanding of the image, and the information contained in the raw images is converted into a structured format that can be shown as time-series data.

Advanced domain knowledge has been digitized in rule-based logic using RulesXpert, the rule logic engine from the DataXpert platform. The logic fuses sensor data collected from the plant and time-series data generated by the machine learning algorithm, leading to significantly enhanced anomaly detection performance.

These improvements could be achieved with no changes to the underlying camera equipment in the field, allowing for easy upgrades of existing TPDS installations by upgrading the server hardware (adding computational resources) and TPDS software only.

Contact: Marc Weydert, Pierre Van Dorpe, Steve Czarnuch, Aissatou Ndiaye BaMadita Bayer