Starting to use artificial intelligence in Foulad Khorasan
Conversation with Fariborz Nashifar, Manager of Technical Services and Support of Foulad Khorasan
The use of artificial intelligence in the steel production process is a significant innovation currently being implemented for the first time at the Khorasan Steel plant. In this regard, we spoke with Mr. Fariborz Nassehi Far, the Technical Services and Support Manager of Khorasan Steel.
In which part of the production process has artificial intelligence been applied?
- In the reduction unit, the prediction of MD and metallic percentage is accurately measured using artificial intelligence. For this purpose, we utilized the latest version of ChatGPT-4, and the error rate in the offline process so far has been below 0.5%.
- Were you able to utilize smart logistics?
- No, our transportation system is not smart; it is implemented in Mobarakeh Steel. There, they have utilized Irancell’s 5G network. However, we do not operate in the same way. But in terms of applying artificial intelligence, I believe we are the only company that has worked on it and achieved results.
- How do the smart control system and sensors function?
- Overall, our control system is intelligent. We have implemented a Japanese system with approximately 2,000 sensors that collect data. Out of these 2,000 sensors, we have selected around 100 sensors for monitoring their data. For example, we have chosen pressure transmitters, flow transmitters, and temperature transmitters, weighted their data, and integrated the factors affecting MD into the smart predictive control system. This system provides us with the MD value and predicts MD for the next six hours.
- To what extent has data transparency impacted your efficiency?
The data is transparent, with a deviation of only 0.5%. This means that if the output MD is 94, the prediction ranges between 93.5 and 94.5. The next step for us is to… The next step for us is to process the data online. We aim to extract the data online, as currently, the data is offline. We have designed a system that can collect the data online. At present, we have stored six months’ worth of data, and we have a system that has stored data from the past ten years.
Soon, this system will be transitioned to real-time (online) mode. With this change, the data will be entered into the system instantly, and predictions will be made immediately. This change will enhance the accuracy of predictions and speed up the production processes. Currently, data is collected and stored from six months to one year ago and is periodically used for future predictions.
The use of neural networks to improve predictions.
To analyze data and improve predictions, deep neural networks (DNNs) are used. The data is entered into the neural networks as time-series data, and more accurate predictive models are built. These networks are capable of automatically simulating and removing unnecessary data to enhance the accuracy of predictions.
Given that neural networks are capable of processing large volumes of data, more accurate predictions for the MD variable in the next six hours are made. Although the accuracy decreases for predictions further than this time frame, the six-hour prediction is currently made with high precision.
Challenges and limitations.
One of the main challenges of this project has been the lack of analysis parameters and the limited number of training samples for the neural networks. This issue has resulted in the prediction models not being fully optimized yet. However, efforts to improve these models and increase the number of training data are ongoing to enhance the accuracy of the predictions.
To improve production processes, pelletizing and sponge iron production are among the most important sections of steel production units. In these units, process data and various outputs are collected and processed to assist in analyzing and optimizing production. The different stages of this process include the following:
✓Output of the pelletizing and sponge iron units: Data related to these units are collected to improve the efficiency of the process.
✓Removal of contaminated data: In this stage, data related to system stoppages and invalid data are removed to ensure more accurate analyses.
✓Collection of information from Midrex systems: Up-to-date information from models and global journals is collected and provided to production experts to update the processes.
✓Reduction of parameters: Using statistical and mathematical methods, the number of system parameters is reduced from five thousand to 300, which results in higher accuracy and facilitates the analysis process.
✓ Training different models: Various models based on statistical methods and machine learning are trained to assist in optimizing the system.
✓Model optimization: After training the models, optimization steps are performed to extract the best possible performance from the system.
Thanks to Fariborz Nassehi Far, the Technical Services and Support Manager of Khorasan Steel, for participating in this interview with the Steel World Review magazine.
Steel World Review
