Productivity Enhancement by Prediction of Liquid Steel Breakout during Continuous Casting Process in Manufacturing of Steel Slabs in Steel Plant Using Artificial Neural Network with Backpropagation Algorithms.

Md Obaidullah Ansari,Szymon Wojciechowski,Shashi Prakash Dwivedi,Somnath Chattopadhyaya,Dominik Walczak,Drazan Kozak,Shubham Sharma,Changhe Li,Joyjeet Ghose,Huseyin Cagan Kilinc,Jolanta B Królczyk

doi:10.3390/ma15020670

Abstract

Breakout is one of the major accidents that often arise in the continuous casting shops of steel slabs in Bokaro Steel Plant, Jharkhand, India. Breakouts cause huge capital loss, reduced productivity, and create safety hazards. The existing system is not capable of predicting breakout accurately, as it considers only one process parameter, i.e., thermocouple temperature. The system also generates false alarms. Several other process parameters must also be considered to predict breakout accurately. This work has considered multiple process parameters (casting speed, mold level, thermocouple temperature, and taper/mold) and developed a breakout prediction system (BOPS) for continuous casting of steel slabs. The BOPS is modeled using an artificial neural network with a backpropagation algorithm, which further has been validated by using the Keras format and TensorFlow-based machine learning platforms. This work used the Adam optimizer and binary cross-entropy loss function to predict the liquid breakout in the caster and avoid operator intervention. The experimental results show that the developed model has 100% accuracy for generating an alarm during the actual breakout and thus, completely reduces the false alarm. Apart from the simulation-based validation findings, the investigators have also carried out the field application-based validation test results. This validation further unveiled that this breakout prediction method has a detection ratio of 100%, the frequency of false alarms is 0.113%, and a prediction accuracy ratio of 100%, which was found to be more effective than the existing system used in continuous casting of steel slab. Hence, this methodology enhanced the productivity and quality of the steel slabs and reduced substantial capital loss during the continuous casting of steel slabs. As a result, the presented hybrid algorithm of artificial neural network with backpropagation in breakout prediction does seem to be a more viable, efficient, and cost-effective method, which could also be utilized in the more advanced automated steel-manufacturing plants.

Highlights

The average weight of liquid steel loss breakout by using variation in multiple thermocouple temperatures
Operational logbook data and Pareto chart show that mold breakout depends on various casting process parameters like casting speed, mold level, and mold/taper
The accuracy curve, loss curve, and testing result show that this system successfully predicts all types of breakouts and even reduces to generate false alarm during casting compared to the existing breakout prediction system (BOPS) system in the Bokaro Steel Plant

Summary

Dropout and Batch-Normalization

The Dropout and Batch-Normalization technique has been used to improve the model performance after each of the two hidden layers. The sigmoid activation function is used majorly in the output layers of binary classification problems It gives a value between 0 to 1, which is the probability prediction of the output, while the ReLu function is one of the most important frequently used activation functions in the hidden layers. It gives a better performance than the sigmoid activation function in the neural networks. From the accuracy curve of the model without dropout and batch normalization, it can be deduced that the model is overfitted if dropout and batch normalization is not used

Selection of Optimizer and Loss Function

Representative Field Application-Based Validation Test

49 CR2 208

Findings

Conclusions