Results of the summarizing process for the arXiv paper: 2206.05818v1

This paper presents a real-time quality control solution for steel-based mass production using machine learning on non-invasive sensor data. Insufficient steel quality in mass production can lead to costly damage to tooling, production downtimes and low-quality products. To address this issue, the authors analyze a high throughput production line of steel-based products where material quality is currently checked using manual destructive testing. However, this method is slow, wasteful and covers only a small fraction of the material. To achieve complete testing coverage, an industrial collaborator developed a contactless and non-invasive electromagnetic sensor that measures all material during production in real-time. The authors' contribution is threefold: firstly, they demonstrate in a controlled experiment that the sensor can distinguish steel with deliberately altered properties; secondly, they fully measure 48 steel coils non-invasively and conduct additional destructive tests on samples to serve as ground truth; they then fit a linear model to predict two key material properties (yield strength and tensile strength) from the non-invasive measurements normally obtained by destructive tests. The performance is evaluated through leave-one-coil-out cross validation. Thirdly, the resulting model is used to analyze the material properties and their relationship with logged product faults on real production data of approximately 108 km of processed material measured with the non-invasive sensor. The model achieves an excellent performance (F3 score of 0.95) predicting when materials run out of specifications for tensile strength. The combination of model predictions and logged product faults shows that if a significant percentage of estimated yield stress values are out of specification there is a high risk of product faults. Overall, this analysis demonstrates promising directions for real time quality control solutions for risk monitoring and fault detection in mass production settings using machine learning on non invasive sensor data. This approach has several advantages over traditional methods such as being automatic, fast and cheap while providing complete testing coverage without damaging the material.

• Steel-based mass production can suffer from insufficient quality, leading to costly damage and low-quality products.
• Manual destructive testing is slow, wasteful, and covers only a small fraction of the material.
• An industrial collaborator developed a contactless and non-invasive electromagnetic sensor that measures all material during production in real-time.
• The authors demonstrate that the sensor can distinguish steel with deliberately altered properties and fit a linear model to predict yield strength and tensile strength from non-invasive measurements obtained by the sensor.
• The resulting model achieves an excellent performance (F3 score of 0.95) predicting when materials run out of specifications for tensile strength.
• If a significant percentage of estimated yield stress values are out of specification, there is a high risk of product faults.
• This approach has several advantages over traditional methods such as being automatic, fast, cheap, and providing complete testing coverage without damaging the material.

Real-Time Quality Control Solutions for Steel-Based Mass Production Using Machine Learning on Non-Invasive Sensor Data

Steel is a widely used material in mass production, but insufficient quality control can lead to costly damage to tooling, production downtimes and low-quality products. To address this issue, researchers have developed a contactless and non-invasive electromagnetic sensor that measures all material during production in real time. This paper presents the authors' research into using machine learning on this data to develop a real-time quality control solution for steel-based mass production.

Background

Currently, steel quality is checked using manual destructive testing which is slow, wasteful and covers only a small fraction of the material. The authors propose an alternative approach: fitting a linear model to predict two key material properties (yield strength and tensile strength) from non-invasive measurements normally obtained by destructive tests. This would provide complete testing coverage without damaging the material while being automatic, fast and cheap.

Experimental Setup

The authors conducted experiments with 48 steel coils where they deliberately altered properties of some of them as well as conducting additional destructive tests on samples to serve as ground truth for their model evaluation. They then fit a linear model to predict yield strength and tensile strength from the non invasive measurements normally obtained by destructive tests. The performance was evaluated through leave one coil out cross validation with an F3 score of 0.95 predicting when materials run out of specifications for tensile strength achieved by the resulting model.

Analysis

The combination of model predictions and logged product faults showed that if a significant percentage of estimated yield stress values are out of specification there is high risk of product faults occurring in real production data consisting approximately 108 km processed material measured with the non invasive sensor system proposed by the authors .

Conclusion

This analysis demonstrates promising directions for real time quality control solutions for risk monitoring and fault detection in mass production settings using machine learning on non invasive sensor data which has several advantages over traditional methods such as being automatic, fast and cheap while providing complete testing coverage without damaging the material