TL;DR:
- Deep learning model developed by Fraunhofer IWM and Schaeffler Technologies enables an objective assessment of steel grain size for rolling bearings.
- Traditional visual inspections by metallographers are subjective and prone to errors, leading to inconsistencies.
- The deep learning model offers high accuracy, ideal reproducibility and can assess large component areas.
- Grain size significantly affects steel’s mechanical properties, with smaller grains resulting in stronger steel.
- The model can distinguish between martensitic and bainitic states and different steel alloys.
- Implementation in the industrial setting promises AI-based and automated defect detection in rolling bearings.
- This model’s potential extends to safety-critical components in various industries.
Main AI News:
In the realm of industrial engineering, rolling bearings play a pivotal role in countless applications, ranging from colossal wind turbines to the smallest of electric toothbrushes. These bearings, predominantly composed of steel components, undergo meticulous scrutiny for quality and application suitability. One of the fundamental determinants of steel performance lies in the grain size, a critical factor that has, until now, been evaluated by metallographers through subjective and error-prone visual inspections.
Breaking free from the limitations of subjective assessment, researchers at the Fraunhofer Institute for Mechanics of Materials IWM, in close collaboration with Schaeffler Technologies AG & Co. KG, have introduced a groundbreaking deep learning model. This model offers an automated and objective means of assessing and determining the grain size in steel materials, thus revolutionizing the quality control processes of rolling bearings.
Rolling bearings predominantly employ surface-hardened steels enriched with carbon, a measure adopted to enhance durability and mitigate the risk of failure, fatigue, and critical crack development resulting from cyclic loading – a move that ultimately safeguards against catastrophic accidents.
Within steel alloys, critical microstructural attributes include non-metallic inclusions and larger-than-average grains, both forming during the steel production process and continually evolving throughout the value chain. Of paramount importance is the grain size, which significantly influences steel’s mechanical properties.
To ensure steadfast quality control, researchers at Fraunhofer IWM, in collaboration with Schaeffler Technologies AG & Co. KG, have introduced a deep learning model tailored for the precise determination of grain size in martensitic and bainitic steels. These steels possess a hardened microstructure due to rapid cooling, rendering their grain size assessments indispensable for safety-critical applications.
Traditionally, metallographers, trained experts, have carried out visual inspections to identify defects, primarily focusing on larger grains and other anomalies, where the potential for failure is most pronounced. However, a recent interlaboratory round robin test has uncovered disparities in grain size evaluations conducted by experts, highlighting the inherent subjectivity and occasional inaccuracy of this method, particularly in safety-relevant scenarios. Additionally, the standard inspection procedure, based on limited sample sizes, is susceptible to errors and impractical for comprehensive component assessments.
The deep learning model for grain size determination, conversely, boasts the ability to evaluate arbitrarily large component areas with remarkable precision and consistency. This achievement is the result of training the model on image data meticulously classified by experts, enabling it to recognize and categorize steel microstructures effectively.
A notable innovation lies in the model’s capacity to assess grain size consistently and objectively. Despite potential annotation noise stemming from variations in the assessments of different metallographers, the model’s optimization effectively filters out this noise. Through continuous exposure to images annotated with both overestimations and underestimations of grain size, the model learns to generate an average representation, bolstering its confidence in microstructure assessments.
Dr. Ali Riza Durmaz, a scientist at Fraunhofer IWM, emphasizes the model’s unique requirements, stating that “neither exceptionally clean data nor large volumes of data are required for training.” Durmaz and his team have developed a web application to visualize the model’s results, incorporating explainable artificial intelligence techniques to enhance transparency in the decision-making process.
The significance of grain size in steel becomes evident – the smaller the grains, the greater the steel’s strength. Increased small grain density results in a higher number of grain boundaries, bolstering the component’s resistance to plastic deformation even under substantial loads. Maintaining smooth bearing operation and optimal frictional properties, as well as ensuring energy efficiency, hinges on preserving the steel’s structural integrity.
In addition to grain size determination, the deep learning model can distinguish between martensitic and bainitic states and differentiate between various steel alloys within the 100Cr6 and C56 families. Currently, Schaeffler Technologies is in the process of implementing the model in an industrial setting, offering unprecedented reproducibility in identifying rolling bearing defects through AI-based automation.
The workflow, which encompasses adapting the AI model for specific materials, integrating it with image processing, and embedding it in user-friendly interfaces, holds immense potential for application in various domains. As Dr. Durmaz aptly concludes, “Our deep learning model paves the way for AI-based and automated qualification in situations involving safety-critical components subject to high and cyclic loads, such as electric drive components or vehicle B-pillars.”
Conclusion:
The introduction of the deep learning model for steel material assessment represents a significant advancement in quality control for rolling bearings and safety-critical components across industries. Its accuracy, reproducibility, and automation capabilities have the potential to enhance product reliability and safety, marking a transformative shift in the market’s approach to quality assurance.
