AI failure analysis of technical elastomers

Identify elastomer damage quickly and easily

Due to the wide variety of damage patterns on elastomer components, it is not always easy to trace the cause of the damage. The Fraunhofer internal project “AI-based damage analysis of technical elastomers – KISTE” discusses the possibilities of objectively evaluating damage cases using artificial intelligence. The resulting process, using simple components as an example, can be transferred to customer-specific applications.

Objectivity and efficiency

Elastomers are exposed to extremely high and complex requirements, such as mechanical and thermal stresses and various media influences. Damage to these components can result from these stresses as well as from aging, deviations in the manufacturing process, and other factors. Damage to components can be analyzed using VDI 3822. However, experience and expertise, as well as similar damage patterns with different causes of damage, lead to a subjective assessment of damage in this process. In the KISTE project, this time-consuming and cost-intensive process is being automated and objectified, and the necessary framework conditions (e.g., input parameters, training database, extrapolation potential, etc.) for practical implementation are being explored.

Automated detection of specific causes of damage: Solid database

In the first group step of objectively evaluating the causes of damage to elastomers in a process chain, the initial focus was on investigating the necessary data basis. Since there are typically not enough damaged parts available to train an AI model, synthetically generated damage data was created and various data augmentation methods, such as geometric transformations and color changes, were investigated and applied to determine their effectiveness. This allows a representative number of damage images for different causes of damage to be created, the process requirements to be systematically explored, and the robustness of the model approach to be increased.

Scope and extrapolation potential

In a second step, the process chain was expanded from a classification model based initially on image data alone to a multi-stage, multimodal classification model. To this end, additional scalar input parameters (e.g., material characteristics) were added to increase the accuracy of damage assessment and make the classification process more robust overall. Further research questions deal with the extrapolatability of the process chain to damage examples outside the training data set of the ML models and the integration of explainable AI methods for optimization for future follow-up projects.