Abstract
Using machine learning (ML) techniques in general and deep learning techniques in specific needs a certain amount of data often not available in large quantities in technical domains. The manual inspection of machine tool components and the manual end-of-line check of products are labor-intensive tasks in industrial applications that companies often want to automate. To automate classification processes and develop reliable and robust machine learning-based classification and wear prognostics models, one needs real-world datasets to train and test the models. The dataset is available under https://doi.org/10.5445/IR/1000129520.
Introduction
The automation of industrial inspection processes is a critical goal for modern manufacturing. Manual inspection of components like ball screw drives (BSDs) is labor-intensive, costly, and prone to human error. While deep learning has shown immense success in general image recognition, its application in technical domains is hindered by the scarcity of high-quality, labeled datasets. Metallic surfaces present unique challenges, including high reflectivity and significant variance in defect appearance. This paper introduces a comprehensive, real-world dataset for surface defect detection on BSDs, specifically focusing on "pitting" wear.
Methodology
To generate realistic wear data, researchers at the Karlsruhe Institute of Technology (KIT) utilized a specialized test bench capable of applying controlled 12kN loads to five spindles simultaneously. A primary challenge in imaging metallic surfaces is managing reflections. To address this, a custom sensor suite was developed.
The system consists of a Raspberry Pi V2 camera integrated into a 3D-printed housing. LED strips provide indirect, diffuse lighting through a "diffusor" to minimize harsh glares, ensuring high-quality image acquisition suitable for machine learning models.
Dataset Description
The released dataset comprises three main components:
- Classification Dataset: 21,853 balanced RGB images (150x150 pixels) categorized into "pitting" and "no pitting."
- Detection/Segmentation Dataset: 1,104 high-resolution images with 394 detailed instance segmentation annotations.
- Wear Progression Sequences: 27 temporal sequences tracking the evolution of specific defects over time.
A key feature of this dataset is the inclusion of "pollution"—soil and metallic debris common in factory environments. This inclusion ensures that models trained on this data are robust to real-world industrial noise. Furthermore, the wear progression sequences enable the development of prognostic models for predicting the remaining useful life of components.
Baseline Results and Discussion
A baseline model using the Mask R-CNN architecture with an Inception ResNet v2 backbone (pretrained on the COCO dataset) was implemented. The model achieved a mean Intersection over Union (mIoU) of 0.316. Analysis indicated that while large defects are reliably detected, the model struggles with "tiny" early-stage pitting and images heavily affected by pollution. This highlights the need for specialized architectures tailored to small-scale industrial defects.
Conclusion
The KIT surface defect dataset provides a vital benchmark for the industrial AI community. By bridging the gap between laboratory conditions and real-world factory floors, it facilitates the development of more reliable automated inspection and wear prognosis systems.
Summary updated