The authors of the BSData: Industrial Machine Tool Element Surface Defect Dataset have compiled a collection of 1,104 channel-3 images accompanied by 394 image annotations, all related to the pitting type of surface damage. They address the problem of the manual inspection of machine tool components and the labor-intensive end-of-line product checks in industrial settings. In order to facilitate the automation of classification processes and the development of robust machine learning-based classification and wear prognostics models, the present dataset is provided.

The Ball Screw Drives (BSD) represent fundamental components in the realm of machine tools, serving the purpose of converting rotary motion into linear motion to enable the movement of machine tool axes. BSDs, as a subset of machine tool components, are notably prone to wear and stand as a primary contributor to the unplanned breakdown of machine tools. These components find extensive application not only within the machine tool industry but also in other sectors, including power generation, medicine, automation, and robotics. Given the tribological nature of BSDs (Sommer, Heinz, Schöfer 2018), wear constitutes a critical tribological aspect. Drawing from the classical bathtub curve (Heise 2002) and insights from (Haberkern 1998), (Schopp 2009), and (Spohrer 2019), BSD failures can be attributed to one of three distinct failure modes:

The dataset is generated through experiments conducted on a test bench at the Institute of Production Science at the Karlsruhe Institute of Technology. Here is a visual representation of the test bench alongside the integrated camera systems:

The design of the test bench accommodates the simultaneous wear testing of up to five spindles. These spindles are arranged in a configuration analogous to the arrangement of dots on a standard six-sided die, with the central spindle being the primary spindle connected to the motor. The remaining four spindles are actuated by a chain drive mechanism linked to the central spindle, ensuring uniform operation across all spindles. The spindles employed are standard 32mm diameter spindles, without any specialized treatment or prestress. Each spindle is preloaded with 70% of the manufacturer-specified 𝐶𝑎 (axial load capacity), where the manufacturer’s stipulated value of 𝐶𝑎 ensures safe operation for 106 revolutions. In this case, the selected 𝐶𝑎 is 12kN. The camera system is programmed to capture an image every four hours in this experimental setup. Following each image capture, the spindle is rotated by an additional 22.5°, and an area measuring 150x150 pixels is automatically cropped from the larger image.

The annotations of the dataset have been created using the annotation tool labelme and are available in JSON format, thus allowing for conversion to VOC and COCO formats. All of the images in this dataset originate from two distinct types of BSDs and 27 wear developments wear_dev. One of the two BSD types referenced in this dataset is represented by 69 images with 55 different image sizes, encompassing clean and soiled conditions. The other BSD type is depicted in 325 images, and these images come in two distinct sizes. Importantly, these images depict the evolution of soiling over time since they have been captured continuously. Additionally, the dataset includes the aforementioned 27 pitting development sequences.