1 Introduction and problem statement Approximately 70 % of all failures and malfunctions in hydraulically operated machines can be traced back to contamination of the hydraulic fluid where solid particles represent the most frequent cause of malfunctions. They lead to material abrasion and increased wear on pumps, motors, valves, and cylinders, thereby significantly reducing the durability of these components. A distinction can be made between internally generated and externally introduced particles. Internally generated particles arise e.g., from friction, corrosion, or cavitation within the hydraulic system. Typical examples include wear particles made of steel from pumps and motors, brass or bronze from sliding bearings, and elastomers from seals or guiding elements. Externally introduced particles enter the system from the environment, for instance sand passing through defective seals or entering during maintenance activities. Particularly small and hard particles have the potential to cause severe damage, since the clearance in hydraulic components is usually less than 10 µm. The concentration of such small particles in hydraulic oil is especially high [1] causing problems. The consequences of wear processes are particularly severe in construction machinery, agricultural and forestry machinery, industrial trucks, and machine tools. Undetected wear can lead to unplanned downtime, production losses, and considerable costs. With the increasing use of proportional and servo hydraulics, the sensitivity of hydraulic systems to sustain damage is increasing [2, 3]. Currently, the condition of hydraulic fluids is monitored by taking oil samples from the machines and analysing them in a laboratory. While these analyses provide detailed results, they are time-consuming, costly, and only reflect the condition at the time of sampling. Inline particle counters represent an alternative, as they continuously measure the number and size of particles in the oil. However, these devices often detect water and air bubbles wrongly as solid particles distorting the measurement results. Furthermore, these devices do not capture particle shape, although this parameter provides crucial information about material, wear mechanisms, and possible sources of origin. In recent years, AI based systems have been developed to automatically detect and classify particles in oil [4, 5]. In [4] a system for detecting and classifying wear particles in engine oils was presented that assigns them to five categories: cutting wear particles, sliding wear particles, fatigue wear particles, spherical particles and nonmetallic particles. The training data required for this system were obtained directly from real oil samples by using an optical particle sensor that extracted oil from Science and Research 71 Tribologie + Schmierungstechnik · volume 72 · issue 3-4/ 2025 DOI 10.24053/ TuS-2025-0023 Concept for the generation of single-type wear particles for training the AI-based image processing of a particle sensor Dennis Jess, Andreas Ligocki* Presented at GfT Conference 2025 Most failures of hydraulically operated machines are caused by solid particles that remove material from components within the hydraulic circuit. If wear remains undetected, unplanned downtime and high costs can result. The HydroVision project therefore develops an AI-based particle sensor that detects, counts, and classifies particles in hydraulic oil, enabling interpretation of wear processes. Suitable training data is required, yet previous approaches are unsuitable because they do not allow generation of single-type training data. Consequently, the relationship between particle shape and material cannot be clearly established, and data labelling is hindered. This paper presents a concept for generating singletype training data. In addition, results from oil particle analyses are presented, which form the basis for defining the fine specifications of the particle sensor. Keywords Hydraulic oil; Wear particles; Particle sensor; Image processing; Artificial intelligence; Training data generation; Predictive maintenance Abstract * M. Eng. Dennis Jess (corresponding author) Prof. Dr.-Ing. Andreas Ligocki Ostfalia University of Applied Sciences Salzdahlumer Str. 46/ 48 38302 Wolfenbüttel, Germany Regional Development Fund (ERDF), Ostfalia University of Applied Sciences in Wolfenbüttel in cooperation with a medium-sized company is pursuing an alternative approach. The objective is likewise the development of a particle sensor that, using optical measurement technology and AI-based image processing, detects the smallest particles in oil circuits, counts them, determines their contours and shapes, and classifies them into different particle types. For training the AI-based image processing, exclusively single-type particle samples are generated to reliably investigate the relationship between particle shape and material and to significantly simplify data labelling. This approach allows better interpretation of wear phenomena in mechanical components of hydraulic systems in construction machinery, agricultural and forestry machinery, industrial trucks, and machine tools, and planning maintenance and repair measures more effectively. In the following, the project-specific concept for generating single-type particles is presented. After this the results from the analysis of oil particles, which form the basis for defining the fine specifications of the particle sensor as well as the requirements for the training data are discussed. 2 Phases of the HydroVision project The HydroVision project is divided into five phases, hereinafter referred to as STEPs. These comprise particle analysis and the establishment of a particle database, the development of test benches, the generation of training Science and Research 72 Tribologie + Schmierungstechnik · volume 72 · issue 3-4/ 2025 DOI 10.24053/ TuS-2025-0023 the main lubrication circuit of an engine and captured images of the particles contained in it. From these images individual particle images were extracted and manually labelled. In [5] an AI based system for monitoring marine lubricants was developed that records size, shape, and number of wear particles to assess the condition of hydraulic systems. For this purpose, oil samples were artificially prepared by adding defined quantities of iron, copper, and aluminium particles to the oil. The samples were then passed through a transparent detection chip and recorded under a microscope using a high-resolution camera. Again, the individual particle images were extracted from the recordings and manually annotated. However, both approaches reveal significant weaknesses: In [4], no separation of particles in the oil samples was carried out to create single-type particles, which restrained the establishment of a relationship between particle shape and material. This made data labelling considerably more difficult and reduced the quality of the datasets. In [5], single-type separation was also lacking. Although in principle the relationship between shape and material could have been investigated, the materials were introduced into a single oil sample instead of being examined in separate samples. In addition, only metallic particles were considered, while non-metallic particles such as elastomers, sand particles, or dust were not considered. As part of the three-year HydroVision research project funded by the N-Bank with resources from the European Figure 1: Overview of the phases of the HydroVision project data and development of AI, the realisation and testing of a prototype, and finally the optimisation of the prototype. The schematic workflow of the HydroVision project is shown in Figure 1. A particular focus of this chapter is the presentation of the concept for generating single-type training data, which is primarily implemented in the first three project phases and forms the basis for the development of the particle sensor. STEP 1 - Particle analysis and establishment of a particle database In the first project phase, fundamental investigations are carried out to assess the conditions of hydraulic systems from different machines in real application cases and target customer segments. For this purpose, oil samples are taken, and the contained particles are analysed under a light microscope. The particles are counted and characterised according to their size, shape, and material. All collected data are documented in a central particle database. Based on this database, the particle types to be detected by the sensor under development are defined. The database thus forms the basis for the artificial and single-type generation of particles. Finally, the fine specifications of the particle sensor are defined based on the identified particle types. STEP 2 - Development of test benches In this phase, the focus lies on the development of various test benches. They serve both to generate training data for the AI and to provide an environment for its training and testing. In total, three test benches are set up. 1.) Particle generation test bench: A compact, modular test bench is designed, engineered, and built to artificially generate single-type wear particles from defined materials with specified sizes and shapes, corresponding to the previously defined particle types. This allows the production of e.g., metallic particles resulting from gear or bearing wear, as well as elastomer particles typically generated by the wear of seals or guiding elements in hydraulic cylinders. 2.) Dry test bench: Afterwards, a dry test bench is designed, engineered, and built. It consists of a linear axis and the components of the Opti module. This module includes a CCD camera and a light source arranged opposite each other. The generated particles are fixed on a movable slide, which is guided along the linear axis between the camera and the light source, thereby simulating the flow in a hydraulic system. During this process, images of the particles are captured. The dry test bench eliminates the influence of hydraulic oil and particle movement. Furthermore, in the initial operating phase, it allows rapid adjustment and easy replacement of all components. Figure 2 shows the schematic setup of the dry test bench. 3.) Wet test bench: Finally, a wet test bench is designed, engineered, and built. It consists of several isolated miniature hydraulic circuits in which the oil is circulated by an external peristaltic pump. These circuits are deliberately loaded with the generated particles. The Opti module (camera and light source) can be flexibly connected to the circuits, enabling images of particles to be captured in the oil flow under realistic conditions. This setup ensures that single-type particles can be examined without cross-contamination. Figure 3 shows the schematic setup of the wet test bench. Science and Research 73 Tribologie + Schmierungstechnik · volume 72 · issue 3-4/ 2025 DOI 10.24053/ TuS-2025-0023 Figure 2: Schematic setup of the dry test bench Figure 3: Schematic setup of the wet bench A compact housing with an integrated measurement cell for particle flow is constructed and manufactured for the sensor components. In addition, a service concept for replacing selected components is developed. A study examines suitable installation positions within the machine’s hydraulic circuit to ensure that particle detection is not impaired. Finally, the prototype is tested in a field trial under real operating conditions, and its performance is evaluated. STEP 5 - Optimisation of the prototype In the final phase, the functionality of the particle sensor is validated and optimised. The evaluation during the field test focuses on the extent to which the specified particles can be detected, counted, and assigned to their potential source. Based on these results, the operational limits of the sensor are fixed, and application scenarios are derived. Furthermore, the collected particles provide the foundation for developing precise recommendations for maintenance. 3 Particle analysis Particle analysis represents the first phase of the Hydro- Vision project and is divided into four steps including oil sampling, membrane filter preparation, particle data acquisition, and documentation and evaluation. The objective of the particle analysis is to use the collected data Science and Research 74 Tribologie + Schmierungstechnik · volume 72 · issue 3-4/ 2025 DOI 10.24053/ TuS-2025-0023 STEP 3 - Generation of training data and development of AI The dry and wet test benches are equipped with the optical components and put into operation in this step. The single-type particles generated in the particle generation test bench are introduced into both test benches. In an iterative process, the test benches, optical setup, image acquisition software, and AI-based image processing are continuously refined. After the hardware and software have been verified, particle images are acquired, labelled, and compiled into training and test datasets. Based on these datasets, suitable AI models are trained, tested, and optimised. The insights gained are then used to further adapt the optical setup and the image acquisition software. Figure 4 illustrates different variations of the optical setup with alternative arrangements of light source and CCD camera, variable light sources in terms of wavelength and intensity, as well as the use of lens systems. This approach enables the continuous improvement of the test benches, optical setup, image acquisition software, and AI-based image processing. STEP 4 - Implementation and testing of a prototype The first functional prototype is developed based on the previous results. The prototype is designed as a plugand-play solution that can be integrated into standard hydraulic circuits. Figure 4: Variations of the optical measurement to define the relevant particle types and the fine specifications for the particle sensor to be developed. 3.1 Oil sampling As part of the sampling process, oil samples were taken from forklift trucks, agricultural machines, forestry machines, construction machines, and machine tools. The sampling was carried out in accordance with the guidelines of Oelcheck GmbH [6] and Bureau Veritas GmbH [7] to ensure the representativeness of the oil samples and to enable comparative analysis assignments. The oil samples were extracted from the tank of each machine using a sampling pump. The extraction was performed from the mid-level of the oil, to avoid disturbances from air at the surface or particles deposited at the bottom. At least five oil samples with a volume of 100 ml were taken from each machine. One sample was analysed in the in-house laboratory, another in an external laboratory, and the remaining samples were kept in stock for potential repetition analyses. Each oil sample was assigned a unique identification number. A sampling protocol documented not only the machine data (e.g. machine type, operating hours) but also oil data (e.g. manufacturer, designation) and sampling data (e.g. sampling date, sample volume). In total, oil samples were taken from three CNC lathes, one CNC milling machine, one excavator, two wheel loaders, three forklift trucks, one portal lifting truck, one forest crawler vehicle, two harvesters, and one tractor. Table 1 provides an overview of the machines from which oil samples were taken. 3.2 Preparation of membrane filters The preparation of the membrane filters for the analyses was carried out according to ISO 4407 [8] using a vacuum filtration system. Figure 5 shows the setup of the ap- Science and Research 75 Tribologie + Schmierungstechnik · volume 72 · issue 3-4/ 2025 DOI 10.24053/ TuS-2025-0023 Figure 5: Setup of the vacuum filtration apparatus p No. Machine type Machine manufacturer Machine name Operating hours [h] Oil usage time [h] 1 CNC lathe Taiwan TAKISAWA NEX-108M 41 066 unknown 2 CNC lathe Taiwan TAKISAWA FX-800 27 243 unknown 3 CNC lathe DMG MORI CLX450 3 252 3 252 4 CNC milling machine MAKINO a51nx 41 264 unknown 5 Excavator Öswag Eurocat 140 HVS 3 248 264 6 Wheel loader Schäffer 2026 1 448 143 7 Wheel loader Liebherr L 506 Compact 807 807 8 Forklift truck Hyster-Yale J2.0XN MWB 12 283 11 878 9 Forklift truck Hyster-Yale H3.5FT 3 796 440 10 Forklift truck Hyster-Yale H3.5FT 3 906 1 039 11 Portal lifting truck Combilift MG3796-05 8 339 1 918 12 Forest crawler vehicle Agria-Werke AGRIA 9700e 134 134 13 Harvester Ponsse Ergo A090247 10 681 10 681 14 Harvester Ponsse Ergo A090853 9 828 925 15 Tractor Deutz Agrotron 150 TT3 6 692 700 Table 1: Overview of machines from which oil samples were taken After microscopic analysis, particle characteristics (e.g., area, maximum diameter, minimum diameter, brightness) were exported as CSV files, and the recorded particle images as PNG files from the microscope software. The determined particle counts were recalculated according to DIN 51455 [9] to the total filter area and, if necessary, to a sample volume of 100 ml. Microscope settings and particle size distributions according to ISO 16232 [10] and ISO 4406 [11] were documented in the accompanying protocol. 3.4 Documentation and evaluation All information obtained in the previous phases was documented in a central particle database. The database contains details on the sample, the associated machine and the oil used, information on the selected analysis parameters and results, as well as the characteristics of the individual particles. Table 2 provides an overview of the stored information and the respective data fields. All entries are linked via the sample ID, ensuring complete traceability from general sample information down to the characteristics of individual particles. Particle size distribution The determination of the particle size distribution in the machines was carried out according to the methodology described in Section 3.3. Table 3 provides an overview of the total number of particles detected and their percentage distribution across the individual size classes. The particle size was defined as the respective maximum diameter. Table 3 shows that most particles are in the size range of 5.1 to 15 µm, with an average percentage of 68.6 %. As Science and Research 76 Tribologie + Schmierungstechnik · volume 72 · issue 3-4/ 2025 DOI 10.24053/ TuS-2025-0023 paratus used. The filter medium was a polyester membrane filter with a diameter of 47 mm and a pore size of 1 µm. A separate membrane filter was inserted into the vacuum filtration system for each oil sample. To ensure a uniform particle distribution, the oil sample was first homogenised. The exact sample volume was then measured and documented before being poured into the funnel of the filtration unit. The negative pressure generated in the suction bottle drew the oil sample through the membrane filter, causing the particles to be deposited on its surface. Filtration continued until the membrane filter was completely dry. The prepared membrane filter was labelled with the identification number of the oil sample and stored in a clean Petri dish until microscopic analysis. 3.3 Acquisition of particle data The microscopic particle analysis was carried out in accordance with DIN 51455 [9] and ISO 16232 [10]. The particles deposited on the membrane filters were analysed using the automated VHX-7000 digital microscope (Keyence Deutschland GmbH). The setup of the VHX-7000 digital microscope is shown in Figure 6. According to DIN 51455 [9], the analysis was performed in reflected light mode within a circular area of 35 mm in diameter at the centre of the membrane filter, with a pixel resolution of < 1 µm per pixel. The acquisition of particle data followed ISO 16232 [10], as this standard considers both light and dark particles. Illumination adjustment, particle counting, particle size classification, acquisition of particle characteristics, and particle image capture were performed automatically. Figure 6: Setup of the VHX-7000 digital microscope particle size increases, the number of particles decreases sharply. Only about 0.9 % of all particles are larger than 200 µm. Since the proportion of particles above 200 µm is negligible and mainly consists of fibres, the upper size limit for particle counting is set at 200 µm. As the particle sensor to be developed must determine the oil contamination level according to ISO 4406 [11], like conventional particle counters, the lower size limit for particle counting is set at 4 µm. It is also noticeable that machines with a high number of particles, such as the CNC milling machine (No. 4), the excavator (No. 5), and the tractor (No. 15), show a higher percentage of particles in the size ranges 15.1 to 50 µm. This confirms the assumption in [12] that with increasing operating time and progressive wear, a greater number of larger particles is generated. On this basis, the lower size limit for particle shape recognition is set at 20 µm, since the progression of wear can be reliably assessed from larger particles. Particle materials Figure 7 illustrates example images of typical metal, sand, elastomer, and fibre particles obtained from the analysed machines. In addition, their characteristic optical and morphological features are listed in bullet points. Metal particles are often shiny and exhibit sharp contours. Sand particles appear matte, with sharp edges and a colour range from light to dark brown. Elastomer particles are matte, with soft contours, and often display irregular surfaces or pores. Fibres, in contrast, usually appear elongated and threadlike. Table 4 shows the particle materials detected in the analysed machines. These include steel (St), zinc (Zn), alu- Science and Research 77 Tribologie + Schmierungstechnik · volume 72 · issue 3-4/ 2025 DOI 10.24053/ TuS-2025-0023 Table sheet Contained data fields Sample information table • Sample data: Sample ID, sampling time, sample volume (ml), sampling location • Machine data: Machine type, machine manufacturer, machine designation, operating hours (h), oil usage time (h), application field • Oil data: Oil manufacturer, oil designation, oil type, kinematic oil viscosity (cSt) • Environment: Environmental influences Sample results table • Microscope data: Microscope manufacturer, microscope designation • Filter data: Diameter (mm), extraction area (mm²), flow-through area (mm²), pore size (μm) • Image acquisition parameters: Illumination type, pixel resolution (μm/ pixel), grayscale threshold values • Particle size distribution: Particle size distribution according to ISO 16232 and ISO 4406 Particle data table • Particle information: Particle ID, particle image name • Particle geometry: Maximum diameter (μm), minimum diameter (μm), aspect ratio, area (μm²) • Optical properties: Average brightness (%) Table 2: Overview of tables in the particle database No. Machine type Machine name Total particle count Particle size distribution [%] 5.1 - 15 [μm] 15.1 - 25 [μm] 25.1 - 50 [μm] 50.1 - 100 [μm] 100.1 - 200 [μm] > 200 [μm] 1 CNC lathe NEX-108M 1 226 69.2 16.9 9.2 3.0 0.5 1.1 2 CNC lathe FX-800 3 017 75.8 16.8 5.7 1.0 0.2 0.5 3 CNC lathe CLX450 1 002 71.5 14.6 8.6 3.6 0.5 1.3 4 CNC milling machine a51nx 7 631 53.3 20.8 19.5 5.2 0.5 0.7 5 Excavator Eurocat 140 HVS 6 447 57.3 23.9 16.3 1.9 0.2 0.4 6 Wheel loader 2026 3 067 82.4 10.3 6.0 0.7 0.1 0.5 7 Wheel loader L 506 Compact 1 379 72.6 14.4 8.6 2.4 0.6 1.5 8 Forklift truck J2.0XN MWB 1 723 63.4 17.2 13.3 4.1 0.5 1.3 9 Forklift truck H3.5FT 1 323 67.6 15.0 12.5 3.4 0.4 1.1 10 Forklift truck H3.5FT 1 699 70.6 14.4 10.3 3.1 0.9 0.8 11 Portal lifting truck MG3796-05 1 114 77.9 17.8 3.8 0.4 0.0 0.9 12 Forest crawler vehicle AGRIA 9700e 3 858 93.3 4.3 1.6 0.3 0.1 0.4 13 Harvester Ergo A090247 1 693 63.6 17.0 12.6 4.8 1.0 0.9 14 Harvester Ergo A090853 1 338 58.4 17.1 16.7 4.8 1.3 1.6 15 Tractor Agrotron 150 TT3 32 836 52.0 21.0 19.6 6.0 0.8 0.6 Average of all machines 68.6 16.1 11.0 3.0 0.5 0.9 Table 3: Overview of machines and particle size distributions the tractor (No. 15). Fabric fibres were present in all oil samples. These results reinforce that the particle sensor to be developed must be capable of distinguishing between metal, sand, elastomer, and fibre particles to reliably interpret wear processes. However, the composition of the particle material spectrum strongly depends on the design and operating conditions of the respective hydraulic system. Science and Research 78 Tribologie + Schmierungstechnik · volume 72 · issue 3-4/ 2025 DOI 10.24053/ TuS-2025-0023 minium (Al), brass (Ms), bronze (Br), sand (Sd), elastomers (El), and fabric fibres (Fb). Steel and aluminium particles were detected in all machines as illustrated in Table 4. Zinc, brass, and bronze particles appeared only sporadically. Sand particles were found in about half of the machines, with a significantly higher proportion observed in the excavator (No. 5) and Particle image Magnification × 800 × 800 × 800 × 1 000 Maximum diameter 72 μm 91 μm 84 μm 70 μm Material Metal (Steel) Metal (Zinc) Metal (Aluminium) Metal (Brass) Features • black, reddish • matte, slightly shiny • sharp edges • light gray, silvery, bluish • shiny • sharp edges • light gray, silvery • shiny • sharp edges • golden • shiny • sharp edges Particle image Magnification × 800 × 800 × 800 × 300 Maximum diameter 68 μm 61 μm 68 μm 234 μm Material Metal (Bronze) Sand Elastomer Fabric Fibre Features • brownish • shiny • sharp edges • light to dark brown • matte • sharp edges • matte • soft edges • often contains holes • light to dark • matte • elongated and thin Figure 7: Example images of particles from different materials p No. Machine type Machine name St Zn Al Ms Br Sd El Fb 1 CNC lathe NEX-108M ✔ - ✔ - - - ✔ ✔ 2 CNC lathe FX-800 ✔ - ✔ - - - ✔ ✔ 3 CNC lathe CLX450 ✔ - ✔ - - - - ✔ 4 CNC milling machine a51nx ✔ - ✔ - - - ✔ ✔ 5 Excavator Eurocat 140 HVS ✔ ✔ ✔ - - ✔ ✔ ✔ 6 Wheel loader 2026 ✔ - ✔ - - ✔ - ✔ 7 Wheel loader L 506 Compact ✔ - ✔ - - ✔ - ✔ 8 Forklift truck J2.0XN MWB ✔ - ✔ - - ✔ ✔ ✔ 9 Forklift truck H3.5FT ✔ - ✔ - - - ✔ ✔ 10 Forklift truck H3.5FT ✔ ✔ ✔ - ✔ ✔ - ✔ 11 Portal lifting truck MG3796-05 ✔ - ✔ - - - ✔ ✔ 12 Forest crawler vehicle AGRIA 9700e ✔ - ✔ - - - - ✔ 13 Harvester Ergo A090247 ✔ - ✔ - ✔ ✔ ✔ ✔ 14 Harvester Ergo A090853 ✔ - ✔ - - ✔ ✔ ✔ 15 Tractor Agrotron 150 TT3 ✔ ✔ ✔ ✔ ✔ ✔ ✔ ✔ Table 4: Overview of the machines and the particle materials identified Shape distribution of metallic particles According to [13], metallic particles can be classified into the basic shapes: lamella, splinter, chip, and sphere. Table 5 summarises the characteristic features and formation mechanisms of these shapes. Figure 8 shows exemplary images of lamellar, splinter, and chip-shaped particles illustrating the differences between these particle shapes. According to the earlier defined restrictions, only particles with a size of 20 µm or larger were considered for shape classification. Not every machine provided enough analysable particle images. Therefore the evaluation was limited to the five machines with the highest number of particles in the size range of 20 to 200 µm. From each of these machines, 100 images of metallic particles were randomly selected, analysed, compared with the features listed in Table 5, and their particle shape was determined. Table 6 provides an overview of the Science and Research 79 Tribologie + Schmierungstechnik · volume 72 · issue 3-4/ 2025 DOI 10.24053/ TuS-2025-0023 Particle shape Features Formation Lamella • flat surface • scratch marks on the surface • large surface area relative to thickness • small thickness • irregular edges • sharp edges (compared to splinters less sharp) • surface fatigue (e.g. pittings) • detachment of material layers • repeated loading Splinter • rough surface • irregular edges • many sharp edges • fracture due to overload • crack formation and spalling • impact or shock loading Chip • elongated shape • fibrous, ribbonor needle-shaped • curved or spiral • sharp edges • may have pointed ends • cutting or abrasive wear • sharp edges on hard surfaces penetrating into soft surfaces Sphere • round or approximately spherical • local melting and solidification (e.g. welding, grinding) Table 5: Overview of particle shapes with their features and formation mechanisms Particle image Magnification × 800 × 800 × 1 000 Maximum diameter 68 μm 68 μm 75 μm Material Zinc Aluminium Aluminium Shape Lamella Splinter Chip Features • flat shape • large surface area relative to thickness • less sharp edges compared to splinters • rough surface • irregular shape • many sharp edges and fracture lines • elongated, fibrous or ribbon-like shape • partly curved • thin • sharp edges Figure 8: Example images of particles of different shapes No. Machine type Machine name Particle shape distribution [%] Lamella Splinter Chip Sphere 1 CNC milling machine a51nx 39 37 24 0 2 Excavator Eurocat 140 HVS 58 18 24 0 3 Forklift truck J2.0XN MWB 69 6 25 0 4 Harvester Ergo A090247 58 21 21 0 5 Tractor Agrotron 150 TT3 56 28 16 0 Average of all machines 56 22 22 0 Table 6: Overview of the machines and the shape distributions of metallic particles lamellar, splinter, and chip-shaped particles to reliably interpret wear processes of metallic components providing a basis for developing a particle sensor that enables reliable condition monitoring of hydraulic systems. In the next step, the requirements for the particle generation test bench will be defined. It will then be designed and built. Subsequently, initial experiments will be carried out to produce particles from selected materials with defined sizes and shapes. By varying the experimental parameters, size and shape of the particles can be specifically influenced. The objective is to identify parameter combinations that allow the production of realistic particles. The generated particles will be examined under a light microscope and compared with particles from real oil samples to verify their similarity. These artificially created, realistic single-type particles will be applied to the dry and wet test benches, forming the basis for the image generation (training data for the sensor under development) using the Opti module. References [1] Karberg & Hennemann GmbH & Co. KG: Ratgeber Öl, 2023. Online verfügbar unter: https: / / www.cjc.de/ wp-content/ uploads/ 2023/ 09/ ratgeber -oel_de.pdf (accessed on August 18, 2025) [2] M. Jocanović, S. Andrić, M. Lazarević, D. Lukić: Example of Good Maintenance Practice for Maintaining the Health of a Hydraulic System. In: M. Rackov, R. Mitrović, M. Čavić (eds), Machine and Industrial Design in Mechanical Engineering. KOD 2021, Mechanisms and Machine Science, vol. 109, Springer, Cham, 2022. https: / / doi.org/ 10.1007/ 978-3-030-88465-9_36 [3] V. Karanovic, M. Jocanović, S. Lalos, B. Z. Knezevic: Oil Cleanliness Class Influence on Wear Intensity of Piston- Cylinder Contact Pair Inside of Hydraulic Distribution Valve, DEMI 2015, 12th International Conference on Accomplishments in Mechanical and Industrial Engineering, Banja Luka, May 2015. [4] S. Fan, T. Zhang, X. Guo, A. Wulamu: FFWR-Net: A feature fusion wear particle recognition network for wear particle classification. Journal of Mechanical Science and Technology, vol. 35, no. 4, pp. 1699-1710, Springer, 2021. https: / / doi.org/ 10.1007/ s12206-021-0333-6 [5] C. Bai, J. Ding, H. Zhang, Z. Xu, H. Liu, W. Li, G. Li, Y. Wei, J. Wang: Research on Abrasive Particle Target Detection and Feature Extraction for Marine Lubricating Oil. Journal of Marine Science and Engineering, 12, 677, MDPI, 2024. https: / / doi.org/ 10.3390/ jmse12040677 [6] OELCHECK GmbH: Ölprobenentnahme. Online verfügbar unter: https: / / de.oelcheck.com/ fileadmin/ user_ upload/ gebrauchsanleitungen/ Anleitung_Allgemein.pdf (accessed on August 18, 2025) [7] Bureau Veritas GmbH: Probenentnahme mit der Probenpumpe, 2020. Online verfügbar unter: https: / / oil-testing.de/ wp-content/ uploads/ 2020/ 10/ BV- Probenentnahme-mit-der-Probenpumpe.pdf (accessed on August 18, 2025) Science and Research 80 Tribologie + Schmierungstechnik · volume 72 · issue 3-4/ 2025 DOI 10.24053/ TuS-2025-0023 examined machines and the shape distribution of metallic particles. Table 6 shows the erratic shape distribution of the metallic particles. The largest proportion consists of lamellar particles, followed by splinter and chip particles, which occur at approximately the same frequency. Spherical particles could not be detected. This demonstrates that the particle sensor under development must be capable of reliably identifying and distinguishing lamellar, splinter, and chip particles to accurately interpret wear processes. Furthermore, the results highlight that real oil samples are only of limited suitability for training AI models, as the shape distribution of metallic particles is highly unbalanced. To ensure that the AI can reliably recognize all particle shapes, a balanced representation of particle shapes in the training dataset is required. 4. Conclusion and outlook As the objective of the HydroVision research project, a particle sensor is being developed that uses optical measurement technology and AI-based image processing to detect, count, and classify particles in hydraulic oils, while also identifying their shapes and contours. For the realisation of the AI-based image processing, suitable training data must first be generated. Previous approaches [4, 5], however, did not allow the generation of single-type training data, preventing a clear link between particle shape and material, and thus making data labelling considerably more difficult. This paper therefore presented a concept that enables the production of single-type training data. This approach allows the targeted investigation of the relationship between particle shape and material, it simplifies data labelling, and ensures a high data quality. Particles from real oil samples of construction machinery, agricultural and forestry machinery, industrial trucks, and machine tools were analysed within the project. Based on these analyses, the relevant particle types were exposed, the functional fine specifications of the particle sensor under development were defined and the foundation for implementing the proposed concept for generating single-type training data was laid. From the analysis results, this indicates that real oil samples are only of limited suitability for training AI models, since the shape distribution of metallic particles is highly unbalanced. To ensure that the AI can reliably recognize all particle shapes, a balanced distribution within the training dataset is required. The sensor under development must be able to detect and count particles in the size range of 4 to 200 µm. Particles larger than 20 µm must be identified by shape and classified into the categories metal, sand, elastomer, and fibre. Metal particles must additionally be classified into [8] ISO 4407: Hydraulic fluid power - Fluid contamination - Determination of particulate contamination by the counting method using an optical microscope. Geneva: International Organization for Standardization, 2002. [9] DIN 51455: Flüssige Mineralölerzeugnisse - Bestimmung der Partikelanzahl und Partikelgröße in Ölen. Berlin: Deutsches Institut für Normung, 2020. [10] ISO 16232: Road vehicles - Cleanliness of components of fluid circuits. Geneva: International Organization for Standardization, 2018. [11] ISO 4406: Hydraulic fluid power - Fluids - Method for coding the level of contamination by solid particles. Geneva: International Organization for Standardization, 2021. [12] B. Bhushan (ed.): Modern Tribology Handbook, 2 Vols. CRC Press, Boca Raton, 2000. https: / / doi.org/ 10.1201/ 9780849377877 [13] F. Bauer (ed.): Tribologie - kompakt und praxisnah. 1 st edition, Springer Fachmedien Wiesbaden (Springer Vieweg), Wiesbaden, 2021. https: / / doi.org/ 10.1007/ 978-3-658- 32920-4 Science and Research 81 Tribologie + Schmierungstechnik · volume 72 · issue 3-4/ 2025 DOI 10.24053/ TuS-2025-0023