Determining the lubrication condition of a sliding bearing using acoustic emission and data-based classification
VTT has participated in the EU-funded and Spanish IKERLAN-coordinated INNTERESTING research project.
Sliding bearings are the oldest known bearing solutions. As the power density of machines and components increases, efforts are being made to reduce the space occupied by rolling bearings, making sliding bearings an interesting alternative. For the useful life of sliding bearings, it is important that the bearing lubrication mode is hydrodynamic (HD) or elastohydrodynamic (EHD), and there are no mechanical contacts between the shaft and bearing surfaces. Mechanical contacts between the surfaces leads to mixed lubrication (ML) and potentially to boundary lubrication (BL), which can lead to wear and destruction of the surfaces. The load-carrying capacity of the lubricating film is affected not only by the dimensions, surface quality, rotational speed, and load but also by the pressure, viscosity, and the temperature of the lubricant.
Figure 1. VTT concept for lubrication regime detection in journal bearing.
Interesting research project
VTT has participated in the EU-funded and Spanish IKERLAN-coordinated INNTERESTING research project ( https://doi.org/10.3030/851245 ). In this project VTT has developed, among other things, the data-based classification method (Figure 1) for observing the operation of a hydrodynamic bearing, the prevailing lubrication situation, as well as the disturbances and abnormal operational conditions. In the experimental study, a hydrodynamic bearing test device designed and built by VTT (Figure 2) was utilized, where the bearing and shaft are interchangeable, and the materials can be selected according to each specific investigation. In the experiments, the torque caused by sliding friction on the bearing was measured in relation to the load and sliding speed.
Figure 2. VTT journal bearing test device.
Measurement data and simulated virtual data were utilized to create a data-driven hybrid model.
Measurement of friction in a journal bearing is a relatively unsensitive method for detecting changes in lubrication conditions. Therefore, the hydrodynamic bearing test device was equipped with a sensitive acoustic emission (AE) sensor to measure the intensity and quantity of elastic waves caused by mechanical interactions. By utilizing a broadband AE sensor, information about variations in the frequency content of the elastic waves was also obtained. AE measurements have previously been used, for example, in the detection of fatigue damage, but recent research results have shown that even the contact and elastic yielding of microscopic surface roughness aperities on the bearing surfaces can cause measurable acoustic emission. By exploiting this phenomenon, changes occurring in the lubrication conditions of the hydrodynamic bearing can be observed.
In the study, the digital twin of a journal bearing test rig was utilized by creating a multi-body simulation model (MBS) using Dassault Systemes/Simpack software. The model included elastic element models (FEM-based models) of key components (Figure 4), as well as the lubrication situation of the journal bearing modeled using HD and EHD models. The validation and sensitivity analysis of the simulation model was performed by comparing the simulation results with experimental results. By utilizing both experimental data and virtual representation, information about the pressure and thickness of the lubricant film, as well as indications of contacting between bearing surfaces, can be obtained. By combining simulated and experimental data, a comprehensive understanding of the friction behaviour of the lubricated journal bearing is obtained up to the boundary lubrication situation.
Figure 3. Components of a journal bearing are 1) lubricant, 2) shaft and 3) bearing cylinder. The load is carried by the lubricating film, which has a pressure distribution and minimum thickness. Bearing friction affects opposite to the direction of rotation. The Stribeck curve shows the friction as lubrication mode changes. The figure has been modified from https://www.tribonet.org/wiki/hydrodynamic-lubrication-regime/ (i.e., Mahadeshwara, M.R., Hydrodynamic Lubrication Regime, cited January 2024).
Data-based classification
Measurement data and simulated virtual data were utilized to create a data-driven hybrid model (Figure 1) that can determine the lubrication situation of the journal bearing as the operating conditions change. Mean shift clustering algorithm was employed in the modeling to divide the measured data into different clusters or groups. The goal was to find calculated features from the experimental data that could serve as a basis for forming the clusters based on the lubrication situation.
The selected features were calculated from the acoustic emission (AE) data, based on previous studies and available experimental data. The selected features were as follows:
• Kurtosis is a measure of distribution tailedness. The high kurtosis indicates a high number of outliers. Kurtosis is a well-known parameter in vibration signal analysis in the field of condition monitoring.
• The coefficient of variation is a statistical measure used to assess the relative variability of a data set. It is calculated as the ratio of the standard deviation to the mean.
• The root mean square (RMS) of AE signal is applied to measure the magnitude of a fluctuating quantity, e.g., vibration levels in mechanical systems. It is calculated by taking the square root of the mean of the squared values of a set of data.
• The dimensionless Hersey number is a function of viscosity, rotational speed, load, and the dimensions of the bearing. This term is used in the study and analysis of lubrication and lubricants, particularly in relation to the performance and efficiency of bearings. It can be used for determining the lubrication requirements and characteristics needed for optimal bearing operation.
Figure 4. Elastic MBS-model (Simpack) for lubricated journal bearing. The model takes account, e.g., surface roughness and temperature effects on lubrication properties.
When referring to friction, the root mean square of friction provides a measure of the average frictional force acting on an object. It considers both the magnitude and direction of the frictional forces experienced by the object.
The developed hybrid method enables in situ monitoring of the lubrication mode of the hydrodynamic bearing (Figure 5). Similar type of development work has been carried out in multiple locations in recent years, for example, Mokhtari et al. (2020), as well as König et al. (2021).
Figure 5. The result of databased clustering (friction curve) and the feature vectors from AE-signal. The green color denotes hydrodynamic lubrication, orange denotes transform towards mixed lubrication, and blue represents danger to boundary lubrication. The classification was based on parameters calculated from AE-signal and Hersey number. The friction was not applied as a parameter for data clustering.
Model-based simulation
VTT's approach also utilizes virtual data, i.e., model-based simulation. The aim of future research is to further develop the concept so that the determination of the lubrication situation can be implemented in larger scale devices. Improving the discriminability between alternative anomalies requires stronger utilization of simulation and determination of additional features. The goal is to create a concept that can be used in real-time pilot-scale as well as in final products for monitoring and control of the lubrication situation of hydrodynamic bearings.
Ideal operating range of a sliding bearing
In a plain bearing, the load is carried by a lubricant film that forms between the rotating shaft and the bearing (Figure 3). The pressure within the lubricant film, which differs from the supply pressure of the lubricant, and the thickness of the lubricant film (h) are the most important parameters describing the operating conditions of hydrodynamic sliding bearings. The hydrodynamic lubrication (HD) situation prevails when the lubricant film completely separates the bearing surfaces. In elastohydrodynamic (EHD) lubrication, the lubricant film also separates the bearing surfaces, but as the contact pressure increases, the bearing surfaces undergo elastic deformations. When the lubricant film thickness decreases, there are contacts between the bearing surfaces in addition to fluid lubrication, leading to a mixed lubrication (ML) situation. When the contacts between the bearing surfaces dominate, a boundary lubrication (BL) situation occurs. The ideal operating range for a sliding bearing is within the hydrodynamic lubrication zone, where friction is low. As the operating conditions change to mixed lubrication and further to boundary lubrication, friction increases and the wear and damage of the bearing surfaces become possible. Typically, boundary and mixed lubrication situations occur during the startup and shutdown of equipment, but for many devices, these conditions can also occur during actual operation at low speeds.
acoustic emission
Acoustic emission refers to the propagation of transient elastic waves in a material, caused by rapid energy release in localized point or points. Since acoustic emission is the motion of sound waves, the propagation of acoustic emission in a material can be described using the equations of normal elasticity theory. Acoustic emission can be caused by various phenomena occurring in the material or external impact-like excitations on the material. These phenomena may include phase transformations in certain metals causing mechanical stresses, crack growth or plastic deformation, and contacts between surfaces in lubricated systems, for example. The detection of acoustic emission, i.e., the intensity of the signal, depends on the type of wave, the medium being measured, and the reflection caused by interfaces. Knowledge of special mounting solutions for installing sensors that measure acoustic emission is required, as well as understanding the effect of the sensor and the measurement method used on the measured signal. Acoustic emission is normally measured in the frequency range of 50 kHz to 1 MHz, depending on the type of sensor. Various quantities can be calculated from the signals, which can be used to characterize the phenomena. With a wideband sensor and an efficient measurement card, it is possible to detect deviations and phenomena also in terms of frequency. It is also possible to measure at lower frequencies, but in this case, the increasing impurity of the signals naturally leads to a stronger need for filtering.
Artificial intelligence
Artificial intelligence has been applied for the diagnosis of machine condition for decades, and the number of applications has been growing rapidly recently. The goal of classifying or clustering data into classes or clusters is to maximize the similarity of data points within the same class or cluster, and, on the other hand, minimize the similarity between different classes or clusters. The key difference between classification and clustering methods is the supervision of learning (supervised or unsupervised learning). If the location of data samples within a specific lubrication interval on the Stribeck curve is known, supervised learning methods could be utilized. In the present study there are no clear boundaries between lubrication regimes and, thus, unsupervised machine learning, has been employed to distinguish data clusters. The most common clustering methods can be categorized into centroid-based, density-based, distribution-based, and hierarchical methods. The commonly mentioned clustering method in learning materials, K-means, belongs to the centroid-based methods. It aims to partition the data into a predetermined number of clusters (k). The method chosen for this work is Mean Shift Clustering (MSC) since the Stribeck curve forms a continuum of lubrication intervals that are difficult to separate. MSC was able to achieve data segmentation with the least amount of information in this case. MSC also belongs to the centroid-based methods, where each data point converges towards the centroid of its cluster by iterating the mean-shift function. The selection of the kernel size/bandwidth of the mean-shift function affects the number of clusters (centroids) and must be done with care. Therefore, a good result cannot be completely achieved purely through machine computation.
VTT Technical Research Centre of Finland Ltd is a Finnish, fully state-owned limited liability company. The special duty of VTT as an independent and impartial research centre is to promote the wide-ranging utilisation and commercialisation of research and technology in commerce and society.
SOURCES:
• Eitzen, D., Wadley, H., Acoustic Emission: Establishing the Fundamentals. Journal of Research of the National Bureau of Standards, Vol. 89, No.1, January-February 1984, pp. 75 – 100.
• Fukunaga, K., Hostetler, L., The estimation of the gradient of a density function, with applications in pattern recognition. IEEE Transactions on Information Theory, vol. 21 (1975), no. 1, ss. 32-40. https://doi.org/10.1109/TIT.1975.1055330
• Halme, J., Parikka, R., Tervo, J., Akustinen emissio ja sen kaytto koneiden ja laitteiden monitoroinnissa ja diagnostiikassa. Julkinen tutkimusraportti BVAL73-001063, VTT, 2001. 28 s.
• Konig, F., Marnheineke, J., Jacobs. G., Sous, C., Zuo, Ming, J., Tian, Z., Data-driven wear monitoring for sliding bearings using acoustic emission signals and long short-term memory neural networks. Wear, Vol 476 (2021), pp. 203616 – 1-7. https://doi.org/10.1016/j.wear.2021.203616
• Mokhtari, N., Pelham, J., Nowoisky, S., Bote-Garcia, J-L, Guhmann, C., Friction and Wear Monitoring Methods for Journal Bearings of Geared Turbofans Based on Acoustic Emission Signals and Machine Learning. Lubricants 2020, 8, 29; https://doi:10.3390/lubricants8030029
• Sato, I., Rotating Machinery Diagnosis with AcousticEmission Techniques. Electrical Engineering in Japan, Vol 110(1990), No. 2, ss. 115 – 127.
• Tribonet, https://www.tribonet.org/wiki/journal-bearing/.
Acknowledgements
The work was carried out in the INNTERESTING
(Innovative Future-Proof Testing Methods for Reliable Critical Components in Wind Turbines)project that received funding from the European Union’s Horizon 2020 –Research and Innovation Framework Programme (2014-2020) in the call H2020-LC-SC3-2019-RES under grant agreement No. 851245.
www.innterestingproject.eu
Jyrki Tervo, Jukka Junttila, Mikko Savolainen, Artur Korostavyi, Helena Ronkainen,
Juha Virtanen (VTT) Ville Lämsä (DIMECC OY)
