*3.2. Drive Technology*

In the field of drive technology, there are several areas of application for using ML for rolling and sliding bearings, seals, brakes, and clutches, which are involved in systems for motion generation and power transmission.

#### 3.2.1. Rolling Bearings

Rolling bearings are among the most important machine elements, locally transmitting large forces via several rolling contacts. The bearing components are also exposed to complex dynamics and friction occurs in numerous contacts influencing the operation. Bearing failures can be of very different nature. Mostly, they are longer lasting processes between first occurrence of damage and fatal failure. However, damage to rolling bearing components can be quickly observed in the operating behavior of machines and systems, for example in the form of increasing friction, heat, vibration, and noise. Therefore, one possible application for ML is condition monitoring and damage detection [65,66]. Most published work was related to vibration theory rather than tribology [67,68], which is why only some representative examples shall be introduced.


**Table 1.** Overview of ML approaches successfully applied in the area of composite materials.


**Table 1.** *Cont.*


**Table 1.** *Cont.*

As such, Subrahmanyam and Sujatha [69] investigated the suitability of two different ANNs, namely multilayered feed forward neural network trained with supervised error back propagation (EBP) technique and an unsupervised adaptive resonance theory-2 (ART2) based neural network, for the diagnosis of local defects in deep groove ball bearings. The input vector consisted of eight parameters that were used to describe the vibration signal and the output was a condition rating for the bearing (good/bad) and, if the condition was classified bad, the defect was pinpointed. The authors concluded from their work that the performance of the ANN with EBP was excellent for recognizing ball bearing states. They reported that defective bearings were distinguished from good ones with 100% confidence, while the ANN had a success rate of over 95% in diagnosing localized defects. The results of the ANN with ART2 were ambivalent: The learning process was about 100 times faster than that of the ANN with EBP and defective bearings also were distinguished from good ones with 100% reliability. Yet, the estimation of localized defects was not satisfactory. Furthermore, Kanai et al. [70] presented a condition monitoring method for ball bearings using both, model-based estimation (MBE) and ANN, to guess the vibration velocity and the defect frequency of the rotor-bearing-system. The authors based their study on a three-layered feed forward neural network trained with EBP, where the input vector consisted of 5 parameters (speed, load, defect volume, radial clearance, number of balls) obtained from rig tests on a self-aligning deep groove ball bearing. According to the authors, the ANN shows satisfactory results compared to MBE and experimental tests.

Apart from condition monitoring, ML approaches have recently been utilized for designing rolling bearing components. Schwarz et al. [71] used different ML methods to

classify possible cage motion modes of rolling bearings and to predict application-related undesired cage instabilities, see Figure 6. The data set was generated from sophisticated rolling bearing dynamics simulations, which were confirmed by means of experimental investigations on a test rig. Based on the simulations, the authors determined metrics that, in combination, reliably characterize the state of the cage condensed in three classes "stable", "unstable" and "circling". They used these metrics to classify cage motion using quadratic discriminant analysis (QDA). QDA is a method of multivariate statistics to separate different classes on the basis of characteristics [16]. It is interesting to note that we could not discover this method in any other article within our literature survey. To predict the class of cage motion, Schwarz et al. applied decision trees as weak learners within an ensemble classification model based on AdaBoostM1 [72] to achieve good results. Furthermore, Wirsching et al. [73] aimed at tailoring the roller face/rib contact in tapered roller bearings. Geometric parameters were sampled by a Latin hypercube sampling (LHS) and the tribological behavior was predicted by means of elastohydrodynamic lubrication (EHL) contact simulations. Key target variables such as pressure, lubricant gap and friction were approximated by a so-called metamodel of optimal prognosis (MOP) [74] and optimization was carried out using an evolutionary algorithm (EA). The MOP fully automatically filtered non-significant variables and various approaches (polynomial regression, moving least squares, isotropic or anisotropic kriging) were trained to derive the most suitable approximation. The applied ML approach provided very good prediction for most geometries and target values, which was reflected in the high prediction coefficients (CoP) in most cases above 90% and the low errors in mostly below 2% of the optimized pairing between the prediction and verification calculations.

**Figure 6.** Global scheme for classifying and predicting rolling bearing cage motion modes based on dynamics simulations and ML following [71].

## 3.2.2. Sliding Bearings

Since the operating behavior of sliding bearings is highly non-linear and depending on numerous parameters, ML methods have been utilized for the analysis and synthesis of the tribosystem. Canbulut et al. [75] analyzed the frictional losses of a hydrostatic slipper

bearing using an ANN fed by experimental test data. Input parameters were the average roughness of the rubbing surfaces, relative velocity, supply pressure, hydrostatic pocket ratio, and capillary tube diameter. Three-layered feed forward neural networks containing 10 neurons in the hidden layer trained with EBP were to be found as suitable. The predictive performance of the ANN was evaluated using six operation cases for the bearing, where an exact match of the ANN predictions with the experimental results was reported. Further, using ANNs, Ünlü et al. [76] analyzed the friction and wear behavior of a radial journal bearing (bronze CuSn10/steel SAE 1050 pairing) under dry and lubricated conditions. The ANN with EBP technique was featured a 3:5:5:3 multilayer architecture for the dry case and 3:4:4:3 for the lubricated case. The input vector was described by time, applied load and rotational speed and the outputs were coefficient of friction, journal and bearing weight loss. Input data were collected from previously published experiments. The ANN predictions show high agreement to the experimental data and the authors stated that such ANNs can effectively reduce the number of future experiments. Furthermore, Moder et al. [77] showed that supervised ML algorithms can be used to predict the lubrication regime of hydrodynamic radial journal bearings based on given torques. Therefore, the torque time series were first analyzed using Fast Fourier Transformation (FFT) and manually assigned to lubrication regimes. Two ML algorithms were used for the classification task: Logistic regression and deep neural networks. Based on their results, the authors concluded that even shallow neural networks as well as logistic regressions are able to reach high accuracy for the given problem. It was indicated that data scaling was essential, while feature scaling, which is often applied in data analysis, was not suitable for the FFT classification. Prost et al. [78] investigated the feasibility of classifying the operating condition (running-in, steady, pre-critical, critical) of a translationally oscillating self-lubricating journal bearing using an ensemble learning algorithm. To this end, the authors applied a semi-supervised random forest classifier (RFC), which was based on the aggregation of a large number of independent decision trees. The RFC was trained with high-resolution force signals from experiments and showed a very high classification accuracy in validation experiments. The authors pointed out, that labeling the data is essential and requires expert knowledge. As this step is very tedious and time-consuming, they suggested a semi-automated process based on principal components analysis and k-means clustering algorithms. Francisco et al. [79] studied how far ML can be used to optimize connecting rod big-end bearings. They combined sophisticated finite element (FE) simulations with a nondominated sorting genetic algorithm, which allowed them to minimize the frictional losses and functioning severity of the bearing by optimizing 10 parameters. The authors concluded that metamodels based on previous simulations and including all relevant parameters allow the optimization of a tribological system in a very time and resource saving way.

#### 3.2.3. Seals

Seals play an important role in mechanical drive technology as they separate lubricants or operating fluids and the environment of the drive train from each other. Contact seals frequently affect the friction behavior in the whole drive train, and they are exposed to wear. Increasing requirements demand more precise descriptions of the tribological behavior of contact seals in design phases as well as condition monitoring [80,81]. Logozzo and Valigi [82] suggested ANNs as an alternative for analytical models to predict friction instabilities and critical angular speeds of face seals during shaft decelerations. The authors studied different feed forward neural networks with 2:*x*:1 architecture (*x* = 6, 8, 10, 12, 15, 16), trained with supervised EBP technique. Thereby, 10 neurons in the hidden layer showed the best training convergence. Input data were collected from experimentally validated tribo-dynamics simulations based on a lumped parameter model with 2 degrees of freedom. The input vector of the ANNs consisted of two parameters (axial and torsional stiffness). The authors pointed out that unlike deterministic models, the ANNs were not able to explain the phenomena of frictional instability but provided a smart

way to define parameters in the design phase for the avoidance of frictional instabilities. Yin et al. [83] used a SVM regression to monitor the status of a gas face seal based on acoustic emissions (AE). Input data as well as validation data were collected from rig tests. To generate the representative vectors with satisfactory experimental agreement, the AE power concentration in several key bands within certain durations were used.

#### 3.2.4. Brakes and Clutches

Brakes and clutches are safety-relevant components and have to work reliably even under extreme conditions. They are usually integrated in closed loop control systems, which makes it necessary to describe and optimize the braking and coupling behavior, involving complex squealing and wear phenomena. Accordingly, ML approaches have been applied in this area as well [84,85]. For example, Aleksendri´c et al. [86,87] applied an ANN to model the speed-dependent cold performance of brakes They considered 18 material composition parameters, five manufacturing parameters and three operating condition parameters as inputs. Friction data was collected from test rig experiments. Since it is not known a priori which model provides the best prediction quality, the authors investigated 18 different architectures with five different learning algorithms (LM, Bayesian regulation, resilient back propagation, scaled conjugate gradient and gradient decent). The best prediction results were provided by a 26:8:4:1 double-hidden-layer architecture trained by a Bayesian regulation algorithm. The authors stated that their ANN has shown sufficient flexibility to generalize the influences of unknown types of friction material on their cold performance. The methodology was later extended to predict materials recovery performance [88] and brake wear [89] by the same authors. Basically, the procedure was similar to the work described above and the best prediction results were attained from a single hidden layer ANN (25:5:1) trained by a Bayesian regulation algorithm. Timur and Aydin [90] investigated whether the friction coefficient of brakes can be predicted by means of ML based upon experimental training data (1050 points). Comparing different regression methods (linear, least median squared linear, Gaussian processes, pace, simple linear, isotonic, SVM) and 10-fold cross-validation, they noted that all algorithms showed a correlation coefficient larger than 0.99 and a root mean squared error below 0.01. However, isotonic regression allowed the fastest model building.

The prediction of friction coefficient for automobile brake as well as clutch materials against steel using ML algorithms was also addressed by Senatore et al. [91], who showed how to obtain a comprehensive view on the influence of the main sliding parameters. Based upon experimental data from pin-on-disk tests with varying sliding speed, acceleration and contact pressure (200 data sets), the authors trained two different supervised feed-forward double-hidden-layer EPB ANNs with a 3:6:3:1 architecture for braking and 3:6:7:1 for the clutch material, respectively. The authors concluded that ANNs have confirmed suitability for valid prediction of friction coefficients, with utility being enhanced by significance as well as sensitivity analysis of input parameters. A possible application could be in more accurate friction maps for electronic control purpose. However, the authors also discussed the limitations of the approach, in particular pointing out extrapolation errors. Comparable findings were obtained by Grzegorzek and Scieszka [92], who used a similar methodology (in this case feed-forward EPB ANN with 6:12:1 architecture) to investigate the friction behavior of industrial emergency brakes from 408 data sets. The authors self-described their work as being at a preliminary stage, yet they were able to demonstrate the performance of ANN against various models of multiply regression analysis.

The works in the area of drive technology are summarized in Table 2 according to the subject, the database, the inputs and outputs, and the ML approach.


**Table 2.** Overview of ML approaches successfully applied in the area of drive technology.


**Table 2.** *Cont.*
