O+P Fluidtechnik 11-12/2024

DICHTUNGEN FORSCHUNG UND

DICHTUNGEN FORSCHUNG UND ENTWICKLUNG PEER REVIEWED T4 Model linear poly lasso mlp svr tree T5 Model linear poly lasso mlp svr tree Baseline results of normalized data using all input parameters Training 0.17 0.00 0.31 0.24 0.13 0.13 Testing 0.48 220.01 0.36 0.33 0.32 0.35 3.3 APPLICATION OF REGRESSION MODELS Training 58.52% 100.00% 23.97% 39.24% 66.71% 68.90% The general process of application is stated above in Section 2.2. In this section, the focus is set on preprocessing the data and the framework in which the different ML models are set up, tuned, and subsequently evaluated. As mentioned before the aim of this work is to implement a framework with all necessary modules for further validation land investigation. The framework consists of several components in order to tackle common drawbacks in data-driven algorithms. Starting with the feature scaling of the input data. This results in an equalization of the features in a common band, resolving the issue of different units and magnitudes in the input data. The scaling can be done by applying min-max normalization, converting all input data to the range 0 to 1. Another method for scaling input data is by using a standard scaler, which normalizes each feature by subtracting its mean and dividing by its standard deviation. This transformation ensures that each feature has a mean of 0 and a standard deviation of 1. [14] In a first step, a general baseline for the prediction performance is obtained by considering all nine features introduced in the previous section. One main enhancement to ML algorithms is the reduction of the input vector to the most crucial features, resulting in faster computation and less data requirement. This is achieved by comparing correlation scores between features and only considering uncorrelated features for further use (features with a correlation score lower than a set limit). Principal component analysis (PCA) further facilitates this process by transforming the original features into a new set of orthogonal (uncorrelated) variables, thereby enabling effective dimensionality reduction. [14] Finding the best parameter of a given model requires hyperparameter tuning, since for example the optimal parameters of a neural network can be suboptimal if the network is not deep/ wide enough. This hyperparameter tuning is labor-intensive and therefore is achieved by automated/heuristic/model-based algorithms. These ensure that not only the best parameter, furthermore the best hyperparameter are used. The framework combines the above-mentioned steps to investigate the elastomer specimen volume change for a given liquid/ seal combination. For an exemplary regression model and a scaled dataset, the framework is visualized in Figure 3. Since the MSE R² Testing -23.13% -52886.43% 7.54% 13.11% 13.57% 13.80% Results of normalized input data using selected input parameters. With absolute difference to baseline ΔBL Train 0.19 0.09 0.32 0.20 0.23 0.13 ΔBL 0.02 0.09 0.01 -0.05 0.10 0.01 MSE R² Test 0.33 0.59 0.33 0.30 0.32 0.39 ΔBL -0.15 -219.42 -0.03 -0.03 -0.01 0.04 Train 52.58% 76.99% 20.83% 51.49% 42.50% 67.58% ΔBL -6% -23% -3% 12% -24% -1% Test 13.06% -51.02% 16.84% 21.23% 19.53% -6.27% ΔBL 36% 52835% 9% 8% 6% -20% number of samples in the entire dataset of this work is small (n = 48), statistical uncertainties are introduced during evaluation, making it difficult to compare different models. For this reason, the training and testing process is repeated on different randomly selected subsets, which is known as k-fold cross-validation [14]. This approach is visualized in section (a) in Figure 3. In this case, the dataset is divided into k = 3 separate subsets that do not overlap. For each fold the best performing model is determined using hyperparameter tuning (see (b)) in Figure 3. For this the HalvingGridSearchCV module is employed [14]. Given the best model and the testing set the model is then evaluated by applying the MSE and R² metric (see (c)). This is repeated for every fold and average scores are calculated for each model, enabling comparison between models. 4 RESULTS In this section, the presented framework is evaluated by applying the MSE and R 2 score to the predictions made by the investigated models. As the data set contains only 48 samples, no statement is made about the actual performance of the investigated models. The focus of the results is on validating the hands-on approach of the framework itself. It must be noted that with each addition of new samples, the database expands, leading to more significant results. The baseline score is obtained by utilising 5-fold cross-validation, hyperparameter tuning by HalvingGridSearchCV and normalizing all nine input parameters using the min-max scaler of the scikit learn toolbox [14]. The averaged baseline scores for the investigated models are shown in Table 4. It becomes apparent that for all models, the MSE of the training process is always lower than the prediction (testing) process. The discrepancy in R 2 values between the training and testing phases, coupled with the overall low or even negative testing R 2 values, leads to the conclusion that the models exhibit a high degree of overfitting, failing to adequately represent the underlying patterns. When investigating the correlation matrix of the input parameters, it becomes apparent that several parameters show a high correlation score (see Figure 4). Applying the same models only on selected parameters with a correlation score lower than 0.80, leads to a small decrease in training accuracy (increase of MSE) while positively affecting the testing accuracy to a greater degree, and thus decreasing the overfitting effect. This is also apparent in the increase in R 2 score during the testing process (see Table 5). Another benefit of reducing the number of input parameters is the accelerated computation time. 5 DISCUSSION Although the overall performance of the single methods cannot be evaluated at this time due to the limited number of data samples, the results indicate the potential for a promising framework that addresses common drawbacks in data-based regression tasks and predicts the elastomers volume change based on known fluid parameters. The initial setup of the framework yields baseline performance scores, to which all future investigations and additions can be compared. It is expected that with more data samples the results become more evident. Those data samples are constantly being generated by conducting more experimental immersion tests. 32 O+P Fluidtechnik 2024/11-12 www.oup-fluidtechnik.de

DICHTUNGEN It is shown that the reduction of dimensionality by only selecting uncorrelated input parameters, the effect of overfitting was reduced. Although a loss of input information was introduced, the overall performance was not affected. Since at this point only a first selection of parameters is chosen to indirectly describe the polarity and steric requirement, further, more substantial, parameters could be introduced to describe the underlying effects of elastomer swelling more accurately. It must be noted that at this stage of development, no special attention was paid to preselecting the most suitable hyperparameter-grid for each model on which the method “halving grid search” was applied to determine the best hyperparameters for the given model and data. It is evident, that providing at least a preliminary selection of suitable hyperparameters would enhance each model’s performance. 6 SUMMARY AND OUTLOOK THE INFLUENCE OF ADDITIVES IN TECHNICAL SEALS CAN BE INVESTIGATED IN THE FUTURE In this work, a hands-on application of ML regression models was developed to precisely plan the further experimental investigation of the material compatibility of bio-hybrid fuels and elastomer sealing materials. Based on the results of previously conducted immersion tests and physics related fluid properties as input parameters the elastomers volume change is predicted with several regression models. Within the developed framework common drawbacks of data-based ML methods were addressed. Due to the limited number of samples available at this stage of development, a baseline score of the different models was computed which builds the foundation in the ongoing development process. The framework allows to be easily extended by means of addition of new samples or input parameters. Therefore, further immersion tests are currently being conducted, firstly focusing solely on SRE-NBR 28/SX and different fuels or liquids. Based on the expanded database the actual performance of the models can be tuned and evaluated. Simultaneously additional fluid properties are considered as input parameters allowing for the use of an iterative selection process of few parameters that describe the underlying physics most accurately. This way, the best performing model and input parameters are selected for SRE-NBR 28/SX, enabling the extension of the investigation to other (standard-reference) elastomeric materials. Given the selected model and understanding of how fluid properties affect the material compatibility, also the influence of additives in technical seals can be investigated in the future. 7 ACKNOWLEDGEMENTS Funded by the Deutsche Forschungsgemeinschaft (DFG, German Research Foun- dation) under Germany´s Excellence Strategy – Exzellenzcluster 86 „The Fuelcience Center” ID: 390919832. Literaturverzeichnis: [1] Hofmeister, M.; Schmitz, K.; Laker, J.; Pischinger, S.; Fischer, M. Neue Herausforderungen an Dichtungswerkstoffe im Hinblick auf bio-hybride Kraftstoffe. Mobility, 2022, 7. [2] Hofmeister, M.; Frische, A.; Grunewald, M.; Reddemann, M.; Gütering, C.; Blank, L.M.; Kneer, R.; Schmitz, K. Tribological properties of PTFE sealing materials with regard to bio-hybrid fuels. In: Sealing technology - old school and cutting edge: International Sealing Conference : 21st ISC, Stuttgart, Germany, October 12-13, 2022; Fachverband Fluidtechnik im VDMA e.V: Frankfurt am Main, 2022; pp. 89–104. [3] Makansi, F.; Schmitz, K. Simulation-Based Data Sampling for Condition Monitoring of Fluid Power Drives. IOP Conf. Ser.: Mater. Sci. Eng., 2021, 1097, 12018. [4] Duensing, Y.; Rodas Rivas, A.; Schmitz, K. Machine Learning for failure mode detection in mobile machinery. In: 11. Kolloquium Mobilhydraulik : Karlsruhe, 10. September 2020. Geimer, M., Synek, P.-M., Eds.; KIT Scientific Publishing, 2020. [5] Okolie, J.A. Introduction of Machine Learning and artificial intelligence in biofuel technology. Current Opinion in Green and Sustainable Chemistry, 2024, 100928. [6] Li, J.; Pan, L.; Suvarna, M.; Tong, Y.W.; Wang, X. Fuel properties of hydrochar and pyrochar: Prediction and exploration with machine learning. Applied Energy, 2020, 269, 115166. [7] DIN e.V. DIN EN 590:2022-05: Kraftstoffe - Dieselkraftstoff - Anforderungen und Prüfverfahren; Beuth Verlag GmbH: 10772 Berlin, 2022-05-00. [8] DIN e.V. DIN EN 228:2017-08: Kraftstoffe - Unverbleite Ottokraftstoffe - Anforderungen und Prüfverfahren; Beuth Verlag GmbH: 10772 Berlin, 2017-08. [9] Voelker, S.; Deutz, S.; Burre, J.; Bongartz, D.; Omari, A.; Lehrheuer, B.; Mitsos, A.; Pischinger, S.; Bardow, A.; Assen, N. von der. Blend for all or pure for few? Well-to-wheel life cycle assessment of blending electricity-based OME 3–5 with fossil diesel. Sustainable Energy Fuels, 2022, 6, 1959–1973. [10] Bardow, A.; Voelker, S.; Groll, N.; Bachmann, M.; Mueller, L.; Neumann, M.; Kossioris, T.; Muthyala, P.; Lehrheuer, B.; Hofmeister, M.; Vorholt, A.; Schmitz, K.; Pischinger, S.; Leitner, W. Towards carbon-neutral and clean combustion with hydroformylated Fischer-Tropsch (HyFiT) fuels, 2023. [11] Ackermann, P.; Braun, K.E.; Burkardt, P.; Heger, S.; König, A.; Morsch, P.; Lehrheuer, B.; Surger, M.; Völker, S.; Blank, L.M.; Du, M.; Heufer, K.A.; RoßNickoll, M.; Viell, J.; Aßen, N. von der; Mitsos, A.; Pischinger, S.; Dahmen, M. Designed to Be Green, Economic, and Efficient: A Ketone-Ester-Alcohol-Alkane Blend for Future Spark-Ignition Engines. ChemSusChem, 2021, 14, 5254–5264. [12] Heitzig, S.; Murrenhoff, H.; Weinebeck, A. Investigation of fluid-seal interaction and their prediction based on the Hansen Parameters. O + P : Fluidtechnik für den Maschinen- und Anlagenbau, 2015, 59. [13] Bishop, C.M. Pattern Recognition and Machine Learning, 1st ed.; Springer New York: New York, NY, 2016. [14] Pedregosa, F.; Varoquaux, G.; Gramfort, A.; Michel, V.; Thirion, B.; Grisel, O.; Blondel, M.; Prettenhofer, P.; Weiss, R.; Dubourg, V.; Vanderplas, J.; Passos, A.; Cournapeau, D.; Brucher, M.; Perrot, M.; Duchesnay, É. Scikit-learn: Machine Learning in Python. Journal of Machine Learning Research, 2011, 12, 2825–2830. [15] DIN e.V. DIN ISO 1817:2016-11: Rubber vulcanized or thermoplastic - Determination of the effect of liquids; Beuth Verlag GmbH: 10772 Berlin, 2016-11. [16] DIN e.V. DIN ISO 13226:2021-06: Elastomere - Standard-Referenz-Elastomere (SREs) zur Charakterisierung des Verhaltens von Flüssigkeiten auf Elastomere; Beuth Verlag GmbH: 10772 Berlin, 06.2021. [17] DIN e.V. DIN 53521:1987-11: Prüfung von Kautschuk und Elastomeren - Bestimmung des Verhaltens gegen Flüssigkeiten, Dämpfe und Gase; Beuth Verlag GmbH: Berlin, 11.1987. [18] Kier, L.B. Quantitation of solvent polarity based on molecular structure. Journal of Pharmaceutical Sciences, 1981, 70, 930–933. [19] Palomar, J.; Torrecilla, J.S.; Lemus, J.; Ferro, V.R.; Rodríguez, F. A COSMORS based guide to analyze/quantify the polarity of ionic liquids and their mixtures with organic cosolvents. Physical chemistry chemical physics : PCCP, 2010, 12, 1991–2000. [20] Hansen, C.M. Hansen solubility parameters: A user’s handbook, 2nd ed.; Taylor & Francis: Boca Raton, 2007. [21] Asfour, A.-F.A.; Saleem, M.; Kee, D. de; Harrison, B.H. Diffusion of saturated hydrocarbons in low density polyethylene (LDPE) films. J of Applied Polymer Sci, 1989, 38, 1503–1514. [22] Hildebrand, J.H.; Scott, R.L. The solubility of nonelectrolytes / Joel H. Hildebrand Robert L. Scott, 3rd ed.; Dover Publ: New York, 1964. [23] Cheméo. Chemical and physical properties from Cheméo. https://www. chemeo.com/ (Accessed April 25, 2024). [24] Linstrom, P. NIST Chemistry WebBook, NIST Standard Reference Database 69, 1997. [25] Wildman, S.A.; Crippen, G.M. Prediction of Physicochemical Parameters by Atomic Contributions. J. Chem. Inf. Comput. Sci., 1999, 39, 868–873. [26] Raschka, S. Model Evaluation, Model Selection, and Algorithm Selection in Machine Learning, November 13, 2018. Autoren: Lukas Boden, M. Sc. , Marius Hofmeister, M. Sc., Faras Brumand-Poor, M. Sc., M. Sc., Linda Pleninger, B.Eng., Univ.-Prof. Dr.-Ing. Katharina Schmitz, all RWTH Aachen University, Institute for Fluid Power Drives and Systems (ifas) www.oup-fluidtechnik.de O+P Fluidtechnik 2024/11-12 33