Science and Research 15 Tribologie + Schmierungstechnik · volume 73 · issue 2/ 2026 DOI 10.24053/ TuS-73-0008 1 Introduction Hydrodynamic journal bearings (JBs) as planetary gear bearings play a crucial role in increasing the torque density and reliability of wind turbine (WT) gearboxes [1]. While they exhibit excellent reliability under hydrodynamic conditions, JBs can experience sudden seizure in the unlikely event of severe overloads, contamination or a lubrication loss [2]. Given the high cost of replacing a WT gearbox, particularly in the offshore sector, avoiding JB damage can reduce operating and maintenance costs, which has a positive impact on the levelized costs of electricity (LCOE). To prevent seizure in planetary JBs (e.g. via a targeted derating and subsequent inspection of the bearings) the aforementioned critical operating conditions need to be detected as they occur. For the detection, condition monitoring systems (CMS) can be applied. For the condition monitoring of JBs numerous approaches exist in the literature, research and even serial application in some industrial fields (e. g. combustion engines [3]). However, for the monitoring of JBs in WT gearboxes no commercially established CMS exists to date. This work addresses the monitoring of hydrodynamic JBs using surface acoustic waves (SAW). This technology has proven to be sensitive to the oil viscosity in a JB [4] and even the operating condition [5]. It is characterized by a particularly high robustness to acoustic disturbance compared to conventional monitoring approaches based on acoustics (e.g. vibration monitoring), since the operating frequency of the SAW can be actively altered. In [6] it was demonstrated, that the propagation behaviour of SAWs is altered in the absence of lubricant in the lubrication gap of a JB or during mixed friction. This causes a measurable modulation in phase shift and amplitude of the SAWs and potentially enables the early Unsupervised machine learning-based approach for real-time monitoring of lubrication loss in journal bearings using surface acoustic waves Thomas Decker, Georg Jacobs, Jonas Betscher, Tim Scholz, Julian Röder* submitted: 29.09.2025 accepted: 30.12.2025 (peer review) Presented at the GfT Conference 2025 With increasing rated power of modern wind turbines, drivetrain components face growing torque density and reliability requirements. Modern wind turbine gearboxes have at least two planetary gear stages with more than three planets in each stage. Planetary journal bearings offer advantages over rolling element bearings due to their compact design and potentially unlimited service life under hydrodynamic lubrication. However, failure causes such as lubrication loss or contamination can lead to sudden seizure and severe component damage. This study presents a method for real-time condition monitoring of journal bearings using surface acoustic waves (SAW) combined with unsupervised machine learning as an approach to prevent sudden damage. A dynamic test procedure is developed on a component test rig to simulate realistic operating conditions, including temporary loss of lubrication. To detect those events in the SAW measurement data the anomaly detection approaches Gaussian mixture model-based clustering and a deep convolutional autoencoder are implemented and compared. While clustering enables basic state differentiation, the autoencoder reliably detects anomalies (e.g. lubrication loss) well before critical increases in friction and temperature. Furthermore, the autoencoder is suitable for real-time application. Thus, the SAW-based monitoring combined with machine learning potentially enables the prevention of sudden damage of planetary journal bearings in wind turbine gearboxes. Keywords Wind turbine drive trains, Journal bearing condition monitoring, Anomaly detection Abstract * Thomas Decker, M.Sc. Prof. Dr.-Ing. Georg Jacobs Jonas Betscher, M.Sc. Tim Scholz, M.Sc. Julian Röder, M.Sc. Chair for Wind Power Drives (RWTH Aachen University), Campus-Boulevard 61, 52074 Aachen Science and Research 16 Tribologie + Schmierungstechnik · volume 73 · issue 2/ 2026 DOI 10.24053/ TuS-73-0008 for condition monitoring of JBs was demonstrated. A correlation between the measured friction torque and the calculated features of the SAW measurements has been demonstrated on a radial JB test rig. Additionally, a multi-layer perceptron neural network was employed to predict the film thickness within the JB [5]. In subsequent work a method for detecting particle contamination and a lubrication loss in a radial JB was presented. This method used an autoencoder to detect these anomalies, but was limited to stationary operating conditions [6]. This work enhances the existing method for real-time capable anomaly detection under dynamic operating conditions. 3 Method The SAW measurement data is obtained from experiments on a JB component test rig. Appropriate unsupervised machine learning methods are selected and implemented. The real-time capability of the best model is subsequently assessed in a third step. a. Experimental test setup The test setup used in this study is identical to the one described in [5]. The bearing material and lubricant used are standard for gearbox applications in WTs. Specifications for the specimen are provided in Table 1. The test rig setup is illustrated in Figure 1. The radial load and rotational speed are applied via a hydraulic actuator and an electric motor, respectively. detection of lubrication loss. These prior investigations were limited to detecting the lubrication loss at constant operating conditions. This is because the measurement results are highly dependent on the operating state, i.e. load and speed. These effects appear superimposed in the measurement, which is why a limit-based detection of lubrication loss is not feasible under dynamic operating conditions. Instead, anomaly detection is a suitable option here. The aim of this work is a method for the detection of the lubrication loss independent of the operating point and dynamic operating conditions. To achieve this, dynamic experiments on a component test rig for JBs are performed and the SAW measurement data is analyzed in detail. 2 State of the research Current research efforts are focused on condition monitoring of JBs, particularly the detection of wear-critical anomalies during operation. Various sensor technologies and algorithms are proposed in the literature for this purpose. Alongside vibration [7], temperature [8] and acoustic emission [9, 10] sensors, using SAW has proven to be a promising alternative: It was found, that SAW sensors can be used to monitor the viscosity [11] and amount [12] of lubricant in a bearing. In addition, SAW proved to be sensitive to changes in the film thickness of a rolling bearing and thus the load [13]. More recently, the suitability of SAW sensors Bearing material 𝐶𝑢𝑆𝑛12𝑁𝑖2 − 𝐶 Bearing diameter 𝐷 = 120 𝑚𝑚 Shaft material 42𝐶𝑟𝑀𝑜4 − 𝑉 Radial bearing clearance 𝑆 = 80 µ𝑚 Bearing width 𝐵 = 30 𝑚𝑚 Lubricant 𝐼𝑆𝑂 𝑉𝐺 320 𝑃𝐴𝑂 Table 1: Specifications of the test setup Figure 1: Experimental setup on the JB component test rig [5] Science and Research 17 Tribologie + Schmierungstechnik · volume 73 · issue 2/ 2026 DOI 10.24053/ TuS-73-0008 Unlike a planetary JB in an actual WT gearbox, the test setup does not experience any gear-related effects such as tilting moments from gear forces or carrier deflection. The specimen is housed within a rotatable enclosure connected to a force transducer, enabling measurement of the friction torque generated during operation. To ensure consistent measurement conditions, the test rig is artificially heated and maintained at a constant surrounding temperature of 50 °C, minimizing temperature-related influences on the SAW signal. b. Surface acoustic waves The SAW sensor system consists of an emitter and a receiver probe. The emitter probe, a piezoelectric transducer, actively excites ultrasonic acoustic waves into the so-called wave guide material (JB bronze material) using the excitation burst shown in Figure 2b. The receiver probe measures the incoming signal (Figure 2c), which consists of several wave groups due to dispersion. The SAW system utilized is identical to the configuration detailed in [5]. The SAW-probes are mounted inside the bearing, as depicted in Figure 2a. The arc length between the SAW-probes is b l ≈ 52 mm. A defined force F is applied to the bearing while the shaft rotates at an angular velocity of ω. The parameterization of the excitation wavelet (EW) (Figure 2b) and feature calculation (FC), based on the measured response signal (Figure 2c), is adapted according to Table 2. The identification of suitable parameterization is done manually. The measured signal is subsequently analyzed and processed further by calculating several features. The amplitude u and the propagation time τ are measured at a particular phase, the gate position t G , marked in Figure 2c. The integral value σ represents the amplitude ener- (b) Figure 2: (a) Schematic depiction of the SAW sensor application in a JB according to [5], (b) excitation wavelet with excitation frequency f E = 200 kHz, (c) exemplary measured response signal corresponding to an excitation frequency of f E = 200 kHz Parameter Symbol Value Unit Frequency of the excitation wavelet (EW) 𝑓 200 𝑘𝐻𝑧 Cycles of the excitation wavelet (EW) 𝑛 5 - Cycle frequency (EW) 𝑓 1000 𝐻𝑧 Gate position (FC) 𝑡 44,5 𝜇𝑠 Time range for integration 1 (FC) 𝑡 , 𝑡 + ∆𝑡 30 − 55 𝜇𝑠 Time range for integration 2 (FC) 𝑡 , 𝑡 + ∆𝑡 58 − 75 𝜇𝑠 Table 2: SAW parameter setting used in this work (a) (b) (c) Science and Research 18 Tribologie + Schmierungstechnik · volume 73 · issue 2/ 2026 DOI 10.24053/ TuS-73-0008 wear, as confirmed by a post-test visual inspection. Figure 3b shows a photograph of the sliding surface of the dismounted bearing: minimal contact on the lower section suggests slight misalignment, while the upper section exhibits extensive damage, including abrasive wear and signs of adhesive wear on the sliding surface. While the quantification of the exact amount of wear that occurred is not within the scope of this work, it is clear that the high friction torque and high temperatures resulting from insufficient lubrication have led to wear in the performed experiment. It is assumed that the periods of operation without active lubrication have led to wear. Detecting this state at an early stage is already of significant value. d. Algorithms for anomaly detection Unsupervised machine learning methods constitute the predominant approach to anomaly detection due to their independence from labelled anomalous data [15]. Models are trained on normal (non-anomalous) data and deviations from the learned distribution are treated as anomalies. The operating points used for model training are shown in Figure 3a. This work investigates clustering and autoencoders. Implementing both enables an evaluation of the complexity required for effective condition monitoring from simple statistical grouping to deep learning-based temporal modelling. A Gaussian Mixture Model (GMM) represents data as a combination of several Gaussian distributions. For anomaly detection, points with a likelihood (probability of belonging to a cluster) below a chosen threshold are flagged as anomalies. Figuratively speaking, these lie outside or at the border of the clusters (see Figure 4a). [16] A Convolutional Autoencoder (CA) (depicted in Figure 4b) learns a compact representation of the input signal. Convolutional layers reduce the temporal dimension while increasing the feature dimensionality. The encoder compresses the input into a low-dimensiogy of the received signal x over a specific time range Δt starting at t 0 as defined in Equation 1. Equation 1 Lastly, the so-called center of energy γ, given by Equation 2, indicates the elapsed time until the measured signal x i decays to half the initially emitted acoustic energy: Equation 2 c. Test procedure The dynamic procedure emulates the real operation of a WT and follows a load curve. The sequence consists of defined operating states (e.g., Start, Stop, Idling, Rated, Noise-reduced, Wind gusts, Emergency Stop) comparable to the definitions that can be found in the IEC 61400 standard for WT bearing design [14]. Each operating point is characterized by specific progressions of sliding speed v S and specific pressure p̅ [1]. Deviations from the load curve occur due to deliberate random variation of v S and p̅ by ± 10 % (± 20 % during gusts) to emulate operational fluctuations and evaluate anomaly detection robustness. Figure 3a illustrates the operating points of the full procedure. To generate wear-critical conditions, the oil supply is intentionally interrupted during the dynamic test procedures at arbitrary points. These interruptions lead to increased friction due to insufficient lubrication and thus to temperature increase. Changes in temperature T and friction torque M Fr are delayed, as a residual amount of oil remains in the system. Prolonged oil-off periods ultimately cause significant rises in both values, indicating the onset of wear-critical states. In a final test, the bearing operates without oil supply until failure. Sustained high friction torque M Fr and temperature T accelerated 𝜎 = 𝑥 ( ∆𝑡 1 ) ( ) 𝛾 = ∑ 𝑡 𝑖 ∙ 𝑥 𝑖2 𝑁𝑖=1 ∑ 𝑥 𝑖2 𝑁𝑖=1 Figure 3: (a) Operating states considered in the test procedure (b) Damaged JB after the component test (a) (b) Science and Research 19 Tribologie + Schmierungstechnik · volume 73 · issue 2/ 2026 DOI 10.24053/ TuS-73-0008 nal latent space. The decoder reconstructs the signal from this representation using transposed convolutions (output). This process forces the model to capture essential patterns and discard noise or irrelevant variations. Anomalies are detected based on the reconstruction error, under the assumption that abnormal patterns cannot be compressed and reconstructed by a model trained only on normal data [15]. The hyperparameters are summarized in Table 5. Unless otherwise stated, standard values are used for initialization and training hyperparameters following best practices for deep convolutional networks. In terms of practical application, real-time capability of the CMS plays a decisive role in preventing failures at an early stage. In this study, real-time capability is defined as the ability of an algorithm to process data faster than it is generated, without data loss. To evaluate this, the computation time per sequence is compared to the sequence’s duration. Additionally, system latency is considered by analysing the time taken to detect a critical state after its occurrence. 4 Results In a preliminary study, input-feature importance is analyzed. Particularly distinct and meaningful results are obtained using the features summarized in Table 3. The performance of the algorithms is evaluated comparing the anomaly label to the complete duration of lubrication loss, as the detection method is expected to identify the entire event. Recall (measuring the models’ ability to detect anomalies, Equation 3) and specificity (measuring the models’ ability to detect normal cases, Equation 4) are used as evaluation metrics, with an emphasis on achieving high specificity to minimize false alarms. Equation 3 Equation 4 The results for both are summarized in Table 4. For both algorithms the recall values are relatively low. Since the recall indicates whether the algorithm can detect the entire period of the lubrication loss (the label is set directly after the oil supply is shut off), it is assumed, that a significant change in the measurement data only occurs after a certain delay, meaning that not the entire event is detected from the beginning. However, the high specificity achieved by the CA indicates that the method is effective at preventing false alarms, making it reliable for identifying critical lubrication loss events. 𝑅𝑒𝑐𝑎𝑙𝑙 = 𝑡𝑟𝑢𝑒 𝑝𝑜𝑠𝑖𝑡𝑖𝑣𝑒𝑠 𝑡𝑟𝑢𝑒 𝑝𝑜𝑠𝑖𝑡𝑖𝑣𝑒 + 𝑓𝑎𝑙𝑠𝑒 𝑛𝑒𝑔𝑎𝑡𝑖𝑣𝑒𝑠 𝑆𝑝𝑒𝑐𝑖𝑓𝑖𝑡𝑦 = 𝑡𝑟𝑢𝑒 𝑛𝑒𝑔𝑎𝑡𝑖𝑣𝑒𝑠 𝑡𝑟𝑢𝑒 𝑛𝑒𝑔𝑎𝑡𝑖𝑣𝑒𝑠 + 𝑓𝑎𝑙𝑠𝑒 𝑝𝑜𝑠𝑖𝑡𝑖𝑣𝑒𝑠 Figure 4: (a) Schematic depiction of the Gaussian Mixture (GMM) clustering Algorithm and (b) visualization of the architecture of a convolutional autoencoder (CA) (a) GMM input features (b) CA input features Amplitude 𝑢 Amplitude 𝑢 Integral 1 𝜎 Center of energy 𝛾 Center of energy 𝛾 Integral 2 𝜎 Sliding speed 𝑣 Sliding speed 𝑣 Propagation time 𝜏 Table 3: Used input features for the GMM approach (a) and the CA (b) (a) (b) Science and Research 20 Tribologie + Schmierungstechnik · volume 73 · issue 2/ 2026 DOI 10.24053/ TuS-73-0008 rors ϵ are found during idling and start/ stop. This is due to highly irregular signal patterns, possibly due to low sliding velocity v S resulting in mixed friction. The second group of anomalies 𝔸 occurs in operating states with higher sliding velocity v S and shows less pronounced reconstruction errors ϵ, indicating that the SAW-signal differs less from rated operation. The smoothed reconstruction error remains mostly below the threshold during oil supply resulting in a correct normal label. Exceptions occur e.g. during wind gusts, which are also labelled as anomalies 𝔸. This can be explained through the stochastic nature of the wind gusts, resulting in signal patterns different to those in training. To solve this issue further tuning of the model’s parameters of the anomaly threshold would be required. The threshold for anomaly detection was set at the 99th percentile of the reconstruction error ε observed while reconstructing the training data. Figure 6 illustrates the effectiveness of this choice. The practical significance of the achieved specificity depends on the actual bearing failure rate. For very low failure rates, even a specificity > 95.9 % may result in a high proportion of false alarms due to the base-rate effect. Since reliable field statistics on the prevalence of detectable damage states in wind turbine plain bearings are currently unavailable, a Bayesian assessment of the positive predictive value cannot be performed. Since the CA algorithm achieved significantly better results according to Table 4, it will be examined in more detail below: In the results of the CA, the detected anomalies A coincide with increasing friction torque M Fr (Figure 5), confirming the method’s general suitability for detecting the critical conditions. Anomalies 𝔸 cease shortly after reactivating the oil supply, as the reformation of the lubrication film stabilizes the SAW-signal. The reconstruction error ϵ is smoothed over a window of 4096 data points, revealing long-term trends while suppressing minor outliers. Two groups of anomalies A are found in periods of lubrication loss. The highest reconstruction er- GMM CA Recall 0,4444 0,6414 Specifity 0,8473 0,9590 Table 4: Summary of the results from the anomaly detection experiments using the GMM and CA approach Figure 5: Dynamic test procedure on the component test rig with two short periods of lubrication loss. Anomaly detection is done using the CA approach Science and Research 21 Tribologie + Schmierungstechnik · volume 73 · issue 2/ 2026 DOI 10.24053/ TuS-73-0008 Most training data points exhibit low reconstruction errors ϵ, while the number of points decreases as the reconstruction error ϵ increases. In contrast, the reconstructed test data displays a higher reconstruction error ϵ. This is expected due to the autoencoder not being trained on it. The number of test data points increases again with rising reconstruction error ϵ near the threshold, consistent with the presence of anomalies with low reconstruction error ϵ. The group of anomalies 𝔸 with high reconstruction error ϵ is found as a more clearly separable anomaly group. Overall, setting the threshold at the 99th percentile of the training reconstruction error effectively separates anomalies from normal behaviour. However, some misclassifications remain unavoidable due to the inherent trade-offs in the threshold selection. a. Comparison and real-time capability of the algorithms To evaluate the time span available for an intervention to prevent bearing damage, the latency between the anomaly detection and the onset of a critical wear state is analyzed. A significant increase in friction torque M Fr is defined as the onset of the critical state. The second derivative M Fr'' serves as a sensitive indicator, as it captures the acceleration of the friction torque increase. A deviation of M Fr'' exceeding one standard deviation σ MFr'' above its mean value m MFr '' is classified as significant, distinguishing real trends from noise according to Equation 5: Equation 5 In view of the aforementioned results, the subsequent investigation of real-time capability is confined to the CA. The CA detects anomalies prior to the critical increase in friction torque M Fr (Equation 5) and temperature T (see Figure 7). Thus, the Autoencoder can identify unusual patterns in the SAW signal regardless of the operating point and potential dynamics before the occurrence of critical operating states. The algorithm re- 𝑀 ≥ 𝑚 + 𝜎 constructs 4.096 seconds of data in a single run, which takes a median of 0.05 seconds (mean: 0.0524 seconds, SD: 0.0102 seconds). This corresponds to a processing speed approximately 80 times faster than real time. As such, the algorithm not only meets real-time requirements but significantly exceeds them, with a latency of just 0.05 seconds. 5 Conclusion and outlook The condition monitoring of JBs in WT gearboxes is subject to current research to improve overall drive train reliability. In this work a SAW-based CMS is used in combination with two anomaly detection approaches to detect lubrication loss in a JB. Clustering distinguishes operating states and elevated friction torque M Fr but is unsuitable for early anomaly detection due to delayed response and low accuracy. In contrast, the autoencoder reliably detects anomalies via increased reconstruction error under lubrication loss. High friction torque M Fr values are consistently found before occurring by over 10 seconds (min. 11.5 s), enabling timely intervention. The method achieves 0.959 specificity and processes 4.096 s of data in just 0.05 s, supporting real-time use. It remains unclear whether initial training is sufficient, or if continuous or periodic retraining is needed due to wear and fatigue over the bearing’s lifetime to maintain long-term model accuracy. Additionally, experiments were conducted under constant temperature condition, which does not reflect real-world variability. Since temperature significantly affects the SAW signals, its influence on model performance must be systematically evaluated. Moreover, to eliminate the possibility for false alarms, the specificity can be improved further. Dynamic reconstruction error thresholds, or alarm triggering after multiple anomaly detections, can improve the overall system performance. In addition, a test of the presented method in a full-size WT gearbox should be done, to show the feasibility of the method under real operating Figure 6: Histogram of the reconstruction error from the CA-approach for the training and test dataset in comparison Science and Research 22 Tribologie + Schmierungstechnik · volume 73 · issue 2/ 2026 DOI 10.24053/ TuS-73-0008 Bearings Using Convolutional Autoencoders. Sensors (Basel, Switzerland) 23 (2023) 22 [8] Paeßens, J., Kratz, K.-H., Gemmeke, T., Kauth, K., Baszenski, T., Lehmann, B. u. Jacobs, G.: Design of a fully integrated sensor system of a plain bearing. Forschung im Ingenieurwesen 88 (2024) 1 [9] König, F., Jacobs, G., Stratmann, A. u. Cornel, D.: Fault detection for sliding bearings using acoustic emission signals and machine learning methods. IOP Conference Series: Materials Science and Engineering 1097 (2021) 1, S. 12013 [10] Mokhtari, N. u. Gühmann, C.: Classification of journal bearing friction states based on acoustic emission signals. tm - Technisches Messen 85 (2018) 6, S. 434-442 [11] Tyreas, G., Dwyer-Joyce, R. S., Notay, R. S. u. Durham, D. J.: Measuring lubricant viscosity at a surface and in a bearing film using shear-horizontal surface acoustic waves. Proceedings of the Institution of Mechanical Engineers, Part J: Journal of Engineering Tribology 236 (2022) 8, S. 1511-1530 [12] Lindner, G., Brückner, C. u. Schmitt, M.: Online Bearing Lubricant Sensing by Mode Conversion of Surface Acoustic Waves. Sensor und Test Conferences 2011 - Sensor proceedings (2011) [13] Chmelar, J., Petr, K., Mossoczy, P. u. Dynybyl, V.: Experimental study of lubrication film monitoring in a roller bearing by utilization of surface acoustic waves. Tribology International 141 (2020), S. 105908 [14] IEC 61400: 2024-07. Design requirements for wind turbine gearboxes [15] Ruff, L., Kauffmann, J. R., Vandermeulen, R. A., Montavon, G., Samek, W., Kloft, M., Dietterich, T. G. u. Müller, K.-R.: A Unifying Review of Deep and Shallow Anomaly Detection (2020) [16] Zeng, H. u. Cheung, Y.-M.: A new feature selection method for Gaussian mixture clustering. Pattern Recognition 42 (2009) 2, S. 243-250Literaturangaben conditions with the noise e.g. from the gear wheels. Finally, practical deployment demands effective post-detection responses, e.g. immediate load shedding to prevent damage and automated diagnostic routines to localize fault causes. Literature [1] Tesini, P., Dzwonczyk-Mertzanis, J. u. Reynders, R.: A new generation of hydrodynamic plain bearings, enabling the next step in gearbox torque density. In: Conference for Wind Power Drives 2023. Konferenzband. Aachen 2023 [2] Marheineke, J.: Ermittlung und Bewertung des Verhaltens von hydrodynamischen Gleitlagern bei. Dissertation. 2024 [3] Hager, G., Laubichler, C., Kiesing, C. u. Bakk, E.: Advanced digital methods for condition monitoring of bearing systems for largebore ICEs. CIMAC Congress 25 - Conference Proceedings (2025) [4] Tyreas, G., Dwyer-Joyce, R. S., Notay, R. S. u. Durham, D. J.: Measuring lubricant viscosity at a surface and in a bearing film using shear-horizontal surface acoustic waves. Proceedings of the Institution of Mechanical Engineers, Part J: Journal of Engineering Tribology 236 (2022) 8, S. 1511-1530 [5] Decker, T., Jacobs, G., Arneth, Philipp, Röder, Julian u. Raddatz, M.: Approach towards the condition monitoring of journal bearings using surface acoustic wave technology. Bearing World Journal Vol. 8 (2025) [6] Decker, T., Jacobs, G., Raddatz, M., Röder, J., Betscher, J. u. Arneth, P.: Detection of particle contamination and lubrication outage in journal bearings in wind turbine gearboxes using surface acoustic wave measurements and machine learning. Forschung im Ingenieurwesen 89 (2025) 1 [7] Ates, C., Höfchen, T., Witt, M., Koch, R. u. Bauer, H.-J.: Vibration-Based Wear Condition Estimation of Journal Figure 7: Timeseries of the anomaly detection using the CA-approach Science and Research 23 Tribologie + Schmierungstechnik · volume 73 · issue 2/ 2026 DOI 10.24053/ TuS-73-0008 Appendix Hyperparameter Value Hyperparameter Value Convolutional layers 5 Learning rate 1e-4 Transposed convolutional layers 5 Batch size 32 Filters per layer 𝑛 𝑖𝑚𝑝𝑢𝑡𝑠 ∙ 2 𝑛 𝑙𝑎𝑦𝑒𝑟𝑠 Epochs 100 Loss function 0,7 SSIM + 0,1 MSE + 0.3 Cos- Similarity + 0,2 FFT Kernel size 5 Batch normalization Yes, before first convolution Stride 4 Padding same Inout window size 4096 Activation function Leaky ReLU Training / validation split 80% / 20% Optimizer Adam Early Stopping Patience = 5 Table 5: Hyperparameters of the convolutional autoencoder