Gas foil bearings (GFBs) are aerodynamic bearings that use ambient gas as a lubricant. Their oil-free operation and high-speed capabilities make them an attractive choice for small turbomachinery such as fuel cell compressors. Despite their wear-free characteristics at desired operating speeds, rotor-bearing contact is inevitable during start-stop cycles or under large, unstable vibrations, which can lead to gradual wear and bearing failure. Therefore, proper lubrication condition monitoring helps prevent such consequences by identifying GFB malfunctions at an early stage. Acoustic emission (AE) measurement offers a suitable basis for such condition monitoring in GFBs. To achieve this, statistical features of the AE signals can be calculated and fed to a classifier that determines the lubrication regime. Training such a classifier requires accurate knowledge of the lift-off speed for state labeling. Electrical contact resistance (ECR) measurement has been used in this study as a lift-off reference. A GFB, accommodating an AE sensor directly on its top foil, is used for monitoring the lubrication regime. The results show a good agreement between the lubrication state determined by the ECR method and the time and frequency-domain features of the corresponding AE signals. This demonstrates the promising potential of real-time lift-off state monitoring using the AE measurement approach.

The contact angle inherently governs the axial-radial load distribution in ball bearings. In many applications, such as Fault Characteristic Frequency calculation and fatigue life ( L 10 ) prediction, this angle is often simplified as a constant initial value, a practice that compromises accuracy. To address this limitation, this paper introduces a Finite Element Analysis (FEA)-based method to determine actual contact angles and investigates their impact on fatigue life prediction. The proposed approach first identifies the major-axis endpoints of the contact ellipse from the FEA-derived contact stress distribution and approximates the contact center as the midpoint. The actual contact angles are then computed through vector operations involving the bearing center, ball center, and contact center. The proposed method is validated against Hertzian contact theory under various simulation conditions, demonstrating strong agreement. Furthermore, the influence of actual contact angles on bearing fatigue life is investigated, showing that incorporating actual contact angles significantly enhances the accuracy of fatigue life estimation.

Gas foil bearings (GFBs) are increasingly utilized in high-speed turbomachinery, where the transition from mixed friction to a fully developed fluid film, termed lift-off, is a critical parameter for operation monitoring and wear assessment. While vibration-based monitoring, particularly using acoustic emission (AE) sensors, offers a means of detecting this transition even in industrial settings, the development of such classification models is limited by the challenges of data labeling. Traditionally, labeling is performed using friction torque (FT) measurements; however, this approach requires specific test rig configurations and can be subject to interpretational variability.In this study, we investigate the utility of contact voltage (CV) measurements as a basis for the automated labeling of AE signals. The CV method utilizes the electrical characteristics of the fluid gap to identify rotor-foil contact, offering a low-latency alternative to traditional mechanical measurements. Using a dataset of thirty experimental runs under varying static loads and speed profiles, we demonstrate a labeling workflow using a simple voltage threshold. A linear Support Vector Machine (SVM) classifier trained on extracted AE features achieved an accuracy of 93.8% on a separate test set. The results suggest that CV-derived labels can effectively capture the transition stages of GFB operation. This approach thus helps lower the barrier to applying vibration-based lift-off monitoring, thereby enabling gas foil bearing condition monitoring in industry.

Large-scale industrial data diagnosability evaluation faces the problems of high computational complexity and insufficient scalability. This paper proposes a framework that combines a down-sampling strategy (DSS) with a result expansion scheme (RES). By preserving the clustering structure on subsets and robustly extending the results to the full dataset, the proposed method significantly reduces the computational cost while preserving diagnostic consistency. Experimental analysis demonstrates that the proposed framework substantially reduces the practical computational and memory burden of diagnosability evaluation under controllable sampling budgets while maintaining strong consistency with full-sample evaluation results. Experimental results on synthetic data and the CWRU industrial bearing dataset confirm that the framework achieves the proposed framework in terms of structure preservation, efficiency improvement, and robustness, demonstrating its practicality and scalability in large-scale industrial scenarios.

Data-driven Remaining Useful Life (RUL) estimation is critical for various industries, yet scarce labeled data poses a significant challenge. Unsupervised Domain Adaptation (UDA) combined with deep learning has emerged as a promising solution by leveraging unlabeled data. However, recent work on deep UDA exhibits notable inconsistencies, including variations in backbone networks and data handling. Such inconsistencies hinder fair comparisons, obscuring if actual advances were made. To address this, we propose CRULE, a comprehensive benchmarking framework that standardizes deep UDA evaluation. CRULE incorporates a unified 1-D CNN backbone architecture, a standardized training scheme with comprehensive hyperparameter tuning, consistent evaluation protocols (both inductive and transductive), and unified performance metrics (RMSE and Score). This ensures fair and reliable comparisons across different UDA methods. Through experiments on three popular RUL datasets, we found, that only one of the evaluated approaches achieves statistically significant improvements over no adaptation. This suggests that deep UDA approaches proposed for RUL estimation may be less reliable under fair evaluation schemes. To catalyze genuine advancements in the field, we open-source CRULE, empowering the research community to develop and consistently benchmark UDA approaches. CRULE is accessible at https://anonymous.4open.science/r/crule-55D1. Note to Practitioners-Predicting the Remaining Useful Life (RUL) of machinery is crucial for minimizing downtime in industries. The challenge of scarce labeled data has led to using Unsupervised Domain Adaptation (UDA) with deep learning, utilizing unlabeled data for better RUL estimation. However, inconsistencies in evaluation methods across studies have clouded the real progress in this field. To address this, our study introduces CRULE, a standardized benchmarking procedure for fair comparison of UDA approaches. Our analysis reveals that while UDA methods often show promise, their performance does not uniformly exceed that of traditional, non-adaptive approaches. CRULE is now openly available, aiming to foster genuine advancements in RUL estimation technologies, benefiting various industries reliant on predictive maintenance.

To achieve both high accuracy and interpretability in battery State-of-Health (SoH) estimation, this study proposes a dynamic time-varying multi-expert fusion network (MEFNet) framework. The framework consists of three specialized experts: a mechanism-based general expert that captures fundamental degradation patterns, an LSTM-based local expert for short-term dynamics, and a Transformer-based global expert for long-term dependencies. These experts are integrated through a novel linear dynamic weighting scheme that adapts to evolving battery health states. This fusion framework balances interpretability and accuracy while accounting for the scarcity of full lifecycle battery data, particularly addressing challenges stemming from limited real-world data collection conditions that typically only cover early-stage operations. The experimental validation demonstrates that the critical end-of-life threshold (SoH = 70% or 80%) typically occurs within the early (0-30%) to middle (30-60%) degradation stages. The proposed MEFNet achieves superior estimation accuracy using only 25% of the lifecycle data, outperforming models trained on complete datasets particularly during early and middle degradation stages.

To enhance safety within the vehicle interior, the New Car Assessment Protocol (NCAP) emphasizes the importance of estimating occupant features, such as age and gender, utilizing integrated sensors like vehicle interior cameras. This work presents a data augmentation strategy designed to enhance the performance of a Convolutional Neural Network (CNN) in estimating the facial age and gender of both the driver and co-driver from vehicle interior images captured in nearinfrared (NIR) and visual (RGB) modalities. The approach was initially tested on publicly available datasets and later evaluated using an in-house dataset derived from a study involving 50 unique test subjects, who acted as drivers and co-drivers within a car. In the in-house dataset, facial images were generated using both coarse and tight cropping strategies to identify the optimal approach. Results indicate that the proposed data augmentation strategy, combined with tighter cropping, significantly enhances gender and age estimation performance in both RGB and NIR modalities while outperforming state of the art methods.

Simulation models and Generative Adversarial Networks (GANs) face challenges in data generation while addressing the data imbalance issue encountered in data-driven fault diagnosis. Traditional simulation models rely on idealized or simplified mechanistic approaches and often fail to capture subtle data variations in complex real-world environments. Moreover, GANs experience inefficient training and slow convergence. To tackle these issues, this study proposes a Physics-informed GAN (PGAN) to replace the conventional random noise with simulated data input to the generator. Additionally, frequency-domain feature discrepancies are incorporated into the generator’s loss function. The PGAN aims to improve training efficiency and the quality of generated data by leveraging the complementary advantages of the two methods. Experimental validation on the Case Western Reserve University (CWRU) bearing dataset proves that PGAN generates realistic vibration signals. Furthermore, the augmented fault dataset significantly improves bearing fault diagnosis accuracy. In conclusion, the proposed PGAN offers an innovative and effective method for generating high-quality fault data, enhancing the diagnostic performance of data-driven models.

Accurate dynamic modeling and fault diagnosis are crucial to ensure the reliability of rotating machinery. PhysicsInformed Neural Networks (PINNs) emerge as a promising approach by integrating physical constraints into deep learning. However, traditional PINNs suffer from structural rigidity and encounter difficulties in embedding complex dynamics. To address these challenges, Particle Swarm Optimization-Improved PINN (PSO-IPINN) is proposed in this paper. First, a parallel structure is employed to extract distinct features, enabling accurate reconstruction of the rolling bearing’s vibration signal through weighted summation. Subsequently, a frequency domain loss function serves as a physical constraint to effectively simulate and predict the dynamic behavior of the rolling bearing. Finally, PSO optimizes both structural parameters and hyperparameters, enhancing prediction accuracy, convergence speed, and generalization. Comparative experiments on outer ring, inner ring, and ball faults confirm that PSO-IPINN outperforms conventional PINNs, achieving higher similarity in the time domain and more accurate Fault Characteristic Frequency (FCF) extraction. This method provides a novel and effective solution for accurately analyzing and predicting rolling bearing vibration characteristics.

The use of machine learning for machine monitoring and fault detection is already quite common to solve some problems within the context of Industry 4.0. However, there are still areas where the use of machine learning techniques is very incipient or almost nonexistent. As an example, one can mention the problem encountered in identifying the lift-off state of shafts in relation to air foil bearings. The main justification for the study’s relevance lies in the absence of a suitable approach for defining the lift-off state, without the need for specially designed test rigs. The state-of–the-art approach needs these test rigs to measure the friction torque of the bearing. The present study employs a convolutional neural network to monitor the lift-off state of a gas foil journal bearing by utilizing experimentally obtained acceleration data, offering a more versatile alternative. The acceleration data from the test rig are analyzed in both the time and frequency domains. The technique used involved converting the signal into vibration images. In all configurations analyzed, the network’s accuracy in identifying the lift-off phenomenon exceeds 90

Gas foil bearings (GFBs) are widely used in high-speed rotating machinery due to their low maintenance requirements, high-speed capabilities, and contamination-free operation. However, challenges such as wear, vibrational, and thermal instabilities require advanced condition monitoring techniques. The present work introduces a sensor-integrating gas foil bearing (SiGFB) designed for real-time monitoring of key state variables, including instantaneous angular speed (IAS), temperature distribution, and friction state. The SiGFB fully incorporates all required electronics: embedded thermocouples, an accelerometer, a microphone, and an acoustic emission sensor, enabling in-situ measurement without compromising its functionality as a bearing. The measurement of acoustic emission in the bearing via a piezoelectric element on the top foil and the measurement of the sound in the gas are also new contributions. The proposed system is evaluated through experimental validation, demonstrating its ability to accurately determine the IAS to within 1 Hz, sound, and current lift-off state. While the temperature measurement is still subject to quantitative uncertainties, the results confirm that the SiGFB provides usable measurement data, laying the groundwork for future autonomous and wireless condition monitoring solutions in high-speed rotating machinery.

This study presents an algorithm for estimating the instantaneous angular speed (IAS) from vibrations in low-power embedded systems, evaluated on the use-case of a gas foil bearing. While there are numerous methods for IAS estimation, they have thus far only been evaluated using industrial-grade sensors and data acquisition systems with high computational resources. To boost the application of IAS estimation methods in embedded applications, like sensor-integrating machine elements, this paper highlights and evaluates an algorithm for low-power embedded systems. Theoretical analysis and benchmarks on various low-power microcontrollers demonstrate its feasibility for real-time application. Furthermore, evaluation with lower sampling rates and ADC resolutions confirms compatibility with small-scale sensors typical in embedded systems. For good performance, a typical Cortex-M4-based MCU (e.g. 1.12 ms estimation time using 1024 data points) and a sensor with 8-bit resolution and 6400 Hz sampling rate would suffice. These findings establish the algorithm's readiness for diverse embedded applications.

Due to the rapidly growing number of functions and sensors or actuators required for this, the vehicle is increasingly developing into a software defined vehicle (SDV). However, the information network of the individual components harbors points of attack for intrusions. The objective is therefore to design a detection system for these intrusions, which is analyzed within the paper on the basis of the accelerator pedal sensor. The focus here is on the symbiosis of fault diagnosis and intrusion detection in order to be able to generate a comprehensive image of the SDV for the secure use of the vehicle and intelligent sensors.

Neural network-based nonlinear system identification is crucial for various multi-step ahead prediction tasks, including model predictive control and digital twins. These applications demand models that are not only accurate but also efficient in training and deployment. While current state-of-the-art neural network-based methods can identify accurate models, they often become prohibitively slow when scaled to achieve high accuracy, limiting their use in resource-constrained or time-critical applications. We propose FranSys, a Fast recurrent neural network-based method for multi-step ahead prediction in non-autoregressive System Identification. FranSys comprises three key innovations: 1) the first non-autoregressive RNN model structure for multi-step ahead prediction that enables much faster training and inference compared to autoregressive RNNs by separating state estimation and prediction into two specialized sub-models, 2) a state distribution alignment training technique that enhances generalizability and 3) a prediction horizon scheduling method that accelerates training by progressively increasing the prediction horizon. We evaluate FranSys on three publicly available benchmark datasets representing diverse systems, comparing its speed and accuracy against state-of-the-art RNN-based multi-step ahead prediction methods. The evaluation includes various prediction horizons, model sizes, and hyperparameter optimization settings, using both our own implementations and those from related work. Results demonstrate that FranSys is 10 to 100 times faster in training and inference with the same and often higher accuracy on test data than state-of-the-art RNN-based multi-step ahead prediction methods, particularly with long prediction horizons. This substantial speed improvement enables the application of larger neural network-based models with longer prediction horizons on resource-constrained systems in time-critical tasks, such as model predictive control and online learning of digital twins. The code of FranSys is publicly available.

Trajectory similarity-based evaluation is the most intuitive method for Remaining Useful Life (RUL) prediction when abundant run-to-failure data are available. However, practical scenarios often present a challenge with limited access to such trajectories. This paper introduces an improved similarity-based method to bridge the gap by employing Monte Carlo to generate more trajectories. To begin, 13 features are extracted from both time and frequency domains, subsequently conducting dimension reduction to build the Health Index (HI) corresponding to the original acceleration measurements. Afterward, HI trajectories are fitted using exponential functions, and three different distribution functions (Gaussian, Gamma, Weibull) are adopted to identify the probability density of the two coefficients (θ, β) in fitted exponential models. Monte Carlo is applied to resample from the coefficient distribution, thereby generating more HI trajectories. Finally, the expanded HI library is used for RUL prediction and validated with two different bearing datasets. Experimental findings reveal that the trajectory expansion achieved through Monte Carlo sampling yields more accurate RUL estimation and reduced uncertainty.

In this paper the amount of data required for reliable and efficient diagnostic development on an application-oriented basis is examined. Embedded in the scalable data-based diagnostic concept, the research question of the amount of data will be analyzed in a kind of Monte Carlo Simulation. The possible influences of the diagnostic methods (Principal Component Analysis (PCA), and Autoencoder (AE)) will also be considered. The evaluation is two-folded, for each of the methods individually and compared to each other. In addition, it will be evaluated whether data availability can be integrated as a further scalability measure and supports the selection of the diagnostic method.

Gas foil bearings (GFBs) are fluid dynamic bearings with applications in high-speed lightweight machinery. One of the important parameters in GFBs is the lift-off speed. It indicates a speed above which there is no occurrence of dry friction and wear between the rotor and bearing, the so-called lift-off state. There are various techniques for determination of the lift-off state. However, they come with major limitations and complexities for use in real-world applications, especially in high-speed machinery. In recent years, monitoring the lift-off condition in journal bearings based on acoustic emissions (AE) has been investigated and shown to be suitable for this type of bearing. Nevertheless, its applicability to GFBs is yet not known, as there has been no practical investigation on AE signals in GFBs. In the present work, the applicability of AE measurements to detection of lift-off state in GFBs is explored experimentally based on previous studies on journal bearings. The results demonstrate that AE measurements are a potential alternative to conventional methods in determining the lift-off state in GFBs. The comparison between features of AE signals, measured at two different locations on the bearing, with the friction torque as the reference exhibits the applicability of AE measurements to detection of lift-off state.

This paper introduces and applies the Scalable Data-based Diagnostic Concept. At its core, the concept consists of (Kernel) Principal Component Analysis (PCA) and Autoencoder (AE), which are used to perform accurate fault diagnosis in technical systems, e.g. in automotive or railroad sectors, including various sub-methods for fault detection, identification and isolation. The analysis of real automotive fault cases is done, where a new smoothed comparative detection chart is presented. The findings prove the necessity of choosing the right method, regarding efficiency and the inherent data structure, which is one of the main objectives of the comprehensive scalable diagnostic concept. Copyright (c) 2024 The Authors. This is an open access article under the CC BY-NC-ND license (https://creativecommons.org/licenses/by-nc-nd/4.0/)

This paper presents the scalable data-based diagnostic concept with a focus on the scalability measures which should support the user of the concept in preselecting an appropriate method. For this purpose, measures for the data distribution as well as measures for the dependency structure of the data are analyzed on the basis of real vehicle data. It is assumed that there is no prior knowledge regarding the data structures, so that only the normal data (without fault case) is available for the analysis. The goal is to investigate the measures under real application and to prepare them for use within the diagnostic concept. Finally, a representation form for the user is presented, which should give a first impression of the data for the selection of the diagnosis method.

This paper introduces an algorithm for the detection of change-points and the identification of the corresponding subsequences in transient multivariate time-series data (MTSD).The analysis of such data has become increasingly important due to growing availability in many industrial fields.Labeling, sorting or filtering highly transient measurement data for training Condition-based Maintenance (CbM) models is cumbersome and error-prone.For some applications it can be sufficient to filter measurements by simple thresholds or finding change-points based on changes in mean value and variation.But a robust diagnosis of a component within a component group for example, which has a complex non-linear correlation between multiple sensor values, a simple approach would not be feasible.No meaningful and coherent measurement data, which could be used for training a CbM model, would emerge.Therefore, we introduce an algorithm that uses a recurrent neural network (RNN) based Autoencoder (AE) which is iteratively trained on incoming data.The scoring function uses the reconstruction error and latent space information.A model of the identified subsequence is saved and used for recognition of repeating subsequences as well as fast offline clustering.For evaluation, we propose a new similarity measure based on the curvature for a more intuitive time-series subsequence clustering metric.A comparison with seven other state-of-the-art algorithms and eight datasets shows the capability and the increased performance of our algorithm to cluster MTSD online and offline in conjunction with mechatronic systems.