In order to perform fault classification using Machine Learning algorithms, sufficient and balanced data is required. Nevertheless, in a lot of use cases data is not available in a sufficient manner to allow a stable and valid training of Machine Learning based algorithms for fault classification. There are various methods for augmenting real measured data, whereby a Generative Adversarial Network (GAN) is one of the most suitable approaches for synthetic data generation. Conditional and non-conditional data augmentation approaches are available for implementation of GAN algorithms for synthetic data generation. In the scope of the proposed paper, the performance of a conditional GAN (cGAN) and traditional GAN is evaluated and compared using vibration data of a rolling bearing measured on a bearing test rig. Both GAN approaches are used with optimized losses, considering the Wasserstein Distance and a gradient penalty in the loss function. The loss function of a cGAN contains label information, enabling the model to be trained as a holistic system for all labels, rather than being trained per label as in a traditional GAN. The training of the both GAN networks show a high performance. The synthetically generated data compared to the real data also shows sufficient similarity. Nevertheless, the distribution of the generated data is different, as the data generated with a non-conditional GAN has a better quality of results. Different methods are used to evaluate the data, for example statistical or cluster analysis, highlighting the differences in the generated data.

This paper presents the development and implementation of digital twin models for a frequency converter in stationary operating points, using machine learning techniques to predict key electrical parameters, including inverter voltage, current, frequency, and IGBT temperature. Four different modeling approaches, each using different machine learning methods, were investigated: a simple neural network structure, a parallel model structure, a serial model structure, and a serial physical model structure. The primary objective of this study was to compare the predictive accuracy of these models and identify the most effective approach. Each model was trained, optimized, and evaluated using real-world data collected from a back-to-back test bench setup. After extensive data pre-processing and normalization, the model performance was evaluated based on error metrics such as mean absolute error, maximum absolute error and mean relative error. The serial physical model structure demonstrated the highest prediction accuracy, achieving a mean relative error of 0.2 % for inverter IGBT temperature, 0.4 % for inverter current, 0.7 % for inverter voltage, and 0.2 % for inverter frequency. This study highlights the importance of incorporating physically meaningful input parameters to improve model performance. Furthermore, unlike many conventional parameter prediction methods, the proposed approach is capable of real-time computation due to its low computational effort. This real-time capability significantly expands its potential applications in various domains. The results highlight the potential of digital twin technologies combined with machine learning to improve the efficiency and reliability of power electronic systems. The proposed methodology supports advanced condition monitoring and system optimization, with potential applications in electromobility and other industrial sectors.

Air compressors are vital in industries, supplying compressed air for industrial applications while ensuring efficiency reduced energy loss, maintenance costs and reliability is critical. These compressors are affected by environmental factors and component health when operating under cyclic conditions. This work presents a data-driven approach to optimize the screw compressor efficiency using feature engineering. Feature selection methods like correlation analysis and Elastic Net identified informative sensors, while feature extraction techniques such as curvature analysis and peak detection quantified deviations of features. To reduce the effort on comparison and faster implementation, an Artificial Neural Network (ANN) Regression with a custom loss function predicts outlet pressure RMSE difference and cycle duration, achieving R-squared scores of 0.97 and 0.99 respectively. The analysis found optimal cycle durations of 193s, 108s, and 94s for high, low, and medium pressures, with RMSE deviations of 0.2, 0.07, and 0.10. The proposed method enables monitoring, alerting users to operational variations for improved compressor efficiency.

This paper presents a systematic, data-driven approach to the multi-fault diagnosis of screw compressors, with a focus on belt and oil level conditions. The paper proposes a systematic approach to data pre-processing and model training. Key steps include feature extraction using root mean square and kurtosis to extract statistical parameters. For feature selection, an embedded method based Elastic Net is used, and selected sensors are generalised for fault detection. This is followed by data augmentation using the Synthetic Minority Oversampling Technique to address imbalanced datasets acquired during operation. The study demonstrates that a single vibration sensor mounted on the motor can effectively detect both fault conditions alongside the existing current sensor. Individual models are then trained using the selected and augmented features with a Random Forest classifier, achieving 0.92 accuracy for both faults. Meanwhile, the ensemble learning-based voting classifier achieved 0.90 accuracy for multi-fault classification. This robust and scalable approach is well suited to industrial applications when developing a multi-fault diagnosis approach, and can be extended to develop low-cost sensor concepts.

This study explores the utilization of a machine learning approach for the prediction of torque in an induction machine. It focuses on the assessment of a model that has demonstrated efficacy in stationary torque prediction, which is evaluated in this study under dynamic conditions. Gaussian process regression is used for this purpose. A key feature of this approach is that the machine learning model is trained with stationary operating points. The experimental results demonstrate that machine learning models encounter specific challenges under dynamic load conditions, particularly due to systematic errors that compromise prediction accuracy. The experimental validation was conducted using a back-to-back test bench configuration with two identical 440kW induction machines. Additionally, a power analyzer was installed at the test setup to record time-synchronous electrical measured variables. The findings offer significant insights into the optimization of a machine learning approach for dynamic operating conditions and underscore the challenges associated with their practical implementation in real drive systems.

Prognostics and Health Management (PHM) systems are evaluated based on different performance criteria, such as result performance and cost efficiency, which may vary depending on the application. The selected features for the PHM application can have a significant influence on the overall PHM system performance. Therefore, this work discusses an approach for multi-objective feature selection for PHM, aiming to optimize overall PHM system performance regarding different criteria. A regularization based embedded approach is tested, which incorporates external evaluation factors for the features in the loss function calculation. The monetary cost of each feature is used to optimize a PHM system for two criteria: Accuracy and cost efficiency. A real test dataset from a monitoring application of a process thermostat is used for evaluation of the approach. It is shown that the presented approach can aid to select suitable subsets of features for multi-objective problems. It can assist in identifying relationships in the dataset and can reduce the dependency on domain knowledge.

With the focus in the realm of automotive testing, process thermostats play a vital role in providing the required operating environment. These process thermostats, with the operating medium of an ethylene glycol-water ratio, play a crucial role in terms of controlling their thermal properties. With an emphasis on identifying the best preprocessing method for classifying these ratios, this study utilises a detailed comparison of statistical methods and the wrapper method-based Genetic Algorithm as a search method for the most relevant feature selections. Given the huge number of existing sensor parameters in the system, thereby emphasising the importance of feature selection criteria for effective analysis and model training. Furthermore, a random forest-based classifier is used with these parameters to predict the accurate ethylene glycol to water ratio.

Condition monitoring plays a crucial role in real-time evaluation of system states, but requires a large amount of measurement data to develop an accurate model. In reality a sufficient amount of data is often not available. Therefore, the proposed approach focuses on data augmentation using a Wasserstein Generative Adversarial Network (WGAN) to augment time series condition data. To enable the generation of synthetic data in the time domain, a Long Short-Term Memory (LSTM) network architecture is used in conjunction with a WGAN. The practical implementation of WGAN with a LSTM architecture is verified with two different datasets, a vibration and an acoustic dataset. Therefore, vibration data of rolling bearings in various system states recorded on a rolling bearing test rig and acoustic data of a welding process are used for synthetic data generation in separately trained networks. This work also focuses on a detailed evaluation of synthetic data in relation to the real data using various methods, such as distribution function analysis, a parameter analysis, and a visual comparison in the time as well as frequency domain. A classification model is also used to classify the real and a combined (real and synthetic) dataset to verify the benefits. The classification of the acoustic data shows an improvement in test accuracy with the combined dataset to 100% compared to the existing real measurement data of 65%. The following paper highlights the potential of the mentioned algorithm for data augmentation as an optimal solution for the described use cases. Synthetic data generation in the time domain is also critically discussed and the difficulties involved are emphasized.

In automotive testing, accurate measurement of torque and speed is critical. However, it is often expensive, requires precise mounting and positioning of sensors, as well as a controlled maintenance. The proposed approach focuses on overcoming these limitations by developing a novel concept of virtual sensors using vibration measurements as a replacement for physical and expensive sensors to measure input shaft speed and output torque with a case study of a two-stage reduction gearbox. The gearbox, mounted on a powertrain test rig, is equipped with multiple sensors, emphasizing sensor positioning and sensor combination as relevant points. A Random Forest-based regression feature importance (RFR-FI) method is used to identify the best performing sensors. A comprehensive time domain analysis and frequency domain comparison is also performed to select the optimal data for model training. Verification of the selected parameters and data is performed by a time-frequency domain analysis using spectrograms. Training two Artificial Neural Network (ANNR) models to predict input shaft speed and output shaft torque results in sufficiently accurate values for R 2 and the Root Mean Squared Error (RMSE). The trained model is evaluated on unseen data with a randomized test cycle, which also yields good results in predicting torque and speed. This method highlights a cost-effective yet reliable estimation of critical parameters for automotive testing, emphasizing the effectiveness and precision of virtual sensor techniques.

In order to perform fault classification using machine learning algorithms, sufficient and balanced data is required. The paper presents a novel approach for data augmentation using a Generative Adversarial Network (GAN) with transient, time-dependent vibration data. The proposed approach aims to generate synthetic data in form of spectrograms that closely resemble the characteristics of the real data sets. Synthetic data generation can be used to improve the training performance of neural networks for vibration analysis if not enough or unbalanced real data is available. The authors demonstrate the effectiveness of this approach by training a vibration-based fault detection model using synthetic data and comparing its performance to a model trained on real data only as well as with a t-distributed stochastic neighbour embedding (t-SNE). Real data is acquired on a bearing test rig and measurements are carried out on bearings with four different system states. The results show that the model trained with the synthetic data set outperforms the model trained with real data only, indicating that the synthetic data generated by the proposed approach can improve the training performance and accuracy of the machine learning model. Overall, the paper highlights the potential of GAN-based data augmentation approach via spectrograms for vibration analysis and offers insights into its practical application to bearings.

Production planning poses major challenges due to high product diversity, high fluctuations in demand and strong dynamic fluctuations in delivery quantities and times. This makes automation of production planning increasingly necessary to enable optimized sequence planning based on current production data. Production control systems (PRS) offer the possibility of automatic corrective actions in the event of deviations from the specified target dates. Production, PRS and data acquisition are considered as a single control loop and are the subject of various development and research topics. However, automatic interventions in planning and control can lead to serious disruptions if there are errors in the software of the PRS. Therefore, testing and validation of PRS is a crucial factor in development, but is hardly possible in real production environments due to the availability of physical manufacturing facilities and the associated risks. Therefore, in the scope of the paper, a concept for a software framework is proposed to provide a test environment for a PRS. A virtual production environment, which can be combined with the PRS and parameterized in form of a discrete-event simulation, offers the possibility to test the PRS. It is also possible to couple the simulation with a reference calculation to show the behavior of the production environment. The entire framework is designed to enable the identification of critical machines and times as well as, in further steps, the detection of software errors in the PRS.

Knowledge of the efficiency of the subsystems in the drive train is essential for development in drive technology. Experiments have confirmed that the efficiency gradient of gearboxes is particularly high in the partial load range. Conventional test planning does not adequately reflect the influence of the efficiency gradient in the partial load range. To adequately represent the partial load range, it must be recorded with a particularly high number of measuring points. Map ranges with a low gradient are recorded with a smaller number of measuring points. This paper describes a method for the design of experiments, not with but for neural networks, with regard to modelling the efficiency of a two-stage spur transmission. In order to determine the test data, an optimized design of experiments is carried out. The test planning is divided into preliminary tests and main tests. The measuring points are optimally distributed in the characteristic diagram for use in neural networks. The measurement data is generated with the planned test data on a road to rig vehicle test rig. Suitable methods are used to prepare the data for further processing. The modelling is done experimentally with the help of neural networks. The result is a function for the continuous description of the efficiency at each operating point. The procedure is compared with the conventional method and validated with statistical methods. It was proven that the optimised experimental design method for neural networks gives better results than the conventional approach.

Condition monitoring of machine elements for end-of-line quality control is an essential part of product assurance in many production areas. Therefore, the monitored machine elements are often separated in different condition classes and are classified via conventional signal processing and machine learning methods. Especially for machine learning algorithms, collection of a sufficient amount of data from each class is particularly relevant so that balanced training datasets, regarding the condition classes, are available. In reality, however, in most cases significantly more data is captured from good system states. This results in unbalanced data sets, which can be counteracted with synthetically generated data. Generative Adversarial Networks (GAN) are a suitable approach to generate synthetic measurement data. In the scope of this paper, a use case is considered, in which bearings are monitored in an end-of-line control via acoustic signals. The generation of such data requires a high computational effort. To reduce this effort, a suitable signal pre-processing method is presented, which allows the reduction of the generated feature dimension. The synthetic data based on low-dimensional features is subsequently evaluated regarding its suitability for the condition classification. It is shown that with synthetically generated data with low-dimensional features a similar classification accuracy can be achieved as with real data, making the synthetic data suitable for data augmentation in the use case.

The increasingly noticeable effects of climate change require action to reduce greenhouse gas emissions. In the transport sector, the increase in electromobility is an important tool for reducing CO2 emissions. In order to meet the essential requirements, a particularly detailed system knowledge, for example on energy efficiency in the powertrain, is needed for the development and improvement of electric vehicles. The present work contributes to this by mapping the efficiency of a planetary gearbox for an electric vehicle. The mapping is done by experimental modelling with machine learning methods based on a standard efficiency description. The data necessary for the neural network training is generated by efficiency experiments on the powertrain test rig. For suitable results the data was pre-processed by applying relevant low-pass filters. A comparison of the Bayesian-Regularization algorithm, the Levenberg-Marquardt algorithm and the scaled-conjugate-gradient algorithm for training exhibited the strengths and weaknesses of the individual algorithms for the efficiency mapping. By comparing each algorithms performance metrics, the one matching the requirements best is chosen for the efficiency mapping. The result is a continuous function that determines the efficiency based on speed- and torque-inputs. Based on a key performance indicator, statistical validation by evaluating the standard deviations is used to ensure the quality of the results. In this paper the suitable algorithm for the given use case was determined. It can be applied for further research due to the significant results shown.