Martin Dazer

Abstract

Using an autoencoder (AE) based approach enables damage detection in a very early stage and does not require labeled training data. The earliest possible damage detection in gearboxes provides a precise insight into the state of health. The aim of this investigation is to use extended training datasets that include multiple operating conditions, thus robust damage detection is investigated independent of the operating state. In a first evaluation the operating states with same speed, but different torques of a healthy gear pair are used as a training dataset. In a second evaluation, operating states of different speeds and torques are used as a training dataset. In addition, the transferability of the algorithm to other gear pairs is examined without training the AE again. The vibration data used for damage detection are recorded on the test bench with a single stage spur gearbox. The test runs are carried out at different damage sizes, speeds and torques. The AE is trained on datasets without damage. The trained AE receives datasets of the tests with and without damage and encodes them. After a subsequent decoding step, the loss of the AE is calculated. The loss varies depending on the damage size. The investigation of the extended training datasets and the transferability of the AE provides valuable insights for the further development of robust damage detection algorithms.

Abstract

In conventional experimental designs, such as the central composite design (CCD), the behavior in terms of statistical test power has been studied. However, these standard experimental designs do not apply to all situations. In the case of restrictions of the experimental effort or available design space, optimal test designs offer an alternative, but their statistical test power properties are not documented. The statistical power is used as an indication of how effectively underlying effects are detected. This study quantifies the statistical test power of A-, D-, and V-optimal designs using Monte Carlo methods using simulations. The findings of this study demonstrate that optimal designs provide a viable alternative to standard test designs, showing superior power of effects compared to a CCD. However, reducing the number of runs indicates a clear negative impact on test power, which needs to be carefully considered in testing applications. Regarding effect detection, differences between optimal designs can be identified. Different optimal designs identify main, interaction, and quadratic effects with varying effectiveness and, thus, a different power. A qualitative relationship between optimality criteria and test power was observed, as improved optimality criteria indicate higher power. However, there is insufficient evidence that improved optimality will yield acceptable test power.

Abstract

In order to generate an empirical data set that is dependent on several loads for reliability or lifetime modeling in the most efficient way, one usually has to implement Accelerated Life Testing (ALT) through Design of Experiments (DOE). The Central Composite Design (CCD) is particularly suitable as one derivative of the Response Surface Methodology (RSM), which is generically performant but can also be efficiently adapted for extrapolation purposes. These become particularly relevant when the Response Surface Design (RSD) response is to be used for prediction at field level based on those End-of-Life (EoL) tests. Various methods for evaluating the prediction quality in terms of its variance and modified designs have been established so far, including the Variance Dispersion Graph (VDG) and the Fraction of Design Space (FDS) plot. However, for reliability engineering in particular, they did not allow direction-dependent assessments of the designs as in the case of extrapolation. Thus, a new variant of assessment tool for prediction, the Extrapolation Variance Graph (EVG), is presented here and its properties and strengths are demonstrated using exemplary configurations.

Abstract

Gearbox failure in offshore wind turbines can be caused by fatigue damages like tooth fank fracture (TFF) or pitting damage. For damage detection in current applications, simple threshold values for vibration amplitudes are used to initiate a shutdown of the wind turbine. At this time significant damage to the gear is already present and an external maintenance is inevitable. The objective of this experimental study is to determine a better sensor concept for detection of gear damage in wind power drives during operation. The damage needs to be detected at a very early stage before the health condition of the gears becomes critical. Evaluation criteria for comparison of sensor concepts include the earliest possible detection of damage, robust and reliable detection under varying operating conditions, as well as practicality and cost-effectiveness. The test results show, vibration measurement in radial direction and in high frequency ranges is most suitable and most robust for pitting detection during operation. With the best sensor concept in place, it is possible to detect gear damage like pitting at a very early stage of 0.3 % pitting size.

Abstract

Condition monitoring with very early gear damage detection enables remaining useful life prediction and health management of drive units. For a reliable detection of tooth flank damage during operation, a holistic approach combines experimental investigations, physical and AI-based data evaluation and is supplemented by FE and multi-body simulation.

Abstract

Early damage detection in gearbox applications enables the implementation of Prognostics and Health Management (PHM). On the one hand, the earliest possible damage detection provides a precise in-sight into the state of health of a gearbox. In addition, early damage detection offers the possibility to slow down the damage progress and extend the remaining useful life (RUL) by intervening in the operating state at an early damage stage. The main contribution of this work is that existing Machine Learning tools are applied to the challenge of very early damage detection in gearboxes. Thus, the need for complex physically based data evaluation is avoided. The aim of this investigation is a comparison of two different machine learning approaches. To investigate the detection possibilities test bench experiments were conducted with a single stage spur gearbox. For a comprehensive investigation, i.e. to detect damage under different operating conditions, the test runs are carried out at different damage sizes, speeds and torques. Based on the recorded vibration data, the damage detection is examined. Two machine learning approaches of anomaly detection are considered: An encoding approach and a loss approach. The same sparse autoencoder is developed for both approaches Both machine learning approaches are able to detect even the smallest damage of about 0.5 % in most operating states. The loss approach allows the different damage stages to be recognized much more clearly than the encoding approach. The comparison of the different approaches provides valuable insights for the further development of robust damage detection algorithms.

Abstract

Absence of failure data is a common challenge for data-driven predictive maintenance, particularly in the context of new or highly reliable systems. This is especially problematic for system level failure prediction of analog electronics since failure characteristics depend on the actual system layout and thus might change with system upgrades. To address this challenge, this work pursues a novel sim\-u\-la\-tion-as\-sist\-ed failure analysis methodology enabling automated and comprehensive evaluation of system level failure effects and failure detectability. While results obtained from simulations are suitable for comparative studies, they are confined to the simulation environment. To overcome this limitation, failure simulations are combined with generative models to generate realistic representations of missing failure data. Preliminary results demonstrate the capability of conditional generative adversarial networks (cGANs) to generate operational data of healthy systems, which accurately reflects correlations present in the source dataset. The proposed approach, using simulations as an additional source for generative models, not only targets the scarcity of failure data for highly reliable electronic systems but also ensures the adaptability of predictive maintenance algorithms to accommodate future system modifications and upgrades.

Abstract

The economic efficiency of machinery operation is significantly impacted by maintenance strategies. In the realm of condition-based or predictive maintenance strategies, the early detection of fatigue-induced damages is crucial. Therefore, this study focuses on the early detection of pitting damages during operation. Experimental investigations are conducted on a test gearbox acquiring acceleration data for different sizes of pitting damages under diverse operating conditions. A successful detection of the pitting damages during operation is achieved at a very early stage of progression, with a minimum size of 0.41% of an active tooth flank area. The utilization of design of experiments techniques facilitates the identification of factors influencing the detectability of pitting damages. The obtained results are analysed to elucidate the physical basis for the reliable detection of pitting damages across diverse operating conditions.

Abstract

Empirical life tests are used for reliability demonstration and determination of the actual reliability of the product. Therefore, engineers are faced with the challenge of selecting the most suitable test strategy out of the possible many and also the optimal parameter setting, e.g. sample size, in order to realize reliability demonstration with limited costs, time and with their available testing resources. It becomes even more challenging due to the stochastic nature of failure times and necessary cost and time being dependent on those. The considerations and guidelines in this paper are intended to simplify this process. Even simple products can fail due to several causes and mechanisms and usually have several components and subsystems. Therefore, this paper provides test planning options for single critical failure mechanisms as well as for systems with multiple failure mechanisms. For this purpose, the Probability of Test Success (Statistical Power of a life test) is used as a central, objective assessment metric. It is capable of indicating the probability of a successful reliability demonstration of a test and thus allows, for example, to answer the question of the required sample size for failure-based tests. The main planning resource is prior knowledge, which is mandatory due to the stochastic lifetime, in order to provide estimates for the Probability of Test Success at all. Therefore, it is also shown how to deal with uncertain prior knowledge and how the underlying information can additionally be used to increase the Probability of Test Success using Bayes’ theorem. The guidelines show how the most efficient test can be identified in the individual case and for individual boundary conditions.

Abstract

Zero Failure Tests are widely used to demonstrate the reliability requirements placed on products. The test is popular in application because it offers a number of advantages. DUTs (device under test) are tested to the service life requirement and then classified in a binary fashion, as they can only assume two states after testing - Intact or Failed. This allows test planning to be based on the binomial distribution and be handled very easily. The sample size to be tested results directly from the reliability and confidence to be demonstrated. Since parts only have to be tested up to the requirement and not beyond, the test also has advantages from a test planning point of view. In addition to cost savings, the test can be planned very well due to its fixed run times. However, there are some major drawbacks. For example, if the Zero Failure Test fails, it may be concluded that the product does need to undergo changes, in order to pass the test. However, this can be the wrong conclusion and a wrong business decision, since changes in the product are often very expensive in the late product development stages. Instead of changing the product, a change in testing can be the more efficient way of dealing with a failed test. In order to do this, a holistic view of available tests and test configuration is needed, as well as an objective measure for assessment. This tutorial provides the basics and differences of life testing strategies and a broad overview of different cases, scenarios, methods, processes and practical aspects of testing for the most important case of reliability demonstration. For several testing scenarios, it is shown how to best deal with test planning, in order to achieve the main target of testing. The tutorial also highlights the implementation practicalities, such as cost constraints, time constraints, and test capacity constraints.

Abstract

Response Surface Methodology (RSM) models may encounter limitations when incorporated with non-orthogonalities in the design matrix. This holds especially when performing end-of-life tests or demonstrating multivariate reliability. Exemplarily applied to CCDs, unexpectedly inexecutable test runs, variations in test point settings, and further factor level shifts can lead to impairments in the design model and prediction capability. This issue is addressed within this work. The consequences of common orthogonality offsets and their implications for predictive modeling relevant to multivariate reliability estimation with confidence intervals are discussed in here. By analyzing selected orthogonality disparities, trends in variance shaping in the parameter space and beyond are captured and described. For this purpose, visualizations like contours and variance dispersions of the test designs are interpreted. Considering the marginal range of predicted values, this approach thus serves to evaluate effects relevant for extrapolation from accelerated parameter ranges with respect to an estimate of their confidence intervals. In this way, the results show that experimental design manipulations justified from practical application exemplarily in a k=2 CCD lead to a more manageable handling of variances in the prediction of model values – which is lastly illustrated in an exemplary layout.

Abstract

In critical industrial applications, gearboxes are already equipped with Condition Monitoring Systems (CMS) based on vibration sensors. However, these systems are only developing their full potential if damages can be detected early and if the accumulated sensor data can be assigned to a type of damage and a damage size. The aim of this investigation is a systematic evaluation of gear damages at different stages to use the CMS sensor data for Prognostics and Health Management (PHM). It is also the objective to evaluate different sensor concepts for data acquisition regarding damage detection in gearbox applications. To investigate the detection possibilities on the test bench, a single stage spur gearbox is developed. The test gearbox is used on a test bench, which is designed as electrical load unit consisting of two electric motors. The tooth flanks of the test gears are manufactured with artificial pitting dam-age with different sizes. The test gearbox is equipped with torque, speed and vibration sen-sors. The experiments are statistically designed by means of Design of Experiments (DOE). The factors pitting damage size, rotational speed, torque and viscosity are varied. This offers a valid basis for developing an empiric model that relates the damage size to the observed sen-sor data. Using machine learning approaches, the vibration signal data are analyzed. By means of anomaly detection, early damage stages can be identified during operation. A con-clusion from the acquired sensor data on the extent of damage is possible through the sys-tematic DOE. This enables PHM which allows a more reliable and sustainable operation of drivetrains in industrial plants and systems.

Abstract

Gear failure caused by pitting is one of the leading reasons of downtime in wind turbines. An adaptive operating strategy applies a load reduction of a damaged tooth by the means of torque variation to increase the remaining useful life. For the highest possible increase of service life, condition monitoring data is used to implement an AI-based prognostic and health management strategy. Measurement data from several gears with different degrees of pitting are recorded with a test gearbox. The labeled data are used to train intelligent neural networks for automatic pitting detection during operation. By using reinforcement learning and artificial intelligence it is possible to identify the degree of pitting and the time of occurrence. Different approaches and methods are investigated. Based on this, an intelligent control can be implemented. In summary, the integration of AI in the control of wind power drives enables the increase of the remaining useful life.

Abstract

Gear failure caused by pitting is one of the leading reasons of downtime in wind turbines. An adaptive operating strategy applies a load reduction of a damaged tooth by the means of torque variation to increase the remaining useful life. For the highest possible increase of service life, a detection of pitting damage at an early stage during operation is necessary. To investigate the detection possibilities on the test rig, a test gearbox is developed. The tooth flanks of test gears are manufactured with artificial pitting damage at different stages. The test gearbox is equipped with various load and vibration sensors mounted at different positions of the housing. The sensors acquire a large amount of data, depending on the size of the damage. The test results are used to train deep neural networks for AI based plant operation. The experiments not only show at which stage pitting damage is detectable, but also form the data basis for AI based condition monitoring of wind power drives.

Abstract

Many failures of wind turbines are due to drive failures caused by pitting. Each failure can be associated with high repair costs and time-consuming repair work. This particularly applies to offshore facilities. For these reasons, increasing the remaining useful life of wind power drives is essential to leave a minimal ecological footprint by simultaneously increasing power output. Pitting damage occurs first on the weakest tooth. Artificial Intelligence is used to apply a local load reduction to a pre-damaged tooth and delay degradation. The other intact teeth compensate for the load reduction in order to achieve a constant average power. To increase the service life of wind power drives and to avoid unexpected failures an adaptive operating strategy can be implemented. With a test gearbox the adaptive operating strategy is examined on a test bench. The test gearbox is equipped with test gears with varying degrees of pre-damage. The objective of the examinations on the test gearbox is to detect pitting damage at the earliest possible stage. The earlier damage is detected, the greater the potential for increasing useful life. For detection, multiple high frequency acceleration sensors are integrated in the gearbox. Using machine learning approaches, the vibration data are analyzed. By means of anomaly detection damage can be identified during operation. Using torque control on the test bench, the load on pre-damaged teeth is minimized depending on the detected damage. In summary, the findings on the test gearbox will provide fundamental knowledge that will enable the implementation of the adaptive operating strategy inwind power drives.

Abstract

Modern vehicles have increasing safety requirements and a need for reliable low-voltage power supply in their on-board power supply systems. Understanding the causes and probabilities of failures in a 12 V power supply is crucial. Field analyses of aged and failed 12 V lead batteries can provide valuable insights regarding this topic. In a previous study, non-invasive electrical testing was used to objectively determine the reasons for failure and the lifetime of individual batteries. By identifying all of the potential failure mechanisms, the Latin hypercube sampling method was found to effectively reduce the required sample size. To ensure sufficient confidence in validating diagnostic algorithms and calculating time-dependent failure rates, all identified aging phenomena must be considered. This study presents a probability distribution of the failure mechanisms that occur in the field, as well as provides insights into potential opportunities, but it also challenges diagnostic approaches for current and future vehicles.

Abstract

An empirical-statistical demonstration of the lifetime of engineering products is necessary in order to assure fatigue and reliability requirements 1. Engineers are faced with the challenge of selecting the most suitable test strategy and also the optimal parameter setting, e.g. sample size, in order to realize reliability demonstration with limited costs, time and with their available testing resources. In this paper, different test strategies such as Sudden Death, time-censored, failurecensored and non-censored tests, and the Success Run Test are compared in a Reliability DOE concerning the necessary time, costs and the Probability of Success of the whole test configuration. The DOE is carried out with standardized cost and time data and the results can therefore be used generally and are transferable to any other product. With the resulting data, the different test strategies are comparable in terms of different boundary conditions. The Probability of Success can be determined and compared for the different strategies with a very low computing time in terms of cost and time. Using these results, engineers are now able to select the most suitable test procedure for their individual reliability requirements. Finally, the applicability is presented in a case study, showing the difference between different test strategies for the demonstration of a specific reliability target.

Abstract

If the focus during product development is on estimating a failure distribution function or a lifetime-quantile, analysis of failure behavior or determination of optimal design alternatives, there is often no alternative to life tests until failure. Alongside the determination of sample size for a desired accuracy of the result, engineers are also faced with the challenge of selecting the most suitable life test type at limited cost, time and test rig capacities. In this paper, results of a comprehensive simulation study are presented. Sudden Death tests, type-I and type-II censored and non-censored life test types are compared for two-parameter Weibull distributed data with respect to accuracy, time, cost and the investigation target. Significant influencing factors and boundary conditions, such as the maximum number of test rigs available, are taken into account to assure the practical applicability of the procedure for planning life tests. The database generated provides recommendations for an appropriate test configuration within an accuracy, time and cost triangle, including optimal test parameters. Furthermore, it enables engineers to substantiate qualitatively known correlations with numbers. This information can be used to describe the influence of a shift in boundary conditions. Finally, the use of the simulation results as basis for planning life tests is illustrated in a case study.