Aufgrund der Fortschritte der Digitalisierung finden Systeme zur Zustandsüberwachung vermehrt Einsatz in der Industrie, um durch eine zustandsbasierte oder eine prädiktive Instandhaltung Vorteile, wie eine verbesserte Zuverlässigkeit und geringere Kosten zu erzielen. Dabei beruhen Zustandsüberwachungssysteme auf den folgenden Bausteinen: Sensorik, Datenvorverarbeitung, Merkmalsextraktion und -auswahl, Diagnose bzw. Prognose sowie einer Entscheidungsfindung basierend auf den Ergebnissen. Jeder dieser Bausteine erfordert individuelle Einstellungen, um ein geeignetes Zustandsüberwachungssystem für die jeweilige Anwendung zu entwickeln. Eine offene Fragestellung im Bereich der Zustandsüberwachung ergibt sich aufgrund der Unsicherheit der Zukunft, die sich in den zukünftigen Betriebs- und Umgebungsbedingungen zeigt. Diese Unsicherheit gilt es in allen Bausteinen zu berücksichtigen. Dieser Beitrag konzentriert sich auf den Baustein Merkmalsextraktion und -selektion, mit dem Ziel anhand geeigneter Merkmale eine Prognose der nutzbaren Restlebensdauer mit hoher Genauigkeit realisieren zu können. Daher werden geeignete Merkmale aus dem Zeitbereich und daraus abgeleitete Zustandsindikatoren für die Restlebensdauerprognose von technischen Systemen vorgestellt. Dabei sind Zustandsindikatoren Kenngrößen zur Beobachtung des Zustands der kritischen Systemkomponenten. Anhand dreier Anwendungsbeispiele wird ihre Eignung evaluiert. Dabei werden Daten aus Lebensdauerversuchen unter instationären Betriebs- und Umgebungsbedingungen ausgewertet. Die auftretenden Unsicherheiten der Zukunft werden somit berücksichtigt. Die Beispielsysteme beruhen auf Gummi-Metall-Elementen und Wälzlagern. Aus den generierten Ergebnissen lässt sich schließen, dass die Zustandsindikatoren aus der betrachteten Zeitreihen-Toolbox auch unter unbekannten Betriebs- und Umgebungsbedingungen robust sind.

Several methods, including order analysis, wavelet analysis and empirical mode decomposition have been proposed and successfully employed for the health state estimation of technical systems operating under varying conditions. However, where information such as the speed of rotating machinery, component specifications or other domain-specific information is unavailable, such methods are often infeasible. Thus, this paper investigates the application of classical time-domain features, features from the medical field and novel features from the highly comparative time-series analysis (HCTSA) package, for the health state estimation of rotating machinery operating under varying conditions. Furthermore, several feature selection methods are investigated to identify features as viable health indicators for the diagnostics and prognostics of technical systems. As a case study, the presented methods are evaluated on real-world and experimentally acquired vibration data of bearings operating under varying speed. The results show that the selected features can successfully be employed as health indicators for technical systems operating under varying conditions.

In the industry 4.0 era, there is a growing need to transform unstructured data acquired by a multitude of sources into information and subsequently into knowledge to improve the quality of manufactured products, to boost production, for predictive maintenance, etc. Data-driven approaches, such as machine learning techniques, are typically employed to model the underlying relationship from data. However, an increase in model accuracy with state-of-the-art methods, such as deep convolutional neural networks, results in less interpretability and transparency. Due to the ease of implementation, interpretation and transparency to both domain experts and non-experts, a rule-based method is proposed in this paper, for prognostics and health management (PHM) and specifically for diagnostics. The proposed method utilizes the most relevant sensor signals acquired via feature extraction and selection techniques and expert knowledge. As a case study, the presented method is evaluated on data from a real-world quality control set-up provided by the European prognostics and health management society (PHME) at the conference’s 2021 data challenge. With the proposed method, our team took the third place, capable of successfully diagnosing different fault modes, irrespective of varying conditions.

In all fields, the significance of a reliable and accurate predictive model is almost unquantifiable. With deep domain knowledge, models derived from first principles typically outperforms other models in terms of reliability and accuracy. When it may become a cumbersome or an unachievable task to build or validate such models of complex (non-linear) systems, machine learning techniques are employed to build predictive models. However, the accuracy of such techniques is not only dependent on the hyper-parameters of the chosen algorithm, but also on the amount and quality of data. This paper investigates the application of classical time series forecasting approaches for the reliable prognostics of technical systems, where black box machine learning techniques might not successfully be employed given insufficient amount of data and where first principles models are infeasible due to lack of domain specific data. Forecasting by analogy, forecasting by analytical function fitting, an exponential smoothing forecasting method and the long short-term memory (LSTM) are evaluated and compared against the ground truth data. As a case study, the methods are applied to predict future crack lengths of riveted aluminium plates under cyclic loading. The performance of the predictive models is evaluated based on error metrics leading to a proposal of when to apply which forecasting approach.