© MEV-Verlag – Jonas Krüger
The project »MADESI« is a project in the context of the BMBF announcement »Richtlinie zur Förderung von Forschungsvorhaben zur automatisierten Analyse von Daten mittels Maschinellen Lernens«. Next to Fraunhofer SCAI, the cooperating partners are Weidmüller Monitoring Systems GmbH, TU Darmstadt and ZF Friedrichshafen AG.
Current sensor technologies make it possible to monitor and evaluate the condition of a given mechanical system constantly by measuring relevant features of the system and reporting them to the respective operator. In doing so, the operator can analyze the received data and check them for anomalies, such that the state of the system can be determined from afar. This gain of information can then be used to predict the remaining life time of the system, or to detect the cause of the anomaly and to organize maintenance operations, only if they are really needed (predictive maintenance). With this approach, the operator effectively can maximize the productivity of the system while minimizing the costs at the same time.
In particular, the MADESI project has the goal to develop tools and methods in order to be able to make sense of data, which are gained from wind turbines. This means detecting failures of the turbines - like imbalances, which are caused by ice formations on the blades, or erosion, caused by hail, or heavy rain - as early as possible. Spotting and fixing these failures at an early stage will then lead to a minimization of the expected power loss generated by the wind turbine.
The corresponding data is stored as time series. In the context of anomaly detection of time series, there exist a number of so called Black-Box Machine-Learning (ML) methods, which are well known and can be trained on a large set of labeled data. However, as these anomalies occur rarely and an artificial creation in this context is very costly, the idea is to generate synthetic data, i.e., simulations and use these for the mathematical analysis and training of the ML models. With simulations one is able to tune a lot of parameters, both of the wind turbine - like turbine type, blade length, blade weigth, etc. - and environmental factors – like wind speed, or temperature – and choose them appropriately for the desired use case. These can be defined in collaberation with engineers and domain experts. By having a controlled setting, the interpretability of the ML-methods shall be increased. In addition, the goodness of the applied methods can be evaluated more reliably. Moreover, as it is known in the wind energy context that there may be massive noise on the data, which complicates the statistical analysis, in the simulations one can handle the influence of the noise factors, especially those produced by strong wind.
Technically speaking, in order to find anomalies in the time series, one has to apply suited metrics, which are able to measure the distance between them and then find the time series with an increased distance from those time series, which behave normal. It is therefore necessary to transform data in a certain way, such that those metrics are applicable. Desired attributes of these transformations are that they simplify the interpretation of the data, as it is for example the case in topological data analysis methods.
Finally, it is favorably to use and analyze the data as efficiently as possible, in the sense that complex, non-linear, dependencies between the different sensors can be found and exploited. The use of multivariate probabilistic models should lead to a better understanding of the whole mechanical system and reduce redundancy. To this end, tractable probabilistic methods, like Sum-Product Networks, and methods for the exploration of non-linear dependencies, such as copulas, will be used.
Having created a large set of simulations, these shall be stored together with real measurement data in a common data base and an interface shall be developed, which makes it possible to find targeted measurments/simulations, with a desired set of parameters and those, which are similar to them. The combination of analysis of real measurement data and controlled simulation data shall then lead to the development of so called Grey Box ML-methods.
The MADESI project is being performed in cooperation between Fraunhofer SCAI, TU Darmstadt, as well as Weidmüller Monitoring Systems and ZF Friedrichshafen.
Project duration: 10/2018 - 09/2021
Machine-Learning Techniques for stochastic-deterministic multi-sensor signals