Representation of Time and Frequency dependent Vibration Data Sensor using Neural Networks
* Presenting author
Abstract:
The derivation of vibration signals, captured by accelerations, is often limited by a high implementation effort for simulation, caused by many issues, e.g. non-linearity. Performing vibration tests for various loads are also quite expensive and laborious and not possible in early stages of the development process of components and machines. Furthermore, many different maneuvers need to be measured to enable reliable and field equivalent load assumptions. With the usage of Deep Neural Networks (DNN) it is possible to simplify and speed up the simulation and testing efforts with the prediction of acceleration signals in the time and frequency domain. In this paper, a framework for vibration analysis using the sensor data is developed. It includes an automatic preprocessing of time series, feature engineering of input data and training of a so-called soft sensor with different types of DNN architectures in time and frequency domains. Examples of the implementation of these approaches of soft sensors are shown and compared to each other.