No abstract

A digital twin test bench was created to demonstrate the digital twin concept for control and prediction of the dynamic behavior of a paper machine roll. The paper presents a complete proof of concept digital twin system with wireless sensoring with flexible measurement patterns, data transfer, storage and visualization in an interactive 3D view. A virtual sensor based on a recurrent neural network was created to predict the middle cross section center point movement of a large flexible rotor based on acceleration and force input measured from bearing housings at the ends of the rotor. The results show that a neural network algorithm is feasible for predicting the dynamic behaviour of the rotor system. Future research at Aalto University aims to apply sensor fusion data as input to non-physics based models with the goal to predict key performance indicators of complex mechanical systems.

The mounting of a rotating machine affects the dynamic behavior of the machine. Typically in large machines, the support structures have lower stiffness on the actual site than in the acceptance tests conducted by the manufacturers. In this research, a method is developed for the support stiffness identification for an in-situ machine using a simulation-data-driven, deep learning algorithm. The novel approach aims to utilize transfer learning to first teach the deep learning algorithm using vibration response data generated from a simulation model of the rotor-bearing-support system, and then test it with measured response. To validate the stiffness estimation of the algorithm for multiple cases, an experimental test rig is used where the horizontal support stiffness can be varied through a range of values. The results from the deep learning algorithm are compared with simpler algorithms such as Linear regression (LR), Artificial Neural Network (ANN), and Support vector regression (SVR) for benchmarking. The models are trained with filtered frequency domain response, and challenges due to measurement uncertainty are analyzed. With proposed pre-processing steps of the frequency domain response and outlier elimination, the deep learningbased virtual sensor can predict the support stiffness with reasonable accuracy, where the limiting factor is the data quality and lack of excitation at critical speed frequencies.

Intelligent fault diagnosis (IFD) models have the potential to increase the level of automation and the diagnosis accuracy of machine condition monitoring systems. Many of the latest IFD models rely on convolutional layers for feature extraction from vibration data. The majority of these models employ batch normalisation (BN) for centring and scaling the input for each neuron. This study includes a novel examination of a competitive approach for layer input normalisation in the scope of fault diagnosis. Network deconvolution (ND) is a technique that further decorrelates the layer inputs reducing redundancy among the learned features. Both normalisation techniques are implemented on three common 1D-CNN-based fault diagnosis models. The models with ND mostly outperform the baseline models with BN in three experiments concerning fault datasets from two different rotor systems. Furthermore, the models with ND significantly outperform the baseline models with BN in the common CWRU bearing fault tests with load domain shifts, if the data from drive-end and fan-end sensors are employed. The results show that whitened features can improve the performance of CNN-based fault diagnosis models.

No abstract