Sensorless Misalignment Detection on Linear Feed Axis with Revised ResNet and Transfer Learning Using Motor Current

Abstract

Abstract Due to ageing populations and a shortage of skilled labour, automatic machine condition monitoring is a powerful tool to ensure smooth operation of production systems with reduced manpower. Automatic condition monitoring enables early detection of machine faults, greatly increasing uptime, reliability, and safety. However, conventional fault detection methods based on vibration require installation of additional sensors, thus bringing up implementation effort and initial costs. The linear feed axis is a machine component whose failure can bring an entire production line to a standstill. Therefore, this study presents a sensorless approach, which uses a linear axis’ motor current for the detection of misalignment. Motor current time series data was encoded as images and then fed to a CNN, more precisely a revised residual neural network (ResNet). A random search hyper-parameter tuning technique was used to optimise the structure of the CNN. Then, transfer learning was used to apply the current signal features already learned to other scenarios. The results showed that both horizontal and vertical misalignments of linear feed axes can be well detected by a revised ResNet using motor current signals, with an accuracy of 99.76%. Using transfer learning, misalignments were detected with an accuracy of 92.67% – even under the influence of external forces.