Fault detection and diagnosis in electric motors using 1d convolutional neural networks with multi-channel vibration signals

[Display omitted] •A methodology for fault detection and diagnosis in electric motor is presented.•A one-dimensional convolution neural network is trained using the vibration signal from two different accelerometers.•A series of experiments with seven different induced faults and operation conditions is performed to validate the method.•t-distributed stochastic neighbor embedding statistical method is used for feature visualization.•The method shows low computational cost, high accuracy and robustness. Fault detection and diagnosis in time series data are becoming mainstream in most industrial applications since the increase of monitoring sensors in machinery. Traditional methods generally require pre-processing techniques before training; however, this task becomes very time-consuming with multiple sensors. Recently, deep learning methods have shown great results on time series data. This paper proposes a multi-head 1D Convolution Neural Network (1D CNN) to detect and diagnose six different types of faults in an electric motor using two accelerometers measuring in two different directions. This architecture was chosen due to each head can deal with each sensor individually, increasing feature extraction. The proposed method is verified through a series of experiments with seven different induced faults and operation conditions. The results show that the proposed architecture is very accurate for multi-sensor fault detection using vibration time series. Since the experiments are based on real electric motors and faults, these results are promising in real applications.