Research was focussed on stator current sensor fault classification algorithm using a Long Short-Term Memory (LSTM) recurrent neural network. The developed method is based on an existing solution (developed by the same authors) that used a multilayer perceptron (MLP) neural network for classification purposes. The use of the LSTM network allowed the number of hyperparameters to be reduced from 441 (MLP) to 251 (LSTM), while maintaining high classification accuracy (LSTM - 96.8%, MLP - 96.8%). Network testing was carried out on data from simulations carried out in the MATLAB/SIMULINK environment in a vector-controlled IM drive system. The test results obtained are important in terms of the optimisation of the size of the neural network, which is directly related to the future hardware implementation.
The attached data gathered in LSTM.json and MLP.json files contain the description and the suitable data for the neural networks exported from the MATLAB environment.
The LSTM.json and MLP.json files contain the description and the data for the neural networks exported from the MATLAB environment. The simulations were conducted for different speeds and load torques, e.g. the folder "data\wm_25_mo_50" means that it contains simulation files performed for 25% of the rated speed and 50% of the rated torque.
The files were saved in the .xlsx format:
isA_mea.xlsx, isB_mea.xlsx: currents measured in phase A and B respectively [p.u.]
isA_est.xlsx, isB_est.xlsx: currents estimated in phase A and B respectively [p.u.]
gamma_A.xlsx, gamma_B.xlsx: angle of the stator current vector relative to the A and B axes of the ABC phase coordinate system [rad]
time.xlsx: simulation time [s]
zero_cross_flagA.xlsx, zero_cross_flagB.xlsx: instants at which a new current period starts [s]
YPred_MLP.xlsx, YPred_LSTM.xlsx: classification results using MLP and LSTM networks
YTrue: true fault class
Fault Classes:
NF (No Fault): No Fault
G (Gain): Gain Change
OFF (Offset): DC Component Appearance
SAT (Saturation): Signal Saturation
OC (Open Circuit): Complete Signal Loss
(2025-02-22)
