In engineering practice

the scarcity of induction motor fault data has become a bottleneck that restricts the widespread application of data-driven diagnostic methods in field applications. In order to overcome the reliance on actual data

a simulation data-driven transfer recognition method for stator and rotor faults of induction motors is proposed. This method firstly generates voltage and current data of the motor under different health states through finite element modeling. Then

it analyzes the performance of different faults in the instantaneous power and converts the fault feature components at different frequencies into trajectories of different colors. The trajectories are thus fused into graphical samples with multiple feature components. The graphical samples reduce the distribution differences between simulation and actual data

thereby decreasing the difficulty of transfer recognition. Finally

using the graphical samples as input

the convolutional neural network trained by the simulation data is directly applied to fault recognition of actual motors. The experimental result shows that the method can still accurately recognize stator and rotor faults in the actual motor in the case of learning only from simulation data samples

with an accuracy as high as 97.7%

meeting engineering requirements.