Convolutional-Attention Hybrid Deep-Learning Model With Transfer Learning for Quality Inspection of DC Motors

Direct current (dc) gear motor is widely used in mechanical and electronic devices, known for the advantages of simple structure, small size, high torque, wide speed range, and high reliability. Accidents caused by motor quality issues or malfunctions can have serious negative consequences, highlighting the need for dc gear motor quality inspections. Addressing the limitations in the generalization and accuracy of current deep-learning-based motor fault detection methods, this work proposes a hybrid convolution and self-attention deep-learning classification network model named ConvNeXt-T-MSA. The model is applied to the quality inspection of dc gear motors, and the time–frequency diagram of the motor vibration signal obtained by a short-time Fourier transform is used as the model input. The model combines convolution and self-attention and optimizes model performance through transfer learning to alleviate the data scarcity issues. The proposed approach achieves effective quality inspection and grade classification for dc gear motors. Multiple 12FN20-37 dc gear motors are used to construct a dc motor quality inspection dataset. For the 12FN20-37 dataset, the proposed method achieves an accuracy of 99.7%, and the accuracy on the public dataset CWRU is 99.9%. Based on the designed network model, we design a dc gear motor quality inspection system. Another 40 12FN20-37 dc geared motors are used to conduct real-time testing. A total of 800 tests are performed and an accuracy rate of 95% is achieved. The average total sampling and detection time for each time is 1.52 s. The results indicate that the system can accomplish the task of dc motor quality inspection with good generalization and accuracy.