When a mechanical failure occurs in three-phase asynchronous motor, a specific frequency of mechanical vibration will be generated, resulting in a series arc fault where poor contact occurs at the electrical point of contact under the action of mechanical vibration. To identify the series arc fault in the motor’s mechanical fault condition, a new multi-scale Convolutional Neural Network model (ECA-IMSCNN) has been developed. An arc fault simulation experiment is carried out on the electrical contact point under the mechanical vibration fault condition of the motor by positioning the arc fault generator near the motor at the back of the frequency converter. The one-dimensional signal is transformed into a two-dimensional data energy heat map using continuous wavelet transform, and then the two-dimensional image is utilized as a sample to train the ECA-IMSCNN fault diagnosis model. Through comparative analysis with other models, it is demonstrated that the model effectively detects arc faults under the mechanical fault condition of motor and can provide initial identification of the motor’s mechanical fault.