Aiming at the problems of weak anti-noise ability and low feature discrimination of traditional time-frequency analysis method in motor fault signal processing under complex working conditions, a motor fault diagnosis method based on improved symmetric point mode feature fusion image is proposed. The variational mode decomposition is used to adaptively decompose the fault signal, and the particle swarm optimization algorithm is used to dynamically adjust the key parameters such as the rotation angle θ of the symmetric plane of the SDP image, the time interval parameter l, and the angle amplification factor ξ to generate the SDP image of multi-modal signal fusion. A lightweight convolutional neural network model is constructed to study motor fault diagnosis. Taking the variable frequency three-phase asynchronous motor as the experimental object, 1200 sets of vibration signal samples were collected for verification under four working conditions : normal, inter-turn short circuit, air gap eccentricity and rotor broken bar. The results show that the proposed method achieves 100 % accuracy in fault diagnosis, which is better than 97.85 % of support vector machine and 95.24 % of random forest. It still maintains 98.76 % accuracy in-6dB strong noise environment, which verifies its robustness. This method improves the diagnostic reliability in noisy environment through parameter optimization and feature fusion, and provides an effective solution for real-time monitoring and intelligent diagnosis of motors.