Abstract: Three-phase asynchronous motors are prone to early faults such as insulation degradation, bearing wear, and rotor imbalance under low-speed operating conditions. This paper proposes a recognition technology scheme based on multi-source signals, starting from the collection of current, vibration, and temperature rise, designing feature parameter calculation methods, constructing a time-frequency model using short time Fourier and wavelet packet decomposition, and combining singular value decomposition for feature dimensionality reduction. In the pattern recognition phase, a hierarchical structure using support vector machines and convolutional neural networks is employed to classify the samples under various operating conditions. Experimental results indicate that the constructed method can accurately extract indicators and provide stable recognition probabilities under light, medium, and heavy loads, forming a feasible path for early fault monitoring of low-speed motors.