In the rapidly developing era of IoT technology, significant progress has been made in enhancing real-time monitoring and data processing capabilities for motor fault diagnosis. Due to the complexity of motor structures and the characteristics of their signals, such as instability, nonlinearity, and large data volume, traditional diagnostic methods have exhibited poor performance. To address these challenges, motor fault signals were transformed into time-frequency spectrograms using the short-time Fourier transform(STFT). A convolutional neural network(CNN) algorithm was constructed to train preprocessed samples, and the Triplet-Center Loss was employed as a loss function to measure the distance between high-dimensional features of fault data. The network was fine-tuned in a supervised manner with labels, and the processed data was transmitted to an IoT cloud platform to enable real-time monitoring and diagnosis. Experimental results demonstrate the effectiveness of this method in efficiently diagnosing motor faults in real-time, thereby overcoming the shortcomings of traditional approaches.