Abstract: To timely capture the abnormal states of machine tool spindles frequently used in the production process and thereby ensure the precision of machined components and personnel safety, a small-sample abnormal state identification network for spindle systems based on constraint generation adversarial network-bi-directional long short-term memory (CGAN-BiLSTM) is proposed. Initially, a feature extraction network under the BiLSTM attention mechanism is adopted to dynamically acquire adaptive characteristics of vibration signals over time, with the attention mechanism effectively capturing the discriminative features of spindle vibration. Subsequently, the latent features are fed into a nonlinear feature mapping. Finally, the distances between hidden layer features are measured in the siamese network architecture to classify the vibration signals of machine tool spindles. Taking the state of the XK7132 machine tool spindle system as the research object, a small sample set of vibration signals is obtained and expanded using the CGAN method. Comparative experiments are conducted with the proposed method and two other methods, namely multiple transformation domain fusion and improved residual dense networks (MTDFIR-Desnet) and multi-size wide kernel convolutional neural network (MWK-net). The results show that the proposed method achieves a spindle system state identification accuracy of up to 94.7%, which is higher than that of the other two methods, and offers more concentrated sample clustering and better robustness.