General aviation flight training exceedance event records are characterized by instantaneous discreteness, data scarcity, and missing features. Associated risk patterns are difficult to be captured by traditional single-factor or continuous-data-based analysis methods. A risk analysis method integrating data augmentation, deep clustering, and association mining was proposed. Flight feature snapshots were augmented with high fidelity by utilizing the Wasserstein generative adversarial network with gradient penalty(WGAN-GP) model. Flight parameter intervals were delineated by introducing the VBGMM. Latent associations among combinations of parameter intervals were mined by applying the FP-Growth algorithm. A total of 259 strong association rules were mined. Four types of core risk patterns, including the speed-attitude bimodal risk of hard landing events and the rate-of-change mismatch of bank angle warning events, were identified with the assistance of knowledge graphs. Specific indicated airspeed intervals were revealed by the cross-event risk network as common risk factors connecting multiple types of exceedance events. Ideas for solving risk analysis problems of similar numerical discrete data were provided. Parameter interval association rules were transformed into topological networks. Theoretical basis and decision support were provided for optimizing the early warning logic of flight quality monitoring, implementing targeted flight training, and improving the efficiency of the risk control system.