Based on sensor data, fault recognition of industrial systems has been a research hotspot in recent years. Establishing a scientific and efficient fault recognition system and making timely adjustments to address the fault type are crucial for restoring the system to its normal operating state when a fault occurs. Compared to normal data, fault data is relatively scarce and suffers from imbalance issues. To address these two problems, the AR-TransWGAN model is proposed for augmenting time series data, and the GA-Transformer model is designed and implemented for fault recognition. Additionally, leveraging the desirable properties of KL points, the design of complex response surfaces is completed to guide the experimental design of the simulation model. The entire scheme is closed-looped into a fault recognition process tailored for sparse and imbalanced fault data. Experimental results on multiple public datasets and the Tennessee Eastman process demonstrate that the proposed comprehensive fault recognition process scheme holds significant reference value for addressing sparse and imbalanced fault recognition tasks.