To ensure the safety of industrial processes, it is crucial to develop accurate fault diagnosis models. However, the labeled samples collected in actual industries are usually limited, while a large number of unlabeled samples are not utilized, resulting in sample label imbalance. This hinders the model from learning the structural information of the samples and reduces the diagnostic performance. To overcome this limitation, this paper proposes an improved FixMatch-assisted semi-supervised conditional Wasserstein generative adversarial network fault diagnosis method (IFixMatch-CWGAN). Firstly, to effectively utilize unlabeled samples, an adaptive smooth dynamic threshold was designed by fusing the quantile threshold and the cosine annealing threshold, thereby optimizing the pseudo-label screening strategy of FixMatch. Secondly, to reduce the risk of overfitting in adversarial or classification single tasks, a shared feature layer is designed in the discriminator to achieve collaborative training of unlabeled samples. Finally, the trained semi-supervised CWGAN is used as a classifier for fault diagnosis. The experimental results on the Tennessee Eastman process show that the proposed method has high diagnostic accuracy in a sample label imbalance environment.