Fault diagnosis is crucial for wind turbine converter, which often faces problems of the unbalanced samples of different categories. Thus, sample generation techniques are often used. However, traditional models like GAN are difficult to generate samples of specific categories. To address this problem, an ACLSGAN (auxiliary classifier least squares generative adversarial networks) model is used to optimize the training data of the fault diagnosis model. Moreover, a sample enhancement algorithm is proposed for anomaly detection of wind turbine converter based on an adversarial network with the least squares loss generation of auxiliary classifiers. In the ACLSGAN algorithm, the least squares loss function is used to alleviate the pattern collapse problem and enhance the stability and generalization ability of the model. With the introduction of auxiliary classifiers, ACLSGAN better learns to generate data distributions related to specific categories, which significantly improves the model ability to learn to generate data distributions and improves the quality of the generated samples. The generated samples are then used as training data to train the DeepForest model to improve its performance and accuracy in the practical application of wind turbine converter fault detection. The experimental results show that compared with other oversampling methods, the ACLSGAN fault detection method achieves lower leakage and false alarm rates while ensuring high accuracy.