With the widespread application of high-voltage shunt reactors in power systems，abnormal phenomena arising during their operation have attracted increasing attention.In existing latent regularization adversarial anomaly detection （LRAAD） methods，the hyperparameter M imposes a hard constraint on the upper bound of the KL divergence of generator-produced spectrograms，which hampers the model ’s ability to effectively distinguish normal from abnormal data in the latent space，leading to training instability and degraded anomaly detection performance.To address this issue，this paper proposes a soft-constrained latent regularization adversarial anomaly detection （Soft-LRAAD） method.The proposed method introduces a soft constraint loss to replace the hard constraint loss，and enhances the discrimination capability in the latent space and the training stability by approximating the upper bound of the KL divergence using a smooth function.Experimental results demonstrate that the proposed method effectively improves the accuracy and robustness of anomaly detection for high-voltage shunt reactors，providing a superior solution for power equipm ent fault diagnosis.