Unsupervised machine learning approaches are particularly valuable in Structural Health Monitoring (SHM), where most sensor data is unlabeled. However, such models often face performance limitation due to the absence of datasets with negative examples. This research aims to address these challenges by incorporating a probabilistic approach into advanced data driven (DD) generative models for unsupervised anomaly detection. In this research, we will utilize a comprehensive dataset of acceleration and strain signals, along with environmental and operational data, collected over five years from prestressed catenary poles in Germany. We have selected Deep Convolutional Generative Adversarial Networks (DCGANs) due to their effectiveness in learning complex data distributions. By integrating Bayesian inference into both the generator and discriminator, we employ minimax and Wasserstein loss functions to enhance the adversarial training process. By defining prior distributions for the weights and updating them based on observed data, the model can adaptively refine its parameters, leading to better anomaly detection and forecasting capabilities. Performance evaluation will be based on metrics like reconstruction error and density-based measures. We anticipate that the integration of Bayesian inference to the DCGANs will significantly enhance the model's performance.