Time series anomaly detection has become a key technology for maintaining the stable operation of Internet of Things (IoT) systems by identifying potential faults and security threats. Existing methods face limitations in capturing the temporal correlations, cross-dimensional spatial dependencies, and local dependencies in the high-dimensional time series data from multi-source sensors. We propose an anomaly detection model, MSSTT-GAN, based on Multi-Scale Spatio-Temporal Transformer (MSSTT) and Cycle Generative Adversarial Network (CycleGAN). The model adopts a cycle generative adversarial architecture, integrating the MSSTT module into the generators. The temporal Transformer captures temporal features, while the spatial Transformer analyzes the topological relationships among multiple sensors. A multi-scale patch mechanism is introduced to extract fine-grained features within patch and across patch spatio-temporal associations, enabling local dependency modeling. The discriminators use a spectral normalization convolutional neural network, applying spectral norm constraints to the convolution layers, thus enhancing the stability of adversarial training. The experimental results show that MSSTT-GAN outperforms existing models in terms of F1 score, with an average detection performance improvement of 6.7%. Meanwhile, the Population Stability Index (PSI) is optimized by 87.9%, and the model stability is significantly enhanced.