Detecting anomalies in multivariate time series (MTS) holds significance for monitoring the behavior of intelligent industrial and Internet of Things (IoT) systems. Nevertheless, due to the absence of anomaly labels in current IoT system data and the high-dimensional complexity of the data, constructing an anomaly detection model that accurately identifies anomalies while maintaining robustness poses a considerable challenge. The paper introduces an unsupervised anomaly detection model, the Multi-channel Heterogeneous Graph Transformer model for IoT Time Series Anomaly Detection (MC-HGT). Our model implements complex non-linear temporal feature extraction using TCN and SENet. The integration of parallel GNN and GAT within a heterogeneous graph neural network, connected with the Transformer network, efficiently discerns local and global temporal feature dependencies. It amplifies the model’s ability to extract temporal features by employing multi-channel feature fusion in conjunction with the Transformer. VAR is employed to capture additional linear features, facilitating the multi-channel fusion of linear and non-linear features to enhance model robustness. We evaluate the detection performance of MC-HGT against five baseline methods on three public datasets. The experimental findings reveal that the MC-HGT algorithm achieves a mean F1 Score of 0.905, surpassing similar algorithms by 6.93%, and a mean AUC of 0.943, exhibiting a 2.35% improvement over comparable algorithms.