Industrial fault diagnosis is a critical challenge in complex systems, where sensor data is often noisy and interdependencies between components are difficult to capture. Traditional methods struggle to effectively model these complexities. This paper presents a novel approach by transforming fault diagnosis into a graph recognition task, using sensor data represented as graph-structured data with the k-nearest neighbors (KNN) algorithm. A Graph Transformer is applied to extract node and graph features, with a combined loss function of cross-entropy and weighted consistency loss to stabilize graph representations. Experiments on the TFF dataset show that Graph Transformer combined with consistency loss outperforms conventional methods in fault diagnosis accuracy, offering a promising solution for enhancing fault detection in industrial systems.