Manufacturing cost anomalies occur through interactions among production processes, resource consumption, equipment conditions, and product quality. Conventional data-driven soft sensors lack interpretable links between sensor measurements and financial outcomes and do not incorporate manufacturing cost relationships. To address these limitations, we developed a counterfactual temporal graph neural network (CF-TGNN) for sensor-data fusion and its use for cost-variance analysis. The model integrates a domain-constrained sensor–cost embedding layer, a production-line similarity graph, temporal graph propagation, and counterfactual normal-cost generation with structural consistency constraints. These elements enable the transformation of heterogeneous sensor data into structured cost representations, capturing relationships among production lines and identifying anomalies through deviations from estimated normal costs. The model was evaluated using 35,464 weekly observations from 341 production lines across 64 manufacturing companies for 104 weeks in China. The CF-TGNN model showed an F1-score of 0.840 and an area under the receiver operating characteristic curve of 0.963, outperforming conventional machine learning, deep learning, and graph neural network baselines. It also achieved an average early-warning horizon of 2.36 weeks, with one-period and two-period early-warning rates of 76.84 and 58.19%, respectively, and a missing-alarm rate of 17.10%. Robustness test results presented consistent performance under alternative anomaly definitions, graph configurations, and selected industry-specific evaluations. By integrating manufacturing cost knowledge, temporal graph learning, and counterfactual inference, we developed a model that conducts sensor-data-driven cost-variance analysis. Using finer-grained data and through deployment in different industries, we can further enhance the model performance.