<p>Data reconciliation is critical for ensuring the integrity of data in process industries. Statistical methods of data reconciliation such as least squares need precise error statistics, whereas deep learning approaches frequently disregard underlying physical constraints or require clean ground-truth labels that are rarely available in practice. To overcome these limitations, we propose the Robust Estimator-guided Neural Network (RENN). This approach is a unified framework that integrates data-driven learning with physics and robust statistics. RENN directly embeds a Quasi-Weighted Least Squares (QWLS) robust estimator into a physics-informed learning objective. The proposed approach enables the network to train end-to-end on noisy industrial measurements alone, without the ground-truth data. By simultaneously minimizing physical constraint violations (e.g., mass balance) and a robust reconciliation loss objective, RENN provides accurate reconciled estimates in a single, feedforward pass, making it computationally suited for online deployment eliminating the need for per-sample iterative optimization. We validated the RENN framework against multiple process benchmarks, including a complex steam metering system. The results show strong alignment with ground truth (R² &gt; 0.95 across most streams), substantial noise reduction (over 80% Mean Squared Error reduction in high-error streams), and a near-perfect mass balance closure (&gt; 99.9%). In comparative analysis, RENN consistently outperformed or matched six traditional estimators and supervised neural network baselines in relative error reduction. The RENN framework provides accurate, and physically consistent solution removing the need for curated training data, facilitating robust process monitoring for applications demanding fast, online data reconciliation.</p>