Abstract: To address the issues of overfitting in purely data-driven predictive models under scenarios with low number of samples and high noise in industrial systems, as well as the heavy reliance of Physical Information Neural Network （PINN） on strict physical equations, this paper proposes a model prediction method based on feature learning enhanced Physics-Informed Autoencoder （PI-AE）. This approach introduces a prediction branch on top of the autoencoder's feature reconstruction, integrating physical monotonicity constraints to design a “reconstruction-physical-prediction” composite loss function. This achieves synergistic optimization of robust feature learning, physical consistency, and accurate prediction. Numerical experiments demonstrate that PI-AE achieves a 46.1% reduction in test-set RMSE compared to PINN and a 97.4% reduction compared to BP neural networks. In acetylene concentration soft measurement tasks, ablation studies systematically validate the progressive performance gains of this multi-task framework. The PI-AE method effectively overcomes limitations of traditional models, establishing a new paradigm for industrial few-shot prediction.