In this paper, Smith-Fuzzy PID is selected as the control strategy to construct a housing heating control system containing acquisition control and monitoring by the upper computer, which realizes the optimal control of heating. Meanwhile, an automatic diagnosis method for heating system faults based on recurrent neural networks and dynamic threshold algorithms is proposed. Using the LSTM network in deep learning, a prediction model for the target sensor observation is constructed, and the error between the model prediction and the actual observation of the sensor is calculated. The DT algorithm determines the abnormality threshold, and if the error falls below that threshold, the system operates normally, and vice versa. It is diagnosed as a system fault. The results show that the Smith-fuzzy PID control is capable of restoring the heating system to stability in only 0.240×104 s at the time of failure, which is significantly better than the PID and fuzzy PID control methods. Meanwhile, the examination of the automatic diagnosis model of heating system faults found that the average F1 values of the LSTM-DT algorithm are all above 90%, and the diagnosis performance is better in the strong noise condition, which has good practical value.