For solving the problems of dispersive data, non-uniform metrics, and not enough explanation of outcomes in yearly performance assessment work of university teachers, a deep belief network model that has missing-value-conscious inputs and double-task outputs was built. According to name-hidden materials from four same-grade undergraduate universities that cover the years 2021 to 2024, this research has collected 1,248 yearly teacher assessment samples. The raw data was systematically organized into five primary dimensions-teaching contributions, research output, student development support, public service, and professional growth-comprising 18 secondary indicators and 64 computable variables. Test results show that the model achieves a Mean Absolute Error (MAE) of 3.31, a Root Mean Square Error (RMSE) of 4.29, an R² of 0.895, an Accuracy of 0.861, and a Macro-F1 of 0.832, outperforming Linear Regression, Random Forest, XGBoost, Backpropagation Neural Networks, and the standard DBN.Compared to DBN-base, the RMSE decreased by 12.1%, and the Macro-F1 score increased by 3.4 percentage points. Robustness experimental outcomes show that when the missing data rate is increasing from 0% to 15%, the value of RMSE is risen only from 3.92 to 4.36; the descending of performance is more obvious when the two items, which are research output and student support variables, all encounter high degree of interference at the same time. The outcome shows that this model can give comparatively steady quantification support and a distinct explanation interface for university teacher performance assessment.
