In order to reduce the experimental cost required for the verification of tensile strength properties of Engineered Cementitious Composites (ECC), and overcome the shortcomings of traditional machine learning models with limited prediction accuracy and relying on manual parameter adjustment, In this study, a prediction method based on Bayesian Optimized Convolutional Neural Network (BO-CNN) is proposed. It is used to predict the tensile strength of PolyEthylene fiber-reinforced cement matrix composite (PE-ECC). Firstly, the complex nonlinear mapping relationship between ECC mix ratio and tensile strength is effectively mined by Convolutional Neural Network (CNN). Secondly, the Bayesian optimization algorithm is used to automatically search for the optimal combination in the high-dimensional hyperparameter space, so as to reduce the time and labor cost of manual parameter tuning. Finally, interpretability analysis methods, including SHAP (SHapley Additive exPlanation) values and Partial Dependence Plots (PDP), are introduced. From the dual dimensions of feature importance ranking and interaction effect visualization, the synergistic enhancement mechanism between the water-to-binder ratio (W/B) and fiber parameters is deeply analyzed, verifying the model's ability to learn the underlying physical nature of the materials. In this study, a database was constructed based on 154 sets of PE-ECC tensile strength experimental data collected from published literature. The results show that: The Coefficient of Determination (R2) of BO-CNN model was 0.911. Its prediction performance is better than grid search optimized CNN model, Artificial Neural Network (ANN) model and 11 compared traditional machine learning models. This method realizes the accurate prediction of the tensile strength of PE-ECC, and provides an effective theoretical tool for the optimization design of ECC mix ratio.