Traditional construction project cost estimation methods rely on expert experience and statistical models, which are difficult to handle complex data and multimodal features effectively and have low prediction precision. This paper constructs an intelligent building engineering cost estimation model that combines subtractive clustering, a self-learning mechanism, and convolutional neural networks (CNN) to address this problem. In the data preprocessing stage, subtractive clustering is applied to optimize multimodal data, screen key features, and eliminate redundant information. Subsequently, the model parameters are dynamically adjusted according to the error feedback through a self-learning mechanism to improve its adaptability to diverse construction projects. In the feature extraction and estimation stage, the CNN module is combined to extract deep features from images, texts, and numerical data to achieve high-precision estimation. The experimental results show that the model in this paper outperforms traditional methods in terms of MSE (mean-square error), MAE (mean absolute error), R² (coefficient of determination), MAPE (mean absolute percentage error), with the mean values being 73.18, 8.33, 0.9477, and 5.33%, respectively. In summary, the model in this paper demonstrates superior precision, adaptability, and robustness in construction project cost estimation.