Petroleum oil products can produce three-dimensional fluorescence spectra with considerable intensity under certain excitation light, allowing the identification and analysis of petroleum pollutants. Because of the complex characteristics and huge data of the fluorescence spectrum of petroleum oil products, it is not easy to be described with a simple mathematical model, nor to rely on artificial observation and analysis. This paper presents a convolutional neural network (CNN) model constructed using raw fluorescence data of three petroleum products (gasoline, oil, diesel), which automatically learns features from training data and classifies petroleum pollutants in water, using its nonlinear computing ability and adaptive representation learning ability. Through various fluorescence spectrum experiments, training and validation spectral datasets of petroleum products are constructed, and the CNN model is established based on the deep learning framework Keras for Python. The CNN model is trained, validated, and tested on the spectral dataset, to classify and discriminate measured oil products. The experimental results show that the classification accuracy of the CNN fluorescence model on the training and validation sets of the three petroleum products is 99.76%, the classification accuracy in a comprehensive test is 82.65%, and the classification accuracy for a single substance is 100%. Additionally, the experimental results confirm the feasibility of combing three-dimensional fluorescence technology with a deep learning algorithm, to distinguish and classify petroleum products accurately and reliably. These results provide technical guidance for further research on creating intelligent identification models for environmental pollutants in water, as well as a new direction for environmental detection methods.