To improve the detection efficiency of the agricultural vehicle engine production line, research has been conducted on the accurate detection of the air‑tightness of the engine water pump by using multi‑channel visual imaging and non‑linear illumination compensation methods. An algorithm based on a hybrid adversarial 3DCNN has been developed to enhance the efficiency and accuracy of bubble video detection. Additionally, the accuracy and speed of small bubble detection have been improved by using the deformable depth convolutional network and transfer learning‑based object detection method. The research results showed that the average precision (mAP), the average recall (AR), and the average accuracy (AP) of the 3DCNN model by using updated datasets were increased to 92.39%, 95.01% and 95.25%, respectively, which indicated that regularly updating the training dataset had a positive impact on improving the recognition accuracy and adaptability of the model. The results indicated that by incorporating TDD—Net into the 3DCNN model for detecting the air‑tightness of the automotive engine water pump casing, the average precision, the average recall, and the average accuracy had been increased to 96.27%, 97.72% and 96.37%, respectively. This demonstrates that deep learning models can effectively enhance the accuracy and robustness of tasks such as image recognition and classification when integrating advanced structures. These results provide a new technical approach and method for the air‑tightness detection of agricultural vehicle engine water pumps, and also provide more reliable and efficient technical support for agricultural production.