Pattern recognition is the core component of the electronic nose (E-nose). Traditional machine learning algorithms highly rely on the feature data selected manually for model training and testing. A complete experiment must be performed before the data can be further processed. To realize the automatic extraction of response features and simplify the model’s training and application process, a multi-task convolutional neural network (MTL-CNN) with a dual-block knowledge-sharing structure is designed to train a model for the E-nose system. This model can simultaneously perform three different classification tasks, for the purposes of target discrimination, concentration prediction, and state judgment. Only a few consecutive seconds of response data are needed to be input into the trained model to obtain various information about the E-nose. With the utilization of an unmanned gas-sensing test system, large-scale measurements of the E-nose can be carried out automatically. A baseline tracking algorithm (BTA) is designed based on the relative changes of short-term data, reducing the impact of long-term shifts. Over thousands of gas response processes and more than 10 million sensing data have participated in the training of the deep learning model. The 5-fold cross-validation method shows that the fully trained model has an outstanding generalization performance. After the baseline is tracked automatically, the accuracy of three tasks towards 12 kinds of volatile organic compounds (VOCs) is about 95% (type recognition: 95.2%, concentration prediction: 92.1%, status judgment: 97.3%) using only 4 s of sensing data during the response status of the E-nose. Our work shows the distinct advantages of combining “big data” and “deep learning” in the gas-sensing field and further proves that the employment of MTL-CNN can significantly improve the training and application efficiency of the E-nose.