Deep neural network(DNN) services are now widely used in industrial defect detection applications. With the increasing number of pre-trained model services on MaaS platforms like HuggingFace and inside smart enterprises, fine-tuning or directly applying DNN services has become a new solution for building intelligent applications. However, selecting appropriate services for tasks with various industrial requirements is also challenging work. Existing DNN model recommendation systems typically categorize models based on a limited set of task types or leverage the training data similarities. However, they fail to reflect the DNN service's native transferability and dynamic ability in the specific industrial scenario (i.e., pragmatics). In this paper, we introduce DSR, a novel pragmatic-information-model-based DNN service recommendation approach, designed to retrieve the most suitable services by incorporating information across the scene of industrial tasks and the ability of services. Through graph convolutional networks, DSR embeds the pragmatic information model of services into unified vectors and applies a regression model for usefulness-oriented recommendation towards specific industrial tasks. Additionally, we established a benchmark dataset with hundreds of customized tasks derived from public datasets with open-source services, on which we evaluate DSR compared to existing methodologies, including ImageDataset2Vec, AutoMRM, and TransferGraph. Our results demonstrate DSR's superior performance in terms of accuracy, efficiency, and generality. We also conduct a case study on an industrial surface defect detection scenario, which illustrates the feasibility of the system.