Modeling of Trailing Suction Hopper Dredger: A Two-Step Method Integrating Deep Neural Network and Human Knowledge

Learning the quantitative and explainable relation between the dredging performance and operating variables from big sensor data is an important yet challenging problem in Trailing Suction Hopper Dredger (TSHD) applications. The deep learning technique cannot be used directly due to two factors, of which one is its black-box nature, and the other is the one-to-many mapping from the dredging performance to the domain of parameters. In this paper we propose a two-step framework combining deep learning and human knowledge for the modeling of dredging performance. The first step is parameter learning, in which deep neural network is adopted to find the implicit relation between operating variables and the performance. Based on this implicit model, we regenerate a new dataset, which is combined with human knowledge in the second step to build the explicit model with human knowledge. The explicit formula between operating variables and dredging performance are obtained by solving a sparse optimization problem. With this formula, we can choose feasible combinations of controllable variables for given performance indices. The effectiveness and advantage of our proposed scheme are verified by practical data collected in a sand extracting project.