Joint distribution adaptation-based transfer learning for status classification of blast furnace gas pipeline network

Blast furnace gas (BFG) is a typical secondary energy resource of the steel industry. Establishing an effective classification model to estimate the status of the BFG pipeline network is of great importance to maintain the system balance. During the production process, the amount of labeled samples for BFG pipeline status classification are very small, and it is rather expensive to re-label a large number of industrial data. Thus, a joint distribution adaptation-based transfer learning framework was proposed in this paper. The preprocessed data of Linz Donawitz converter gas (LDG) pipeline network were taken as an auxiliary training data set to improve classification accuracy of the BFG pipeline network. Firstly, an offset value between the source domain and the target domain was calculated and removed to improve the similarity of marginal data distribution between them. Then a Kernel Mean Matching based Label (LKMM) algorithm was proposed to estimate sample weights of the source domain for the conditional distribution differences between different domains. The experimental results of real industrial data demonstrated that, the proposed method could avoid the negative transfer and improve the classification accuracy. Our approach provides the reliable status information to control the balance of the BFG system.