关键词: COD, 在线检测, UVVis, 水体分布差异, 深度学习, 域对抗训练网络

Abstract:  For aquaculture water, the traditional spectral method is based on the establishment of Chemical Oxygen Demand (COD) detection model based on the independent and identical distribution of spectral data between the source domain (the existing sample library) and the target domain (the water in the detection site). However, when the distribution of the source domain is different from that of the target domain, the low error model derived from the source domain often performs worse in the target domain. Aiming at this problem, this paper puts forward the UVVis spectrum domain oriented training network (DAUVwpNet), the distribution of different source domain and target domain data mapping to the same feature space, make it as close as possible in the space distance, thus in the feature space to the source domain training objective function can also be migrated to the target domain in order to reduce the error of the model in the target domain. Experimental results show that for the same test data, the prediction error of DAUVWPNET is 0.78, which is lower than that of the traditional model (0.85). The correlation coefficient between the predicted and the measured value of DAUVwpPnet is 0.95, higher than that of the traditional model (0.89). It is shown that this network can efficiently align the distribution of characteristic spatial data of the two domains and reduce the COD detection error caused by the difference of distribution.

Key words: COD, online detection, UVVis, water distribution difference, deep learning, domainadversarial training network