Long-term stagnation monitoring using machine learning: comparison of artificial neural network model and convolution neural network model

In this study, a device to diffuse the flow of water in a horizontal direction was installed over a small river connected to Nakdonggang River and the dissolved oxygen (DO) concentration within the range of its influence was monitored. A DO probe was installed and operated at three depths of water; the surface layer, middle layer and deep layer. In order to judge stagnant water by operating and controlling the device automatically, an artificial neural network model that worked through profiling by logics and expert learning was applied. For expert learning, the number of all cases generated from DO data was labeled based on expert judgment. In other words, when DO concentration was divided into 7 levels, the number of cases was 343, the experts were requested to determine whether each case was a stagnant water. Machine learning was carried out targeting labelling by experts with the artificial neural network (ANN) and the convolution neural network (CNN). The target datasets for learning were 3 × 1 based on numbers from 1 to 7 and 7 × 7 based on the dot graph. The correct ratio for the ANN model learning result based on the graph was only 29.2%, so it was excluded. The correct ratio for the ANN model learning result based on numbers was 87.2%. The correct ratio for the CNN based on the graph was 94.2%. When machine learning was carried out with 30 to 300 randomly selected targeted graphs, the ANN model showed 74.6% as the correct ratio for up to 150 graphs, which was somewhat low, while the CNN showed 84.3% for 30 graphs and 94.2% for 200 graphs, a gradual increase with results comparable to the total number of graphs. By applying the relevant control logics to actual monitoring results, 91.5% and 87.4% was judged to be stagnant water from points directly and indirectly affected by the device, respectively.