A series of cases on coke quality prediction . The background of this service is that the raw coal excavated from the ground needs to go through a series of processes for coking and refining into coking coal before it can be used in production and life. The process of coking coal blending can be compared to prescribing Chinese medicine. The difficulty lies in the formula industry experience. After digitization, the key lies in the coke quality prediction function: input the coal blending list with different coal ratios to predict the output coke quality. As long as the coke quality can be predicted, the coal blending work can be carried out with a targeted approach.
The challenges faced in this scenario mainly include data heterogeneity and small samples:
Due to the existence of different coal types, processes and other scenarios in different factories, the prediction accuracy of some scenarios is difficult to meet production requirements using traditional artificial intelligence methods, and the average prediction accuracy is >0.95.
When considering large models, it is difficult to build the model kuwait mobile phone number list due to insufficient initial samples.
After the model was deployed for a period of time, the prediction accuracy gradually decreased due to changes in coal sources, processes and other scenarios, affecting business production.
7.png
Figure 6 Coke transportation and coking process
In the coke quality prediction series of cases, KubeEdge-Ianvs and KubeEdge-Sedna developed edge-cloud collaborative lifelong learning features to deal with the small sample problem first. Edge-cloud collaborative lifelong learning handles the small sample problem through incremental and migration mechanisms, small sample learning algorithms, and hierarchical architectures. At the same time, facing the problem of heterogeneous data at each edge node, it is solved by calling different task models under different working conditions.