Affected by the changing factors such as raw material properties, equipment consumption and operating conditions, the real industrial process generally has dynamic and time-varying characteristics, and the soft sensor model needs to be updated online to adapt to the change of characteristics. Due to the lack of on-line analyzer of operation index, the operation index data obtained by periodic manual test is sparse and the value is difficult to be guaranteed, which limits the on-line updating of the model. For this reason, a light active soft sensor method based on dual dynamic memory of data stream is proposed. We design data memory update strategies for process data and label data respectively, to adapt to the change of characteristics by dynamically maintaining two sets of cache data. Ebinghaus memory law is introduced to design and update cache process data, and then highly representative process data are screened based on cache process data. a micro-cluster dynamic memory strategy is designed to update the cache of label data, and the uncertainty of data is evaluated based on micro-cluster information to form a mixed sampling strategy combining representativeness and uncertainty. Finally, the updated soft sensor model is established based on the memorized label data, and the active soft sensor method of the industrial process is realized in the case of only light data storage, the method considers the limited storage and computing power in real industrial systems, especially in the current context of Industry 4.0. In the process of transformation and development of industrial systems to end-cloud-edge collaboration, the problem of whether the model can be deployed on the edge is considered, avoiding the additional deployment cost of upgrading equipment. The experimental research is carried out by using the field coal preparation production data to prove the effectiveness of this method.
 

industrial process; dynamic memory; light active soft sensor model