MRNEC recently developed within COGNIPLANT a methodology based on deep learning for predicting and preventing stops of parallel flow regenerative shaft kilns which produce lime at Fornaci Calce Grigolin S.p.A. Already a vanilla LSTM neural network appears to be able to capture, learn and embed the observed process states as multi-variate time-series concerning kiln states, lime quality, fuel aspects, set points for e.g. lime production and heat gas emissions, alarms, stops, automated or operator exerted control actions. After framing (40 +1) parallel contiguous time series of relevant process parameters (including KPI = time_start_to_stop until next stop) for supervised learning of predicting one step ahead prediction of time_start_to_stop given 10 time lags / time steps of 41 process parameters (including time_start_to_stop parameter) as input, the neural network predictsreasonably well time-start-to-stop of the kiln are reasonable if compared with the actual ones; neural network does embed the time-start-to-stop of kiln correlation with the other process parameters. Of course, there are overshoots and deviations as the model as such is not focusing one particular cause for a stop nor has it been adapted to the underlying timescales. As part of the new deep learning methodology MRNEC also developed various types of deep learning models and neural networks for predicting or rather inferring changes in set points to reduce reduce kiln stops and/or improve lime quality at appropriate time-scales.
top of page
Deep Reinforcement Learning
​
Nonlocal Multi-Scale Complex Interaction Network Analytics and Predictive Distributed Control
Critical Infrastructures
Health and Home Care
Business Intelligence
Autonomous Transport and Logistics
Smart Regions, Industry and Nations
​
​
bottom of page
コメント