| In the processing industry,most complex devices or systems have strong non-linearity,high stochasticity,and long time delay.These properties lead to the difficulties in analyzing system mechanisms,identifying dynamic equations,and controling critical operational indices.Mechanistic-based dynamic system identification methods are not ideal to sufficiently fit the complexity of the system.Deep learning methods ignore the limitations from system mechanisms and prior knowledge,which demands a larger size of training data and lacks generalization in different scenarios.From the continuous-time domain,this dissertation focuses on the application of deep differential equation networks,a combination of continuous time systems and deep learning,in the identification and control optimization of complex systems.(1)For the modeling of non-linear and long-delayed complex industrial systems which are essentially continuous-time processes,this dissertation proposes a continuoustime deep system identification model based on the ODE-net backbone.The model learns the auto-regressive changes of industrial system outputs and the non-linear and timedelayed effects from inputs on outputs in continuous-time domain.The evaluation conducted on real industrial datasets of paste thickeners indicates that the model achieves comparable performances on both long-time and short-time predictions.(2)For the modeling of the complex systems with stochasticity,this dissertation proposes the Ordinary Differential Equation Recurrent State Space Model(ODE-RSSM).On the basis of Recurrent Neural Network,an additional stochastic path defined by latent variable is embedded in the state transition for modeling stochasticity.Furthermore,the ODERSSM incorporates an ordinary differential equation network to model the continuoustime evolution of latent states between adjacent time points.Inspired from the equivalent linear transformation on integration limits,this dissertation also propose an efficient reparameterization method for solving batched ODEs with non-uniform time spans in parallel for efficiently training the ODE-RSSM with irregularly sampled sequences.Extensive experiments on three datasets,including the paste thickening dataset and two public system identification datasets,demonstrate that the ODE-RSSM outperforms other baselines in predicting stochastic system in open loop when the time spans of predicted points are uneven.(3)For the controling of complex industrial systems without prior knowledge of mechanism,this dissertation proposes a novel continuous-time model-based reinforcement learning control algorithm,Heuristic Critic Network Value Iteration(HCNVI).The method models the dynamic system with differential equation neural network and introduces the integral reinforcement learning to approximate the adaptive critic function under the control law.Depending on the evaluation of critic network and the prediction of model network,the controlling action is determined by stochastic gradient descent algorithms.Experiments conducted on thickening simulation system verify that the proposed HCNVI outperforms the other control algorithms.In the meanwhile,the method is also deployed in a thickening control system of a real paste backfilling station.In comparison to the original rules-based controlling strategy,the algorithm is effective to control the underflow concentration in a stable range.(4)To address the problem of modeling periodic multi-staged system,this dissertation proposes continuous-time Autonomous Jump Ordinary Differential Equation(AJODE-Net)to learn continuous-time period jump systems.Given sequential inputs and outputs belonged to labeled stages,the model is able to independently learn the subdynamics in different stages from non-uniform sampled data.The model also introduces a stage transition predictor to realize the autonomous stage transformation in open-loop prediction.By evaluating the model with dataset from an cement addition system in paste production,the results indicates that the AJ-ODE-Net can accurately predict the system outputs and the duration time in each stage.The open-loop prediction of paste system runtime also assists in optimizing 8.2%cost of backfilling. |
