Modeling Research On Particle Filter Network For Aluminum Electrolysis Manufacturing System

Aluminum and its alloys are widely used in infrastructure projects and strategic emerging industries due to their excellent performance,low production cost and mature manufacturing technology.It is mainly produced in batch by electrolysis with alumina as the solute and cryolite as the solvent,in which material replacement and energy exchange occur constantly.Meanwhile,the external environment is also frequently accompanied by anode replacement,bus adjustment and other processes alternately.The energy consumption of traditional aluminum smelting process is often high.The energy utilization rate is less than 50% and the pollution is serious,which is incompatible with the production concept of energy-saving and emission-reduction in China.Therefore,the research on energy-saving and emission-reduction technology in aluminum electrolysis manufacturing system(AEMS)has important engineering application value to improve production efficiency,reduce energy consumption and protect natural environment.Aluminum electrolysis intelligent manufacturing based on machine learning,intelligent modeling and optimization strategy has become a trend of energy saving and emission reduction and green development in aluminum electrolysis process equipment.How to deeply integrate the AEMS with the advanced artificial intelligence technology is studied,which has a significant strategic position in promoting the level of intelligent aluminum electrolysis manufacturing.Due to the complex characteristics of AEMS,such as parameter redundancy,dynamic time variance and a non-Gaussian distribution of process data,how to accurately establish a dynamic adaptive model that can reflect the real aluminum electrolysis process has become an urgent difficulty to realize the intelligent decision optimization of the system.However,the intelligent level,data processing ability and model prediction accuracy of AEMS are not up to the ideal requirements.The main reasons are as follows:(1)The mechanism of AEMS is extremely complex,and it is difficult to express clearly in a concise form.Even if the presupposition is given,it is still difficult to guarantee the consistency between the obtained model and the actual process.(2)The noise of aluminum electrolysis process often presents the complex characteristics of non-linear and non-Gaussian,so it is still a technical problem to explore a strong robust model suitable for various noise interference.(3)The decision variables of aluminum electrolysis process influence each other and redundancy is serious.How to deal with the coupling problem of decision variables in high-dimensional space is also a research focus.(4)The AEMS needs to exchange material,energy and information with the external environment continuously.If the model is only static,it is difficult to adapt to environmental changes.To solve these problems,this paper has carried out the following research work:(1)Non-mechanism dynamic model of particle filter networkWhen the mechanism of the system is fuzzy,this paper constructs a novel particle filter neural network model.Based on the research data itself,the model takes the neural network's(NN's)weights and thresholds as the particle filter's(PF's)state variables,and the NN's outputs as the PF's measurement variables.The dynamic approximation of particle filter is used to adjust the NN's weights and thresholds in real time.The experimental results show that the particle filter network model has a certain degree of robustness and scalability,which not only broadens the filtering neural network system,but also provides a model basis for subsequent optimization.(2)High precision model of hybrid annealed particle filter networkAiming at the negative effects of the PF's inherent characteristics,based on the particle filter network,the standard particle filter is replaced by hybrid annealed particle filter,and the hybrid proposal distribution is used as the importance function to replace the posterior distribution in the standard particle filter.By exploring the annealing factor to adjust the relationship between state noise and measurement noise,the hybrid proposal distribution is closer to the likelihood distribution.The experimental results show that the hybrid annealed particle filter network improves the defect of particle filter and improves the prediction accuracy of the model.(3)Clustering optimization strategy based on geometric manifold energyIn order to further tap the above model's prediction potential,this paper firstly uses the local linear embedding(LLE)algorithm to reduce the data dimension in high-dimensional space.Then,the geometric curvature of reduced-dimension data is used to represent the manifold energy,so that the manifold energy can be minimized to get the boundary points,and the clustering region can be divided.Finally,the particle sparseness mesh resampling is introduced to improve particle diversity loss.Based on the ablation design of a novel type of aluminum electrolysis cell,the experimental results verify the rationality and effectiveness of the proposed optimization strategy.