Fault Diagnosis And Advanced Process Control Strategy Research Application Of Tobacco Industrial Drying Process

Industrial fault diagnosis refers to judge and analyze whether the industrial process has abnormal working characteristics or abnormal forms according to the information obtained from the test,finding the location where the abnormal state occurs,and isolating the fault.With the increasing scale of industrial production,the probability of direct or indirect equipments fault increases.To make sure the normal operation of the production processes,it is of great significance to find and eliminate faults in time to ensure the stable and safe operation.The drying process is a typical industrial process with high complexity,strong coupling and high energy consumption.The drying process contains physical and chemical reactions,and the affecting factors of the drying process are characterized by uncertainty.Therefore,when fault occurs in the drying process,it is not practical to judge the fault solely by a sign or expert prior knowledges.Bayesian Network(BN)is one of the effective methods to deal with uncertain knowledge and causal inference.BN is a method to express the dependency relationship between variables and provide knowledge visualization.It combines graph theory and probability organically.It can not only reveal the hidden relationship between data,but also deal with the uncertainty of industrial process under the limited data set.Taking the tobacco industry drying process as the research object,this paper proposes the fault diagnosis method of tobacco drying process based on bayesian network and the control strategy of tobacco drying process caused by components faults and time-varying delay changes.Around the tobacco industry drying process,there are many kinds of equipments,complex process flows and many parts prone to fault,it is a huge task to find the location of the fault and deal with the fault,constructing bayesian network fault model by putting the leaf tobacco drying process network,learning the fault network parameters under the limited data set,and deducing the reason and probability of the fault.In view of the problems that the delay caused by the fault is easy to make the system unstable,the controlled variable deviates from the set ideal value,and the anti-disturbance inhibition effect is delayed,the positive dynamic compensation of the delay is used to offset the negative effect of the delay,so as to achieve the goal of stabilizing the system.This paper mainly studies fault diagnosis based on bayesian network and advanced process control strategy of tobacco industry drying process.The main research contents include fault diagnosis and advanced process control.The fault diagnosis part studies bayesian network structure learning,parameter learning and fault diagnosis of industrial tobacco drying process based on bayesian network,builds a fault model of tobacco drying process and carries out accurate reasoning.Process control studies the control strategy of delay caused by fault.Aiming at the large delay caused by faults in the system,the controller is designed by dynamic forward compensation,and the system oscillation caused by disturbances is eliminated.The specific research contents and innovations of this paper are as follows:A Recursive Hyper-graph Decomposition Principal Sub-Block Bayesian Network Structure Learning Method Based on Integer Programming Is ProposedIn view of the large number of nodes,the traditional structure learning method is easy to make the structure search fall into local optimization,and the network structure learning needs large sample data.The bayesian network structure learning problem under small sample data set is studied.Based on the moral graph decomposition theory,a fully conditional independent recursive hyper-graph decomposition main sub-block graph algorithm based on integer programming is proposed,and adds polyhedron approximation constraints to the decomposed low-dimensional small network,constantly seeks the highest scores of nodes in the network and their different parent nodes,selects the pair with the highest score as the optimal structure of the low-dimensional network,and merge all the best networks with the highest score according to the causal relationship to achieve the best network with the highest degree of data consistency.This algorithm proves theoretically that the recursive hyper-graph decomposition algorithm based on full conditional independence is an equivalent decomposition of the original structure,does not change the properties of the original structure,and the segmented cluster is the largest main sub-block.Moreover,compared with the previous Max-min Hill Climbing(MMHC)with previous excellent performance and Markov equivalent class structure recursive learning algorithm(REC)based on graph model decomposition,the accuracy of the new algorithm proposed in this paper is improved by about 10%,and the running time is reduced by about 20% on average,the structure accuracy is higher,the time required to generate the optimal structure is shorter,and the balance between accuracy and computational efficiency is better achieved.A Bayesian Network Parameters Leaning Integrating Monotonicity Constraints and Information Entropy Is ProposedIt is difficult to obtain the exact conditional probability table of discrete bayesian network parameters based on data-driven traditional parameters learning meth-ods under limited data sets.This paper studies the parameter learning problem of bayesian networks based on information in small sample data sets,a bayesian network parameter learning method with monotonicity constraint and information entropy fusion is proposed based on entropy theory.On the premise of knowing the structure of bayesian network,the parameters are taken as the uncertainty information entropy and discussed in the global scope,the monotonicity between parameters is used to construct the maximum information entropy function about parameters,and the monotonicity of parameters is used to analyze the information,the entropy function is used to constrain,and the convex optimization problem is transformed into an unconstrained problem through re-parameterization,finally,the information entropy function is solved by the KKT method and the parameters are obtained.The method proposed in this chapter theoretically proves that the monotonicity constraint is a set of relational constraints applied to parameter estimation.As a kind of expert prior knowledge,the monotonicity constraint can be used as a condition to constrain the upper and lower bounds of parameters under small sample data sets,experiments with different sample capacities on nine standard bayesian networks are carried out,and compared with previous parameter learning methods such as MLE,MAP,CO and CML,the learning accuracy of the proposed algorithm is improved by about 30%.For small networks with few nodes,the learning performance is improved by about 15%,while for medium and large networks with many nodes,the learning performance is improved by about 5%.A Fault Diagnosis Method of Drying Process Based on Bayesian Network Is ProposedIndustrial drying is an uncertain and strongly nonlinear process,especially for the system containing mechanical and electrical equipments,which has correlation and coupling between components,contains a lot of uncertain factors and information,and the generated faults are also interrelated.The traditional single fault diagnosis method cannot solve the data tolerance problems caused by large volume and strong disturbance,the fault diagnosis method based on bayesian network is studied.The tobacco re-baking drying process and cutting tobacco drying process is constructed with the fault symptom as the node,the fully conditional independent recursive decomposition hyper-graph algorithm proposed in Chapter 2 is used to construct the fault model based on bayesian network for the process of industrial tobacco drying,then,the parameters of the fault network are learned by using the algorithm with monotonicity constraint which integrates information entropy proposed in Chapter 3.On the basis of considering the network scale and data characteristics,the fault tree and bayesian network are combined and transformed to calculate the probability of fault occurrence.Based on the fault tree,the fault model was accurately reasoned by bayesian network,and the possible causes and probability of the fault in the drying process of industrial tobacco leaf were calculated and reasoned.In theory,the fault factors and the probability of the fault of the main equipment in tobacco drying process were deduced by using bayesian network accurate reasoning through fault tree.In practical application,it is proved that the proposed method has good reliability in fault diagnosis,and can infer the probability of failure of each component in industrial tobacco drying process,so as to provide reasonable and accurate information for operators to take necessary measures for maintenance.Advanced Control Strategy of Industrial Drying Process with Faults Is StudiedIn view of the large delay and instability of industrial drying process caused by components faults,external disturbances and self-factors,the control problem of industrial tobacco drying process with large delay was studied.According to the cascade control theory,an improved series dynamic delay compensation control strategy based on cascade control is proposed in this paper,dynamic compensation is added in the inner and outer loops respectively to offset the delay link in the form of compensation.An disturbance suppressor is designed in the inner loops,and a set-point tracking controller is designed in the outer loop to eliminate disturbance and overshoot.According to the experts priori knowledge,reasonable adjustment of parameters is tuned,and the control goal of good robustness and stability is achieved.Taking the drying process of cutting tobacco drying as an example,a simulation of moisture content of drying process was built,and the advantages of dynamic compensation structure were proved in three cases.Firstly,the control strategy with dynamic compensation is compared with the previously existing method(without dynamic compensation strategy),and the dynamic response is better and the secondary disturbance can be well suppressed.Secondly,compared with the existing methods(without dynamic compensation strategy),the control strategy with dynamic compensation element has a good performance in dynamic response and overshoot in predicted PI and IMC.Thirdly,under model mismatch,the control method with dynamic compensation structure not only has good dynamic response,but also is stable and has no overshoot,while the previous method(without dynamic compensation strategy)is out of control and the response curve oscillates violently.Finally,this paper summarized the related achievements of the research on fault diagnosis and advanced process control strategy of industrial tobacco drying process,and looks forward to the future development trend and problems to be solved in this field.