A robust adaptive large-scale pneumatic conveying control method

Abstract

The present application provides a robust adaptive large-scale pneumatic conveying control method, including: receiving detection data including temperature, humidity, air pressure and air flow in the pneumatic conveying system; transmitting the detection data to a deep learning network, and outputting state parameters after operation of the deep learning network; sampling the state parameters, decision variables and optimization targets, and obtaining complete population data through an evolutionary algorithm; based on the complete population data, calculating and obtaining control parameters through a Monte Carlo tree search algorithm. The present application deploys multiple sensors in a pneumatic conveying system, and uses a neural network to analyze sensor parameters to obtain the state of the pipeline itself. Based on the state of the pipeline, the population in the swarm optimization algorithm is used as a node to construct a Monte Carlo tree, ensuring that a high-quality solution is searched in a large-scale decision space, thereby improving the robustness of the solution while ensuring that the pneumatic conveying system can optimize the target according to its own real-time state.

Description

Technical Field

[0001] The present application claims protection for a data processing technology, and more particularly, relates to a robust adaptive large-scale pneumatic conveying control method. Background Art

[0002] Pneumatic conveying is a technology that uses airflow to transport materials in a closed pipeline. It is widely used in the field of material transportation. The advantages of pneumatic conveying are long conveying distance, large conveying volume and high conveying speed. However, on the other hand, it also faces problems such as high energy consumption due to large gas volume, high conveying pressure due to high gas volume fluctuation, and system safety caused by frequent opening and closing of pipelines. Intelligent control is an effective means to solve such problems.

[0003] At present, the research on pneumatic conveying systems has made a lot of progress. Especially with the advancement of intelligent manufacturing, a series of new methods and means such as intelligent valves have emerged, making intelligent control of pneumatic conveying systems possible. However, the current pneumatic conveying systems have not yet solved the following key problems: first, how to efficiently detect the state of the conveying system and calculate and optimize the scheme of the pneumatic conveying system based on this; second, how to improve the robustness of the scheduling scheme, that is, to ensure the robustness of multiple solutions.

[0004] The core of the pneumatic conveying system is how to maintain the optimal pressure in the pipeline. Too high or too low pressure will either lead to excessive energy consumption, excessive wear of materials and pipelines, or cause pipeline blockage. However, how to maintain the optimal pressure involves two aspects: 1. How to judge the state in the pipeline, that is, whether the inside of the pipeline is "gas phase" or "solid phase"; 2. How to design an algorithm based on the state inside the pipeline and comprehensively consider multiple factors to efficiently adjust the pressure.

[0005] In the existing technology, different sensors, such as pressure, temperature and humidity, combined with deep learning algorithms can extract the characteristics of the internal state of the pipeline, thereby predicting the internal state of the pipeline in the current situation - whether it is in the "gas phase" or the "solid phase". On this basis, efficient pressure adjustment can be regarded as a multi-objective optimization problem, so as to achieve a comprehensive consideration of multiple factors and achieve multiple goals such as energy saving, safe production and completion of production tasks.

[0006] Multi-objective optimization algorithms based on swarm intelligence are the mainstream solution. However, the structure of the pneumatic conveying problem leads to a huge search space. There is a major problem in using swarm intelligence multi-objective optimization algorithms to solve the scheduling of pneumatic conveying systems. The randomness of swarm intelligence algorithms as heuristic algorithms in the optimization process leads to fluctuations in the quality of the solutions for each run, that is, the robustness of the algorithm is low.

[0007] It is not difficult to find that the problems of existing technologies include:

[0008] 1. It is difficult to accurately predict the state inside the pipeline.

[0009] 2. It is difficult to solve large-scale pipeline control problems.

[0010] 3. Problems such as large volatility of algorithm results.

[0011] Therefore, a solution method for pneumatic conveying system control with high responsiveness and robustness that can be put into use in actual industrial production is needed. Summary of the invention

[0012] In order to solve the above problems in the prior art, the present application provides a robust adaptive large-scale pneumatic conveying control method.

[0013] The present application also provides a robust adaptive large-scale pneumatic conveying control method, comprising:

[0014] Detecting the detection data of temperature, humidity, air pressure and air flow at each node in the pneumatic conveying system, and dividing the detection data into a plurality of input data sets according to a preset rule;

[0015] Inputting the input data set into a deep learning network, iterating the input data using a hybridization algorithm, and obtaining state parameters of the pneumatic conveying system;

[0016] Sampling the state parameters, decision variables and optimization objectives, and obtaining complete population data through an evolutionary algorithm;

[0017] Based on the complete population data, the control parameters are calculated by using a Monte Carlo tree search algorithm.

[0018] Optionally, the steps of the hybridization algorithm include:

[0019] Step 1: Initialize the population;

[0020] Step 2: Perform a fast non-dominated sort on the population to obtain the number of layers F for each individual i ;

[0021] Step 3: Select individuals from the first layer until they are placed in the Fth layer. i+1 When the number of individuals in the layer is just greater than N, only F1 to F i The number of individuals in the layer is n;

[0022] Step 4: For F i+1 The individuals in the layer calculate the crowding distance and select the first Nn F i+1 The individuals of the first layer are put into the population, and the population size is N at this time;

[0023] Step 5: Select, crossover, and mutate to generate offspring population;

[0024] Step 6: Check whether the optimization stop condition is met. If not, go to step 2.

[0025] Optionally, it also includes sampling processing of the data, including:

[0026] Given a node N, there is a population P of n solutions, each of which has d decision variables. For each child node of N, the algorithm extracts d solutions from the d decision variables. n The decision variables that have not been sampled are optimized, and the other decision variables that have not been sampled remain unchanged to generate this child node.

[0027] Optionally, the calculating and obtaining the control parameter by using a Monte Carlo tree search algorithm includes:

[0028] Build a search tree;

[0029] Based on the search tree, the Tree optimization algorithm Policy stage and the Default optimization algorithm Policy stage are used to iterate data;

[0030] A Monte Carlo tree search algorithm is used to improve the robustness of the pneumatic conveying system control.

[0031] Optionally, the robustness algorithm includes:

[0032] Step 1: Initialize the population as the root node, and set the evaluation Δ of the root node to 0.

[0033] Step 2: If the parent node has k child nodes, select the node with the highest evaluation until the number of child nodes of node N is less than k.

[0034] Step 3: Use decision variable sampling to optimize and construct a new node optimization algorithm Nl.

[0035] Step 4: Estimate the hypervolume value of the child node Nl as the evaluation value Δ, then backpropagate Δ and update the UCB value.

[0036] Step 5: If the super capacity of the child node optimization algorithm Nl optimization algorithm is better than the archive node, the archive node is updated with the new node.

[0037] Step 6: Repeat step 2 to optimize algorithm until computing resources are exhausted.

[0038] Optionally, in the pneumatic conveying system, the control time of each pipeline includes: start and stop time and and Satisfy constraints and and

[0039] Optionally, the control time is expressed as follows:

[0040]

[0041]

[0042] in represents the scheduling plan of the i-th pipeline at time j, 1 means running, 0 means stopping; represents the peak value within the production cycle T, which needs to be minimized. Indicates the difference between the peak and valley pressures of the pneumatic conveying system.

[0043] The present application provides a robust adaptive large-scale pneumatic conveying control device, comprising:

[0044] monitor;

[0045] Memory;

[0046] The processor retrieves the program of the robust adaptive large-scale pneumatic conveying control method from the memory and executes the following steps: receiving detection data including temperature, humidity, air pressure and air flow in the pneumatic conveying system; transmitting the detection data to the deep learning network, and outputting state parameters after calculation by the deep learning network; sampling the state parameters, as well as decision variables and optimization targets, and obtaining complete population data through an evolutionary algorithm; and calculating and obtaining control parameters based on the complete population data through a Monte Carlo tree search algorithm.

[0047] Optionally, it also includes: an algorithm library for storing the Monte Carlo tree search algorithm.

[0048] The present application also provides a robust adaptive large-scale pneumatic conveying control device, comprising:

[0049] A data module, used to receive detection data including temperature, humidity, air pressure and air flow in the pneumatic conveying system;

[0050] An analysis module, used for transmitting the detection data to a deep learning network and outputting state parameters through calculation by the deep learning network;

[0051] An optimization module, used for sampling the state parameters, decision variables and optimization targets, and obtaining complete population data through an evolutionary algorithm;

[0052] The result module is used to calculate and obtain the control parameters through the Monte Carlo tree search algorithm based on the complete population data.

[0053] The advantages of this application compared to the prior art are:

[0054] The present application also provides a robust adaptive large-scale pneumatic conveying control method, including: detecting the detection data of temperature, humidity, air pressure and air flow at each node in the pneumatic conveying system, and dividing the detection data into multiple input data sets according to preset rules; inputting the input data set into a deep learning network, iterating the input data using a hybrid algorithm, and obtaining the state parameters of the pneumatic conveying system; sampling the state parameters, decision variables and optimization targets, and obtaining complete population data through an evolutionary algorithm; based on the complete population data, calculating and obtaining control parameters through a Monte Carlo tree search algorithm. The present application deploys multiple sensors in the pneumatic conveying system, and uses a neural network to analyze the sensor parameters to obtain the state of the pipeline itself. Based on the state of the pipeline, the population in the population optimization algorithm is used as a node to construct a Monte Carlo tree, ensuring that a high-quality solution is searched in a large-scale decision space, thereby ensuring that the pneumatic conveying system can optimize the target according to its own real-time state while improving the robustness of the solution. BRIEF DESCRIPTION OF THE DRAWINGS

[0055] Figure 1 It is a flow chart of adaptive large-scale pneumatic conveying control in this application.

[0056] Figure 2 It is a schematic diagram of the deep neural network for state analysis of the pneumatic conveying system in this application.

[0057] Figure 3 It is a schematic diagram of the decision variable sampling optimized in this application.

[0058] Figure 4 It is a flowchart of the UCT iteration in this application.

[0059] Figure 5 It is a data processing flow chart in this application.

[0060] Figure 6 It is a schematic diagram of convergence analysis in this application. DETAILED DESCRIPTION

[0061] The following contents are all examples of specific implementation processes provided for describing in detail the technical solutions to be protected by this application. However, this application may also be implemented in other ways different from the descriptions herein. Persons skilled in the art may, under the guidance of the concepts of this application, adopt different technical means to implement this application. Therefore, this application is not limited to the specific embodiments below.

[0062] Embodiment 1:

[0063] The present application also provides a robust adaptive large-scale pneumatic conveying control method, including: detecting the detection data of temperature, humidity, air pressure and air flow at each node in the pneumatic conveying system, and dividing the detection data into multiple input data sets according to preset rules; inputting the input data set into a deep learning network, iterating the input data using a hybrid algorithm, and obtaining the state parameters of the pneumatic conveying system; sampling the state parameters, decision variables and optimization targets, and obtaining complete population data through an evolutionary algorithm; based on the complete population data, calculating and obtaining control parameters through a Monte Carlo tree search algorithm. The present application deploys multiple sensors in the pneumatic conveying system, and uses a neural network to analyze the sensor parameters to obtain the state of the pipeline itself. Based on the state of the pipeline, the population in the population optimization algorithm is used as a node to construct a Monte Carlo tree, ensuring that a high-quality solution is searched in a large-scale decision space, thereby ensuring that the pneumatic conveying system can optimize the target according to its own real-time state while improving the robustness of the solution.

[0064] Figure 1 It is a flow chart of adaptive large-scale pneumatic conveying control in this application.

[0065] Please refer to Figure 1 As shown, S101 detects the detection data of temperature, humidity, air pressure and air flow at each node in the pneumatic conveying system, and divides the detection data into multiple input data sets according to preset rules.

[0066] Various and multiple sensors are deployed in the pneumatic conveying system to detect the parameters of the pneumatic conveying system, and the sensors at least include a temperature sensor, an air pressure sensor, a humidity sensor, and an air flow sensor.

[0067] The preset rule may be based on a preset time limit, and the data detected by each sensor may be summarized into the same data set according to the preset time limit, and the data set may be input as a data. The data input expression is as follows:

[0068] S={S1,S2,……,S l}

[0069] Wherein, S represents the data input, l represents the number of sensors, S1, S2, ..., S l Represents the data of each sensor.

[0070] Please refer to Figure 1 As shown, S102 inputs the input data set into a deep learning network, and uses a hybrid algorithm to iterate the input data to obtain state parameters of the pneumatic conveying system.

[0071] A deep neural network is constructed, and the input data is input into the deep neural learning network for calculation to obtain the operating status of the pneumatic conveying system.

[0072] The deep neural network model is designed based on the number of sensors, the scale of the pneumatic conveying system, etc. Figure 2 shown.

[0073] The training samples of the deep neural network can be obtained by testing the pneumatic conveying system and collecting sensor data. The labels of the samples are the operating status of the pipeline, and the labels include: gas phase or solid phase.

[0074] The operating state can be expressed as: ensuring that the pipeline is at the optimum temperature and optimum degree, while minimizing the energy consumption of the pneumatic conveying system, minimizing the operating pressure peak of the pneumatic conveying system (minimizing high pressure), minimizing the pressure peak-valley difference of the pneumatic conveying system (minimizing high and low pressure difference), maximizing the flow rate of the pneumatic conveying system, etc. The above optimization problem satisfies the following expression:

[0075] minimizeF(x)= <f1(x),f2(x),…,f M (x)>subjecttox∈X

[0076] Among them, x is the adjustable parameter of pipeline operation, that is, the decision variable of the optimization problem, X is the search space of decision variables, and f m (x) is the mth optimization objective that needs to be optimized in the pipeline. Among them, large-scale multi-objective optimization problems mainly refer to the case where the dimension of the decision variables is greater than 100, that is, the dimension of x is greater than 100.

[0077] In the pneumatic conveying system control, there are m pipelines with a production cycle T (unit: s). For the i-th pipeline, there is a running time requirement t i .

[0078] The algorithm finally assigns start and stop times to each pipeline and and Satisfy constraints and and

[0079] Based on the above constraints, the chromosome of the problem can be encoded as: And satisfy the constraints The corresponding It can be expressed by the formula calculate.

[0080] For example, if the optimization is to minimize the pressure peak of the pneumatic conveying system and minimize the difference between the pressure peak and the pressure valley, the expressions of these two objectives are:

[0081]

[0082]

[0083] in represents the scheduling plan of the i-th pipeline at time j, 1 means running, 0 means stopping; represents the peak value within the production cycle T, which needs to be minimized. Indicates the difference between the peak and valley pressures in the pneumatic conveying system.

[0084] Based on the above chromosome encoding, the process of optimizing the algorithm is as follows:

[0085] Step 1: Initialize the population;

[0086] Step 2: Perform a fast non-dominated sort on the population to obtain the number of layers F for each individual i ;

[0087] Step 3: Select individuals from the first layer until they are placed in the Fth layer. i+1 When the number of individuals in the layer is just greater than N, only F1 to F i The number of individuals in the layer is n;

[0088] Step 4: For F i+1 The individuals in the layer calculate the crowding distance and select the first Nn F i+1 The individuals of the first layer are put into the population, and the population size is N at this time;

[0089] Step 5: Select, crossover, and mutate to generate offspring population;

[0090] Step 6: Check whether the optimization stop condition is met. If not, go to step 2.

[0091] Please refer to Figure 1 As shown, S103 samples the state parameters, decision variables and optimization objectives, and obtains complete population data through an evolutionary algorithm.

[0092] The above conventional hybrid forwarding is suitable for the case where the data volume is small. If the data volume is large, the data needs to be sampled.

[0093] Figure 3 It is a schematic diagram of the decision variable sampling optimized in this application.

[0094] Please refer to Figure 3As shown in the figure, each rounded box is a solution, and the circles are decision variables. The optimization algorithm first randomly samples a subset of decision variables, then reduces the dimension of each solution, and then uses an evolutionary algorithm to optimize these decision variables. Finally, all solutions are restored to a complete population based on the sampled decision variables to form child nodes.

[0095] Specifically, given a node N, there is a population P of n solutions, each solution has d decision variables, and for each child node of N, the algorithm extracts d from the d decision variables. n The decision variables that have not been sampled are optimized, and the other decision variables that have not been sampled remain unchanged to generate this child node.

[0096] Based on the sampled decision variables, each child node will optimize a different decision variable based on the parent population.

[0097] After obtaining the sampled d n After the decision variables are determined, the traditional multi-objective optimization algorithm is used to perform evolutionary search on these decision variables. In this case, each solution of the parent population is first reduced to d according to the sampled decision variables. n dimension.

[0098] The algorithm performs an evolution strategy on the reduced population according to the optimization algorithm used and then recovers to a d-dimensional solution.

[0099] Please refer to Figure 1 As shown, S104 calculates and obtains control parameters through a Monte Carlo tree search algorithm based on the complete population data.

[0100] include:

[0101] assessing offspring and performing environmental selection;

[0102] After several iterations, the final population is used to construct the child node N l .

[0103] After the above processing, the data is constructed into a multi-level data set, and further data processing will be carried out next.

[0104] Figure 5 It is a data processing flow chart in this application.

[0105] Please refer to Figure 5 As shown, S501 constructs a search tree. This application uses the UCT method. UCT (UpperConfidenceBoundApplytoTree) is an extended implementation of the UpperConfidenceBound (UCB) of the search tree. The overall idea of ​​UCT is to treat each node in the search tree as a proxy. When selecting a node, we select it based on the UCB value, such as Figure 3As shown, select the node with the highest UCB value that is not fully expanded; expand the node; evaluate the node using the default strategy; propagate the evaluation to all parent nodes to update the UCB value.

[0106] S502 iteration, the iteration process is divided into two stages: the Tree optimization algorithm Policy stage and the Default optimization algorithm Policy stage. In the Tree optimization algorithm Policy stage, starting from the root node, the child node with the largest UCB value is selected each time, unless the node is fully expanded; then a new node is expanded from the node. In the Default optimization algorithm Policy optimization algorithm stage, the expanded new node is simulated (also called optimization algorithm roll-out); then the simulated value is continuously propagated backward to update the value of each node on the path. Repeated iteration until the computing resource budget is reached, and finally the child node with the highest value under the root node is returned.

[0107] In step S502, TreePolicy is part of the large-scale multi-objective optimization described in the first part. The purpose of DefaultPolicy is to evaluate the suitability of nodes for further search. The quality of the population can be measured by the hypervolume (HV) value of the population. The HV is an indicator to guide the search process. This indicator gives the hypervolume of the dominant part of the target space and currently represents the only strictly monotonic type of set quality measure. Whenever one Pareto set approximation dominates another Pareto set approximation, the indicator value of the former will be greater than the indicator value of the latter. This property makes it very suitable as an evaluation strategy in the Monte Carlo tree search framework.

[0108] S503 Robustness algorithm: In this application, a Monte Carlo tree search algorithm is used to improve the robustness of the pneumatic conveying system control. The steps of this algorithm in this application are as follows:

[0109] Step 1: Initialize the population as the root node, and set the evaluation Δ of the root node to 0.

[0110] Step 2: If the parent node has k child nodes, select the node with the highest evaluation until the number of child nodes of node N is less than k.

[0111] Step 3: Use decision variable sampling to optimize and construct a new node optimization algorithm Nl.

[0112] Step 4: Estimate the hypervolume value of the child node Nl as the evaluation value Δ, then backpropagate Δ and update the UCB value.

[0113] Step 5: If the super capacity of the child node optimization algorithm Nl optimization algorithm is better than the archive node, the archive node is updated with the new node.

[0114] Step 6: Repeat step 2 to optimize algorithm until computing resources are exhausted.

[0115] Finally, the control parameters are derived according to the updated UCB values.

[0116] Embodiment 2:

[0117] Sensors are deployed at various nodes of the pneumatic conveying system, and one or more of the following sensors are deployed as needed: temperature sensor, humidity sensor, and air pressure sensor.

[0118] Run a trial run and collect samples and data to train the neural network.

[0119] The description of the pneumatic conveying system is given as follows: for a production cycle T (unit: s), m pipelines, for the i-th pipeline, there is a running time requirement t i , pipeline gas consumption q i Through the algorithm, each pipeline is assigned a start and stop time. and and Satisfy constraints and and

[0120] Based on this, the encoding of the problem can be defined as:

[0121] Optimization goal:

[0122] 1. Minimize peaks: in It represents the scheduling plan of the i-th pipeline at time j, 1 means running, and 0 means stopping.

[0123] 2. Minimize volatility:

[0124] This application sets up two processes corresponding to small-scale data and large-scale data respectively. Algorithm process 1 corresponds to small-scale data; algorithm process 2 corresponds to large-scale data.

[0125] Algorithm flow 1:

[0126] Step 1: Initialize the population as the root node, and set the evaluation Δ of the root node to 0;

[0127] Step 2: If the parent node has k child nodes, select the node with the highest evaluation until the number of child nodes of node N is less than k;

[0128] Step 3: Use decision variable sampling to optimize and construct a new node Nl;

[0129] Step 4: Estimate the hypervolume value of child node Nl as the evaluation Δ, then back-propagate Δ and update the UCB value;

[0130] Step 5: If the excess capacity of child node Nl is better than that of the archive node, update the archive node with the new node;

[0131] Step 6: Repeat step 2 until computing resources are exhausted.

[0132] Algorithm flow 2:

[0133] Step 1: Randomly sample some decision variables;

[0134] Step 2: Select, crossover, and mutate the sampled decision variables to generate the offspring population, while the other decision variables remain unchanged;

[0135] Step 3: Perform a quick non-dominated sort on the population to obtain the number of layers F for each individual i ;

[0136] Step 3: Select individuals from the first layer until they are placed in the Fth layer. i+1 When the number of individuals in the layer is just greater than N, only F1 to F i The number of individuals in the layer is n;

[0137] Step 4: For F i+1 The individuals in the layer calculate the crowding distance and select the first Nn F i+1 The individuals of the first layer are put into the population, and the population size is N at this time;

[0138] Step 5: Generate child nodes

[0139] Stop condition:

[0140] The number of generations of evolutionary computation is set as the stopping condition. At the same time, statistical analysis methods can be used to determine whether convergence has been achieved.

[0141] The present application also provides a robust adaptive large-scale pneumatic conveying control device, comprising:

[0142] monitor;

[0143] Memory;

[0144] The processor retrieves the program of the robust adaptive large-scale pneumatic conveying control method from the memory and executes the following steps: receiving detection data including temperature, humidity, air pressure and air flow in the pneumatic conveying system; transmitting the detection data to the deep learning network, and outputting state parameters after calculation by the deep learning network; sampling the state parameters, as well as decision variables and optimization targets, and obtaining complete population data through an evolutionary algorithm; and calculating and obtaining control parameters based on the complete population data through a Monte Carlo tree search algorithm.

[0145] The present application also includes: an algorithm library for storing the Monte Carlo tree search algorithm.

[0146] The present application also provides a robust adaptive large-scale pneumatic conveying control device, comprising:

[0147] A data module, used to receive detection data including temperature, humidity, air pressure and air flow in the pneumatic conveying system;

[0148] An analysis module, used for transmitting the detection data to a deep learning network and outputting state parameters through calculation by the deep learning network;

[0149] An optimization module, used for sampling the state parameters, decision variables and optimization targets, and obtaining complete population data through an evolutionary algorithm;

[0150] The result module is used to calculate and obtain the control parameters through the Monte Carlo tree search algorithm based on the complete population data.

[0151] Embodiment 3:

[0152] Please refer to Figure 5 As shown, based on the calculation of Example 1 or 2 20 times, the result is compared with the calculation of the existing mainstream technology, and the calculation results are as follows:

[0153]

[0154] Table 1 IGD obtained by various compared algorithms on 54 test instances of LSMOP test

[0155] The statistical results of the average IGD values ​​of 20 runs are shown in Table 1. Overall, the solution method outperforms seven state-of-the-art algorithms.

[0156] Table 1 lists the IGD values ​​of the eight algorithms. LMOMCTS obtains 20 best results out of 54 test cases, outperforming the other seven compared methods. MOEA / DVA and MOEA / PSL achieve 10 and 6 dominant cases, respectively, and the other algorithms obtain less than 5 best results.

[0157] In the test, the proposed LMOMCTS also outperformed the other seven comparison algorithms in terms of convergence performance. LMOMCTS obtained 18 out of 54 best results. The results of other algorithms on the HV index were only 4 or less, which was significantly inferior to the results of this application.

[0158] The above experimental results show that for most test problems, this application is a method with good convergence and diversity. It can be seen that the proposed algorithm can achieve competitive performance on the LSMOP test problem, which also confirms the effectiveness of the algorithm in solving large-scale decision variable problems.

[0159]

[0160] Table 2 Variance of IGD after running various competing algorithms 20 times

[0161] This application shows the variance of IGD after running all competing algorithms 20 times in Table 2. In general, the variance of this application on IGD is very good, which verifies that the designed method improves the robustness of the algorithm optimization results.

[0162] The present application obtained a total of 17 best results, which means that it was run 20 times on 18 instances with the smallest variance, and has better robustness than other algorithms. Combined with Table 1, the present application not only has an advantage in performance, but also the robustness of the algorithm is much better. MOEA / DVA has 9 better instances than LMOMCTS in Table 1, only 4 of which are significantly better than the algorithm of the present application in Table 2, and the variance is worse when solving LSMOP3, LSMOP6 and LSMOP7. The robustness of other algorithms in their results is also poor, such as CCGDE3 for solving LSMOP9, LMEA for solving LSMOP3, LMOCSO for solving LSMOP9, and GMOEA for solving LSMOP2, LSMOP3 and LSMOP9. In these mentioned instances, the variance is greater than 1, or even greater than 10, which is unacceptable in real-world practice. The variance of the algorithm of the present application on all test instances is kept at a very small order of magnitude, ensuring the robustness of the algorithm in practical applications.

[0163] Another important aspect of robustness is to maintain good results when the algorithm stops in unexpected situations. In this regard, the changes in the IGD values ​​achieved by all compared algorithms on the LSMOP1-LSMOP9 problems with three objectives are shown in Figure 5 As shown in the figures. It can be seen from these figures that the proposed algorithm does not always dominate, but it shows higher stability and robustness. For example, LMEA shows obvious fluctuations when solving LSMOP1, LSMOP2, and LSMOP3, MOEA / DVA shows serious performance fluctuations when solving LSMOP6, and CCGDE3 also shows fluctuations in LSMOP8 and MOEA / PSL when solving LSMOP9.

[0164] In addition to robustness, the running time of the algorithm is also critical in practical applications. The average running time of all compared algorithms on LSMOP1-LSMOP9 in this application, the decision variables range from 100 to 5,000. The LMOMCTS proposed in this application has a relatively short computation time. Among them, the time complexity of MOEA / PSL and MOEA / DVA, which perform well in the IGD metric comparison, is very high, and increases sharply due to the increase in the number of decision variables. However, the algorithm of this application maintains a relatively low time complexity in different dimensions.

[0165] The above contents are all examples of specific implementation processes provided for describing in detail the technical solutions to be protected by this application. However, this application may also be implemented in other ways different from the descriptions herein. Persons skilled in the art may, under the guidance of the concepts of this application, adopt different technical means to implement this application. Therefore, this application is not limited to the specific embodiments below.

Claims

1. A robust adaptive large-scale pneumatic conveying control method, characterized in that: Detecting the detection data of temperature, humidity, air pressure and air flow at each node in the pneumatic conveying system, and dividing the detection data into a plurality of input data sets according to a preset rule; Inputting the input data set into a deep learning network, iterating the input data using a hybridization algorithm, and obtaining state parameters of the pneumatic conveying system; Sampling the state parameters, decision variables and optimization objectives, and obtaining complete population data through an evolutionary algorithm; Based on the complete population data, a control parameter is calculated by a Monte Carlo tree search algorithm; The steps of the hybridization algorithm include: Step 1: Initialize the population; Step 2: Perform a quick non-dominated sort on the population to obtain the number of layers for each individual ; Step 3: Select individuals from the first layer until they are placed in the When the number of individuals is just greater than , then only put arrive The number of individuals in the layer is ; Step 4: The individuals in the layer calculate the crowding distance and select the first indivual The individuals of the first layer are put into the population, and the population size is ; Step 5: Select, crossover, and mutate to generate offspring population; Step 6: Check whether the optimization stop condition is met, if not, go to step 2; The steps of the hybridization algorithm also include sampling the data, including: Given a node ,have The population of solutions , each solution has decision variables, for For each child node of Decision variables are extracted The decision variables are optimized, and the other unsampled decision variables remain unchanged to generate this child node; Robust algorithms include: Step 1: Initialize the population as the root node, and set the evaluation Δ of the root node to 0; Step 2: If the parent node has k child nodes, select the node with the highest evaluation until the number of child nodes of node N is less than k; Step 3: Use decision variable sampling to optimize and construct a new node optimization algorithm Nl; Step 4: Estimate the hypervolume value of the child node optimization algorithm Nl optimization algorithm as the evaluation optimization algorithm Δ, then back-propagate the optimization algorithm Δ optimization algorithm and update the optimization algorithm UCB optimization algorithm value; Step 5: If the super capacity of the child node optimization algorithm Nl optimization algorithm is better than the archive node, update the archive node with the new node; Step 6: Repeat step 2 to optimize algorithm until computing resources are exhausted.

2. The robust adaptive large-scale pneumatic conveying control method according to claim 1, characterized in that: The control parameters obtained by calculating the Monte Carlo tree search algorithm include: Build a search tree; Based on the search tree, the Tree optimization algorithm Policy stage and the Default optimization algorithm Policy stage are used to iterate data; A Monte Carlo tree search algorithm is used to improve the robustness of the pneumatic conveying system control.

3. The robust adaptive large-scale pneumatic conveying control method according to claim 1, characterized in that: In the pneumatic conveying system, the control time of each pipeline includes: start and stop time and , and Satisfy constraints ,and and ; Where T is the production cycle, m is the number of pipelines, t i is the running time of the ith pipeline.

4. The robust adaptive large-scale pneumatic conveying control method according to claim 3, characterized in that: The control time is related to the expression of minimizing the system pressure peak and minimizing the pressure peak valley: Among them, q i is the gas consumption of the ith pipeline, Indicates Root pipe in The scheduling plan at the time, 1 means running, 0 means stopping; Indicates production cycle The peak value within needs to be minimized. Indicates the difference between the peak and valley pressures of the pneumatic conveying system.

5. A robust adaptive large-scale pneumatic conveying control device, characterized in that: include: monitor; Memory; The processor retrieves the program of the robust adaptive large-scale pneumatic conveying control method from the memory for execution, thereby completing the method as described in any one of claims 1 to 4, and performing the following steps: receiving detection data including temperature, humidity, air pressure and air flow in the pneumatic conveying system; transmitting the detection data to the deep learning network, and outputting state parameters after calculation by the deep learning network; sampling the state parameters, decision variables and optimization targets, and obtaining complete population data through an evolutionary algorithm; Based on the complete population data, the control parameters are calculated by using a Monte Carlo tree search algorithm.

6. The robust adaptive large-scale pneumatic conveying control device according to claim 5, characterized in that: Also includes: The algorithm library is used to store the Monte Carlo tree search algorithm.

7. A robust adaptive large-scale pneumatic conveying control device, characterized in that: For executing the method according to any one of claims 1 to 4, the device comprises: A data module, used to receive detection data including temperature, humidity, air pressure and air flow in the pneumatic conveying system; An analysis module, used for transmitting the detection data to a deep learning network and outputting state parameters through calculation by the deep learning network; An optimization module, used for sampling the state parameters, decision variables and optimization targets, and obtaining complete population data through an evolutionary algorithm; The result module is used to calculate and obtain the control parameters through the Monte Carlo tree search algorithm based on the complete population data.

Images