Optimization design method of solid-liquid hydrocyclone based on GA-BP neural network
By adding triangular separation blocks to the inlet channel of the hydrocyclone and combining it with the GA-BP neural network optimization design method, the problem of separation difficulties of solid-liquid hydrocyclones when separating materials with small density and particle size differences was solved, and efficient and low-cost optimization design and separation effects were achieved.
Patent Information
- Application Number
- CN202510804312.6
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2025-06-17
- Publication Date
- 2025-09-12
AI Technical Summary
Existing solid-liquid hydrocyclones have difficulty in separating coking coal materials with small density and particle size differences. Their design relies on experience, resulting in low performance, low separation efficiency, and high experimental costs.
An optimization design method based on GA-BP neural network is adopted. By adding triangular separation blocks in the feed flow channel, combining mixing level design and additional test points, computational fluid dynamics (CFD) simulation and genetic algorithm are used to optimize the structural parameters of the hydrocyclone, establish a nonlinear mapping relationship, and achieve multi-objective collaborative optimization.
It significantly improves separation efficiency and accuracy, reduces experimental costs, simplifies the design process, and improves the robustness and generalization ability of the model, making it suitable for efficient operation under a wide range of working conditions.
Smart Images
Figure CN120633087A_ABST
Abstract
Description
Technical Field
[0001] The invention belongs to the technical field of hydrocyclone separation, and in particular relates to an optimization design method for a solid-liquid hydrocyclone based on a GA-BP neural network. Background Art
[0002] A hydrocyclone is a device that uses the difference in particle size and density between immiscible media to achieve effective separation and classification. Hydrocyclones are widely used in mining, chemical industry, coal and petroleum, environment, agriculture and other fields. However, it is still difficult to separate coking coal with small density and particle size differences. Solid-liquid hydrocyclones with traditional structures cannot achieve effective separation. At the same time, due to the complexity of the internal flow field of the solid-liquid hydrocyclone and the many influencing factors, and the different structures of the solid-liquid hydrocyclones used by researchers, they are extremely dependent on the experience of the designers and can only rely on a large number of experimental designs to find the best. At present, the design of solid-liquid hydrocyclones mainly relies on the empirical formulas accumulated by predecessors or references to existing molded products. Therefore, the existing design of solid-liquid hydrocyclones has problems such as low performance, large particle size of separated solid phase particles, and high pressure drop, and cannot well meet the needs of the separation equipment industry. Summary of the Invention
[0003] The present invention aims to overcome the shortcomings and deficiencies of existing technologies by providing a method for optimizing the geometric parameters of a solid-liquid hydrocyclone, with the goal of achieving narrow density separation. This design method can improve the performance of the solid-liquid hydrocyclone while simplifying the optimization design process for conventional solid-liquid hydrocyclones.
[0004] Technical solution: In order to achieve the above-mentioned purpose, the present invention proposes an optimization design method for a solid-liquid hydrocyclone based on a GA-BP neural network, which comprises the following steps:
[0005] Step 1: adding a triangular separation block to the inlet channel of the hydrocyclone to obtain a pre-separation hydrocyclone;
[0006] Step 2: Design multiple groups of experiments for the pre-separation hydrocyclone using a mixed level design combined with additional test points, and select multiple key structural parameters as input variables of the BP algorithm, including column length L, cone angle θ, overflow pipe insertion depth L. insert , Overflow pipe diameter D over , bottom flow opening diameter D under and bottom flow outlet length L under ;
[0007] Step 3: After the hydrocyclone is running stably, the difference in separation efficiency between gangue and coal particles at the bottom flow outlet is selected as the output variable μ oi , then the L, θ, Linsert 、D over 、D under 、L under 、μ oi Constitute a set of training samples;
[0008] Step 4: Use computational fluid dynamics (CFD) to perform numerical simulation calculations on the structural parameters obtained in step (2) to form a database corresponding to the difference between different hydrocyclone structural parameters and the gangue separation efficiency, and train the GA-BP neural network prediction model to output the hydrocyclone structural parameters corresponding to the maximum difference in gangue separation efficiency, that is, to obtain the optimal structural parameters for gangue separation.
[0009] Furthermore, step 4 includes the following steps:
[0010] Step 4.1: Define the mathematical transformation rules between different network layers in the BP neural network, convert the weighted input of the neuron into nonlinear or linear output, use the hidden layer transfer function to normalize the input data, enhance the nonlinear expression ability, and define tanh as the hidden layer transfer function:
[0011]
[0012] Among them, φ(n) is the neuron output value after the hidden layer transfer function, and n is the weighted sum of the inputs of the hidden layer neurons, that is, Among them, i is the group order of each sample, P is the total number of features, that is, the total number of hydrocyclone structural parameters, w i is the weight corresponding to the i-th input feature, x i Input features include column length L, cone angle θ, overflow pipe insertion depth L insert , Overflow pipe diameter D over , bottom flow opening diameter D under and bottom flow outlet length L under , b is the neuron bias term;
[0013] In step 4.2, use the transfer function of the output layer to restrict the output to a probabilistic form, map the parameter values of the input structure to the interval (0, 1), and define the Sigmoid function as the transfer function of the output layer:
[0014]
[0015] Among them, φ(k) is the final predicted value of the output layer, k is the weighted sum of the input of the output layer neurons, and the calculation method is the same as that of the hidden layer neurons, that is, Among them, H is the number of hidden neurons, w j is the weight from the jth neuron in the hidden layer to the output layer, h jis the output value of the jth neuron in the hidden layer, and c is the bias term of the output layer;
[0016] In step 4.3, a genetic algorithm (GA) is introduced to perform global search and optimize the initial weights of the BP neural network to obtain the optimal structural parameters for gangue separation.
[0017] Furthermore, the process of step 4.3 is as follows:
[0018] Step 4.3.1, perform GA encoding of the initial population;
[0019] Concatenate all weights w and bias terms b into real vectors in layer order:
[0020] In this chromosome, the input layer X=6, the hidden layer H=3, and the output layer Q=1. A complete chromosome represents a complete set of hydrocyclone structural parameter prediction models, where: is the connection weight from the mth neuron in the input layer to the nth neuron in the hidden layer, mapping the hydrocyclone structural parameters to the hidden layer features. m ranges from 1 to X, and n ranges from 1 to H. is the activation threshold of the i-th neuron in the hidden layer, representing the bias of the hidden layer, is the connection weight from the i-th neuron in the hidden layer to the output layer, which represents the difference in separation efficiency between the intermediate hidden layer features μ oi The contribution weight of i ranges from 1 to H; is the output layer neuron adjustment term, representing the bias of the efficiency prediction layer;
[0021] Initialize and randomly generate N pop chromosomes, that is, the population size, N pop The value is 50-200, and each gene value is sampled from the uniform distribution U(-1,1) to ensure that the initial neural network initialization iteration can cover the full design space of the hydrocyclone parameters;
[0022] Step 4.3.2, fitness calculation
[0023] The fitness function is calculated by decoding each vector in the chromosome into the weights and biases of the neural network to quantify the separation performance of the current parameter combination:
[0024]
[0025] Among them, J is the total loss value between the BP neural network model prediction and the database result;
[0026] Step 4.3.3 Select an action
[0027] Each time, t individuals are randomly selected from the population, and the individual with the highest fitness is selected to enter the next generation, and this process is repeated until the size of the new population is the same as the original population.
[0028] Step 4.3.4 Crossover Operation
[0029] Randomly select two crossover points in the chromosome and exchange the middle segments of the parent chromosome. The chromosome segments represent the weights and biases of the neural network. The structural parameter combination of the cyclone is mapped according to the network weights and biases. By exchanging the middle segments, the advantageous features of different structural parameters can be recombined.
[0030] Step 4.3.5 Mutation Operation
[0031] Use Gaussian mutation with mutation probability p m = 0.001~0.1 Randomly perturb the genes in the chromosome:
[0032] w new =w old +N(0,σ)
[0033] Among them, w old is the current weight value; N(0,σ) is a Gaussian distribution random perturbation, which means fine-tuning the weight, σ is 0.1, and p is used in the neural network prediction model. m = 0.001~0.1 mutation probability enters the new structure parameter iteration; w new is the new weight value after perturbation, representing the local optimization of the parameter mapping relationship;
[0034] Step 4.3.6: Retain the best individuals. The 10% individuals with the highest fitness in each generation are retained and directly enter the next generation to prevent the loss of excellent genes. That is, the optimal structural parameter combination is retained. If the fitness improvement is less than 1% for 10 consecutive iterations, the system is terminated early, indicating that a stable optimal structure has been found.
[0035] Step 4.3.7 Weight import and BP fine-tuning
[0036] Optimal chromosome decoding: Select the individual with the highest fitness from the final population and decode it into a weight matrix and bias vector, that is, extract the optimal weight value found by the GA algorithm to obtain the initial neural network prediction model;
[0037] BP network initialization: import the decoded weights into the neural network, replace the random initial values of the model itself; call the built-in Adam optimizer in Python, use the adaptive learning rate a=0.001~0.01 optimization algorithm, set the number of iterations E epochs =200-500, balancing training cost and accuracy; set early stopping: if the validation set loss does not decrease for 20 consecutive rounds, terminate training.
[0038] Furthermore, the number of neurons in the hidden layer is determined by the following formula:
[0039]
[0040] Among them, β is a constant, Y is the number of input neurons, and X is the number of output neurons, which corresponds to the final optimization goal, that is, the difference in coal gangue separation efficiency.
[0041] Furthermore, β is set to a value of 5 to 10.
[0042] Furthermore, the square of the error is defined as the loss function to calculate the error between the true value of the database and the prediction of the neural network model:
[0043]
[0044] Among them, J is the total loss value between the BP neural network model prediction and the database result, l is the total number of samples, y i is the output after normalization of the difference in coal gangue separation efficiency in the database, o i is the prediction output of the BP neural network.
[0045] Beneficial effects: Compared with the prior art, the technical solution of the present invention has the following beneficial technical effects:
[0046] (1) Significantly improve separation efficiency and accuracy. Pre-separation block design: By adding triangular separation blocks to the inlet pipe, the density and size differences of mineral particles are used to achieve pre-separation, effectively reducing the ash content in the clean coal and improving the quality of the clean coal. Experimental data shows that the ash content of the clean coal of the optimized structure is reduced from 31.75% to 9.82% compared to the unoptimized hydrocyclone structure, which is an effective reduction of 32.96% compared to the ash content of the raw coal of 42.78%.
[0047] (2) Significantly reduce experimental and design costs, using mixed-level design and additional test points: Through scientific experimental design methods, the large number of physical experiments required by traditional trial-and-error methods can be reduced, saving 30% to 50% of R&D time and costs. Data-driven modeling: Using numerical simulation (ANSYS Fluent) and machine learning to replace empirical formulas reduces reliance on designers' own experience, reduces the number of experimental groups, and shortens the design cycle.
[0048] (3) Enhance model robustness and generalization capabilities. Global search: Genetic algorithm (GA) prevents BP neural network from falling into local optimality, ensuring that weight initialization is closer to the global optimal solution. Local fine-tuning: BP algorithm further optimizes parameters through gradient descent, improving the model's generalization ability under complex flow conditions. Regularization and early stopping: Introducing L2 regularization term in the loss function, combined with early stopping method, effectively prevents overfitting, and improves the model's prediction stability under unknown working conditions by 20%.
[0049] (4) Achieve multi-objective collaborative optimization and multi-parameter dynamic mapping: A nonlinear relationship between structural parameters (six input variables) and separation efficiency is established through a neural network to achieve multi-objective collaborative optimization. Extrapolation prediction capability: By generating and verifying new operating condition data (e.g., 50 sets of extrapolated samples), the model can quickly recommend the optimal parameter combination to support efficient operation of the equipment under a wide range of operating conditions.
[0050] (5) Improve engineering applicability and automation level, and standardize processes: From pre-separation design, experimental data collection to intelligent optimization, form a standardized technical route, reduce dependence on designer experience, and guide engineering improvements. BRIEF DESCRIPTION OF THE DRAWINGS
[0051] Figure 1 It is a flow chart of the optimization design method of the solid-liquid hydrocyclone of the present invention. DETAILED DESCRIPTION
[0052] In order to more clearly illustrate the technical solutions in the embodiments of the present invention, the following briefly introduces the drawings required for use in the description of the embodiments. Obviously, the drawings described below are only some embodiments of the present invention. For ordinary technicians in this field, other drawings can be obtained based on these drawings without creative work.
[0053] In order to make the purpose, technical solutions and advantages of the present invention more clear, the specific implementation of the present invention is further described below in conjunction with the accompanying drawings and embodiments, which are intended to explain rather than limit the present invention.
[0054] The present invention will be described in further detail below with reference to the accompanying drawings and specific embodiments.
[0055] like Figure 1 As shown, the present invention proposes an optimization design method for a solid-liquid hydrocyclone based on a GA-BP neural network, which includes the following steps:
[0056] Step 1: adding a triangular separation block to the inlet channel of the hydrocyclone to obtain a pre-separation type hydrocyclone;
[0057] Step 2: Design multiple groups of experiments for the pre-separation hydrocyclone using a mixed level design combined with additional test points, and select multiple key structural parameters as input variables of the BP algorithm, including column length L, cone angle θ, overflow pipe insertion depth L. insert , Overflow pipe diameter D over , bottom flow opening diameter D under and bottom flow outlet length L under ;
[0058] Step 3: After the hydrocyclone is running stably, the difference in separation efficiency between gangue and coal particles at the bottom flow outlet is selected as the output variable μ oi , then the L, θ, L insert 、D over 、D under 、L under 、μ oi Constitute a set of training samples;
[0059] Step 4: Use computational fluid dynamics (CFD) to perform numerical simulation calculations on the structural parameters obtained in step (2) to form a database corresponding to the difference between different hydrocyclone structural parameters and the gangue separation efficiency, and train the GA-BP neural network prediction model to output the hydrocyclone structural parameters corresponding to the maximum difference in gangue separation efficiency, that is, to obtain the optimal structural parameters for gangue separation.
[0060] Example
[0061] Select the column length L, cone angle θ, and overflow pipe insertion depth L of the solid-state hydrocyclone insert , Overflow pipe diameter D over , bottom flow opening diameter D under and bottom flow outlet length L under As the optimization variable X, then X=[L i ,θ i ,L insert,i ,D over,i ,D under,i ,L under,i ];
[0062] Multiple groups of experiments were designed using a mixed level design combined with additional test points. The same separation blocks were added to the inlet channel of each group of experiments. The constraints for optimizing the structural parameters in the first step were set as follows:
[0063] Table 1 Constraints of cyclone structural parameters
[0064]
[0065] In the third step, a fluid domain model was created based on the 72 sets of corresponding structural parameters, and the same mesh size was used for meshing. Numerical simulations were performed using ANSYS Fluent 2021R1, and the finite volume method (FVM) was used to solve the governing equations and boundary conditions. Table 2 lists the simulation settings, and Table 3 provides the boundary conditions. In the Mixture multiphase flow model, water is the primary phase, and particles and air are the secondary phases. The simulation results of the steady-state flow field are used as the initial solution for the unsteady flow field calculation, and the time step is set to 1×10 -4 s, the convergence residual is set to 1×10 -5The convergence criterion is based on the average solid phase flow balance at the inlet and outlet. The difference in the comprehensive separation efficiency of coal and gangue particles at the bottom flow outlet is μ oi As an evaluation index, the initial database is obtained;
[0066] Table 2 Simulation settings
[0067]
[0068] Table 3 Boundary conditions
[0069]
[0070] The obtained data graph is shown in the attached Figure 1 The optimal structural parameters are obtained on the Python programming software platform.
[0071] Through the above steps, it is possible to optimize multiple geometric parameters of a hydrocyclone for narrow density grade coking coal sorting. This method can also be extended to enhance the classification effect of the hydrocyclone and improve the working performance of the solid-liquid hydrocyclone.
[0072] The following is a comparison table 4 of the experimental results of slurry separation between the optimal structure hydrocyclone obtained by the optimization design method of the present invention and the general structure hydrocyclone:
[0073] Table 4 Comparison of experimental results of general hydrocyclone and optimal hydrocyclone
[0074]
[0075] From the data in the table, it can be seen that when the GA-BP neural network algorithm is used to design the pre-separation hydrocyclone for narrow density grade particle separation, the mineral recovery rate is significantly improved, and the working performance of the pre-separation hydrocyclone for solid-liquid narrow density grade particle separation is improved to meet the production needs of the industry.
[0076] The above embodiments are preferred implementation modes of the present invention, but the implementation modes of the present invention are not limited to the above embodiments. Any other changes, modifications, substitutions, combinations, and simplifications that do not deviate from the spirit and principles of the present invention should be considered as equivalent replacement methods and are included in the scope of protection of the present invention.
Claims
1. An optimization design method for a solid-liquid hydrocyclone based on a GA-BP neural network, characterized in that: The method comprises the following steps: Step 1: adding a triangular separation block to the inlet channel of the hydrocyclone to obtain a pre-separation hydrocyclone; Step 2: Design multiple groups of experiments for the pre-separation hydrocyclone using a mixed level design combined with additional test points, and select multiple structural parameters as input variables of the BP algorithm, including column length L, cone angle θ, overflow pipe insertion depth L. insert , Overflow pipe diameter D over , bottom flow opening diameter D under and bottom flow outlet length L under ; Step 3: After the hydrocyclone is running stably, the difference in separation efficiency between gangue and coal particles at the bottom flow outlet is selected as the output variable μ oi , then the L, θ, L insert 、D over 、D under 、L under 、μ oi Constitute a set of training samples; Step 4: Use computational fluid dynamics (CFD) to perform numerical simulation calculations on the structural parameters obtained in step (2) to form a database corresponding to the difference between different hydrocyclone structural parameters and the gangue separation efficiency, and train the GA-BP neural network prediction model to output the hydrocyclone structural parameters corresponding to the maximum difference in gangue separation efficiency, that is, to obtain the optimal structural parameters for gangue separation.
2. The optimization design method of a solid-liquid hydrocyclone based on a GA-BP neural network according to claim 1, characterized in that: This step 4 includes the following steps: Step 4.1: Define the mathematical transformation rules between different network layers in the BP neural network, convert the weighted input of the neuron into nonlinear or linear output, use the hidden layer transfer function to normalize the input data, enhance the nonlinear expression ability, and define tanh as the hidden layer transfer function: Among them, φ(n) is the neuron output value after the hidden layer transfer function, and n is the weighted sum of the inputs of the hidden layer neurons, that is, Among them, i is the group order of each sample, P is the total number of features, that is, the total number of hydrocyclone structural parameters, w i is the weight corresponding to the i-th input feature, x i Input features include column length L, cone angle θ, overflow pipe insertion depth L insert , Overflow pipe diameter D over , bottom flow opening diameter D under and bottom flow outlet length L under , b is the neuron bias term; In step 4.2, use the transfer function of the output layer to restrict the output to a probabilistic form, map the parameter values of the input structure to the interval (0, 1), and define the Sigmoid function as the transfer function of the output layer: Among them, φ(k) is the final predicted value of the output layer, k is the weighted sum of the input of the output layer neurons, and the calculation method is the same as that of the hidden layer neurons, that is, Among them, H is the number of hidden neurons, w j is the weight from the jth neuron in the hidden layer to the output layer, h j is the output value of the jth neuron in the hidden layer, and c is the bias term of the output layer; In step 4.3, a genetic algorithm (GA) is introduced to perform a global search to optimize the initial weights of the BP neural network and obtain the optimal structural parameters for gangue separation.
3. The optimization design method of a solid-liquid hydrocyclone based on a GA-BP neural network according to claim 2, characterized in that: The process of step 4.3 is as follows: Step 4.3.1, perform GA encoding of the initial population; Concatenate all weights w and bias terms b into real vectors in layer order: In this chromosome, the input layer X=6, the hidden layer H=3, and the output layer Q=1. A complete chromosome represents a complete set of hydrocyclone structural parameter prediction models, where: is the connection weight from the mth neuron in the input layer to the nth neuron in the hidden layer, mapping the hydrocyclone structural parameters to the hidden layer features, where m ranges from 1 to X and n ranges from 1 to H; is the activation threshold of the i-th neuron in the hidden layer, representing the bias of the hidden layer, is the connection weight from the i-th neuron in the hidden layer to the output layer, which represents the difference in separation efficiency between the intermediate hidden layer features μ oi The contribution weight of i ranges from 1 to H; is the output layer neuron adjustment term, representing the bias of the efficiency prediction layer; Initialize and randomly generate N pop chromosomes, that is, the population size, N pop The value ranges from 50 to 200, and each gene value is sampled from the uniform distribution U(-1,1) to ensure that the initialization iteration of the initial neural network covers the full design space of the hydrocyclone parameters; Step 4.3.2, fitness calculation Each vector in the chromosome is decoded into the weights and biases of the neural network, and the fitness function is calculated to quantify the separation performance of the current parameter combination: Among them, J is the total loss value between the BP neural network model prediction and the database result; Step 4.3.3 Select an action Each time, t individuals are randomly selected from the population, and the individual with the highest fitness is selected to enter the next generation, and this process is repeated until the size of the new population is the same as the original population. Step 4.3.4 Crossover Operation Randomly select two crossover points in the chromosome and exchange the middle segments of the parent chromosome. The chromosome segments represent the weights and biases of the neural network. The structural parameter combination of the cyclone is mapped according to the network weights and biases. By exchanging the middle segments, the advantageous features of different structural parameters can be recombined. Step 4.3.5 Mutation Operation Use Gaussian mutation with mutation probability p m = 0.001~0.1 Randomly perturb the genes in the chromosome: w new =w old +N(0,σ) Among them, w old is the current weight value; N(0,σ) is the Gaussian distribution random perturbation, which represents the adjustment of the weight, σ is 0.1, and p is used in the neural network prediction model. m = 0.001~0.1 mutation probability enters the new structure parameter iteration; w new is the new weight value after perturbation, representing the local optimization of the parameter mapping relationship; Step 4.3.6: Retain the best individuals. In each generation, the 10% individuals with the highest fitness are retained and directly enter the next generation. That is, the optimal structural parameter combination is retained. If the fitness improvement is less than 1% for 10 consecutive iterations, it is terminated early, indicating that a stable optimal structure has been found. Step 4.3.7 Weight import and BP fine-tuning Optimal chromosome decoding: Select the individual with the highest fitness from the final population and decode it into a weight matrix and bias vector, that is, extract the optimal weight value found by the GA algorithm to obtain the initial neural network prediction model; BP network initialization: import the decoded weights into the neural network, replace the random initial values of the model itself; call the built-in Adam optimizer in Python, use the adaptive learning rate a=0.001~0.01 optimization algorithm, set the number of iterations E epochs =200~500, balancing training cost and accuracy; if the validation set loss does not decrease for 20 consecutive rounds, training is terminated.
4. The optimization design method of a solid-liquid hydrocyclone based on a GA-BP neural network according to claim 2, characterized in that: The number of neurons in the hidden layer is determined by the following formula: Among them, β is a constant, Y is the number of input neurons, and X is the number of output neurons, which corresponds to the final optimization goal, that is, the difference in coal gangue separation efficiency.
5. The optimization design method of a solid-liquid hydrocyclone based on a GA-BP neural network according to claim 4, characterized in that: The value of β is 5 to 10.
6. The optimization design method of a solid-liquid hydrocyclone based on a GA-BP neural network according to claim 2, characterized in that: The square of the error is defined as the loss function to calculate the error between the true value of the database and the prediction of the neural network model: Among them, J is the total loss value between the BP neural network model prediction and the database result, l is the total number of samples, y i is the output after normalization of the difference in coal gangue separation efficiency in the database, o i is the prediction output of the BP neural network.