Main Nozzle Optimization Method Based on Hybrid Network and Multi-Strategy Particle Swarm Algorithm

Abstract

The present invention discloses a main nozzle optimization method based on a hybrid network and a multi-strategy particle swarm algorithm, belonging to the technical field of gas dynamics data processing. The main nozzle optimization method based on a hybrid network and a multi-strategy particle swarm algorithm utilizes a deep learning neural network combined with a machine learning algorithm to construct a regression model of the main nozzle structure parameters and velocity. The improved Newton-Raphson (INRBO) optimization algorithm is used to optimize the hyperparameters of the convolutional neural network-attention mechanism-random forest (CNN-Attention-RF) regression model. The velocity of the main nozzle jet at different monitoring points along the axis is predicted by the improved hybrid model, and then the heterogeneous improved dynamic particle swarm (HIDMS-PSO) optimization algorithm is used to find the main nozzle structure parameters corresponding to the highest velocity predicted by the surrogate model, improving the weft insertion efficiency of air-jet looms, reducing the R & D cost and cycle, and having important theoretical significance and engineering value.

Description

Technical Field

[0001] The present invention relates to the technical field of aerodynamic data processing, and particularly to a main nozzle optimization method based on a hybrid network and a multi-strategy particle swarm algorithm. Background Art

[0002] The main nozzle is a key component for air-jet weft insertion in air-jet looms. The jet characteristics of the main nozzle directly affect the efficiency, energy consumption of the air-jet loom, and the final quality of the fabric. To ensure high-speed air-jet weft insertion in the air-jet loom, the main nozzle jet needs to transport the weft yarn to the relay area in the profile reed at high speed and stably. The main nozzle structure design adopts a small throat section, a profiled hole flow guide groove, and a yarn guide tube with a large length-diameter ratio to meet the weft insertion requirements of "high speed and small flow rate". Due to the small internal space and complex structure of the main nozzle, it is difficult to measure the internal flow field of the main nozzle using devices such as pitot tubes and hot-wire anemometers. Therefore, multi-physics computational fluid dynamics (CFD) software is used to perform an integrated numerical simulation of the internal and external flow fields of the main nozzle, obtain the velocity on the centerline of the main jet flow and the velocity characteristic curves at different cross-sections, and reveal the influence law of the main nozzle structure parameters on the jet velocity distribution. Traditional optimization methods use empirical formulas or response surface optimization methods, which have insufficient fitting for complex fluids, low processing efficiency for high-dimensional variables, or fall into local optima, and cannot obtain the optimal main nozzle structure parameters. Summary of the Invention

[0003] The purpose of the present invention is to provide a main nozzle optimization method based on a hybrid network and a multi-strategy particle swarm algorithm, which uses a deep learning neural network combined with a machine learning algorithm to construct a regression model of the main nozzle mechanism parameters and velocity, uses an improved INRBO optimization algorithm to optimize the hyperparameters of the CNN-Attention-RF regression model, predicts the velocity of different points along the axis of the main nozzle jet through the improved model, and then uses the HIDMS-PSO algorithm to find the main nozzle structure parameters corresponding to the highest velocity predicted by the surrogate model, improving the weft insertion efficiency of the air-jet loom, reducing the R & D cost and cycle, and having important theoretical significance and engineering value.

[0004] To achieve the above purpose, the present invention provides a main nozzle optimization method based on a hybrid network and a multi-strategy particle swarm algorithm, including the following steps:

[0005] S1. Establish a three-dimensional flow field network model of the main nozzle, select the data of the design parameters, simulate the distribution laws of the internal and external flow fields of the three-dimensional flow field network models corresponding to different main nozzle structures, extract the center axis velocity curve, and design an experimental device to verify the velocity of the center axis monitoring points, obtain a data sample containing the axis velocity and design parameters, and preprocess the data sample;

[0006] S2. Construct a CNN-Attention-RF neural network model;

[0007] S3. Optimize the hyperparameters of the CNN-Attention-RF neural network model based on the INRBO algorithm;

[0008] S4. The INRBO-CNN-Attention-RF model predicts the air flow velocity of the main nozzle axis with different structures through training;

[0009] S5. Obtain the optimal structural parameters of the main nozzle based on the multi-strategy particle swarm optimization algorithm.

[0010] Preferably, in the S1, the acquisition of data samples is specifically as follows: According to the design parameters of the main nozzle, several different structural data are taken by using advanced Latin hypercube sampling within its corresponding value range. The general solver (Fluent or Star ccm+) of computational fluid dynamics (CFD) is used to set specific parameters such as boundary conditions and initial conditions. The jet axis velocity of different main nozzles is obtained through CFD numerical simulation to construct the training set and test set samples for deep learning. And a pitot tube experimental device is designed to test the main jet axis velocity, and the numerical results are verified to construct the validation set for deep learning.

[0011] Preferably, in the S2, in the CNN-Attention-RF neural network model, the convolutional neural network (CNN) is used to perform convolutional processing on the input structural parameter data and output axis jet velocity of the main nozzle, extract new data features, and enhance the correlation between the input data and the output data; the attention mechanism (Attention) is used to perform weighted fusion on the features extracted by the CNN so that the network can automatically focus on the information more useful for prediction; the random forest regression model (RF) balances the errors of unevenly distributed data in the sample and performs regression prediction.

[0012] Preferably, in the S3, the INRBO algorithm specifically includes the following steps:

[0013] 1) Initialize parameters;

[0014] 2) Initialize the population with a chaotic sequence;

[0015] 3) Calculate the adaptive inertia weight and use a dynamically adjusted inertia weight;

[0016] 4) Apply the Newton-Raphson search strategy (NRSR) and adjust the position update of the vector according to the difference between the optimal solution and the worst solution;

[0017] 5) Determine whether the random number is less than the difference factor (DF); if the random number < the difference factor, perform TAO update, if the random number ≥ the difference factor, perform local optimum detection. When the improvement of the optimal solution detected by the local optimum detection is less than the preset threshold, increase the stagnation counter. When the stagnation counter reaches a certain number of times, it is determined that the current search has fallen into the local optimum, trigger chaotic perturbation, use the chaotic perturbation Logistic mapping to update the position of the solution, apply perturbation to the current position, so that the particle jumps out of the local optimum region and resumes the global search ability;

[0018] 6) Determine whether the iteration terminates. If the iteration terminates, update the convergence curve to record the best and output the result.

[0019] Preferably, in the above (2), initializing the population with a chaotic sequence is specifically:

[0020] Use the Tent chaotic mapping to initialize the position of the vector. By introducing a chaotic sequence, a more random and unpredictable initialization position is generated to improve the exploration ability of the algorithm;

[0021] For each vector position , randomly generate within the interval ([0,1]): ;

[0022] Apply the Tent mapping to each randomly generated value: ;

[0023] Map the chaotic sequence to the specified search space range: ;

[0024] Wherein, UB is the upper bound of the search space, LB is the lower bound of the search space, is the parameter of the Tent mapping.

[0025] Preferably, in the above (3),

[0026] Inertia weight formula: ;

[0027] Wherein, is the inertia weight of the current iteration, is the current iteration number, is the maximum iteration number, is the maximum value of the inertia weight, is the minimum value of the inertia weight, is to control the steepness of the Sigmoid function, is to control the center position of the Sigmoid function.

[0028] Preferably, in the above (4), the NRSR search specifically includes:

[0029] Calculate , ;

[0030] Calculate the first NRSR,

[0031] , is a random number sampled from a standard normal distribution.

[0032] Propose three difference operators to improve the parameter , guide the population to the correct direction, improve the calculation efficiency of the NRBO algorithm, and the proposed expression of the parameter is as follows:

[0033] ;

[0034] ;

[0035] ;

[0036] ;

[0037] In this formula, and are different integers randomly selected from the population, is the best vector position , is the current position of the vector , and it will be continuously updated.

[0038] Update the position , ;

[0039] Calculate and , ;

[0040] Calculate the second NRSR, ;

[0041] Among them, is a random perturbation used to adjust the current solution,

[0042] NRSR is a search rule based on random perturbation,

[0043] Respectively generate new search directions by randomly combining Position with random weights,

[0044] Position update​ To select different update strategies according to the values of positions

[0045] The second NRSR is to further adjust the search direction and increase the diversity and flexibility of the search.

[0046] Preferably, in S5, the multi-strategy particle swarm algorithm divides the particle swarm into several units, each unit contains a main particle and several slave particles, and the particles within the same unit interact with each other and exchange information with other units.

[0047] Preferably, the initial population of the multi-strategy particle swarm algorithm is divided into two equal homogeneous subgroups and a heterogeneous subgroup. The homogeneous subgroups use the update equation of the classical PSO algorithm, and the heterogeneous subpopulation forms a unit structure and adopts an inward movement strategy or an outward movement strategy. The inward movement strategy guides the particles to the same unit, and the outward movement strategy uses samples from different units to guide the particles.

[0048] Preferably, the inward movement strategy provides guidance for the particles by using the position information obtained from all unit members.

[0049] The main particles randomly select any one of the following three equations to update their velocities:

[0050] ;

[0051] ;

[0052] ;

[0053] Where: is the velocity of the main particle, is the position of the main particle at time , is the slave particle with the smallest position similarity in the th unit, is the position of the most suitable slave particle in the th unit, is the average position of all slave particles within the current unit of the main particle, is the inertia weight, the constant and are acceleration coefficients, and are random vectors ∈ [0, 1];

[0054] The slave particles move according to the following formula:

[0055] ;

[0056] Where, For the particle velocity, For the particle at time t The best position found, For the particle position, For the Main particle position of the

[0057] The outward movement strategy provides guidance for the particles from different cells in the population while maintaining the hierarchy,

[0058] The main particles randomly select any one of the following three equations to update their velocities:

[0059] ;

[0060] ;

[0061] ;

[0062] Where, Is the main particle velocity, For the Average position of the particles in the Main particle position of the randomly selected cell, Is the average position of the members of the particle's own cell;

[0063] The slave particles move from the particles to the same type of random slave particles belonging to another cell using the following velocity update equation: .

[0064] The main nozzle optimization method based on the hybrid network and multi-strategy particle swarm algorithm of the present invention uses a deep learning neural network combined with a machine learning algorithm to construct a regression model of the main nozzle mechanism parameters and velocity, and uses an improved INRBO optimization algorithm to optimize the hyperparameters of the CNN-Attention-RF regression model. The velocity of the main nozzle jet at different points along the axis is predicted by the improved model, and then the main nozzle structure parameters corresponding to the highest velocity predicted by the surrogate model are found using the HIDMS-PSO algorithm, which improves the weft insertion efficiency of the air-jet loom, reduces the R & D cost and cycle, and has important theoretical significance and engineering value.

[0065] The technical solution of the present invention will be further described in detail below with reference to the drawings and embodiments. Description of the Drawings

[0066] Figure 1 Is the network architecture diagram of the HIDMS-PSO-CNN-Attention-RF for optimizing the main nozzle flow rate in the embodiment of the present invention;

[0067] Figure 2 Structural diagram of the main nozzle according to an embodiment of the present invention;

[0068] Figure 3 Structural diagram of the nozzle core according to an embodiment of the present invention;

[0069] Figure 4 Cross-sectional structural diagram of the nozzle core according to an embodiment of the present invention;

[0070] Figure 5 Three-dimensional flow field network model of the main nozzle according to an embodiment of the present invention;

[0071] Figure 6 CNN convolutional neural network architecture according to an embodiment of the present invention;

[0072] Figure 7 Structural diagram of the convolutional attention module CBAM according to the present invention;

[0073] Figure 8 RF network structure according to an embodiment of the present invention;

[0074] Figure 9 Structural diagram of the CNN-Attention-RF model according to an embodiment of the present invention;

[0075] Figure 10 INRBO flow chart according to an embodiment of the present invention;

[0076] Figure 11 Degree of fitting of training samples according to an embodiment of the present invention;

[0077] Figure 12 Schematic diagram of model training according to an embodiment of the present invention;

[0078] Figure 13 Topological structure diagram of heterogeneous particle swarm according to an embodiment of the present invention;

[0079] Figure 14 Nozzle optimization algorithm flow chart of the heterogeneous improved multi-strategy particle swarm algorithm according to an embodiment of the present invention;

[0080] Figure 15 Comparison diagram of results before and after optimization according to an embodiment of the present invention.

[0081] Reference numerals

[0082] 1. Nozzle core; 2. Nozzle body; 3. Conical sleeve; 4. Yarn guide tube; 5. Weft yarn inlet; 6. High-pressure air inlet; 7. Annular air chamber; 8. Conical air chamber. Detailed description of the specific implementation mode

[0083] The following will describe in detail the embodiments of the present invention with reference to the accompanying drawings.

[0084] Embodiment

[0085] As shown Figure 2 in the figure, the main nozzle consists of a nozzle core 1, a nozzle body 2, a conical sleeve 3, a yarn guide tube 4, etc. High-pressure air flow enters the annular air chamber 7 through the high-pressure air inlet 6 of the main nozzle, and after passing through the diversion groove into the conical air chamber 8, the wake region and the subsonic acceleration region, it jets out from the yarn guide tube 4. The weft yarn enters from the weft yarn inlet 5. The throat of the main nozzle is located in the conical region between the small-diameter end face of the conical surface of the nozzle core 1 and the inner wall of the yarn guide tube 4. The air flow accelerates at the throat, and the Mach number reaches or even exceeds 1. In the weft yarn inlet, the dynamic pressure increases from small to large until the conical negative pressure region, which is beneficial to sucking the weft yarn into the yarn guide tube 4. After passing through the conical region, the dynamic pressure change region becomes stable. The air flow outside the yarn guide tube 4 of the main nozzle belongs to a free submerged circular jet. Due to entrainment and diffusion effects, the velocity of the main nozzle jet decreases along the central axis direction, resulting in a sharp drop.

[0086] The structure of the nozzle core 1 can change the jet velocity of the main nozzle. By optimizing the structure of the nozzle core 1, the highest jet velocity Z can be obtained. The structure of the nozzle core 1 is as Figure 3 、 Figure 4 shown. The main structural design parameters of the nozzle core 1 include: the outer diameter Dh of the annular air chamber 7, the length Ld of the diversion groove, the large-end diameter Dz1 of the conical surface of the nozzle core 1, the small-end diameter Dz2 of the conical surface of the nozzle core 1, the length Lz1 of the conical section of the nozzle core 1, and the diameter E of the circular hole diversion groove. In order to find the optimal structural parameters, the above six design parameters are taken as the optimization objects.

[0087] As Figure 1 shown. An optimization method for the main nozzle based on a hybrid network and a multi-strategy particle swarm algorithm includes the following steps:

[0088] S1. Establish a three-dimensional flow field network model of the main nozzle, as Figure 5 shown. Select the data of the main nozzle design parameters, simulate the internal and external flow field distribution laws corresponding to the three-dimensional flow field network models of different main nozzle structures, extract the central axis velocity curve, and design an experimental device to verify the velocity of the central axis monitoring points, obtain a data sample containing the axis velocity and design parameters, and preprocess the data sample.

[0089] The acquisition of the data sample is specifically as follows: According to the above six design parameters, 1000 groups of different data are taken by using advanced Latin hypercube sampling within their corresponding value ranges. Set specific parameters such as boundary conditions and initial conditions in the Fluent or Star ccm+ multi-physics computational fluid dynamics solver, including temperature, total pressure, static pressure, turbulent kinetic energy, and turbulent kinetic energy dissipation rate. Through CFD numerical simulation, the internal and external flow field characteristics corresponding to the above three-dimensional flow field network models are obtained, and 1000 main nozzle axis velocity curves are obtained.

[0090] Combine the above 1000 sets of design parameters with their corresponding 1000 sets of main nozzle axis velocities to obtain 1000 sets of data samples. Perform standardization and normalization processing on the data samples, and normalize the data samples into the interval [0, 1]. Divide the data samples into a 78% training set, a 15% validation set, and a 7% test set. Design a pitot tube experimental device to measure the main jet axis velocity, verify the numerical results, and construct a validation set for deep learning.

[0091] S2. Construct a CNN-Attention-RF neural network model.

[0092] In the CNN-Attention-RF neural network model, first use the convolutional neural network CNN to perform convolutional processing on the input parameter data of the main nozzle shape structure and the output axis jet velocity to better extract new data features and enhance the correlation between the input data and the output data. Then use the attention mechanism Attention to perform weighted fusion on the features extracted by CNN so that the network can automatically focus on information that is more useful for prediction. Introduce the random forest regression model RF to balance the errors of unevenly distributed data in the sample, and it has good tolerance for outliers and noise. RF can not only handle large-scale multi-dimensional data but also is suitable for non-linear problems.

[0093] The prediction method based on the CNN-Attention-RF hybrid neural network combines three network structures: the convolutional neural network (CNN), the attention mechanism (Attention), and the random forest (RF), which improves the accuracy and efficiency of model prediction.

[0094] The model architecture of the CNN-Attention-RF hybrid neural network is as Figure 9 shown.

[0095] As Figure 6 shown. Input layer: It includes the number of data samples contained in a batch, the number of time steps contained in each data sample, and the feature dimension of the input data.

[0096] Convolutional layer (CNN): Use a 2D convolutional layer to convert the input feature dimension into a higher feature dimension. After the convolutional operation, the output shape.

[0097] The calculation through the convolutional layer can be expressed as: Z = Conv2D(X), where the shape of Z is (batch_size, 64, sequence_length), indicating that for each input sample, the convolutional layer generates 64 feature maps, and the length of each feature map is the same as the length of the input sequence. These feature maps capture different feature patterns in the input sequence for further processing by subsequent layers.

[0098] The CNN is mainly used to capture local features in sequential data. The multi-dimensional features of the input simulation data are used to extract features through the convolutional layer.

[0099] Attention layer: Apply the Attention mechanism to weight the output of the CNN, enabling the model to focus on different parts of the input sequence.

[0100] Such as Figure 7 shown. CBAM mechanism: The specific calculation of the attention mechanism of the convolutional module can be abstracted and summarized into three stages:

[0101] In the first stage, global max pooling and global average pooling are respectively performed on the width and height, and the features are output through the MLP. The channel attention features and the input features are subjected to an element-wise weighting operation, and then passed through the sigmoid activation function to generate the channel attention features. And the channel attention features and the input features are subjected to an element-wise multiplication operation to generate the input features of the spatial attention module.

[0102] In the second stage, the spatial attention mechanism operation is performed to compress the channels, and average pooling and max pooling are respectively performed on the channel dimension.

[0103] In the third stage, the features extracted previously are merged into a 2-channel feature to generate the output features.

[0104] The relevant formulas are as follows:

[0105] ;

[0106] ;

[0107] Among them, is the channel attention weight, is the spatial attention weight, Performs average pooling on the convolutional layer, Performs max pooling on the convolutional layer.

[0108] The attention mechanism is used to perform weighted fusion on the features extracted by the CNN so that the network can automatically focus on the information that is more useful for prediction.

[0109] Such as Figure 8As shown. In the third process, according to the random forest, the feature vectors generated by the CNN feature extraction network are used as the input features of the RF regressor and divided into a test set and a training set. The bootstrap sampling method in RF is used to sample on the training set, generate its corresponding training subset for each decision tree, and train the decision tree. The optimal attribute is selected for splitting at each internal node to generate the final output.

[0110] S3. Optimize the hyperparameters of the CNN-Attention-RF neural network model based on the improved INRBO algorithm.

[0111] The improved algorithm proposed based on the Newton-Raphson optimizer (NRBO) is INRBO.

[0112] The NRBO optimization algorithm is based on two key mechanisms: the Newton-Raphson search rule (NRSR) and the trap avoidance operator (TAO). NRSR accelerates the update of the position by using the first and second derivatives, thus improving the search ability and convergence speed of the algorithm. TAO helps the algorithm avoid falling into local optimal solutions by increasing the randomness and diversity of the solutions.

[0113] As Figure 10 shown. The specific steps of the INRBO algorithm are as follows:

[0114] 1) Initialize the parameters, including the population size, the maximum number of iterations, the difference factor, the upper limit of the hyperparameter vector, and the lower limit of the hyperparameter vector.

[0115] 2) Initialize the population with a chaotic sequence.

[0116] Specifically:

[0117] Use the Tent chaotic map to initialize the positions of the particles. This method can enhance the diversity of the initial population and avoid the algorithm falling into local optimal solutions at the initial stage. By introducing a chaotic sequence, more random and unpredictable initialization positions are generated, improving the exploration ability of the algorithm.

[0118] For each particle position , randomly generate within the interval ([0, 1]): ;

[0119] Apply the Tent map to each randomly generated value: ;

[0120] Map the chaotic sequence to the specified search space range: ;

[0121] where UB is the upper bound of the search space,LB is the lower bound of the search space, and is the parameter of the Tent mapping.

[0122] 3) Adaptive inertia weight calculation, using a dynamically adjusted inertia weight. This weight gradually changes with the number of iterations. In the initial stage of the algorithm, the inertia weight is large, tending to global search, thus increasing the exploration space; in the later stage of iteration, the inertia weight decreases, being more inclined to local exploitation. This strategy helps to widely explore the solution space in the early stage and converge to the optimal solution in the later stage.

[0123] Inertia weight formula: ;

[0124] where, is the inertia weight of the current iteration, is the current iteration number, is the maximum number of iterations, is the maximum value of the inertia weight, is the minimum value of the inertia weight, is to control the steepness of the Sigmoid function, and is to control the center position of the Sigmoid function.

[0125] 4) Apply NRSR search, adjusting the position update of particles according to the difference between the optimal solution and the worst solution.

[0126] The NRSR search specifically includes:

[0127] Calculate , ;

[0128] Calculate the first NRSR,

[0129] , where rand is a random number sampled from the standard normal distribution.

[0130] Introduce the exponential decay factor , , the linear decay factor

[0131] ;

[0132] ; ;

[0133] Use , , to dynamically generate the parameter

[0134] ;

[0135] Randomly select four individual indices for difference, and combine the parameters , update the position , ;

[0136] Calculate and , ;

[0137] Calculate the second NRSR, ;

[0138] Among them, is a random perturbation used to adjust the current solution,

[0139] NRSR is a search rule based on random perturbation,

[0140] Position update selects different update strategies according to the value of the position,

[0141] , respectively generate new search directions by combining Position with random weights, The second NRSR further adjusts the search direction to increase the diversity and flexibility of the search

[0142] ity.

[0143] 5) Judge whether the random number is less than the difference factor. If the random number < difference factor, perform TAO update. If the random number ≥ difference factor, perform local optimum detection. When the local optimum detection detects that the improvement of the optimal solution is less than the preset threshold, increase the stagnation counter. When the stagnation counter reaches a certain number of times, it is determined that the current search has fallen into a local optimum, thus triggering chaotic perturbation. Use the chaotic perturbation Logistic map to update the position of the solution, apply perturbation to the current position, so that the particle jumps out of the local optimum region and restores the global search ability.

[0144] Iterate the Logistic map for 10 times to enhance the chaotic effect: .

[0145] Logistic map parameter Take 4.

[0146] The number limit of local optimum detection is set to 10, and the improvement threshold is 1×10 -6 .

[0147] 6) Judge whether the iteration terminates. If the iteration terminates, update the convergence curve to record the best and output the result.

[0148] S4. The INRBO-CNN-Attention-RF model predicts the air flow velocity of the main nozzle axis with different structures through training.

[0149] The INRBO algorithm is used to optimize the hyperparameters of the CNN-Attention-RF model: the size of the convolutional kernel, the number of neurons in each convolutional layer, the number of samples per iteration and the number of training epochs, the number of trees in the random forest, the maximum depth of the trees, the minimum number of samples required for each internal node, the minimum number of samples required for each leaf node, and the learning rate. First, set the parameters of the INRBO algorithm, the maximum number of iterations, the population size, the deep learning epoch Epoch, the objective function, the minimum and maximum value ranges of the hyperparameter vector. Configure the batch size [32, 128] and the convolutional kernel size [1, 2] for CNN1; the batch size [16, 64] and the convolutional kernel size [1, 2] for CNN2; the number of trees in the random forest [10, 200], the maximum depth of the trees [3, 30], the minimum number of samples for each internal node [2, 10], and the minimum number of samples required for each leaf node [1, 4] to build and train the hybrid network model, and use the root mean square error (RMSE) as the loss function for the regression task. RMSE measures the difference between the predicted value and the actual value to find the optimal hyperparameters of the hybrid model.

[0150] ;

[0151] The performance of the hybrid network model is evaluated by the mean absolute error (MAE), the root mean square error (RMSE), and the coefficient of determination to determine the effectiveness of each parameter configuration.

[0152] Through multiple iterations (multiple epochs, in each epoch, forward propagation is performed, the output is calculated, the loss is calculated, backpropagation is performed, and the model parameters are updated), the INRBO algorithm gradually optimizes the hyperparameters of the hybrid model. In each iteration, the algorithm evaluates the performance of the current parameter settings, monitors the values of the loss function (RMSE) and evaluation metrics (R 2 、MAE) on the validation set, and evaluates the performance of the hybrid network model. And adjust according to the natural rhythm strategy to explore better solutions. Finally, build a hybrid CNN-Attention-RF network using the optimized hyperparameter configuration. The parameter settings of the INRBO-optimized CNN-Attention-RF hybrid network model are shown in Table 1.

[0153] Table 1 Parameter settings of the INRBO-CNN-Attention-RF hybrid network model

[0154]

[0155] The fitting degree of the training samples of the CNN-Attention-RF hybrid network model optimized by INRBO is as follows Figure 11 shown, and the model training process is as follows Figure 12 shown.

[0156] Using simulation data, the hyperparameters of the CNN-Attention-RF model were optimized by the improved INRBO algorithm. The comparison of evaluation metrics of the INRBO-CNN-Attention-RF, CNN, RF, and CNN-Attention prediction models is shown in Table 2.

[0157] Table 2 Comparison of evaluation metrics of different prediction models

[0158]

[0159] The results show that: compared with the individual CNN network or RF machine learning regression model and the CNN-Attiention deep learning model, the CNN-Attention-RF hybrid model has a smaller prediction error and higher prediction accuracy for the jet velocity of the main nozzle axis.

[0160] The trained hybrid network can be used for velocity prediction. Inputting new structural parameter data, the network will output the corresponding velocity prediction value.

[0161] S5. Obtain the optimal structural parameters of the main nozzle based on the multi-strategy particle swarm algorithm.

[0162] Combining deep learning and machine learning to construct the surrogate model INRBO-CNN-Attention-RF, predicting the airflow velocity of the main nozzle with different structures through training. Based on the trained prediction set, using the heterogeneous improved multi-strategy particle swarm algorithm HIDMS-PSO, the best output, the highest outlet velocity, is found according to the given input value range, and the corresponding input main nozzle structural parameters are the optimal structure. The flow chart of the HIDMS-PSO algorithm is as follows Figure 14 shown.

[0163] As Figure 13 shown. The heterogeneous improved multi-strategy particle swarm algorithm HIDMS-PSO is an advanced metaheuristic optimization algorithm based on the particle swarm optimization algorithm (PSO). This algorithm utilizes heterogeneous inertia weights and dynamic multi-subgroup strategies to improve search efficiency and convergence performance. Specifically, HIDMS-PSO divides the particle swarm into several units, each unit containing a main particle and three slave particles. The particles within different units not only interact with each other but also can exchange information with other units to avoid premature convergence.

[0164] The unit structure hierarchy among particles and the hypothesized differences among subordinate particles are used to restrict and control the flow of information between particles, rather than global or arbitrary information exchange. The communication model can be summarized by the following rules: 1. The master particle of the th unit does not communicate directly with the master particle of the th unit, and communication is established only through subordinate particles. 2. Master particles can only exchange information with subordinate particles within the unit. 3. Subordinate particles can only communicate with other unit subordinate particles of the same type, so they cannot communicate with subordinate particles within the unit.

[0165] (1) Classical Particle Swarm Optimization

[0166] The swarm consists of N particles, each with a velocity of , at position at time , with the individual best position being , and the global best position . The particles move in a -dimensional search space by using three vectors , , and , i.e., the position of the global best particle in the swarm. The standard PSO algorithm uses the following two equations to update the velocity and position of the particles:

[0167] ;

[0168] ;

[0169] where is the inertia weight, which affects the inertia of the particles. The constants and are acceleration coefficients, which control the velocity of the particles flying towards and , respectively. and are random vectors ∈ [0, 1].

[0170] (2) Search Method

[0171] In the HIDMS-PSO algorithm, the initial population is divided into two equal subpopulations, one is homogeneous and the other is heterogeneous, and each subpopulation adopts a different motion strategy. The homogeneous subpopulation uses the update equation of the classical PSO algorithm, while the heterogeneous subpopulation forms N unit structures and adopts inward and outward strategies. The particles in the heterogeneous subgroup randomly select one of the strategies. The inward-oriented strategy aims to guide the particles to the same unit. On the contrary, the outward-oriented strategy uses samples from different units to guide the particles.

[0172] Inward movement strategy: The inward movement strategy uses the position information obtained from all unit members to guide the particles. This strategy essentially allows each unit to explore and exploit local solutions simultaneously.

[0173] The master particles randomly select any one of the following three equations to update their velocities:

[0174] ;

[0175] ;

[0176] ;

[0177] Where: is the velocity of the master particle, is the position of the master particle at time the position, is the slave particle with the smallest position similarity in the th unit, is the position of the most suitable slave particle in the th unit, is the average position of all slave particles within the current unit of the master particle;

[0178] The slave particles move according to the following formula:

[0179] ;

[0180] Where, is the velocity of the slave particle, is the best position found by the particle at time the position, is the position of the slave particle, is the th unit of the master particle position;

[0181] Outward movement strategy: Compared with the inward movement strategy, the outward movement strategy provides guidance for the particles from different units in the population while maintaining the hierarchy.

[0182] The master particles randomly select any one of the following three equations to update their velocities:

[0183] ;

[0184] ;

[0185] ;

[0186] Where, is the velocity of the master particle, is the The average position of particles in a cell, is the position of the main particle of a randomly selected cell, which is the average position of the members of the cell to which the particle itself belongs;

[0187] The particle moves from the particle to a randomly selected slave particle of the same type belonging to another cell using the following velocity update equation:

[0188] .

[0189] By combining these two strategies and allowing a part of the population to perform classical PSO position updates, the entire population is divided into homogeneous groups and heterogeneous groups, and the heterogeneous population is further divided into cells for effective exploration and exploitation. The homogeneous population focuses on a single solution output, while the heterogeneous population focuses on multiple solutions.

[0190] The optimal parameter design is obtained by using the main nozzle optimization method based on the hybrid network and multi-strategy particle swarm algorithm described in the present invention. The comparison between the optimal design parameters and other parameters is shown in Table 3.

[0191] Table 3 Comparison between the optimal structural parameter design and other parameters

[0192]

[0193] As can be seen from Table 3, under the condition of the same air supply pressure of 0.4 MPa for different structural main nozzles, through numerical simulation and Pitot tube experiment tests, the highest velocity along the axis of the main jet is obtained. It can be seen from Table 3 that the structural parameter of the main nozzle in Group No. 1 is the optimized one, and the maximum outlet velocity is 348.42 m / s, which is higher than the jet velocities of the main nozzles in other group numbers, and the flow rate is , and the air consumption is smaller than that of the main nozzles in Group No. 2, 3, 4, and 5. The results in Table 3 show that: under the same pressure condition, the jet velocity along the axis of the main nozzle structure optimized by the HIDMS-PSO algorithm increases significantly, and the air consumption is reduced to a certain extent. The production unit uses the optimized main nozzle on the air-jet loom, which can improve the weft insertion efficiency, reduce the air consumption per unit time, thereby reducing the power consumption and the production cost of weaving.

[0194] Figure 15 This is the comparison chart of the results before and after optimization of the embodiment of the present invention. As can be seen from Figure 13 it, after the main nozzle parameters are optimized, the outlet jet velocity of the main nozzle is significantly improved, and the outlet flow rate is reduced.

[0195] Therefore, by adopting the main nozzle optimization method based on the hybrid network and multi-strategy particle swarm algorithm of the present invention, a regression model of the main nozzle mechanism parameters and velocity is constructed by combining the deep learning neural network with the machine learning algorithm.

[0196] The improved INRBO optimization algorithm is used to optimize the hyperparameters of the CNN-Attention-RF regression model. The velocity of the main nozzle jet at different points along the axis is predicted by the improved model. Then, the HIDMS-PSO algorithm is used to find the main nozzle structure parameters corresponding to the highest velocity predicted by the surrogate model, which improves the weft insertion efficiency of the air-jet loom, reduces the R & D cost and cycle, and has important theoretical significance and engineering value.

[0197] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention and are not intended to limit them. Although the present invention has been described in detail with reference to the preferred embodiments, those of ordinary skill in the art should understand that they can still modify or equivalently replace the technical solutions of the present invention, and these modifications or equivalent replacements cannot make the modified technical solutions deviate from the spirit and scope of the technical solutions of the present invention.

Claims

1. A main nozzle optimization method based on hybrid network and multi-strategy particle swarm algorithm, characterized in that: The following steps are involved: S1. Establish a three-dimensional flow field numerical model of the main nozzle, select the data of the main nozzle design parameters, simulate the internal and external flow field distribution laws of the three-dimensional flow field model corresponding to different main nozzle structures, extract the central axis velocity curve, and design an experimental device to verify the central axis monitoring point velocity, obtain data samples containing the axis velocity and the main nozzle structure design parameters, and pre-process the data samples; S2. Build a CNN-Attention-RF hybrid model; S3. Optimize the hyperparameters of the CNN-Attention-RF hybrid network model based on the INRBO algorithm; S4, INRBO-CNN-Attention-RF model predicts the airflow velocity along the axis of the main nozzle with different structures through training; S5. Based on the multi-strategy particle swarm optimization algorithm, the optimal structural parameters of the main nozzle are obtained to increase the jet speed of the main nozzle, thereby improving the weft insertion efficiency; In S3, the INRBO algorithm specifically includes the following steps: 1) Initialization parameters; 2) Initialize the population using chaotic sequences; 3) Adaptive inertia weight calculation, using dynamically adjusted inertia weight; 4) Apply the Newton-Raphson search strategy to adjust the position update of the vector according to the difference between the optimal solution and the worst solution; 5) Determine whether the random number is less than the difference factor; if the random number is less than the difference factor, perform TAO update; if the random number is greater than or equal to the difference factor, perform local optimal detection. When the local optimal detection detects that the improvement of the optimal solution is less than the preset threshold, increase the stagnation counter. When the stagnation counter reaches a certain number of times, it is determined that the current search is trapped in the local optimal state, triggering chaotic perturbation. Use chaotic perturbation Logistic mapping to update the position of the solution, apply perturbation to the current position, so that the particle jumps out of the local optimal area and restores the global search capability. 6) Determine whether the iteration is terminated. If the iteration is terminated, update the convergence curve to record the best result and output the result.

2. The main nozzle optimization method based on hybrid network and multi-strategy particle swarm algorithm according to claim 1 is characterized in that: In S1, the data samples are obtained specifically as follows: according to the main nozzle design parameters, a number of different structure data are sampled using advanced Latin hypercube sampling within the corresponding value range, boundary conditions and initial conditions are set using a general solver of computational fluid dynamics, and the velocity along the axis of the main nozzle jet is obtained through numerical simulation using the general solver of computational fluid dynamics, training set and test set samples for deep learning are constructed, and a Pitot tube experimental device is designed to test the axis velocity of the main jet, the numerical results are verified, and a verification set for deep learning is constructed.

3. The main nozzle optimization method based on hybrid network and multi-strategy particle swarm algorithm according to claim 2 is characterized in that: In S2, in the CNN-Attention-RF neural network model, the convolutional neural network is used to perform convolution processing on the main nozzle input structural parameter data and the output jet axis velocity, extract new data features, and enhance the correlation between the input data and the output data; The attention mechanism is used to perform weighted fusion on the features extracted by the convolutional neural network so that the network can automatically focus on information that is more useful for prediction; the random forest regression model balances the errors of unevenly distributed data in the samples and performs regression prediction.

4. The main nozzle optimization method based on hybrid network and multi-strategy particle swarm algorithm according to claim 3 is characterized in that: In 2), the population is initialized using a chaotic sequence as follows: The position of the initialization vector is initialized using the Tent chaotic map. By introducing a chaotic sequence, a more random and unpredictable initialization position is generated, which improves the exploration ability of the algorithm. For each particle position , randomly generated in the interval ([0,1]): ; Apply the Tent map to each randomly generated value: ; The chaotic sequence Map to the specified search space range: ; in, UB is the upper bound of the search space, LB is the lower bound of the search space, It is the parameter of Tent mapping.

5. The main nozzle optimization method based on hybrid network and multi-strategy particle swarm algorithm according to claim 4 is characterized in that: In the above 3), Inertia weight formula: ; in, is the inertia weight of the current iteration, is the current iteration number, is the maximum number of iterations, is the maximum value of the inertia weight, is the minimum value of the inertia weight, To control the steepness of the Sigmoid function, To control the center position of the Sigmoid function.

6. The main nozzle optimization method based on hybrid network and multi-strategy particle swarm algorithm according to claim 5 is characterized in that: In 4), NRSR is a search rule based on random perturbations, specifically including: calculate , ; Introducing three operators, modifying Parameters guide the population to update in the correct direction. The expression is: ； ； ； ； is a random number that follows a standard normal distribution. are four individual indices randomly selected from the population for differential calculation. is the best vector position , is the current position of the vector , is the exponential decay factor, is the linear attenuation factor; Calculate the first NRSR, ; Update Location , ; calculate and , ; Calculate the NRSR of the second one, ； in, is a random perturbation used to adjust the current solution, are different integers randomly selected from the population; NRSR is a search rule based on random perturbations. Location Updates To select different update strategies based on the value of the position, They are respectively combined by random weights Position, Generate new search directions, The second NRSR is to further adjust the search direction and increase the diversity and flexibility of the search.

7. The main nozzle optimization method based on hybrid network and multi-strategy particle swarm algorithm according to claim 6 is characterized by: In S5, the multi-strategy particle swarm algorithm divides the particle swarm into several units, each unit includes a master particle and several slave particles, and the particles in the same unit interact with each other and exchange information with other units.

8. The main nozzle optimization method based on hybrid network and multi-strategy particle swarm algorithm according to claim 7 is characterized in that: The initial population of the multi-strategy particle swarm algorithm is divided into two equal isomorphic subgroups and heterogeneous subgroups. The isomorphic subgroup uses the classic PSO algorithm to update the equation, and the heterogeneous subgroup forms The inward moving strategy guides particles to the same unit, and the outward moving strategy uses samples from different units to guide particles.

9. The main nozzle optimization method based on hybrid network and multi-strategy particle swarm algorithm according to claim 8, characterized in that: The inward movement strategy uses the position information obtained from all unit members to provide guidance for the particles. The master particles randomly choose any of the following three equations to update their velocity: ； ； ； in: is the main particle velocity, The main particle at time location, For the The particle with the smallest position similarity among the units, It is The unit is best suited for the particle position, is the average position of all slave particles in the current unit of the master particle, is the inertia weight, constant and is the acceleration factor, and is a random vector ; The slave particles are moved using the following formula: ； in, is the particle velocity, For a particle at time The best location found, is the particle position, For the The main particle position of each unit; The outward movement strategy provides particles with guidance from different units in the population while maintaining the hierarchy. The master particles randomly choose any of the following three equations to update their velocity: ； ； ； in, is the main particle velocity, For the The average position of particles in a cell, is the position of the primary particle of the randomly selected unit, is the average position of the particle's own unit members; and is a random vector ; A slave particle moves toward a random slave particle of the same type belonging to another cell using the following velocity update equation: .

Images