A laser cutting process design method and system

Abstract

The application provides a laser cutting process design method and system, comprising: determining a trained neural network model; comprising: a BP hidden layer and an ELM hidden layer; determining a target roughness of a laser cutting product, and preliminarily designing multiple groups of process parameters according to the target roughness by using a teach-and-learn algorithm; inputting each group of the preliminarily designed process parameters into the trained neural network model to respectively predict the upper and lower surface roughness of the product cut according to the process parameters; combining the target roughness and the roughness predicted by each group of the process parameters, and selecting an optimal group of process parameters from the multiple groups of the preliminarily designed process parameters by using a fuzzy optimization method, so as to design corresponding process parameters for the target roughness again based on the teach-and-learn algorithm, and cyclically performing the roughness prediction, target function selection and process parameter redesign process until the predicted upper and lower surface roughness reaches the target roughness. The application improves the quality of the laser cutting part.

Description

Technical Field

[0001] This invention belongs to the field of laser cutting, and more specifically, relates to a laser cutting process design method and system. Background Technology

[0002] Laser cutting utilizes a high-power-density laser beam to irradiate the material being cut, rapidly heating it to its vaporization temperature and causing it to evaporate and form a hole. As the beam moves across the material, the hole continuously forms a narrow kerf, completing the cut. Compared to other thermal cutting methods, it offers advantages such as high cutting quality, high efficiency, and high speed. However, with continuous technological advancements, the requirements for cutting quality of some high-precision materials are becoming increasingly stringent, with surface roughness being a crucial evaluation indicator. In laser cutting, surface roughness is influenced by multiple process parameters, resulting in complex relationships. Optimizing the process through numerous repetitive basic experiments can easily lead to resource waste.

[0003] With the development of information technology and artificial intelligence, simulating the results of various processing technologies using algorithms and then recommending processing parameters based on the results is considered an effective way to save resources and improve production efficiency. Currently, metaheuristic algorithms such as teaching and learning algorithms and particle swarm optimization play an important role in computer simulations. How to improve the search capabilities of these algorithms, prevent them from getting trapped in local optima, and complete the computation of multi-objective problems is a topic worthy of serious research. Furthermore, although algorithm simulation does not require actual experiments, it does require a mathematical description of the experimental laws to complete the simulation. However, many experimental laws often cannot be described by specific mathematical expressions; therefore, a single algorithm cannot simulate and recommend processing technologies. Summary of the Invention

[0004] In view of the shortcomings of the prior art, the purpose of this invention is to provide a laser cutting process design method and system, which aims to solve the problem that the prior art cannot simulate and recommend the laser cutting process, and therefore cannot effectively design the laser cutting process.

[0005] To achieve the above objectives, in a first aspect, the present invention provides a laser cutting process design method, comprising the following steps:

[0006] A trained neural network model is determined; the neural network model is used to predict the roughness of the upper and lower surfaces of the product at the cut point based on the laser cutting process parameters, and includes: a BP hidden layer and an ELM hidden layer; the BP hidden layer is used to perform nonlinear mapping on the information related to the laser cutting process parameters to extract effective information; the ELM hidden layer is used to extract information related to the roughness of the upper and lower surfaces based on the effective information, so as to predict the roughness of the upper and lower surfaces of the product at the cut point.

[0007] The target roughness of the laser-cut product is determined, and multiple sets of process parameters are initially designed based on the target roughness using a teaching and learning algorithm; the target roughness includes the target roughness of the upper surface and the target roughness of the lower surface.

[0008] The preliminary design of each set of process parameters is input into the trained neural network model to predict the roughness of the upper and lower surfaces of the product cut according to each set of process parameters.

[0009] Combining the target roughness and the roughness predicted by each set of process parameters, the fuzzy optimization method is used to select an optimal set of process parameters from the multiple sets of process parameters in the preliminary design, which is used as the optimal objective function of the teaching and learning algorithm. Based on the teaching and learning algorithm, the corresponding process parameters are designed again for the target roughness. The above process of roughness prediction, objective function selection and redesign of process parameters is repeated until the predicted roughness of the upper and lower surfaces reaches the target roughness.

[0010] In an optional example, the process of roughness prediction, objective function selection, and redesign of process parameters is repeated until the predicted roughness of the upper and lower surfaces reaches the target roughness, specifically:

[0011] The current design uses multiple sets of process parameters as input to the neural network model to predict the upper and lower surface roughness of the corresponding products.

[0012] The predicted upper and lower surface roughnesses are subtracted from the target upper and lower surface roughnesses, respectively, to obtain two objective functions for each set of process parameters. Multiple sets of process parameters are divided into two subsets, in which the objective function value of the upper surface roughness is minimized in one subset and the objective function value of the lower surface roughness is minimized in the other subset.

[0013] When both objective functions are not zero, the optimal set of process parameters is selected from multiple sets of process parameters using the fuzzy optimization method, with the set of process parameters having the smallest weight value of the objective function.

[0014] The optimal set of process parameters is used as the teacher, and the other sets of process parameters in the multiple sets of process parameters are used as the students. The teaching and learning algorithm is used and the greedy mechanism is used to update the values ​​of each student again, which are used as the new design process parameters for each student. After designing process parameters for each student again, a set of process parameters that belong to the same subset as the newly designed process parameters are randomly selected as learning partners for learning. Based on the greedy mechanism, it is decided whether to design the corresponding process parameters again to replace the newly designed process parameters.

[0015] The process parameters of the newly designed process are learned together with the historical values ​​of the process parameters of the newly designed process. A greedy mechanism is used to decide whether to redesign the corresponding process parameters to replace the process parameters of the newly designed process. The historical values ​​of the process parameters of the newly designed process are updated by random numbers to improve the ability of the teaching and learning algorithm to search for the optimal solution and avoid getting trapped in local optima.

[0016] The process parameters of the newly designed system, updated in the last update, are used as input to the neural network model. The process of roughness prediction, objective function selection, and new design process parameters is executed again until the value of either objective function is 0.

[0017] In an optional example, if the loop has reached its maximum number of iterations but the predicted roughness of the upper and lower surfaces has not yet reached the target roughness, then all solutions in the current global non-dominated solution set are output.

[0018] In an optional example, the optimal set of process parameters is selected from multiple sets of process parameters using the fuzzy optimization method, specifically:

[0019] According to the Pareto rule, a globally non-dominated solution set is calculated from multiple sets of process parameters; the globally non-dominated solution set includes N sets of process parameters; N is an integer greater than 1.

[0020] The N sets of process parameters are divided into two subsets; the objective function value of the upper surface roughness is minimized in one subset, and the objective function value of the lower surface roughness is minimized in the other subset.

[0021] The total weight of each group of process parameters in the first and second subsets is calculated separately. The total weight is the sum of the weight of the upper surface roughness objective function and the weight of the lower surface roughness objective function. The weight of the upper surface roughness objective function of a group of process parameters is obtained by subtracting the minimum upper surface roughness objective function value in its subset from the upper surface roughness objective function value of that subset, and then dividing by the difference between the maximum and minimum upper surface roughness objective function values ​​in its subset. Similarly, the weight of the lower surface roughness objective function of a group of process parameters is obtained by subtracting the minimum lower surface roughness objective function value in its subset from the lower surface roughness objective function value of that subset, and then dividing by the difference between the maximum and minimum lower surface roughness objective function values ​​in its subset.

[0022] The group of process parameters with the smallest total weight among multiple groups of process parameters is the optimal group of process parameters.

[0023] If the value with the smallest total proportion among multiple sets of process parameters corresponds to more than one set of process parameters, then one set is randomly selected from the multiple sets of process parameters as the optimal set.

[0024] In an optional example, the neural network model further includes an input layer and an output layer;

[0025] The input layer is used to receive laser cutting process parameters and convert them into relevant information;

[0026] The output layer is used to predict the roughness of the upper and lower surfaces based on the information related to the roughness of the upper and lower surfaces, and output the predicted roughness of the upper and lower surfaces. The output layer has two nodes, which output the roughness of the upper surface and the roughness of the lower surface respectively.

[0027] The input layer is connected to the BP hidden layer, the BP hidden layer is connected to the ELM hidden layer, and the ELM hidden layer is connected to the output layer.

[0028] In an optional example, the training process of the neural network model is as follows:

[0029] Multiple sets of process parameters for laser cutting are designed using the response surface methodology; these process parameters include: laser power, cutting speed, defocusing amount, and gas pressure.

[0030] Based on multiple sets of process parameters, the samples were cut into specimens, and the surface roughness of the upper and lower surfaces of the cut products was measured.

[0031] Multiple sets of process parameters and the corresponding measured surface roughness of the upper and lower surfaces were used as training data.

[0032] The neural network model is trained based on the training data.

[0033] Secondly, the present invention provides a laser cutting process design system, comprising:

[0034] A prediction model determination unit is used to determine a trained neural network model. The neural network model is used to predict the roughness of the upper and lower surfaces at the product cutting point based on laser cutting process parameters. It includes a BP hidden layer and an ELM hidden layer. The BP hidden layer is used to perform nonlinear mapping on information related to laser cutting process parameters to extract effective information. The ELM hidden layer is used to extract information related to the roughness of the upper and lower surfaces based on the effective information, so as to predict the roughness of the upper and lower surfaces at the product cutting point.

[0035] The process parameter design unit is used to determine the target roughness of the laser-cut product and to preliminarily design multiple sets of process parameters based on the target roughness using a teaching and learning algorithm; the target roughness includes the target roughness of the upper surface and the target roughness of the lower surface.

[0036] The roughness prediction unit is used to input the preliminary design of each set of process parameters into the trained neural network model, and predict the roughness of the upper and lower surfaces of the product cut according to each set of process parameters.

[0037] The process parameter optimization unit is used to combine the target roughness and the roughness predicted by each set of process parameters, and use fuzzy optimization method to select an optimal set of process parameters from the multiple sets of process parameters in the preliminary design as the optimal objective function of the teaching and learning algorithm, so that the corresponding process parameters can be designed again for the target roughness based on the teaching and learning algorithm, and the above roughness prediction, objective function selection and redesign of process parameters are executed cyclically until the predicted roughness of the upper and lower surfaces reaches the target roughness.

[0038] In an optional example, the process parameter optimization unit iteratively executes the above-described process of roughness prediction, objective function selection, and redesign of process parameters until the predicted upper and lower surface roughnesses reach the target roughness. Specifically, the following steps are taken: Multiple sets of currently designed process parameters are used as inputs to a neural network model to predict the upper and lower surface roughness of the corresponding product; the predicted upper and lower surface roughnesses are subtracted from the target upper and lower surface roughnesses, respectively, to obtain two objective functions for each set of process parameters; the multiple sets of process parameters are divided into two subsets, where the objective function value of the upper surface roughness is minimized in one subset, and the objective function value of the lower surface roughness is minimized in the other subset; when both objective functions are not zero, the process parameters with the smallest objective function weight value among the multiple sets of process parameters are selected as the optimal set of process parameters using a fuzzy optimization method; the optimal set of process parameters is used as the teacher, and the other sets of process parameters are used as students, employing a teaching and learning approach. The algorithm then uses a greedy mechanism to update the values ​​of each student again, which are used as the new design process parameters for each student. After redesigning the process parameters for each student, it randomly selects a set of process parameters belonging to the same subset as the newly designed process parameters as learning partners for learning, and decides whether to redesign the corresponding process parameters to replace the newly designed process parameters based on the greedy mechanism. The algorithm learns together the last updated newly designed process parameters with the historical values ​​of the newly designed process parameters, and decides whether to redesign the corresponding process parameters to replace the newly designed process parameters based on the greedy mechanism. It also uses random numbers to decide whether to update the historical values ​​of the newly designed process parameters to improve the teaching and learning algorithm's ability to search for optimal solutions and avoid getting trapped in local optima. Finally, it uses the last updated newly designed process parameters as input to the neural network model and executes the above roughness prediction, objective function selection, and new design process parameter process again until the value of either objective function is 0.

[0039] In an optional example, if the process parameter optimization unit performs the maximum number of iterations, but the predicted roughness of the upper and lower surfaces still does not reach the target roughness, then all solutions in the current global non-dominated solution set are output.

[0040] In an optional example, the process parameter optimization unit uses a fuzzy optimization method to select the group of process parameters with the smallest objective function weight value from multiple groups of process parameters as the optimal group of process parameters. Specifically, according to the Pareto rule, a global non-dominated solution set is calculated from multiple groups of process parameters; the global non-dominated solution set includes N groups of process parameters; N is an integer greater than 1; the N groups of process parameters are divided into two subsets; the objective function value of the upper surface roughness is minimized in one subset, and the objective function value of the lower surface roughness is minimized in the other subset; the total weight of each group of process parameters in the first subset and the second subset is calculated respectively; the total weight is the sum of the weight of the objective function of the upper surface roughness and the weight of the objective function of the lower surface roughness; the upper surface roughness of a group of process parameters... The weight of the surface roughness objective function of a set of process parameters is obtained by subtracting the minimum value of the surface roughness objective function in its subset from the surface roughness objective function value of the set of process parameters, and then dividing by the difference between the maximum and minimum values ​​of the surface roughness objective function in its subset. Similarly, the weight of the surface roughness objective function of a set of process parameters is obtained by subtracting the minimum value of the surface roughness objective function in its subset from the surface roughness objective function value of the set of process parameters, and then dividing by the difference between the maximum and minimum values ​​of the surface roughness objective function in its subset. The set of process parameters with the smallest total weight among multiple sets of process parameters is selected as the optimal set of process parameters. Furthermore, if the minimum total weight among multiple sets of process parameters corresponds to more than one set of process parameters, then one set is randomly selected from these multiple sets of process parameters as the optimal set.

[0041] In an optional example, the neural network model further includes an input layer and an output layer;

[0042] The input layer is used to receive laser cutting process parameters and convert them into relevant information;

[0043] The output layer is used to predict the roughness of the upper and lower surfaces based on the information related to the roughness of the upper and lower surfaces, and output the predicted roughness of the upper and lower surfaces. The output layer has two nodes, which output the roughness of the upper surface and the roughness of the lower surface respectively.

[0044] The input layer is connected to the BP hidden layer, the BP hidden layer is connected to the ELM hidden layer, and the ELM hidden layer is connected to the output layer.

[0045] In an optional example, the system further includes: a model training unit;

[0046] The model training unit is used to design multiple sets of process parameters for laser cutting using the response surface design method. The process parameters include: laser power, cutting speed, defocusing amount and gas pressure. Based on the multiple sets of process parameters, the sample is cut into specimens and the roughness of the upper and lower surfaces of the cut product is measured. The multiple sets of process parameters and the corresponding measured upper and lower surface roughness are used as training data. The neural network model is trained based on the training data.

[0047] In summary, the technical solutions conceived by this invention have the following beneficial effects compared with the prior art:

[0048] This invention provides a laser cutting process design method and system, offering a method for optimizing the surface roughness of laser-cut products under the influence of various process parameters. The method first improves the BP neural network by introducing an ELM hidden layer, enhancing the convergence speed and prediction accuracy. It also improves the teaching and learning algorithm commonly used for solving single-objective functions by introducing a fuzzy optimization method for the teaching phase. This method categorizes all individuals according to their advantageous objectives, calculates the total weight of each individual within its category, and uses this total weight as the criterion for selecting the "teacher," thus completing the selection for multi-objective problems while preserving the strengths of each individual. Furthermore, a historical learning mechanism is added after the learning phase to improve the algorithm's search ability and prevent premature convergence and getting trapped in local optima. The results obtained from an experimental scheme designed based on the response surface methodology are then used as the training set to train the improved BP neural network. The trained BP neural network provides objective function values ​​for subsequent improved teaching and learning algorithms. Finally, the improved teaching and learning algorithm recommends processes based on the desired target roughness value. This invention improves the quality of laser-cut parts and increases laser cutting production efficiency. Attached Figure Description

[0049] Figure 1 This is a flowchart of the laser cutting process design method provided in the embodiments of the present invention;

[0050] Figure 2 This is a flowchart of the laser cutting process design optimization method provided in the embodiments of the present invention;

[0051] Figure 3 This is a system architecture diagram of laser cutting process design provided in an embodiment of the present invention. Detailed Implementation

[0052] To make the objectives, technical solutions, and advantages of this invention clearer, the invention will be further described in detail below with reference to the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are merely illustrative and not intended to limit the invention.

[0053] Understandably, artificial neural networks consist of an input layer, hidden layers, and an output layer. They are characterized by strong predictive ability and high accuracy, but suffer from poor interpretability, long training time, high requirements for training samples, and unidirectional prediction. Therefore, training neural networks based on actual experimental results, improving their structure to enhance prediction accuracy and convergence speed, increasing the search capability of metaheuristic search algorithms, and then combining both to jointly optimize and recommend process parameters holds great research potential.

[0054] Figure 1 This is a flowchart of the laser cutting process design method provided in an embodiment of the present invention; as shown below. Figure 1 As shown, it includes the following steps:

[0055] S101, Determine the trained neural network model; the neural network model is used to predict the roughness of the upper and lower surfaces of the product at the cut point based on the laser cutting process parameters, and includes: a BP hidden layer and an ELM hidden layer; the BP hidden layer is used to perform nonlinear mapping on the information related to the laser cutting process parameters to extract effective information; the ELM hidden layer is used to extract information related to the roughness of the upper and lower surfaces based on the effective information, so as to predict the roughness of the upper and lower surfaces of the product at the cut point.

[0056] S102, determine the target roughness of the laser-cut product, and use the teaching and learning algorithm to preliminarily design multiple sets of process parameters based on the target roughness; the target roughness includes the target roughness of the upper surface and the target roughness of the lower surface;

[0057] S103, input the preliminary design of each set of process parameters into the trained neural network model, and predict the roughness of the upper and lower surfaces of the product cut according to each set of process parameters.

[0058] S104, combining the target roughness and the roughness predicted by each set of process parameters, the optimal set of process parameters is selected from the multiple sets of process parameters in the preliminary design using the fuzzy optimization method, and used as the optimal objective function of the teaching and learning algorithm, so that the corresponding process parameters can be designed again for the target roughness based on the teaching and learning algorithm, and the above process of roughness prediction, objective function selection and redesign of process parameters is executed cyclically until the predicted roughness of the upper and lower surfaces reaches the target roughness.

[0059] In an optional example, the process of roughness prediction, objective function selection, and redesign of process parameters is repeated until the predicted roughness of the upper and lower surfaces reaches the target roughness, specifically:

[0060] The current design uses multiple sets of process parameters as input to the neural network model to predict the upper and lower surface roughness of the corresponding products.

[0061] The predicted upper and lower surface roughnesses are subtracted from the target upper and lower surface roughnesses, respectively, to obtain two objective functions for each set of process parameters. Multiple sets of process parameters are divided into two subsets, in which the objective function value of the upper surface roughness is minimized in one subset and the objective function value of the lower surface roughness is minimized in the other subset.

[0062] When both objective functions are not zero, the optimal set of process parameters is selected from multiple sets of process parameters using the fuzzy optimization method, with the set of process parameters having the smallest weight value of the objective function.

[0063] The optimal set of process parameters is used as the teacher, and the other sets of process parameters in the multiple sets of process parameters are used as the students. The teaching and learning algorithm is used and the greedy mechanism is used to update the values ​​of each student again, which are used as the new design process parameters for each student. After designing process parameters for each student again, a set of process parameters that belong to the same subset as the newly designed process parameters are randomly selected as learning partners for learning. Based on the greedy mechanism, it is decided whether to design the corresponding process parameters again to replace the newly designed process parameters.

[0064] The process parameters of the newly designed process are learned together with the historical values ​​of the process parameters of the newly designed process. A greedy mechanism is used to decide whether to redesign the corresponding process parameters to replace the process parameters of the newly designed process. The historical values ​​of the process parameters of the newly designed process are updated by random numbers to improve the ability of the teaching and learning algorithm to search for the optimal solution and avoid getting trapped in local optima.

[0065] The process parameters of the newly designed system, updated in the last update, are used as input to the neural network model. The process of roughness prediction, objective function selection, and new design process parameters is executed again until the value of either objective function is 0.

[0066] In an optional example, if the loop has reached its maximum number of iterations but the predicted roughness of the upper and lower surfaces has not yet reached the target roughness, then all solutions in the current global non-dominated solution set are output.

[0067] In an optional example, the optimal set of process parameters is selected from multiple sets of process parameters using the fuzzy optimization method, specifically:

[0068] According to the Pareto rule, a globally non-dominated solution set is calculated from multiple sets of process parameters; the globally non-dominated solution set includes N sets of process parameters; N is an integer greater than 1.

[0069] The N sets of process parameters are divided into two subsets; the objective function value of the upper surface roughness is minimized in one subset, and the objective function value of the lower surface roughness is minimized in the other subset.

[0070] The total weight of each group of process parameters in the first and second subsets is calculated separately. The total weight is the sum of the weight of the upper surface roughness objective function and the weight of the lower surface roughness objective function. The weight of the upper surface roughness objective function of a group of process parameters is obtained by subtracting the minimum upper surface roughness objective function value in its subset from the upper surface roughness objective function value of that subset, and then dividing by the difference between the maximum and minimum upper surface roughness objective function values ​​in its subset. Similarly, the weight of the lower surface roughness objective function of a group of process parameters is obtained by subtracting the minimum lower surface roughness objective function value in its subset from the lower surface roughness objective function value of that subset, and then dividing by the difference between the maximum and minimum lower surface roughness objective function values ​​in its subset.

[0071] The group of process parameters with the smallest total weight among multiple groups of process parameters is the optimal group of process parameters.

[0072] If the value with the smallest total proportion among multiple sets of process parameters corresponds to more than one set of process parameters, then one set is randomly selected from the multiple sets of process parameters as the optimal set.

[0073] In an optional example, the neural network model further includes an input layer and an output layer;

[0074] The input layer is used to receive laser cutting process parameters and convert them into relevant information;

[0075] The output layer is used to predict the roughness of the upper and lower surfaces based on the information related to the roughness of the upper and lower surfaces, and output the predicted roughness of the upper and lower surfaces. The output layer has two nodes, which output the roughness of the upper surface and the roughness of the lower surface respectively.

[0076] The input layer is connected to the BP hidden layer, the BP hidden layer is connected to the ELM hidden layer, and the ELM hidden layer is connected to the output layer.

[0077] In an optional example, the training process of the neural network model is as follows:

[0078] Multiple sets of process parameters for laser cutting are designed using the response surface methodology; these process parameters include: laser power, cutting speed, defocusing amount, and gas pressure.

[0079] Based on multiple sets of process parameters, the samples were cut into specimens, and the surface roughness of the upper and lower surfaces of the cut products was measured.

[0080] Multiple sets of process parameters and the corresponding measured surface roughness of the upper and lower surfaces were used as training data.

[0081] The neural network model is trained based on the training data.

[0082] Figure 2 This is a flowchart of the laser cutting process design optimization method provided in the embodiments of the present invention; as follows: Figure 2 As shown, it includes the following steps:

[0083] S1: Design of experimental process parameters for laser cutting using the response surface methodology;

[0084] S2: The process performance indicators of the experimental group are obtained by characterizing the samples through performance testing experiments;

[0085] S3: Based on the data in S2, take the experimental process parameters as input and the process performance index as output to train the improved neural network model.

[0086] S4: By improving the teaching and learning algorithm, the process parameters required to achieve the target process performance value are solved, and the process recommendation for laser cutting is realized, thereby obtaining the optimal process.

[0087] The response surface design used in S1 is a central composite experimental design. The designed process parameters include laser power, cutting speed, defocusing amount and gas pressure; finally, a total of 31 experimental groups with four factors and five levels were obtained.

[0088] Furthermore, the cutting height in S1 is 1mm-3mm, and the other process parameters such as integral frequency, cutting process temperature, grinding head speed and moving speed are the machine default values;

[0089] In one example, the process parameters designed in step S1 are shown in Table 1:

[0090] Table 1 Selection of Process Parameter Levels

[0091]

[0092] Furthermore, the process performance index mentioned in S2 refers to the roughness of the upper and lower surfaces of the cut part of the sample, which is measured by a roughness tester; the specific sample process parameters are shown in Table 2:

[0093] Table 2 Sample Process Parameter Design Table

[0094]

[0095] The neural network used in S3 is an improved BP-ELM neural network model, which consists of an input layer, a BP hidden layer, an ELM hidden layer, and an output layer. The process parameters, including laser power, scanning speed, scanning spacing, and powder layer thickness, are the required input data, while the output data is the roughness of the upper and lower surfaces at the cut. The final structure of the neural network is determined to be [4-XY-2], representing that the neural network model has 4 nodes in the input layer (i.e., 4 process parameters), X BP hidden layer nodes, Y ELM hidden layer nodes, and 2 output nodes (i.e., upper and lower surface roughness). X is the number of BP hidden layer nodes determined based on the characteristics of the feature index. First, a BP neural network with a three-layer structure [4-X-1] is constructed. The value of X is determined to be 4-16 according to an empirical formula. Then, BP neural networks with different X values ​​are trained using the PimaDiabetes dataset, and the accuracy of these BP neural networks is compared. Finally, the value of X is determined to be 6. Then, the input layer and hidden layer of the BP neural network are frozen, and an ELM hidden layer is connected to the end. Using the output of the hidden layer of the BP neural network as input, the output of a node in the hidden layer of the ELM is calculated using the following formula:

[0096]

[0097] in, It is the output of the i-th node in the hidden layer, a j and b j It can be randomly generated, j = 1, 2…k (k is the number of hidden layer nodes in ELM).

[0098] The hidden layer output weight matrix of ELM is calculated using the following formula:

[0099]

[0100] Where H is the output matrix of the hidden layer node in ELM, and Y is the desired output value.

[0101] Finally, the data obtained from the experiment designed using the response surface methodology are used to train the entire neural network model, resulting in the final BP-ELM neural network model.

[0102] Furthermore, the normalization and transfer function of the constructed BP neural network model are selected as hyperbolic tangent functions with a range of [-1, 1]. Random errors are eliminated by averaging the model indicators after multiple training cycles (e.g., 5 times), so that the model indicators are close to the model performance at optimal convergence. The expected error is set to 0.0001-0.001, and the momentum coefficient is set to 0.001-0.01. The learning algorithm is selected from one of the following: the fastest gradient descent optimization algorithm, the Levenberg-Marquardt (LM) algorithm based on nonlinear least squares, and the Bayesian regularization (BR) algorithm.

[0103] In step S4, the teaching and learning algorithm generates target process parameters and inputs them into the neural network model to obtain corresponding performance data prediction values. Specifically, the prediction values ​​of the neural network model trained in step S3 are used as the objective function for improving the teaching and learning algorithm. Then, iterative optimization is performed continuously, generating new students to form a new population. When the function value of the obtained student reaches the required process target value, the optimization ends, and the desired result is output. If, after a set number of iterations, no student with the function value of the target process value is found, the student whose result is closest is taken as the output result. The specific process is as follows:

[0104] S41. Based on the actual situation, set the boundaries of laser power to [500, 6000], cutting speed to [1, 30], defocusing amount to [-5, 0], and air pressure to [5, 35], and then randomly generate an initial population. Each population includes multiple sets of process parameters, and the current population is used as the initial value for the historical populations.

[0105] S42. Using the current population as input to the BP-ELM neural network, the difference between the calculated upper surface roughness y1 and lower surface roughness y2 and the required y1 and y2 is used as the individual's objective function. First, according to the Pareto rule, the global non-dominated solution is calculated as the teacher. Then, the evaluation index of each individual on the two objective functions relative to the minimum value of the objective function is used, that is, the evaluation index of particle i on the objective function j relative to the optimal value. Evaluation indicators The larger the value, the more advantageous particle i is in objective function j. Therefore, feasible solutions with the largest evaluation index for the same objective are grouped into the same subset d. i (i=1,2). The calculation method is as follows:

[0106]

[0107] Where F j Let j be the function value of the individual on the j-th objective function, where j = 1, 2

[0108] S43. Since there may be multiple non-dominated solutions, but an individual can only choose one teacher for instruction, a fuzzy optimization method is used. This involves calculating the difference between each individual's objective function in the non-dominated solution set and the minimum value of the corresponding objective function in the solution set, and then summing these sums relative to the difference between the maximum and minimum values ​​of that objective function in the solution set (if an individual's objective function value equals the maximum value of that objective function in the solution set, then the sum is 1). This summates the total weight value B, identifying the solution that shows a more balanced performance across the two objective functions. The individual with the smallest total weight value B is then selected as the teacher. If multiple individuals still have the same total weight value, a random selection is made. After instruction is completed, a greedy mechanism is used to determine whether to update the individual.

[0109] S44. After completing the teaching, each individual will be in their respective subset d. i In the process, a learning partner is randomly selected for learning, and a greedy mechanism is used to decide whether to update the individual.

[0110] S45. After completing the learning phase, each individual will continue to learn from its historical values ​​and use a greedy mechanism to decide whether to update the individual.

[0111] S46. Randomly update the historical population, that is, by generating random numbers, decide whether to replace the historical value of each individual with the current value.

[0112] S47. Determine if there are any individuals in the current population whose objective function values ​​are all 0. If so, end the algorithm and output the corresponding individual values, which are the recommended process parameters. If not, determine if the maximum number of iterations has been reached. If the maximum number of iterations has been reached, output all solutions in the current Pareto solution set. If not, start the loop from S42.

[0113] Figure 3 This is a system architecture diagram of the laser cutting process design provided in an embodiment of the present invention, such as... Figure 3 As shown, it includes:

[0114] The prediction model determination unit 310 is used to determine the trained neural network model. The neural network model is used to predict the roughness of the upper and lower surfaces of the product at the cut point based on the laser cutting process parameters. It includes a BP hidden layer and an ELM hidden layer. The BP hidden layer is used to perform nonlinear mapping on the information related to the laser cutting process parameters to extract effective information. The ELM hidden layer is used to extract information related to the roughness of the upper and lower surfaces based on the effective information so as to predict the roughness of the upper and lower surfaces of the product at the cut point.

[0115] The process parameter design unit 320 is used to determine the target roughness of the laser-cut product and to preliminarily design multiple sets of process parameters based on the target roughness using a teaching and learning algorithm; the target roughness includes the target roughness of the upper surface and the target roughness of the lower surface.

[0116] Roughness prediction unit 330 is used to input the preliminary design of each set of process parameters into the trained neural network model, and predict the roughness of the upper and lower surfaces of the product cut according to each set of process parameters.

[0117] The process parameter optimization unit 340 is used to combine the target roughness and the roughness predicted by each set of process parameters, and use fuzzy optimization method to select an optimal set of process parameters from the multiple sets of process parameters in the preliminary design as the optimal objective function of the teaching and learning algorithm, so that the corresponding process parameters can be designed again for the target roughness based on the teaching and learning algorithm, and the above roughness prediction, objective function selection and redesign of process parameters are executed cyclically until the predicted roughness of the upper and lower surfaces reaches the target roughness.

[0118] The model training unit 350 is used to design multiple sets of process parameters for laser cutting using the response surface design method. The process parameters include: laser power, cutting speed, defocusing amount and gas pressure. Based on the multiple sets of process parameters, the sample is cut into specimens and the roughness of the upper and lower surfaces of the cut product is measured. The multiple sets of process parameters and the corresponding measured upper and lower surface roughness are used as training data. The neural network model is trained based on the training data.

[0119] It should be noted that, Figure 3 For detailed implementation of the functions of each unit, please refer to the description in the aforementioned method embodiments, which will not be repeated here.

[0120] Those skilled in the art will readily understand that the above description is merely a preferred embodiment of the present invention and is not intended to limit the present invention. Any modifications, equivalent substitutions, and improvements made within the spirit and principles of the present invention should be included within the scope of protection of the present invention.

Claims

1. A laser cutting process design method, characterized in that, Includes the following steps: S101. Determine the trained neural network model; The neural network model is used to predict the roughness of the upper and lower surfaces of the product at the cut point based on laser cutting process parameters. It includes a BP hidden layer and an ELM hidden layer. The BP hidden layer is used to perform nonlinear mapping on information related to laser cutting process parameters to extract effective information. The ELM hidden layer is used to extract information related to the roughness of the upper and lower surfaces based on the effective information in order to predict the roughness of the upper and lower surfaces of the product at the cut point. S102. Determine the target roughness of the laser-cut product, and use the teaching and learning algorithm to preliminarily design multiple sets of process parameters based on the target roughness; the target roughness includes the target roughness of the upper surface and the target roughness of the lower surface; S103. Input the preliminary design of each set of process parameters into the trained neural network model to predict the surface roughness of the upper and lower surfaces of the product cut according to each set of process parameters. S104. Combining the target roughness and the roughness predicted by each set of process parameters, the optimal set of process parameters is selected from the multiple sets of process parameters in the preliminary design using the fuzzy optimization method. This set is used as the optimal objective function of the teaching and learning algorithm. Based on the teaching and learning algorithm, the corresponding process parameters are designed again for the target roughness. The process of roughness prediction, objective function selection, and redesign of process parameters is repeated until the predicted roughness of the upper and lower surfaces reaches the target roughness.

2. The method according to claim 1, characterized in that, The process of repeatedly performing roughness prediction, objective function selection, and redesigning process parameters continues until the predicted roughness of the upper and lower surfaces reaches the target roughness. The specific steps are as follows: The current set of process parameters are used as inputs to the neural network model to predict the upper and lower surface roughness of the corresponding products. The predicted upper and lower surface roughnesses are subtracted from the target upper and lower surface roughnesses, respectively, to obtain two objective functions for each set of process parameters in the current design. The multiple sets of process parameters in the current design are divided into two subsets, in which the objective function value of the upper surface roughness is minimized in one subset and the objective function value of the lower surface roughness is minimized in the other subset. When both objective functions are not 0, the fuzzy optimization method is used to select the set of process parameters with the smallest weight of the objective function from the multiple sets of process parameters currently designed as the optimal set of process parameters. The optimal set of process parameters is taken as the teacher, and the other sets of process parameters in the current design are taken as the students. The teaching and learning algorithm is used and the greedy mechanism is used to update the values ​​of each student again, which are taken as the new design process parameters for each student. After designing process parameters for each student again, a set of process parameters that belong to the same subset as the new design process parameters are randomly selected as learning partners for learning. Based on the greedy mechanism, it is decided whether to design the corresponding process parameters again to replace the new design process parameters. The algorithm learns from the last updated process parameters along with their historical values ​​and decides whether to redesign the corresponding process parameters to replace the last updated process parameters based on a greedy mechanism; and it also decides whether to update the historical values ​​of the newly designed process parameters using random numbers, in order to improve the teaching and learning algorithm's ability to search for optimal solutions and avoid getting trapped in local optima. The last updated process parameters are used as input to the neural network model, and the above-mentioned roughness prediction, objective function selection, and process parameter redesign process are executed again until the value of either of the two objective functions is 0.

3. The method according to claim 2, characterized in that, If the loop reaches its maximum number of iterations, but the predicted roughness of the upper and lower surfaces still fails to reach the target roughness, then all solutions in the current global non-dominated solution set are output.

4. The method according to claim 2, characterized in that, The above-mentioned method of using fuzzy optimization to select the set of process parameters with the smallest objective function weight from multiple sets of process parameters in the current design as the optimal set of process parameters is as follows: According to the Pareto rule, a global non-dominated solution set is calculated from multiple sets of process parameters in the current design; the global non-dominated solution set includes N sets of process parameters; N is an integer greater than 1. The N sets of process parameters are divided into two subsets; wherein, the objective function value of the upper surface roughness is minimized in one subset, and the objective function value of the lower surface roughness is minimized in the other subset; The total weight of each group of process parameters in the first and second subsets is calculated separately. The total weight is the sum of the weight of the upper surface roughness objective function and the weight of the lower surface roughness objective function. Specifically, the weight of the upper surface roughness objective function of a group of process parameters is obtained by subtracting the minimum upper surface roughness objective function value in its subset from the upper surface roughness objective function value of that subset, and then dividing by the difference between the maximum and minimum upper surface roughness objective function values ​​in its subset. Similarly, the weight of the lower surface roughness objective function of a group of process parameters is obtained by subtracting the minimum lower surface roughness objective function value in its subset from the lower surface roughness objective function value of that subset, and then dividing by the difference between the maximum and minimum lower surface roughness objective function values ​​in its subset. The process parameters with the smallest total weight among the multiple sets of process parameters in the current design are taken as the optimal set of process parameters; if the value with the smallest total weight among the multiple sets of process parameters in the current design corresponds to more than one set of process parameters, then one set is randomly selected from the multiple sets of process parameters as the optimal set of process parameters.

5. The method according to any one of claims 1 to 4, characterized in that, The neural network model further includes: an input layer and an output layer; The input layer is used to receive laser cutting process parameters and convert them into relevant information; The output layer is used to predict the roughness of the upper and lower surfaces based on the information related to the roughness of the upper and lower surfaces, and output the predicted roughness of the upper and lower surfaces. The output layer has two nodes, which output the roughness of the upper surface and the roughness of the lower surface respectively. The input layer is connected to the BP hidden layer, the BP hidden layer is connected to the ELM hidden layer, and the ELM hidden layer is connected to the output layer.

Images