A method, medium and system for predicting edge roughness of steel structure laser cutting
Through the combination of Bayesian network and multi-branch neural network, the roughness of the laser cutting edge of steel structures is analyzed and predicted, and the problem of insufficient comprehensive parameter correlation analysis is solved, achieving in-depth understanding of the impact of various process parameters and efficient prediction.
Patent Information
- Application Number
- CN202411154872.3
- Authority / Receiving Office
- CN · China
- Patent Type
- Patents(China)
- Current Assignee / Owner
- Filing Date
- 2024-08-22
- Publication Date
- 2024-11-01
- Estimated Expiration
- 2044-08-22
AI Technical Summary
In the prediction of laser cutting edge roughness of steel structures, the parameter correlation analysis is not comprehensive enough, and it is difficult to fully grasp the degree of influence of each process parameter on roughness.
The Bayesian network model is used to calculate the correlation index of each parameter, establish a weight vector, set up the sensitivity matrix, and construct a roughness equation system, combine the multi-branch neural network model for feature extraction and fusion to build an accurate roughness prediction model.
By comprehensively analyzing the impact of each process parameter on roughness, the prediction accuracy and applicability and robustness of the model are improved, and efficient prediction under new process conditions are achieved.
Smart Images

Figure CN118690177B_ABST
Abstract
Description
Technical Field
[0001] The present invention belongs to the technical field of steel structure laser cutting, and in particular, relates to a method, medium and system for predicting edge roughness of steel structure laser cutting. Background Art
[0002] Laser cutting is a high-precision and high-efficiency metal processing method, which is widely used in automobile manufacturing, aerospace, light industry manufacturing and other industries. In the laser cutting process, the surface roughness of the cutting edge is an important quality indicator, which directly affects the performance of the workpiece and the subsequent processing process. For example, too high roughness will reduce the mechanical strength and increase the wear of parts; too low roughness requires more subsequent polishing, which increases the production cost. Therefore, how to accurately predict and control the surface roughness in the laser cutting process has always been a hot issue in the industry.
[0003] The existing roughness prediction methods mainly fall into the following categories:
[0004] 1. Empirical formula method: Through a large number of experimental accumulations, an empirical formula model between steel properties, cutting parameters and roughness is established. This method is simple and intuitive, but its scope of application is limited and it is difficult to extend to new process conditions.
[0005] 2. Data-driven method: Use machine learning techniques, such as artificial neural networks and support vector machines, to directly learn the roughness prediction model from historical data. This method can cover a wider parameter space, but requires a large amount of high-quality training data and is highly dependent on data.
[0006] 3. Physical model method: Based on the physical process of laser cutting, theoretical models such as cutting temperature field and molten pool dynamics are established to derive the roughness prediction formula. This method has strong generalizability, but requires a lot of prior knowledge and the model complexity is high.
[0007] The above methods have their own advantages and disadvantages, and in practical applications they often need to be selected and combined according to specific circumstances. But in general, the existing technology in the prediction of edge roughness of steel structure laser cutting has the technical problem that the parameter correlation analysis is not comprehensive enough and it is difficult to fully grasp the influence of various process parameters on roughness. Summary of the invention
[0008] In view of this, the present invention provides a method, medium and system for predicting the edge roughness of laser cutting of steel structures, which can solve the technical problems in the prior art of predicting the edge roughness of laser cutting of steel structures, that is, the parameter correlation analysis is not comprehensive enough and it is difficult to fully grasp the influence of each process parameter on the roughness.
[0009] The present invention is achieved in that:
[0010] A first aspect of the present invention provides a method for predicting edge roughness of steel structure laser cutting, which comprises the following steps:
[0011] S10, obtaining multiple sets of historical steel structure laser cutting data sets, recorded as historical data sets, including steel structure parameters, cutting parameters and roughness parameters;
[0012] S20, calculating the correlation index of each steel structure parameter, each cutting parameter and each roughness parameter using a Bayesian network model according to the historical data set;
[0013] S30, for each steel structure parameter, establishing a weight vector for representing the correlation index between the steel structure parameter and each roughness parameter; for each cutting parameter, establishing a weight vector for representing the correlation index between the cutting parameter and each roughness parameter;
[0014] S40, calculating a sensitivity matrix according to the correlation index, and setting power or interaction terms and nonlinear terms according to the sensitivity matrix, and establishing a roughness equation group considering steel structure parameters, cutting parameters and error terms, including an average roughness equation, a maximum roughness equation, a waviness equation, a kurtosis equation, a skewness equation, a variance equation, a central moment equation and a comprehensive roughness equation;
[0015] S50, fitting the roughness equation group using the historical data set to obtain a fitting equation group;
[0016] S60, setting different steel structure parameters and cutting parameters, using the fitting equation group to calculate corresponding roughness parameters, and establishing an extended data set including steel structure parameters, cutting parameters and roughness parameters;
[0017] S70, after weighted normalization processing, each set of data in the extended data set and the historical data set are used as a training set and a validation set respectively, and a multi-branch neural network model is trained to obtain a roughness prediction model; wherein the input of the training is the steel structure parameter and the cutting parameter, and the output of the training is the roughness parameter;
[0018] S80, obtaining steel structure parameters and cutting parameters of the current steel structure to be cut, and inputting them into the roughness prediction model to obtain the predicted edge roughness of the steel structure laser cut.
[0019] The steel structure parameters include steel thickness, strength grade, heat treatment status, surface condition, residual stress, and grain size.
[0020] The cutting parameters include laser power, cutting speed, focal position, auxiliary gas type, auxiliary gas pressure, nozzle diameter, pulse frequency, and duty cycle.
[0021] The roughness is represented by multiple texture features of multi-angle images of the cutting surface, including average roughness, maximum roughness, waviness, kurtosis, skewness and central moment; the roughness is a multidimensional vector.
[0022] The weights used in the weighted normalization process are weight vectors corresponding to each steel structure parameter or each cutting parameter.
[0023] The multi-branch neural network model includes a steel structure feature extractor, a cutting feature extractor, a feature fusion network and a roughness prediction network.
[0024] Furthermore, the steel structure feature extractor performs feature extraction according to the input steel structure parameters and their corresponding weight vectors to obtain steel structure features;
[0025] The cutting feature extractor performs feature extraction according to the input cutting parameters and their corresponding weight vectors to obtain cutting features;
[0026] The feature fusion network is used to fuse the steel structure features and the cutting features. Its input is the steel structure features and the cutting features. It uses the attention mechanism to perform feature fusion through a multi-layer fully connected network and outputs a fused feature vector.
[0027] The roughness prediction network is used to predict the roughness according to the fusion features, and its input is the output of the feature fusion network. The prediction is performed through a multi-layer fully connected network to output the roughness.
[0028] The structure of the feature fusion network is specifically:
[0029] Input layer: receiving steel structure features and cutting features;
[0030] Attention layer: Use self-attention mechanism to weight features;
[0031] Fully connected layer 1: 128 neurons, ReLU activation function;
[0032] Fully connected layer 2: 64 neurons, ReLU activation function;
[0033] Output layer: 32 neurons, output fusion feature vector;
[0034] The structure of the roughness prediction network is specifically:
[0035] Input layer: receives fused feature vector;
[0036] Fully connected layer 1: 64 neurons, ReLU activation function;
[0037] Fully connected layer 2: 32 neurons, ReLU activation function;
[0038] Fully connected layer 3: 16 neurons, ReLU activation function;
[0039] Output layer: output multidimensional roughness vector;
[0040] Residual connection: connect the input layer directly to the output layer.
[0041] Furthermore, the roughness prediction network adopts a residual connection structure to improve prediction accuracy.
[0042] Wherein, the step S10 specifically includes: collecting a large amount of steel structure laser cutting data from relevant experiments or production processes, including steel structure parameters, cutting parameters and roughness parameters; the steel structure parameters include steel thickness, strength grade, heat treatment state, surface state, residual stress and grain size, etc.; the cutting parameters include laser power, cutting speed, focus position, auxiliary gas type and pressure, nozzle diameter, pulse frequency and duty cycle, etc.; the roughness parameters are represented by texture features of multi-angle images, including average roughness, maximum roughness, waviness, kurtosis, skewness and central moment, etc.; each set of data corresponds to a multi-dimensional roughness vector; by comprehensively collecting and organizing these historical data, a foundation is laid for subsequent modeling and training.
[0043] Wherein, the step S20 specifically includes: constructing a Bayesian network structure, taking the steel structure parameters and cutting parameters as parent nodes, and the roughness parameters as child nodes; using a structural learning algorithm (such as the K2 algorithm) to determine the network structure; using the maximum likelihood estimation (MLE) or Bayesian estimation method to learn the conditional probability table (CPT) of each node; for each pair of parameters (Steel structure parameters or cutting parameters) and roughness parameters , calculate their correlation index ,The calculation formula of correlation index is: ,in yes and The mutual information between yes The entropy of these probability values can be obtained from the conditional probability table of the Bayesian network; by calculating the correlation index, the degree of association between each parameter and roughness can be quantified.
[0044] Specifically, the steps to calculate the relevance index using the Bayesian network model are:
[0045] Step 1: Construct a Bayesian network structure: take steel structure parameters and cutting parameters as parent nodes and roughness parameters as child nodes; use a structural learning algorithm (such as K2 algorithm) to determine the network structure;
[0046] Step 2: Parameter learning: Use maximum likelihood estimation (MLE) or Bayesian estimation to learn the conditional probability table (CPT);
[0047] Step 3: Calculate the correlation index:
[0048] For each pair of parameters (Steel structure parameters or cutting parameters) and roughness parameters :
[0049] ;
[0050] in:
[0051] is the correlation index;
[0052] is the mutual information, and the calculation formula is:
[0053] ;
[0054] yes The entropy is calculated as:
[0055] ;
[0056] , , It can be obtained from the CPT of the Bayesian network.
[0057] Wherein, the step S30 specifically includes: for each steel structure parameter , according to its relationship with each roughness parameter The correlation index , construct a weight vector:
[0058]
[0059] Similarly, for each cutting parameter , and also construct a weight vector:
[0060]
[0061] These weight vectors reflect the correlation between each parameter and the roughness feature, and will play a role in subsequent feature extraction.
[0062] Wherein, the step S40 specifically includes: calculating the sensitivity matrix ,in Representation parameters Roughness parameters The sensitivity of is the element of the sensitivity matrix, which can be approximately calculated by the finite difference method; according to the sensitivity matrix S, different powers, interaction terms and nonlinear terms are set:
[0063] According to the sensitivity matrix , setting different powers, interaction terms, and nonlinear terms:
[0064] if (high sensitivity), add quadratic term ;
[0065] if (Medium sensitivity), add interaction term ;
[0066] if (low sensitivity), only keep linear terms .
[0067] in , and is a predefined threshold value that can be adjusted according to actual conditions. Based on this, a set of equations containing multiple roughness characteristic equations is established, including:
[0068] Average roughness equation:
[0069] ;
[0070] Maximum roughness equation:
[0071] ;
[0072] Waviness equation:
[0073] ;
[0074] Kurtosis equation:
[0075] ;
[0076] Skewness equation:
[0077] ;
[0078] Variance equation:
[0079] ;
[0080] Central moment equation:
[0081] ;
[0082] Comprehensive roughness equation:
[0083] ;
[0084] in: Indicates steel structure parameters or cutting parameters; , , , , , , , is the coefficient to be determined; arrive is the error term.
[0085] Wherein, the step S50 specifically includes: using the roughness equation group established in step S40 as the model to be fitted, using the complete historical data set collected in step S10 as a training sample, and using the least squares method or other regression algorithms to optimize the fitting of the coefficients of each equation; through this process, a set of calibrated roughness prediction equations can be obtained, laying the foundation for the subsequent construction of an extended data set and neural network training.
[0086] Wherein, the step S60 specifically includes: setting a series of different combinations of steel structure parameters and cutting parameters, substituting them into the fitting equation group obtained in step S50, and calculating the corresponding roughness parameters; in this way, a larger extended data set is constructed, including steel structure parameters, cutting parameters and roughness parameters; this extended data set will play an important role in subsequent neural network training.
[0087] Among them, the step S70 specifically includes: performing weighted normalization processing on the historical data set of step S10 and the extended data set of step S60, and for each steel structure parameter or cutting parameter, using its corresponding weight vector (obtained in step S30) as the weight for normalization; dividing the normalized data set into a training set and a validation set; constructing a neural network model with a multi-branch structure, including a steel structure feature extractor, a cutting feature extractor, a feature fusion network and a roughness prediction network; using the training set to perform end-to-end training on the model, and using the validation set to perform performance evaluation, and finally obtaining a mature roughness prediction model.
[0088] Among them, the formula of the weight vector is expressed as follows:
[0089] Parameters for steel structures :
[0090]
[0091] For cutting parameters :
[0092]
[0093] in is the correlation index calculated previously.
[0094] 5. Formula representation of steel structure feature extractor:
[0095] Input: Steel structure parameter vector and the corresponding weight vector ;
[0096] ;
[0097] In the formula, Represents element-wise multiplication; Is a nonlinear activation function, such as ReLU: ; are the extracted steel structure features.
[0098] Among them, the formula of the cutting feature extractor is expressed as follows:
[0099] Input: Cutting parameter vector and the corresponding weight vector ;
[0100] ;
[0101] In the formula, Represents element-wise multiplication; Is a nonlinear activation function, such as ReLU: ; is the extracted cutting feature.
[0102] Wherein, the step S80 specifically includes: obtaining the steel structure parameters and cutting parameters of the current steel structure to be cut, and inputting these parameters into the roughness prediction model trained in step S70. The model will output the predicted roughness of the laser cutting edge of the steel structure based on the input steel structure and cutting parameters; through the prediction model, the expected roughness can be estimated before actual cutting, providing a basis for the optimization of process parameters.
[0103] Among them, during the training process of the roughness prediction model, a residual connection structure is used to improve the prediction accuracy; the residual connection structure refers to directly connecting the input layer to the output layer, thereby enhancing the fitting ability and robustness of the model.
[0104] Among them, the feature fusion network adopts the attention mechanism through a multi-layer fully connected network to fuse the steel structure features and cutting features into a feature vector; the attention mechanism can adaptively assign different weights to different features, which helps to improve the effectiveness of feature fusion.
[0105] Among them, the nonlinear activation function used in the steel structure feature extractor and the cutting feature extractor is the ReLU function, that is, f(x) = max(0,x) and g(x) = max(0,x); the ReLU function has good gradient characteristics and is conducive to the training convergence of the network.
[0106] A second aspect of the present invention provides a computer-readable storage medium, wherein the computer-readable storage medium stores program instructions, and when the program instructions are executed, they are used to execute the above-mentioned method for predicting edge roughness of laser cutting of steel structures.
[0107] A third aspect of the present invention provides a steel structure laser cutting edge roughness prediction system, which includes the above-mentioned computer-readable storage medium.
[0108] Compared with the prior art, the method, medium and system for predicting edge roughness of steel structure laser cutting provided by the present invention have the following beneficial effects:
[0109] 1. Make full use of historical data: Use the Bayesian network to conduct in-depth analysis of the correlation between parameters, fully grasp the influence of each process parameter on roughness, and lay the foundation for establishing an accurate prediction model.
[0110] 2. Improve the generalization ability of the model: Based on the fitting of the roughness equation group, an extended data set was constructed to enrich the training samples and improve the adaptability of the model under new process conditions.
[0111] 3. Enhanced multi-source information fusion: The multi-branch neural network structure can effectively integrate steel structure characteristics and cutting characteristics, fully explore the potential connection between the two types of parameters, and improve prediction accuracy.
[0112] 4. Improve prediction performance: Based on feature fusion and residual connection, the prediction model of the present invention has strong fitting and generalization capabilities, and the prediction results are highly consistent with the measured values.
[0113] In general, the roughness prediction method proposed in the present invention overcomes the shortcomings of the existing technology. Through parameter correlation analysis, data expansion and multi-source information fusion, a prediction model integrating scientific modeling and data-driven is constructed. While improving the prediction accuracy, the applicability and robustness of the model are also enhanced. It solves the technical problem that the parameter correlation analysis of the existing technology in the prediction of the edge roughness of laser cutting of steel structures is not comprehensive enough, and it is difficult to fully grasp the influence of each process parameter on the roughness. BRIEF DESCRIPTION OF THE DRAWINGS
[0114] Figure 1 A flow chart of the method provided by the present invention. DETAILED DESCRIPTION
[0115] In order to make the purpose, technical solution and advantages of the embodiments of the present invention more clear, the technical solution in the embodiments of the present invention will be clearly and completely described below in conjunction with the accompanying drawings in the embodiments of the present invention.
[0116] like Figure 1 FIG. 1 is a flow chart of a method for predicting edge roughness of steel structure laser cutting provided by the first aspect of the present invention. The method comprises the following steps:
[0117] S10, obtaining multiple sets of historical steel structure laser cutting data sets, recorded as historical data sets, including steel structure parameters, cutting parameters and roughness parameters;
[0118] S20, based on the historical data set, using the Bayesian network model, calculating the correlation index of each steel structure parameter, each cutting parameter and each roughness parameter;
[0119] S30. For each steel structure parameter, a weight vector is established to represent the correlation index between the steel structure parameter and each roughness parameter; for each cutting parameter, a weight vector is established to represent the correlation index between the cutting parameter and each roughness parameter;
[0120] S40, calculating a sensitivity matrix according to the correlation index, and setting power or interaction terms and nonlinear terms according to the sensitivity matrix, and establishing a roughness equation group considering steel structure parameters, cutting parameters and error terms, including an average roughness equation, a maximum roughness equation, a waviness equation, a kurtosis equation, a skewness equation, a variance equation, a central moment equation and a comprehensive roughness equation;
[0121] S50, fitting the roughness equation group using the historical data set to obtain a fitting equation group;
[0122] S60, setting different steel structure parameters and cutting parameters, using a fitting equation group to calculate corresponding roughness parameters, and establishing an extended data set including steel structure parameters, cutting parameters and roughness parameters;
[0123] S70, after weighted normalization processing, each set of data in the extended data set and the historical data set are used as a training set and a validation set respectively, and a multi-branch neural network model is trained to obtain a roughness prediction model; wherein the input of the training is the steel structure parameter and the cutting parameter, and the output of the training is the roughness parameter;
[0124] S80, obtaining steel structure parameters and cutting parameters of the current steel structure to be cut, and inputting them into a roughness prediction model to obtain the predicted edge roughness of the steel structure laser cut.
[0125] The specific implementation methods of the above steps are described in detail below:
[0126] The specific implementation method of step S10 is: obtain multiple groups of historical data sets of steel structure laser cutting. First, collect a large amount of steel structure laser cutting data from relevant experiments or production processes, including steel structure parameters, cutting parameters and roughness parameters. Steel structure parameters include steel thickness, strength grade, heat treatment state, surface state, residual stress and grain size. Cutting parameters include laser power, cutting speed, focal position, auxiliary gas type and pressure, nozzle diameter, pulse frequency and duty cycle. Roughness parameters are represented by texture features of multi-angle images, including average roughness, maximum roughness, waviness, kurtosis, skewness and central moment. Each set of data corresponds to a multi-dimensional roughness vector. By comprehensively collecting and organizing these historical data, the foundation for subsequent modeling and training is laid.
[0127] The specific implementation of step S20 is: using the Bayesian network model to calculate the correlation index of the parameters. First, a Bayesian network structure is constructed, with the steel structure parameters and cutting parameters as parent nodes and the roughness parameters as child nodes. Then, a structural learning algorithm (such as the K2 algorithm) is used to determine the network structure. Next, the conditional probability table (CPT) of each node is learned using the maximum likelihood estimation (MLE) or Bayesian estimation method.
[0128] For each pair of parameters (Steel structure parameters or cutting parameters) and roughness parameters , calculate their correlation index The calculation formula of correlation index is:
[0129] ;
[0130] in, yes and The mutual information between them is calculated as:
[0131] ;
[0132] yes and The entropy is calculated as:
[0133] ;
[0134] These probability values , and It can be obtained from the conditional probability table of the Bayesian network. By calculating the correlation index, the degree of association between each parameter and roughness can be quantified.
[0135] The specific implementation of step S30 is: establishing a weight vector of steel structure parameters and cutting parameters. , according to its relationship with each roughness parameter The correlation index , construct a weight vector:
[0136] Similarly, for each cutting parameter , and also construct a weight vector:
[0137] These weight vectors reflect the correlation between each parameter and the roughness feature, and will play a role in the subsequent feature extraction.
[0138] The specific implementation of step S40 is: calculate the sensitivity matrix and establish the roughness equation group. First, calculate the sensitivity matrix ,in Representation parameters Roughness parameters This can be approximated by the finite difference method.
[0139] According to the sensitivity matrix , setting different powers, interaction terms, and nonlinear terms:
[0140] if (high sensitivity), add quadratic term ;
[0141] if (Medium sensitivity), add interaction term ;
[0142] if (low sensitivity), only keep linear terms .
[0143] in , and It is a predefined threshold and can be adjusted according to actual conditions.
[0144] Finally, based on the above analysis, a set of equations containing multiple roughness characteristic equations is established:
[0145] Average roughness equation: ;
[0146] Maximum roughness equation: ;
[0147] Waviness equation: ;
[0148] Kurtosis equation: ;
[0149] Skewness equation: ;
[0150] Variance equation: ;
[0151] Central moment equation: ;
[0152] Comprehensive roughness equation: ;
[0153] in , , , , , , , is the coefficient to be determined, arrive is the error term. These equations will play an important role in subsequent fitting and training.
[0154] The specific implementation of step S50 is: fitting the roughness equation group using the historical data set. First, the roughness equation group established in step S40 is used as the model to be fitted. Then, the complete historical data set collected in step S10 is used as a training sample, and the coefficients of each equation are optimized and fitted using the least square method or other regression algorithms. Through this process, a set of calibrated roughness prediction equations can be obtained, laying the foundation for the subsequent construction of the extended data set and neural network training.
[0155] The specific implementation of step S60 is: construct an extended data set. First, set a series of different steel structure parameters and cutting parameter combinations. Then, substitute the fitting equations obtained in step S50 to calculate the corresponding roughness parameters. In this way, a larger extended data set is constructed, including steel structure parameters, cutting parameters and roughness parameters. This extended data set will play an important role in the subsequent neural network training.
[0156] The specific implementation of step S70 is: training a multi-branch neural network model. First, the historical data set of step S10 and the extended data set of step S60 are weighted and normalized. For each steel structure parameter or cutting parameter, the corresponding weight vector (obtained in step S30) is used as the weight for normalization. This ensures that the contribution of each parameter to the training is more reasonable.
[0157] Next, the normalized dataset is divided into a training set and a validation set. The training set is used to train the model, and the validation set is used to evaluate the model performance.
[0158] Then, a multi-branch neural network model is constructed. The model includes the following parts:
[0159] 1. Steel structure feature extractor: Based on the input steel structure parameters and their weight vectors , extract steel structure features .in is a non-linear activation function, such as ReLU.
[0160] 2. Cutting feature extractor: based on the input cutting parameters and their weight vector , extract cutting features .in It is also a non-linear activation function.
[0161] 3. Feature fusion network: Use the attention mechanism through a multi-layer fully connected network to fuse the steel structure features and cutting features into a feature vector.
[0162] 4. Roughness prediction network: Based on the fusion features, the roughness parameters are predicted through a multi-layer fully connected network, and the residual connection structure is used to improve the prediction accuracy.
[0163] The input of the entire neural network model is the steel structure parameters and cutting parameters, and the output is the predicted roughness parameters. The model is trained end-to-end using the training set, and the performance is evaluated using the validation set, ultimately obtaining a mature roughness prediction model.
[0164] The specific implementation of step S80 is: use the trained roughness prediction model to make predictions. First, obtain the steel structure parameters and cutting parameters of the current steel structure to be cut. Then, input these parameters into the roughness prediction model trained in step S70. The model will output the predicted roughness of the laser cutting edge of the steel structure based on the input steel structure and cutting parameters. Through this prediction model, the expected roughness can be estimated before actual cutting, providing a basis for the optimization of process parameters.
[0165] In general, this roughness prediction method based on Bayesian network, sensitivity analysis and multi-branch neural network makes full use of the information of historical data, and through parameter correlation analysis and feature extraction, a roughness prediction model integrating steel structure characteristics and cutting characteristics is constructed. This model can effectively predict the roughness of the edge of steel structure laser cutting and provide a reliable basis for the optimization of process parameters.
[0166] A second aspect of the present invention provides a computer-readable storage medium, wherein the computer-readable storage medium stores program instructions, and when the program instructions are executed, they are used to execute the above-mentioned method for predicting edge roughness of laser cutting of steel structures.
[0167] A third aspect of the present invention provides a steel structure laser cutting edge roughness prediction system, which includes the above-mentioned computer-readable storage medium.
[0168] Specifically, the principle of the present invention is:
[0169] First, by analyzing historical data through Bayesian network, the correlation between each process parameter and roughness can be fully quantified. Bayesian network is a probabilistic graphical model that can effectively describe the conditional dependency between variables. In the present invention, steel structure parameters and cutting parameters are used as parent nodes, and roughness parameters are used as child nodes. Through structural learning and parameter learning, the correlation index between each parameter and roughness is obtained. This lays the foundation for the subsequent parameter weight setting and equation construction.
[0170] Secondly, based on the results of correlation analysis, a set of roughness prediction equations considering steel structure parameters, cutting parameters and error terms were established. By calculating the sensitivity matrix, the influence of each parameter on the roughness was determined, and reasonable power, interaction terms and nonlinear terms were set accordingly. This modeling method based on physical mechanism can better reflect the complexity of the actual process and improve the interpretability of the model.
[0171] Finally, a multi-branch neural network model was constructed to adaptively fuse the steel structure features and cutting features, and the residual connection structure was used to improve the prediction accuracy. This multi-source information fusion method fully utilizes the powerful representation ability of neural networks and can effectively capture the complex relationship between different parameters, thereby obtaining more accurate roughness prediction results.
[0172] In summary, the key points of the scheme of the present invention are: 1) parameter correlation analysis based on Bayesian network to fully grasp the influence of various process parameters on roughness; 2) roughness equation based on physical mechanism is established to reflect the complexity of the process; 3) feature fusion method of multi-branch neural network is adopted to explore the potential connection between steel structure and cutting parameters. The innovation points of these three aspects are organically combined to form the core technical scheme of the roughness prediction method of the present invention, which can effectively solve the technical problems existing in the prior art that parameter correlation analysis is not comprehensive enough and it is difficult to fully grasp the influence of various process parameters on roughness.
[0173] In order to better understand and implement the present invention, a specific embodiment 1 of the present invention is provided below: The steps included in this embodiment 1 are specifically described as follows:
[0174] The specific implementation of step S10 is: obtaining multiple sets of historical data sets of steel structure laser cutting. First, a large amount of steel structure laser cutting data is collected from relevant experiments or production processes, including steel structure parameters, cutting parameters and roughness parameters.
[0175] Steel structure parameters Including: Steel thickness Strength level ; Heat treatment status ; Surface state ; Residual stress ; Grain size ;...;
[0176] Cutting parameters Includes: Laser power ; Cutting speed ;Focus position ; Auxiliary gas type ; Auxiliary gas pressure ; Nozzle diameter ; Pulse frequency ; Duty cycle ;...;
[0177] Roughness parameters The texture features of multi-angle images are used to represent it, including:
[0178] Average roughness ; Maximum roughness ; Waviness ; Kurtosis ; Skewness ;variance ; Center moment ; Comprehensive roughness ;
[0179] Each set of data corresponds to a multidimensional roughness vector By comprehensively collecting and organizing these historical data, we can lay the foundation for subsequent modeling and training.
[0180] The specific implementation of step S20 is: using the Bayesian network model to calculate the correlation index of the parameters. First, a Bayesian network structure is constructed to convert the steel structure parameters into and cutting parameters As a parent node, roughness parameter As child nodes. Then, a structural learning algorithm (such as the K2 algorithm) is used to determine the network structure. Next, the conditional probability table (CPT) of each node is learned using the maximum likelihood estimation (MLE) or Bayesian estimation method.
[0181] For each pair of parameters and roughness parameters , calculate their correlation index The calculation formula of correlation index is:
[0182] ;
[0183] in, yes and The mutual information between them is calculated as:
[0184] ;
[0185] yes The entropy is calculated as:
[0186] ;
[0187] These probability values , and It can be obtained from the conditional probability table of the Bayesian network. By calculating the correlation index, the degree of association between each parameter and roughness can be quantified.
[0188] The specific implementation of step S30 is: establishing a weight vector of steel structure parameters and cutting parameters. , according to its relationship with each roughness parameter The correlation index , construct a weight vector:
[0189] ;
[0190] Similarly, for each cutting parameter , and also construct a weight vector:
[0191] ;
[0192] These weight vectors reflect the correlation between each parameter and the roughness characteristics, and will play a role in subsequent feature extraction.
[0193] The specific implementation of step S40 is: calculate the sensitivity matrix and establish the roughness equation group. First, calculate the sensitivity matrix ,in Representation parameters Roughness parameters This can be approximated by the finite difference method.
[0194] According to the sensitivity matrix , set different powers, interaction terms and nonlinear terms:
[0195] if (high sensitivity), add quadratic term ;
[0196] if (Medium sensitivity), add interaction term ;
[0197] if (low sensitivity), only keep linear terms .
[0198] in , and It is a predefined threshold value, which can be adjusted according to the actual situation. The default setting is 1.5 times the average value, The average and 0.66 times the average.
[0199] Finally, based on the above analysis, a set of equations containing multiple roughness characteristic equations is established:
[0200] Average roughness equation:
[0201] ;
[0202] Maximum roughness equation:
[0203] ;
[0204] Waviness equation:
[0205] ;
[0206] Kurtosis equation:
[0207] ;
[0208] Skewness equation:
[0209] ;
[0210] Variance equation:
[0211] ;
[0212] Central moment equation:
[0213] ;
[0214] Comprehensive roughness equation:
[0215] ;
[0216] in , , , , , , , is the unknown coefficient, arrive is the error term. These equations will play an important role in subsequent fitting and training.
[0217] The specific implementation of step S50 is: fitting the roughness equation group using the historical data set. First, the roughness equation group established in step S40 is used as the model to be fitted. Then, the complete historical data set collected in step S10 is used as a training sample, and the coefficients of each equation are optimized and fitted using the least square method or other regression algorithms. Through this process, a set of calibrated roughness prediction equations can be obtained, laying the foundation for the subsequent construction of the extended data set and neural network training.
[0218] The specific implementation of step S60 is: construct an extended data set. First, set a series of different steel structure parameters and cutting parameters Then, substitute the fitting equations obtained in step S50 to calculate the corresponding roughness parameters This resulted in a larger extended data set including steel structure parameters , cutting parameters and roughness parameters . This extended dataset will play an important role in subsequent neural network training.
[0219] The specific implementation of step S70 is: training a multi-branch neural network model. First, the historical data set of step S10 and the extended data set of step S60 are weighted and normalized. For each steel structure parameter or cutting parameters , using its corresponding weight vector or (obtained in step S30) is normalized as a weight. This ensures that the contribution of each parameter to the training is more reasonable.
[0220] Then, the normalized dataset is divided into a training set and a validation set. The training set is used to train the model, and the validation set is used to evaluate the model performance.
[0221] Next, build a multi-branch neural network model, including the following parts:
[0222] 1. Steel structure feature extractor:
[0223] According to the input steel structure parameters and its weight vector , extract steel structure features .in Is a nonlinear activation function, such as the ReLU function: .
[0224] 2. Cut feature extractor:
[0225] According to the input cutting parameters and its weight vector , extract cutting features .in It is also a non-linear activation function.
[0226] 3. Feature fusion network:
[0227] Use the attention mechanism through a multi-layer fully connected network to integrate the steel structure features and cutting features Fusion into a feature vector .
[0228] 4. Roughness prediction network:
[0229] Based on fusion features , through a multi-layer fully connected network to perform roughness The residual connection structure is used to improve the prediction accuracy.
[0230] The input of the entire neural network model is the steel structure parameters and cutting parameters , the output is the predicted roughness parameter The model is trained end-to-end using the training set, and the performance is evaluated using the validation set, ultimately obtaining a mature roughness prediction model.
[0231] Through the detailed implementation process of the above 7 steps in this embodiment 1, it can be seen that this roughness prediction method based on parameter correlation analysis and multi-branch neural network makes full use of the information of historical data, and through feature extraction and fusion, a roughness prediction model integrating steel structure features and cutting features is constructed. This model can effectively predict the roughness of the edge of steel structure laser cutting and provide a reliable basis for the optimization of process parameters.
[0232] In order to better understand and implement the present invention, Example 2 of a specific application scenario of the present invention is provided below. This Example 2 is specifically described as follows: A certain automobile parts manufacturing company has been engaged in the production of steel automobile body structural parts for a long time. The company uses laser cutting technology to perform high-precision processing on steel plates, requiring that the roughness index of the cutting edge must meet certain requirements to ensure the performance and appearance quality of the parts. In order to optimize the laser cutting process parameters and improve the predictability and controllability of the roughness, the company decided to adopt the roughness prediction method based on Bayesian network and multi-branch neural network proposed in the present invention.
[0233] 1. Data Collection and Preprocessing
[0234] First, the company collected a large amount of steel structure laser cutting data from historical production records, including steel parameters, cutting parameters and corresponding roughness measurement results. Specifically, they include:
[0235] Steel parameters:
[0236] Steel thickness :2~10mm;
[0237] Strength level :Q235, Q345, 16Mn;
[0238] Heat treatment status : Normalizing, quenching and tempering;
[0239] Surface state : Bright, degrease;
[0240] Residual stress :0~300MPa;
[0241] Grain size :10~50μm;
[0242] Cutting parameters:
[0243] Laser power :2~6kW;
[0244] Cutting speed :2~8m / min;
[0245] Focus position :-2~2mm;
[0246] Auxiliary gas type : Nitrogen, oxygen;
[0247] Auxiliary gas pressure :0.2~0.6MPa;
[0248] Nozzle diameter :2~4mm;
[0249] Pulse frequency :500~2000Hz;
[0250] Duty Cycle :30%~70%;
[0251] Roughness parameters:
[0252] Average roughness :1.6~6.3μm;
[0253] Maximum roughness :10~40μm;
[0254] Waviness :0.2~2.0;
[0255] Kurtosis :2~5;
[0256] Skewness :-1~1;
[0257] variance :1~25;
[0258] Central Moment :2~20;
[0259] Comprehensive roughness :5~25μm;
[0260] The company collected a total of 2,000 sets of complete laser cutting data, covering the roughness measurement results under different steel materials and cutting process conditions. In order to ensure data quality, the company also conducted strict inspections and cleaning of these historical data, eliminating samples with obvious abnormalities or missing data.
[0261] 2. Parameter Correlation Analysis
[0262] Based on the above historical data, the company first built a Bayesian network model to transform the steel parameters and cutting parameters As a parent node, roughness parameter As a child node. Through the K2 algorithm, the enterprise determined the network structure and used the maximum likelihood estimation method to learn the conditional probability table (CPT) of each node.
[0263] Then, the firm calculated for each pair of parameters and The correlation index between , the specific formula is as follows:
[0264] ;
[0265] ;
[0266] in, and is mutual information, is the entropy, which can be obtained from the CPT of the Bayesian network.
[0267] After calculation, the company obtained the correlation between each parameter and the roughness index. The results are shown in Table 1:
[0268] Table 1 Correlation degree table
[0269]
[0270] From the above table, we can see that the thickness of steel and cutting speed The influence on various roughness indexes is the most significant, and the correlation index is high. and nozzle diameter The relative importance of these parameters will be reflected in subsequent modeling.
[0271] 3. Establishment of roughness equation
[0272] Based on the above correlation analysis results, the company constructed a set of roughness prediction equations that consider steel parameters and cutting parameters. First, the sensitivity matrix of each parameter to roughness was calculated. ,in .
[0273] Based on the sensitivity analysis, the company set the following parameters:
[0274] when When adding the quadratic term ;
[0275] when When adding interaction terms ;
[0276] when When , only the linear term is retained ;
[0277] Finally, the company obtained a set of prediction models containing multiple roughness characteristic equations:
[0278] Average roughness equation:
[0279] ;
[0280] Maximum roughness equation:
[0281] ;
[0282] Waviness equation:
[0283] ;
[0284] Kurtosis equation:
[0285] ;
[0286] Skewness equation:
[0287] ;
[0288] Variance equation:
[0289] ;
[0290] Central moment equation:
[0291] ;
[0292] Comprehensive roughness equation:
[0293] ;
[0294] in, , , , , , , , is the unknown coefficient, arrive is the error term. These equations will play an important role in subsequent fitting and training.
[0295] 4. Model training and verification
[0296] Based on the above roughness prediction equations, the company first used historical data sets to optimize the fitting of the coefficients of each equation and obtained a set of more reliable prediction models.
[0297] Next, the company built a larger extended data set. Specifically, based on the original 2,000 sets of historical data, the company set up 1,000 sets of different steel parameters. and cutting parameters Combination, substitute into the above fitting equation, calculate the corresponding roughness parameters This results in an extended dataset containing 3,000 sets of data.
[0298] In order to make full use of these data for model training, the company performed weighted normalization on all samples. Specifically, for each steel parameter or cutting parameters , using its corresponding correlation index or Normalize as weights. This ensures that each parameter can be reasonably reflected in the subsequent feature extraction and fusion.
[0299] Then, the company randomly divides the normalized data set into a training set and a validation set in a ratio of 7:3. The training set is used to train a multi-branch neural network model end-to-end, and the validation set is used to evaluate the performance. The neural network model includes the following main parts:
[0300] 1. Steel structure feature extractor:
[0301] According to the input steel parameters and its weight vector , extract steel structure features ,in is the ReLU activation function.
[0302] 2. Cut feature extractor:
[0303] According to the input cutting parameters and its weight vector , extract cutting features ,in It is also the ReLU function.
[0304] 3. Feature fusion network:
[0305] Using the attention mechanism, the steel structure features are integrated through a multi-layer fully connected network. and cutting features Fusion into a feature vector .
[0306] 4. Roughness prediction network:
[0307] Based on fusion features , through a multi-layer fully connected network and residual connection structure, various roughness parameters are predicted .
[0308] After repeatedly adjusting the network structure and hyperparameters, the company finally obtained a roughness model with high prediction accuracy. The mean absolute error (MAE) on the validation set is as follows:
[0309] Average roughness :MAE = 0.21 μm;
[0310] Maximum roughness :MAE = 1.8 μm;
[0311] Waviness :MAE = 0.09;
[0312] Kurtosis :MAE = 0.15;
[0313] Skewness :MAE = 0.07;
[0314] variance :MAE = 1.2;
[0315] Central Moment :MAE = 1.5;
[0316] Comprehensive roughness :MAE = 1.0 μm;
[0317] In general, the prediction accuracy of this prediction model is significantly improved compared with existing methods, and it can better adapt to different process conditions.
[0318] In general, the roughness prediction method proposed in this invention fully utilizes the advantages of Bayesian network, feature fusion neural network and other technologies to realize intelligent modeling and control of laser cutting process. This method not only has high prediction accuracy, but also has strong adaptability and robustness, providing strong technical support for the optimization management of metal processing processes such as laser cutting.
[0319] The above description is only a specific implementation mode of the present invention, but the protection scope of the present invention is not limited thereto. Any technician familiar with the technical field can easily think of changes or substitutions within the technical scope disclosed by the present invention, which should be covered by the protection scope of the present invention.
Claims
1. A method for predicting edge roughness of steel structure laser cutting, characterized in that: The following steps are involved: S10, obtaining multiple sets of historical steel structure laser cutting data sets, recorded as historical data sets, including steel structure parameters, cutting parameters and roughness parameters; S20, calculating the correlation index of each steel structure parameter, each cutting parameter and each roughness parameter using a Bayesian network model according to the historical data set; S30, for each steel structure parameter, establishing a weight vector for representing the correlation index between the steel structure parameter and each roughness parameter; for each cutting parameter, establishing a weight vector for representing the correlation index between the cutting parameter and each roughness parameter; S40, calculating a sensitivity matrix according to the correlation index, and setting power or interaction terms and nonlinear terms according to the sensitivity matrix, and establishing a roughness equation group considering steel structure parameters, cutting parameters and error terms, including an average roughness equation, a maximum roughness equation, a waviness equation, a kurtosis equation, a skewness equation, a variance equation, a central moment equation and a comprehensive roughness equation; S50, fitting the roughness equation group using the historical data set to obtain a fitting equation group; S60, setting different steel structure parameters and cutting parameters, using the fitting equation group to calculate corresponding roughness parameters, and establishing an extended data set including steel structure parameters, cutting parameters and roughness parameters; S70, after weighted normalization processing, each set of data in the extended data set and the historical data set are used as a training set and a validation set respectively, and a multi-branch neural network model is trained to obtain a roughness prediction model; wherein the input of the training is the steel structure parameter and the cutting parameter, and the output of the training is the roughness parameter; S80, obtaining steel structure parameters and cutting parameters of the current steel structure to be cut, and inputting them into the roughness prediction model to obtain the predicted edge roughness of the steel structure laser cut.
2. A method for predicting edge roughness of steel structure laser cutting according to claim 1, characterized in that: The steel structure parameters include steel thickness, strength grade, heat treatment state, surface state, residual stress, and grain size.
3. The method for predicting edge roughness of steel structure laser cutting according to claim 1, characterized in that: The cutting parameters include laser power, cutting speed, focus position, auxiliary gas type, auxiliary gas pressure, nozzle diameter, pulse frequency, and duty cycle.
4. The method for predicting edge roughness of steel structure laser cutting according to claim 1, characterized in that: The roughness is represented by multiple texture features of multi-angle images of the cutting surface, including average roughness, maximum roughness, waviness, kurtosis, skewness and central moment; the roughness is a multidimensional vector.
5. The method for predicting edge roughness of steel structure laser cutting according to claim 1, characterized in that: The weights used in the weighted normalization process are weight vectors corresponding to each steel structure parameter or each cutting parameter.
6. The method for predicting edge roughness of steel structure laser cutting according to claim 1, characterized in that: The neural network model of the multi-branch structure includes a steel structure feature extractor, a cutting feature extractor, a feature fusion network and a roughness prediction network.
7. The method for predicting edge roughness of steel structure laser cutting according to claim 6, characterized in that: The steel structure feature extractor performs feature extraction based on input steel structure parameters and their corresponding weight vectors to obtain steel structure features; The cutting feature extractor performs feature extraction according to the input cutting parameters and their corresponding weight vectors to obtain cutting features; The feature fusion network is used to fuse the steel structure features and the cutting features. Its input is the steel structure features and the cutting features. It uses the attention mechanism to perform feature fusion through a multi-layer fully connected network and outputs a fused feature vector. The roughness prediction network is used to predict the roughness according to the fusion features, and its input is the output of the feature fusion network. The prediction is performed through a multi-layer fully connected network to output the roughness.
8. The method for predicting edge roughness of steel structure laser cutting according to claim 7, characterized in that: The roughness prediction network adopts a residual connection structure to improve the prediction accuracy.
9. A computer-readable storage medium, characterized in that: The computer-readable storage medium stores program instructions, and when the program instructions are executed, they are used to execute the method for predicting edge roughness of laser cutting of steel structures according to any one of claims 1 to 8.
10. A steel structure laser cutting edge roughness prediction system, characterized in that: A computer-readable storage medium comprising the computer-readable storage medium of claim 9.
