A method and system for evaluating the quality of a metal product

Abstract

The application discloses a kind of metal product production quality evaluation method and system, it is related to production quality evaluation technical field, comprising the following steps, original data is processed, obtains fusion feature vector, surface quality error value and component compliance error value, according to surface quality error value and component compliance error value, construct objective function, based on fusion feature vector, initial population is generated by improved MSPO algorithm, and initial population is as the current population of initial iteration round, based on improved MSPO algorithm, the current population is iteratively optimized by adaptive adjustment, obtain new population and the optimal feature vector of new population, based on the optimal feature vector of new population, by CatBoost model is handled, and quality evaluation result is obtained, significantly improve the accuracy, stability and high efficiency of metal product production quality evaluation, provide reliable support for accurate control production quality.

Description

Technical Field

[0001] This invention relates to the field of production quality assessment technology, and in particular to a method and system for assessing the production quality of metal products. Background Technology

[0002] In recent years, with the rapid development of metal products, the requirements for the quality stability and reliability of metal products have become increasingly stringent. Quality assessment in the production process has become a core link in ensuring product performance and reducing defect rates. Quality assessment technologies based on visual inspection and component analysis are gradually replacing traditional manual inspection. The application of high-precision inspection equipment such as laser three-dimensional scanning and laser-induced breakdown spectroscopy has made it possible to obtain multi-dimensional data on the surface morphology and composition of metal products, promoting the transformation of quality assessment towards digitalization and intelligence. Accurate quality assessment can not only promptly detect process deviations in the production process and prevent unqualified products from flowing into downstream links, but also provide data support for optimizing production parameters. This has important practical significance for enhancing the core competitiveness of enterprises and ensuring the safety of equipment in key areas.

[0003] However, existing metal product quality assessment technologies still have many shortcomings. First, the fusion of multi-source data is insufficient, and there is a lack of effective correlation analysis between surface features in visual data and material features in composition data. This can easily lead to redundant features interfering with the assessment accuracy and makes it difficult to form a fusion feature that comprehensively reflects product quality. Second, feature weight configurations are mostly set based on experience or use traditional optimization algorithms, which cannot be dynamically and adaptively adjusted according to feature differences and quality errors, resulting in inaccurate selection of the optimal feature combination. Third, the assessment model is not adaptable enough to the optimized features, making it difficult to fully explore the mapping relationship between features and quality levels, thus affecting the reliability of the assessment results. Summary of the Invention

[0004] The technical problem solved by this invention is that in the existing metal product production quality assessment, there are difficulties in fusing multi-source heterogeneous features, the feature weights cannot be adaptively adjusted, and the optimized features have poor adaptability to the assessment model, resulting in low assessment accuracy and insufficient stability.

[0005] To solve the above-mentioned technical problems, the present invention provides the following technical solution: a method for evaluating the production quality of metal products.

[0006] As a preferred embodiment of the metal product production quality assessment method of the present invention, step S1 involves processing the original data to obtain a fused feature vector.

[0007] Step S1 includes steps S101, S102, S103, S104 and S105;

[0008] Step S101: Obtain the raw data of the metal product, including visual data and composition data;

[0009] The visual data was acquired using a laser 3D scanner, and the component data was acquired using a laser-induced breakdown spectrometer.

[0010] Step S102: Extract surface feature sets and material feature sets from visual data and composition data based on the CNFS algorithm;

[0011] Step S103: The surface feature set and material feature set are processed through the feature correlation analysis layer to obtain the correlation degree matrix;

[0012] Each row and column of the correlation matrix corresponds to a single feature resulting from the combination of the surface feature set and the material feature set. The matrix elements are the correlation values ​​between the single feature in each row and the single feature in each corresponding column, with values ​​ranging from [0,1].

[0013] Step S104: The correlation matrix is ​​processed through a redundant feature filtering layer to obtain a simplified feature set;

[0014] Step S105: The simplified feature set is processed through the feature importance evaluation layer to obtain the ranking result;

[0015] The optimal feature combination is obtained by processing the ranking results through an optimal feature combination selection layer.

[0016] Based on the optimal feature combination, a fused feature vector is obtained;

[0017] Step S106: For each feature in the surface feature set, calculate the deviation value between each feature and the corresponding preset standard feature value, and perform a weighted average to obtain the surface quality error value.

[0018] For each feature in the material feature set, calculate the deviation value between each feature and the corresponding preset standard range, and perform a weighted average to obtain the composition compliance error value;

[0019] Each feature in the surface feature set corresponds to a unique preset standard feature value, and each feature in the material feature set corresponds to a unique preset standard range.

[0020] As a preferred embodiment of the metal product production quality assessment method of the present invention, in step S2, an objective function is constructed based on the surface quality error value and the composition compliance error value, and processed by the improved MSPO algorithm based on the fused feature vector to obtain an initial population, and the initial population is used as the current population of the initial iteration round;

[0021] Step S2 includes steps S201, S202, S203 and S204;

[0022] Step S201, the objective function expression is,

[0023] ;

[0024] in, Represents the fused feature vector. This represents the surface quality error value, with a range of [0,1]. This represents the component compliance error value, with a range of [0,1]. , Represents the weighting coefficient, and + =1.

[0025] As a preferred embodiment of the metal product production quality assessment method of the present invention, in step S202, a d-dimensional search space is constructed based on the dimension of the fused feature vector, wherein each dimension in the d-dimensional search space corresponds to the weight of each feature in the fused feature vector, and the value range is [0,1].

[0026] Step S203: In the d-dimensional search space, randomly generate the initial position vector and initial velocity vector of N particles to construct the initial population, and denot it as the current population;

[0027] In this context, the initial position vector of each particle represents a set of feature weight configurations;

[0028] Step S204: Set the initial position vector of each particle to the corresponding individual optimal position vector;

[0029] Simultaneously, the initial position vector of each particle is multiplied element-wise with the corresponding fused feature vector to obtain a weighted feature vector. This weighted feature vector is then substituted into the objective function to calculate the objective function value, and the fitness value is calculated using the fitness function, whose expression is as follows:

[0030] ;

[0031] in, Indicates the fitness value;

[0032] The position vector corresponding to the particle with the highest fitness value is selected as the optimal position vector of the subgroup.

[0033] As a preferred embodiment of the metal product production quality assessment method of the present invention, step S3 involves iteratively optimizing the current population through adaptive adjustment based on the improved MSPO algorithm to obtain a new population and the optimal feature vector of the new population.

[0034] Step S3 includes step S301,

[0035] Step S301: Input the optimal position vectors of the current population individuals and the optimal position vectors of the subpopulation into the user-defined parameter network. Output the particle swarm inertia weights, individual learning factors, global learning factors, and P groups of candidate action parameters. Establish the mapping relationship between the user-defined parameter network and the particle swarm inertia weights, individual learning factors, global learning factors, and P groups of candidate action parameters. The expression is as follows:

[0036] ;

[0037] in, The inertial weights of the particle swarm are represented. Represents individual learning factors. Represents the global learning factor. This represents the candidate action parameters for group P. This represents the process by which the parametric network processes the input. Represents the optimal position vector of an individual. Represents the optimal position vector of the subgroup;

[0038] The range is [0,1].

[0039] As a preferred embodiment of the metal product manufacturing quality assessment method described in this invention, step S302 involves converting the individual optimal position vector... optimal position vector of subgroup The concatenated feature vector is obtained by concatenating the features. This concatenated feature vector is then processed through hidden layer 1 to obtain a 64-dimensional feature vector. ;

[0040] Will The 64-dimensional feature vector is obtained by processing through hidden layer 2. and will Normalization is performed by activating the Sigmoid function;

[0041] Convert the inertia weights and learning factors into actual inertia weights and actual learning factors;

[0042] The expression for the actual inertia weight is,

[0043] ;

[0044] in, Indicates the actual inertia weight. This represents the inertial weight value output by the parameter network and after normalization.

[0045] The expression for the actual learning factor is,

[0046] ;

[0047] in, Represents the actual learning factor. This represents the value of the i-th learning factor, output by the parameter network and after normalization. .

[0048] As a preferred embodiment of the metal product manufacturing quality assessment method described in this invention, step S303 involves calculating the iterative linear change parameters based on the actual inertia weight and the actual learning factor, expressed as follows:

[0049] ;

[0050] ;

[0051] in, The parameter representing the iterative linear change of the actual inertia weight. The parameter representing the iterative linear change of the learning factor. Indicates the current iteration number. This indicates the total number of iterations.

[0052] In a preferred embodiment of the metal product manufacturing quality assessment method described in this invention, step S304 involves weighted fusion of the actual inertia weight and the actual learning factor with the corresponding iterative linear change parameters of the actual inertia weight and the learning factor, respectively, to obtain the final inertia weight and the final learning factor, expressed as follows:

[0053] ;

[0054] ;

[0055] in, Indicates the final inertia weight. express , , These are weighted fusion coefficients, all ranging from [0,1].

[0056] Step S305, based on the final inertia weight Final learning factor The candidate action parameters of group P are used to adaptively update the velocity vector and position vector of each particle in the current population through the improved MSPO algorithm, generate a new population, and use the new population as the current population for the next iteration.

[0057] Step S306: Multiply the position vector of each particle in the new population element by element with the fusion feature vector to obtain a weighted feature vector. Substitute the weighted feature vector into the objective function to calculate the objective function value, and calculate the fitness value through the fitness function.

[0058] Compare the current fitness value of each particle in the new population with the fitness value of the individual optimal position vector corresponding to each particle, and retain the position vector with the higher fitness value as the updated individual optimal position vector;

[0059] The fitness values ​​corresponding to the updated individual optimal position vectors of all particles are compared globally, and the position vector with the highest fitness value is selected as the updated optimal position vector of the subgroup.

[0060] Step S307: Repeat steps S301 to S306 until the iteration termination condition is met, stop the iteration, and multiply the optimal position vector of the subgroup at the time of stopping the iteration with the fused feature vector element by element to obtain the optimal feature vector.

[0061] The iteration termination condition is that the fitness value corresponding to the optimal position vector of the subgroup changes less than a preset threshold ε for K consecutive iterations, where K is the preset number of stable iterations.

[0062] As a preferred embodiment of the metal product manufacturing quality assessment method of the present invention, step S4 involves processing the optimal feature vector using a CatBoost model to obtain the quality assessment result.

[0063] The optimal feature vector is processed through a feature preprocessing layer to obtain a standardized feature vector;

[0064] The standardized feature vectors are processed through a tree ensemble inference layer to obtain the ensembled prediction probability distribution.

[0065] The predicted probability distribution is processed through the output layer to obtain the final quality assessment result.

[0066] A metal product manufacturing quality assessment system includes a processing module, a generation module, an adjustment module, and an assessment module;

[0067] The processing module processes the raw data to obtain the fused feature vector, surface quality error value, and composition compliance error value.

[0068] The generation module constructs an objective function based on surface quality error value and composition compliance error value. Based on the fused feature vector, it generates an initial population through an improved MSPO algorithm and uses the initial population as the current population for the initial iteration round.

[0069] The adjustment module, based on the improved MSPO algorithm, iteratively optimizes the current population through adaptive adjustment to obtain a new population and its optimal feature vector.

[0070] The evaluation module, based on the optimal feature vector of the new population, processes it using the CatBoost model to obtain the quality evaluation results.

[0071] The beneficial effects of this invention are as follows: It accurately extracts multi-source features using the CNFS algorithm and organically fuses these features based on the feature correlation matrix and domain adaptation threshold. The optimal feature combination is obtained through correlation analysis and redundancy filtering, overcoming the shortcomings of traditional feature fusion methods, such as high redundancy and poor effectiveness. An improved MSPO algorithm is employed, dynamically generating inertial weights and learning factors through a custom parameter network and fusing them with linear decay weighting to achieve adaptive optimization of the population iteration, overcoming the rigid parameter adjustment and low iteration optimization efficiency of traditional algorithms. Combined with the ordered boosting mechanism and probability threshold determination of the CatBoost model, it achieves collaborative evaluation between the feature layer and the model layer, improving the generalization ability and prediction accuracy of the evaluation model. This forms a complete causal chain of data-feature-weight-evaluation, significantly improving the accuracy, stability, and efficiency of metal product production quality evaluation. It possesses substantial technological innovation and industrial application value, providing reliable support for precise control of production quality. Attached Figure Description

[0072] Figure 1 This is a schematic diagram of the basic process of a method for evaluating the production quality of metal products according to an embodiment of the present invention.

[0073] Figure 2 This is a schematic flowchart of a metal product manufacturing quality assessment system provided in one embodiment of the present invention. Detailed Implementation

[0074] To make the above-mentioned objects, features and advantages of the present invention more apparent and understandable, the specific embodiments of the present invention will be described in detail below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments.

[0075] Example 1, referring to Figure 1 As an embodiment of the present invention, a method for evaluating the production quality of metal products is provided, comprising the following steps:

[0076] Step S1: Process the raw data to obtain the fused feature vector, surface quality error value, and composition compliance error value;

[0077] Step S2: Construct an objective function based on the surface quality error value and the composition compliance error value. Based on the fused feature vector, generate an initial population using the improved MSPO algorithm, and use the initial population as the current population for the initial iteration round.

[0078] Step S3: Based on the improved MSPO algorithm, the current population is iteratively optimized through adaptive adjustment to obtain a new population and its optimal feature vector.

[0079] Step S4: Based on the optimal feature vector of the new population, the CatBoost model is used to process the data to obtain the quality assessment results.

[0080] In one embodiment, the raw data is processed to obtain a fused feature vector, surface quality error value, and composition compliance error value. Then, an objective function is constructed based on the surface quality error value and composition compliance error value. An initial population is generated based on the fused feature vector using an improved MSPO algorithm, and the initial population is used as the current population for the initial iteration round. Subsequently, the current population is adaptively adjusted and iteratively optimized using the improved MSPO algorithm to obtain a new population and the optimal feature vector. Finally, the quality assessment result is obtained by processing the optimal feature vector through the CatBoost model. Through multi-dimensional feature fusion and adaptive iterative optimization, the accuracy and evaluation efficiency of feature selection are effectively improved. Combined with the strong fitting ability of the CatBoost model, the accuracy and reliability of the quality assessment result are significantly enhanced, realizing efficient and accurate control of the production quality of metal products.

[0081] It should be noted that this invention is based on the analysis of the inherent logic of metal product manufacturing quality assessment, identifying core and interrelated technical difficulties, specifically including: the physical correlation between multi-source heterogeneous features (visual surface features and spectral composition features) is not effectively utilized, resulting in high redundancy of fused features; the feature weight configuration cannot be dynamically and adaptively adjusted with quality errors, and the parameters of traditional optimization algorithms are fixed, making it difficult to match the nonlinear coupling relationship between surface and composition errors of metal products; there is an adaptation gap between the optimized features and the CatBoost model, and conventional models cannot fully explore the deep mapping between features and quality levels. This solution constructs a progressive technical path of feature association simplification - adaptive weight optimization - orderly improvement of assessment, with clear input-output dependencies and synergy between each step.

[0082] Step S1: Process the original data to obtain the fused feature vector;

[0083] Step S1 includes steps S101, S102, S103, S104 and S105;

[0084] Step S101: Obtain the raw data of the metal product, including visual data and composition data;

[0085] Visual data is acquired using a laser 3D scanner, while compositional data is acquired using a laser-induced breakdown spectrometer.

[0086] Step S102: Extract surface feature sets and material feature sets from visual data and composition data based on the CNFS algorithm;

[0087] Step S103: The surface feature set and material feature set are processed through the feature correlation analysis layer to obtain the correlation degree matrix;

[0088] Each row and column of the correlation matrix corresponds to a single feature resulting from the combination of the surface feature set and the material feature set. The matrix elements are the correlation values ​​between the single feature in each row and the single feature in the corresponding column, with values ​​ranging from [0,1].

[0089] Step S104: The correlation matrix is ​​processed through a redundant feature filtering layer to obtain a simplified feature set;

[0090] Step S105: The simplified feature set is processed through the feature importance evaluation layer to obtain the ranking result;

[0091] The optimal feature combination is obtained by processing the ranking results through an optimal feature combination selection layer.

[0092] Based on the optimal feature combination, a fused feature vector is obtained;

[0093] Step S106: For each feature in the surface feature set, calculate the deviation value between each feature and the corresponding preset standard feature value, and perform a weighted average to obtain the surface quality error value.

[0094] For each feature in the material feature set, calculate the deviation value between each feature and the corresponding preset standard range, and perform a weighted average to obtain the composition compliance error value;

[0095] Each feature in the surface feature set corresponds to a unique preset standard feature value, and each feature in the material feature set corresponds to a unique preset standard range.

[0096] In one embodiment, step S1 aims to accurately extract data value and lay the foundation for evaluation. Through layered processing, it transforms raw data into effective features and error indicators. Step S101 uses a laser 3D scanner and a laser-induced breakdown spectrometer to collect visual and compositional data of the metal product, respectively. The high-precision detection capabilities of the laser ensure the data's accuracy and comprehensiveness, providing a multi-dimensional data source for subsequent feature extraction. Step S102 uses an adaptive CNFS algorithm that locally enhances the edge features of surface defects in the metal product and performs baseline drift calibration on the compositional spectral data. The feature extraction window size is set to 5×5, and the noise suppression coefficient is 0.0. 3. Extract surface feature sets such as surface defect density and flatness from visual data, and extract material feature sets such as the content ratio and distribution uniformity of each element from composition data to achieve preliminary data dimensionality reduction and accurate capture of key information; Step S103 calculates the Pearson correlation coefficient between features and normalizes it to the [0,1] interval to obtain the correlation matrix, clarifying the intrinsic relationship between features and providing a quantitative basis for redundancy filtering; Step S104 sets the correlation threshold to 0.7. This threshold is determined based on the industry-standard feature redundancy judgment in the field of metal product quality assessment and the adaptability analysis of the CNFS algorithm in this scheme. Redundant features with a correlation greater than or equal to 0.7 are eliminated to obtain a simplified feature set. To improve subsequent processing efficiency and avoid model overfitting; step S105 sorts and filters simplified features based on an information gain threshold of 0.15. This threshold is set based on the critical value of the contribution of features to the quality assessment result. Features with an information gain less than 0.15 improve the assessment accuracy by less than 5%. The optimal feature combination selection layer selects the top 50% of features with the highest information gain to form the optimal feature combination, which is then fused into a fused feature vector to provide efficient input for subsequent algorithm optimization; step S106 sets the surface feature standard value and material feature standard range according to the national industry standard for the corresponding metal products, using a relative deviation formula, specifically, deviation value = |actual value - standard value| / standard value. The standard value is calculated by weighting the deviation values ​​of each feature with a surface defect weight of 0.6 and a component weight of 0.4. The weight allocation is determined by combining expert scores with historical quality failure data statistics. Surface defects account for 60% of quality problems, resulting in surface quality error values ​​and component compliance error values. These provide core indicators for constructing the objective function. Each step forms a closed-loop logic of data collection, feature extraction, correlation analysis, redundancy filtering, optimal fusion, and error quantification. This progressive approach ensures the accuracy of data processing and the seamless connection of subsequent processes. It comprehensively solves the shortcomings of traditional assessments, such as low data utilization, feature redundancy, and inaccurate quantification. This lays a high-quality foundation for subsequent improvements to the MSPO algorithm and CatBoost-type assessments, ensuring the accuracy and reliability of the entire assessment scheme from the top level.

[0097] Step S2: Construct an objective function based on the surface quality error value and the composition compliance error value. Based on the fused feature vector, process it through the improved MSPO algorithm to obtain the initial population, and use the initial population as the current population for the initial iteration round.

[0098] Step S2 includes steps S201, S202, S203 and S204;

[0099] Step S201, the objective function expression is,

[0100] ;

[0101] in, Represents the fused feature vector. This represents the surface quality error value, with a range of [0,1]. This represents the component compliance error value, with a range of [0,1]. , Represents the weighting coefficient, and + =1.

[0102] Step S202: Based on the dimension of the fused feature vector, construct a d-dimensional search space. Each dimension in the d-dimensional search space corresponds to the weight of each feature in the fused feature vector, and the value range is [0,1].

[0103] Step S203: In the d-dimensional search space, randomly generate the initial position vector and initial velocity vector of N particles to construct the initial population, and denot it as the current population;

[0104] In this context, the initial position vector of each particle represents a set of feature weight configurations;

[0105] Step S204: Set the initial position vector of each particle to the corresponding individual optimal position vector;

[0106] Simultaneously, the initial position vector of each particle is multiplied element-wise with the corresponding fused feature vector to obtain a weighted feature vector. This weighted feature vector is then substituted into the objective function to calculate the objective function value, and the fitness value is calculated using the fitness function, whose expression is as follows:

[0107] ;

[0108] in, Indicates the fitness value;

[0109] The position vector corresponding to the particle with the highest fitness value is selected as the optimal position vector of the subgroup.

[0110] In one embodiment, step S2 focuses on constructing an optimization objective and generating an initial population for adaptation, providing direction and a foundation for subsequent iterative optimization. Step S201 is based on surface quality error values. Error value of ingredient compliance The objective function is constructed, and the processing method is weighted summation. The surface quality error value is quantified by weight allocation. Error value of ingredient compliance Regarding the impact on quality assessment, the core optimization objective is clearly defined as minimizing the overall error. The constraints of the objective function are as follows: weighting coefficients... and satisfy + =1, and surface quality error value Error value of ingredient compliance The values ​​of are all in the range [0,1], and the optimization objective is to minimize . Objective function and fused feature vector The necessary relationship is that, It is the only valid input basis for the objective function. As the optimal feature combination obtained after extraction, correlation analysis, and redundancy filtering by the CNFS algorithm, it comprehensively characterizes the surface and composition quality of metal products. Secondly, the core calculation of the objective function depends on... The completeness of the fusion feature vector can only be achieved when the fusion feature vector is obtained. Only under these conditions can accurate calculations be made. and This provides the objective function with a clear optimization direction and quantitative basis, supporting the adaptive optimization of feature weights in subsequent improvements to the MSPO algorithm. The design aims to provide optimization criteria for improving the MSPO algorithm. , The values ​​of 0.6 and 0.4, as mentioned earlier, are based on expert ratings combined with historical failure data, with surface defects accounting for 60% of quality problems. and All values ​​are [0,1], ensuring that the function output can be quantified and compared;

[0111] Step S202 constructs a d-dimensional search space based on the fused feature vector dimension d, with each dimension corresponding to a feature weight and taking values ​​[0,1]. The processing method is dimension matching mapping, which transforms the feature weight optimization problem into a space search problem to define the optimization range. The design is adapted to the subsequent particle swarm search logic. For example, if the fused feature vector is 12-dimensional, a 12-dimensional search space is constructed, with each dimension corresponding to the weight value of one feature. Step S203 randomly generates the initial position and velocity vectors of N=50 particles in the d-dimensional space. The N value is set based on the industry standard initial population size of particle swarm algorithms combined with the feature dimension adaptability analysis of this scheme. When the feature dimension is 10-20, N=50 can balance the search. To ensure comprehensiveness and computational efficiency, the initial velocity vector is set to a range of [0, 0.3], based on the adaptation of the feature weights [0, 1] to avoid excessively large initial velocities causing particles to exceed the search space. An initial population is constructed using random sampling, which improves optimization comprehensiveness by covering the search space with multiple particles. This step aims to provide initial samples for optimization and serves as a population foundation for iterative optimization. Each particle's position vector represents a set of feature weights. In step S204, the initial position vector is set to the individual optimal value. A weighted feature vector is obtained by element-wise multiplication of the position and the fused feature vector. When the weighted feature vector is substituted into the objective function, optimization is achieved through the weighted feature vector. and The evaluation accuracy, i.e., the weighted core features more accurately reflect the quality deviation, is improved by converting the calculated function value into a fitness value using a fitness function. This transforms the problem of minimizing the error into a problem of maximizing the fitness, adapting to the algorithm's optimization logic. The position vector corresponding to the optimal fitness value is selected as the optimal subgroup. The purpose of this step is to establish an initial optimization benchmark, providing individual and global references for subsequent iterations. Each step forms a progressive logic of goal setting, space construction, population generation, and benchmark setting. The objective function in step S201 clarifies the optimization direction and determines the dimensions and core of the search space construction in step S202. Step S20... The search space in step S203 defines the boundary for the initial population generation, ensuring that the particle configuration meets the feature weight requirements. The initial population in step S203 serves as the basis for the baseline setting in step S204, which provides a basis for comparison in subsequent iterative optimizations. Step S201 achieves comprehensive error quantification by clarifying the weights and function forms, solving the problem of the one-sidedness of single error evaluation. Steps S202-S203 ensure the adaptability and comprehensiveness of the initial population, avoiding optimization from getting trapped in local optima. The baseline setting in step S204 improves the efficiency of subsequent iterations. Overall, step S2 connects the fusion of feature vectors and surface quality error values. and ingredient compliance error value This provides a clear objective, reasonable space, and reliable initial conditions for improving the MSPO algorithm optimization, and solves the shortcomings of traditional optimization such as vague objectives and poor initial population fitness, thus ensuring the accuracy and efficiency of the subsequent optimization process from the top level.

[0112] Step S3: Based on the improved MSPO algorithm, the current population is iteratively optimized through adaptive adjustment to obtain a new population and its optimal feature vector.

[0113] Step S3 includes step S301,

[0114] Step S301: Input the optimal position vectors of the current population individuals and the optimal position vectors of the subpopulation into the user-defined parameter network. Output the particle swarm inertia weights, individual learning factors, global learning factors, and P groups of candidate action parameters. Establish the mapping relationship between the user-defined parameter network and the particle swarm inertia weights, individual learning factors, global learning factors, and P groups of candidate action parameters. The expression is as follows:

[0115] ;

[0116] in, The inertial weights of the particle swarm are represented. Represents individual learning factors. Represents the global learning factor. This represents the candidate action parameters for group P. This represents the process by which the parametric network processes the input. Represents the optimal position vector of an individual. Represents the optimal position vector of the subgroup;

[0117] The range is [0,1].

[0118] Step S302, the individual optimal position vector optimal position vector of subgroup The concatenated feature vector is obtained by concatenating the features. This concatenated feature vector is then processed through hidden layer 1 to obtain a 64-dimensional feature vector. ;

[0119] Will The 64-dimensional feature vector is obtained by processing through hidden layer 2. and will Normalization is performed by activating the Sigmoid function;

[0120] Convert the inertia weights and learning factors into actual inertia weights and actual learning factors;

[0121] The expression for the actual inertia weight is,

[0122] ;

[0123] in, Indicates the actual inertia weight. This represents the inertial weight value output by the parameter network and after normalization.

[0124] The expression for the actual learning factor is,

[0125] ;

[0126] in, Represents the actual learning factor. This represents the value of the i-th learning factor, output by the parameter network and after normalization. .

[0127] Step S303: Calculate the iterative linear change parameters based on the actual inertia weights and the actual learning factor, expressed as follows:

[0128] ;

[0129] ;

[0130] in, The parameter representing the iterative linear change of the actual inertia weight. The parameter representing the iterative linear change of the learning factor. Indicates the current iteration number. This indicates the total number of iterations.

[0131] Step S304: The actual inertia weight and actual learning factor are weighted and fused with the corresponding iterative linear change parameters of the actual inertia weight and the iterative linear change parameters of the learning factor, respectively, to obtain the final inertia weight and final learning factor, expressed as follows:

[0132] ;

[0133] ;

[0134] in, Indicates the final inertia weight. express , , These are weighted fusion coefficients, all ranging from [0,1].

[0135] Step S305, based on the final inertia weight Final learning factor The candidate action parameters of group P are used to adaptively update the velocity vector and position vector of each particle in the current population through the improved MSPO algorithm, generate a new population, and use the new population as the current population for the next iteration.

[0136] Step S306: Multiply the position vector of each particle in the new population element by element with the fusion feature vector to obtain a weighted feature vector. Substitute the weighted feature vector into the objective function to calculate the objective function value, and calculate the fitness value through the fitness function.

[0137] Compare the current fitness value of each particle in the new population with the fitness value of the individual optimal position vector corresponding to each particle, and retain the position vector with the higher fitness value as the updated individual optimal position vector;

[0138] The fitness values ​​corresponding to the updated individual optimal position vectors of all particles are compared globally, and the position vector with the highest fitness value is selected as the updated optimal position vector of the subgroup.

[0139] Step S307: Repeat steps S301 to S306 until the iteration termination condition is met, stop the iteration, and multiply the optimal position vector of the subgroup at the time of stopping the iteration with the fused feature vector element by element to obtain the optimal feature vector.

[0140] The iteration termination condition is that the fitness value corresponding to the optimal position vector of the subgroup changes less than a preset threshold ε for K consecutive iterations, where K is the preset number of stable iterations.

[0141] In one embodiment, step S3 aims to obtain the optimal feature weight configuration through adaptive iterative optimization to improve the quality of feature vectors and support accurate evaluation. Specifically, step S301 involves obtaining the optimal position vector of individuals in the current population. optimal position vector of subgroup Input a custom parameter network, which is a 3-layer fully connected structure. The input layer dimension is 2d, where d is the dimension of the fused feature vector. and The concatenated vector has a dimension of 3+. Both hidden layers use the ReLU activation function, and the output layer has a dimension of 3+. To match And the output requirements for candidate motion parameters in group P, output inertia weights. Individual learning factors Global learning factor And P=10 sets of candidate action parameters, where P is set based on the breadth of balance action exploration and computational efficiency. After optimization and verification by the metal product quality assessment algorithm, the mapping relationship is expressed as follows: ,in The range is [0,1], and the processing method is dual-optimal vector driven parameter output. This is achieved by dynamically generating parameters guided by individual and global optimal features. This step provides adaptive adjustment parameters for algorithm iteration, addressing the inherent limitation of fixed and rigid parameters in traditional MSPO algorithms. It should be clarified that the MSPO algorithm in this scheme refers to a particle swarm optimization algorithm based on the standard Particle Swarm Optimization (PSO) algorithm, with its velocity and position updates using standard updates. Step S302 will... and The concatenated vectors are then processed through two 64-dimensional hidden layers to obtain the feature vectors. After normalization to the [0,1] interval by the Sigmoid function, and then by the formula Converted to actual inertia weights, Convert to learning factors, set =0.4、 =0.9, =0.5、 =2.0, set based on the industry-standard range of particle swarm optimization (PSO) parameters combined with the feature dimension adaptability analysis of this solution. The processing method is vector concatenation-feature mapping-parameter transformation. The principle is to achieve nonlinear adaptive adjustment of parameters through neural networks. The purpose of this step is to generate actual adjustment parameters that meet the iteration requirements, and the design improves the algorithm's adaptability to the search space. Step S303 calculates the iterative linear change parameters of the actual inertia weight and the iterative linear change parameters of the learning factor based on the actual parameters, and sets the total number of iterations. The design is based on balancing optimization accuracy and computation time, and the processing method is linear decay calculation. This simulates the gradual process of algorithm iteration from global exploration to local development. This step provides a linear adjustment benchmark for parameter fusion and is designed to ensure convergence stability in the later stages of iteration. Step S304 weights and fuses the actual parameters with the linearly changing parameters, and sets the weighting fusion coefficient. = =0.6, set based on prioritizing the adaptive characteristics of parameters while considering the stability of linear adjustment. The processing method is weighted summation, which combines the advantages of adaptive and linear parameters. This step generates the final iterative adjustment parameters, and the design avoids the algorithm from getting trapped in local optima. The final inertia weight and final learning factor calculated in step S304 are directly substituted into the corresponding positions in the standard update of the particle swarm optimization algorithm, forming the core of the improved MSPO algorithm in this scheme. Regarding the P groups of candidate action parameters, their function is: at each velocity update, a group is randomly selected from the P groups of candidate action parameters to dynamically adjust the value of the learning factor. Specifically, the final individual learning factor and the global learning factor are multiplied by the adjustment coefficient in the corresponding candidate action parameter, which enhances the diversity of particle swarm search and avoids the algorithm from getting trapped in local optima too early. It should be noted that if the candidate action parameters are not used (i.e., the adjustment coefficient is always kept at 1), this scheme can still achieve adaptive optimization. Step S305 is based on the final inertia weight, learning factor and candidate action parameters. As parameters, the velocity and position vectors of the population particles are adaptively updated by improving the MSPO algorithm to generate a new population. The processing method is parameter-driven particle position iteration, and intelligent particle search is achieved by dynamically adjusting the parameters. This step completes a single round of population optimization iteration, and the design pushes the feature weights to approach the optimal configuration. Step S306 multiplies the new population particle position vector with the fused feature vector element by element to obtain a weighted feature vector, substitutes it into the objective function to calculate the function value, and then converts it into a fitness value through the fitness function. The optimal position vector of the individual and the optimal position vector of the subgroup are compared and updated. The processing method is weighted feature evaluation-fitness comparison-optimal update. The fitness value quantifies the quality of the particle position. This step establishes the optimal benchmark for this round of iteration and provides an optimization starting point for the next round of iteration. Step S307 repeats steps S301-S306 until the iteration termination condition is met. The iteration termination condition is that the change of the optimal fitness value of the subgroup for K consecutive iterations is less than a preset threshold ε, where K is set to 10 and ε is set to 0.001, K and ε are set to ensure the stability of the optimization results. After stopping the iteration, the optimal position vector of the final subgroup is multiplied element-wise with the fused feature vector to obtain the optimal feature vector. The processing method is iterative convergence judgment - optimal feature generation. The reliability of the output feature is ensured by the stable convergence condition. This step outputs the optimal feature vector after the algorithm optimization. The design provides high-quality input for the subsequent CatBoost model evaluation. Each step forms a closed-loop optimization logic of parameter generation - parameter transformation - parameter fusion - population iteration - optimal update - convergence judgment. The parameter network output of step S301 is the basis for parameter transformation in step S302 and determines the initial characteristics of the adjusted parameters. The actual parameters of step S302 and the linear parameters of step S303 together support the weighted fusion in step S304. The final parameters after fusion directly drive the population iteration in step S305. The new population in step S305 is Step S306 serves as the carrier for optimal updating, and the updated optimal vector then becomes the input for step S301 to initiate the next iteration. Step S307's convergence determination is the endpoint control of the entire optimization process, ensuring the output of the optimal feature vector. Steps S301-S304, through adaptive parameter adjustment and fusion, address the shortcomings of the traditional MSPO algorithm, such as fixed parameters and susceptibility to local optima. Steps S305-S306, through population iteration and optimal updating, improve the accuracy of feature weight configuration. Step S307's convergence determination ensures the stability of the optimization results. Overall, step S3, by improving the adaptive iterative optimization of the MSPO algorithm, achieves intelligent optimization of feature weights, overcoming the shortcomings of subjective and low-accuracy traditional feature weight configuration. This provides high-quality optimal feature vectors for subsequent CatBoost model quality evaluation, ensuring the accuracy and reliability of the entire evaluation scheme.

[0142] It should be noted that the custom parameter network obtains its internal connection weights and bias parameters through offline pre-training. The training data construction method specifically includes: collecting quality assessment samples from several batches in the history of metal product production; running the standard particle swarm optimization algorithm on each sample; recording the individual optimal position vector and the subgroup optimal position vector at each iteration; and simultaneously using a combination of grid search and experimental verification to determine the optimal inertia weights, individual learning factors, global learning factors, and multiple sets of candidate action parameters for the final feature weight configuration. These combinations are used as the network output labels, and the paired input vectors (individual optimal...) are... The training samples are formed by concatenating the position vector with the optimal position vector of the subgroup and the corresponding output label. The loss function of the parameter network adopts the mean squared error loss, which is the sum of squares of the differences between the network output inertia weight, individual learning factor, global learning factor and the difference between each group of candidate action parameters and the corresponding label. During training, the mini-batch stochastic gradient descent algorithm is used, with the batch size set to 32 and the learning rate set to 0.001. Iterative training is performed until the loss function value converges. After training, the weight parameters of the parameter network are fixed and saved. In the actual metal product production quality evaluation process, only forward calculation is performed and no online updates are performed.

[0143] In summary, a fused feature vector is obtained based on feature correlation analysis and redundancy filtering; an objective function is constructed based on surface quality error values ​​and composition compliance error values; inertia weights and learning factors are dynamically generated through a custom parameter network; the final inertia weights and final learning factors are obtained by weighted fusion of adaptive parameters and linear decay parameters; the final parameters are substituted into the velocity and position update rules of the standard particle swarm optimization algorithm for population iterative optimization; and the optimal feature vector is evaluated using the CatBoost model, thus forming a complete causal chain from data fusion to adaptive optimization and then to quality assessment.

[0144] Step S4: The optimal feature vector is processed by the CatBoost model to obtain the quality assessment result;

[0145] The optimal feature vector is processed through a feature preprocessing layer to obtain a standardized feature vector;

[0146] The standardized feature vectors are processed through a tree ensemble inference layer to obtain the ensembled prediction probability distribution.

[0147] The predicted probability distribution is processed through the output layer to obtain the final quality assessment result.

[0148] In one embodiment, step S4 aims to achieve accurate quality assessment of metal products based on the optimized feature vector. The optimal feature vector is input into the feature preprocessing layer and processed using the Z-score normalization method to obtain a standardized feature vector. This processing method is statistical normalization, which eliminates the influence of dimensional differences and numerical ranges between different features, improving the consistency of the feature vector. The design is adapted to the CatBoost model's requirements for the distribution of input features. The standardized feature vector is then input into the tree integration inference layer. This layer uses 100 gradient boosting trees to construct the integration model. The number of trees is set to balance assessment accuracy and computational efficiency. Preliminary experiments verified that when the number of trees is greater than or equal to 100, the evaluation accuracy improvement is less than 2%. The depth of each tree is set to 6 to avoid overfitting. Combined with the adaptability analysis of the feature dimension d of metal products being less than or equal to 20, the catBoost model's unique ordered boosting algorithm is used to reduce the category feature bias. The output is an integrated predicted probability distribution, such as the probability proportions of qualified, unqualified, and pending re-inspection. The processing method is ensemble tree inference, which uses the fusion of prediction results from multiple decision trees to reduce generalization error and improve the accuracy and robustness of quality assessment. The design ensures the reliability of the evaluation results. The predicted probability distribution is input to the output layer, and a probability threshold is set. , The threshold of 0.2 is set based on the requirement that the misjudgment rate in the metal products quality assessment industry should be less than or equal to 3%. Experiments have verified that the misjudgment rates for both qualified and unqualified products are below 2% under this threshold. If the probability of a certain quality grade is greater than or equal to... If the probability of any level is less than a certain value, then it is directly classified as that level. If the probability is not met, it is determined to be pending re-inspection; otherwise, the highest probability level is output as the final quality assessment result. The processing method is probability threshold determination, which clarifies the assessment result by quantifying the probability boundary, transforming abstract probability into an intuitive and applicable assessment conclusion. The design meets the practical needs of quality judgment in industrial scenarios. Each step forms a progressive logic of feature standardization, ensemble reasoning, and probability determination. Standardization processing eliminates interference through tree ensemble reasoning, ensuring that the model focuses on the intrinsic correlation of features. Predicting the probability distribution provides a quantitative basis for probability threshold determination, ensuring the objectivity of the result judgment. Based on the CatBoost model, it achieves accurate mapping from optimized features to assessment results, solving the model bias problem caused by inconsistent feature dimensions. Ensemble learning improves assessment accuracy and noise resistance, avoiding the limitations of a single model. Standardized output of assessment results is achieved by clarifying the threshold, reducing the difficulty of practical operation. Based on the optimal feature vector, the advantages of the CatBoost model are fully utilized through hierarchical processing, solving the shortcomings of poor generalization ability and ambiguous results of traditional quality assessment models. It completes the closed loop of feature optimization and model evaluation, ensuring the implementation and accuracy of the entire assessment scheme from the top level, and providing a reliable quantitative basis for the quality control of metal products.

[0149] Example 2, refer to Figure 2 In another embodiment of the present invention, which differs from the first embodiment, a metal product manufacturing quality assessment system is provided, including a processing module, a generation module, an adjustment module, and an assessment module;

[0150] The processing module processes the raw data to obtain the fused feature vector, surface quality error value, and composition compliance error value.

[0151] The generation module constructs an objective function based on surface quality error value and composition compliance error value. Based on the fused feature vector, it generates an initial population through an improved MSPO algorithm and uses the initial population as the current population for the initial iteration round.

[0152] The adjustment module, based on the improved MSPO algorithm, iteratively optimizes the current population through adaptive adjustment to obtain a new population and its optimal feature vector.

[0153] The evaluation module, based on the optimal feature vector of the new population, processes it using the CatBoost model to obtain the quality evaluation results.

[0154] In one embodiment, the raw data is processed to obtain a fused feature vector, surface quality error value, and composition compliance error value. An objective function is then constructed based on the surface quality error value and the composition compliance error value. An initial population is generated using an improved MSPO algorithm based on the fused feature vector, and this initial population is used as the current population for the initial iteration round. Subsequently, the improved MSPO algorithm is used to adaptively adjust and iteratively optimize the current population to obtain a new population and the optimal feature vector. Finally, the quality assessment result is obtained by processing the optimal feature vector using a CatBoost model. Through multi-dimensional feature fusion and adaptive iterative optimization, the accuracy and efficiency of feature selection are effectively improved. Combined with the strong fitting ability of the CatBoost model, the accuracy and reliability of the quality assessment result are significantly enhanced, achieving efficient and precise control over the production quality of metal products.

[0155] This invention accurately extracts multi-source features using the CNFS algorithm and organically fuses them based on the feature correlation matrix and domain adaptation threshold. It obtains the optimal feature combination through correlation analysis and redundancy filtering, overcoming the shortcomings of high redundancy and poor effectiveness in traditional feature fusion. An improved MSPO algorithm is employed, dynamically generating inertial weights and learning factors through a custom parameter network and fusing them with linear decay weighting to achieve adaptive optimization of the population iteration, overcoming the rigid parameter adjustment and low iteration efficiency of traditional algorithms. Combining the ordered boosting mechanism and probability threshold determination of the CatBoost model enables collaborative evaluation between the feature layer and the model layer, improving the generalization ability and prediction accuracy of the evaluation model. This forms a complete causal chain of data-feature-weight-evaluation, significantly improving the accuracy, stability, and efficiency of metal product production quality evaluation. It possesses substantial technological innovation and industrial application value, providing reliable support for precise control of production quality.

[0156] Those skilled in the art will understand that embodiments of the present invention can be provided as methods, systems, or computer program products. Therefore, the present invention can take the form of a completely hardware embodiment, a completely software embodiment, or an embodiment combining software and hardware aspects. Furthermore, the present invention can take the form of a computer program product implemented on one or more computer-usable storage media containing computer-usable program code. The storage medium can be implemented by any type of volatile or non-volatile storage device or a combination thereof, such as Static Random Access Memory (SRAM), Electrically Erasable Programmable Read-Only Memory (EEPROM), Erasable Programmable Read Only Memory (EPROM), Programmable Red-Only Memory (PROM), Read-Only Memory (ROM), magnetic storage, flash memory, magnetic disk, or optical disk. These computer program instructions may also be stored in a computer-readable storage medium that can direct a computer or other programmable data processing device to function in a particular manner, such that the instructions stored in the computer-readable storage medium produce an article of manufacture including instruction means, which are implemented in a process Figure 1 One or more processes and / or boxes Figure 1 The function specified in one or more boxes.

[0157] It should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention and are not intended to limit it. Although the present invention has been described in detail with reference to preferred embodiments, those skilled in the art should understand that modifications or equivalent substitutions can be made to the technical solutions of the present invention without departing from the spirit and scope of the technical solutions of the present invention, and all such modifications or substitutions should be covered within the protection scope of the present invention.

Claims

1. A method for evaluating the production quality of metal products, characterized in that, Includes the following steps, Step S1: Process the raw data to obtain the fused feature vector, surface quality error value, and composition compliance error value; Step S2: Construct an objective function based on the surface quality error value and the composition compliance error value. Based on the fused feature vector, generate an initial population using the improved MSPO algorithm, and use the initial population as the current population for the initial iteration round. Step S3: Based on the improved MSPO algorithm, the current population is iteratively optimized through adaptive adjustment to obtain a new population and its optimal feature vector. Step S4: Based on the optimal feature vector of the new population, the CatBoost model is used to process the data to obtain the quality assessment results.

2. The method for evaluating the production quality of metal products as described in claim 1, characterized in that, Step S1: Process the original data to obtain the fused feature vector; Step S1 includes steps S101, S102, S103, S104 and S105; Step S101: Obtain the raw data of the metal product, including visual data and composition data; The visual data was acquired using a laser 3D scanner, and the component data was acquired using a laser-induced breakdown spectrometer. Step S102: Extract surface feature sets and material feature sets from visual data and composition data based on the CNFS algorithm; Step S103: The surface feature set and material feature set are processed through the feature correlation analysis layer to obtain the correlation degree matrix; Each row and column of the correlation matrix corresponds to a single feature resulting from the combination of the surface feature set and the material feature set. The matrix elements are the correlation values ​​between the single feature in each row and the single feature in each corresponding column, with values ​​ranging from [0,1]. Step S104: The correlation matrix is ​​processed through a redundant feature filtering layer to obtain a simplified feature set; Step S105: The simplified feature set is processed through the feature importance evaluation layer to obtain the ranking result; The optimal feature combination is obtained by processing the ranking results through an optimal feature combination selection layer. Based on the optimal feature combination, a fused feature vector is obtained; Step S106: For each feature in the surface feature set, calculate the deviation value between each feature and the corresponding preset standard feature value, and perform a weighted average to obtain the surface quality error value. For each feature in the material feature set, calculate the deviation value between each feature and the corresponding preset standard range, and perform a weighted average to obtain the composition compliance error value; Each feature in the surface feature set corresponds to a unique preset standard feature value, and each feature in the material feature set corresponds to a unique preset standard range.

3. The method for evaluating the production quality of metal products as described in claim 2, characterized in that, Step S2: Construct an objective function based on the surface quality error value and the composition compliance error value. Based on the fused feature vector, process it through the improved MSPO algorithm to obtain the initial population, and use the initial population as the current population for the initial iteration round. Step S2 includes steps S201, S202, S203 and S204; Step S201, the objective function expression is, ; in, Represents the fused feature vector. This represents the surface quality error value, with a range of [0,1]. This represents the component compliance error value, with a range of [0,1]. , Represents the weighting coefficient, and + =1.

4. The method for evaluating the production quality of metal products as described in claim 3, characterized in that, Step S202: Based on the dimension of the fused feature vector, construct a d-dimensional search space. Each dimension in the d-dimensional search space corresponds to the weight of each feature in the fused feature vector, and the value range is [0,1]. Step S203: In the d-dimensional search space, randomly generate the initial position vector and initial velocity vector of N particles to construct the initial population, and denot it as the current population; In this context, the initial position vector of each particle represents a set of feature weight configurations; Step S204: Set the initial position vector of each particle to the corresponding individual optimal position vector; Simultaneously, the initial position vector of each particle is multiplied element-wise with the corresponding fused feature vector to obtain a weighted feature vector. This weighted feature vector is then substituted into the objective function to calculate the objective function value, and the fitness value is calculated using the fitness function, whose expression is as follows: ； in, Indicates the fitness value; The position vector corresponding to the particle with the highest fitness value is selected as the optimal position vector of the subgroup.

5. The method for evaluating the production quality of metal products as described in claim 4, characterized in that, Step S3: Based on the improved MSPO algorithm, the current population is iteratively optimized through adaptive adjustment to obtain a new population and its optimal feature vector. Step S3 includes step S301, Step S301: Input the optimal position vectors of the current population individuals and the optimal position vectors of the subpopulation into the user-defined parameter network. Output the particle swarm inertia weights, individual learning factors, global learning factors, and P groups of candidate action parameters. Establish the mapping relationship between the user-defined parameter network and the particle swarm inertia weights, individual learning factors, global learning factors, and P groups of candidate action parameters. The expression is as follows: ; in, The inertial weights of the particle swarm are represented. Represents individual learning factors. Represents the global learning factor. This represents the candidate action parameters for group P. This represents the process by which the parametric network processes the input. Represents the optimal position vector of an individual. Represents the optimal position vector of the subgroup; The range is [0,1].

6. The method for evaluating the production quality of metal products as described in claim 5, characterized in that, Step S302, the individual optimal position vector optimal position vector of subgroup The concatenated feature vector is obtained by concatenating the features. This concatenated feature vector is then processed through hidden layer 1 to obtain a 64-dimensional feature vector. ; Will The 64-dimensional feature vector is obtained by processing through hidden layer 2. and will Normalization is performed by activating the Sigmoid function; Convert the inertia weights and learning factors into actual inertia weights and actual learning factors; The expression for the actual inertia weight is, ; in, Indicates the actual inertia weight. This represents the inertial weight value output by the parameter network and after normalization. The expression for the actual learning factor is, ; in, Represents the actual learning factor. This represents the value of the i-th learning factor, output by the parameter network and after normalization. .

7. The method for evaluating the production quality of metal products as described in claim 6, characterized in that, Step S303: Calculate the iterative linear change parameters based on the actual inertia weights and the actual learning factor, expressed as follows: ; ; in, The parameter representing the iterative linear change of the actual inertia weight. The parameter representing the iterative linear change of the learning factor. Indicates the current iteration number. This indicates the total number of iterations.

8. The method for evaluating the production quality of metal products as described in claim 7, characterized in that, Step S304: The actual inertia weight and actual learning factor are weighted and fused with the corresponding iterative linear change parameters of the actual inertia weight and the iterative linear change parameters of the learning factor, respectively, to obtain the final inertia weight and final learning factor, expressed as follows: ; ; in, Indicates the final inertia weight. express , , These are weighted fusion coefficients, all ranging from [0,1]. Step S305, based on the final inertia weight Final learning factor The candidate action parameters of group P are used to adaptively update the velocity vector and position vector of each particle in the current population through the improved MSPO algorithm, generate a new population, and use the new population as the current population for the next iteration. Step S306: Multiply the position vector of each particle in the new population element by element with the fusion feature vector to obtain a weighted feature vector. Substitute the weighted feature vector into the objective function to calculate the objective function value, and calculate the fitness value through the fitness function. Compare the current fitness value of each particle in the new population with the fitness value of the individual optimal position vector corresponding to each particle, and retain the position vector with the higher fitness value as the updated individual optimal position vector; The fitness values ​​corresponding to the updated individual optimal position vectors of all particles are compared globally, and the position vector with the highest fitness value is selected as the updated optimal position vector of the subgroup. Step S307: Repeat steps S301 to S306 until the iteration termination condition is met, stop the iteration, and multiply the optimal position vector of the subgroup at the time of stopping the iteration with the fused feature vector element by element to obtain the optimal feature vector. The iteration termination condition is that the fitness value corresponding to the optimal position vector of the subgroup changes less than a preset threshold ε for K consecutive iterations, where K is the preset number of stable iterations.

9. The method for evaluating the production quality of metal products as described in claim 8, characterized in that, Step S4: The optimal feature vector is processed by the CatBoost model to obtain the quality assessment result; The optimal feature vector is processed through a feature preprocessing layer to obtain a standardized feature vector; The standardized feature vectors are processed through a tree ensemble inference layer to obtain the ensembled prediction probability distribution. The predicted probability distribution is processed through the output layer to obtain the final quality assessment result.

10. A metal product manufacturing quality assessment system, applied to a metal product manufacturing quality assessment method as described in any one of claims 1-9, characterized in that, It includes a processing module, a generation module, an adjustment module, and an evaluation module; The processing module processes the raw data to obtain the fused feature vector, surface quality error value, and composition compliance error value. The generation module constructs an objective function based on surface quality error value and composition compliance error value. Based on the fused feature vector, it generates an initial population through an improved MSPO algorithm and uses the initial population as the current population for the initial iteration round. The adjustment module, based on the improved MSPO algorithm, iteratively optimizes the current population through adaptive adjustment to obtain a new population and its optimal feature vector. The evaluation module, based on the optimal feature vector of the new population, processes it using the CatBoost model to obtain the quality evaluation results.

Images