Improved product key process quality prediction method under small sample data
By combining grey relational analysis and residual connection network model with multi-objective co-evolutionary algorithm, the problem of quality prediction caused by numerous process parameters in product assembly under small sample data is solved, and efficient and accurate product quality prediction and production optimization are achieved.
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Patents(China)
- Current Assignee / Owner
- Filing Date
- 2022-10-11
- Publication Date
- 2026-04-07
AI Technical Summary
With small sample data, the product assembly process involves numerous and complex process parameters, making it difficult to extract process quality parameters. This results in difficulties in predicting product quality and poor accuracy. Existing technologies, such as population co-evolutionary algorithms, have low debugging efficiency and are difficult to achieve efficient optimization.
Key process parameters are extracted using grey relational analysis, a residual connection network model is constructed, and a multi-objective co-evolutionary global fast gradient optimization algorithm is designed. The process parameters are optimized through a reverse elitist strategy and a Levy flight strategy to achieve global optimization.
The optimization algorithm for product quality prediction with small sample data has been improved in terms of convergence speed and global optimization efficiency, thereby enhancing the accuracy of product quality prediction and production efficiency.
Smart Images
Figure CN115906399B_ABST
Abstract
Description
Technical Field
[0001] This invention relates to the design of a product key process quality prediction method under small sample data, and to the establishment of a product key process quality learning meta-prediction model under residual connection network under small sample data, the optimization design of key process parameters of multi-objective quality prediction factors, and the design of a multi-objective co-evolutionary global fast gradient quality prediction optimization solution algorithm. Background Technology
[0002] The product key process quality prediction method based on small sample data, as addressed in this invention, is an important technical means to improve product quality and achieve "zero" defects in production. However, in actual product assembly processes, due to the large number of process parameters, the cumbersome acquisition process, the difficulty in manually extracting potential variation characteristics of processes and quality, and the insufficient amount of product quality data making global optimization of process quality parameters difficult, the development of related technologies such as product key process quality prediction methods based on small sample data has been slow. However, with the development of computer science and technology, intelligent algorithms are gradually being applied to process optimization and monitoring analysis in product manufacturing, providing possibilities for freeing up human resources in the production process and improving worker safety and production efficiency. Therefore, developing product key process quality prediction methods based on small sample data has practical significance, and how to utilize intelligent algorithms for product key process quality prediction based on small sample data has attracted considerable attention from scholars both domestically and internationally. The literature [A. Slowik and H. Kwasnicka, "Nature Inspired Methods and Their Industry Applications—Swarm Intelligence Algorithms," in IEEE Transactions on Industrial Informatics, vol. 14, no. 3, pp. 1004-1015, March 2018] studies swarm intelligence algorithms such as swarm co-evolution and applies swarm co-evolution algorithms to product process quality prediction, achieving good results. However, conventional swarm co-evolution algorithms suffer from problems such as the time-consuming nature of manually modifying the algorithm when solving new problems such as multi-objective control. To address the inefficiencies in debugging and modifying population co-evolutionary algorithms, the literature [CLCamacho-Villalón and M. Dorigo, "PSO-X: A Component-Based Framework for the Automatic Design of Particle Swarm Optimization Algorithms," in IEEE Transactions on Evolutionary Computation, vol. 26, no. 3, pp. 402-416, June 2022] proposes a method using automatic design to overcome the limitations of manually searching for and executing population co-evolutionary algorithms, thus improving the debugging efficiency of these algorithms. Furthermore, to address the difficulty in predicting process quality under numerous and complex process parameters in product assembly with small sample data, and to improve production efficiency, ensure production safety, and reduce production costs, an improved method for predicting key process quality under small sample data is of great significance to industrial production. Summary of the Invention
[0003] This invention aims to address the problems of difficulty in predicting product process quality and poor accuracy in product assembly processes with small sample data, due to the limited amount of product data, numerous process parameters, and the difficulty in manually extracting potential variation characteristics of processes and quality.
[0004] The method employed in this invention to solve the aforementioned problems is based on the extraction of key process parameters using grey relational analysis, the construction of a learning meta-prediction model for key process quality of products using small sample data under a residual connection network, and the design of a multi-objective co-evolutionary global fast gradient optimization algorithm for product quality prediction. By using the key process parameter extraction method based on grey relational analysis, key process parameters for product quality prediction are obtained. A learning meta-prediction model for key process quality of products using a residual network based on small sample data is constructed. Constraints on corresponding process parameters for product quality are integrated to optimize multi-objective quality factors under key process parameters for product quality prediction. Furthermore, based on a reverse elitist strategy and a global optimization iteration strategy for process parameters under Levy fly, a multi-objective co-evolutionary global fast gradient optimization algorithm for product quality prediction is designed, thereby obtaining key process quality parameters with global optimization characteristics under small sample data. This invention can effectively improve the convergence speed and global optimization efficiency of the product quality prediction optimization algorithm under small sample data.
[0005] First, based on the grey relational analysis method, the correlation degree is calculated using process parameters and quality data during product assembly:
[0006]
[0007] Where x i Let y(k) represent the i-th feature of the k-th data point, and y(k) represent the label of the k-th data point. Then, the correlation coefficient ξ between feature i of the k-th data point and the label of the k-th data point is... i (k), |·| represents the absolute value of the relevant variable, and ρ is the grey relational coefficient, which is usually 0.5. and Let |y(k)-x| represent the values of k and i that satisfy the conditions. i (k)|The maximum and minimum values that can be obtained. The correlation threshold is used to classify the correlation degree of product quality process parameters, thereby identifying those with higher correlation degrees as key process parameters to be optimized for quality prediction.
[0008] Secondly, key process parameters related to product quality are used as input training data for the model. The data is divided into multiple batches and multiple task product quality data. The model parameters are updated through alternating parallel gradient updates of the inner and outer loops, constructing a meta-prediction model for key process quality based on residual networks for small sample data. The formulas for the inner and outer loop gradient updates of the model parameters and the loss function are as follows:
[0009]
[0010] Where θ is a parameter of the product quality prediction model. It is a classification task. It is a task batch containing multiple categorized tasks, θ i 'Is a task' Model parameters updated by inner loop gradient, θ″ i It is a task Model parameters updated by the outer loop gradient; α is the inner loop learning rate, and β is the outer loop learning rate. It is a task loss function, It is the task of inner loop gradient update. The gradient direction of the loss function, It is a batch task during outer ring gradient update. Each category task The sum of the gradient direction vectors of the loss function; y (j) It is data x (j) The tag, It is data x (j) The predicted probability Indicates task Each data x (j) and its tag y (j) The sum of cross-entropy operations.
[0011] Furthermore, key process quality parameters related to product quality are used as target optimization parameters to establish a corresponding multi-factor optimization objective function for product quality. Combining a learning meta-prediction model for key process quality under small sample data and the physical characteristic constraints of the key process quality parameters, the multi-objective quality factors under key process parameters for product quality prediction are optimized. The objective function for the multi-objective quality factors under key process parameters for product quality prediction under small sample data is as follows:
[0012]
[0013] Where, {x1,x2,…,x n} are key process parameters. These are constraints imposed by the physical characteristics of the product's key process quality parameters. This means that f(x1,x2,…,x) satisfies n When (x1, x2, ..., x) is minimized n The value of f. u The u-th objective function can be expressed as follows:
[0014] f u (x1,x2,…,x n )=f model (x1,x2,…,x n )-y u (4)
[0015] Among them, f model This is a product quality prediction model based on small sample data, {x1,x2,…,x n} is the key process parameter, y u It is the expected output of the u-th objective function.
[0016] Finally, an improved global fast gradient optimization algorithm based on co-evolutionary objectives is used to achieve global optimization of key process quality parameters for products under small sample data. The main steps are as follows:
[0017] First, to reduce the search space and improve the convergence speed, a reverse population is generated using a reverse elitist strategy:
[0018]
[0019] Where: i∈(1,m), m is the number of individuals in the population; j∈(1,n), n is the number of key process parameters. X ij (t) represents the component of the i-th individual in the population at the j-th critical process parameter, and k is a random number between 0 and 1. For X ij The elite inverse solution corresponding to (t).
[0020] Second, the values of key process quality parameters of the product are updated by incorporating an improved population co-evolutionary algorithm based on the Levy flight strategy:
[0021]
[0022] v i (k) and x i (k) represents the velocity and position of the i-th individual in the k-th iteration, v i (k+1) and x i (k+1) represents the velocity and position of the i-th individual in the (k+1)-th iteration. w is the inertia coefficient, and c1 and c2 are self-cognitive factors and social cognitive factors. Let G be the optimal fitness value of the i-th individual during the iteration process.best This represents the optimal fitness value for all individuals. levy(s) is a random step size, and its formula is as follows:
[0023]
[0024] in: Γ is the gamma function, and γ is a constant, usually taken as γ = 1.5. levy(s)~s means that levy(s) is equivalent to the step size s. ν represents the mathematical expectation as 0 and the variance as . The normal distribution μ is equivalent to having a mathematical expectation of 0 and a variance of 0. It follows a normal distribution.
[0025] Third, to avoid the algorithm getting stuck in local optima, a global gradient optimization strategy is used to determine the probability that a new solution for key process parameters will replace an individual historical best solution:
[0026]
[0027] Wherein, g(x) i (k) is the fitness value of the i-th individual at the k-th iteration. Let be the historical best fitness value of the i-th individual, and T(k) be the temperature in the current iteration. This is determined by comparing the randomly generated probabilities p and p... i The size of (k), if p > p i If (k), then the individual's optimal solution is replaced by the individual's new solution. Attached Figure Description
[0028] Figure 1 This is an overall flowchart of the product key process quality prediction method under small sample data according to the present invention.
[0029] Figure 2 This is a flowchart illustrating the construction of a learning meta-prediction model for key process quality of products based on small sample data using residual networks, as described in this invention.
[0030] Figure 3 This is a flowchart illustrating the design of the global fast gradient optimization algorithm for multi-objective co-evolution in this invention. Detailed Implementation
[0031] The technical solution of the present invention will be further described in detail below with reference to the accompanying drawings.
[0032] like Figure 1As shown, the product key process quality prediction method designed in this invention under small sample data mainly includes: obtaining product key process quality parameters based on grey relational analysis; establishing a product key process quality learning meta-prediction model under small sample data using residual connection networks; optimizing key process parameters of multi-objective quality prediction factors; and designing a multi-objective co-evolutionary global fast gradient quality prediction optimization solution algorithm, as detailed below:
[0033] a. Obtaining key process quality parameters of products based on grey relational analysis
[0034] First, key process quality parameters of the product are extracted using grey relational analysis. For the small sample dataset of product process quality, each data point contains multiple features and a label, denoted by x. i Let y(k) represent the i-th process parameter of the k-th data point, and y(k) represent the label of the k-th data point. Then, the correlation coefficient ξ between the process parameter i of the k-th data point and the label of the k-th data point is... i (k) is represented as: Where |·| represents the absolute value of the relevant variable, and ρ is the grey relational coefficient, which is usually 0.5; and Let |y(k)-x| represent the values of k and i that satisfy the conditions. i (k)|The maximum and minimum values that can be obtained. Set the correlation threshold ξ(k), if a certain process parameter ξ i If (k) ≥ ξ(k), then this process parameter is considered a critical process parameter; conversely, if a certain process parameter ξ... i If (k)≤ξ(k), then this process parameter is considered not a critical process quality parameter. Therefore, parameters with higher correlation are identified as undetermined critical process parameters for quality prediction and optimization. Using the above method, the final critical process quality parameters are determined as: {x1,x2,…,x...} n}
[0035] b. Small-sample data product key process quality meta-learning prediction model under residual connection network
[0036] like Figure 2 As shown, considering the limited sample size and insignificant feature variations in small sample data, this paper utilizes grey relational analysis to obtain key process quality parameters for products. A residual network learning prediction model for key process quality is then developed, using the data of key process parameters related to product quality as training input. Parallel, multi-batch, and multi-task residual network learning is employed to accurately obtain product quality prediction models under small sample data conditions. The detailed steps are as follows:
[0037] b1. First, normalize the original dataset of a small sample of products and construct a residual connection network as the network structure for training the model.
[0038] b2. Divide the normalized product dataset D into training set D. train and test set D test .
[0039] b3. In D respectively train and D test Random sampling is used. The dataset is processed in multiple batches and tasks, meaning N classes are randomly selected from all data categories, and Q samples are drawn from each class. K samples from each class are called the support set, and the remaining QK samples are called the query set. Therefore, a classification task... There are N*K samples. Several classification tasks are involved. Composed of batch tasks
[0040] b4. Gradient updates during model training are implemented using loss functions, and the loss functions for each classification task are as follows:
[0041]
[0042] Where y (j) It is data x (j) The tags, It is data x (j) The predicted probability It is a categorized task. Indicates task Each data x (j) and its tag y (j) The sum of cross-entropy operations.
[0043] b5. The parameters of the inner loop task in the model are updated using the gradient formula:
[0044]
[0045] Where θ is a parameter of the product quality prediction model, θ i 'Is a task' Model parameters updated by inner loop gradient. It is a classification task, where α is the inner-loop learning rate. It is a task loss function, It is the task of inner loop gradient update. gradient direction of the loss function
[0046] b6. After optimizing the inner loop batch task parameters of the model, the outer loop task parameters of the model can be updated according to the following formula:
[0047]
[0048] Where θ″ i It is a task The model parameters are updated after the outer loop gradient, where β is the outer loop learning rate. It is a task loss function, It is a batch task during outer ring gradient update. Each category task The gradient direction vector of the loss function.
[0049] b7. Repeat steps b5 and b6 to obtain the product key process quality learning meta-prediction model f based on residual network under small sample data. model .
[0050] c. Optimization design of key process parameters for multi-objective quality prediction factors
[0051] Key process parameters related to product quality are used as target optimization parameters. A corresponding multi-factor optimization objective function for product quality is established. Combining a meta-prediction model for key process quality under small sample data and the physical characteristic constraints of the key process parameters, the multi-objective quality factors under key process parameters for product quality prediction are optimized. The objective function for the multi-objective quality factors under key process parameters for product quality prediction under small sample data is as follows:
[0052]
[0053] Where, {x1,x2,…,x n} are key process parameters. These are constraints imposed by the physical characteristics of the product's key process quality parameters. This means that f(x1,x2,…,x) satisfies n When (x1, x2, ..., x) is minimized n The value of f. u The u-th objective function can be expressed as follows:
[0054] f u (x1,x2,…,x n )=f model (x1,x2,…,x n )-y u (13)
[0055] Among them, f model This is a product quality prediction model based on small sample data, {x1,x2,…,x n} is the key process parameter, y u It is the expected output of the u-th objective function.
[0056] d. Design of a multi-objective co-evolutionary global fast gradient quality prediction and optimization algorithm
[0057] like Figure 3 As shown, an improved multi-objective co-evolutionary global fast gradient optimization algorithm is used to optimize key process parameters for multi-objective product quality prediction factors under small sample data. The detailed steps are as follows:
[0058] d1. Set the range of critical process parameter values for the population based on the constraints of the critical process parameters for quality, and set the population size m, the number of critical process parameters n, and the maximum number of iterations k. max and the number of iterations T at each temperature max .
[0059] d2. Randomly generate the initial values and initial update directions for all key process parameters in the population.
[0060] d3. Generate elite reverse individuals based on the elite reverse strategy:
[0061]
[0062] Where: i∈(1,m), m is the population size; j∈(1,n), n is the number of key process parameters. X ij (t) represents the component of the i-th individual in the population at the j-th critical process parameter, and k is a random number between 0 and 1. For X ij (t) corresponds to the elite reverse solution. After calculating the elite reverse solution for the key process parameters, it is first necessary to determine whether it exceeds the bounds. If it does, the boundary value is used instead. The individual with the best fitness value among the current individual and the elite reverse individuals is selected as the next generation population.
[0063] d4. Initial screening of the non-dominated solution set: After generating elite reverse individuals, when an individual is not dominated by other individuals (i.e., no other individual has a multi-objective fitness value superior to this individual), it is placed into the non-dominated solution set. Before particle updates, a particle is randomly selected from the non-dominated solution set as G. best And store the fitness of the i-th individual in the population. middle.
[0064] d5. Introduce the Mitropolis criterion from the global gradient optimization strategy into the iteration. This is achieved by using the population's optimal fitness g(G) best Set an initial temperature, and decrease it by a certain cooling factor μ after each iteration. The initial temperature of the global gradient optimization strategy for each iteration is set according to the following formula:
[0065]
[0066] Where α is the cooling coefficient, typically taken as α = 0.95, g(G best ) represents the optimal fitness of the population at the k-th iteration. T(k) represents the temperature at the k-th iteration, and T(k-1) represents the temperature at the (k-1)-th iteration.
[0067] d6. Update the non-dominated solution set, that is, when a new particle is not dominated by other particles or particles in the current non-dominated solution set, put the new particle into the non-dominated solution set.
[0068] d7. Determine the probability p that a new solution to the critical process parameters replaces the individual historical optimal solution based on the Mitropolis criterion. i (k), whose expression is as follows:
[0069]
[0070] Where T(k) is the temperature at the k-th iteration, and g(x) i (k) is the fitness value of the i-th individual at the k-th iteration. It is the historical best fitness value of the i-th individual. This is determined by comparing the randomly generated probabilities p with p... i The size of (k), if p i If (k) > p, then the individual optimal solution of the process parameters is replaced by the new solution of the key process parameters.
[0071] d8. Based on the Levy flight strategy and the swarm co-evolutionary algorithm, update the values and directions of key process parameters. The formula is:
[0072]
[0073] Among them, v i (k) and x i (k) represents the value and update direction of the i-th individual in the k-th iteration, v i (k+1) and x i (k+1) represents the value and update direction of the i-th individual in the (k+1)-th iteration. Let G be the optimal fitness value of the i-th individual during the iteration process. best Let w(k) and w(k+1) be the optimal fitness values for all individuals, respectively, and let α be the inertia weights in the k-th and k+1-th iterations of the population. α is the decay coefficient of the inertia weights. c1(k) and c2(k) are the self-cognition factor and social cognition factor in the k-th iteration, representing the self-search capability and global search capability of the key process parameters, respectively. c1(k+1) and c2(k+1) are the self-cognition factor and social cognition factor in the k-th iteration, respectively. β1 and β2 are the decay coefficients of c1 and c2. Levy(s), or Levy flight, is essentially a random step size that follows a Levy distribution.
[0074]
[0075] In the formula:
[0076]
[0077] Where Γ is the gamma function, and γ is a constant, usually taken as γ = 1.5. levy(s)~s means that levy(s) is equivalent to the step size s. ν represents the mathematical expectation as 0 and the variance as . The normal distribution μ is equivalent to having a mathematical expectation of 0 and a variance of 0. The parameters follow a normal distribution. After updating the values and direction of the process parameters, it is necessary to first determine whether the values of key process parameters have exceeded the limits. If they have, the boundary values are used instead.
[0078] d9. Update the individual optimum. That is, update the current new individual and the individual optimum. In the process of selecting a dominant particle, if neither of the two individuals is a dominant particle, then a single individual is randomly selected as the optimal individual.
[0079] d10. Update the non-dominated solution set, that is, when a new individual is not dominated by other individuals or individuals in the current non-dominated solution set, put the new individual into the non-dominated solution set.
[0080] d11. Determine if the required number of iterations T for each temperature has been reached. max and maximum number of iterations k max If the condition is not met, return to step 5.
[0081] d12. Obtain the dominant solution set, output the current optimal value of the product's key process quality parameters, and the algorithm terminates.
Claims
1. A global optimization method for key process quality parameters of a product based on a multi-objective population co-evolutionary algorithm on small sample data, characterized in that, First, based on the grey relational analysis method, the correlation degree is calculated using process parameters and quality data during product assembly: in This represents the i-th process parameter of the k-th data point. The label represents the k-th data item. This represents the correlation coefficient between the process parameter i of the k-th data point and the label of the k-th data point. Represents the absolute value of the relevant variable. The grey relational coefficient is... and These represent the values of k and i that satisfy the conditions, respectively. The maximum and minimum values that can be obtained; setting the correlation threshold. If the correlation coefficient of a certain process parameter If this process parameter is deemed critical, then the critical process quality parameter is ultimately determined as follows: ; Secondly, key process parameters related to product quality are used as input training data for the model. The data is divided into multiple batches and multiple tasks of product quality data. The model parameters are updated through alternating parallel gradient updates of the inner and outer loops. A meta-prediction model for key process quality based on residual networks is constructed for small sample data. The formulas for the inner and outer loop gradient updates of the model parameters and the loss function are as follows: in, These are the parameters of the product quality prediction model. It is a classification task. It is a batch of tasks that includes multiple categories of tasks. It is a task Model parameters updated by inner loop gradient. It is a task Model parameters updated by outer loop gradient; The inner loop learning rate, For the outer loop learning rate, It is a task loss function, It is the task of inner loop gradient update. The gradient direction of the loss function, It is a batch task during outer ring gradient update. Each category task The sum of the gradient direction vectors of the loss function; It is data The tag, It is data The predicted probability Indicates task Each data and its tags The sum of cross-entropy operations; Furthermore, key process quality parameters related to product quality are used as target optimization parameters to establish corresponding multi-factor optimization objective functions for product quality. Combining a learning meta-prediction model for key process quality under small sample data and the physical characteristic constraints of key process quality parameters, the multi-objective quality factors under key process parameters for product quality prediction are optimized. The objective function for the multi-objective quality factors under key process parameters for product quality prediction under small sample data is as follows: in, These are constraints imposed by the physical characteristics of the product's key process quality parameters. Indicates satisfaction Minimum The value that can be taken; and For the first The objective function is expressed by the following formula: in, It is a product quality prediction model based on small sample data. These are key process parameters. It is the first The expected output of each objective function; Finally, an improved global fast gradient optimization algorithm based on co-evolutionary objectives is used to achieve global optimization of key process quality parameters of the product under small sample data. The steps are as follows: First, to reduce the search space and improve the convergence speed, a reverse population is generated using a reverse elitist strategy: in: , where m is the number of individuals in the population; , where n is the number of key process parameters. Let i be the component of the i-th individual in the population at the j-th critical process parameter. It is between random numbers, for The corresponding elite reverse solution; Second, the values of key process quality parameters of the product are updated by incorporating an improved population co-evolutionary algorithm based on the Levy flight strategy: in and These are the velocity and position of the i-th individual in the k-th iteration. and These are the velocity and position of the i-th individual in the (k+1)-th iteration. It is the coefficient of inertia. and For self-cognitive factors and social cognitive factors, Let be the optimal fitness value of the i-th individual during the iteration process. This represents the optimal fitness value for all individuals, while It is a random step size, and its formula is as follows: in , , For gamma function, It is a constant. This means that levy(s) is equivalent to step size s. express This is equivalent to having a mathematical expectation of 0 and a variance of 0. The normal distribution express This is equivalent to having a mathematical expectation of 0 and a variance of 0. The normal distribution; Third, to avoid the algorithm getting stuck in local optima, a global gradient optimization strategy is used to determine the probability that a new solution for key process parameters will replace an individual historical best solution: in, It is the fitness value of the i-th individual at the k-th iteration. It is the historical best fitness value of the i-th individual. For the temperature in the current iteration, compare the probabilities of random generation. and The size, if If the individual's optimal solution is not found, then the new solution of the individual is used instead of the individual's optimal solution.