A Dynamic Disassembly Line Balancing Method Based on Support Vector Regression and Gaussian Inverse Model

Abstract

This invention provides a dynamic dismantling line balancing method based on support vector regression and Gaussian inverse model, comprising the following steps: obtaining the Pareto optimal solution set and corresponding target value set for all historical environments of the dismantling line; the optimal solution includes the robot-to-workstation assignment vector, the task-to-robot assignment vector, and the task sequence vector; the target values ​​include the workstation cycle time and the number of robots used; training the target value set using support vector regression to predict the target value in the new environment; using the Pareto optimal solution set and the corresponding target value set as training samples to establish a Gaussian inverse model; inputting the target value in the new environment into the inverse model to obtain the initial population in the new environment; performing evolutionary calculations on the generated initial population to output the optimal solution in the new environment. This invention effectively enhances the rapid response capability of the dynamic dismantling line to environmental changes and significantly improves the quality of the solution.

Description

Technical Field

[0001] This invention belongs to the field of disassembly line balancing technology, specifically involving a dynamic disassembly line balancing method based on support vector regression and Gaussian inverse model. Background Technology

[0002] Recycling and reusing waste products is an effective way for manufacturing enterprises to save resources and protect the environment. Product disassembly is an important means of realizing the recycling and reuse of waste products and a necessary link in achieving the integrity and closure of the product life cycle. Disassembly lines are the best choice for large-scale disassembly, therefore, the effective design and balancing of disassembly lines are crucial to improving disassembly efficiency. The Disassembly Line Balancing Problem (DLBP) considers how to allocate disassembly operations to operators on the disassembly line to optimize predetermined objectives, such as cycle time.

[0003] In reality, disassembly lines involve various environmental uncertainties, such as uncertain product quality. These uncertainties make DLBP essentially a dynamic multi-objective optimization problem (DMOP). However, most current research focuses on deterministic disassembly environments and cannot address the various uncertainties present in actual industrial settings.

[0004] For dynamic multi-objective optimization problems, researchers have proposed many specialized dynamic optimization algorithms. Among them, prediction-based algorithms have attracted widespread attention due to their high performance. The basic idea behind prediction-based dynamic multi-objective optimization algorithms is to help the optimization algorithm quickly find the Pareto optimal solution set in a new environment by reusing the knowledge from high-quality solutions found. They typically select labeled solutions from the historical environment and unlabeled solutions from the new environment as training samples to learn various knowledge reuse operators, such as prediction models and feature latent spaces based on statistical machine learning. The learned operators are used to predict good solutions in the new environment, thereby helping the optimization algorithm to search for time-varying Pareto front positions in a timely manner. However, existing prediction-based decision analysis uses solutions from the new environment as training samples to learn knowledge reuse operators, indicating that operator learning only begins when the new environment arrives. The learning process is both time-consuming and resource-intensive. Therefore, existing prediction-based dynamic multi-objective optimization algorithms still have room for improvement in terms of rapid response to environmental changes. Summary of the Invention

[0005] The purpose of this invention is to address the shortcomings of the aforementioned background technology by providing a dynamic disassembly line balancing method based on support vector regression and Gaussian inverse model. This method learns common knowledge from historical environments to generalize to new environments, thereby enhancing the ability to respond quickly to environmental changes and solving the difficulty of quickly tracking the optimal solution set in the new environment after environmental changes, thus significantly improving the quality of the solution.

[0006] The technical solution adopted in this invention is: a dynamic disassembly line balancing method based on support vector regression and Gaussian inverse model, comprising the following steps:

[0007] Obtain the Pareto optimal solution set and the corresponding target value set for all historical environments of the disassembly line; a single optimal solution in the Pareto optimal solution set includes the robot-to-workstation assignment vector, the task-to-robot assignment vector, and the task sequence vector; a single target value in the target value set includes the workstation cycle time and the number of robots used.

[0008] The model is trained using support vector regression to predict target values ​​in a new environment.

[0009] Using the Pareto optimal solution set and the corresponding objective value set as training samples, an inverse model based on Gaussian processes is established.

[0010] The target value under the new environment is input into the inverse model to obtain the initial population under the new environment; each individual in the initial population includes 3 solution vectors and 1 target vector; the solution vectors include the robot-to-workstation allocation vector, the task-to-robot allocation vector, and the task sequence vector; the target vector consists of the workstation cycle time and the number of robots used.

[0011] Evolutionary calculations are performed on the generated initial population to output the optimal solution under the new environment.

[0012] In the above technical solution, the process of establishing an inverse model based on a Gaussian process by using the Pareto optimal solution set and the corresponding objective value set as training samples includes: determining the parameters of the Gaussian process model; constructing the inverse model of the Gaussian process model; and training the inverse model by using the Pareto optimal solution set and the corresponding objective value set as training samples.

[0013] In the above technical solution, the process of using support vector regression to train a model on the target value set to predict the target value in the new environment includes: based on the changes in the new environment, selecting the cycle time of the workstation or the number of robots used in the fixed target value set, and predicting the target value parameter of one of them.

[0014] In the above technical solution, the process of using support vector regression to train a model on the target value set to predict the target value in the new environment includes: introducing a hyperparameter q; and training the model from historical sequences. We obtain (tq) samples to form a training set, expressed as follows:

[0015]

[0016] The set of target values ​​in all historical environments of the disassembly line is represented as {POF(1), POF(2), ..., POF(t)}; the set of cycle times of the workstations or the number of robots used is represented as: Each target set contains N target vectors; t represents the number of target value sets; x i Indicates the model input, y i This represents the model output.

[0017] In the above technical solution, the process of using the Pareto optimal solution set and the corresponding target value set as training samples includes:

[0018] The Pareto optimal solution set for all historical environments of the dismantling line is represented as follows:

[0019] {POS(1), POS(2), ..., POS(t)}, each optimal solution set is represented as... N represents the population size, and d represents the dimension of each solution; Used as training samples.

[0020] The beneficial effects of this invention are as follows: In actual industrial disassembly environments, similar products often exhibit similar quality. Support Vector Regression (SVR) can effectively identify the variation patterns of workstation cycle times in similar disassembly environments, predicting workstation cycle times with low error without requiring information about the new environment. In SVR, a nonlinear mapping is introduced to map the input data to a higher-dimensional feature space, making them linearly correlated in the higher-dimensional space. SVR adds slack variables to SVM, thereby relaxing the interval requirements of the function and increasing the generalization ability of the regression model. In this invention, the optimization target value can be calculated through the solution vector, and the mapping relationship between the target value and the solution vector is established through the Gaussian Process Regression (GPR) model. This allows the invention to generate a high-quality initial population before the arrival of a new environment, greatly accelerating the evolution process of the new environment and improving the quality of the solution.

[0021] Furthermore, in this invention, the target space is composed of the total number of robots used and the cycle time of the workstation. When the environment changes, it often only affects a certain dimension of the target. This invention improves the overall calculation speed of the model by fixing the unaffected target value parameters.

[0022] Furthermore, premature values ​​in the historical context may contribute little to the prediction. This invention introduces a hyperparameter to control the number of relevant historical context values. Since only q previous values ​​are associated with the next value, more samples can be obtained for model training by sliding the context window forward.

[0023] Furthermore, this invention employs a random grouping method, which on the one hand reduces the number of Gaussian models to be built, thus lowering computational complexity. On the other hand, for decision variables assigned to the same group, these correlations can be implicitly considered, thereby mitigating the inaccuracies caused by the independence assumptions of decision variables required when decomposing the multivariate probability distributions of m inputs and n outputs into univariate probability distributions. Attached Figure Description

[0024] Figure 1 This is a flowchart of the method of the present invention;

[0025] Figure 2 This is a schematic diagram illustrating the principle of the SVR model;

[0026] Figure 3 This is a schematic diagram illustrating the principle of the Gaussian inverse model.

[0027] Figure 4 Here is an example of disassembling a product: a cross-sectional view of a flashlight;

[0028] Figure 5 This is a diagram showing the disassembly sequence of the flashlight in environment l and the priority relationship between disassembly operations.

[0029] Among them, 1-cover, 2-glass, 3-light bulb, 4-head shell, 5-main shell, 6-spring, 7-battery. Detailed Implementation

[0030] The present invention will be further described in detail below with reference to the accompanying drawings and specific embodiments to facilitate a clear understanding of the present invention, but these descriptions do not constitute a limitation on the present invention.

[0031] like Figure 1 As shown, this invention provides a dynamic disassembly line balancing method based on support vector regression and Gaussian inverse model, comprising the following steps:

[0032] S1, obtain the Pareto optimal solution set and the corresponding target value set for all historical environments of the disassembly line; a single optimal solution in the Pareto optimal solution set includes the robot-to-workstation allocation vector, the task-to-robot allocation vector, and the task sequence vector; a single target value in the target value set includes the workstation cycle time and the number of robots used.

[0033] S2 uses support vector regression to train a model on the target value set in order to predict the target value in the new environment;

[0034] S3, using the Pareto optimal solution set and the corresponding objective value set as training samples, establishes an inverse model based on a Gaussian process;

[0035] S4, input the target value under the new environment into the inverse model to obtain the initial population under the new environment; each individual in the initial population includes 3 solution vectors and 1 target vector; the solution vector includes the robot-to-workstation allocation vector, the task-to-robot allocation vector and the task sequence vector, and the target vector consists of the workstation cycle time and the number of robots used;

[0036] S5, perform evolutionary calculations on the generated initial population and output the optimal solution under the new environment. The evolutionary algorithm can be any population-based multi-objective optimization algorithm. In this invention, the NSGAII algorithm is used for evolutionary calculations. The NSGAII algorithm is derived from [H.Li and Q.Zhang, "Multiobjective Optimization Problems With Complicated Pareto Sets, MOEA / D and NSGA-II," in IEEE Transactions on Evolutionary Computation, vol.13, no.2, pp.284-302, April 2009.].

[0037] The purpose of this invention is to provide a dynamic disassembly line balancing method based on support vector regression and the inverse Gaussian process model. This allows for rapid response to environmental changes when product quality, operator efficiency, etc., evolve, enabling the replanning of the disassembly scheme to track the time-varying Pareto optimal set, thereby improving disassembly efficiency.

[0038] In this invention, the Pareto solution consists of three vectors: the robot-to-workstation assignment vector RW, the task-to-robot assignment vector TR, and the task sequence vector TS. RW = [j1…,j…] r …,j |R| [] is an integer array of length |R| used to determine the robot's allocation to the workstation, where j rLet ∈ J∪{-1}, r∈R, where J and R are the sets of workstation indices and robot indices, respectively. The r-th element j in RW... r A value of "-1" indicates that robot r has not been assigned to any workstation. TR = [r1, ..., r d …,r |D| [] is an array of |D| integer values ​​used to determine the task to be assigned to the robot, where r d ∈R, d=1…,|D|. d represents the set of all selected disassembly operations. Where B represents the set of all disassembly operations. TS = [b1, ..., b d …,b |D| [] represents the permutation of the selected operation |D| set. The order in which each selected operation appears in TS indicates its processing priority. Note that the processing priority of operations in TS cannot violate the priority constraint of disassembling all disassembled products. Furthermore, elements at the same position in TS and TR are in one-to-one correspondence. In other words, operation b in TS... d Robot r assigned to TR d To execute. In this invention, the objective value of the dynamic disassembly line balancing problem optimization consists of the workstation cycle time and the number of robots used. In this invention, the initial population size is 100, and each individual in the population consists of three solution vectors (RW, TR, TS) and one objective vector, where the objective vector consists of the workstation cycle time and the number of robots used.

[0039] It is worth noting that the target space in this invention is composed of the total number of robots used and the cycle time of the workstations. Within the same batch of products, the objective dimension of the total number of robots is often the same across non-dominated solutions in different environments. For example, in environment 0, the total number of robots in its non-dominated solution set is 4, 5, 6, 7, and 8. In environment 1, changes in the quality of some products affect the time it takes for robots to disassemble product components. In other words, changes in product quality often only affect the type of robot assigned, not the number of robots assigned.

[0040] Therefore, in step S2 of this specific embodiment, the total number of robots is fixed as a target quantity to predict the cycle time of the workstation in the new environment. This invention assumes that the stored target set obtained in the previous environment is represented as {POF(1), POF(2), ..., POF(t)}, where each target set contains N target vectors, and N represents the population size. Each individual in the population consists of a solution vector and a target vector. That is... However, not all previous values ​​are consistent with Strong correlation, meaning that premature values ​​in the historical context may have an impact on... The predictions contribute almost nothing. Therefore, a hyperparameter q is introduced to control the relationship between the prediction and the prediction. The number of relevant historical contexts, represented by the input as The output is This can constitute a sample (x) t-q+1 y t-q+1 ).

[0041] Since only q previous values ​​are related to the next value, more samples can be used for training if the context window is slid forward. From historical sequences... We obtain (tq) samples to train the SVR model. The training set can be represented by the following formula:

[0042]

[0043] It should be noted that, due to the characteristics of the problem, the number of robots is fixed in this specific embodiment experiment. However, it is not necessary to strictly adhere to a particular target. For example, in other new environments, the historical robot count can be used to predict the robot count in the new environment, or the historical cycle time can be used to predict the cycle time in the new environment. In other problems, any set of targets can be selected to predict their values ​​in the new environment, based on the characteristics of the problem.

[0044] SVR (Separate Dynamic Regression) is an important application branch of SVM (Separate Dynamics Machine). In the real world, nonlinear correlations between historical and current solutions are more common in practice. Because nonlinear regression problems are more difficult to handle, SVR is primarily used to address them. In SVR, a nonlinear mapping is introduced to map the input data to a higher-dimensional feature space, making them linearly correlated in that space. A specific implementation in SVM aims to find a separating hyperplane by maximizing the margin, such that the vast majority of sample points lie outside the two decision boundaries. SVR also considers maximizing the margin, but it focuses on points within the decision boundaries, aiming to keep as many sample points as possible within the margin. Figure 2 As shown. The final optimization objective can be transformed into the following dual problem, as shown in the formula below:

[0045]

[0046]

[0047] In the formula, α i Let denote the Lagrange multiplier, and C denote the penalty coefficient.

[0048] Use a trained SVR model to predict target values ​​in a new environment.

[0049] like Figure 3 As shown, step S3 specifically includes the following steps:

[0050] (3.1) Determine the parameters of the Gaussian process model.

[0051] Because the Gaussian distribution is extremely common in nature, we can consider the original n y's as following a Gaussian distribution, and then introduce a new X. n+1 These n+1 y's still follow a joint normal distribution. A characteristic of Gaussian processes is that for each x, there is a corresponding Gaussian distribution, and for a set X = {x1, x2, ..., x...}, the distribution follows a Gaussian distribution. t}, that is, there exists a joint Gaussian distribution that satisfies in.

[0052]

[0053] X = {x1, x2, ..., x} t Original distribution:

[0054]

[0055] If a new sample x(t+1) arrives, then the Gaussian distribution is:

[0056]

[0057] Where k = [k(x) t+1 ,x1) k(x t+1 x2) ... k(x t+1 x t )]

[0058] Then the posterior distribution of f(t+1) is calculated as follows:

[0059]

[0060] μ t (x t+1 )=k T K -1 f 1：t

[0061]

[0062] For the covariance matrix K above, k(x,y) represents the kernel function. The kernel function is the core of a Gaussian process, determining its properties. The kernel function generates a covariance matrix (correlation coefficient matrix) in the Gaussian process to measure the "distance" between any two points. Different kernel functions have different measurement methods, resulting in different properties of the Gaussian process. Because the Gaussian kernel function has more complex and diverse boundaries, it can most accurately distinguish data samples, and the numerically calculated K value fluctuates less. Therefore, the Gaussian kernel function is chosen as the kernel function for the Gaussian process; it is also called the radial basis function (RBF). Its basic form is as follows:

[0063]

[0064] Where σ and l are the hyperparameters of the Gaussian kernel.

[0065] Hyperparameter tuning: Maximize the probability of y occurring under these two hyperparameters. The optimal parameters are found by maximizing the marginal log-likelihood, which is expressed as:

[0066]

[0067] (3.2) The multivariate conditional probability distribution of m inputs and n outputs is decomposed into m×n univariate conditional probability distributions. Theoretically, m different inverse models (probability distribution models) can be built for each decision variable, which would make the algorithm computationally very intensive. To solve this problem, a random grouping method is used to reduce the number of Gaussian models to be built. Given m objectives, m groups of inverse models are built, where the j-th (1≤j≤m) group of models uses the index of the j-th objective f j As variables. In each group, L decision variables will be randomly assigned to it to use f j Used as variables to construct the inverse model, where L << n.

[0068] In this invention, the total number of robots does not usually change with the environment, so it is only necessary to establish a 1×n univariate conditional probability distribution with the disassembly workstation cycle time as the objective.

[0069] This invention assumes that the Pareto solution set stored in the preceding environment is represented as {POS(1), POS(2), ..., POS(t)}, where each solution set contains N solution vectors. N represents the population size, and d represents the dimension of each solution. Use these samples to train the Gaussian inverse model.

[0070] This invention uses Visual Studio to implement the proposed novel method as an executable program; test instances are generated from eight D-TAOGs, each based on one of eight existing TAOGs. All ordinary nodes in these eight TAOGs have three states: normal, damaged, and missing. A TAOG refers to the disassembly sequence of a product and the priority relationships between all disassembly operations within that sequence. D-TAOGs, based on TAOGs, assign three states to the ordinary nodes of all TAOGs. For ordinary node B... b Processing time in damaged state exist The time is randomly generated between these parameters, and the processing time in the lost state is... exist It is randomly generated from among them. B represents the normal state b The processing time. To better understand the dynamic disassembly line balancing model under uncertain product quality, consider a small-scale example.

[0071] like Figure 4 As shown. In this case, the flashlight is completely detached from the disassembly line. Figure 5 A cross-sectional view of the flashlight and its D-TAOG plot in the first environment are shown. Figure 4 The flashlight in the image consists of seven parts and their connections. For example... Figure 5 As shown, the sub-components consisting of parts and disassembly operations are modeled as artificial nodes (labeled A). a ) and ordinary nodes (labeled as B) b The number sequence next to each artificial node indicates the set of parts that make up the component. For example, artificial node A1 represents sub-assembly "3 / 7", which consists of bulb 3, head housing 4, main housing 5, spring 6, and battery 7. Ordinary node B1 indicates that flashlight "1 / 7" (A0) is broken down into sub-assembly "3 / 7" (A1) and sub-assemblies "1,2" (A5). For simplicity, sub-assemblies with only one part are not shown. Figure 5 This includes all possible disassembly sequences, such as [B1, B4, B7, B8, B9, B10] and [B2, B5, B7, B8, B9, B10], etc. Product quality inspection allows us to determine the state of each part in every product under any environment, as well as the state of each common node. For example, in environment l, since part 4 (head housing) is detected as damaged, related parts B4 and B6 are also determined to be damaged. The processing time for common nodes varies depending on the state.

[0072] Based on these eight D-TAOGs, a set of test instances with different scales and environmental similarities were constructed. Specifically, two or three and four or five D-TAOGs were randomly selected from the eight to form small-scale and large-scale instances, respectively. Specifically, six small-scale instances (P1-P6) and six large-scale instances (P7-P12) were randomly generated.

[0073] The specific implementation further expands the test instance set regarding environmental similarity. Each instance is specified to contain ten time-varying disassembly environments (l = 0..., 9). The state of some common nodes in the D-TAOGs may differ in different environments, resulting in varying processing times. For each environment in the test instance, the specific implementation randomly places 5% and 5% of the common nodes in the corresponding D-TAOG into damaged and missing states, respectively. The remaining common nodes remain in a normal state, thus forming 12 high-similarity test instances (P1H-P12H).

[0074] Similarly, in this specific embodiment, the proportions of normal nodes in damaged and missing states are adjusted to 10% and 10%, respectively, thus constructing 12 medium similarity test instances (P1M-P12M). The specific embodiment further adjusts the proportions of the two normal nodes to 15% and 15%, constructing 12 low similarity test instances (P1L-P12L). Ultimately, the generated instance set contains 36 test instances with different sizes and environmental similarities.

[0075] To analyze the performance of the Dynamic Disassembly Line Equilibrium Method (DT-DMOEA) based on support vector regression and Gaussian process inverse model of the present invention and its comparative algorithms, a specific embodiment uses two metrics to compare their results: mean reverse generation distance (MIGD) and mean hypervolume (MHV). MIGD and MHV are variants of IGD and HV, respectively. MIGD is defined as the average IGD value obtained by the algorithm in all environments, and MHV is defined as the average HV value in all environments.

[0076] IGD is the average distance from the obtained solution to the nearest solution in the optimal solution set. The smaller the IGD of the solution set, the better the approximation and distribution. Note that the optimal solution set for the problem considered in the specific implementation cannot be obtained in advance. Therefore, the Dynamic Disassembly Line Equilibrium Method (DT-DMOEA) based on support vector regression and Gaussian process inverse model and its comparative algorithm are run independently 30 times, with each iteration evolving for 200 generations. All the obtained solution sets are merged, and then all the non-dominated solutions are taken as the optimal solution set. In addition, the objective function values ​​of all solutions are normalized to [0,1], and then the IGD value is calculated. HV mainly calculates the volume formed by the obtained solutions and the reference point containing the two maximum objective values ​​in all solution sets. The larger the hypervolume of the solution set, the better the approximation and distribution. When calculating the hypervolume, the objective function values ​​of all solutions are normalized to [0,1], and the reference point is set to (1.1,1.1).

[0077] The proposed Dynamic Disassembly Line Balancing Method (DT-DMOEA) based on support vector regression and Gaussian process inverse model is compared with several state-of-the-art dynamic multi-objective optimization algorithms, with the better average values ​​for each instance highlighted in bold.

[0078] In the table, and The results show that the proposed method significantly outperforms and is comparable to its competitors. Table 1 shows the performance of the proposed method DT-DMOEA in MIGD across small-scale test instances. Table 1 shows that the proposed method performs better in five, five, and four out of six small-scale test instances with high, medium, and low similarity, respectively. For MIGD, the proportions in which the proposed method outperforms MOEA / D-SVR, DMOEA-DVC, SGEA, KT-DMOEA, and Tr-RM-MEDA are 18 / 18, 13 / 18, 15 / 18, 15 / 18, and 18 / 18, respectively.

[0079] The MOEA / D-SVR algorithm is derived from [L.Cao, L.Xu, ED.Goodman, C.Bao, and S.Zhu, “Evolutionary dynamic multiobjective optimization assisted by a support vector regression predictor,” IEEE Trans. Evol. Comput., vol.24, no.2, pp.305-319, 2020.].

[0080] The DMOEA-DVC algorithm originates from [Z.Liang,T.Wu,X.Ma,Z.Zhu and S.Yang, “A Dynamic Multiobjective Evolutionary Algorithm Based on Decision Variable Classification,”IEEE Trans.Cybern.,doi:10.1109 / TCYB.2020.2986600,2020.].

[0081] The SGEA algorithm originates from [S. Jiang and S. Yang, “A steady-state and generational evolutionary algorithm for dynamic multi-objective optimization,” IEEE Trans. Evol. Comput., vol. 21, no. 1, pp. 65-82, 2017.].

[0082] The KT-DMOEA algorithm comes from [M.Jiang, Z.Wang, H.Hong, and GGYen, "Knee point-based imbalanced transfer learning for dynamic multiobjective optimization," IEEE Trans.Evol.Comput., vol.25, no.1, pp.117-129, 2021.],

[0083] The TrRM-MEDA algorithm originates from [M.Jiang, Z.Huang, L.Qiu, W.Huang, and GGYen, "Transfer learning-based dynamic multiobjective optimization algorithms," IEEE Trans. Evol. Comput., vol.22, no.4, pp.501-514, 2018.].

[0084] Table 1. Mean and standard deviation of MIGD results for DT-DMOEA and competitors in small-scale test cases.

[0085]

[0086] Table 2 shows the MIGD performance of the proposed DT-DMOEA method in large-scale test instances. Table 2 demonstrates that the proposed method outperforms competitors in all large-scale test instances. For MIGD, the proportions in which the proposed method outperforms MOEA / D-SVR, DMOEA-DVC, SGEA, KT-DMOEA, and Tr-RM-MEDA are 18 / 18, 18 / 18, 17 / 18, 15 / 18, and 18 / 18, respectively.

[0087] Table 2 shows the mean and standard deviation of MIGD results for DT-DMOEA and its competitors in large-scale test instances.

[0088]

[0089] Table 3 shows the MHV of the proposed method DT-DMOEA in small-scale test instances. Table 3 shows that the proposed method performs better in four, six, and five out of six small-scale test instances with high, medium, and low similarity, respectively. For MHV, the proportions of test instances where the proposed method outperforms MOEA / D-SVR, DMOEA-DVC, SGEA, KT-DMOEA, and Tr-RM-MEDA are 18 / 18, 15 / 18, 14 / 18, 18 / 18, and 18 / 18, respectively.

[0090] Table 3 shows the mean and standard deviation of MHV results for DT-DMOEA and its competitors in small-scale test cases.

[0091]

[0092] Table 4 shows the MHV of the proposed method DT-DMOEA in large-scale test instances. Table 4 shows that the proposed method performs better in five, six, and six of the six high-, six medium-, and six low-similarity small-scale test instances, respectively. For MHV, the proportions of test instances where the proposed method outperforms MOEA / D-SVR, DMOEA-DVC, SGEA, KT-DMOEA, and Tr-RM-MEDA are 18 / 18, 18 / 18, 14 / 18, 18 / 18, and 18 / 18, respectively.

[0093] Table 4 shows the mean and standard deviation of MHV results for DT-DMOEA and its competitors in large-scale test instances.

[0094]

[0095]

[0096] In summary, the method proposed in this invention can not only significantly improve the quality of the obtained solution, but also significantly enhance the responsiveness to environmental changes, thereby effectively addressing various uncertainties in actual industrial applications and improving disassembly efficiency.

[0097] The contents not described in detail in this specification are existing technologies known to those skilled in the art.

Claims

1. A dynamic disassembly line balancing method based on support vector regression and Gaussian inverse model, characterized in that: Includes the following steps: Obtain the Pareto optimal solution set and the corresponding target value set for all historical environments of the disassembly line; a single optimal solution in the Pareto optimal solution set includes the robot-to-workstation assignment vector, the task-to-robot assignment vector, and the task sequence vector; a single target value in the target value set includes the workstation cycle time and the number of robots used. The model is trained using support vector regression to predict target values ​​in a new environment. Using the Pareto optimal solution set and the corresponding objective value set as training samples, an inverse model based on Gaussian processes is established. The target value under the new environment is input into the inverse model to obtain the initial population under the new environment; each individual in the initial population includes 3 solution vectors and 1 target vector; the solution vectors include the robot-to-workstation allocation vector, the task-to-robot allocation vector, and the task sequence vector; the target vector consists of the workstation cycle time and the number of robots used. Evolutionary calculations are performed on the generated initial population to output the optimal solution under the new environment.

2. The method according to claim 1, characterized in that: The process of establishing an inverse model based on a Gaussian process, using the Pareto optimal solution set and the corresponding objective value set as training samples, includes: determining the parameters of the Gaussian process model; constructing the inverse model of the Gaussian process model; and training the inverse model using the Pareto optimal solution set and the corresponding objective value set as training samples.

3. The method according to claim 2, characterized in that: The process of training a model using support vector regression to predict target values ​​in a new environment includes: based on the changes in the new environment, selecting the cycle time of a workstation or the number of robots used in a fixed set of target values, and predicting the target value parameters of one of them.

4. The method according to claim 3, characterized in that: The process of training a model using support vector regression to predict target values ​​in a new environment includes: introducing a hyperparameter q; and training the model from historical sequences. We obtain (tq) samples to form a training set, expressed as follows: The set of target values ​​in all historical environments of the disassembly line is represented as {POF(1), POF(2), ..., POF(t)}; the set of cycle times of the workstation or the number of robots used is represented as: Each target set contains N target vectors; t represents the number of target value sets; x i Indicates the model input, y i This represents the model output.

5. A method according to claim 4, characterized in that: The process of using the Pareto optimal solution set and the corresponding objective value set as training samples includes: The Pareto optimal solution set for all historical environments of the dismantling line is represented as follows: {POS(1), POS(2), ..., POS(t)}, each optimal solution set is represented as... N represents the population size, and d represents the dimension of each solution; Used as training samples.

Images