A multi-working-condition complex assembly line resource configuration evaluation method based on instant learning

Abstract

The application provides a multi-working-condition complex assembly line resource configuration evaluation method based on instant learning, which comprises the following steps: determining key resource variables affecting the production rhythm based on a complex assembly line production mode and a process flow, preparing complex assembly line resource configuration data to construct a sample database, dividing the working conditions of the complex assembly line resource configuration by adopting a Kmeans clustering method, classifying resource configuration schemes with similar production characteristics, constructing a LWPLS-Kmeans dynamic fusion algorithm model based on instant learning, training and evaluating the model by using the resource configuration data, and inputting different complex assembly line resource configuration schemes to be evaluated into the resource configuration evaluation model to accurately evaluate the performance of various resource configuration schemes. The application introduces an instant learning algorithm and a Kmeans clustering algorithm, optimizes a similarity sample subset screening mechanism, establishes a dynamic fusion resource configuration evaluation model based on the LWPLS-Kmeans instant learning algorithm, realizes accurate prediction of the production rhythm, and thus improves the accuracy of resource configuration evaluation.

Description

Technical Field

[0001] This invention belongs to the field of resource allocation evaluation for complex assembly lines, and specifically designs a multi-condition complex assembly line resource allocation evaluation method based on real-time learning. Background Technology

[0002] The intelligent transformation of the manufacturing industry is accelerating, and market demands are changing rapidly. As a core production unit in high-end manufacturing, the ability of complex assembly lines to quickly respond to changes in market demand directly impacts a company's competitiveness. Cycle Time (CT) is a key indicator that directly reflects the rationality of resource allocation and is often used to evaluate the effectiveness of resource allocation.

[0003] However, due to the nonlinear coupling relationship and multi-condition characteristics between complex assembly line resource allocation data and production cycle time, it poses a challenge to evaluate resource allocation by accurately predicting production cycle time. The published literature "Sun Jinhao, Yang Yi, Du Rui, et al. Design and analysis of closed-loop control model for aircraft final assembly line cycle time [J]. Aviation Manufacturing Technology, 2023, 66(8): 38–46." constructs an aircraft final assembly line capacity prediction model using a Markov model and feeds the prediction data back to the cycle time control system in real time, thereby adjusting the production plan by adjusting resource allocation. Although traditional discrete event simulation methods can evaluate the optimal resource allocation scheme based on production cycle time, the algorithm in this published literature has high complexity and low solution efficiency. To overcome this limitation, advanced technologies such as machine learning and digital twin algorithms are used. The published paper "Zhang Qi, Jiang Changjian, Han Jiawei, et al. Optimization of Material Configuration Scheme for Aircraft Assembly Stations Based on GRU Neural Network and Genetic Algorithm [J]. Aviation Manufacturing Technology, 2024, 67(17): 78-82, 92" uses a GRU neural network to construct a simulation proxy model for evaluating material configuration schemes. Taking the material configuration scheme as input and the expected completion time of the assembly run and the average dwell time of key materials as outputs, it efficiently evaluates the material configuration scheme. Although this published paper has achieved significant results in reducing computational complexity and improving the accuracy of resource configuration evaluation, the global modeling method is difficult to adapt to the dynamic changes of complex assembly lines and cannot maintain accurate evaluation precision for all resource configurations. Therefore, it is evident that the accuracy of existing resource configuration evaluation methods for complex assembly lines still has room for further improvement. Summary of the Invention

[0004] To address the insufficient accuracy of resource allocation assessment in existing technologies for complex assembly lines with multiple operating conditions, this invention proposes a resource allocation assessment method for complex assembly lines with multiple operating conditions based on real-time learning. This method introduces real-time learning algorithms and K-means clustering algorithms, and optimizes the similarity sample subset selection mechanism to construct a dynamic fusion resource allocation assessment model based on real-time learning and LWPLS-Kmeans. By employing a dynamic weighted fusion strategy for prediction results, this model can achieve accurate prediction of production cycle time, thereby improving the accuracy of resource allocation assessment.

[0005] The technical solution of this invention is as follows:

[0006] A resource allocation evaluation method for complex assembly lines under multiple working conditions based on real-time learning includes the following steps:

[0007] Step 1: Based on the production mode and process flow of the complex assembly line, identify the key resource variables that affect the production cycle time, prepare resource configuration data for the complex assembly line, and construct a sample database.

[0008] Step 2: Use the K-means clustering method to classify the working conditions of the resource allocation of the complex assembly line and categorize the resource allocation with similar production characteristics.

[0009] Step 3: Construct a resource allocation evaluation model based on real-time learning and local weighted partial least squares-k-means clustering dynamic fusion, and train and optimize the evaluation model using resource allocation data;

[0010] Step 4: Input the resource allocation schemes for different complex assembly lines into the evaluation model in Step 3 to obtain the production cycle prediction results for each resource allocation scheme. Based on the production cycle prediction results, the effectiveness of each resource allocation scheme can be accurately evaluated.

[0011] Furthermore, in step 1, the Pearson correlation coefficient method is used to analyze the correlation of resource characteristics in complex assembly lines and to evaluate the correlation between different resources and production cycle time. Based on the ranking results of feature correlation, the variable combinations of resource configurations with higher correlation rankings are selected as input variables for the resource configuration evaluation model. Based on the actual operation process, a complex assembly line simulation model is established using discrete-time simulation software, and the simulation is run to obtain the production cycle time of different resource configurations.

[0012] Furthermore, in step 2, the resource configuration sample dataset X=[x1,x2,…,x] is processed. N Standardization preprocessing is performed, where sample x i =[x i1 ,x i2 ,…,x imLet ] represent the i-th resource configuration sample of a complex assembly line, N be the total number of samples, and m be the dimension of the resource configuration variable. The resource configuration sample dataset is clustered, and the silhouette coefficient is used as the evaluation index. By comparing the clustering quality under different clustering numbers k, the optimal number of clusters is determined. K-means clustering is then performed on the standardized resource configuration data using the optimal clustering number k to obtain the work condition division result {G1, G2, ..., G...}. k}

[0013] Furthermore, the expression for the contour coefficient is:

[0014]

[0015] in For sample x i The average intra-cluster distance; For sample x i The average distance of the nearest neighbor cluster.

[0016] Furthermore, the objective function for K-means clustering is:

[0017]

[0018] Where k is the number of clusters; v j C is the center vector of the j-th cluster; j Let represent the sample set of the j-th cluster.

[0019] Furthermore, in step 3, a resource configuration dataset D=[X,Y] is constructed and standardized, where Y=[y1,y2,…,y…]. N ], y i Let i be the production cycle time of the complex assembly line for the i-th sample;

[0020] For the evaluation model, in the offline phase, the optimal parameter combination (h, R, p) is determined by minimizing the prediction error using a standardized resource allocation dataset and a leave-one-out cross-validation method combined with a grid search method; where h and R are the core parameters of the local partial least squares algorithm, and p is the distance threshold.

[0021] During the online phase, for each resource configuration sample, the Euclidean distance between the sample and each cluster center is calculated, and a similar sample subset Q=(Q1,Q2,…,Q) is obtained by filtering through a distance threshold p. s ), and for each similar sample subset Q j A local weighted partial least squares model is used to construct an instant learning model, and the corresponding production cycle prediction value y is output. j Based on the predicted value y j Euclidean distance d from the cluster center to the production cycle time jThe prediction results of the real-time learning model corresponding to each similar sample subset are weighted and fused to output the predicted production cycle y.

[0022] Furthermore, in step 3, the process of predicting the production cycle time using a locally weighted partial least squares model is as follows:

[0023] Step 3.1: According to the formula

[0024]

[0025] Calculate the distance D = [d1, d2, ..., d] between the predicted point and each training sample. N ], and then through the formula

[0026]

[0027] Calculate the weights corresponding to each training sample, and then construct a diagonal matrix based on the weight calculation results. :

[0028]

[0029] Where x q =[x q1 ,x q2 ,…,x qm ], x q Assign assembly line resources for the predicted points; d i For the prediction point x q and training sample x i The Euclidean distance between them;

[0030] Step 3.2: According to the formula

[0031] X r = X − 1 n [ x 1 ¯ , x 2 ¯ , x 3 ¯ , … , x m ¯ ]

[0032]

[0033] x q r = x q − [ x 1 ¯ , x 2 ¯ , x 3 ¯ , … , x m ¯ ]

[0034]

[0035]

[0036] Perform data mean centering processing, where and The weighted average of the input and output of the training samples, 1 n Let X be an n-dimensional column vector of all 1s. r Y r and x q,rThese are the input, output, and predicted point vectors of the centered training sample, respectively.

[0037] Step 3.3: Input the training samples and the predicted samples x respectively. q,r Extracting latent variables:

[0038]

[0039]

[0040] In the formula, t r and t q,r They are X r and x q,r The corresponding r-th latent variable; w r yes The eigenvector corresponding to the largest eigenvalue;

[0041] According to the formula

[0042]

[0043]

[0044] Calculate the prediction point x q The load vector p r and the regression coefficient vector q of the model r Establish a regression equation

[0045]

[0046] Update the prediction point output; if the current number of latent variables r reaches the set value, output the predicted production cycle time value of the prediction point. Otherwise, let r = r + 1, and use the formula

[0047]

[0048] After updating the input, output, and predicted point vectors, return to step 3.3 to continue iterating.

[0049] Furthermore, in step 3, the process of filtering for a subset of similar samples based on the distance threshold p is as follows:

[0050] Through formula

[0051]

[0052] Calculate the Euclidean distance between the predicted point and each cluster center, and compare it with the distance threshold p;

[0053] If the distance is within the threshold range, then the corresponding cluster dataset is selected as the similar sample subset Q=(Q1,Q2,…,Q…) for the prediction point. s):

[0054]

[0055] .

[0056] Furthermore, in step 3, based on the predicted value y j Euclidean distance d from the cluster center to the production cycle time j The process of weightedly fusing the prediction results of the real-time learning model corresponding to each similar sample subset and outputting the predicted production cycle y is as follows:

[0057]

[0058] in The value is a very small positive value; the prediction results of the real-time learning model corresponding to each similar sample subset are multiplied by their corresponding weights, and then weighted and summed to obtain the final production cycle prediction value:

[0059] .

[0060] Beneficial effects

[0061] The advantages of this invention are:

[0062] 1. This method introduces the instant learning algorithm and K-means clustering algorithm to construct a resource allocation evaluation model, which can accurately predict the production cycle of resource allocation in complex assembly lines and further improve the accuracy of resource allocation evaluation.

[0063] 2. This invention employs a locally weighted partial least squares model for production cycle time prediction. In the similarity sample selection stage, a subset of samples with high similarity is filtered using a distance threshold p, and a locally weighted partial least squares model is constructed based on the filtered subsets. The production cycle time prediction results are dynamically weighted and fused using the Euclidean distance between the sample point to be tested and the cluster center. This invention achieves high-precision prediction of production cycle time, improving the accuracy of resource allocation assessment for complex assembly lines.

[0064] Additional aspects and advantages of the invention will be set forth in part in the description which follows, and in part will be obvious from the description, or may be learned by practice of the invention. Attached Figure Description

[0065] The above and / or additional aspects and advantages of the present invention will become apparent and readily understood from the description of the embodiments taken in conjunction with the following drawings, in which:

[0066] Figure 1 This invention provides a flowchart of a multi-condition complex assembly line resource allocation evaluation method based on real-time learning.

[0067] Figure 2 The present invention provides a flowchart for constructing a complex assembly line resource allocation evaluation model using the LWPLS-Kmeans dynamic fusion algorithm based on real-time learning.

[0068] Figure 3 This invention provides a schematic diagram of the final assembly line process for a certain type of aircraft.

[0069] Figure 4 The resource allocation variable correlation coefficient diagram provided by this invention;

[0070] Figure 5 The resource configuration input variable diagram provided by this invention;

[0071] Figure 6 This invention provides a sample library of eight resource configurations.

[0072] Figure 7 The model parameter setting diagram provided by this invention;

[0073] Figure 8 The present invention provides six different model evaluation index diagrams;

[0074] Figure 9 Comparison of evaluation results of SVM model (left) and LSTM model (right) under 8 resource allocation schemes provided by this invention;

[0075] Figure 10 Comparison of evaluation results of PLS ​​model (left) and LWPLS model (right) under 8 resource allocation schemes provided by this invention;

[0076] Figure 11 Comparison of the evaluation results of the LWPLS-Kmeans model (left) and the model of this invention (right) under 8 resource allocation schemes provided by this invention; Detailed Implementation

[0077] The embodiments of the present invention are described in detail below. These embodiments are exemplary and intended to explain the present invention, and should not be construed as limiting the present invention.

[0078] Please see Figure 1 This invention provides a method for evaluating resource allocation in complex assembly lines under multiple working conditions based on real-time learning, comprising the following steps:

[0079] Step 1: Based on the complex assembly line production mode and process flow, identify the key resource variables that affect the production cycle time, prepare complex assembly line resource configuration data, and construct a sample database.

[0080] In this embodiment, the Pearson correlation coefficient method is used to analyze the correlation of resource characteristics in complex assembly lines and to evaluate the correlation between different resources and production cycle time, including human resources and equipment resources.

[0081] Based on the feature correlation ranking results, the variable combinations of resource allocation with higher correlation rankings are selected as input variables for the resource allocation evaluation model.

[0082] Based on the actual operation process, a complex assembly line simulation model was established using discrete-time simulation software, and simulation was run to obtain the production cycle time with different resource configurations; thus, a sample database was constructed.

[0083] Step 2: Use the K-means clustering method to classify the resource allocation of complex assembly lines into different working conditions and categorize resource allocations with similar production characteristics.

[0084] In this embodiment, the resource configuration sample dataset X=[x1,x2,…,x] is used. N Standardization preprocessing is performed, where sample x i =[x i1 ,x i2 ,…,x im Let ] represent the i-th resource configuration sample in a complex assembly line, N be the total number of samples, and m be the dimension of the resource configuration variable. In this embodiment, the silhouette coefficient is used as an evaluation metric. By comparing the clustering quality under different values ​​of the number of clusters k, the optimal number of clusters is determined. The expression for the silhouette coefficient is:

[0085] (1)

[0086] In formula (1), For sample x i The average intra-cluster distance; For sample x i The average distance of the nearest neighbor cluster.

[0087] The standardized resource allocation data and the optimal number of clusters k are input into the K-means clustering algorithm. The objective function is minimized through continuous iterative optimization. The objective function of the K-means algorithm is:

[0088] (2)

[0089] In formula (2), k is the number of clusters; v j C is the center vector of the j-th cluster; j Let {G1, G2, ..., Gj} represent the sample set of the j-th cluster. The final output is the load condition partitioning result {G1, G2, ..., Gjj}. k} and cluster center C:

[0090] C = [ c 1 , 1 c 1 , 2 c 2 , 1 c 2 , 2 … c 1 , m … c 2 , m ... c k , 1 … c k , 2 … … … c k , m ]

[0091] Step 3: Construct a resource allocation evaluation model based on real-time learning and dynamically fused locally weighted partial least squares-k-means clustering (LWPLS-Kmeans). Train and optimize the evaluation model using resource allocation data.

[0092] In this embodiment, a resource configuration dataset D=[X,Y] is constructed and standardized, where Y=[y1,y2,…,y N ], y i Let be the production cycle time of the complex assembly line for the i-th sample.

[0093] For the evaluation model, in the offline phase, using the standardized resource allocation dataset, leave-one-out cross-validation combined with grid search is employed to determine the optimal parameter combination (h, R, p) by minimizing the prediction error. Here, h and R are the core parameters of the local partial least squares algorithm, and p is the distance threshold.

[0094] During the online phase, for each resource configuration sample, the Euclidean distance between the sample and each cluster center is calculated, and a similar sample subset Q=(Q1,Q2,…,Q) is obtained by filtering through a distance threshold p. s ), and for each similar sample subset Q j A local weighted partial least squares model is used to construct an instant learning model, and the corresponding production cycle prediction value y is output. j Based on the predicted value y j Euclidean distance d from the cluster center to the production cycle time j The prediction results of the real-time learning model corresponding to each similar sample subset are weighted and fused to output the predicted production cycle time y. After the prediction is completed, the constructed real-time learning model is discarded, and the real-time learning model is reconstructed when new prediction data arrives.

[0095] Specifically, the process of predicting production cycle time using a locally weighted partial least squares model is as follows:

[0096] Step 3.1: According to the formula

[0097] (3)

[0098] Calculate the distance D = [d1, d2, ..., d] between the predicted point and each training sample. N ], and then through the formula

[0099] (4)

[0100] Calculate the weights corresponding to each training sample, and then construct a diagonal matrix based on the weight calculation results. :

[0101] (5)

[0102] In formula (3), x q =[x q1 ,x q2 ,…,x qm ], x q Assign assembly line resources for the predicted points; d i For the prediction point x q and training sample x i The Euclidean distance between them.

[0103] Step 3.2: According to the formula

[0104] X r = X − 1 n [ x 1 ¯ , x 2 ¯ , x 3 ¯ , … , x m ¯ ] (6)

[0105] (7)

[0106] x q r = x q − [ x 1 ¯ , x 2 ¯ , x 3 ¯ , … , x m ¯ ] (8)

[0107] (9)

[0108] (10)

[0109] Data mean centering is performed, where the weighted average of the input and output of the training samples is calculated using equations (9) and (10). and In the formula, 1 n Let X be an n-dimensional column vector of all 1s. r Y r and x q,r These are the input, output, and predicted point vectors of the centered training sample, respectively.

[0110] Step 3.3: Input the training samples and the predicted samples x respectively. q,r Extracting latent variables:

[0111] (11)

[0112] (12)

[0113] In the formula, t r and t q,r They are X r and x q,r The corresponding r-th latent variable; wr yes The eigenvector corresponding to the largest eigenvalue.

[0114] According to the formula

[0115] (13)

[0116] (14)

[0117] Calculate the prediction point x q The load vector p r and the regression coefficient vector q of the model r Establish a regression equation

[0118] (15)

[0119] Update the prediction point output. If the current number of latent variables r reaches the set value, output the predicted production cycle time value of the prediction point. Otherwise, let r = r + 1, and use the formula...

[0120] (16)

[0121] After updating the input, output, and predicted point vectors, return to step 3.3 to continue iterating.

[0122] In step 3, the process of filtering similar sample subsets based on the distance threshold p is as follows:

[0123] Through formula

[0124] (17)

[0125] Calculate the Euclidean distance between the predicted point and each cluster center, and compare it with the distance threshold p.

[0126] If the distance is within the threshold range, then the corresponding cluster dataset is selected as the similar sample subset Q=(Q1,Q2,…,Q…) for the prediction point. s ):

[0127] (18)

[0128] (19)

[0129] In step 3, based on the predicted value y j Euclidean distance d from the cluster center to the production cycle time j The process of weightedly fusing the prediction results of the real-time learning model corresponding to each similar sample subset and outputting the predicted production cycle y is as follows:

[0130] The Euclidean distance d between the predicted point and the corresponding production cycle time of the cluster center. j Calculate the weight of each cluster:

[0131] (20)

[0132] in The value is a very small positive value. The prediction results of the real-time learning model corresponding to each similar sample subset are multiplied by their corresponding weights and then summed in a weighted manner to obtain the final production cycle prediction value.

[0133] (twenty one)

[0134] Step 4: Input the resource allocation schemes for different complex assembly lines into the evaluation model in Step 3 to obtain the production cycle prediction results for each resource allocation scheme. Based on the production cycle prediction results, the effectiveness of each resource allocation scheme can be accurately evaluated.

[0135] See Figures 2-3 This embodiment uses an aircraft final assembly line as an example, which has a total of 5 workstations. The first workstation is responsible for the docking of complex components; the second and third workstations are responsible for the installation of mechanical and special systems and internal equipment; the fourth workstation is responsible for power-on testing; and the fifth workstation is responsible for delivery inspection. During production, work-in-process flows sequentially according to the process route. After completing the work at a single station and undergoing quality inspection, it is transferred to the next workstation by the automated system until the entire assembly process is completed. The tasks of each workstation have different characteristics, and the amount of resources required also varies. Therefore, the rationality of resource allocation directly affects the production cycle and overall efficiency. Based on on-site research, 10 resources that may be unavailable in a timely manner during production are selected as the key resources in this case. Each workstation has a clear functional division, and personnel are generally not allowed to move between workstations.

[0136] The Pearson correlation coefficient method was used to analyze the correlation of 50 characteristics, identifying key resource variables that were significantly correlated with production cycle time. (See also...) Figure 4 Pearson correlation coefficient analysis showed that different resource allocation variables had significantly different impacts on production cycle time. The number of hoisting devices at workstation 1 had the strongest correlation with production cycle time, while the power system testing equipment at workstation 2 had the weakest correlation. Based on the feature correlation ranking results, combinations of 8 and 9 resource allocation variables with high correlation rankings were selected as input variables for the resource allocation evaluation model, with production cycle time as the output variable. For detailed information on each variable, please refer to [link / reference]. Figure 5 Based on the actual operation process, 1880 sets of 8-dimensional and 2500 sets of 9-dimensional resource allocation sample data were obtained using the discrete-time simulation software Simio (operation time 8 hours). The database structure of the 8 types of resource allocation samples is described in [reference needed]. Figure 6 Through system parameter optimization calculations (see model parameter settings), Figure 7 The optimal parameter combinations for the eight variables were ultimately determined to be: k=19, h=0.9, R=5, and p=2.397; the optimal parameter combinations for the nine variables were determined to be: k=19, h=1, R=8, and p=1.929. These parameter combinations will be used to construct the LWPLS-Kmeans dynamic fusion model. In this embodiment, in addition to the resource allocation evaluation methods described above, global models SVM and LSTM, and local models PLS, LWPLS, and LWPLS-Kmeans are also used to establish the resource allocation evaluation model. Please refer to the model result comparison chart. Figure 8 .

[0137] The LWPLS models with 8 resource allocation variables (RMSE=6.5831, R²=0.8838) and 9 resource allocation variables (RMSE=6.3033, R²=0.8677) significantly outperformed the SVM and LSTM models in predicting production cycle time, demonstrating the advantages of local modeling methods in handling complex assembly line resource allocation data. The LWPLS model, by introducing a sample weighting strategy, linearizes the local features of the data, thus significantly improving the prediction performance of production cycle time for both 8 and 9 resource allocation variables. Traditional PLS methods, due to the inherent limitations of linear modeling, struggle to capture the nonlinear relationships in complex assembly line resource allocation data, resulting in poor performance across all three metrics for both the 8 and 9 resource allocation variable PLS models. The LWPLS-Kmeans model, optimized with K-means clustering, shows a significant performance improvement. Based on this, the model of this invention further reduces the RMSE of 8 resource allocation variables and 9 resource allocation variables to 2.7430 and 2.8041 by combining cluster analysis and dynamic fusion strategy.

[0138] To see a more detailed comparison of the performance of various models, 100 samples were randomly selected from the validation sets of 8 resource configurations, and the evaluation results of the 6 models were visually compared in the form of line graphs. Please refer to [link / reference]. Figure 9 , Figure 10 and Figure 11 It can be observed that the model of this invention has the smallest error between the actual value and the predicted value, and the best fitting effect, which proves that the modeling method has certain advantages in the resource allocation evaluation of complex assembly lines under multiple working conditions.

[0139] This embodiment introduces a real-time learning algorithm and a K-means clustering algorithm, and optimizes the similarity sample subset selection mechanism to construct a dynamic fusion resource allocation evaluation model based on real-time learning and LWPLS-Kmeans. By adopting a dynamic weighted fusion strategy for prediction results, this model can achieve accurate prediction of production cycle time, thereby improving the accuracy of resource allocation evaluation. It can be used in complex assembly line resource allocation evaluation scenarios and has broad practical application value.

[0140] Although embodiments of the present invention have been shown and described above, it is understood that the above embodiments are exemplary and should not be construed as limiting the present invention. Those skilled in the art can make changes, modifications, substitutions and variations to the above embodiments within the scope of the present invention without departing from the principles and spirit of the present invention.

Claims

1. A resource allocation evaluation method for complex assembly lines under multiple working conditions based on real-time learning, characterized in that: Includes the following steps: Step 1: Based on the production mode and process flow of the complex assembly line, identify the key resource variables that affect the production cycle time, prepare resource configuration data for the complex assembly line, and construct a sample database. Step 2: Apply the K-means clustering method to classify the resource allocation of complex assembly lines into different working conditions, grouping resource allocations with similar production characteristics; specifically: For the resource configuration sample dataset X=[x1,x2,…,x N Standardization preprocessing is performed, where sample x i =[x i1 ,x i2 ,…,x im Let ] represent the i-th resource configuration sample of a complex assembly line, N be the total number of samples, and m be the dimension of the resource configuration variable. The resource configuration sample dataset is clustered, and the silhouette coefficient is used as the evaluation index. By comparing the clustering quality under different clustering numbers k, the optimal number of clusters is determined. K-means clustering is then performed on the standardized resource configuration data using the optimal clustering number k to obtain the work condition division result {G1, G2, ..., G...}. k The expression for the contour coefficient is: in For sample x i The average intra-cluster distance; For sample x i The average distance between the nearest neighbor clusters; The objective function for K-means clustering is: Where k is the number of clusters; v j C is the center vector of the j-th cluster; j Let represent the sample set of the j-th cluster; Step 3: Construct a resource allocation evaluation model based on real-time learning and dynamic fusion of locally weighted partial least squares-k-means clustering. Train and optimize the evaluation model using resource allocation data. Specifically: Construct a resource allocation dataset D=[X,Y] and perform standardization, where Y=[y1,y2,…,y N ], y i Let i be the production cycle time of the complex assembly line for the i-th sample; For the evaluation model, in the offline phase, the optimal parameter combination (h, R, p) is determined by minimizing the prediction error using a standardized resource allocation dataset and a leave-one-out cross-validation method combined with a grid search method; where h and R are the core parameters of the local partial least squares algorithm, and p is the distance threshold. During the online phase, for each resource configuration sample, the Euclidean distance between the sample and each cluster center is calculated, and a similar sample subset set Q = (Q1, Q2, ..., Qn) is obtained by filtering through a distance threshold p. s ), and for each similar sample subset Q j A local weighted partial least squares model is used to construct an instant learning model, and the corresponding production cycle prediction value y is output. j Based on the predicted value y j Euclidean distance d from the cluster center to the production cycle time j The prediction results of the real-time learning model corresponding to each similar sample subset are weighted and fused to output the predicted production cycle y. Step 4: Input the resource allocation schemes for different complex assembly lines into the evaluation model in Step 3 to obtain the production cycle prediction results for each resource allocation scheme. Based on the production cycle prediction results, the effectiveness of each resource allocation scheme can be accurately evaluated.

2. The resource allocation evaluation method for complex assembly lines under multiple working conditions based on real-time learning according to claim 1, characterized in that: In step 1, the Pearson correlation coefficient method is used to analyze the correlation between resource characteristics of complex assembly lines and to evaluate the correlation between different resources and production cycle time. Based on the feature relevance ranking results, the variable combinations of resource allocation with higher relevance rankings are selected as input variables for the resource allocation evaluation model. Based on the actual operation process, a complex assembly line simulation model was established using discrete-time simulation software, and simulation runs were conducted to obtain the production cycle time with different resource configurations.

3. The resource allocation evaluation method for complex assembly lines under multiple working conditions based on real-time learning according to claim 1, characterized in that: In step 3, the process of predicting the production cycle time using the locally weighted partial least squares model is as follows: Step 3.1: According to the formula Calculate the distance D = [d1, d2, ..., d] between the predicted point and each training sample. N ], and then through the formula Calculate the weights corresponding to each training sample, and then construct a diagonal matrix based on the weight calculation results. : Where x q =[x q1 ,x q2 ,…,x qm ], x q Assign assembly line resources for the predicted points; d i For the prediction point x q and training sample x i The Euclidean distance between them; Step 3.2: According to the formula Perform data mean centering processing, where and The weighted average of the input and output of the training samples, 1 n Let X be an n-dimensional column vector of all 1s. r Y r and x q,r These are the input, output, and predicted point vectors of the centered training sample, respectively. Step 3.3: Input the training samples and the predicted samples x respectively. q,r Extracting latent variables: In the formula, t r and t q,r They are X r and x q,r The corresponding r-th latent variable; w r yes The eigenvector corresponding to the largest eigenvalue; According to the formula Calculate the prediction point x q The load vector p r and the regression coefficient vector q of the model r Establish a regression equation Update the prediction point output; if the current number of latent variables r reaches the set value, output the predicted production cycle time value of the prediction point. Otherwise, let r = r + 1, and use the formula After updating the input, output, and predicted point vectors, return to step 3.3 to continue iterating.

4. The resource allocation evaluation method for complex assembly lines under multiple working conditions based on real-time learning according to claim 1, characterized in that: In step 3, the process of filtering and obtaining a subset of similar samples based on the distance threshold p is as follows: Through formula Calculate the Euclidean distance between the predicted point and each cluster center, and compare it with the distance threshold p; If the distance is within the threshold range, then the corresponding cluster dataset is selected as the similar sample subset Q=(Q1,Q2,…,Q…) for the prediction point. s ): 。 5. The resource allocation evaluation method for complex assembly lines under multiple working conditions based on real-time learning according to claim 1, characterized in that: In step 3, based on the predicted value y j Euclidean distance d from the cluster center to the production cycle time j The process of weightedly fusing the prediction results of the real-time learning model corresponding to each similar sample subset and outputting the predicted production cycle y is as follows: in The value is a very small positive value; the prediction results of the real-time learning model corresponding to each similar sample subset are multiplied by their corresponding weights, and then weighted and summed to obtain the final production cycle prediction value: 。

Images