Assembly line adaptation method based on process simulation data

Through three-dimensional assembly process simulation and machine learning algorithms, the assembly line prediction method has solved the problem of large workload and low reliability in small batches and diversified products, and achieved low cost and high accuracy assembly time prediction and station balance optimization.

CN119024775APending Publication Date: 2024-11-26CRRC YONGJI ELECTRIC CO LTD
View PDF 0 Cites 2 Cited by

Patent Information

Application Number
CN202410973914.X
Authority / Receiving Office
CN · China
Patent Type
Applications(China)
Current Assignee / Owner
Filing Date
2024-07-19
Publication Date
2024-11-26

AI Technical Summary

Technical Problem

Existing assembly line prediction methods such as the predetermined time standard method (MOD) have problems such as the problem of large workload, high cost and difficult to guarantee reliability in small batches, diversified and complex products.

Method used

Data is collected through three-dimensional assembly process simulation software, and assembly time prediction models of the same and similar devices are constructed. Time prediction is performed using twin neural convolution networks and support vector machines, and station balance optimization is performed in combination with genetic gene algorithms to achieve smoothing coefficients of assembly lines. Minimize.

Benefits of technology

Reduces workload and cost, improves prediction accuracy and reliability, and is especially suitable for small batches of diverse and complex products, ensuring optimal planning of assembly lines.

✦ Generated by Eureka AI based on patent content.

Smart Images

  • Figure CN119024775A_ABST
    Figure CN119024775A_ABST
Patent Text Reader

Abstract

The invention relates to the technical field of assembly lines, in particular to an assembly line adaptation method based on process simulation data, and mainly solves the technical problems that an existing MOD method is large in workload, high in cost and not suitable for small-batch diversified complex products, and the structural reliability is difficult to guarantee. According to the method, process simulation is carried out on a three-dimensional model assembly sequence to form assembly process data information, training is carried out by combining production procedures and working hour data of typical products, an assembly time prediction model of devices of the same specification and similar specifications is established, and before brand new product trial-manufacture, assembly time prediction is realized according to a simulation result, so that the assembly time prediction efficiency is improved. And station balance optimization is carried out according to the predicted man-hour data, so that the optimal planning of the assembly process of the new product is realized. The method can be permanently used after the model is trained, the workload is small, the cost is low, the method is particularly suitable for small-batch diversified complex products, and the structural reliability depends on the model precision and is easy to guarantee.
Need to check novelty before this filing date? Find Prior Art

Description

Technical Field

[0001] The present invention relates to the technical field of assembly lines, and in particular to an assembly line adaptation method based on process simulation data. Background Art

[0002] Mechanical products need to go through multiple production processes from roughcast to finished products. Assembly includes the process of assembling various semi-finished parts into complete products, so assembly is the last and most critical production link. The quality and efficiency of assembly directly affect the quality and output of the final product. As current product orders gradually develop in small batches, complexity, and diversification, the requirements for assembly line balance are also gradually increasing. How to achieve relatively accurate step and working time prediction through data analysis of process simulation and process planning, so as to achieve assembly line station balance design planning, has become a key issue in current research.

[0003] The existing method of predicting and guiding assembly lines with the scheduled time standard method (MOD) is to convert the normal operation time using the scheduled basic action time. The principle is that all the work, operations and operations involved by all people can be divided into some basic action elements, and all complex actions can be composed of the most basic actions. Different people do the same basic action (under the same conditions). Through a large number of experiments, it is shown that the time value is basically equal, and different people can be regarded as the same worker. When different parts of the body move, the time value used for the action is proportional to each other (such as: the wrist movement takes twice the time required for the finger movement), and the proportional calculation can be performed through different parts. However, the MOD method has the following defects: First, the MOD method must test the task completion time through a large number of repeated experiments, and the calculation of standard working hours is large and the implementation cost is high; second, it is not suitable for small batches of diversified and complex products; third, because the experiment is based on the participants' actions to determine the calculation combination, the reliability of the experimental results is subject to the subjective initiative of the participants, and the reliability is difficult to guarantee. Summary of the invention

[0004] In order to overcome the technical defects of the existing MOD method for predicting and guiding assembly lines, such as large workload, high cost, unsuitability for small batches of diversified and complex products, and difficulty in ensuring structural reliability, the present invention provides an assembly line adaptation method based on process simulation data.

[0005] The assembly line adaptation method based on process simulation data provided by the present invention comprises the following steps:

[0006] S1. Collect data:

[0007] Collecting assembly process data by means of three-dimensional assembly process simulation software, wherein the assembly process data includes assembly activity data, process step data, and workstation activity data;

[0008] S2. Build the model:

[0009] Constructing an assembly time prediction model for identical devices, and using specification code and simulated assembly time as feature data and actual assembly time as label data to train the assembly time prediction model for identical devices, thereby obtaining a prediction model M1;

[0010] Constructing a similar device assembly time prediction model, and using the model feature similarity, quality ratio and simulation assembly time ratio as feature data and the actual assembly time ratio as label data to train the similar device assembly time prediction model to obtain a prediction model M2;

[0011] S3. Assembly time prediction:

[0012] For the same device, the prediction model M1 is called to complete the prediction of assembly time;

[0013] For similar devices, the prediction model M2 is called to complete the prediction of assembly time;

[0014] In this way, the assembly time prediction of each component of the entire product is completed;

[0015] S4. Workstation balance:

[0016] Taking the assembly line smoothing coefficient SI as the objective function, the genetic algorithm is used to iterate the assembly line station plan. The result is optimal when the assembly line smoothing coefficient SI takes the minimum value, where:

[0017]

[0018] Where, T emax is the takt time; T ei is the homework time of unit i;

[0019] When SI takes the minimum value, it is regarded as the optimal solution, and the corresponding assembly line workstation solution is the optimal solution.

[0020] Optionally, the method for solving the model feature similarity is as follows:

[0021] 1) Generate three-view images and 45° isometric images of two similar devices A1 and A2 respectively to form two sample groups;

[0022] 2) The twin neural convolutional network is used to extract the features of the two sample groups and form two one-dimensional feature vectors W1 and W2 of size N. After mapping the two feature vectors, a one-dimensional vector W3 is obtained. The mapping formula is:

[0023]

[0024] The calculation formula of the loss function L is:

[0025]

[0026] Among them, y is the sample label, margin represents the threshold,

[0027] 3) Compare the front view of device A1 with the three views of device A2, and use the loss function L to solve the similarity, and output ε1, ε2 and ε3;

[0028] Compare the top view of device A1 with the three views of device A2, and use the loss function L to solve the similarity, and output ε4, ε5 and ε6;

[0029] Compare the side view of device A1 with the three views of device A2, and use the loss function L to solve the similarity, and output ε7, ε8 and ε9;

[0030] Compare the 45° isometric views of devices A1 and A2, and use the loss function to solve the similarity and output ε0;

[0031] 4) The calculation formula of the model feature similarity is:

[0032]

[0033] Optionally, the objective function of the similar device assembly time prediction model is expressed as:

[0034]

[0035] Where: is a nonlinear mapping function; ω is the weight coefficient; C is the penalty coefficient, indicating the degree of penalty for misclassified samples; ξ i Adjust for slack variables;

[0036] The constraint condition of the similar device assembly time prediction model is expressed as:

[0037] y i (ω×x i +b)-1+ξ i ≥0,(i=1,2,…,n)

[0038] Where: y i The result label.

[0039] Optionally, the objective function of the similar device assembly time prediction model is transformed into an expression after introducing Lagrange multipliers:

[0040]

[0041] Where: Q(a) is the geometric interval, ai is the Lagrange coefficient, a j is the derivation of the Lagrange coefficient, y i is the result label;

[0042] The transformed expression of the constraint conditions of the similar device assembly time prediction model is:

[0043]

[0044] Optionally, a support vector machine is used to train the similar device assembly time prediction model to obtain the prediction model M2.

[0045] Optionally, in step S4, the constraints include two points: first, in the workstation balance, the sequence constraint of the previous and next processes cannot be broken; second, the time of each workstation cannot be greater than the cycle time of the workstation.

[0046] Optionally, the expression of the genetic algorithm is:

[0047]

[0048] In the formula: C represents the individual encoding scheme, E represents the individual fitness evaluation function, P0 represents the initial population, M represents the population size, represents the selection operator, Г represents the crossover operator, ψ represents the mutation operator, T represents the termination condition of the genetic algorithm, N represents the number of iterations, G represents the number of workstations, and I represents the standard working time of each process.

[0049] The technical solution provided by the present invention has the following advantages compared with the prior art:

[0050] The assembly line adaptation method based on process simulation data provided by the present invention performs process simulation on the assembly sequence of the three-dimensional model to form data information such as the assembly sequence, simulation action time, and assembly trajectory of the assembly process. By combining the production process and working time data of typical products for training, an assembly time prediction model for devices of the same specification and similar specifications is established. Before the trial production of a new product, the assembly time is predicted according to the simulation results, and then the workstation balance is optimized according to the predicted working time data to achieve the optimal planning of the assembly process of the new product. This method can be used permanently after the model is trained, with small workload and low cost. It is particularly suitable for small batches of diversified and complex products, and the structural reliability depends on the model accuracy and is easier to guarantee. BRIEF DESCRIPTION OF THE DRAWINGS

[0051] The accompanying drawings, which are incorporated in and constitute a part of this specification, illustrate embodiments consistent with the invention and, together with the description, serve to explain the principles of the invention.

[0052] In order to more clearly illustrate the embodiments of the present invention or the technical solutions in the prior art, the drawings required for use in the embodiments or the description of the prior art will be briefly introduced below. Obviously, for ordinary technicians in this field, other drawings can be obtained based on these drawings without paying any creative work.

[0053] Figure 1 A flow chart showing an assembly line adaptation method according to an embodiment of the present invention. DETAILED DESCRIPTION

[0054] In order to more clearly understand the above-mentioned objectives, features and advantages of the present invention, the scheme of the present invention will be further described below. It should be noted that the embodiments of the present invention and the features in the embodiments can be combined with each other without conflict.

[0055] In the following description, many specific details are set forth to facilitate a full understanding of the present invention, but the present invention may also be implemented in other ways different from those described herein; it is obvious that the embodiments in the specification are only part of the embodiments of the present invention, rather than all of the embodiments.

[0056] The specific embodiments of the present invention are described in detail below with reference to the accompanying drawings.

[0057] Reference Figure 1 This embodiment provides an assembly line adaptation method based on process simulation data, including: S1. collecting data; S2. building a model; S3. predicting assembly time; S4. balancing workstations.

[0058] S1. Collect data: Collect assembly process data through 3D assembly process simulation software. The assembly process data includes assembly activity data, process step data and workstation activity data.

[0059] It is easy to understand that assembly activity data includes information on assembly objects (thumbnail, specification model, quantity) and assembly actions (trajectory, simulation duration); process step data consists of one or more assembly activity data, usually representing assembly-related activities of a class or similar devices. After editing by simulation software, a process step flow chart of a single process step or multiple process steps consisting of assembly activities is formed, which expresses the relationship between the process steps; workstation activity data consists of one or more process step data, representing an operating workstation on the actual assembly line.

[0060] Specifically, the assembly object information includes step name, device name, device material, device unit weight, device quantity and device thumbnail.

[0061] S2. Build a model: Build an assembly time prediction model for the same device, and use the specification code and simulated assembly time as feature data, and the actual assembly time as label data to train the assembly time prediction model for the same device, and obtain a prediction model M1; build an assembly time prediction model for similar devices, and use the model feature similarity, quality ratio and simulated assembly time ratio as feature data, and the actual assembly time ratio as label data to train the assembly time prediction model for similar devices, and obtain a prediction model M2.

[0062] Prediction model M1:

[0063] It should be noted that, since each simulated assembly action corresponds to an actual assembly action, the corresponding data is found to train the assembly time prediction model. For example, the following table is a specific training data feature table of this embodiment:

[0064]

[0065] It should be noted that the specification code is the internal material code corresponding to the device of this specification model, which is in the form of pure digital coding to facilitate the training of the subsequent model; after training, a prediction model M1 for various devices of the same specification model is formed.

[0066] Prediction model M2:

[0067] It should be noted that products that have been mass-produced are used for model training, while products that have not been mass-produced are used for model prediction.

[0068] Specifically, the method for solving the model feature similarity is as follows:

[0069] 1) Generate three-view images and 45° isometric images of two similar devices A1 and A2 respectively to form two sample groups;

[0070] 2) The twin neural convolutional network is used to extract the features of the two sample groups and form two one-dimensional feature vectors W1 and W2 of size N. After mapping the two feature vectors, a one-dimensional vector W3 is obtained. The mapping formula is:

[0071]

[0072] The calculation formula of the loss function L is:

[0073]

[0074] Among them, y is the sample label, margin represents the threshold,

[0075] 3) Compare the front view of device A1 with the three views of device A2, and use the loss function L to solve the similarity, and output ε1, ε2 and ε3;

[0076] Compare the top view of device A1 with the three views of device A2, and use the loss function L to solve the similarity, and output ε4, ε5 and ε6;

[0077] Compare the side view of device A1 with the three views of device A2, and use the loss function L to solve the similarity, and output ε7, ε8 and ε9;

[0078] Compare the 45° isometric views of devices A1 and A2, and use the loss function to solve the similarity and output ε0;

[0079] 4) The calculation formula for model feature similarity is:

[0080]

[0081] Through the above operations, it is ensured that the model similarity comparison is performed from the most similar angle of the two models, and it is found that the coefficient ε output after multiple comparisons of the models is the same.

[0082] Specifically, if the mass of device A1 is m1 and the mass of device A2 is m2, the mass ratio m=m1 / m2.

[0083] Specifically, if the simulated assembly time of device A1 is T1 and the simulated assembly time of device A2 is T2, then the simulated assembly time ratio T=T1 / T2.

[0084] Specifically, if the actual assembly time of device A1 is t1 and the actual assembly time of device A2 is t2, then the actual assembly time ratio t=t1 / t2.

[0085] It should be noted that data training is performed on all related similar devices to increase the sample set of products, thereby increasing the accuracy of the results. For example, the following table is a specific training data feature table of this embodiment:

[0086]

[0087] Specifically, the objective function of the similar device assembly time prediction model is expressed as:

[0088]

[0089] Where: is a nonlinear mapping function; ω is the weight coefficient; C is the penalty coefficient, indicating the degree of penalty for misclassified samples; ξ i Adjust for slack variables;

[0090] The constraint expression of the similar device assembly time prediction model is:

[0091] y i (ω×x i +b)-1+ξ i ≥0,(i=1,2,…,n)

[0092] Where: y i The result label.

[0093] More specifically, the objective function of the similar device assembly time prediction model is transformed into the expression after introducing the Lagrange multiplier:

[0094]

[0095] Where: Q(a) is the geometric interval, a i is the Lagrange coefficient, a j is the derivation of the Lagrange coefficient, y i is the result label;

[0096] The transformed expression of the constraint conditions of the similar device assembly time prediction model is:

[0097]

[0098] Specifically, a support vector machine is used to train a similar device assembly time prediction model to obtain a prediction model M2.

[0099] S3. Assembly time prediction: For the same device, call the prediction model M1 to complete the prediction of the assembly time; for similar devices, call the prediction model M2 to complete the prediction of the assembly time; in this way, the assembly time prediction of each device of the entire product is completed.

[0100] Specifically, after the three-dimensional process assembly simulation of the new product, the assembly time of each device is predicted. If there are models with the same specifications, the prediction model M1 is directly called to predict the assembly time. If there is no historical data with the same specifications, the prediction model M2 is called for device C. By selecting similar typical devices A x Perform appearance similarity comparison to obtain ε x , mass comparison to obtain m x , the simulation assembly time is compared to get T x , and finally get the actual assembly time ratio t x , the actual assembly time of device C is typical of product A x The actual assembly time divided by the actual assembly time ratio t x The predicted assembly time of each component of the entire product can be obtained through the prediction models M1 and M2.

[0101] S4. Workstation balance: Taking the assembly line smoothing coefficient SI as the objective function, the genetic algorithm is used to iterate the assembly line workstation plan. The result is optimal when the assembly line smoothing coefficient SI takes the minimum value, where:

[0102]

[0103] Where, T emax is the takt time; T ei is the operation time of the i-th unit; when SI takes the minimum value, it is regarded as the optimal solution, and the corresponding assembly line workstation plan is the optimal plan.

[0104] Specifically, the constraints include two points: first, in the workstation balance, the sequence constraints of the previous and next processes cannot be broken; second, the time of each workstation cannot be greater than the cycle time of the workstation.

[0105] Specifically, the expression of the genetic algorithm is:

[0106]

[0107] Where: C represents the individual coding scheme, E represents the individual fitness evaluation function, P0 represents the initial population, M represents the population size, φ represents the selection operator, Г represents the crossover operator, ψ represents the mutation operator, T represents the termination condition of the genetic algorithm, N represents the number of iterations, G represents the number of workstations, and I represents the standard working time of each process.

[0108] In actual calculation, the process-based individual coding scheme C is generated through binary coding. Binary coding is used, 0 means there is no process before the process, and 1 means there is a process before the process; the objective function SI is set as the individual fitness function E; the initial population number M is set; M initial populations P0 are randomly generated; the random selection operator φ is set; the crossover probability is set to Pc; the mutation probability is set to Pm; the termination condition is the minimum SI value; the number of iterations N is set; the number of workstations G and the process time matrix I. After the population is selected by the selection operator φ, the crossover operator Г and the mutation operator ψ, the optimal solution is finally obtained, which is the optimal assembly line workstation scheme.

[0109] The above is only a specific implementation of the present invention, which enables those skilled in the art to understand or implement the present invention. Although detailed descriptions are given with reference to the aforementioned embodiments, those skilled in the art should understand that they can still modify the technical solutions recorded in the aforementioned embodiments, or replace some or all of the technical features therein by equivalents; and these modifications or replacements do not make the essence of the corresponding technical solutions deviate from the scope of the technical solutions of the embodiments, and they should all be covered by the protection scope of the claims.

Claims

1. An assembly line adaptation method based on process simulation data, characterized in that: The steps include: S1. Collect data: Collecting assembly process data by means of three-dimensional assembly process simulation software, wherein the assembly process data includes assembly activity data, process step data, and workstation activity data; S2. Build the model: Constructing an assembly time prediction model for identical devices, and using specification code and simulated assembly time as feature data and actual assembly time as label data to train the assembly time prediction model for identical devices, thereby obtaining a prediction model M1; Constructing a similar device assembly time prediction model, and using the model feature similarity, quality ratio and simulation assembly time ratio as feature data and the actual assembly time ratio as label data to train the similar device assembly time prediction model to obtain a prediction model M2; S3. Assembly time prediction: For the same device, the prediction model M1 is called to complete the prediction of assembly time; For similar devices, the prediction model M2 is called to complete the prediction of assembly time; In this way, the assembly time prediction of each component of the entire product is completed; S4. Workstation balance: Taking the assembly line smoothing coefficient SI as the objective function, the genetic algorithm is used to iterate the assembly line station plan. The result is optimal when the assembly line smoothing coefficient SI takes the minimum value, where: Where, T emax is the takt time; T ei is the homework time of unit i; When SI takes the minimum value, it is regarded as the optimal solution, and the corresponding assembly line workstation solution is the optimal solution.

2. The assembly line adaptation method based on process simulation data according to claim 1, characterized in that: The method for calculating the similarity of the model features is as follows: 1) Generate three-view images and 45° isometric images of two similar devices A1 and A2 respectively to form two sample groups; 2) The twin neural convolutional network is used to extract the features of the two sample groups and form two one-dimensional feature vectors W1 and W2 of size N. After mapping the two feature vectors, a one-dimensional vector W3 is obtained. The mapping formula is: The calculation formula of the loss function L is: Among them, y is the sample label, margin represents the threshold, 3) Compare the front view of device A1 with the three views of device A2, and use the loss function L to solve the similarity, and output ε1, ε2 and ε3; Compare the top view of device A1 with the three views of device A2, and use the loss function L to solve the similarity, and output ε4, ε5 and ε6; Compare the side view of device A1 with the three views of device A2, and use the loss function L to solve the similarity, and output ε7, ε8 and ε9; Compare the 45° isometric views of devices A1 and A2, and use the loss function to solve the similarity and output ε0; 4) The calculation formula of the model feature similarity is:

3. The assembly line adaptation method based on process simulation data according to claim 1, characterized in that: The expression of the objective function of the similar device assembly time prediction model is: Where: is a nonlinear mapping function; ω is the weight coefficient; C is the penalty coefficient, which indicates the degree of penalty for misclassified samples; ξ i Adjust for slack variables; The constraint condition of the similar device assembly time prediction model is expressed as: y i (ω×x i +b)-1+ξ i ≥0,(i=1,2,…,n) Where: y i The result label.

4. The assembly line adaptation method based on process simulation data according to claim 3 is characterized in that: The objective function of the similar device assembly time prediction model is transformed into the expression after introducing the Lagrange multiplier: Where: Q(a) is the geometric interval, a i is the Lagrange coefficient, a j is the derivation of the Lagrange coefficient, y i is the result label; The transformed expression of the constraint conditions of the similar device assembly time prediction model is:

5. The assembly line adaptation method based on process simulation data according to claim 3 or 4, characterized in that: The similar device assembly time prediction model is trained using a support vector machine to obtain the prediction model M2.

6. The assembly line adaptation method based on process simulation data according to claim 1, characterized in that: In step S4, the constraints include two points: first, in the workstation balance, the sequence constraints of the previous and next processes cannot be broken; second, the time of each workstation cannot be greater than the cycle time of the workstation.

7. The assembly line adaptation method based on process simulation data according to claim 1, characterized in that: The expression of the algorithm of the genetic gene is: In the formula: C represents the individual encoding scheme, E represents the individual fitness evaluation function, P0 represents the initial population, M represents the population size, represents the selection operator, Г represents the crossover operator, ψ represents the mutation operator, T represents the termination condition of the genetic algorithm, N represents the number of iterations, G represents the number of workstations, and I represents the standard working time of each process.

Citation Information

Cited By

  • Model and method for predicting biological traits by integrating genetic background and variation information

    CN120108509A

  • A model and method for predicting biological traits by integrating genetic background and variation information

    CN120108509B