Automobile parts assembly production line expensive multi-objective scheduling optimization method and system

Abstract

This invention discloses an expensive multi-objective scheduling optimization method and system for automotive parts assembly lines. The method includes the following steps: collecting basic constraints, multi-objective indicators, and dynamic interference data of the production line; after cleaning, feature encoding and standardization, and feature selection and construction, dividing the data into training, validation, and test sets; constructing a GAN architecture consisting of a multi-objective generator and dual discriminators; using multi-objective constraint loss for the generator and dual discriminators to evaluate the feasibility and superiority of the proposed solutions; and converging the generator through multiple rounds of adversarial training; real-time monitoring of dynamic interference such as equipment failures and emergency order insertions, encoding these as interference vectors and inputting them into the generator to generate and optimize adjustment schemes; evaluating the schemes from the dimensions of feasibility, superiority, and economy, and converting the optimal scheme into MES system instructions for execution. This invention significantly reduces evaluation costs and provides a fast dynamic response speed for scheme adjustments.

Description

Technical Field

[0001] This invention relates to the intersection of automobile manufacturing and intelligent optimization technology, and in particular to an expensive multi-objective scheduling optimization method and system for automobile parts assembly production lines. It is applicable to the scenario of multi-variety, small-batch automobile parts final assembly, and takes into account the generation and optimization of scheduling schemes with multiple objectives such as production efficiency, equipment utilization, and cost control. Background Technology

[0002] The automotive manufacturing industry is currently developing towards "customization, multi-variety, and short-cycle" production, and automotive parts assembly lines (such as chassis parts, engine components, and interior component assembly lines) are facing significant scheduling complexity.

[0003] 1. Significant conflicts among multiple objectives: It is necessary to simultaneously optimize "shortest production cycle" (to meet the urgent orders of car manufacturers), "maximum equipment utilization" (to avoid idle high-priced equipment such as welding robots and CNC machining centers), "minimize production costs" (to reduce material waste and labor overtime costs), and "maximize product qualification rate" (to avoid rework caused by assembly errors), and there are strong conflicts among these objectives (such as shortening the cycle may increase equipment load and costs).

[0004] 2. High evaluation costs: The verification of traditional scheduling solutions relies on "simulation testing + on-site trial and error". Among them, building a digital twin simulation model of the production line requires the integration of data from multiple systems such as MES (Manufacturing Execution System) and ERP (Enterprise Resource Planning), and a single simulation takes more than 8 hours. If an unreasonable solution is directly implemented, it may lead to production line shutdown (such as material backlog caused by process connection errors). The loss of one hour of shutdown of a single production line can reach tens of thousands of yuan, which is a typical "expensive multi-objective optimization problem".

[0005] 3. Frequent dynamic interference: During the production process, there are often emergency order insertions (such as car companies temporarily adding orders for a certain model of chassis), equipment failures (such as sudden failures of welding robots), and material delays (such as fluctuations in the supply chain of imported bolts). Traditional static scheduling methods (such as manual scheduling and offline optimization by genetic algorithms) are slow to respond and difficult to adjust in real time.

[0006] Currently, the technologies for solving production line scheduling problems are mainly divided into three categories, none of which can effectively address the "expensive multi-objective optimization requirements".

[0007] 1. Traditional heuristic algorithms, such as genetic algorithms and particle swarm optimization, can solve multi-objective optimization problems, but they require a large number of iterations to search for the optimal solution. Each iteration requires calling a simulation model to evaluate the scheme, resulting in high computational costs (e.g., solving 100 schemes takes more than 24 hours) and they are prone to getting trapped in local optima (e.g., optimizing only the equipment utilization rate while ignoring the delivery cycle).

[0008] 2. Shallow Neural Network Method: Although there are cases of using shallow neural networks to achieve single-objective scheduling optimization, shallow neural networks can only fit linear mapping relationships and cannot capture the nonlinear coupling relationship of "equipment status-process time-cost" in production line scheduling. Furthermore, they are difficult to handle the generation of Pareto optimal solutions under multi-objective conflicts.

[0009] 3. Multimodal optimization techniques: Some studies use multimodal models to process scheduling data (such as combining equipment vibration data and order text data), but they have not designed optimization mechanisms for the "expensive evaluation" characteristics, and still need to rely on a lot of field testing to verify the scheme, thus failing to reduce optimization costs. Summary of the Invention

[0010] The purpose of this invention is to provide an expensive multi-objective scheduling optimization method and system for automotive parts assembly production lines, which addresses the problems of high evaluation costs, difficulty in balancing multiple objectives, and slow dynamic response in traditional scheduling methods, and achieves low-cost and efficient scheduling decisions.

[0011] To address this, the present invention provides a GAN-based method for optimizing expensive multi-objective scheduling in automotive parts assembly lines, comprising the following steps: S1, cleaning, encoding discrete features and standardizing continuous features, and filtering effective features from historical data collected from automotive parts assembly lines, and then dividing the data into training, validation, and test sets. This historical data includes basic constraint data, multi-objective index data, and dynamic interference data; S2, constructing a Generative Adversarial Network (GAN) model consisting of a multi-objective generator and a dual discriminator. The multi-objective generator incorporates an attention mechanism and a multi-objective constraint loss, while the dual discriminator evaluates the feasibility and superiority of the proposed solutions. The multi-objective generator is converged through multiple rounds of adversarial training; S3, real-time monitoring of production line interference events, encoding these events into interference vectors, and inputting them into the generator to generate adjustment schemes. The optimal scheme is then ranked and selected based on the interference impact and multi-objective loss; S4, evaluating the schemes from three dimensions: feasibility, superiority, and economy. The optimal scheme is then transformed into MES system instructions for implementation and scheduling optimization.

[0012] Compared with the prior art, the present invention has the following significant advantages:

[0013] 1. Significantly reduce evaluation costs: By learning from historical data through GAN models, the reliance on real simulations / field trial and error is reduced. Experimental verification shows that the cost of scheme evaluation is reduced by 60% to 70% (traditional methods require 20 simulation verifications, while this invention only requires 5 to 6).

[0014] 2. Significant multi-objective optimization results: In a test on an automobile chassis assembly line, the production cycle was shortened by 25%–30% (from 56 hours to 42 hours), equipment utilization increased by 15%–20% (from 72% to 88%), production costs decreased by 12%–15% (from 2350 yuan / unit to 1920 yuan / unit), and the product qualification rate remained stable at over 99.2%.

[0015] 3. Fast dynamic response speed: In the face of interference such as equipment failure and emergency orders, the solution adjustment time is shortened from 2-4 hours in the traditional way to 10-30 seconds, and the delivery delay rate of emergency orders is reduced from 12% to below 3%;

[0016] 4. High versatility: It can be adapted to assembly lines of different types of automotive parts (such as engine components and interior components). Only the data input and target constraint parameters need to be adjusted. There is no need to reconstruct the model. Its adaptability covers more than 80% of automotive parts manufacturing scenarios.

[0017] In addition to the objectives, features, and advantages described above, the present invention has other objectives, features, and advantages. The invention will now be described in further detail with reference to the figures. Attached Figure Description

[0018] The accompanying drawings, which form part of this application, are used to provide a further understanding of the invention. The illustrative embodiments of the invention and their descriptions are used to explain the invention and do not constitute an undue limitation of the invention. In the drawings:

[0019] Figure 1 This is a flowchart of the expensive multi-objective scheduling optimization method for automotive parts assembly production lines of the present invention;

[0020] Figure 2 This is a schematic diagram of the network structure of the GAN core optimization module of the present invention;

[0021] Figure 3 This is a structural block diagram of the expensive multi-objective scheduling optimization system for automotive parts assembly production lines of the present invention. Detailed Implementation

[0022] The present invention will now be described in detail with reference to the accompanying drawings and embodiments.

[0023] The technical solution of this invention comprises four parts: a data preprocessing module, a GAN model core optimization module, a dynamic scheduling and adjustment module, and a scheme evaluation and output module. These modules are logically connected and functionally complementary.

[0024] The data preprocessing module is used to build high-quality training datasets.

[0025] This module implements end-to-end processing from raw production data to model input vectors, ensuring data quality and adaptability. Specific steps include:

[0026] (1) Data acquisition: Collecting data from the automotive parts assembly line, including three types of core data:

[0027] (1.1) Basic constraint data: equipment parameters (such as welding robot working power, CNC machining center machining accuracy, equipment maximum daily capacity), process data (sequence of assembly processes for each part, standard time for a single process, process constraints), order data (part model, quantity, delivery date). The sequence of assembly processes for parts is expressed through a pre-constraint matrix. In the matrix, Aij=1 indicates that process i must be completed before process j, Aij=0 indicates that i and j are unconstrained and can be parallel, and Aij=-1 is used for internal error correction, indicating that j is the process preceding i. Process constraints include the requirement that the cooling station must follow the welding station.

[0028] (1.2) Multi-objective indicator data: production cycle (time from production to completion of a single batch of parts), equipment utilization rate (actual working time of each piece of equipment / planned working time), cost data (material loss cost, labor cost, equipment maintenance cost), and pass rate (number of qualified parts in a single batch / total number).

[0029] (1.3) Dynamic interference data: historical equipment fault records (fault type, repair time), emergency order insertion records (order quantity, insertion time), and material delay records (delay time, affected processes).

[0030] (2) Data cleaning: To address the issues of duplication, missing data, and anomalies that are prone to occur during the data collection process on the production line, a three-step cleaning process is adopted:

[0031] (2.1) Data deduplication: Using “order number + process number + collection timestamp” as the composite primary key, delete completely duplicate records; for records with duplicate fields but inconsistent key information (such as processing time), retain the entry with the smallest deviation from the same batch of data to avoid duplicate data interfering with model learning;

[0032] (2.2) Missing value handling: "filling with the average of similar parameters of the equipment" (e.g., when the fault data of a certain welding robot is missing, it is filled with the average fault duration of the same model of robot in the past 30 days) and "process time interpolation to complete" (e.g., when the processing time of a certain process is missing, it is filled by linear interpolation based on the processing time of the preceding and following processes and the equipment capacity); records with missing key constraint data such as order delivery date and core parameters of equipment are directly removed (the proportion is usually ≤ 1%, which does not affect the integrity of the dataset);

[0033] (2.3) Outlier handling: The “3σ principle + business rule verification” dual screening is adopted. For continuous data such as production cycle, equipment vibration amplitude, and cost, values ​​that deviate from the mean by more than 3 standard deviations are removed and replaced with the median of the feature. Combined with business rule verification (such as single process processing time ≤ 8 hours, unit cost ≥ material base cost), abnormal records that violate common sense are deleted to ensure data authenticity.

[0034] (3) Feature encoding and standardization: the cleaned data is transformed into structured features that the model can process, eliminating dimensional differences and semantic biases.

[0035] (3.1) Discrete feature encoding: One-hot encoding is used for discrete features with priority, such as order priority, equipment failure type, and part model; label encoding is used for discrete features without priority, such as equipment number.

[0036] (3.2) Standardization of continuous features: For continuous features such as production cycle, equipment utilization rate, cost, and vibration amplitude, Min-Max normalization is used to map them to the [0,1] interval. The formula is as follows:

[0037]

[0038] In the formula, The normalized feature result will be mapped to the interval [0,1], where x is the original data of a certain continuous feature. It is the minimum value of this feature in the entire dataset. It is the maximum value of the feature in the entire dataset, eliminating the unit difference between "hour" and "yuan" and avoiding model bias towards features with large vector dimensions.

[0039] (4) Feature selection and dataset partitioning: optimize feature quality, extract information valuable for scheduling, reduce model computational complexity, and split the data according to the "training-validation-testing" logic:

[0040] (4.1) Redundant features are removed using the Pearson correlation coefficient, as shown in the formula:

[0041]

[0042] In the formula, The Pearson correlation coefficient represents the correlation between features X and Y, and its value ranges from [-1, 1]. A Pearson correlation coefficient greater than 0 indicates that X and Y are positively correlated, and vice versa. When the value is greater than 0.8, it indicates a high correlation, and one of them needs to be removed to reduce redundancy;

[0043] (4.2) Invalid features are removed using the variance threshold method, the formula is as follows:

[0044]

[0045] In the formula, the sample variance ≥ 0 indicates that the larger the value, the more volatile the feature; the closer the value is to 0, the more stable the feature value. < When that happens, it should be removed.

[0046] (4.3) Dataset partitioning: The dataset is divided in a 7:2:1 ratio: training set (1278 samples, used for adversarial training of GAN model to learn scheduling rules), validation set (365 samples, used to adjust the hyperparameters of generator and discriminator, such as learning rate and loss function weights), and test set (182 samples, used to verify generalization ability with data not used in training to avoid overfitting). When partitioning, it is necessary to ensure that "order data in the same batch do not cross sets" to avoid the process data of a certain order appearing in both the training set and the test set at the same time, which may lead to model misjudgment.

[0047] 2. The core optimization module of GAN generates optimal scheduling schemes for multiple objectives.

[0048] This module is the core execution unit, employing an improved GAN architecture of "multi-object generator + dual discriminator." It achieves efficient generation of high-quality scheduling schemes through adversarial training, specifically including architecture design, training process, and convergence determination.

[0049] (1) Network architecture design

[0050] (1.1) Multi-objective generator (G), focusing on the accurate generation of scheduling schemes to adapt to the multi-objective requirements of the production line:

[0051] (1.1.1) Network structure: The network adopts a progressive architecture of “input layer → fully connected layer #1 → fully connected layer #2 → output layer”. The input layer receives a 128-dimensional fusion vector (64-dimensional order demand vector + 64-dimensional equipment status vector), and the output layer outputs a 64-dimensional scheduling scheme vector (including “process-equipment allocation” 20-dimensional, “process time window” 24-dimensional, “material delivery node” 12-dimensional, and “personnel configuration” 8-dimensional).

[0052] (1.1.2) Activation function: LeakyReLU was used for fully connected layer #1 (256 neurons) and fully connected layer #2 (128 neurons), and the formula is as follows:

[0053]

[0054] In the formula, the slope of 0.01 is adapted to the characteristics of production data;

[0055] The output layer uses a sigmoid function, with the following formula:

[0056]

[0057] The output is mapped to the [0,1] interval to facilitate subsequent score calculation;

[0058] (1.1.3) Loss Function: A multi-objective constraint term is embedded in the generator loss function, as shown in the following formula:

[0059]

[0060] In the formula, To mitigate losses, Wasserstein distance is used for calculation.

[0061]

[0062] in, It represents the distribution of "truly high-quality scheduling schemes" throughout the history of the production line. It is the distribution of the "candidate scheduling schemes" output by the GAN. From A genuine, high-quality solution selected from the pool. From One of the generation schemes is extracted from the middle.

[0063] In the loss function It is a multi-target loss function, using weighted Euclidean distance.

[0064]

[0065] Here, the value of k ranges from 1 to 4, corresponding to four types of objectives, including production cycle: k=1 minimizing production cycle, k=2 maximizing equipment utilization, k=3 minimizing production cost, and k=4 maximizing product qualification rate. The weights for each objective are set according to the company's production priorities. For predicted values, This is the ideal threshold.

[0066] It is the constraint loss, which is 10 when the process or equipment constraint is violated, and 0 otherwise. α, β, and γ are weighting coefficients (set to 0.4, 0.3, and 0.3 respectively after optimization on the validation set).

[0067] (1.2) Dual discriminators (D1, D2) are used to perform dual verification of the "feasibility + excellence" of the solution, ensuring its feasibility for implementation.

[0068] (1.2.1) Feasibility Discriminator (D1): Anchored to process and equipment constraints, the network structure is "input layer (64-dimensional scheme vector) → 3×3 convolutional layer (64 feature maps) → max pooling layer → fully connected layer (64 neurons) → output layer (1-dimensional score)"; the activation functions are LeakyReLU (convolutional layer) and Sigmoid (output layer); the loss function is binary cross-entropy.

[0069]

[0070] In the formula, For a truly feasible solution, To generate a solution, an output score ≥ 0.8 is considered "feasibility met".

[0071] (1.2.2) Superiority Discriminator (D2): Focuses on multi-objective balancing effects. The input is a "64-dimensional scheme vector + 16-dimensional multi-objective index vector". The network structure is "input layer (80 dimensions) → fully connected layer #1 (128 neurons) → batch normalization layer → fully connected layer #2 (64 neurons) → output layer (1-dimensional score)". The activation functions used are ReLU and Sigmoid, where the ReLU formula is...

[0072]

[0073] The loss function is the mean squared error:

[0074]

[0075] In the formula Distribution of feasible compliance solutions For an unbalanced target distribution, an output score ≥ 0.85 is considered "excellent and meets the standard".

[0076] (2) Combat training process

[0077] The training process revolves around "data adaptation → alternating training → convergence determination," adapting to the characteristics of automotive parts production line data, and involves a total of 5000 iterations.

[0078] (2.1) Pre-training requirements:

[0079] (2.1.1) Data adaptation: The preprocessed 128-dimensional input vector and 64-dimensional real solution vector are loaded into the model, where the real solutions are selected from the training set as "multi-objective compliance samples" (production cycle ≤ 48 hours, equipment utilization rate ≥ 85%).

[0080] (2.1.2) Model initialization: The generator and dual discriminator parameters are initialized using the He normal distribution, and the hyperparameters are set as batch size = 32, lr_G (generator learning rate) = 0.0001, lr_D (discriminator learning rate) = 0.0002, β1 = 0.5 (Adam optimizer);

[0081] (2.1.3) Generator pre-training: Fix the discriminator parameters and pre-train the generator for 100 rounds (each round traverses 57 batches) to force the feasibility score of the generation scheme to be ≥ 0.6, so as to avoid the discriminator being "one-sided" in the initial training.

[0082] (2.2) Iterative training (each round contains two phases):

[0083] Phase 1: Fix G, train D1 + D2:

[0084] ① Batch sampling: Randomly select 32 sets of input vectors and generate 32 sets of candidate solutions through G; Simultaneously extract 32 sets of real high-quality solutions and corresponding multi-objective index vectors;

[0085] ②D1 Training: Input "real feasible solutions (positive samples) + generated solutions (including 20% ​​non-compliant negative samples)", calculate Update the D1 parameter (each update increment ≤ 0.01), and the accuracy rate of identifying violation schemes after a single round is ≥ 88%;

[0086] ③D2 Training: Select feasible solutions with a D1 score ≥ 0.8, and input them into D2 for calculation. Update D2 parameters; after a single round, the accuracy rate of identifying the qualified solution is ≥ 85%.

[0087] Phase 2: Fix D1+D2, train G:

[0088] ① Sample Generation: Input 32 new 128-dimensional vectors. G increases the feature weights of "order urgency" and "equipment load rate" (by 0.35 and 0.25 respectively) through an attention layer, outputting 32 solutions:

[0089] ② Loss Calculation: Input the solution into D1 and D2 to obtain a score, and combine it with... , , calculate;

[0090] ③ Parameter update: Adjust the fully connected layer parameters of G through backpropagation using the Adam optimizer.

[0091] (2.3) Convergence criterion:

[0092] (2.3.1) Core indicators meet the standards: In 10 consecutive rounds of verification, the average score of D1 of the generated solutions is ≥0.95, the average score of D2 is ≥0.9, and the generation time of a single batch of solutions is ≤5 seconds;

[0093] (2.3.2) Loss stability: LG fluctuation ≤ 5%, , Stable below 0.1;

[0094] (2.3.3) Save the model parameters after the target is met.

[0095] 3. Dynamic scheduling and adjustment module: Real-time response to production disruptions.

[0096] This module connects the core optimization module of GAN with the actual production system, enabling rapid response to interference events. The specific process is as follows:

[0097] (3.1) Interference Detection: The system collects equipment status (vibration amplitude, temperature), order status (emergency order insertion, delivery date change), and material status (delayed delivery) data every second through the MES system interface. Preset trigger thresholds are set for: equipment failure duration ≥ 2 hours, emergency order quantity ≥ 100 pieces, and material delay duration ≥ 4 hours. Dynamic adjustment is initiated when these thresholds are triggered. Vibration amplitude and temperature are used as early warning signals for potential equipment failures. For example, when the vibration amplitude deviates from the mean by 3 standard deviations, the system predicts that the equipment is about to fail and triggers dynamic interference vector coding for 'equipment failure' in advance, thereby realizing the transformation from 'passive post-maintenance scheduling' to 'proactive preventative scheduling'.

[0098] (3.2) Interference information encoding: The interference event is converted into an 8-16 dimensional interference vector (e.g., “1#CNC mechanical failure, repair for 3 hours” is encoded as [1,0,0,0,1,0,0.3,0]), and concatenated with the original 128 dimensional input vector to form a new 136 dimensional input;

[0099] (3.3) Rapid update of the scheme: The new input vector is fed into the trained generator G, and 3 to 5 sets of adjustment schemes are output within 10-30 seconds (such as transferring the process to the backup equipment when the equipment fails, and optimizing the order sequence when an emergency order is inserted).

[0100] (3.4) Priority ranking: Calculate the comprehensive score according to "interference impact (weight 0.4) + multi-objective loss (weight 0.6)" and select the scheme with the highest score (such as "equipment failure adjustment scheme A" with a score of 0.92, which is higher than scheme B's 0.88) to ensure that the production loss after adjustment is minimized.

[0101] 4. Solution Evaluation and Output Module: Visualization and Implementation Integration

[0102] This module performs the final verification and production deployment of the scheduling scheme, and integrates with the enterprise's existing systems:

[0103] (4.1) Multi-dimensional evaluation: The evaluation scheme is evaluated from three dimensions: "feasibility" (satisfaction of process constraints, such as process sequence and equipment load), "excellence" (compliance rate of multiple target indicators, such as production cycle ≤ 48h and cost ≤ 2000 yuan / unit), and "economic efficiency" (adjustment costs, such as equipment switching costs and material transfer costs), and an evaluation report is generated.

[0104] (4.2) System integration: The optimal solution is converted into instruction codes that the MES system can recognize (such as welding robot start and stop signals, material delivery order numbers), and pushed synchronously to the equipment controller and production dashboard to achieve seamless integration with the existing production line and ensure that the solution is implemented within 10 minutes.

[0105] (a) Implementation Scenarios

[0106] Taking the chassis assembly production line of an automotive parts manufacturing company as the implementation target, the production line is equipped with 12 core pieces of equipment (3 welding robots, 2 CNC machining centers, 4 assembly machines, and 3 testing equipment), adopts a "multi-variety, small-batch" production mode, and processes an average of 5-8 batches of orders per day (covering 3 chassis models). The core optimization objectives are: production cycle ≤ 48 hours, equipment utilization rate ≥ 85%, unit cost ≤ 2000 yuan / unit, and pass rate ≥ 99%. The production line has deployed an MES system (Siemens Opcenter Execution), which has the ability to collect data in real time and issue equipment commands.

[0107] (II) Implementation Steps

[0108] (1) Data preparation (corresponding to the data preprocessing module)

[0109] (1.1) Data collection: Historical data of the production line from January 2023 to June 2024 were collected, totaling 1825 sets of valid scheduling records, including equipment parameters (welding robot power 15kW, CNC machining accuracy ±0.02mm), order data (single order quantity of 50-200 units for model A chassis), multi-objective index data (average production cycle 56 hours, equipment utilization rate 72%), and dynamic interference data (average annual equipment failures 32 times, emergency order insertion rate 15%).

[0110] (1.2) Data cleaning: Handling missing values ​​(accounting for 3.2%), such as filling missing data with the average failure time of the same type of welding robot (2.5 hours); removing outliers (such as records of "single process processing time 12 hours" and "unit cost of 1800 yuan is lower than material cost of 2000 yuan").

[0111] (1.3) Feature processing: One-hot encoding is used for “part model (A / B / C)”, and Min-Max normalization is used for “production cycle (30-72 hours)”; derived features such as “equipment load rate” and “order urgency” are constructed; the training set (1278 groups), validation set (365 groups), and test set (182 groups) are divided into three sets according to a ratio of 7:2:1.

[0112] (2) GAN model training (corresponding to the core optimization module of GAN)

[0113] (2.1) Training environment:

[0114] The personal computer is equipped with a 13th Gen Intel(R) Core(TM) i7-13700K (3.40 GHz) processor, Intel(R) UHD Graphics 770 graphics card, 64GB DDR5 4800MHz memory (16GB×4), and an ASUS ROG STRIX Z790-A GAMING WIFI motherboard. The model is implemented based on the PyTorch 2.0 framework with CUDA version 11.7.

[0115] (2.2) Model training: Set the number of iteration rounds to 5000, batch size=32, and the learning rate to the above parameters; the model converged when training reached the 3800th round, the average score of the validation set D1 was 0.96, the average score of D2 was 0.92, and the LG was stable at 0.08~0.084;

[0116] (2.3) Model validation: The model was validated using 182 test samples. The average production cycle of the generated solution was 42 hours, the equipment utilization rate was 88%, and the cost was 1920 yuan / unit, all of which met the target requirements. The average generation time of the solution was 4.2 seconds.

[0117] (3) Conventional scheduling applications (corresponding to the scheme evaluation and output module)

[0118] Typical scheduling case 1: Batch assembly of model A chassis

[0119] 1. Data Input and Characterization: The MES system issues an order for "100 Type A chassis, delivery cycle of 48 hours, high priority". The system collects the real-time status of the current No. 1 CNC machining center being idle and the No. 2 welding robot being at high load (60%), and encodes this data along with the process path into a 128-dimensional input vector.

[0120] 2. Generating Schemes (GAN Execution): The trained generator G generates 5 different candidate scheduling schemes (including equipment allocation, material delivery nodes, etc.) within 5 seconds.

[0121] 3. Dual Verification and Optimization: Feasibility discriminator D1 found that Scheme 3 violated the "welding before testing" process sequence and eliminated it. D1 screened out 4 feasible schemes (score ≥ 0.9). Excellence discriminator D2 scored the remaining schemes and found that "Scheme 1" achieved the highest score (0.93) while ensuring 88% equipment utilization and requiring only 42 hours of production cycle.

[0122] 4. Implementation: The system converts Solution 1 into instructions, and the solution into MES instructions. Welding robots #2 and #3 are started at 10:00, CNC #1 is started at 10:30, and materials are delivered to the assembly station at 11:00. The actual completion time is 8:30 on July 17, 1.5 hours ahead of schedule.

[0123] 5. Dynamic Response: Production reached 14:00 on July 17th when the system detected abnormal vibration of CNC #1, indicating a potential fault (vibration amplitude exceeding 3σ, estimated repair time 3 hours), triggering the dynamic adjustment module. The interference vector was concatenated with the original input to form a 136-dimensional vector. Generator G re-outputted Solution A within 12 seconds: transfer the process to CNC #2, delaying by 20 minutes, with a comprehensive score of 0.92, determined as the optimal solution. The system terminated the task of CNC #1, sent a start command to CNC #2, smoothly transferred subsequent processes to CNC #2, and simultaneously adjusted the assembly process time to 17:00, ultimately resulting in a delay of only 1.5 hours, without affecting overall delivery.

[0124] Typical scheduling case 2: A hybrid production line needs to assemble 50 pure electric cars (EVs) and 50 gasoline SUVs (ICEs) at the same time.

[0125] Multiple objectives: 1. Cycle time: Total production time for 100 vehicles ≤ 48 hours. 2. Efficiency: Utilization rate of key stations (such as chassis assembly station) ≥ 88%. 3. Cost: Lowest material transfer and process changeover costs. 4. Quality: First-pass yield rate ≥ 99%.

[0126] 1. Data Feature Modification (Input Layer)

[0127] The MES system collects the current production line status and encodes it as a 128-dimensional input vector: Order characteristics: the mixing ratio of EV models and ICE models, and the urgency of delivery (e.g., EVs are export orders, with higher priority). Static constraints: Based on the preceding constraint matrix, it is clearly stated that "engine assembly" only applies to ICE, "battery pack installation" only applies to EVs, and both must be after "interior assembly".

[0128] 2. Scheme generation (GAN generator G)

[0129] Generator G focuses on "order urgency" through an attention mechanism and quickly outputs 3-5 sets of full scheduling tables; the plan details which vehicle enters the "assembly station" and which vehicle enters the "inspection station" every minute, as well as the corresponding personnel configuration plan.

[0130] 3. Scheme evaluation and correction (dual discriminators D1 & D2)

[0131] Feasibility Verification (D1): The system automatically identified that in "Scheme 2," an EV sedan was incorrectly assigned to the "fuel and oil filling station." Discriminator D1 assigned a low score (0.2) to this scheme based on the process matrix (Aij constraint) and eliminated it. Excellence Screening (D2): Among the remaining feasible schemes, D2 found that "Scheme 1," by alternating between EV and ICE models, balanced the load on the battery installation station and the fuel system installation station, shortening the production cycle to 42 hours, and received the highest score (0.95).

[0132] 4. Real-time adjustment under dynamic disturbances

[0133] Interference Trigger: During the 20th hour of production, the chassis assembly robot suddenly triggered a temperature warning (imminent malfunction). Second-Level Response: The system encoded the interference signal as a 16-dimensional vector (1717). The GAN model updated its solution within 12 seconds: High-load tasks were temporarily switched to redundant workstations, and the order sequence was reordered. Implementation: The MES system received the instruction code, automatically switched equipment status, and updated the Kanban information, ensuring uninterrupted production with a final delay of less than 1.5 hours.

[0134] (III) Effect Verification

[0135] The scheduling performance of this invention is compared with that of traditional genetic algorithms, and the results are shown in the table below:

[0136] The verification results show that the present invention is significantly better than the traditional method in terms of multi-objective optimization effect, generation efficiency and dynamic response capability, and fully meets the scheduling requirements of automotive parts assembly production line.

[0137] The above description is merely an embodiment of the present invention and is not intended to limit the invention. Various modifications and variations can be made to the present invention by those skilled in the art. Any modifications, substitutions, or improvements made within the spirit and principles of the present invention should be included within the scope of protection of the present invention.

Claims

1. A GAN-based method for optimizing the expensive multi-objective scheduling of automotive parts assembly lines, characterized in that, Includes the following steps: S1. After cleaning, encoding discrete features and standardizing continuous features, and filtering valuable features, the historical data collected from the automotive parts assembly line is divided into training set, validation set and test set. The historical data includes basic constraint data, multi-objective index data and dynamic interference data. S2. Construct a Generative Adversarial Network (GAN) model consisting of a multi-target generator and a dual discriminator. The multi-target generator incorporates an attention mechanism and a multi-target constraint loss. The dual discriminator is used to evaluate the feasibility and superiority of the scheme. The multi-target generator is converged through multiple rounds of adversarial training. S3. Monitor production line interference events in real time, encode the interference events into interference vectors and input them into the generator to generate adjustment plans. Then, sort and select the optimal plan according to the interference impact and multi-objective loss. S4. Evaluate the solution from three dimensions: feasibility, excellence, and economy. Transform the optimal solution into MES system instructions for implementation and scheduling optimization.

2. The expensive multi-objective scheduling optimization method for automotive parts assembly lines based on GAN as described in claim 1, characterized in that, The feature encoding and standardization in step S1 include: using one-hot encoding for discrete features with priority, and label encoding for discrete features without priority; and using Min-Max normalization for continuous features, as shown in the following formula: , In the formula, The normalized features will eventually map to the [0,1] interval. It is the original data of a certain continuous feature. It is the minimum value of this feature in the entire dataset. It is the maximum value of this feature across the entire dataset.

3. The expensive multi-objective scheduling optimization method for automotive parts assembly lines based on GAN as described in claim 1, characterized in that, The loss function of the multi-objective generator G in step S2 is: , In the formula, L adv It is to combat loss so that the generated solution approximates the real high-quality solution, L obj It is a multi-objective loss, used to calculate the deviation between the generated scheme and the ideal multi-objective index, L. const It is a constraint loss, used to punish schemes that violate the process sequence or exceed equipment load limits, with α, β, and γ as weighting coefficients.

4. The expensive multi-objective scheduling optimization method for automotive parts assembly lines based on GAN as described in claim 1, characterized in that, The dual discriminator mentioned in step S2 includes: The feasibility discriminator D1 uses binary cross-entropy loss, as shown in the formula: , In the above formula, L D1 It is the total loss value of the feasibility discriminator. It is the mathematical expectation operator, and D1() is the output score of the feasibility discriminator for the actual scheduling scheme. Here, z represents the probability distribution of these realistically feasible solutions, and z is the input noise vector of the generator. It is the probability distribution of noise. It is the output score of the feasibility discriminator on the generated scheme. The superiority discriminator D2 evaluates the loss using the mean square error form, as shown in the formula: , In the above formula, L D2 is the total loss value of the excellence discriminator, x is the multi-objective achievement scheme, and y is the multi-objective index vector. It is the probability distribution of the qualified sample pairs. It is a multi-objective imbalance scheme. It is the probability distribution of imbalanced sample pairs.

5. The expensive multi-objective scheduling optimization method for automotive parts assembly lines based on GAN according to claim 1, characterized in that, The adversarial training includes a pre-training phase, an alternating training phase, and a convergence determination phase.

6. The expensive multi-objective scheduling optimization method for automotive parts assembly lines based on GAN according to claim 1, characterized in that, Data cleaning includes: Using the order number, process number, and data collection timestamp as a combined primary key, completely duplicate records are deleted; for records with duplicate fields but inconsistent key information, the entry with the smallest deviation from the data in the same batch is retained. For missing data, the average value is used to fill in the parameters of the same type of equipment and the interpolation is used to complete the process time. Records with missing key constraint data, including order delivery date and core parameters of equipment, are directly removed. For continuous data including production cycle, equipment vibration amplitude, and cost, values ​​that deviate from the mean by more than 3 standard deviations are removed and replaced with the median of the feature. Business rules are used for verification, and abnormal records that violate common sense are deleted.

7. The expensive multi-objective scheduling optimization method for automotive parts assembly lines based on GAN according to claim 1, characterized in that, The valuable feature screening includes removing redundant features using the Pearson correlation coefficient and removing invalid features using the variance threshold method.

8. A costly multi-objective scheduling optimization system for automotive parts assembly lines based on GAN, characterized in that, include: The data preprocessing module is used to clean, encode discrete features and standardize continuous features, and filter effective features of historical data collected from automotive parts assembly lines. After that, the data is divided into training set, validation set and test set. The historical data includes basic constraint data, multi-objective index data and dynamic interference data. The GAN model training module is used to construct a generative adversarial network (GAN) model consisting of a multi-objective generator and a dual discriminator. The multi-objective generator incorporates an attention mechanism and a multi-objective constraint loss, while the dual discriminator is used to evaluate the feasibility and superiority of the scheme. The multi-objective generator is converged through multiple rounds of adversarial training. The dynamic scheduling and adjustment module is used to monitor production line interference events in real time. After encoding the interference events into interference vectors, the modules input them into the generator to generate adjustment plans. Then, the optimal plan is sorted and selected according to the interference impact and multi-objective loss. The solution evaluation and output module evaluates solutions from three dimensions: feasibility, excellence, and economy. It then transforms the optimal solution into MES system instructions for implementation and scheduling optimization.

9. The expensive multi-objective scheduling optimization system for automotive parts assembly lines based on GAN according to claim 8, characterized in that, The GAN model also includes a multi-objective generator, a feasibility discriminator, and a superiority discriminator. The multi-objective generator includes a 128-bit input layer, a fully connected layer, and a 64-dimensional output layer. The feasibility discriminator includes convolutional layers and pooling layers to verify process constraints. The superiority discriminator includes a fully connected layer and a batch normalization layer to evaluate the multi-objective performance.

10. The expensive multi-objective scheduling optimization system for automotive parts assembly lines based on GAN according to claim 8, characterized in that, The scheme evaluation and output module is used to interface with the MES system of the automotive parts assembly line, and to convert the optimal scheduling scheme into instructions from the MES system.

Images