A self-adaptive optimization method for process forming scheduling workshop process parameters

Abstract

The application provides a self-adaptive optimization method for process forming scheduling workshop process parameters, comprising the following steps: feature extraction and feature alignment are performed on multi-modal data of the process forming scheduling workshop to obtain feature vectors of the modal data; the contribution degrees of the modal data to the target are used to calculate the modal weight, and the feature vectors are weighted and fused; the process parameters and the processes are respectively coded into chromosomes, an adaptability function is constructed according to the weighted and fused feature vectors, the influence weight of the process parameters is calculated based on the sensitivity of the process parameters and the modal contribution degree, and the cross probability and the mutation probability of each parameter gene are adaptively adjusted according to the influence weight, so that the optimal process parameter and the process scheduling scheme are solved; the optimal process parameter and the process scheduling scheme are input into a simulation system and compared with online measured results, if the error between the predicted defect rate and the actual defect rate exceeds a threshold value, the step length of genetic search and the contribution degrees of the modes are updated and then re-iterated, otherwise, the optimal process parameter and the process scheduling scheme are issued to the production line for execution.

Description

Technical Field

[0001] This invention belongs to the field of manufacturing process system optimization technology, specifically relating to an adaptive optimization method for process parameters in a process forming scheduling workshop. Background Technology

[0002] In complex systems such as intelligent manufacturing, medical diagnosis, and smart cities, information is distributed across different sensors, processes, and business systems, forming multimodal data with diverse forms. To overcome the limitations of single-modality data, researchers have proposed multimodal data integration techniques: by aggregating data from different sources and using deep learning models to extract cross-modal latent features, a more comprehensive and three-dimensional characterization of the system state can be achieved. However, this high-dimensional heterogeneous information fusion brings new optimization challenges—a huge search space and numerous local extrema, making it difficult for traditional single optimization strategies to simultaneously address global exploration and fine-tuning of local areas.

[0003] To address these challenges, the industry has developed several complementary technological approaches. At the data integration level, existing methods primarily include three architectures: ETL (Extract, Transform, Load) data processing, data warehouse and data lake collaboration, and data federation. ETL uses specialized tools to extract, clean, and transform data from various data sources before loading it into a data warehouse for subsequent analysis. Data warehouse and data lake collaboration achieves elastic storage and parallel computing by having the warehouse manage structured data in a schema-driven manner, while the data lake accommodates structured, semi-structured, and unstructured data. Data federation employs a virtual inheritance approach, dynamically integrating multiple heterogeneous data sources during queries to provide a unified view without centralized storage. However, ETL methods suffer from complex preprocessing, insufficient real-time performance, and difficulty adapting to frequent changes in data sources. While data warehouse and data lake collaboration offers strong scalability, the warehouse's reliance on predefined schemas and the potential for data corruption within the data lake lead to inconsistent data quality. Data federation, while avoiding centralized storage, struggles with updating and maintaining consistency across multi-source heterogeneous data. Therefore, the current mainstream approach is to define the global optimization space using a genetic algorithm as the main framework, followed by gradient optimization for refined local searches, thereby quickly converging to the optimal solution. Genetic algorithms, based on population evolution mechanisms of natural selection and genetic variation, possess global search capabilities and can locate potential optimal regions in a vast solution space; gradient descent algorithms, on the other hand, have local refinement capabilities by iteratively adjusting parameters and minimizing the loss function.

[0004] Despite the widespread application of the above-mentioned approaches, certain problems remain. While genetic algorithms can perform global searches, their computational cost is enormous in high-dimensional, multimodal scenarios, and the population size and number of generations are difficult to set. Gradient descent, while allowing for fine-tuning of parameters, is prone to getting trapped in local minima and is sensitive to the learning rate and initial point. Although a hybrid approach is complementary, there is no unified theoretical guidance on how to dynamically switch search strategies and coordinate global and local trade-offs, resulting in a significant contradiction between convergence speed and solution quality. Summary of the Invention

[0005] This invention proposes an adaptive optimization method for process parameters in the process forming scheduling workshop, which solves the problem of global and local search imbalance in existing technologies under multimodal conditions.

[0006] To address the aforementioned technical problems, this invention provides an adaptive optimization method for process parameters in a process forming scheduling workshop, comprising the following steps:

[0007] Step S1: Collect multimodal data from the process forming scheduling workshop, extract and align features from the multimodal data to obtain feature vectors for each modality;

[0008] Step S2: Calculate the modal weights based on the contribution of each modal data to the target, and perform weighted fusion of the feature vectors based on the modal weights;

[0009] Step S3: Use a genetic algorithm to solve for the optimal process parameters and process scheduling scheme in the process forming scheduling workshop: encode the process parameters and processes into chromosomes respectively, construct a fitness function based on the weighted fused feature vector, calculate the influence weight of the process parameters based on the sensitivity and modal contribution of the process parameters, and dynamically adjust the crossover probability and mutation probability of the process parameters based on the influence weight;

[0010] Step S4: Input the optimal process parameters and process scheduling scheme into the simulation system, calculate the error between the predicted defect rate and the actual defect rate. If the error exceeds the set error threshold, update the step size of the genetic algorithm and the contribution of each mode, and return to step S2. Otherwise, send the optimal process parameters and process scheduling scheme to the production line of the process forming scheduling workshop for execution.

[0011] Preferably, the expression for the contribution of the modal data to the target in step S2 is:

[0012] ;

[0013] In the formula, For the first i The contribution of each modality of data; The system response cycle; For the first iModal displacements of modal data; Expressing the request The first derivative; The weights of the i-th modality data; This represents the number of types of modal data.

[0014] Preferably, the expression for the modal weights in step S2 is:

[0015] ;

[0016] In the formula, For the first i Weights of modal data; For the first i The contribution of each modality of data.

[0017] Preferably, the expression for constructing the fitness function based on the weighted fused feature vector in step S3 is:

[0018] ;

[0019] ;

[0020] In the formula, The fitness function; For the first i Weights of modal data; For process parameter combinations and scheduling strategy Next i Performance metrics for modal data; This represents the total number of modal data types. This is a penalty term for violating the parameter feasible region and scheduling constraints; This is the penalty coefficient; This represents the total number of types of process parameters; For the first k One process parameter; , The first k Upper and lower limits of each process parameter; This is a penalty term for scheduling constraints.

[0021] Preferably, the expression for the sensitivity of the process parameters in step S3 is:

[0022] ;

[0023] ;

[0024] In the formula, For the first k The sensitivity of each process parameter; Let the policy fitness function be used. For the first k One process parameter; , , These are the weighting coefficients for quality priority, energy consumption priority, and delivery priority, respectively. For process parameter combinations and scheduling strategy Predict the product defect rate; For process parameter combinations and scheduling strategy The total energy consumption is predicted below; For scheduling strategy The maximum completion time.

[0025] Preferably, the calculation of the influence weight of the process parameters based on the sensitivity of the process parameters and the contribution of each modal data in step S3 includes the following steps:

[0026] Step S31: Construct the correlation matrix between modal data and process parameters :

[0027] ;

[0028] ;

[0029] In the formula, Let be the correlation strength between the i-th mode and the k-th process parameter, where , This represents the total number of modal data types. , This represents the total number of types of process parameters; This indicates the calculation of the correlation coefficient; Let i be the feature vector of the i-th modality data; This refers to the k-th process parameter;

[0030] Step S32: Map the contribution of modal data to process parameters:

[0031] ;

[0032] In the formula, The total modal contribution of the kth process parameter;

[0033] Step S33: Calculate the influence weights of process parameters:

[0034] ;

[0035] In the formula, For the first kThe influence weight of each process parameter; This is the balance coefficient; For the first k The sensitivity of each process parameter.

[0036] Preferably, the expression for the crossover probability of dynamically adjusting the process parameters according to the influence weights in step S3 is:

[0037] ;

[0038] In the formula, For the adjusted number k Crossover probabilities corresponding to each process parameter; For the first k The original crossover probabilities corresponding to each process parameter; This is the crossover probability adjustment coefficient; For the first k The influence weight of each process parameter;

[0039] The expression for the variation probability of dynamically adjusting process parameters based on the aforementioned influence weights is as follows:

[0040] ;

[0041] In the formula, For the adjusted number k The probability of variation corresponding to each process parameter; For the first k The original variation probability corresponding to each process parameter; This is the variation probability adjustment coefficient.

[0042] Preferably, after the genetic algorithm performs the crossover operation in step S3, a new population is obtained. Several locally optimal solutions with the highest fitness function values ​​are selected from the new population for local fine-tuning. The expression for local fine-tuning is:

[0043] ;

[0044] ;

[0045] In the above formula, For the first Process parameters at the next iteration; For the first Process parameters at the next iteration; For the first The learning rate for each iteration; For the first The gradient of the fitness function at the next iteration; The initial learning rate; This is the decay factor of the learning rate; This represents the current iteration number.

[0046] Preferably, in step S3, when using a genetic algorithm to solve for the optimal process parameters of the process forming scheduling workshop, every [time period]... N The variable neighborhood search algorithm is used to search for the local optimal solution.

[0047] Preferably, in step S1, the alignment of multimodal features in the semantic space is achieved through a cross-attention mechanism.

[0048] The beneficial effects of the present invention include at least the following:

[0049] 1. By collecting multimodal data, information about the manufacturing process in the workshop can be obtained from multiple perspectives. After feature extraction and feature alignment of the multimodal data, feature vectors of each modality are obtained, enabling data from different sources to be fused and analyzed in the same feature space. This fully utilizes the effective information contained in various data and avoids information loss that may be caused by a single data modality.

[0050] 2. Calculate modal weights based on the contribution of each modal data to the target, and perform weighted fusion of feature vectors based on the weights. This contribution-based weighted fusion method can highlight data modalities that are more valuable for process optimization and scheduling, thereby more accurately reflecting key factors in the fused feature vectors and providing more accurate feature inputs for subsequent optimization steps.

[0051] 3. Calculate the influence weight of process parameters based on their sensitivity and modal contribution, and dynamically adjust the crossover and mutation probabilities of process parameters so that the genetic algorithm can perform more targeted searches based on the importance of parameters and their contribution to the data. Attached Figure Description

[0052] Figure 1 This is a schematic diagram of the method flow according to an embodiment of the present invention. Detailed Implementation

[0053] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the protection scope of the present invention.

[0054] like Figure 1 As shown in the figure, this embodiment of the invention provides an adaptive optimization method for process parameters in a process forming scheduling workshop, including the following steps:

[0055] Step S1: Collect multimodal data from the process forming scheduling workshop, extract and align features from the multimodal data to obtain the feature vectors of each modality.

[0056] Specifically, vision, force, current, and position sensors are installed on the equipment in the process forming scheduling workshop to collect multimodal raw data in real time, including images, mechanical and electrical parameters, and historical process records. After cleaning, denoising, and synchronizing each type of data, a complete dataset is constructed in the form of feature-labels and proportionally divided into training, validation, and test sets to provide a consistent and reproducible data foundation for subsequent modeling.

[0057] To address the characteristics of different modalities, the images are first fed into a convolutional neural network to extract local spatial features, and the text data and structured sensor data are converted into sequential semantic vectors through a Transformer encoder.

[0058] In a shared semantic space, a cross-modal attention mechanism is introduced to enable image features and text semantics to mutually focus on and align within the same semantic space, thus eliminating intermodal differences.

[0059] ;

[0060] In the formula, This is the attention matrix for image-to-text conversion; For activation functions; Features of the original image; The projection matrix; Features of the original text; For key projection;

[0061] The three types of feature vectors are paired and interacted using a cross-attention mechanism to obtain the aligned feature representations:

[0062] ;

[0063] In the formula, For value projection.

[0064] The aligned modal features are projected together into a unified embedding space to form a joint feature representation with consistent dimensions and semantic alignment.

[0065] ;

[0066] In the formula, For joint features; Features of the aligned image; Features of the aligned text.

[0067] In this embodiment of the invention, the image is obtained after aligning with the structured data. After aligning the image and text records, the result is... After aligning structured data with text records, the following is obtained: The generated joint feature representation This allows for direct use in subsequent fusion, weight calculation, and adaptive optimization of process parameters.

[0068] Step S2: Calculate the modal weights based on the contribution of each modal data to the target, and perform weighted fusion of the feature vectors based on the modal weights.

[0069] Specifically, for the feature vector of each mode Each modal module is constructed separately. The modal module quantifies the contribution of each mode to the gradient effect on the optimization objective and calculates the initial weights.

[0070] ;

[0071] ;

[0072] In the formula, For the first i Weights of modal data; For the first i The contribution of each modality of data; The system response cycle; For the first i The changing pattern of modal characteristics over time; Expressing the request The first derivative; For the first Weights of modal data; This represents the number of types of modal data.

[0073] The weights are strictly mapped to the positive number space through an exponential transformation to avoid interference from zero or negative values. The modal features are then weighted and summed based on the transformed weights to obtain the fused feature representation. :

[0074] .

[0075] At the start of optimization, the weights of each mode are evenly distributed. After each round of optimization, the system dynamically re-evaluates the contribution based on real-time changes in operating conditions and prediction errors, and adjusts the weights online to achieve continuous adaptive updates.

[0076] Step S3: Use a genetic algorithm to solve for the optimal process parameters and process scheduling scheme in the process forming scheduling workshop: encode the process parameters and processes into chromosomes respectively, construct a fitness function based on the weighted fused feature vector, calculate the influence weight of the process parameters based on the sensitivity and modal contribution of the process parameters, and dynamically adjust the crossover probability and mutation probability of the process parameters based on the influence weight.

[0077] Specifically, the optimal process parameters and process scheduling scheme for the process forming scheduling workshop are solved using a genetic algorithm, including the following steps:

[0078] Step S31: Map process parameters to chromosome gene fragments using real number encoding, and map processes to chromosome gene fragments using a combination of process sequence, start time and equipment assignment encoding. One chromosome represents the scheduling scheme and process parameters. Initialize the population size according to the preset process upper and lower limits and scheduling constraints to ensure that all process parameter combinations are within the feasible domain and all scheduling strategies satisfy the scheduling constraints.

[0079] Step S32: Construct a prediction model by using the weighted fusion of multimodal feature vectors as input to the prediction model and constructing a fitness function using the performance metrics output by the prediction model. :

[0080] ;

[0081] ;

[0082] In the formula, For process parameter combinations and scheduling strategy The first prediction model output i Performance metrics for modal data; This represents the total number of modal data types. This is a penalty term for violating the parameter feasible region and scheduling constraints; This is the penalty coefficient; This represents the total number of types of process parameters; For the first k One process parameter; , The first k Upper and lower limits of each process parameter; This is a penalty term for scheduling constraints.

[0083] The prediction model in this embodiment of the invention learns common, generalized feature representations across different tasks by sharing a low-level feature extraction layer. The prediction model includes an input layer, a shared feature extraction layer, and an output layer. The input layer receives multimodal fusion data; the shared feature extraction layer further extracts higher-level, task-general feature patterns from the fusion features; and the output layer learns to map the shared features to a specific performance metric, enabling the model to learn multiple related tasks simultaneously.

[0084] The prediction model is trained using historical data, with the input being the fusion features corresponding to historical process parameters. The model is trained end-to-end using the backpropagation algorithm in the cloud to achieve a one-to-one correspondence between performance and fitness. When the process parameters exceed the preset feasible range, a penalty term is introduced, which significantly reduces the fitness of the solution.

[0085] Step S33: Retain the one with the highest fitness. k A select few individuals directly enter the next generation, while the remaining individuals are selected in a roulette wheel based on their fitness ratio, ensuring population diversity.

[0086] Step S34: For the selected elite individuals, calculate the influence weight of the process parameters based on the modal contribution of the modal data corresponding to the sensitivity and process parameters, and dynamically adjust the crossover probability and mutation probability of the selected chromosomes according to the influence weight, including the following steps:

[0087] Step S341: Introduce a strategy determination module to the cloud platform. This module dynamically allocates multi-objective weights based on the current production task's quality urgency, delivery urgency, and energy consumption urgency indicators, and constructs a strategy fitness function.

[0088] ;

[0089] ;

[0090] ;

[0091] ;

[0092] In the formula, Let the policy fitness function be used. , , These are the weighting coefficients for quality priority, energy consumption priority, and delivery priority, respectively. For process parameter combinations and scheduling strategy Predict the product defect rate; For process parameter combinations and scheduling strategy The total energy consumption is predicted below; For scheduling strategy Maximum completion time; The required quality level for the order; The average quality level; Remaining time for the order; Standard delivery cycle; Current energy consumption cost; This represents the energy consumption cost threshold.

[0093] like If the value is greater than 0.7, then delivery takes priority. If the first condition is not met, the second condition is checked. If the value is greater than 0.6, then quality priority is applied; if neither of the above two conditions is met, then energy consumption priority is applied.

[0094] The weight coefficients in the policy fitness function are calculated using a cloud-based platform, and then mapped to trainable parameters.

[0095] ;

[0096] ;

[0097] ;

[0098]

[0099] In the formula, The fitness of the scheduling scheme; This is the delivery rate weighting coefficient; This is the utilization rate weighting coefficient; Energy consumption rate weighting coefficient; For order delay rate; This represents the average equipment utilization rate. This represents the standardized total energy consumption. For the first Utilization rate of equipment; K Total number of devices; Total energy consumption; This represents the highest energy consumption in history.

[0100] , , All allocations are dynamically determined by the cloud-based policy judgment module.

[0101] Step S342: Calculate the sensitivity coefficients of the process parameters using Local Sensitivity Analysis (Local SA):

[0102] ;

[0103] In the formula, For the firstk The sensitivity of each process parameter.

[0104] Step S343: Calculate the modal contribution of each modal data based on the system response energy:

[0105] .

[0106] Step S344: Construct the correlation matrix between modal data and process parameters :

[0107] ;

[0108] ;

[0109] In the formula, For the first i Type 1 mode and the first k The correlation strength between the process parameters, among which , This represents the total number of modal data types. , This represents the total number of types of process parameters; This indicates the calculation of the correlation coefficient; For the first i Feature vectors of modal data; For the first k Each process parameter.

[0110] Step S345: Map the contribution of modal data to process parameters:

[0111] ;

[0112] In the formula, For the first k Total modal contribution of each process parameter.

[0113] Step S346: Calculate the influence weights of process parameters:

[0114] ;

[0115] In the formula, For the first k The influence weight of each process parameter; This is the balance coefficient; For the first k The sensitivity of each process parameter.

[0116] Step S347: Perform a crossover operation on the selected individuals, that is, exchange gene segments in a random interval to generate a new population, and set the adaptive crossover probability as follows:

[0117] ;

[0118] In the formula, For the adjusted number k Crossover probabilities corresponding to each process parameter; For the first k The original crossover probabilities corresponding to each process parameter; The crossover probability adjustment coefficient is used to reduce the crossover rate for high-weight process parameters in order to retain superior genes, and to increase the crossover rate for low-weight process parameters in order to enhance exploration. For the first k The influence weight of each process parameter.

[0119] Step S348: Set the adaptive mutation probability:

[0120] ;

[0121] In the formula, For the adjusted number k The probability of variation corresponding to each process parameter; For the first k The original variation probability corresponding to each process parameter; The mutation probability adjustment coefficient is used to increase the mutation amplitude for high-weight process parameters to escape local optima and prepare for subsequent fine-grained search, while reducing the mutation amplitude for low-weight process parameters to reduce invalid calculations.

[0122] Step S35: From the newly generated population, select several local optima with the highest fitness function values ​​for local fine-tuning. Calculate the gradient of their fitness function using numerical differencing, and update the parameters along the negative gradient direction.

[0123] ;

[0124] ;

[0125] In the above formula, For the first Process parameters at the next iteration; For the first Process parameters at the next iteration; For the first The learning rate for each iteration; For the first The gradient of the fitness function at the next iteration; The initial learning rate; This is the decay factor of the learning rate; This represents the current iteration number.

[0126] For individuals with high fitness, the gradient of the fitness function is calculated using numerical differencing. Learning rate It no longer relies solely on the number of iterations, but dynamically adjusts based on the continuous changes in the fitness function; when Increase the learning rate when it is continuously decreasing, and decrease it if there is fluctuation or increase.

[0127] In the early stages, a larger step size is used to quickly approach the optimal region, while in the later stages, the step size is reduced to improve convergence stability and accuracy. Individuals that have undergone gradient updates are reinserted into the population to replace those with low fitness, and elite individuals optimized by gradients are retained to directly enter the next generation, avoiding the loss of high-quality solutions.

[0128] Step S36: Each N Introducing a variable neighborhood perturbation randomly resets some parameters, thus avoiding the problems of discrete solution space and isolated optimal regions, and enhancing the diversity of parameter sets; as the number of iterations increases... N Decreasing, the amplitude of the disturbance decreases synchronously:

[0129] Coarse-grained: Randomly resets approximately 40% of parameters, allowing for large-scale jumps to escape local extrema;

[0130] Medium granularity: Randomly reset 10%-40% of parameters, and reposition with a moderate amplitude;

[0131] Fine-grained: Only a few key genes are fine-tuned, improving local precision.

[0132] Step S37: Repeat steps S32 to S36 until the maximum number of iterations or the fitness improvement threshold is reached, and output the globally optimal combination of process parameters.

[0133] Step S4: Input the optimal process parameters into the simulation system, calculate the error between the predicted defect rate and the actual defect rate. If the error exceeds the set error threshold, update the step size of the genetic algorithm and the contribution of each mode. After each round or N rounds of optimization, train based on the latest data in a small batch of samples to dynamically adjust the contribution of each mode. Return to step S2. Otherwise, output the optimal process parameters to the production line of the process forming scheduling workshop.

[0134] Specifically, the optimized process parameters are input into a high-precision simulation system to predict the defect rate, the actual defect rate is collected online, and compared with the simulation prediction value; if the deviation between the two exceeds the set threshold, the current parameter is determined to be invalid.

[0135] Once the deviation exceeds the limit, the process parameters are automatically updated, and the step size of the genetic-gradient hybrid algorithm and the contribution of each mode are adjusted simultaneously, and the iterative optimization is restarted.

[0136] Modal weights are treated as trainable parameters and optimized using gradient descent based on the final prediction error. A trainable adaptive weight parameter is initialized for each modality, and after each or N rounds of optimization, it is trained using a mini-batch of the latest data to dynamically adjust the contribution of each modality.

[0137] Each time a new round of optimization parameters is obtained, they are immediately sent to the production line for real-time monitoring; if the product quality still does not meet the target, the simulation-actual measurement-correction closed loop is triggered again to continuously improve the product.

[0138] When the simulation-predicted defect rate is consistently lower than the specified value and the actual verification is consistent, the iteration is terminated, and the process parameters are officially put into mass production application.

[0139] To illustrate the effectiveness of the method provided in this embodiment of the invention, this method is used to perform multimodal adaptive optimization of the dynamic scheduling decision parameters of a composite material component molding workshop, including the following steps:

[0140] Step 1: Data Integration and Preprocessing

[0141] The following sensors are deployed on each autoclave:

[0142] Real-time sensor data: used to collect temperature, pressure, vacuum, current and voltage data.

[0143] Process document data: Obtain the process formula document for each production task from the MES system.

[0144] Equipment status data: Obtain the autoclave's maintenance records and health status scores from the equipment management system.

[0145] Order data: This includes order delivery date, priority, component type, and other data.

[0146] Data preprocessing:

[0147] Sensor data is filtered using a moving average to eliminate noise. Unstructured process document data is parsed into structured process curve feature vectors. The amount of data collected in this embodiment is shown in Table 1.

[0148] Table 1 Data Collection Volume

[0149]

[0150] Step 2: Multimodal Feature Extraction and Alignment

[0151] A one-dimensional convolutional neural network (1D-CNN) is used to extract deep features that characterize the current operating state of the device, and outputs a 32-dimensional feature vector.

[0152] A Transformer encoder is used to process the process curve features, outputting a 32-dimensional process recipe feature vector. A cross-attention mechanism is constructed to align the equipment state features and process recipe features in the semantic space, with the aim of predicting the total estimated time and energy consumption for running the recipe on this equipment. The aligned joint feature has a dimension of 64, which serves as the input for subsequent optimization algorithms.

[0153] Step 3: Adaptive Modal Weight Adjustment Mechanism

[0154] Initial weights: process formulation mode =0.7, Device state mode =0.3. During operation, it was found that for a high-performance piece of equipment that had just undergone a major overhaul, its heating rate was much faster than the standard value. The system automatically increased the weight of the equipment state mode to 0.3. =0.5, =0.5, thus more accurately predicting the process time that the equipment can shorten.

[0155] Step 4: Global Exploration Phase of Genetic Algorithm

[0156] Encoding: A single chromosome contains the task start time, equipment assignment, and key process parameters. There are 50 tasks, so the chromosome length is 50. A hybrid encoding method is used. A single chromosome contains the scheduling information (start time, equipment assignment) and key process parameters for all tasks. Each task is assigned 3 gene loci.

[0157] Initialize the population: Randomly generate 150 scheduling schemes, i.e., population size = 150, ensuring that the schemes meet the constraint that the task must start when the equipment is available and the capacity allows.

[0158] The weights are dynamically adjusted and the weighted features are updated to ensure that different modal information is considered in each round. The combined weights of each modality after adjustment are shown in Table 2.

[0159] Table 2. Overall weights of modes

[0160]

[0161] Step 5: Gradient-guided local fine-tuning stage:

[0162] After the genetic algorithm runs for 30 generations, the 20 individuals with the highest fitness in the population are selected for local fine-tuning.

[0163] Gradient calculation: For the selected individual, the gradient of the fitness function with respect to the start time of each task is calculated by numerical difference.

[0164] Parameter update: Fine-tune the start time along the negative gradient direction (i.e., the direction that reduces the fitness value).

[0165] Population update: Replace the finely tuned high-quality individuals with the lowest fitness in the population.

[0166] Step 6: VNS perturbation

[0167] Every N Each iteration introduces a perturbation, and as the iteration continues, the perturbation frequency decreases and the perturbation amplitude decreases.

[0168] Coarse-grained perturbation: Randomly select 30% of the tasks and reassign them to devices.

[0169] Medium-grained perturbation: Randomly select 15% of the tasks and randomly offset their start times within a window of [-4 hours, +4 hours].

[0170] Fine-grained perturbation: Randomly select two high-priority tasks and swap their order with that of adjacent tasks on the current device.

[0171] Step 7: Feedback Adjustment and Closed-Loop Update

[0172] The optimal scheduling plan is sent to the MES system for execution. The system monitors the actual temperature and pressure curves of each autoclave in real time. Once a deviation between the actual process curve and the standard curve lasts for more than 10 minutes, a new emergency order with priority 1 is inserted, or a sudden equipment failure alarm is triggered, the re-optimization process is immediately initiated. Using the actual status of all tasks at the current moment as the new initial conditions, a new optimized scheduling plan is quickly generated within 5 minutes and executed on a rolling basis.

[0173] The comparison between the actual detection results and the model prediction results is shown in Table 3.

[0174] Table 3 Comparison between actual detection results and model prediction results

[0175]

[0176] The final optimized parameters are fed back to the Manufacturing Execution System (MES). Table 4 shows a comparison of the overall system performance before and after process parameter optimization.

[0177] Table 4 Comparison of Overall System Effectiveness

[0178]

[0179] Traditional methods often employ ETL, data warehouses / data lakes, or data federation technologies for multimodal data integration, which suffer from insufficient real-time performance and poor data quality. In contrast, the dynamic multimodal data fusion method used in this invention employs cross-modal feature alignment and adaptive weight allocation techniques to fuse multimodal weight matrices in real time, thereby improving the accuracy and robustness of multimodal feature representation.

[0180] Genetic algorithms suffer from slow convergence due to low global search efficiency, while gradient descent algorithms are prone to getting trapped in local optima. This invention employs a hybrid gradient genetic algorithm to drive a closed-loop compensation system for process parameters. The genetic algorithm framework globally explores the solution space to locate potential optimal regions, and then introduces VNS perturbations for dynamic adjustment. Gradient descent is switched within the neighborhood of a high-quality solution, utilizing the gradient information of the objective function for refined iteration, minimizing processing errors, energy consumption, and time. The genetic algorithm extensively explores the solution space, while the gradient descent algorithm performs fine-tuning near the discovered high-quality solution, thus rapidly converging to the accurate solution.

[0181] The technical features of the above embodiments can be combined arbitrarily. For the sake of brevity, not all possible combinations of the technical features in the above embodiments are described; only preferred embodiments of the present invention are illustrated. The descriptions are relatively specific and detailed, but they should not be construed as limiting the scope of the present invention. As long as the combination of these technical features does not contradict each other, it should be considered within the scope of this specification.

[0182] It should be noted that those skilled in the art can make various modifications and improvements without departing from the inventive concept, and these all fall within the scope of protection of this invention. Therefore, the scope of protection of this invention should be determined by the appended claims.

Claims

1. A method for adaptive optimization of process parameters in a process shaping dispatch shop, characterized in that, The method comprises the following steps: Step S1: Collecting multi-modal data of a process forming scheduling workshop, performing feature extraction and feature alignment on the multi-modal data to obtain feature vectors of each modal data; Step S2: Calculating modal weights according to the contribution of each modal data to the target, and performing weighted fusion on the feature vectors according to the modal weights; Step S3: Solving the optimal process parameters and process scheduling scheme of the process forming scheduling workshop by using a genetic algorithm: encoding the process parameters and processes into chromosomes, constructing a fitness function according to the feature vectors after weighted fusion, calculating the influence weight of the process parameters according to the sensitivity of the process parameters and the modal contribution, and dynamically adjusting the crossover probability and mutation probability of the chromosomes according to the influence weight; the calculation of the influence weight of the process parameters according to the sensitivity of the process parameters and the modal contribution comprises the following steps: Step S31: constructing a correlation matrix between modal data and process parameters : ； ； In the formula, For the first i Type 1 mode and the first k The correlation strength between the process parameters, among which , This represents the total number of modal data types. , This represents the total number of types of process parameters; This indicates the calculation of the correlation coefficient; For the first i Feature vectors of modal data; For the first k One process parameter; Step S32: Mapping the contribution of the modal data to the process parameters: ； In the formula, is the total modal contribution degree of the first k process parameter; is the contribution degree of the first i modal data; Step S33: Calculating the influence weight of the process parameters: ； wherein is the influence weight of the jth k process parameter; is the balancing factor; is the sensitivity of the jth k process parameter; Step S4: Inputting the optimal process parameters and process scheduling scheme into a simulation system to calculate the error between the predicted defect rate and the actual defect rate, if the error exceeds the set error threshold, updating the step length of the genetic algorithm and the contribution of each modal, returning to step S2, otherwise, issuing the optimal process parameters and process scheduling scheme to the production line of the process forming scheduling workshop for execution.

2. The method of claim 1, wherein the method is characterized by: The expression of the contribution of the modal data to the target in step S2 is: ； In the formula, For the first i The contribution of each modality of data; The system response cycle; For the first i Modal displacements of modal data; Expressing the request The first derivative; For the first i Weights of modal data; This represents the number of types of modal data.

3. The method of claim 1, wherein: the process shaping schedule is a process shaping schedule for a job shop; and the process parameters are process parameters for the job shop. The expression of the modal weight in step S2 is: ； In the formula, is the weight of the first i modal data; is the contribution degree of the first i modal data.

4. The adaptive optimization method for process parameters in a process forming scheduling workshop according to claim 1, characterized in that: The expression of constructing the fitness function according to the feature vectors after weighted fusion in step S3 is: ； ； In the formula, is a fitness function; is a weight of the i th modality data; is a combination of process parameters and scheduling strategy , the performance index of the i th modality data output by the prediction model with the fusion feature vector as the input; is the total number of types of modality data; is a penalty term for violating the parameter feasible region and scheduling constraints; is a penalty coefficient; is the total number of types of process parameters; is the k th process parameter; , are respectively the upper limit and the lower limit of the k th process parameter; is a penalty term for scheduling constraints.

5. The adaptive optimization method for process parameters in a process forming scheduling workshop according to claim 1, characterized in that: The expression of the sensitivity of the process parameters in step S3 is: ； ； wherein, is the sensitivity of the jth process parameter; k is the strategy fitness function; is the jth process parameter; k , , are the weight coefficients for quality priority, energy consumption priority, and delivery priority, respectively; is the predicted product defect rate for the process parameter combination and the scheduling strategy ; is the predicted total energy consumption for the process parameter combination and the scheduling strategy ; is the maximum completion time for the scheduling strategy .​​ 6. The method of claim 1, wherein: The expression of dynamically adjusting the crossover probability of the chromosomes according to the influence weight in step S3 is: ； In the formula, For the adjusted number k Crossover probabilities corresponding to each process parameter; For the first k The original crossover probabilities corresponding to each process parameter; This is the crossover probability adjustment coefficient; For the first k The influence weight of each process parameter; The expression of dynamically adjusting the mutation probability of the chromosomes according to the influence weight is: ； In the formula, is the adjusted variation probability corresponding to the i th process parameter; k is the adjusted variation probability corresponding to the i th process parameter; is the original variation probability corresponding to the i th process parameter; k is the original variation probability corresponding to the i th process parameter; is the variation probability adjustment coefficient.

7. The method of claim 1, wherein: After the genetic algorithm in step S3 performs a crossover operation, a new population is obtained, and a number of local optimal solutions with the highest fitness function values are selected from the new population for local fine-tuning, and the expression of the local fine-tuning is: ； ； In the above formula, is the process parameter at the first iteration; is the process parameter at the first iteration; is the process parameter at the first iteration; is the process parameter at the first iteration; is the learning rate for the first iteration; is the learning rate for the first iteration; is the gradient of the fitness function at the first iteration; is the gradient of the fitness function at the first iteration; is the initial learning rate; is the decay coefficient for the learning rate; is the current iteration number.

8. The method of claim 1, wherein: In step S3, the genetic algorithm is used to solve the optimal process parameters of the process forming scheduling workshop. Every 50 generations N The variable neighborhood search algorithm is used for local search of the optimal solution.

9. The method of claim 1, wherein: the process shaping schedule is a process shaping schedule for a job shop; and the process parameters are process parameters for the job shop. In step S1, the multi-modal features are aligned in the semantic space through a cross-attention mechanism.

Images