A hardware manufacturing process parameter optimization method based on machine learning

Abstract

This invention discloses a machine learning-based method for optimizing manufacturing process parameters in hardware products, belonging to the field of industrial manufacturing process optimization technology. The method includes: constructing a dynamic benchmark dataset integrating raw material properties and generating a raw material state representation vector; constructing and training a quality prediction model with process parameters and raw material state as dual inputs; using multi-objective Bayesian optimization to iteratively optimize within the process parameter space, with minimizing thickness deviation and surface roughness as dual objectives, to obtain a Pareto optimal solution set; implementing parameter optimization and collecting actual production data; and dynamically updating the dataset and prediction model using qualified production data in a closed-loop manner. This invention explicitly considers the impact of raw material fluctuations, achieves automatic collaborative optimization of multiple quality objectives, and enables the system to continuously adapt to changes in the production process through an online learning mechanism, significantly improving the robustness, accuracy, and long-term effectiveness of process optimization.

Description

Technical Field

[0001] This invention relates to the field of industrial manufacturing process optimization technology, and in particular to a method for optimizing manufacturing process parameters of hardware products based on machine learning. Background Technology

[0002] The manufacturing of hardware products, such as metal parts produced through plastic forming processes like stamping, forging, extrusion, and die casting, relies heavily on the quality of the final product (e.g., dimensional accuracy, mechanical strength, surface finish) to the process parameters set during manufacturing. These parameters typically include, but are not limited to, mold temperature, material feed rate, forming pressure, holding time, punch speed, and lubrication conditions. In current production practices, setting these process parameters primarily depends on engineers' experience and recommendations from process manuals. When faced with new materials, new molds, or new quality requirements, companies often employ a "trial and error" approach to parameter adjustment. This involves setting several sets of parameters based on experience, conducting trial production, testing product quality, and then repeatedly adjusting the parameters until the product meets specifications. This method has three major drawbacks: First, the debugging cycle is long, consuming a lot of raw materials, energy and labor, resulting in high production costs; second, the obtained parameter combinations are often only "feasible solutions" rather than "optimal solutions", making it difficult to maximize production efficiency or minimize energy consumption while ensuring quality; third, the process has poor stability. When the performance of raw material batches fluctuates (such as slight changes in the hardness and ductility of metal ingots) or the equipment status drifts slowly, the original "optimal" parameters may become invalid, leading to an increase in the defect rate, and re-debugging will trigger a new round of cost losses.

[0003] To address the aforementioned issues, existing technologies have introduced parameter optimization methods based on statistical models (such as response surface methodology) or traditional machine learning models (such as support vector machines and random forests). The general process of these methods involves collecting historical production data (process parameters and corresponding quality indicators), training a predictive model describing the "parameter-quality" relationship, and then using optimization algorithms to find the optimal parameter combination in the parameter space that yields the best predicted quality. However, these methods still have fundamental shortcomings when dealing with complex hardware manufacturing scenarios. First, they are typically based on an implicit and unrealistic assumption: that all other conditions in the production process, especially the properties of raw materials, are constant. In reality, raw materials for hardware products (such as aluminum ingots and steel coils) exhibit inherent batch variations. This variation acts as a powerful perturbation variable, severely affecting the stability of the "process parameter-product quality" mapping relationship. Models that ignore this variable will suffer significant reductions in predictive accuracy and robustness. Second, these methods are mostly "static" or "one-off" optimizations. Even after training a model with historical data, this model is then fixed and used to guide future production. However, the manufacturing process is time-varying. Factors such as equipment wear and environmental changes cause process characteristics to drift slowly, and static models will gradually deviate from the actual working conditions, leading to the failure of optimization suggestions. Finally, existing methods often only construct a single quality prediction model, without fully considering the potential conflicts and coupling relationships between different quality indicators (such as size and strength). Their optimization objectives are singular, making it difficult to achieve synergistic optimization of multiple quality objectives.

[0004] Therefore, there is an urgent need for an intelligent process parameter optimization method that can overcome the above-mentioned defects, especially one that can explicitly handle raw material fluctuations and has continuous self-evolution capabilities, in order to achieve robust, precise and continuously optimal operation of the hardware manufacturing process. Summary of the Invention

[0005] To achieve the above objectives, this invention provides a method for optimizing manufacturing process parameters of hardware products based on machine learning, comprising the following steps:

[0006] Step 1: Construct a dynamic benchmark dataset and raw material state representation vector; Step 1 includes: For each production batch, synchronously collect the process parameter set, online quality index set, and raw material batch information, and store them together; wherein, the process parameter set includes die temperature, stamping speed, holding pressure, and lubrication flow; the online quality index set includes thickness deviation and surface roughness; the raw material batch information includes yield strength and elongation; select qualified historical data to form an initial benchmark dataset; calculate the dynamic raw material state representation vector for the current batch of raw materials;

[0007] Step 2: Construct a dual-channel quality prediction model that integrates the raw material state and train it using the initial benchmark dataset; the input of the dual-channel quality prediction model includes the set of process parameters to be evaluated and the dynamic raw material state representation vector, and the output is the predicted thickness deviation and surface roughness.

[0008] Step 3: Real-time optimization of process parameters based on multi-objective Bayesian optimization; Step 3 includes: using the trained dual-channel quality prediction model as the evaluator, minimizing the predicted thickness deviation and minimizing the predicted surface roughness as the dual objectives, iterative optimization is performed within the safe operating range of the process parameters, and finally the Pareto optimal solution set composed of non-dominated solutions is obtained.

[0009] Step 4: Implement the recommended optimal combination of process parameters and collect and record verification data after actual production;

[0010] Step 5: Perform online updates of the dynamic benchmark dataset and prediction model; Step 5 includes: when the verification data record is qualified, add it to the dynamic benchmark dataset; when the dataset update meets the preset conditions, trigger incremental learning of the dual-channel quality prediction model.

[0011] Preferably, the construction of the dynamic benchmark dataset and raw material state representation vector in step 1 specifically includes the following sub-steps:

[0012] Step 1.1: Data Acquisition. During the stable production phase of each production batch, real-time values ​​of the mold temperature, stamping speed, holding pressure, and lubrication flow rate are acquired through the production line control system and online sensors, serving as the set of process parameters. The thickness of multiple consecutive workpieces is acquired using a high-precision thickness gauge, and the statistical value of thickness deviation is calculated. The surface profile of multiple consecutive workpieces is acquired using a surface profiler, and the statistical value of surface roughness is calculated, collectively serving as the set of online quality indicators. Before batch production, the raw materials are sampled and subjected to standard mechanical property tests to obtain the measured values ​​of yield strength and elongation, serving as the batch information of the raw materials.

[0013] Step 1.2: Data association and storage. The process parameter set, online quality index set and raw material batch information collected from the same production batch are associated to form a complete data record and stored in the production database.

[0014] Step 1.3: Select qualified data to form the initial benchmark dataset. Set the upper limit threshold for thickness deviation and the upper limit threshold for surface roughness. From all historical data records in the production database, select data records where both thickness deviation and surface roughness are lower than the corresponding upper limit threshold. The set of qualified historical data is the initial benchmark dataset.

[0015] Step 1.4: Calculate the dynamic raw material state characterization vector. For the current batch to be optimized, the yield strength and elongation of its raw materials are known. From the initial benchmark dataset, calculate the absolute difference between the yield strength of the raw materials in all historical data records and the yield strength of the current batch, as well as the absolute difference between the elongation of the raw materials and the elongation of the current batch. Add these two absolute differences to obtain the comprehensive difference value between each historical record and the current batch of raw materials. Select the top N historical data records in ascending order of comprehensive difference value, where N is a preset positive integer. Calculate the average value of each parameter in the process parameter set of these N historical data records to obtain the historical process parameter mean vector. Calculate the average value of each indicator in the online quality indicator set of these N historical data records to obtain the historical quality indicator mean vector. Sequentially concatenate the historical process parameter mean vector, the historical quality indicator mean vector, and the current raw material performance vector composed of the yield strength and elongation of the current batch of raw materials to generate the dynamic raw material state characterization vector.

[0016] Preferably, step 2, which involves constructing and training a dual-channel quality prediction model that integrates the raw material states, specifically includes the following sub-steps:

[0017] Step 2.1: Model input definition. The dual-channel quality prediction model has two input ports. The first input port is used to receive a four-dimensional vector, which corresponds to the set of process parameters to be evaluated, including four parameter values: mold temperature, stamping speed, holding pressure, and lubrication flow rate. The second input port is used to receive a vector of a specific dimension generated in step 1.4, which is the dynamic raw material state characterization vector.

[0018] Step 2.2: Model Structure Construction. The dual-channel quality prediction model adopts a feedforward neural network architecture. The network channel processing the first input port is called the process parameter channel, which includes a first fully connected layer, a second fully connected layer, and a third fully connected layer. The first fully connected layer has 32 neurons, the second fully connected layer has 16 neurons, and the third fully connected layer has 8 neurons. The network channel processing the second input port is called the raw material status channel, which includes a fourth fully connected layer and a fifth fully connected layer. The fourth fully connected layer has 16 neurons, and the fifth fully connected layer has 8 neurons. The output of the third fully connected layer of the process parameter channel and the output of the fifth fully connected layer of the raw material status channel are fused in a splicing layer. The fused feature vector is input to the sixth fully connected layer, which has 12 neurons. The output of the sixth fully connected layer is connected to the seventh fully connected layer, which has 2 neurons. The output of this layer is the final output of the dual-channel quality prediction model.

[0019] Step 2.3: Model training. The dual-channel quality prediction model is trained under supervision using all qualified historical data records in the initial benchmark dataset. For each sample in the training dataset, a dynamic raw material state representation vector corresponding to the sample is generated according to the raw material batch information of the sample, as described in Step 1.4, and used as the input to the second input port of the model. The process parameter set of the sample is used as the input to the first input port of the model. The online quality index set of the sample is used as the target output of the model training. The loss function used in the training process is the mean squared error loss function, and the optimization algorithm is the Adam optimization algorithm. The training process continues until the prediction error of the model on the validation set no longer decreases significantly.

[0020] Preferably, step 3, which involves real-time optimization of process parameters based on multi-objective Bayesian optimization, specifically includes the following sub-steps:

[0021] Step 3.1: Define the optimization problem. The optimization variables are mold temperature, stamping speed, holding pressure, and lubrication flow rate. Each optimization variable has its corresponding lower and upper limits of safe operating range. There are two optimization objectives: the first objective is to minimize the thickness deviation predicted by the dual-channel quality prediction model, and the second objective is to minimize the surface roughness predicted by the dual-channel quality prediction model.

[0022] Step 3.2: Initialize the optimization process. Using the Latin hypercube sampling method, M sets of initial process parameter combinations are uniformly generated within a four-dimensional hypercube space defined by the lower and upper limits of the safe operating range of each optimization variable, where M is a preset positive integer. Each of the M sets of initial process parameter combinations, along with the fixed dynamic raw material state representation vector generated in Step 1.4 for the current production batch, is input into the trained dual-channel quality prediction model to obtain the predicted thickness deviation and predicted surface roughness corresponding to each set of parameters. The M sets of process parameter combinations and their corresponding two predicted target values ​​together constitute the initial observation sample set.

[0023] Step 3.3: Construct surrogate models. Based on the current set of observed samples, construct a Gaussian process regression model as a surrogate model for the first objective and the second objective respectively; the Gaussian process regression model uses the Marting 5 / 2 kernel function as the covariance function;

[0024] Step 3.4: Iterative optimization to calculate the expected hypervolume improvement acquisition function; the expected hypervolume improvement acquisition function is based on the two Gaussian process surrogate models and is used to quantify the contribution of any new sampling point in the process parameter space to the expected improvement of the current Pareto front; by maximizing the expected hypervolume improvement acquisition function through the optimization algorithm, the next combination of process parameters to be evaluated is found within the safe operating range of the process parameters;

[0025] Step 3.5: Evaluation and Update. Input the combination of process parameters found in Step 3.4 and the fixed dynamic raw material state characterization vector into the dual-channel quality prediction model to obtain the predicted thickness deviation and predicted surface roughness corresponding to the point. Add this new combination of process parameters and its predicted target value as a new sample to the observation sample set.

[0026] Step 3.6: Loop and terminate, repeat steps 3.3 to 3.5 until the number of iterations reaches the preset maximum number of iterations T;

[0027] Step 3.7: Extract the Pareto optimal solution set. After the optimization loop terminates, compare the Pareto dominance of all samples in the final observation sample set. For any two samples, if one sample is not inferior to the other sample in both the first and second objectives, and is strictly superior to the other sample in at least one objective, then the former is said to dominate the latter. All samples that are not dominated by any other sample constitute the Pareto optimal solution set.

[0028] Preferably, step 4 involves implementing the recommended optimal process parameter combination and collecting verification data records after actual production. Specifically, this includes: manually selecting a set of process parameter combinations as the recommended optimal process parameter combination from the Pareto optimal solution set, based on the different priority requirements of thickness deviation and surface roughness for the actual production task; transmitting the specific values ​​of mold temperature, stamping speed, holding pressure, and lubrication flow rate from the recommended optimal process parameter combination to the programmable logic controller (PLC) of the hardware manufacturing production line, which then drives the actuator to operate according to the set parameters; after completing at least one production shift or one minimum production batch, re-collecting the actual process parameter set, online quality index set, and batch information of the raw materials during that production period to form a new verification data record; the format of the verification data record is exactly the same as the data record format described in step 1.2.

[0029] Preferably, step 5 involves online updates of the dynamic benchmark dataset and the prediction model, specifically including the following sub-steps:

[0030] Step 5.1: Verify the validity of the data. Check the online quality index set in the verification data record to determine whether the thickness deviation is lower than its upper limit threshold and whether the surface roughness is lower than its upper limit threshold. If both conditions are met, the verification data record is determined to be a valid and successful sample and marked as usable for updating.

[0031] Step 5.2: Update the dynamic benchmark dataset by adding valid and successful samples that are marked as usable for updating to the current dynamic benchmark dataset; set the maximum capacity limit for the dynamic benchmark dataset; when the number of data records in the dynamic benchmark dataset exceeds the maximum capacity limit, remove the earliest data record according to the order in which the data records were entered into the database to maintain a constant dataset capacity;

[0032] Step 5.3: Trigger incremental learning of the model. Set a model update trigger counter with an initial value of 0. Whenever a new valid successful sample is added to the dynamic benchmark dataset, the model update trigger counter is incremented by 1. When the value of the model update trigger counter reaches a preset trigger threshold K, the incremental learning process is executed. The incremental learning process is as follows: using the updated dynamic benchmark dataset as the training set, using the current network weight parameters of the dual-channel quality prediction model as the initial values, and using a learning rate lower than that in the initial training phase of Step 2.3, a new round of training is performed on the dual-channel quality prediction model. After training is completed, the original network weight parameters are completely replaced with the newly obtained network weight parameters, and the model update trigger counter is reset to 0.

[0033] Preferably, the value range of the preset positive integer N in step 1.4 is 50 to 200; the dimension of the dynamic raw material status representation vector is 10-dimensional, wherein the historical process parameter mean vector is 4-dimensional, the historical quality index mean vector is 2-dimensional, and the current raw material performance vector is 2-dimensional, which are sequentially concatenated to obtain a 10-dimensional vector.

[0034] Preferably, in the model structure construction of step 2.2, the first fully connected layer, the second fully connected layer, the third fully connected layer, the fourth fully connected layer, the fifth fully connected layer, and the sixth fully connected layer are all connected with modified linear unit activation functions; the seventh fully connected layer is not connected with any nonlinear activation function.

[0035] Preferably, the value range of the preset positive integer M in step 3.2 is 10 to 30; and the value range of the preset maximum number of iterations T in step 3.6 is 50 to 150.

[0036] Preferably, the maximum capacity of the dynamic benchmark dataset in step 5.2 is limited to 5,000 to 10,000 records; and the preset trigger threshold K in step 5.3 ranges from 20 to 100.

[0037] The beneficial effects of this invention are:

[0038] 1. This invention employs a unique "dynamic raw material state representation vector" generation step, explicitly encoding raw material properties as core background information and constructing a "dual-channel quality prediction model." This model enables the model to learn the differentiated mapping relationships of process parameters under different raw material backgrounds. This allows optimization recommendations to automatically adapt to the current material characteristics, fundamentally solving the problem of unstable optimization results caused by material fluctuations, and improving the universality and reliability of the method in actual production.

[0039] 2. This invention simultaneously optimizes thickness deviation and surface roughness, employing a "multi-objective Bayesian optimization" framework and utilizing an "expected hypervolume improvement" acquisition function to efficiently explore the process parameter space. This method automatically finds a series of "Pareto optimal solutions," which clearly demonstrate the optimal trade-offs between different quality objectives, providing a rich and scientific selection of non-inferior solutions for production decisions, and achieving a leap from single-point optimization to multi-objective collaborative optimization.

[0040] 3. This invention constructs a complete closed loop of "production-verification-learning." By automatically incorporating qualified production data into a "dynamic benchmark dataset" and periodically triggering "incremental learning" of the model, the system can continuously update itself using newly generated production knowledge. This transforms the system from a static tool into a dynamic learning system, capable of tracking and adapting to production process drift, maintaining the accuracy and advancement of optimization suggestions over the long term, and significantly reducing the cost of subsequent manual maintenance and remodeling. Attached Figure Description

[0041] To more clearly illustrate the technical solutions in this invention or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, those skilled in the art can obtain other drawings based on these drawings without creative effort.

[0042] Figure 1 This is a flowchart of the steps of the method of the present invention;

[0043] Figure 2 A flowchart illustrating the steps of constructing a dynamic benchmark dataset and raw material state representation vectors according to the method of this invention;

[0044] Figure 3 This is a flowchart illustrating the steps of the method of the present invention for online updating of the dynamic benchmark dataset and the prediction model. Detailed Implementation

[0045] The present invention will now be described in detail with reference to the accompanying drawings and specific embodiments. It should also be noted that, to make the embodiments more comprehensive, the following embodiments are the best and preferred embodiments, and those skilled in the art can use other alternative methods to implement some well-known technologies; moreover, the accompanying drawings are only for more specific description of the embodiments and are not intended to specifically limit the present invention.

[0046] Please see Figures 1-3 This invention provides a machine learning-based method for optimizing manufacturing process parameters in hardware products. The core of this invention lies in constructing an intelligent closed-loop system for optimizing process parameters that can explicitly integrate raw material state information, perform multi-objective automatic optimization, and possess online self-evolution capabilities. The method sequentially executes the following five main steps: constructing a dynamic benchmark dataset and raw material state representation vectors; constructing and training a dual-channel quality prediction model that integrates raw material states; performing real-time optimization of process parameters based on multi-objective Bayesian optimization; implementing recommended process parameter combinations and collecting verification data; and updating the dynamic benchmark dataset and prediction model online. The following provides a detailed explanation of each step.

[0047] Step 1: Construct a dynamic benchmark dataset and raw material state representation vectors.

[0048] This step aims to establish a high-quality historical production knowledge base and generate a background vector that can quantitatively represent the current production conditions (the core being the state of raw materials), laying a data foundation for subsequent intelligent modeling and optimization.

[0049] Step 1.1: Data Acquisition. In the continuous production process of hardware products (such as stamped parts and forgings), three types of data need to be collected synchronously for each independent production batch. The first type is the process parameter set, which contains several key controllable process variables that have a significant impact on product quality. In a specific embodiment of this invention, the process parameter set includes die temperature, stamping speed, holding pressure, and lubrication flow rate. Data acquisition should be carried out after the production batch enters a stable production stage. The average value over a period of time is directly read or calculated by sensors and the control system on the production line to ensure the representativeness and stability of the data. For example, the die temperature is measured by a thermocouple embedded in the die, the stamping speed comes from the set value of the press controller or encoder feedback, the holding pressure is collected by a pressure sensor to obtain the peak value, and the lubrication flow rate is monitored by a flow meter.

[0050] The second category is the online quality indicator set, which contains key product quality characteristics acquired in real time through automated inspection equipment deployed upstream or downstream of the production line. In one specific embodiment of the invention, the online quality indicator set includes thickness deviation and surface roughness. Thickness deviation can be measured by a laser thickness gauge or machine vision system to determine the difference between the actual and nominal dimensions at a specific location on the workpiece, and the statistical value (e.g., root mean square error) of this deviation for multiple consecutive workpieces can be calculated. Surface roughness can be measured by a contact or non-contact surface profilometer, and the average roughness parameter (e.g., Ra value) for multiple consecutive workpieces can be calculated.

[0051] The third category is raw material batch information. Before each batch of raw materials is put into production, representative samples are drawn from that batch and subjected to laboratory testing according to national or industry standards to obtain key parameters that characterize its mechanical or physical properties. In one specific embodiment of the present invention, the raw material batch information includes yield strength and elongation. These two parameters are important indicators for measuring the formability of metallic materials.

[0052] Step 1.2: Data Association and Storage. The process parameter set, online quality indicator set, and raw material batch information collected in Step 1.1 for the same production batch are forcibly associated to form a production data record with complete context. This record is stored in a dedicated production database or manufacturing execution system database, and each record contains a unique batch identifier and the specific values ​​of the above three types of data.

[0053] Step 1.3: Select qualified data to form the initial benchmark dataset. Set the upper limit threshold for each quality indicator in the online quality indicator set. For example, set the upper limit threshold for thickness deviation to 0.05 mm and the upper limit threshold for surface roughness to 2.0 micrometers. Traverse all historical data records in the production database and select records where the measured value of each indicator in the online quality indicator set is lower than its corresponding upper limit threshold. The set of all selected qualified historical data records constitutes the initial benchmark dataset. This dataset is the knowledge source for subsequent model training, ensuring that the model learns the inherent laws of "successful production".

[0054] Step 1.4: Calculate the dynamic raw material state representation vector. When process optimization is required for a new production batch, first obtain the performance parameters of the raw materials for that batch, i.e., the current raw material performance vector, such as [yield strength, elongation]. Then, find the historical production experience most similar to the current raw material from the initial benchmark dataset. The specific method is: calculate the absolute difference between the raw material performance parameters of each historical record in the initial benchmark dataset and the corresponding parameters in the current raw material performance vector, and sum these absolute differences to obtain the comprehensive difference value between the historical record and the current batch of raw materials. Sort all historical records in ascending order of comprehensive difference value, and select the top N records, where N is a preset positive integer, with a value range, for example, between 50 and 200.

[0055] Next, the statistics for these N most similar historical records are calculated: the average value of each parameter in the process parameter set of these N records is calculated to obtain the historical process parameter mean vector; the average value of each indicator in the online quality indicator set of these N records is calculated to obtain the historical quality indicator mean vector. Finally, the historical process parameter mean vector, the historical quality indicator mean vector, and the current raw material performance vector are concatenated in a preset order to generate a dynamic raw material state representation vector. This vector is a comprehensive background descriptor, and its dimension is the sum of the dimensions of its components. For example, if the process parameter set is 4-dimensional, the quality indicator set is 2-dimensional, and the raw material performance is 2-dimensional, then the generated dynamic raw material state representation vector will be 8-dimensional.

[0056] Step 2: Construct a dual-channel quality prediction model that integrates the state of raw materials.

[0057] This step aims to build a neural network model that can simultaneously consider controllable process parameters and uncontrollable raw material background, and accurately predict product quality under given process parameters.

[0058] Step 2.1: Model Input Definition. The dual-channel quality prediction model is designed with two independent input ports. The first input port is used to receive the set of process parameters to be evaluated, i.e., a vector containing specific values ​​related to die temperature, stamping speed, holding pressure, and lubrication flow rate. The second input port is used to receive the dynamic raw material state representation vector generated in Step 1.4 for the current optimization task.

[0059] Step 2.2: Model Structure Construction. The model adopts a feedforward neural network architecture, containing two parallel processing channels. The channel processing the process parameter set is called the process parameter channel, and its structure includes a first fully connected layer, a second fully connected layer, and a third fully connected layer. The number of neurons in the first fully connected layer is set to 32, the second fully connected layer to 16, and the third fully connected layer to 8. The channel processing the dynamic raw material state representation vector is called the raw material state channel, and its structure includes a fourth fully connected layer and a fifth fully connected layer. The number of neurons in the fourth fully connected layer is set to 16, and the fifth fully connected layer to 8. A modified linear unit activation function is applied after each fully connected layer to introduce nonlinear transformation capability.

[0060] The output features of the third fully connected layer in the process parameter channel and the output features of the fifth fully connected layer in the raw material state channel are fed into a splicing layer for fusion, resulting in a higher-dimensional fused feature vector. This fused feature vector is then input to a sixth fully connected layer with 12 neurons, also using a modified linear unit activation function. The output of the sixth fully connected layer is connected to a seventh fully connected layer with 2 neurons, corresponding to the two quality indicators to be predicted: thickness deviation and surface roughness. No nonlinear activation function is connected after the seventh fully connected layer to achieve direct regression output for continuous values.

[0061] Step 2.3: Model Training. The dual-channel quality prediction model is trained in a supervised manner using the initial benchmark dataset constructed in Step 1.3. During training, for each sample in the dataset, a dynamic raw material state representation vector corresponding to the sample needs to be generated according to the method described in Step 1.4 (finding the N most similar historical records in the dataset centered on the sample's own raw materials), based on the sample's raw material batch information. This vector serves as the input to the second input port of the model. The process parameter set of the sample serves as the input to the first input port. The true online quality index set of the sample serves as the target output for model training.

[0062] The training loss function is the mean squared error loss function, which measures the difference between the model's predictions and the true values. The optimization algorithm used is the Adam optimization algorithm, a variant of stochastic gradient descent with an adaptive learning rate. The training process consists of multiple rounds, in which the training data is divided into several batches and input into the model in each round. Training continues until the model's performance on an independent validation dataset no longer shows significant improvement; at this point, the model is considered to have fully learned the mapping relationships in the data, and training is complete.

[0063] Step 3: Real-time optimization of process parameters based on multi-objective Bayesian optimization.

[0064] This step utilizes a trained prediction model to automatically find the optimal or near-optimal combination of process parameters, taking into account the conflict of multiple quality objectives.

[0065] Step 3.1: Define the optimization problem. The optimization variables are those included in the process parameter set: die temperature, stamping speed, holding pressure, and lubrication flow rate. Each variable has a preset safe operating range, determined by the physical limits of the equipment and basic process knowledge; for example, the die temperature range is 100 degrees Celsius to 200 degrees Celsius. There are two optimization objectives: the first objective is to minimize the thickness deviation predicted by the dual-channel quality prediction model; the second objective is to minimize the surface roughness predicted by the model. This is a typical bi-objective minimization problem.

[0066] Step 3.2: Initialize the optimization process. Using experimental design methods, such as Latin hypercube sampling, M sets of initial process parameter combinations are uniformly generated within a multidimensional space comprised of the safe operating ranges of each optimization variable. M is a preset positive integer, for example, between 10 and 30. For each set of initial process parameters, it is input together with the dynamic raw material state representation vector generated in Step 1.4, which is fixed for the current production batch, into the trained dual-channel quality prediction model to obtain the corresponding predicted thickness deviation and predicted surface roughness. These M sets of process parameters and their corresponding two predicted target values ​​together constitute the initial observation sample set for the Bayesian optimization process.

[0067] Step 3.3: Constructing surrogate models. Based on the current set of observed samples, Gaussian process regression models are constructed as surrogate models for the two optimization objectives (thickness deviation and surface roughness). A Gaussian process is a nonparametric probabilistic model that can provide an estimate of the uncertainty of the predicted values. The Gaussian process regression model uses the Martin 5 / 2 kernel function as its covariance function, which is suitable for modeling functions with moderate smoothness.

[0068] Step 3.4: Iterative Optimization. Calculate the expected hypervolume improvement acquisition function. This function is the core of Bayesian optimization of multi-objective problems; it quantifies the expected improvement that any new sampling point in the parameter space may bring to the currently known Pareto front. The Pareto front refers to a set of solutions that are not dominated by any other solution (i.e., a solution is no worse than another solution in all objectives, and is better than at least one objective). The expected hypervolume improvement is calculated using methods such as Monte Carlo sampling; the larger its value, the more likely that the point is to significantly advance the Pareto front. Maximize this acquisition function using an internal optimization algorithm (such as differential evolution or quasi-Newton methods) to find the next most worthwhile combination of process parameters for model evaluation.

[0069] Step 3.5: Evaluation and Update. The process parameter combination obtained in Step 3.4, along with the fixed dynamic raw material state representation vector, is input into the dual-channel quality prediction model for "evaluation" to obtain the predicted target value for that point. Subsequently, this new "process parameter-prediction target" sample is added to the observation sample set.

[0070] Step 3.6: Looping and Termination. Repeat steps 3.3 (updating the surrogate model), 3.4 (addressing new points), and 3.5 (evaluation and update) to form an iterative optimization loop. The optimization loop terminates when the number of iterations reaches a preset maximum number of iterations T, where T can be between 50 and 150, for example.

[0071] Step 3.7: Extract the Pareto optimal solution set. After the optimization loop terminates, compare the Pareto dominance relationships of all samples in the final observation sample set pairwise. Select all samples that are not dominated by any other samples; the set of these samples constitutes the Pareto optimal solution set obtained in this optimization. This solution set provides multiple optimal process parameter schemes that offer different trade-offs between the two objectives of thickness deviation and surface roughness.

[0072] Step 4: Implement the recommended optimal combination of process parameters and collect and record verification data after actual production.

[0073] This step applies the optimization results to actual production and collects production feedback data, forming a key verification step in the optimization closed loop.

[0074] From the Pareto optimal solution set obtained in step 3.7, production personnel manually select a set of process parameter combinations as the final recommended optimal combination based on the specific priority of the current production task (e.g., whether dimensional accuracy or surface quality is more important). The specific values ​​of each parameter in this combination are then distributed to the corresponding programmable logic controllers, temperature controllers, servo drives, and other execution units on the production line via the industrial network, driving the production equipment to operate according to the new parameters.

[0075] After completing a specific production unit (such as a shift or a minimum production batch) using the recommended parameters, the actual set of process parameters and online quality indicators for that production period must be re-collected strictly according to the data collection specifications described in step 1.1, and the batch information of the raw materials used must be recorded. These three data points are correlated to form a new validation data record. This record is used to verify the optimization effect and provide new samples for system learning.

[0076] Step 5: Perform online updates of the dynamic benchmark dataset and prediction model.

[0077] This step enables the entire system to continuously learn from actual production, thereby adapting to the slow time-varying nature of the process and maintaining long-term optimized performance.

[0078] Step 5.1: Verify the validity of the verification data. Check the online quality index set in the verification data record generated in Step 4. Determine whether the thickness deviation is below its upper limit threshold and whether the surface roughness is below its upper limit threshold. Only when both conditions are met is the verification data record considered a valid successful production sample and marked as usable for updating the system knowledge base.

[0079] Step 5.2: Update the dynamic benchmark dataset. Records marked as valid validation data are added to the dynamic benchmark dataset. To prevent the dataset from growing indefinitely and to ensure the model focuses more on recent process characteristics, a maximum capacity limit needs to be set for the dynamic benchmark dataset, such as 5,000 to 10,000 records. When the number of records in the dataset exceeds this maximum capacity limit, remove one or more records that were first added to the dataset based on their timestamps to maintain a constant dataset size.

[0080] Step 5.3: Trigger incremental model learning. Set a model update trigger counter and preset a trigger threshold K, where K ranges from 20 to 100. Each time a new valid successful sample is added to the dynamic benchmark dataset, the counter increments by 1. When the cumulative counter value reaches the preset trigger threshold K, the incremental learning process for the dual-channel quality prediction model is automatically triggered.

[0081] The incremental learning process is as follows: The updated dynamic benchmark dataset containing new knowledge is used as the training set. The currently deployed network weight parameters of the dual-channel quality prediction model are used as the initial training values, rather than being randomly initialized. A new round of training is conducted on the model using a learning rate lower than that used in the initial training phase in step 2.3. This setup aims to fine-tune the model, allowing it to smoothly integrate new knowledge and avoiding catastrophic forgetting of existing knowledge. After training, the original weight parameters of the online model are completely replaced with the newly acquired network weight parameters. Finally, the model update trigger counter is reset to 0 to prepare for the next learning cycle.

[0082] By cyclically executing steps 1 to 5, this invention constructs a complete closed loop from data acquisition, model building, intelligent optimization, production verification to self-updating, making the optimization of hardware product manufacturing process parameters a continuously autonomously evolving intelligent process.

[0083] Example: Optimization of stamping process parameters for automotive door hinge reinforcement plates

[0084] This embodiment is applied to the metal stamping workshop of an automotive parts manufacturer, specifically for the product "automotive door hinge reinforcement plate". This component requires high dimensional accuracy to ensure assembly quality, and also has certain requirements for the smoothness of its visible surfaces. The main equipment on the production line includes a 2000kN mechanical press, molds with heating and cooling channels, an automatic lubrication system, and an online inspection station (equipped with a laser thickness gauge and a contact surface profilometer). The raw material is cold-rolled steel sheet coils. Due to slight differences in smelting composition and rolling process, the mechanical properties of different batches of steel sheets may fluctuate within the standard allowable range.

[0085] This embodiment fully implements the five steps of the method described in this invention.

[0086] Step 1: Construct a dynamic benchmark dataset and raw material state representation vectors.

[0087] Step 1.1: Data Acquisition. Whenever a new steel coil (one production batch) is put into production, a sample is first taken from the head of the coil and sent to the laboratory for testing according to GB / T 228.1-2021 "Metallic Materials - Tensile Testing - Part 1: Test at Room Temperature" to obtain the key mechanical property parameters of this batch of raw materials: yield strength. And elongation after fracture (A). For example, the test results for the current batch are: , .

[0088] Once this batch of production reaches a stable state (after approximately 100 units have been produced continuously), a set of key process parameters is collected through the production line monitoring system:

[0089] Mold temperature: Measured by a thermocouple embedded in the mold cavity, and the average value is taken over 10 consecutive minutes, for example... .

[0090] Pressing speed: Read the slide speed setting value from the press controller, for example... .

[0091] Holding pressure: The peak pressure is read by a pressure sensor, for example... .

[0092] Lubrication flow rate: The flow rate during the spraying stage is monitored by a flow meter, for example... .

[0093] Meanwhile, the online inspection station inspects 50 consecutive workpieces during the stable production phase:

[0094] Thickness deviation: The actual thickness of a workpiece measured at a specified point by a laser thickness gauge, compared to the nominal thickness. Compare and calculate the root mean square value of the thickness deviation of these 50 workpieces. ,For example .

[0095] Surface roughness: A contact surface profilometer measures the surface roughness Ra values ​​of 50 workpiece surfaces in a specified area and calculates the arithmetic mean. ,For example .

[0096] Step 1.2: Data Association and Storage. The collected data is associated as a single record: {Batch ID, Raw Material: [355,21], Process Parameters: [150,500,800,15], Quality Index: [0.025,1.2]}, and stored in the workshop's Manufacturing Execution System database.

[0097] Step 1.3: Select qualified data to form the initial benchmark dataset. Based on the product's technical specifications, set a quality acceptance threshold: thickness deviation. Surface roughness All production records from the past 18 months were extracted from the database. Historical data that met both of these conditions were filtered out, resulting in 5127 qualified records, which formed the initial baseline dataset. .

[0098] Step 1.4: Calculate the dynamic raw material state representation vector. Assume the current requirement is a new batch of raw materials ( , The production process of the reinforcing plate was optimized.

[0099] First, calculate the initial benchmark dataset. The overall difference between each historical record and the current batch of raw materials. :

[0100]

[0101] Then, according to Sort the records in ascending order and select the first N=100 closest historical records. Then, calculate the statistics for these 100 records:

[0102] Historical process parameter mean vector Assume the calculation result is [152,510,795,16].

[0103] Historical quality index mean vector Assume the calculation result is [0.026, 1.3].

[0104] Finally, these two mean vectors are compared with the current raw material performance vector. The vectors are then concatenated to generate a 10-dimensional dynamic representation vector of the raw material state. :

[0105]

[0106] This vector encapsulates the key background information of "the level of technology and the quality achieved in historically successful production with similar raw materials".

[0107] Step 2: Construct a dual-channel quality prediction model that integrates the state of raw materials.

[0108] Step 2.1: Model Input Definition. The model requires two inputs. The first input is the set of 4D process parameters to be evaluated. The second input is the 10-dimensional dynamic raw material state representation vector generated in step 1.4. .

[0109] Step 2.2: Model Structure Construction. This example uses the PyTorch framework to build a feedforward neural network. The network structure is as follows:

[0110] Process parameter channels: Input layer (4 nodes) → First fully connected layer (32 nodes, using ReLU activation function) → Second fully connected layer (16 nodes, using ReLU activation function) → Third fully connected layer (8 nodes, using ReLU activation function).

[0111] Raw material state channel: Input layer (10 nodes) → Fourth fully connected layer (16 nodes, using ReLU activation function) → Fifth fully connected layer (8 nodes, using ReLU activation function).

[0112] Feature Fusion and Output: The 8-dimensional features output from the process parameter channel and the 8-dimensional features output from the raw material status channel are concatenated in a stitching layer to obtain a 16-dimensional fused feature vector. This vector is then input to the sixth fully connected layer (12 nodes, using the ReLU activation function), and finally connected to the seventh fully connected layer (2 nodes, without an activation function, used for direct regression prediction). The output is the predicted thickness deviation and surface roughness.

[0113] Step 2.3: Model Training. Using the initial benchmark dataset obtained in Step 1.3. (5127 samples) were used for training. For each sample in the dataset, a corresponding sample was generated based on its own raw material properties, following the method in step 1.4 (but finding the nearest neighbor centered on itself). This data serves as part of the input for this sample. The dataset is randomly divided into a training set (4101 records) and a validation set (1026 records) in an 8:2 ratio. The training configuration is as follows: the optimizer is Adam, and the initial learning rate is... The batch size is 64, and the loss function is mean squared error. The training process lasts for 200 epochs. When the loss on the validation set no longer decreases within 20 consecutive epochs, training is stopped early, and the parameters of the best-performing model on the validation set are saved.

[0114] Step 3: Real-time optimization of process parameters based on multi-objective Bayesian optimization.

[0115] Step 3.1: Define the optimization problem. The optimization variables and their safe operating ranges are set based on equipment capabilities and process knowledge:

[0116] Mold temperature T:

[0117] stamping speed V:

[0118] Holding pressure :

[0119] Lubrication flow :

[0120] The optimization objective is: , .

[0121] Step 3.2: Initialize the optimization process. Latin hypercube sampling is used to generate M=20 initial sample points in the 4-dimensional parameter space. Each set of process parameters is then compared with the parameters calculated in Step 1.4 for the current batch (…). (a fixed vector) Together, input the model trained in step 2 to obtain 20 sets. This constitutes the initial observation sample set O.

[0122] Step 3.3: Construct the surrogate model. Based on the current observation set O, for two objectives... and Construct Gaussian process surrogate models respectively and Both of these Gaussian processes use the Matérn 5 / 2 kernel function, which can model functions with moderate smoothness very well.

[0123] Steps 3.4 & 3.5: Iterative optimization and update. Use the expected hypervolume improvement as the acquisition function to select the next evaluation point.

[0124] Example of Pareto dominance: Suppose there are two points in the current observation set: point A ( Point B Point A Better than point B, but Inferior to point B, the two do not dominate each other. Point C ( If point C dominates both points A and B, then point C dominates both points A and B, because its and None of them are inferior to A and B, and at least one of them is better.

[0125] Expected Hypervolume Improvement: This function quantifies how much a new point can increase the "hypervolume" bounded by the current Pareto front and the reference point. This volume represents the size of the space of all solutions superior to the current front. Maximizing this function efficiently finds points that simultaneously improve multiple objectives.

[0126] By maximizing the acquisition function through optimization algorithms (such as differential evolution), the next combination of process parameters to be evaluated can be found. To combine with Input the prediction model, obtain the predicted value, add the new sample to the observation set O, and update the Gaussian process surrogate model.

[0127] Step 3.6: Looping and Termination. Repeat steps 3.3 to 3.5 for T=80 iterations. Including the initial 20 points, a total of 100 points were evaluated.

[0128] Step 3.7: Extract the Pareto optimal solution set. After optimization, compare the Pareto dominance relationships between each pair of the final 100 samples to filter out all non-dominated solutions, forming the Pareto optimal solution set. Assume that 15 non-dominated solutions are ultimately obtained, which constitute the solution set within the thickness deviation (range, for example...). arrive ) and surface roughness (range, e.g. arrive The "optimal frontier" is a trade-off between ( ) and ( ).

[0129] Step 4: Implement the recommended optimal combination of process parameters and collect and record verification data after actual production.

[0130] Based on the extremely high assembly precision requirements of this order, the production engineer selected the solution with the smallest predicted thickness deviation from the Pareto set. A solution to ) has the following parameters: , , , The parameter set is then sent to the press and auxiliary system control system. After completing the production of this batch (approximately 3000 pieces), actual data is collected according to the method in step 1.1 to form a verification record: {Batch ID, Raw Materials: [358, 20], Process Parameters: [164.5, 552, 848, 17.9], Quality Indicators: [0.019, 1.55]}. This record shows that the actual quality meets the requirements and is close to the predicted value.

[0131] Step 5: Perform online updates of the dynamic benchmark dataset and prediction model.

[0132] Step 5.1: Verify data validity. Check the quality indicators of the new records: thickness deviation. Surface roughness This record is determined to be a valid and successful sample.

[0133] Step 5.2: Update the dynamic benchmark dataset. Add the new record to the dynamic benchmark dataset. In this embodiment, the maximum data set size is set to 6000 records. Since there were already 5127 records before the addition and 5128 records after the addition, which does not exceed the limit, there is no need to remove the old data.

[0134] Step 5.3: Trigger incremental learning. Set the update trigger threshold K=50. Each time a new valid successful sample is added, the counter increments by 1. In this example, the counter changes from 0 to 1. Since 1 < 50, incremental learning is not triggered yet. Assuming the system continues to run, and 49 more valid successful samples are accumulated in subsequent production runs, causing the counter to reach 50, the incremental learning process is triggered: with the current record containing 5128 + 49 = 5177 records... As the training set, the weights of the original prediction model are used as initial values, and a small learning rate is adopted. The model undergoes 50 rounds of fine-tuning training. After training, the online model is updated with the new weights, and the counter is reset to 0. This allows the model to continuously adapt to slow changes in the production process.

[0135] Comparison of proportions and effects

[0136] To demonstrate the effectiveness of this invention, the following two comparative examples were designed to compare it with the method of this invention:

[0137] Comparative Example 1 (Traditional Trial and Error Method): An experienced process engineer, based on the certificate of conformity for the new batch of raw materials (… Initial parameters were set based on past experience. A production-inspection-adjustment cycle was continued until 50 consecutive qualified products were produced. All consumption during the debugging process was recorded.

[0138] Comparative Example 2 (Static Single-Objective Model Optimization Method): This method employs an existing technique that ignores raw material factors. A single neural network model is trained using historical batch data (regardless of raw material) (with only four process parameters as input). With minimizing thickness deviation as the single objective, a genetic algorithm is used to find the optimal set of parameters within the same parameter space, which is then applied to production.

[0139] Under the same production conditions, for the same batch of new materials ( Three methods were applied to obtain and implement process parameters, and the production of five consecutive batches of different raw materials (with fluctuating performance) was tracked. The comparison results are shown in the table below:

[0140] Comparison Projects Method of the present invention Comparative Example 1 (Traditional Trial and Error Method) Comparative Example 2 (Static Single-Objective Model) Parameter debugging stage / / / Number of trial production runs required to obtain the first set of usable parameters 0 times (given directly by optimization) An average of 6 times 1 time (given by optimization) Material / energy consumption during commissioning phase Extremely low (simulation calculation only) High (approximately equivalent to production losses over 6 shifts) Low (simulation calculation only) First batch production quality Thickness deviation RMSE 0.019 mm 0.045 mm (Result of the 6th trial production) 0.038 mm Surface roughness Ra 1.55μm 1.9μm (Result of the 6th trial production) 2.2μm (exceeds the threshold) Stability across multiple production batches (5 subsequent batches) / / / Quality pass rate 100%(5 / 5) 60% (3 / 5, 2 batches require re-adjustment) 40% (2 / 5, 3 batches exceeded the quality standard) Process parameter adjustment frequency No adjustment required High (2 batches require manual readjustment) High (3 batches need to be re-run and optimized) Overall benefits / / / Adaptability Strong, the model input includes the state of raw materials, and it adapts automatically. Weak, entirely dependent on engineers' experience The model is weak; it does not consider raw materials, and the optimization results are unstable. Long-term maintenance costs Low, closed-loop learning with automatic updates High, and continues to rely on expert intervention. In the process, all data needs to be collected periodically to retrain the model.

[0141] Conclusion: As shown in the table above, the method of this invention significantly outperforms traditional trial-and-error methods and static model optimization methods that ignore raw material fluctuations in terms of debugging costs, first-batch quality, stability in multi-batch production, and long-term adaptability. It is the first to achieve robust, multi-objective collaborative automatic optimization and continuous evolution of process parameters under raw material fluctuations.

[0142] This invention encompasses any substitutions, modifications, equivalent methods, and solutions made within the spirit and scope of this invention. To provide the public with a thorough understanding of this invention, specific details are described in detail in the following preferred embodiments; however, those skilled in the art will fully understand the invention even without these details. Furthermore, to avoid unnecessary misunderstanding of the essence of this invention, well-known methods, processes, procedures, components, and circuits are not described in detail.

[0143] The above description is only a preferred embodiment of the present invention. It should be noted that for those skilled in the art, several improvements and modifications can be made without departing from the principle of the present invention, and these improvements and modifications should also be considered within the scope of protection of the present invention.

Claims

1. A method for optimizing manufacturing process parameters of hardware products based on machine learning, characterized in that, Includes the following steps: Step 1: Construct a dynamic benchmark dataset and raw material state representation vector; Step 1 includes: For each production batch, synchronously collect the process parameter set, online quality index set, and raw material batch information, and store them together; wherein, the process parameter set includes die temperature, stamping speed, holding pressure, and lubrication flow; the online quality index set includes thickness deviation and surface roughness; the raw material batch information includes yield strength and elongation; select qualified historical data to form an initial benchmark dataset; calculate the dynamic raw material state representation vector for the current batch of raw materials; Step 2: Construct a dual-channel quality prediction model that integrates the raw material state and train it using the initial benchmark dataset; the input of the dual-channel quality prediction model includes the set of process parameters to be evaluated and the dynamic raw material state representation vector, and the output is the predicted thickness deviation and surface roughness. Step 3: Real-time optimization of process parameters based on multi-objective Bayesian optimization; Step 3 includes: using the trained dual-channel quality prediction model as the evaluator, minimizing the predicted thickness deviation and minimizing the predicted surface roughness as the dual objectives, iterative optimization is performed within the safe operating range of the process parameters, and finally the Pareto optimal solution set composed of non-dominated solutions is obtained. Step 4: Implement the recommended optimal combination of process parameters and collect and record verification data after actual production; Step 5: Perform online updates of the dynamic benchmark dataset and prediction model; Step 5 includes: when the verification data record is qualified, add it to the dynamic benchmark dataset; when the dataset update meets the preset conditions, trigger incremental learning of the dual-channel quality prediction model.

2. The method for optimizing hardware product manufacturing process parameters based on machine learning according to claim 1, characterized in that, Step 1, which involves constructing a dynamic benchmark dataset and raw material state representation vectors, specifically includes the following sub-steps: Step 1.1: Data Acquisition. During the stable production phase of each production batch, real-time values ​​of the mold temperature, stamping speed, holding pressure, and lubrication flow are collected through the production line control system and online sensors, serving as the set of process parameters. The thickness of multiple consecutive workpieces is collected by a high-precision thickness gauge and the statistical value of the thickness deviation is calculated. The surface profile of multiple consecutive workpieces is collected by a surface profiler and the statistical value of the surface roughness is calculated. Together, they form the online quality index set. Before batch production, the raw materials are sampled and standard mechanical property tests are performed to obtain the measured values ​​of the yield strength and the elongation, which serve as the batch information of the raw materials. Step 1.2: Data association and storage. The process parameter set, online quality index set and raw material batch information collected from the same production batch are associated to form a complete data record and stored in the production database. Step 1.3: Select qualified data to form the initial benchmark dataset. Set the upper limit threshold for thickness deviation and the upper limit threshold for surface roughness. From all historical data records in the production database, select data records where both thickness deviation and surface roughness are lower than the corresponding upper limit threshold. The set of qualified historical data is the initial benchmark dataset. Step 1.4: Calculate the dynamic raw material state representation vector. For the current batch to be optimized, the yield strength and elongation of its raw materials are known. From the initial benchmark dataset, calculate the absolute difference between the yield strength of the raw materials in all historical data records and the yield strength of the raw materials in the current batch, as well as the absolute difference between the elongation of the raw materials and the elongation of the raw materials in the current batch. Add these two absolute differences to obtain the comprehensive difference value between each historical record and the raw materials in the current batch. Select the top N historical data records in order of comprehensive difference value from smallest to largest, where N is a preset positive integer. Calculate the average value of each parameter in the process parameter set of these N historical data records to obtain the historical process parameter mean vector; Calculate the average value of each indicator in the online quality indicator set of these N historical data records to obtain the historical quality indicator mean vector; sequentially concatenate the historical process parameter mean vector, the historical quality indicator mean vector, and the current raw material performance vector composed of the yield strength and elongation of the current batch of raw materials to generate the dynamic raw material state characterization vector.

3. The method for optimizing hardware product manufacturing process parameters based on machine learning according to claim 2, characterized in that, Step 2, which involves constructing and training a dual-channel quality prediction model that integrates the states of raw materials, specifically includes the following sub-steps: Step 2.1: Model input definition. The dual-channel quality prediction model has two input ports. The first input port is used to receive a four-dimensional vector, which corresponds to the set of process parameters to be evaluated, including four parameter values: mold temperature, stamping speed, holding pressure, and lubrication flow rate. The second input port is used to receive a vector of a specific dimension generated in step 1.4, which is the dynamic raw material state characterization vector. Step 2.2: Model Structure Construction. The dual-channel quality prediction model adopts a feedforward neural network architecture. The network channel processing the first input port is called the process parameter channel, which includes a first fully connected layer, a second fully connected layer, and a third fully connected layer. The first fully connected layer has 32 neurons, the second fully connected layer has 16 neurons, and the third fully connected layer has 8 neurons. The network channel processing the second input port is called the raw material status channel, which includes a fourth fully connected layer and a fifth fully connected layer. The fourth fully connected layer has 16 neurons, and the fifth fully connected layer has 8 neurons. The output of the third fully connected layer of the process parameter channel and the output of the fifth fully connected layer of the raw material status channel are fused in a splicing layer. The fused feature vector is input to the sixth fully connected layer, which has 12 neurons. The output of the sixth fully connected layer is connected to the seventh fully connected layer, which has 2 neurons. The output of this layer is the final output of the dual-channel quality prediction model. Step 2.3: Model training. The dual-channel quality prediction model is trained under supervision using all qualified historical data records in the initial benchmark dataset. For each sample in the training dataset, a dynamic raw material state representation vector corresponding to the sample is generated according to the method in Step 1.4 based on the raw material batch information of the sample, and used as the input to the second input port of the model. The process parameter set of the sample is used as the input to the first input port of the model. The online quality index set of the sample is used as the target output of the model training. The loss function used during training is the mean squared error loss function, and the optimization algorithm is the Adam optimization algorithm. The training process continues until the prediction error of the model on the validation set no longer decreases significantly.

4. The method for optimizing hardware product manufacturing process parameters based on machine learning according to claim 3, characterized in that, Step 3, which involves real-time optimization of process parameters based on multi-objective Bayesian optimization, specifically includes the following sub-steps: Step 3.1: Define the optimization problem. The optimization variables are mold temperature, stamping speed, holding pressure, and lubrication flow rate. Each optimization variable has its corresponding lower and upper limits of safe operating range. There are two optimization objectives: the first objective is to minimize the thickness deviation predicted by the dual-channel quality prediction model, and the second objective is to minimize the surface roughness predicted by the dual-channel quality prediction model. Step 3.2: Initialize the optimization process. Using the Latin hypercube sampling method, M sets of initial process parameter combinations are uniformly generated within a four-dimensional hypercube space defined by the lower and upper limits of the safe operating range of each optimization variable, where M is a preset positive integer. Each of the M sets of initial process parameter combinations, along with the fixed dynamic raw material state representation vector generated in Step 1.4 for the current production batch, is input into the trained dual-channel quality prediction model to obtain the predicted thickness deviation and predicted surface roughness corresponding to each set of parameters. The M sets of process parameter combinations and their corresponding two predicted target values ​​together constitute the initial observation sample set. Step 3.3: Construct surrogate models. Based on the current set of observed samples, construct a Gaussian process regression model as a surrogate model for the first objective and the second objective respectively; the Gaussian process regression model uses the Marting 5 / 2 kernel function as the covariance function; Step 3.4: Iterative optimization to calculate the expected hypervolume improvement acquisition function; the expected hypervolume improvement acquisition function is based on the two Gaussian process surrogate models and is used to quantify the contribution of any new sampling point in the process parameter space to the expected improvement of the current Pareto front; by maximizing the expected hypervolume improvement acquisition function through the optimization algorithm, the next combination of process parameters to be evaluated is found within the safe operating range of the process parameters; Step 3.5: Evaluation and Update. Input the combination of process parameters found in Step 3.4 and the fixed dynamic raw material state characterization vector into the dual-channel quality prediction model to obtain the predicted thickness deviation and predicted surface roughness corresponding to the point. Add this new combination of process parameters and its predicted target value as a new sample to the observation sample set. Step 3.6: Loop and terminate, repeat steps 3.3 to 3.5 until the number of iterations reaches the preset maximum number of iterations T; Step 3.7: Extract the Pareto optimal solution set. After the optimization loop terminates, compare the Pareto dominance of all samples in the final observation sample set. For any two samples, if one sample is not inferior to the other sample in both the first and second objectives, and is strictly superior to the other sample in at least one objective, then the former is said to dominate the latter. All samples that are not dominated by any other sample constitute the Pareto optimal solution set.

5. The method for optimizing manufacturing process parameters of hardware products based on machine learning according to claim 1, characterized in that, Step 4 involves implementing the recommended optimal process parameter combination and collecting verification data records after actual production. Specifically, this includes: manually selecting a set of process parameter combinations as the recommended optimal process parameter combination from the Pareto optimal solution set, based on the different priority requirements of thickness deviation and surface roughness for the actual production task; transmitting the specific values ​​of mold temperature, stamping speed, holding pressure, and lubrication flow rate from the recommended optimal process parameter combination to the programmable logic controller (PLC) of the hardware manufacturing production line, which then drives the actuator to operate according to the set parameters; after completing at least one production shift or one minimum production batch, re-collecting the actual process parameter set, online quality index set, and batch information of the raw materials for that production period to form a new verification data record; the format of the verification data record is exactly the same as the data record format described in step 1.

2.

6. The method for optimizing hardware product manufacturing process parameters based on machine learning according to claim 5, characterized in that, Step 5 involves online updates of the dynamic benchmark dataset and the prediction model, specifically including the following sub-steps: Step 5.1: Verify the validity of the data. Check the online quality index set in the verification data record to determine whether the thickness deviation is lower than its upper limit threshold and whether the surface roughness is lower than its upper limit threshold. If both conditions are met, the verification data record is determined to be a valid and successful sample and marked as usable for updating. Step 5.2: Update the dynamic benchmark dataset by adding valid successful samples that are marked as usable for updating to the current dynamic benchmark dataset; Set a maximum capacity limit for the dynamic benchmark dataset; when the number of data records in the dynamic benchmark dataset exceeds the maximum capacity limit, remove the earliest data record that was entered into the database in the order of data entry time to maintain a constant dataset capacity; Step 5.3: Trigger incremental learning of the model. Set a model update trigger counter with an initial value of 0. Whenever a new valid successful sample is added to the dynamic benchmark dataset, the model update trigger counter is incremented by 1. When the value of the model update trigger counter reaches the preset trigger threshold K, the incremental learning process is executed. The incremental learning process is as follows: using the updated dynamic benchmark dataset as the training set, using the current network weight parameters of the dual-channel quality prediction model as the initial value, and using a learning rate lower than that in the initial training stage in step 2.3, a new round of training is performed on the dual-channel quality prediction model. After training is complete, the original network weight parameters are completely replaced with the newly obtained network weight parameters, and the model update trigger counter is reset to 0.

7. The method for optimizing hardware product manufacturing process parameters based on machine learning according to claim 2, characterized in that, The preset positive integer N in step 1.4 has a value range of 50 to 200; the dynamic raw material status representation vector has a dimension of 10, wherein the historical process parameter mean vector has a dimension of 4, the historical quality index mean vector has a dimension of 2, and the current raw material performance vector has a dimension of 2, which are sequentially concatenated to obtain a 10-dimensional vector.

8. The method for optimizing hardware product manufacturing process parameters based on machine learning according to claim 3, characterized in that, In the model structure construction in step 2.2, the first fully connected layer, the second fully connected layer, the third fully connected layer, the fourth fully connected layer, the fifth fully connected layer, and the sixth fully connected layer are all connected to a modified linear unit activation function; the seventh fully connected layer is not connected to any nonlinear activation function.

9. The method for optimizing hardware product manufacturing process parameters based on machine learning according to claim 4, characterized in that, In step 3.2, the preset positive integer M ranges from 10 to 30; in step 3.6, the preset maximum number of iterations T ranges from 50 to 150.

10. The method for optimizing manufacturing process parameters of hardware products based on machine learning according to claim 6, characterized in that, In step 5.2, the maximum capacity of the dynamic benchmark dataset is limited to 5,000 to 10,000 records; in step 5.3, the preset trigger threshold K ranges from 20 to 100.

Images