Multi-objective collaborative optimization and decision-making method for laser cladding process parameters

Abstract

The invention relates to the technical field of machine learning and intelligent optimization, and discloses a multi-objective collaborative optimization and decision-making method for laser cladding process parameters, and the method comprises the steps: an initialization stage: constructing a physical correlation model and a Gaussian process agent model; in the fast loop optimization stage, the estimation performance is obtained by utilizing the physical correlation model so as to quickly update the Gaussian process agent model; in a slow loop calibration stage, concurrently obtaining foundation truth value data of the key points; and a slow loop reconstruction stage: reversely updating the physical correlation model by using the ground-based truth value, and globally reconstructing the Gaussian process proxy model. According to the method, a double-ring optimization architecture with cooperative work of fast-ring fast exploration and slow-ring concurrent calibration is constructed, so that dynamic correction of system deviation of the low-cost physical correlation model by using a high-cost foundation truth value is realized, and parameter exploration efficiency is greatly improved while optimization precision is ensured; the problem that an existing optimization method is difficult to balance among speed, cost and precision is solved.

Description

Technical Field

[0001] This invention relates to the field of machine learning and intelligent optimization technology, specifically to a multi-objective collaborative optimization and decision-making method for laser cladding process parameters. Background Technology

[0002] Laser cladding is an advanced material surface modification technology that uses a high-energy laser beam to melt alloy powder and coat it onto the surface of a substrate to prepare a cladding layer with specific properties. The final quality of the cladding layer, such as microstructure, hardness, wear resistance, and crack sensitivity, is highly dependent on the complex interaction of key process parameters such as laser power, scanning speed, and powder feeding rate. Therefore, achieving precise optimization of laser cladding process parameters to obtain the optimal parameter combination that meets multiple performance requirements is the core technical problem in this field.

[0003] Currently, optimization methods for laser cladding process parameters are mainly divided into two categories. The first category relies on experimental methods based on high-fidelity ground truth data, such as traditional orthogonal experimental design and response surface methodology. These methods involve conducting a finite number of physical experiments within the parameter space and constructing a statistical regression model to find the optimal solution. In recent years, some studies have also adopted machine learning surrogate models trained on ground truth data, such as neural networks or Gaussian process models, to fit the nonlinear relationship between process parameters and performance and guide subsequent experiments. The second category relies on low-cost estimation models, such as numerical simulation models based on finite element analysis, to simulate the thermal flow field and phase transition process during cladding, thereby predicting the process results.

[0004] However, the aforementioned optimization methods have certain shortcomings in practical applications: On the one hand, the first type of method, which relies entirely on ground truth data, is limited by high experimental costs and long testing cycles, resulting in low optimization iteration efficiency and difficulty in fully exploring high-dimensional and complex parameter spaces within a limited cost budget. On the other hand, the second type of method, which relies entirely on low-cost estimation models, often relies on simplified assumptions in numerical simulations or physical correlation models, leading to inherent systematic biases between predicted results and physical truths. If optimization is entirely dependent on such biased estimation models, the optimization process may converge to a spurious optimal solution, failing to guide actual production and thus negating the fundamental meaning of optimization. Therefore, achieving an effective balance between optimization efficiency and accuracy, and establishing a mechanism that can dynamically correct estimation model biases, is a pressing technical challenge in this field. Summary of the Invention

[0005] To address the shortcomings of existing technologies, this invention provides a multi-objective collaborative optimization and decision-making method for laser cladding process parameters. It aims to solve the problems of existing optimization methods struggling to balance optimization efficiency and accuracy, as well as the systematic bias in the estimation model that cannot be dynamically corrected.

[0006] The first aspect of this invention provides a multi-objective collaborative optimization and decision-making method for laser cladding process parameters, comprising the following steps: In the initialization phase, a physical correlation model is first defined to establish the relationship between in-service sensor signals and process performance estimates, and it has adjustable internal parameters.

[0007] Next, based on the initial dataset generated by the physical correlation model, a heteroscedastic Gaussian process surrogate model is constructed to fit the relationship between laser cladding process parameters and estimated process performance values. In this process, the estimation uncertainty vector corresponding to each data point is used as the variance of the observation noise of its respective data point, thereby enabling the surrogate model to effectively handle the heteroscedasticity characteristics of the estimation noise.

[0008] During the fast-loop asynchronous sequential optimization phase, the current Gaussian process surrogate model is used, and the expected hypervolume improvement is adopted as the acquisition function. A numerical optimization algorithm is then used to find its maximum value across the entire parameter space to determine the next optimal process parameter to be tested. This optimal process parameter is placed in the fast-loop execution queue, awaiting asynchronous acquisition and execution by the physical equipment.

[0009] During the physical cladding experiment, estimated performance data was acquired using signals from in-service sensors through a physical correlation model. This data was then used to trigger incremental background updates of the Gaussian process proxy model. This phase, through asynchronous operations and rapid model updates, enabled rapid exploration and optimization of the parameter space.

[0010] In the slow-loop concurrent calibration decision-making and truth value acquisition phase, concurrently with the fast-loop optimization, the Pareto optimal solution set for the current prediction is calculated based on the posterior predicted mean of the Gaussian process surrogate model. The optimal calibration process parameters are determined by finding the point with the greatest overall uncertainty within this Pareto optimal solution set. These optimal calibration process parameters are placed in a separate slow-loop calibration queue to prompt operator intervention. The operator performs corresponding physical cladding experiments supplemented by offline high-precision measurements to obtain the ground truth performance data for the optimal calibration process parameters. This stage, by strategically acquiring high-precision ground truth values, provides a reliable basis for model calibration.

[0011] During the self-calibration and global reconstruction phase of the slow loop model, the newly acquired ground truth performance data are used to update the internal adjustable parameters of the physical correlation model by solving optimization problems in reverse order, so as to correct its system bias.

[0012] Subsequently, using the updated physical correlation model, the in-service sensor signals of all historical sample points stored in the master dataset are recalculated in batches to globally correct all acquired estimated performance data, resulting in a globally corrected master dataset.

[0013] Finally, all old Gaussian process surrogate models were discarded, and all Gaussian process surrogate models were completely reconstructed based on the globally corrected master dataset. This stage achieved dynamic self-calibration of the physical correlation model and global reconstruction of the surrogate model, ensuring the long-term accuracy and reliability of the optimization process.

[0014] A second aspect of the present invention provides a multi-objective collaborative optimization and decision-making system for laser cladding process parameters, used to execute the method, the system comprising: The model processing module is responsible for building and maintaining the physical correlation model and the surrogate model. It constructs the initial model during system initialization and performs model self-calibration and global reconstruction upon receiving the ground truth data from the data management module. This module also provides the latest surrogate model to the fast loop optimization module and the slow loop calibration module, and provides the physical correlation model (or its access interface) to the physical execution and sensing module.

[0015] The fast-loop optimization module is responsible for receiving the proxy model provided by the model processing module, determining the optimal test point based on it, and sending the optimal test point instruction to the data management module.

[0016] The slow-loop calibration module runs concurrently with the fast-loop optimization module. It receives a proxy model provided by the model processing module, determines the optimal calibration point based on it, and sends the optimal calibration point instruction to the data management module.

[0017] The physical execution and sensing module is responsible for obtaining test point and calibration point instructions from the data management module. When executing the test point instruction, it calls the physical correlation model of the model processing module to calculate the estimated performance and sends the estimated performance data to the data management module. When executing the calibration point instruction, it obtains the true value of the foundation and sends the true value data of the foundation to the data management module.

[0018] The data management module acts as the system's data and communication hub. It sets up and manages the fast-loop execution queue, slow-loop calibration queue, master dataset, and foundation truth dataset, and provides corresponding data and queue read interfaces for other modules. When new data is written to the master dataset, this module sends a "master dataset update" trigger signal to the model processing module.

[0019] This invention achieves intelligent fusion and dynamic correction of the prediction results from high-cost foundation ground truth and low-cost physical correlation models by constructing a dual-loop optimization architecture that coordinates fast-loop rapid exploration and slow-loop concurrent calibration. This mechanism overcomes the difficulty of balancing speed, cost, and accuracy in existing methods, and helps improve the efficiency of laser cladding process parameter optimization and the reliability of final performance.

[0020] This invention provides a multi-objective collaborative optimization and decision-making method for laser cladding process parameters. It has the following beneficial effects: 1. This invention sets up a fast-loop asynchronous sequential optimization stage, uses the physical correlation model to obtain estimated performance data to quickly update the Gaussian process surrogate model, and adopts a fast-loop execution queue to realize asynchronous decision-making and physical execution. This helps to avoid waiting for expensive and time-consuming ground truth measurements in each iteration, enabling the optimization process to iterate quickly, thereby improving the efficiency of parameter space exploration and optimization speed.

[0021] 2. This invention sets up a slow-loop model self-calibration and global reconstruction stage, uses concurrently acquired ground truth data to back-update the physical correlation model, and globally corrects all historical data based on the updated model, thereby completely reconstructing the Gaussian process surrogate model. This self-calibration and reconstruction mechanism can dynamically correct the initial error of the physical correlation model, effectively prevent error accumulation from causing optimization deviation, and help improve the model accuracy and the reliability of the final optimization results.

[0022] 3. This invention achieves intelligent allocation of high-cost measurement resources and low-cost estimation data through the coordinated work of rapid exploration by the fast loop and concurrent calibration by the slow loop. The fast loop uses the estimation data for rapid exploration, while the slow loop makes decisions based on uncertainty, only acquiring the ground truth value of the most critical Pareto solution at a high cost. This dual-loop structure ensures the accuracy of optimization convergence while reducing the dependence on expensive offline measurements to a certain extent, achieving an effective balance between speed, accuracy and cost. Attached Figure Description

[0023] Figure 1 This is a schematic diagram of the method flow of the present invention; Figure 2 This is a schematic diagram of the dual-loop optimized data flow of the present invention; Figure 3 This is a schematic diagram of the physical association model of the present invention; Figure 4 This is a schematic diagram illustrating the construction of the heteroscedastic Gaussian process proxy model of the present invention; Figure 5 A schematic diagram illustrating the determination of the optimal test point for the fast-loop optimization of this invention; Figure 6 A schematic diagram illustrating the determination of the optimal calibration point for the slow loop calibration of this invention; Figure 7 This is a flowchart of the slow loop model self-calibration and global reconstruction process of the present invention; Figure 8 This is a system module architecture diagram of the present invention. Detailed Implementation

[0024] The technical solutions in 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 scope of protection of the present invention.

[0025] Please see the appendix Figure 1 -Appendix Figure 7 This invention provides a multi-objective collaborative optimization and decision-making method for laser cladding process parameters, comprising the following steps: S1: Initialization phase.

[0026] This stage establishes the basic model and initial data for subsequent optimization cycles.

[0027] S101: Define a physical correlation model and its internal adjustable parameters. This model is used to establish an estimated relationship between in-service sensor signals and final process performance.

[0028] S102: Perform a set of (e.g.) The physical cladding experiment of initial sample points (number) was conducted to collect signals from in-service sensors and calculate the initial dataset containing estimated performance and estimation uncertainty through the physical correlation model defined in S101.

[0029] S103: Based on the initial dataset obtained in S102, construct heteroscedastic Gaussian process surrogate models for each optimization objective; this step aims to provide a complete... For each of the process performance optimization objectives, an independent heteroscedastic Gaussian process GP proxy model is constructed. During construction, S103 utilizes the initial dataset generated by S102 and extracts the estimated performance values ​​corresponding to each objective (i.e., Each component of the performance vector is estimated and used as training observations for the corresponding surrogate model.

[0030] S2: Fast Loop Asynchronous Sequential Optimization.

[0031] This step is the main body of this method. It is started after S1 initialization is completed or after S4 global reconstruction is completed. It is executed cyclically to quickly explore the parameter space and approach the optimal solution.

[0032] S201: Based on the current Gaussian process proxy model provided by S103 or S404, determine the next optimal test point by calculating the multi-objective acquisition function.

[0033] S202: Place the optimal test point into a fast loop execution queue, waiting for the physical device to asynchronously capture and execute it.

[0034] S203: During the cladding process of the test point in S202 by the physical equipment, the estimated performance data (including estimated value and estimation uncertainty) of this experiment is obtained in real time through the physical correlation model.

[0035] S204: Asynchronously add the (single) estimated performance data obtained in S203 to the main dataset and trigger an (incremental) background update of the Gaussian process surrogate model.

[0036] S205: Repeatedly execute the fast-loop steps from S201 to S204 to continuously generate new test points and update the model.

[0037] S3: Slow Loop Concurrent Calibration Decision and Truth Acquisition.

[0038] This step is executed concurrently with S2, and its goal is to identify the parameter points that are most worthy of high-cost offline detection, so as to provide a basis for correction in S4.

[0039] S301: Concurrent with S201, based on the prediction uncertainties of the same set of Gaussian process surrogate models (from S103 or S404), determine an optimal calibration point from the current Pareto optimal solution set of the predictions.

[0040] S302: Place the optimal calibration point into a separate slow-loop calibration queue to prompt the operator for manual intervention.

[0041] S303: The operator executes the physical cladding experiment at the optimal calibration point according to the instructions in the S302 queue, and obtains the true performance data of the foundation through offline high-precision measurement methods (such as metallographic analysis and hardness testing). The output of this step (the true foundation value) is the input signal that triggers S4.

[0042] S4: Slow loop model self-calibration and global reconstruction.

[0043] This step is triggered after S303 produces new foundation truth values, performs self-correction of the base model (physical correlation model), and reconstructs the upper-level model (Gaussian process model) based on the corrected model.

[0044] S401: Add the true foundation performance data and its corresponding parameter points obtained in S303 to a dedicated true foundation dataset.

[0045] S402: Based on the entire foundation truth dataset updated in step S401, the internal parameters of the physical association model defined in S101 are updated in reverse by solving an optimization problem (e.g., least squares method).

[0046] S403: Using the updated physical association model from S402, batch recalculate the in-service sensor signals of all historical sample points stored in the master dataset to obtain a globally corrected master dataset.

[0047] S404: Discard all old Gaussian process surrogate models and completely reconstruct all Gaussian process surrogate models based on the globally corrected master dataset generated in S403. The output of this step (a completely new set of Gaussian process surrogate models) will be immediately used by S2 and S3 for their subsequent iterative decisions.

[0048] The following will describe in detail each of the above steps in the method of the embodiments of the present invention.

[0049] In step S1, the initialization phase is performed. This phase establishes the basic model and initial data for subsequent optimization loops.

[0050] S101: Define a physical correlation model and its internal adjustable parameters. This step establishes a mathematical model in the system to convert the physical signals that can be measured in real time during the laser cladding process into an estimate of the final process performance. This model is called the physical correlation model and has a set of internal parameters that can be adjusted through the subsequent step S402.

[0051] The physical association model can be: Simplified analytical models based on physical mechanisms (e.g., heat conduction estimation models based on the Rosenthal equation). Or a physical information neural network based on domain knowledge; Alternatively, a Bayesian neural network pre-trained based on historical simulation data.

[0052] The function of the above mathematical model is to receive a set of in-service sensor signals as input and output two results: One is the estimated value of the final process performance corresponding to this set of signals, and the other is the degree of uncertainty in this estimation. The mathematical relationship of this mathematical model can be expressed as: ; In the formula, for 3D process parameter vector; for Dimensional in-service sensor signal vectors; This is the functional expression for the physical correlation model; This is the set of internal adjustable parameters for the physical correlation model; for Dimensional estimation performance vector; for Dimensionally estimate the uncertainty vector.

[0053] estimation uncertainty The acquisition method corresponds to the model type: like If a Bayesian neural network or a Gaussian process model with equal probability is used, then The predicted variance can be obtained directly from the model (e.g., Monte Carlo dropout method); like If a standard neural network or analytical model is used, then This can be obtained through an additional error model used to fit the estimated error, or based on... Signal stability (such as variance) is indirectly obtained as a heuristic indicator.

[0054] During this step, parameters Assigned a set of initial values This initial value can be preset based on existing technical documents or empirical data.

[0055] S102: Perform a set of (e.g.) The physical cladding experiment of initial sample points (number) was conducted to collect signals from in-service sensors and calculate the initial dataset containing estimated performance and estimation uncertainty through the physical correlation model defined in S101.

[0056] This step provides the necessary training data for building the Gaussian process surrogate model for S103 through a set of pre-designed physical experiments. This step specifically includes: exist Within the 3D process parameter space, space-filling sampling methods (e.g., Latin hypercube sampling) are used to generate... initial sample points .

[0057] For each sample point A control system will vectorize process parameters. Input the controller of the laser cladding equipment and drive the equipment to perform physical cladding operations.

[0058] First, during the cladding process, the in-service sensing units deployed on the equipment synchronize with the control system clock to collect data. 3D physical signal vector .

[0059] The aforementioned in-service sensing units include (but are not limited to) one or more combinations of the following: High-speed pyrometers for measuring molten pool temperature, visual sensors (such as CCD / CMOS cameras) for acquiring molten pool morphology, acoustic emission sensors for monitoring the stability of the cladding process, or spectrometers for acquiring plasma spectral information. This refers to the raw signals or extracted feature vectors obtained from these sensors.

[0060] Then, the collected and the initial value set in S101 The data are input together into the physical correlation model for calculation, and the estimated performance vector corresponding to the sample point is obtained. and estimating the uncertainty vector .

[0061] Finally, all Group experimental results, i.e., data triples The initial dataset is formed by summarizing the data. .

[0062] S103: Based on the initial dataset obtained in S102, construct heteroscedastic Gaussian process surrogate models for each optimization objective; this step utilizes the initial dataset generated in S102 to construct surrogate models for all... For each of the process performance optimization objectives, an independent heteroscedastic Gaussian process GP surrogate model is constructed. Assuming that the total number of process performance optimization objectives is k, this step aims to construct an independent heteroscedastic Gaussian process GP surrogate model for each of these k objectives.

[0063] The aforementioned GP surrogate model is a data-driven regression model used to fit the complex nonlinear relationship between process parameters and performance estimates, and to quantify the model's own predictive uncertainty in this relationship. indivual( From 1 to The target performance is modeled as following a Gaussian process, as shown in the following formula: ; In the formula, For the first The true value of the target performance; Represent a Gaussian process; Let be the mean function of the Gaussian process; Let be the covariance function of the Gaussian process.

[0064] When building this model, the initial dataset is used. As training data, the model's input is a vector of process parameters, and the observed values ​​are the corresponding estimated performance. The heteroscedasticity of the model is reflected in: The estimation uncertainty corresponding to each data point is used as the variance of the observation noise for that data point. After training, for any new, untested point in the parameter space, the GP surrogate model can provide a prediction result in the form of a Gaussian distribution; the mean and variance of this prediction distribution are calculated using the following formulas: ; In the formula, For GP proxy model to new points First The posterior prediction mean of each target performance; For GP proxy model to new points First The posterior prediction variance of the target performance; This represents the number of data points in the current training set. The vector of new process parameters to be predicted; For one 3D matrix; For one 3D column vector; For one Dimension covariance matrix; For one 3D column vector; for Its own variance; For one A diagonal matrix, the elements of which are composed of the estimated uncertainty variance corresponding to each sample point in the training set.

[0065] The output of S103 is k GP surrogate models that have completed initial training. These GP surrogate models will serve as the initial basis for the S2 fast loop optimization and S3 slow loop calibration decisions.

[0066] This implementation method defines a physical correlation model in S101, collects initial data in S102, and constructs a heteroscedastic Gaussian process surrogate model in S103, thus realizing a complete initialization process. This helps to provide the necessary benchmark model and data foundation for subsequent fast loop optimization in S2 and slow loop calibration in S3 / S4.

[0067] In step S2, fast-loop asynchronous sequential optimization is performed. This step is the main body of the method and is started after the initialization of S1 or the global reconstruction of S4. It is executed cyclically to quickly explore the parameter space and approach the optimal solution.

[0068] S201: Based on all current Gaussian process proxy models, determine the next optimal test point by calculating the multi-objective acquisition function.

[0069] This step is the decision-making step of the fast-loop cycle, based on the current k Gaussian process surrogate models provided by S103 or S404. Within the dimensional process parameter space, the next most valuable test point is determined. This decision is achieved through a multi-objective acquisition function that balances the exploration of the parameter space with the utilization of the optimal region.

[0070] The Expected Hypervolume Improvement (EHVI) is used as the acquisition function. A numerical optimization algorithm (such as L-BFGS-B) is used to find the maximum value of this acquisition function across the entire parameter space, thus obtaining the next optimal measurement point. ; ; In the formula, This will be the next optimal test point. Let be a candidate point in the parameter space; for 3D process parameter space; Candidate points The expected improvement in supervolume; for 3D target space; This is the set of Pareto front points predicted by the current GP proxy model; for 3D target performance vector; for One goal is Joint posterior probability distribution of points; Indicates in Integrate over the target space.

[0071] S202: Place the optimal test point into a fast loop execution queue, waiting for the physical device to asynchronously capture and execute it.

[0072] This step is used to decouple the decision calculation of S201 from the execution of the physical experiment. The optimal test point calculated by S201... It is treated as an instruction and placed into a fast-loop execution queue data structure that follows the first-in-first-out (FIFO) principle.

[0073] The control system of the laser cladding physical equipment independently monitors the queue. When the equipment is idle, it automatically retrieves a test point instruction from the head of the queue and prepares to execute it, thus realizing asynchronous pipelined operation of decision calculation and physical execution.

[0074] S203: During the cladding process of the test point in S202, the estimated performance data of this experiment is obtained in real time through the physical correlation model.

[0075] This step involves performing physical experiments and initial data conversions. The physical device controller then uses the data retrieved from the queue. The instructions set process parameters such as laser power and scanning speed, and executed the cladding operation. During the operation, the in-service sensing unit synchronously collected data. 3D physical signal vector .

[0076] The signal vector is immediately input into the physical correlation model defined in S101 (using the current parameters), and the results of this experiment are calculated in real time. Dimensional estimation performance vector and dimensional estimation uncertainty vector .

[0077] S204: Asynchronously add the newly acquired estimated performance data to the main dataset and trigger a background update of the Gaussian process surrogate model.

[0078] This step integrates the experimental results obtained in S203 into the main dataset, including... The data triples are added to the master dataset.

[0079] The above-mentioned addition operation is an event that triggers an independent computation thread. This thread incrementally updates or retrains the Gaussian process proxy model corresponding to all optimization objectives established in S103 based on the updated master dataset. This update process is executed in the background and does not block S201 from continuing to calculate the next test point.

[0080] S205: Execute the fast loop steps from S201 to S204 in a loop. This step defines the continuous operation mode of the fast loop. The system continuously and cyclically executes the steps S201 (decision), S202 (enqueue), S203 (execution and estimation), and S204 (data integration and model update) until the preset termination conditions are met (e.g., reaching the maximum number of iterations, Pareto front convergence, or manual termination by the operator).

[0081] This implementation method achieves a complete fast-loop asynchronous sequential optimization process through the acquisition function decision in S201, asynchronous queue scheduling in S202, physical execution and real-time estimation in S203, and data integration and incremental model update in S204. This helps to quickly iterate and approach the Pareto optimal set of multi-objective performance without waiting for high-cost offline detection.

[0082] In step S3, slow loop concurrent calibration decision and truth value acquisition are performed. This step is performed concurrently with the fast loop optimization in S2. Its goal is to identify the parameter points that most need to obtain the foundation truth value through high-cost offline detection from the existing optimal solutions based on the uncertainty of the model, so as to provide input for the subsequent model self-calibration in S4.

[0083] S301: Based on the prediction uncertainty of the Gaussian process surrogate model, determine an optimal calibration point from the current Pareto optimal solution set of the prediction.

[0084] This step is executed concurrently with S201, using the same set of k Gaussian process surrogate models provided by S103 or S404.

[0085] First, based on the posterior prediction mean of k GP surrogate models, the Pareto optimal solution set of the current prediction is calculated.

[0086] Secondly, the optimal calibration point is determined by finding the point with the greatest overall uncertainty in the Pareto optimal solution set: ; In the formula, This is the optimal calibration point; This represents the currently predicted Pareto optimal solution set; for A candidate process parameter vector; For the first The target is at point The posterior prediction variance at the location is given by the GP surrogate model constructed by S103 or S404; The target total number.

[0087] S302: Place the optimal calibration point into a separate slow-loop calibration queue to prompt the operator for manual intervention. This step is used for asynchronous communication between the decision result of S301 and the manual execution of S303.

[0088] S301 calculates the optimal calibration point. It is placed as an instruction into a dedicated slow-loop calibration queue data structure that follows the first-in-first-out principle. The contents of the queue are presented to an operator interface, waiting for operator confirmation and execution.

[0089] S303: The operator executes the experiment at the optimal calibration point according to the queue instructions and obtains the true performance data of the foundation through offline high-precision measurement. This step is the process of obtaining high-reliability data and is the data source for the S4 calibration step. The specific steps include: First, the operator retrieves the optimal calibration point from the slow-loop calibration queue of S302. .

[0090] Secondly, the operator controls the laser cladding physical equipment according to the optimal calibration point. Perform cladding experiments and prepare the corresponding physical samples.

[0091] Next, the operator uses offline high-precision measuring instruments (e.g., metallographic microscope, microhardness tester) to inspect the sample.

[0092] Finally, the measurement results were compiled into dimensional foundation true value performance vector and its corresponding It is passed to step S401 as an input signal to trigger the self-calibration of the S4 slow loop model.

[0093] This implementation method achieves a set of concurrent ground truth acquisition processes through the optimal calibration point decision in S301, the asynchronous calibration queue in S302, and the operator offline measurement in S303. This helps to provide highly reliable data input for model self-calibration in step S4 without interrupting the fast loop optimization in S2.

[0094] In step S4, slow loop model self-calibration and global reconstruction are performed. This step is triggered after new ground truth values ​​are produced in S303. Self-correction of the basic model (physical correlation model) is performed, and the upper-level model (Gaussian process model) is reconstructed based on the corrected model.

[0095] S401: Store the true foundation performance data and its corresponding parameter points obtained in S303 into a dedicated true foundation dataset. This step is the data accumulation step for slow loop calibration.

[0096] Obtained in S303 dimensional foundation true value performance vector and its corresponding 3D process parameter vector It was used as a data pair It was added to a foundation truth dataset specifically designed for storing high-precision offline measurement data.

[0097] S402: Based on the entire foundation truth dataset, the internal parameters of the physical correlation model defined in S101 are updated in reverse by solving an optimization problem. This step uses all the foundation truth data accumulated in S401 to update the internal parameters of the physical correlation model.

[0098] The update is achieved by solving an optimization problem that aims to find a new set of parameters. This minimizes the error between the prediction results of the physical correlation model based on historical in-service sensor signals and the true values ​​of the foundation. The solution process for this optimization problem is expressed as follows: ; In the formula, For the updated physical association model parameters; These are the model parameter variables to be optimized; This is a true dataset of foundation values; This is a data pair in the foundation truth dataset; In order to adjust process parameters The collected in-service sensor signal vectors; For the physical association model in parameters Below sensor signal The predicted output; This represents the square of the L2 norm, i.e., the sum of squared errors.

[0099] S403: Using the updated physical association model from S402, batch recalculate the in-service sensor signals of all historical sample points stored in the master dataset to obtain a globally corrected master dataset.

[0100] This step utilizes the updated physical correlation model to globally correct all historical data accumulated during the fast-track optimization process. The specific operation is as follows: Iterate through each historical sample point in the main dataset and extract its corresponding in-service sensor signal vector. ,Will The new model parameters obtained in S402 are input into the physical correlation model to recalculate a new set of estimated performance vectors. and estimating the uncertainty vector Replace the original dataset in the master dataset with this new set of data. and This ultimately generates a globally corrected master dataset.

[0101] S404: Discard all old Gaussian process surrogate models and completely rebuild all Gaussian process surrogate models based on the globally corrected master dataset generated by S403.

[0102] This step involves a complete reconstruction of the upper-layer proxy model based on the globally corrected data. Specific steps include: First, the system discards the Gaussian process surrogate models corresponding to all current objectives. Then, using the globally corrected master dataset generated by S403 as the new training set, the construction process of S103 is completely repeated. For each optimization objective, a brand new heteroscedastic Gaussian process surrogate model is reconstructed. In this step, a new set of GP surrogate models will immediately replace the old models and be used by the S2 fast loop optimization and S3 slow loop calibration decisions for its subsequent iterations.

[0103] This implementation method achieves a complete slow-loop model self-calibration and global reconstruction process by storing the true value of the foundation in S401, updating the physical model parameters in S402, globally correcting historical data in S403, and completely reconstructing the surrogate model in S404. This helps to eliminate optimization deviations caused by the accumulation of initial errors in the physical model, and improves the accuracy of the entire sequential optimization process and the reliability of the final convergence result.

[0104] Please see the appendix Figure 8 This invention provides a multi-objective collaborative optimization and decision-making system for laser cladding process parameters, the system comprising: The model processing module is used to build and maintain the physical association model and the proxy model; it builds the initial model during system initialization and performs model self-calibration and global reconstruction when it receives the foundation truth data from the data management module.

[0105] The fast-loop optimization module receives the proxy model provided by the model processing module, determines the optimal test point based on it, and sends the optimal test point instruction to the data management module.

[0106] The slow-loop calibration module is used to concurrently receive the proxy model provided by the model processing module, determine the optimal calibration point based on it, and send the optimal calibration point instruction to the data management module.

[0107] The physical execution and sensing module is used to obtain the test point and calibration point instructions from the data management module; when executing the test point instruction, it calls the physical association model of the model processing module to calculate the estimated performance and sends the estimated performance data to the data management module; when executing the calibration point instruction, it obtains the foundation true value and sends the foundation true value data to the data management module.

[0108] The data management module stores the instruction queues passed from the fast loop optimization module and the slow loop calibration module, stores the estimated performance data and foundation true value data passed from the physical execution and sensing module, and provides corresponding data and queue reading interfaces for the physical execution and sensing module and the model processing module.

[0109] In the model processing module, the following operations are performed during the system initialization phase: First, define a physical association model with adjustable internal parameters; Secondly, the initial dataset is obtained by interacting with the physical execution and sensing module; Finally, an initial Gaussian process proxy model is constructed based on this initial dataset.

[0110] During system operation, the model processing module is responsible for updating and reconstructing the proxy model, specifically as follows: When the "master dataset update" trigger signal is received from the data management module, the model processing module performs an incremental update of the Gaussian process model. When the new ground truth data is received from the data management module, the model processing module triggers slow loop self-calibration, which means updating the internal parameters of the physical correlation model in reverse and requesting the data management module to provide historical sensor signals for batch recalculation, and finally globally reconstructing all Gaussian process proxy models.

[0111] The model processing module provides the updated physical correlation model (or its access interface) to the physical execution and sensing module, and provides the latest Gaussian process proxy model to the fast loop optimization module and the slow loop calibration module.

[0112] In the fast-loop optimization module, the core decision-making for executing the fast-loop asynchronous sequential optimization is carried out. The fast-loop optimization module receives the latest Gaussian process surrogate model from the model processing module in a loop. Based on the fast-loop optimization module, it calculates the multi-objective acquisition function, solves in the process parameter space, and determines the next optimal test point. After the decision is made, the optimal test point is immediately sent as an instruction to the fast-loop execution queue in the data management module without waiting for the instruction to be physically executed.

[0113] The slow-loop calibration module is responsible for executing slow-loop concurrent calibration decisions. The slow-loop calibration module runs concurrently with the fast-loop optimization module and receives the same set of the latest Gaussian process surrogate models from the model processing module.

[0114] The slow loop calibration module calculates the Pareto optimal solution set based on the posterior prediction mean of the surrogate model. Then, by analyzing the prediction uncertainty of all solutions in the solution set, it determines an optimal calibration point. After the decision is made, the module sends the optimal calibration point as an instruction to the slow loop calibration queue in the data management module.

[0115] In the physical execution and sensing module, it acts as the interface between physical execution and data acquisition. The physical execution and sensing module independently monitors the fast loop execution queue and slow loop calibration queue in the data management module.

[0116] When the optimal test point is obtained from the fast loop execution queue, the physical execution and sensing module drives the laser cladding equipment to perform physical cladding and collects signals through the in-service sensing unit. Then, it calls the physical correlation model provided by the model processing module to calculate the estimated performance and estimation uncertainty in real time, and sends the set of estimated performance data to the data management module.

[0117] When the optimal calibration point is obtained from the slow-loop calibration queue, the physical execution and sensing module drives the device to perform cladding to prepare a physical sample for the operator to perform offline high-precision measurements to obtain the true value of the foundation, and sends the set of true foundation data to the data management module.

[0118] In the data management module, it acts as the system's data and communication hub.

[0119] The data management module sets up a fast loop execution queue and a slow loop calibration queue, which are used to receive the optimal test point instruction from the fast loop optimization module and the optimal calibration point instruction from the slow loop calibration module, respectively, and provide an interface for the physical execution and sensing module to read these two queues.

[0120] Meanwhile, the data management module also sets up a master dataset and a foundation true value dataset, which are used to store the estimated performance data and foundation true value data passed in by the physical execution and sensing module, respectively.

[0121] The data management module provides the model processing module with an interface to read the ground truth dataset and the master dataset (especially historical sensor signals). When new data is written to the master dataset, the data management module sends a "master dataset update" trigger signal to the model processing module.

Claims

1. A multi-objective collaborative optimization and decision method of laser cladding process parameters, characterized in that, The method comprises the following steps: An initialization stage: Defining a physical correlation model for establishing the relationship between the in-service sensor signals and the process performance estimates; Based on the initial data set generated by the physical correlation model, a Gaussian process surrogate model is constructed for fitting the relationship between the laser cladding process parameters and the process performance estimates; A fast ring asynchronous sequential optimization stage: Determine the optimal process parameters to be tested based on the Gaussian process surrogate model; Perform a physical cladding experiment of the optimal process parameters to be tested to collect in-service sensor signals, and use the in-service sensor signals to obtain estimated performance data of the optimal process parameters to be tested through the physical correlation model, and update the Gaussian process surrogate model using the estimated performance data; A slow ring concurrent calibration decision and true value acquisition stage: Concurrently determine the optimal calibration process parameters based on the Gaussian process surrogate model, and obtain the ground truth performance data of the optimal calibration process parameters by performing a physical cladding experiment of the optimal calibration process parameters and supplementing with offline high-precision measurement; A slow ring model self-calibration and global reconstruction stage: Update the physical correlation model based on the ground truth performance data, and use the updated physical correlation model to perform batch recalculation on the collected in-service sensor signals to globally correct the obtained estimated performance data, and reconstruct the Gaussian process surrogate model based on the corrected estimated performance data.

2. The multi-objective collaborative optimization and decision method of laser cladding process parameters according to claim 1, characterized in that, In the initialization stage, the physical correlation model has internal adjustable parameters for outputting process performance estimates and estimated uncertainty vectors, and the step of constructing the Gaussian process surrogate model comprises: For each optimization objective, an anisotropic Gaussian process surrogate model is constructed, and the estimated uncertainty vector corresponding to each data point is used as the variance of the observation noise of the respective data point.

3. The multi-objective collaborative optimization and decision method of laser cladding process parameters according to claim 1, characterized in that, In the fast ring asynchronous sequential optimization stage, the step of determining the optimal process parameters to be tested comprises: Use the expected hypervolume improvement as the acquisition function, and solve the maximum value of the acquisition function in the entire parameter space by a numerical optimization algorithm to obtain the optimal process parameters to be tested.

4. The multi-objective collaborative optimization and decision method of laser cladding process parameters according to claim 1, characterized in that, The fast ring asynchronous sequential optimization stage further comprises: Put the optimal process parameters to be tested into a fast ring execution queue and wait for the physical device to asynchronously grab and execute; The step of updating the Gaussian process surrogate model using the estimated performance data is specifically triggering the incremental background update of the Gaussian process surrogate model.

5. The multi-objective collaborative optimization and decision method of laser cladding process parameters according to claim 1, characterized in that, In the slow ring concurrent calibration decision and true value acquisition stage, the step of determining the optimal calibration point comprises: Based on the posterior predictive mean of the Gaussian process surrogate model, a set of currently predicted Pareto optimal solutions is calculated; The optimal calibration process parameters are determined by finding the point with the maximum comprehensive uncertainty in the set of Pareto optimal solutions.

6. The multi-objective collaborative optimization and decision method of laser cladding process parameters according to claim 1, characterized in that, The slow ring concurrent calibration decision and true value acquisition stage further comprises: Put the optimal calibration point into a separate slow ring calibration queue to prompt the operator for manual intervention; The step of obtaining the ground truth performance data of the optimal calibration process parameters is specifically obtained by offline high-precision measurement means.

7. The multi-objective collaborative optimization and decision method of laser cladding process parameters according to claim 1, characterized in that, In the slow ring model self-calibration and global reconstruction stage, updating the physical correlation model comprises: Based on a dedicated ground truth dataset, the internal tunable parameters of the physical correlation model are updated in reverse by solving an optimization problem.

8. The multi-objective collaborative optimization and decision method of laser cladding process parameters according to claim 1, characterized in that, In the global reconstruction phase, the step of globally correcting historical data specifically includes: Using the updated physical correlation model, the in-service sensor signals of all historical sample points stored in the main dataset are batch recalculated to obtain a globally corrected main dataset.

9. The multi-objective collaborative optimization and decision method of laser cladding process parameters according to claim 8, characterized in that, The step of reconstructing the Gaussian process surrogate model specifically includes: All old Gaussian process surrogate models are discarded, and all Gaussian process surrogate models are completely reconstructed based on the globally corrected main dataset.

10. A multi-objective collaborative optimization and decision system of laser cladding process parameters, characterized in that, A system for performing the multi-objective collaborative optimization and decision-making of a laser cladding process parameter according to any one of claims 1-9, the system comprising: A model processing module for constructing and maintaining physical correlation models and surrogate models; constructing an initial model at system initialization, and performing model self-calibration and global reconstruction when receiving ground truth data transmitted by the data management module; A fast loop optimization module for receiving surrogate models provided by the model processing module, deciding optimal test points therefrom, and sending the optimal test point instructions to the data management module; A slow loop calibration module for concurrently receiving surrogate models provided by the model processing module, deciding optimal calibration points therefrom, and sending the optimal calibration point instructions to the data management module; A physical execution and sensing module for obtaining test point and calibration point instructions from the data management module; calling the physical correlation model of the model processing module to calculate estimated performance when executing the test point instructions, and sending the estimated performance data to the data management module; obtaining ground truth when executing the calibration point instructions, and sending the ground truth data to the data management module.

Images