A method and apparatus for optimizing granular stir friction additive feed frequency

Abstract

This invention discloses a method and apparatus for optimizing the feeding frequency of particulate friction stir additive manufacturing. First, feeding experiments are conducted under different operating conditions to collect pressure and displacement data. A mechanism-empirical hybrid model of the dynamic coupling relationship between pressure, displacement, and feeding frequency is established. An uncertainty set is defined. A robust optimization model for the feeding frequency is established. The dual objective is transformed into a single objective through weighted summation. The inner worst-case scenario is solved. The outer optimization employs adsorption kinetics-enhanced gradient descent. This invention uses an improved gradient descent algorithm as the core optimizer, requiring only one calculation of the inner worst-case scenario and gradient per generation. The computational load is significantly less than that of population-based algorithms, making it suitable for embedding in real-time control systems for online optimization. Combined with adaptive learning rate, dynamic diffusion intensity, and desorption probability adjustment mechanisms, the algorithm achieves rapid convergence while maintaining global search capability, providing an efficient and feasible technical solution for online dynamic tuning of the feeding frequency.

Description

Technical Field

[0001] This invention belongs to the field of additive manufacturing technology, specifically relating to a method and apparatus for optimizing the feeding frequency of granular friction stirring additive manufacturing. Background Technology

[0003] Particulate friction stir deposition (FSDD), a type of solid-state additive manufacturing, utilizes frictional heat generated between a stirring head and the material to induce a plastic flow state, thereby enabling the layer-by-layer deposition of metal wires. This avoids defects such as hot cracking, porosity, and element loss found in traditional welding processes, and is particularly suitable for the rapid near-net-shape forming of high-strength lightweight alloys such as aluminum and magnesium. This technology utilizes the high-speed rotation of the stirring head and the frictional heat generated by the wire to induce plastic flow in a thermoplastic state, forming a strong metallurgical bond with the substrate or previous layer, thus achieving continuous deposition of complex components.

[0004] In actual feeding processes, stable control of the feeding frequency is crucial for ensuring forming quality and process consistency. The feeding frequency directly affects the material supply rate, frictional heat input, and temperature field distribution in the plastic zone, thus influencing the forming accuracy, microstructure, and mechanical properties of the deposited layer. In existing algorithms, gradient descent achieves rapid convergence by iteratively calculating the gradient of the objective function and updating parameters along the negative gradient direction. However, when dealing with complex engineering optimization problems, it is prone to getting trapped in local optima, slow convergence speed, sensitivity to initial points, and lack of global search capabilities.

[0005] Therefore, how to overcome the limitations of traditional gradient descent is the technical problem faced by this invention. Summary of the Invention

[0007] To address the problems of pressure instability and feeding interruption caused by changes in material properties and external disturbances in existing technologies, the present invention aims to provide a method and apparatus for optimizing the feeding frequency of granular friction-stirring additive manufacturing. This invention proposes an adsorption kinetics-enhanced gradient descent algorithm, which effectively overcomes the limitations of traditional gradient descent by introducing biomimetic mechanisms such as diffusion and desorption, combined with a bilayer robust optimization structure.

[0008] This invention provides a method for optimizing the feeding frequency of granular friction stirring additive manufacturing, comprising the following steps:

[0009] S1: An S-type beam sensor is installed on the upper end of the push rod to read the axial pressure parameters during the feeding process;

[0010] S2: The magnetic scale is installed parallel to one side of the push rod body, and the reading head is connected to the moving part of the push rod to accurately read the real-time displacement information of the push rod;

[0011] S3: Conduct feeding experiments under different working conditions, collect pressure and displacement data, and perform preprocessing to obtain preprocessed pressure and displacement data;

[0012] S4: Establish a mechanism-empirical hybrid model for the dynamic coupling relationship between pressure, displacement, and feeding frequency. The model is as follows:

[0013]

[0014] in, This represents the total axial pressure. For mechanical mechanism terms, This is an experience-based compensation item.

[0015] The mechanical mechanism term is:

[0016]

[0017]

[0018]

[0019]

[0020] The experience compensation items are:

[0021]

[0022]

[0023]

[0024]

[0025]

[0026] S5: Define the uncertainty set: Determine the fluctuation range of key uncertainty parameters affecting the feeding process, namely wire diameter and friction coefficient, to form a box-type uncertainty set. ;

[0027] S6: Establish a robust optimization model for the feeding frequency, as follows:

[0028]

[0029]

[0030] in, For actual pressure, target pressure =100-300N, The integral time is the continuous and stable feeding time. Total running time The ratio is used as an indicator of feeding continuity;

[0031] S7: Transform the dual objectives (minimizing pressure fluctuations and maximizing feeding continuity) into a single objective through weighted summation:

[0032]

[0033] S8: Solving the worst-case scenario for the inner layer:

[0034] That is, for any given We need to find the parameter that maximizes the uncertainty of the objective function:

[0035] S9: Outer layer optimization employs adsorption kinetics to enhance gradient descent: the algorithm incorporates decision variables... Analogous to an adsorbate, it undergoes an alternating motion of three stages in the lost landscape: diffusion, relaxation, and desorption; including parameter initialization, iterative process, adsorption stage judgment, and convergence check;

[0036] S10: Apply the optimal frequency to the servo motor to control the feeding process.

[0037] In the method described, in step S4, the equivalent mass Take 0.5-2kg, damping coefficient Take 0.1-0.5 N·s / mm as the stiffness coefficient. Take 10-50 N / mm, shear strength Take 20-60MPa, wire diameter Take a diameter of 1.0-2.4mm. For push rod displacement, friction coefficient Take 0.2-0.4, positive pressure Use 50-200N; empirical coefficient Take a value of 0.1-0.3, an empirical coefficient. Take 0.5-2, temperature coefficient Take 0.01-0.05℃ -1 , For real-time temperature, Reference temperature, frequency coupling coefficient Take a value of 0.05-0.15. Phase angle, initial wear amount Take 0.1-0.5N, wear rate Take 0.01-0.05h -1 linear wear coefficient Take 0.001-0.01 N / h. For time, The shear force generated by the plastic flow of the filament. Based on friction, The additional frictional force caused by the change in the coefficient of friction. Additional force caused by temperature change The additional force caused by fluctuations in the wire feeding frequency. This refers to the additional force caused by wear and tear.

[0038] In the method described, step S7, the weighting coefficients... When pressure stability is of greater importance, a smaller value should be used; when feeding continuity is of greater importance, a larger value should be used.

[0039] In the method described above, step S10 involves optimizing the optimal feeding frequency using a differential evolution algorithm. and the corresponding push rod displacement The following closed-loop control architecture is applied to the servo motor drive system to achieve precise, stable, and adaptive control of the feeding process; firstly, the optimal parameter set... The core process setting value is input to the motion controller; among which, frequency The speed command is sent to the servo drive's port. Based on this command and the real-time speed feedback from the motor encoder, the drive's internal control loop calculates the precise torque output using a proportional-integral control algorithm. This drives the motor, via a coupling, to move the ball screw, thus converting the rotary motion into a linear feeding action of the push rod. Simultaneously, the optimal displacement... The given value of the position loop is compared with the actual displacement fed back in real time by a high-precision magnetic scale linear encoder. The deviation is used to feed forward compensation or fine-tune the speed command, forming a position-speed cascaded control loop to ensure the dynamic accuracy of the feeding length and speed.

[0040] The method described above, wherein the parameter initialization involves randomly selecting an initial point within the feasible region or based on experience. Set the base learning rate Attenuation factor diffusion intensity Desorption step size Maximum number of iterations Convergence threshold and the threshold for stopping counts .

[0041] The method described, the iterative process: for According to the current A hybrid strategy is used for computation. Record the corresponding target value and gradient ;

[0042] The method described herein, wherein the adsorption stage is determined as follows: if and Then the stagnation count ;otherwise =0; if If the system is found to be stuck in a local optimum, it enters the decoupling phase; that is, when stagnation is detected, a decoupling operation is performed: a new point is randomly sampled within a large neighborhood centered on the current point.

[0043] ,in The feasible region can be 10% to 30%. Calculate the worst-case target value for the new point. ;

[0044] like < Then accept desorption. Otherwise, keep the original position and increase it appropriately. Retry. Reset stall count. and appropriately increase the diffusion intensity. To enhance subsequent exploration.

[0045] Otherwise, select the update method based on the current stage, initially in the relaxation stage. Use an adaptive learning rate optimizer to calculate the update direction. The basic step size is Perform gradient descent update: At the same time, with probability Entering the diffusion stage: Diffusion probability It can adapt and change with iteration.

[0046] The method described above, the convergence check, if and If the iteration fails multiple times consecutively, stop; otherwise, continue. After the iteration ends, output the historical best solution, and take the historical best solution: the minimum. corresponding As a final result, it is used in subsequent closed-loop control systems.

[0047] A stir friction additive manufacturing apparatus employing any of the methods described above.

[0048] To address dynamic disturbances and material uncertainties during the process, the system integrates an adaptive fine-tuning mechanism based on multi-sensor fusion. The controller continuously acquires signals from the axial pressure sensor integrated at the end of the push rod. When the measured pressure value continuously deviates from the ideal pressure range predicted by the mechanistic-empirical hybrid model, it indicates that friction conditions or material rheology may have changed. At this point, the system can adjust the operating frequency according to preset fuzzy rules or an online-identified small-amplitude correction model. Dynamic, small-amplitude real-time compensation is performed. If the performance deviates beyond the preset safety threshold, the system can automatically trigger a rapid online optimization process. Using the current parameters as initial values, a simplified version of the differential evolution algorithm is executed to quickly find the locally optimal parameters that adapt to the new operating conditions. The parameters are then updated to the actuator through a parameter smoothing transition algorithm to avoid sudden process changes.

[0049] Furthermore, the system is designed with multi-level safety and fault-tolerance strategies to ensure control robustness. The current, speed, and position loop parameters of the servo drive are all identified and optimized by the system, and strict torque, speed, and position limit protections are provided. All control commands, status feedback, process parameters, and performance indicators are synchronously recorded and stored, and can be monitored in real time and traced historically through a human-machine interface. This constitutes a complete feeding process control system that combines stability, adaptability, and reliability, from intelligent optimization decision-making to physical precision execution.

[0050] This invention provides a method for optimizing the feeding frequency of granular friction stirring additive manufacturing. It has the following beneficial effects:

[0051] (1) This invention enhances and improves the traditional gradient descent algorithm by introducing diffusion, relaxation, and desorption mechanisms from adsorption kinetics, enabling the optimization process to have stronger global search capabilities and the ability to escape local optima. The diffusion mechanism simulates random walks, enhancing the breadth of parameter space exploration; the relaxation mechanism utilizes gradient information to achieve precise local convergence; and the desorption mechanism actively escapes local extrema when stagnation is detected, effectively avoiding the defect of traditional gradient descent being prone to getting trapped in local optima, and significantly improving the global optimization capability and robustness of feed frequency optimization.

[0052] (2) This invention constructs a mechanism-empirical hybrid model reflecting the dynamic coupling relationship between pressure, displacement, and frequency, and establishes a multi-objective robust optimization model that balances minimizing pressure fluctuations with maximizing feeding continuity. Through the synergistic optimization of worst-case solution in the inner layer and gradient descent enhanced by adsorption kinetics in the outer layer, it can effectively cope with uncertainties such as time-varying material properties and friction fluctuations. The system can identify changes in process status in real time and trigger adaptive parameter correction, significantly enhancing the additive manufacturing process's resistance to interference and process stability under complex conditions.

[0053] (3) This invention uses an improved gradient descent algorithm as the core optimizer. Each generation only needs to calculate the worst-case scenario and gradient of the inner layer once, and the computational cost is much less than that of population-based algorithms, making it suitable for embedding in real-time control systems for online optimization. By combining the adjustment mechanisms of adaptive learning rate, dynamic diffusion intensity and desorption probability, the algorithm achieves fast convergence while ensuring global search capability, providing an efficient and feasible technical solution for online dynamic tuning of feeding frequency.

[0054] (4) The present invention directly applies the optimized feeding frequency and displacement parameters obtained by optimization to the closed-loop control system of the servo motor, realizing the whole-process closed-loop control from intelligent optimization decision-making to physical precision execution. The system has a multi-level safety fault tolerance strategy, which can effectively suppress pressure fluctuations, ensure feeding continuity, and significantly improve the forming accuracy, microstructure uniformity and mechanical properties of the deposited layer, making it suitable for high-performance additive manufacturing scenarios of complex components. Attached Figure Description

[0056] Figure 1 This is a flowchart illustrating a method for optimizing the feeding frequency of granular friction stir additive manufacturing according to the present invention.

[0057] Figure 2 This is a schematic diagram of the device structure used in the method for optimizing the feeding frequency of granular stirring friction additive manufacturing according to the present invention. Detailed Implementation

[0059] The present invention will be described in detail below with reference to specific embodiments.

[0060] A method for optimizing the feeding frequency of granular friction-additive mixing, such as Figure 1 The technical flowchart shown includes the following specific steps:

[0061] Step 1: Sensor Installation and Data Acquisition

[0062] The present invention implements improvements on the apparatus for manufacturing graded functional materials based on particle-based triboelectric stirring deposition (2026100844559), as follows:

[0063] The push rod 002 (i.e., the push rod 302 in the aforementioned patent 2026100844559) is connected to the other end of the S-type beam sensor 001. The push rod 002 performs the action of cutting the wire and pushing the particles. The S-type beam sensor 001 (range 0-500N, accuracy ±0.05%FS) is used to read the axial pressure parameters in real time during the feeding process. The magnetic grating ruler 004 (resolution 0.1μm, accuracy ±1μm / m) is installed parallel to one side of the push rod body, the ruler body is fixed to the base, and the reading head 003 is fixed to the moving part of the push rod. Feeding experiments were conducted under different working conditions (feeding frequency 5-20Hz, stirring head speed 800-2000rpm, feeding speed 10-50mm / s). The axial pressure on the push rod during the feeding process was measured in real time by the S-type beam sensor 001, and the real-time displacement of the push rod was measured with high precision by the magnetic grating ruler 004. This was used for position closed-loop control and feeding length feedback. The sampling frequency was ≥2kHz, and the wavelet threshold denoising method was used for preprocessing.

[0064] Step 2: Construction of a hybrid mechanism-empirical model:

[0065] Mechanism-Empirical Hybrid Model for Establishing the Dynamic Coupling Relationship between Pressure, Displacement, and Feeding Frequency:

[0066]

[0067] in, This represents the total axial pressure. For mechanical mechanism terms, This is an experience-based compensation item.

[0068] The mechanical mechanism term is:

[0069]

[0070]

[0071]

[0072] The experience compensation items are:

[0073]

[0074]

[0075]

[0076]

[0077]

[0078] Among them, equivalent quality Take 0.5-2kg, damping coefficient Take 0.1-0.5 N·s / mm as the stiffness coefficient. Take 10-50 N / mm, shear strength Take 20-60MPa, wire diameter Take a diameter of 1.0-2.4mm. For push rod displacement, friction coefficient Take 0.2-0.4, positive pressure Use 50-200N; empirical coefficient Take a value of 0.1-0.3, an empirical coefficient. Take 0.5-2, temperature coefficient Take 0.01-0.05℃ -1 , For real-time temperature, Reference temperature, frequency coupling coefficient Take a value of 0.05-0.15. Phase angle, initial wear amount Take 0.1-0.5N, wear rate Take 0.01-0.05h -1 linear wear coefficient Take 0.001-0.01 N / h. For time, The shear force generated by the plastic flow of the filament. Based on friction, The additional frictional force caused by the change in the coefficient of friction. Additional force caused by temperature change The additional force caused by fluctuations in the wire feeding frequency. This refers to the additional force caused by wear and tear.

[0079] Step 3, Definition of Uncertainty Set:

[0080] wire diameter and coefficient of friction Considered a key uncertainty parameter, its fluctuation range is determined based on actual operating conditions: ∈[0.95 , 1.05 ], ∈[0.8 , 1.2 This constitutes a box-shaped set of uncertainties. .

[0081] Step 4: Establishment of a multi-objective robust optimization model:

[0082] Establish a dual-objective optimization model:

[0083]

[0084]

[0085] in, For actual pressure, target pressure =100-300N, The integral time is the continuous and stable feeding time. Total running time The ratio is used as an indicator of feeding continuity.

[0086] Step 5, weighted transformation with dual objectives:

[0087] The dual objectives are transformed into a single objective through weighted summation:

[0088]

[0089] Among them, the weighting coefficient The specific value can be adjusted according to the process requirements: take a smaller value when more emphasis is placed on pressure stability, and take a larger value when more emphasis is placed on feeding continuity.

[0090] Step 6: Solve the worst-case scenario for the inner layer:

[0091] For a given decision variable Solve for the uncertainty parameters in the worst case:

[0092]

[0093] This inner-level maximization problem is solved using either particle swarm optimization or pattern search, and the worst-case objective value is recorded. and gradient vector .

[0094] Step 7: Optimize gradient descent to enhance outer layer adsorption kinetics.

[0095] 7.1 Parameter Initialization: Randomly select an initial point within the feasible region. Set the base learning rate Attenuation factor diffusion intensity Desorption step size Take 20% of the feasible region (e.g., ±1.5Hz for a frequency range of 5-20Hz, ±10mm for a displacement range of 0-100mm), and set the maximum number of iterations. Convergence threshold Stagnation Count Threshold =10.

[0096] 7.2 Iterative Process: For

[0097] Calculate the current point Worst case parameters Record the corresponding target value and gradient

[0098] Adsorption stage judgment: If Then the stagnation count ;otherwise =0;

[0099] like If the system is found to be trapped in a local optimum, an unattachment operation is performed.

[0100]

[0101] Calculate the worst-case target value for the new point. .like < (For minimization problems), then decoupling is accepted. Otherwise, keep the original position and increase it appropriately. Retry, up to 3 times. Reset stall count. and appropriately increase the diffusion intensity. Up to 1.2 times the original value.

[0102] Otherwise, it enters the relaxation / diffusion phase:

[0103] Calculate the adaptive learning rate The Adam optimizer is used to calculate the update direction. .

[0104] Perform gradient descent update:

[0105]

[0106] At the same time, with probability (Initially 0.1, decaying to 0.01 with each iteration) Perform diffusion operation:

[0107]

[0108] 7.3 Convergence check: If and If it occurs 5 times consecutively, the iteration stops.

[0109] 7.4 Output Results: Output the historical best solution This serves as the process setpoint for subsequent closed-loop control.

[0110] Step 8: Optimal Frequency Application and Closed-Loop Control

[0111] The optimal feeding frequency obtained through optimization (e.g., 12.5Hz) and corresponding push rod displacement (e.g., 85mm) is input to the motion controller. The servo drive uses position-speed cascade control: the speed loop receives... The command, compared with the encoder feedback speed, is then used by PI control to output torque; the position loop will... Compared with the displacement feedback from the magnetic scale, the deviation is used for feedforward compensation. A pressure sensor integrated into the end of the push rod continuously acquires signals; when the measured pressure deviates from the predicted value of the hybrid model by more than ±10%, fuzzy rule processing is triggered. Fine-tune (adjustment range ≤ ±5%); if the deviation exceeds ±20%, automatically trigger rapid online optimization, run a simplified differential evolution algorithm (population size 20, iterations 50) with the current parameters as the initial value to find the best value again.

[0112] The system has a multi-level safety strategy: torque limit of 150% of the rated value, speed limit of 120% of the rated value, and displacement soft limit of ±5mm.

[0113] The optimization method proposed in this invention significantly improves the stability and forming quality of the feeding process in particulate friction stir additive manufacturing. Before optimization, due to fluctuations in material properties and frictional disturbances, the feeding pressure fluctuated by ±18% (±36N at the target pressure of 200N), the feeding continuity index was only about 82%, and the traditional gradient descent algorithm was prone to getting trapped in local optima, requiring an average of 320 iterations for convergence. After adopting the gradient descent algorithm enhanced by adsorption kinetics and a pressure-displacement-frequency hybrid model, the pressure fluctuation was reduced to ±5% (within ±10N), the feeding continuity was improved to 96%, the number of algorithm iterations was reduced to 110, and the global optimization capability was significantly enhanced. Combined with closed-loop control and adaptive fine-tuning, the system can still respond quickly under changing material conditions, ensuring the compactness and mechanical consistency of the deposited layer.

[0114] It should be understood that those skilled in the art can make improvements or modifications based on the above description, and all such improvements and modifications should fall within the protection scope of the appended claims.

Claims

1. A method for optimizing the feeding frequency of granular friction stirring additive manufacturing, characterized in that, Includes the following steps: S1: An S-type beam sensor is installed on the upper end of the push rod to read the axial pressure parameters during the feeding process; S2: The magnetic scale is installed parallel to one side of the push rod body, and the reading head is connected to the moving part of the push rod to accurately read the real-time displacement information of the push rod; S3: Conduct feeding experiments under different working conditions, collect pressure and displacement data, and perform preprocessing to obtain preprocessed pressure and displacement data; S4: Establish a mechanism-empirical hybrid model for the dynamic coupling relationship between pressure, displacement, and feeding frequency. The model is as follows: ； in, This represents the total axial pressure. For mechanical mechanism terms, This is an experience-based compensation item; The mechanical mechanism term is: ; ； ； ； The experience compensation items are: ; ; ; ; ; S5: Define the uncertainty set: Determine the fluctuation range of key uncertainty parameters affecting the feeding process, namely wire diameter and friction coefficient, to form a box-type uncertainty set. ; S6: Establish a robust optimization model for the feeding frequency, as follows: ; ； in, For actual pressure, target pressure =100-300N, The integral time is the continuous and stable feeding time. Total running time The ratio is used as an indicator of feeding continuity; S7: Transform the dual objectives (minimizing pressure fluctuations and maximizing feeding continuity) into a single objective through weighted summation: ; S8: Inner worst-case solution: that is, for any given... We need to find the parameter that maximizes the uncertainty of the objective function: ; S9: Outer layer optimization employs adsorption kinetics to enhance gradient descent: the algorithm incorporates decision variables... Analogous to an adsorbate, it undergoes an alternating motion of three stages in the lost landscape: diffusion, relaxation, and desorption; including parameter initialization, iterative process, adsorption stage judgment, and convergence check; S10: Apply the optimal frequency to the servo motor to control the feeding process.

2. The method according to claim 1, characterized in that, In step S4, equivalent mass Take 0.5-2kg, damping coefficient Take 0.1-0.5 N·s / mm as the stiffness coefficient. Take 10-50 N / mm, shear strength Take 20-60MPa, wire diameter Take a diameter of 1.0-2.4mm. For push rod displacement, friction coefficient Take 0.2-0.4, positive pressure Use 50-200N; empirical coefficient Take a value of 0.1-0.3, an empirical coefficient. Take 0.5-2, temperature coefficient Take 0.01-0.05℃ -1 , For real-time temperature, Reference temperature, frequency coupling coefficient Take a value of 0.05-0.

15. Phase angle, initial wear amount Take 0.1-0.5N, wear rate Take 0.01-0.05h -1 linear wear coefficient Take 0.001-0.01 N / h. For time, The shear force generated by the plastic flow of the filament. Based on friction, The additional frictional force caused by the change in the coefficient of friction. Additional force caused by temperature change The additional force caused by fluctuations in the wire feeding frequency. This refers to the additional force caused by wear and tear.

3. The method according to claim 1, characterized in that, In step S7, the weighting coefficients When pressure stability is of greater importance, a smaller value should be used; when feeding continuity is of greater importance, a larger value should be used.

4. The method according to claim 1, characterized in that, In step S10, the optimal feeding frequency is obtained through differential evolution algorithm optimization. and the corresponding push rod displacement The following closed-loop control architecture is applied to the servo motor drive system to achieve precise, stable, and adaptive control of the feeding process; firstly, the optimal parameter set... The core process setting value is input to the motion controller; among which, frequency The speed command is sent to the servo drive's port. Based on this command and the real-time speed feedback from the motor encoder, the drive's internal control loop calculates the precise torque output using a proportional-integral control algorithm. This drives the motor, via a coupling, to move the ball screw, thus converting the rotary motion into a linear feeding action of the push rod. Simultaneously, the optimal displacement... The given value of the position loop is compared with the actual displacement fed back in real time by a high-precision magnetic scale linear encoder. The deviation is used to feed forward compensation or fine-tune the speed command, forming a position-speed cascaded control loop to ensure the dynamic accuracy of the feeding length and speed.

5. The method according to claim 1, characterized in that, The parameter initialization involves randomly selecting an initial point within the feasible region or based on experience. Set the base learning rate Attenuation factor diffusion intensity Desorption step size Maximum number of iterations Convergence threshold and the threshold for stopping counts .

6. The method according to claim 1, characterized in that, The iterative process described above: For According to the current A hybrid strategy is used for computation. Record the corresponding target value and gradient .

7. The method according to claim 1, characterized in that, The adsorption stage determination: if and Then the stagnation count ; otherwise =0; if If the system is found to be stuck in a local optimum, it enters the decoupling phase; that is, when stagnation is detected, a decoupling operation is performed: a new point is randomly sampled within a large neighborhood centered on the current point. ,in The feasible region can be 10% to 30%. Calculate the worst-case target value for the new point. ;like < Then accept desorption. Otherwise, keep the original position and increase it appropriately. Retry. Reset stall count. and appropriately increase the diffusion intensity. To enhance subsequent exploration; otherwise, select the update method based on the current stage, initially in the relaxation stage. Use an adaptive learning rate optimizer to calculate the update direction. The basic step size is Perform gradient descent update: At the same time, with probability Entering the diffusion stage: Diffusion probability It can adapt and change with iteration.

8. The method according to claim 1, characterized in that, The aforementioned convergence check, if and If the iteration fails multiple times consecutively, stop; otherwise, continue. After the iteration ends, output the historical best solution, and take the historical best solution: the minimum. corresponding As a final result, it is used in subsequent closed-loop control systems.

9. A stir friction additive manufacturing apparatus employing the method described in any one of claims 1-8.

Images