A high-cost multi-parameter coupling welding cross-domain parameter adaptive method and system based on reinforcement learning

Abstract

The application discloses a high-cost multi-parameter coupling welding cross-domain parameter adaptive method and system based on reinforcement learning. In order to solve the problems of high cost and long cycle of parameter optimization welding in the pressure cabin, the initial parameter mapping from normal pressure to high pressure is established by collecting land (normal pressure) and high-pressure dry multi-modal data, and then the high-pressure welding process is modeled as MDP, safety reinforcement learning and multi-modal quality prediction are introduced, and online adaptive parameter adjustment is carried out under the constraint of rolling prediction control (MPC), and the high-pressure recommended process parameters and their confidence interval are output. The reward function takes surface forming and molten pool morphology as the core, and considers the action cost, supports the generalization application of multi-material, multi-bevel and multi-cabin pressure level. Compared with the prior art, the application significantly reduces the number of cabin welding and consumables, and makes the high-pressure weld comprehensive performance close to the land level, with high stability and traceability.

Description

Technical Field

[0001] This invention belongs to the field of intelligent welding control and underwater engineering manufacturing technology, and relates to parameter optimization and quality control of high-pressure dry welding. Specifically, it relates to a method and system for transferring land-based welding process parameters to high-pressure dry welding scenarios and achieving adaptive parameter tuning by using reinforcement learning and other methods. Background Technology

[0002] In marine engineering, nuclear power, and deep-sea equipment manufacturing, high-pressure dry welding has attracted widespread attention due to its weld quality approaching that of land-based welding. However, the high-pressure environment leads to significant changes in gas properties, arc behavior, heat-fluid-mass coupling transmission, and metallurgical behavior, manifested as arc contraction, changes in anode-cathode height difference, and fusion ratio mismatch. To achieve the desired surface shape, traditional methods require extensive trial welding and parameter adjustments within a pressurized chamber, which is costly, time-consuming, and its reproducibility is greatly affected by the operator's experience.

[0003] Enterprises and laboratories have often accumulated a large amount of multimodal data (welding current, welding voltage, arc voltage, surface formation, penetration depth / width, etc.) of onshore (atmospheric pressure) welding. How to efficiently and reliably transfer this "atmospheric pressure knowledge" to the "high pressure dry method" scenario and form a closed-loop adaptive parameter adjustment capability is a key challenge currently facing intelligent underwater welding in industry.

[0004] Reinforcement learning is a machine learning method for sequential decision optimization that has demonstrated groundbreaking capabilities in fields such as autonomous driving and resource scheduling. The core mechanism of reinforcement learning is that an agent continuously interacts with its environment, dynamically adjusting its strategy based on rewards to ultimately maximize its returns. Unlike supervised learning, which is driven by static data, reinforcement learning is based on Markov decision processes, autonomously exploring optimal behavioral paths in unknown dynamic environments. Its significant advantage lies in its ability to handle complex, high-dimensional, continuous conditions, providing intelligent solutions for process stability control. Summary of the Invention

[0005] To reduce repeated trial welding in the pressurized chamber and ensure stable and qualified weld surface formation, this invention proposes a method and system for automatic parameter adjustment in high-pressure dry welding based on reinforcement learning: using onshore parameters as a reference, the system automatically adjusts the welding current, output voltage, welding speed, wire feed speed and shielding gas flow rate within the safety boundary, and achieves rapid convergence through short-track trial welding and online formation evaluation, outputting recommended parameters under the target chamber pressure.

[0006] The technical solution adopted in this invention is:

[0007] This invention proposes a method and system for cross-domain parameter adaptation in high-pressure dry welding based on reinforcement learning. The method includes the following steps:

[0008] Step 1: Utilize existing land and underwater high-pressure dry welding data (parameters, environment, materials, process signals, forming / defect results) to establish the correspondence and initial mapping between "process parameters - forming / defects" as the initial reference for the system.

[0009] Step 2: Set the target forming parameters (such as melt depth, melt width, excess height, porosity, undercut, spatter, etc.) in the interface, and input the land reference parameter a0.

[0010] Step 3: The system provides the first set of parameters a under different chamber pressure conditions based on the reference mapping. {p,0} .

[0011] Step 4: The experiment acquires data on the welding process under both atmospheric and high pressure. Real-time data collection includes welding current, welding voltage, welding speed, wire feed speed, shielding gas flow rate, arc voltage, and welding images, forming an atmospheric pressure dataset D0 and a high pressure dataset D10. {p,0} .

[0012] The data includes: Atmospheric pressure dataset D0: process parameters (current I, voltage U, duty cycle / polarity ratio, welding speed v, wire feed speed f, shielding gas flow rate Q, etc.), environmental parameters (temperature T). a Humidity H a ), structural / material parameters (base metal / welding wire / groove), process information (arc voltage V(t), current I(t), welding images, etc.), and output quality (surface finish index S). High-voltage dataset D {p,0} Recorded in the same dimension as D0.

[0013] Step 5: Calculate the forming deviations from the target (such as melt depth, melt width, and excess height) and defect indicators (porosity, undercut, spatter, etc.) to obtain the comprehensive deviation E. form With defect degree D defect .

[0014] Step 6: Explore the welding parameters within the safety boundary set using reinforcement learning to obtain the next set of parameters a. {k+1} a {p, k+1} And record the corresponding data D {k+1} D {p,k+1} .

[0015] Step 7: Repeat steps 4 to 6 to obtain parameter a. n ,a {p, n} and dataset D n D {p,n} .

[0016] Step 8: Output recommended parameter window a*, compliance evidence, and process traceability data.

[0017] Furthermore, in step 1, the original data is time-synchronized, denoised and anomaly removed, and process features are extracted from the process signals to construct the initial mapping.

[0018] Furthermore, in step 2, tolerances and weights are set for each indicator, and a data collection protocol is determined to form a unified evaluation standard for the future.

[0019] Furthermore, in step 3, the first set of parameters a is generated for different chamber pressures. {p,1} A safety step size recommendation and a feasible domain check should be included.

[0020] Furthermore, in step 4, the ambient temperature T a and ambient humidity H a The temperature and humidity are measured by a humidity sensor, the current / voltage is collected synchronously by a current / voltage sensor and a data acquisition card, and the image of the welding area is acquired by a high-speed camera.

[0021] Furthermore, in step 5, when the melt depth is difficult to measure directly online, a melt depth proxy quantity calibrated with features such as contour / heat input can be used instead. This invention does not limit the measurement method.

[0022] Furthermore, in step 6, the reinforcement learning strategy can be any implementation such as constrained policy gradient, model-based rolling optimization, or Bayesian optimization, with the common feature being that I, U, v, f, and Q are updated in small steps adaptively within the safe set.

[0023] Furthermore, in step 7, when the convergence criterion: E is satisfied... form ≤ε and D defect ≤δ twice consecutively, or the maximum number of iterations N is reached. max And stop when the current optimal point is reached.

[0024] Furthermore, in step 8, the process traceability data can be automatically exported as experimental number, chamber pressure, original and derived features, key images and evaluation results, which facilitates auditing and reuse.

[0025] This invention proposes a method and system for cross-domain parameter adaptation in high-pressure dry welding based on reinforcement learning. The system includes:

[0026] Data acquisition and synchronization module: humidity and temperature sensor, current / voltage sensor, data acquisition card, high-speed camera, process and environmental information recording;

[0027] Cross-domain mapping and feature engineering module;

[0028] Reinforcement learning and security constraint optimization module;

[0029] Quality prediction and MPC coordination module;

[0030] Interface for chamber pressure control and welding execution;

[0031] Human-computer interaction and traceability module (parameter suggestions, risk alarms, report export, batch traceability).

[0032] The beneficial effects of this invention are as follows: it transfers a large amount of onshore atmospheric pressure process knowledge to high-pressure dry welding, significantly reducing the number of trial welds and material consumption in the pressurized chamber; through the combination of safety reinforcement learning and MPC, it achieves rapid and repeatable parameter convergence under the condition of meeting safety boundaries such as arc stability and upper limit of heat input; the reward function with surface forming as the core enables the performance of high-pressure welds to approach the onshore level; the modular system architecture is adaptable to multiple welding power sources, multiple materials, multiple bevels, and multiple chamber pressure levels, and has good engineering versatility and scalability; it supports online learning and knowledge accumulation, reducing the strong dependence on operational experience.

[0033] The present invention will now be described in detail with reference to specific embodiments. Attached Figure Description

[0034] Figure 1 The system architecture diagram of the present invention is shown. The overall system framework (100) includes: a data acquisition and synchronization module (110), a cross-domain mapping module (120), a reinforcement learning and security constraint optimization module (130), a quality prediction and MPC coordination module (140), an execution and chamber pressure control interface (150), and a human-machine interaction and traceability module (160); the equipment side (200) includes: a welding power source (201), an electrical signal acquisition system (202), a high-pressure chamber and welding platform (203), a temperature and humidity monitoring system (204), a visual monitoring system (205), and a chamber pressure controller (206). Detailed Implementation

[0035] It should be noted that in the above embodiments, as long as the technical solutions are not contradictory, they can be permuted and combined. Those skilled in the art can exhaust all possibilities based on the mathematical knowledge of permutation and combination. Therefore, the present invention will not describe the technical solutions after permutation and combination one by one, but it should be understood that the technical solutions after permutation and combination have been disclosed by the present invention.

[0036] The above description is merely a preferred embodiment of the present invention and is not intended to limit the invention. Various modifications and variations can be made to the present invention by those skilled in the art. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the scope of protection of the present invention.

[0037] Example 1: This implementation example provides a method and system for cross-domain parameter adaptation in high-pressure dry welding based on reinforcement learning, specifically including the following steps:

[0038] Step 1: Configure the hardware equipment, which includes a welding power source, an electrical signal acquisition system, a high-pressure chamber and welding platform, a temperature and humidity monitoring system, a visual monitoring system, and a chamber pressure controller. The welding power source is used to regulate the welding current and voltage, the high-pressure chamber and welding platform are used to regulate the welding position, the chamber pressure controller is used to control the pressure inside the chamber, the electrical signal acquisition system is used to acquire current and voltage signals, the temperature and humidity monitoring system is used to monitor the ambient temperature and humidity, and the visual monitoring system is used to monitor the welding images.

[0039] Step 2: Input the land reference parameters in the interface, set the target forming (melt depth, melt width, excess height, porosity, undercut, spatter), and give the tolerance and weight.

[0040] Step 3: The system automatically generates the first set of parameters for the corresponding compartment based on the existing data and reference mapping, and provides a safe step size suggestion.

[0041] Step 4: Perform a short weld bead of 100-150 mm and collect process parameters, environmental parameters, structural / material parameters, process information, and output quality information.

[0042] Step 5: Based on the image and contour, obtain the melt depth, melt width, remaining height, porosity, undercut, and spatter, and give the qualified / unqualified and simple deviation values.

[0043] Step 6: Explore welding parameters based on reinforcement learning within the set of safety boundaries, and train the model under the condition of satisfying safety boundaries such as arc stability and upper limit of heat input.

[0044] Step 7: After the model achieves the target twice consecutively, training stops, the parameter window is output, key images, curves and parameter records are saved, and a traceable process card is formed.

Claims

1. A high-cost, multi-parameter coupled welding cross-domain parameter adaptive method based on reinforcement learning, characterized in that, include: Construct onshore and high-pressure datasets for process parameters, environmental / material parameters, process signals, and quality indicators; The initial parameter mapping from normal pressure to high pressure is obtained based on the migration / domain adaptive model, and the initial parameters for high pressure are given. The high-pressure dry welding process is modeled as a Markov decision process, with the state including parameters, process signal statistics, and online quality prediction, and the action being the parameter increment. Within the safety boundary defined by the upper limit of current, the upper limit of heat input, the arc stability threshold, and the chamber pressure fluctuation threshold, iterative optimization is performed using safety reinforcement learning. Construct a reward function that focuses on surface shaping and molten pool morphology and includes action cost penalties; Candidate actions are evaluated and screened by combining multimodal quality prediction and rolling predictive control; Through online evaluation until convergence, the recommended process parameters and their allowable windows under the target chamber pressure are output, and traceability data is generated.

2. The method according to claim 1, wherein, The process signals include at least one of arc voltage, visual parameters, and molten pool profile parameters.

3. The method according to claim 1, wherein, The initial parameter mapping employs supervised learning combined with at least one of the following domain adaptation methods: maximum mean difference (MMD), CORAL, adversarial domain adaptation, or meta-learning.

4. The method according to claim 1, wherein, The security reinforcement learning is one of constrained policy gradient, constrained deep deterministic policy gradient, hierarchical reinforcement learning, or Bayesian reinforcement learning, and satisfies constraints such as current / heat input / arc stability through barrier functions or Lagrangian constraints.

5. The method according to claim 1, wherein, The reward function is a weighted sum of surface forming deviation and melt depth / melt width deviation, and the weights are adaptively tuned according to chamber pressure, material or joint structure.

6. The method according to claim 1, wherein, A limited number of trial welds with a length of 100–150 mm were used for closed-loop optimization. The process stopped and output parameters and evidence when the overall deviation of surface forming and the degree of defect met the preset threshold for two consecutive times, or when the maximum number of iterations was reached.

7. A cross-domain parameter adaptive system for implementing the method of claim 1 in high-pressure dry welding, characterized in that, include: The data acquisition and synchronization module is used to acquire and align electrical signals, visual data, and process / environmental data. Cross-domain mapping and feature engineering module; The module includes reinforcement learning and safety constraint optimization; multimodal quality prediction and rolling predictive control coordination module; execution and cabin pressure control interface; and human-machine interaction and traceability module.

8. The system according to claim 7, wherein, The data acquisition and synchronization module includes a timestamp alignment and drift correction unit, which can achieve sub-millisecond synchronization accuracy.

9. The system according to claim 7, wherein, The multimodal quality prediction method uses a fusion of a temporal model and a graph network or Transformer structure to support online incremental learning and output uncertainty estimates for action selection.

10. A computer-readable storage medium storing a program, which, when executed on a processor, is used to implement the steps of the method according to any one of claims 1 to 6, and to cooperate with the system according to any one of claims 7 to 9 to output recommended process parameters, parameter windows, and traceability reports.

Images