Parameter analysis system and analysis method suitable for welding control system

Abstract

The invention relates to the field of welding management, in particular to parameter analysis optimization design of a welding control system in a welding global process. According to the parameter analysis system and analysis method suitable for the welding control system, a pre-welding process prediction planning model and a welding process prediction regulation and control model are organically integrated into a collaborative agent, and when the models face changing welding conditions and complex working conditions, the welding process prediction planning model and the welding process prediction regulation and control model are integrated into the collaborative agent; the method has the capability of carrying out comprehensive judgment based on context, the decision has perspectiveness and current adaptability, and high-quality welding seams are stably produced in a complex dynamic environment.

Description

Technical Field

[0001] This invention relates to the field of welding management, and more specifically to the parameter analysis and optimization design of welding control systems in the overall welding process. Background Technology

[0002] Welding process parameters refer to a series of key physical quantities that are pre-set or controlled in real time to guide welding equipment in performing welding operations in order to ensure welding quality. These parameters collectively determine the physicochemical behaviors of energy input, material melting, and crystallization during the welding process, and are direct control factors of welding quality. In this invention, the core welding process parameters mainly include core variables in the welding process such as welding current, welding voltage, welding speed, and welding torch oscillation amplitude.

[0003] With the development of intelligent and automated processes, automatic welding equipment is gradually replacing welding work that was originally done manually. In practice, the welding process parameters mentioned above are input into the control equipment, enabling the automatic welding equipment to complete the welding process based on these parameters. Therefore, for intelligent welding, the calculation and generation of welding process parameters is the core step.

[0004] In the existing technology, the generation of welding process parameters mainly follows two relatively independent technical routes: pre-welding calculation and in-welding calculation.

[0005] Pre-welding computational models predict parameters based on offline, static working condition information. Before welding, this type of model inputs the characteristic dimension vectors of the bevel into an artificial neural network (ANN) model. The model then maps and outputs a set of basic welding process parameters as the initial instructions for welding execution. In-welding computation, on the other hand, involves detecting and collecting data on the welding process using one or more sensing methods such as machine vision, electrical signals, and arc sound. The welding process parameters are then directly predicted and output using artificial neural networks (ANNs), convolutional neural networks (CNNs), or multimodal neural networks, enabling the control of the welding process.

[0006] However, both of these technical approaches have significant drawbacks. For pre-weld calculations, the welding process relies entirely on fixed parameters planned before welding, resulting in a fundamentally static and rigid system. For post-weld calculations, adjustments are made based on real-time sensor data during welding, making the system prone to short-sightedness and chaos. Sensor signals at the welding site, especially visual and arc signals, are typically quite noisy. The models used for in-weld calculations are highly susceptible to overreacting to this noise, leading to frequent parameter oscillations, instability in the welding process, and ultimately, compromised weld quality. Summary of the Invention

[0007] The purpose of this invention is to provide a parameter analysis system and method suitable for welding control systems, which organically integrates the pre-welding process prediction and planning model with the welding process prediction and control model into a collaborative intelligent agent. When faced with changing welding conditions and complex working conditions, the model has the ability to make comprehensive judgments, and its decision-making is both forward-looking and adaptable to the current situation, so as to stably produce high-quality welds in complex dynamic environments.

[0008] This invention is achieved using the following technical solution: a parameter analysis method suitable for welding control systems, characterized by comprising: The pre-welding model outputs one set of process parameters and stores pre-welding fusion information; Based on the aforementioned process parameter one, the welding equipment begins automatic welding; The welding model acquires welding information and pre-welding fusion information during the automatic welding process, outputs the predicted weld value through positive prediction calculation, and determines the weld size difference based on the predicted weld value and the target weld value. The welding model determines the amount of process parameter adjustment based on the weld size difference and the pre-weld fusion information; The process parameter adjustment amount is used in conjunction with process parameter one to form process parameter two.

[0009] As a preferred embodiment of the present invention, during the process of outputting process parameters and storing pre-welding fusion information in the pre-welding model, the input data input to the pre-welding model includes bevel feature data and welding condition input data.

[0010] As a preferred embodiment of the present invention, the pre-weld fusion information is stored in the intermediate layer of the pre-weld model.

[0011] As a preferred embodiment of the present invention, in the process of acquiring welding information and pre-welding fusion information during the automatic welding process by the welding model, the acquisition of the welding information is multimodal, and the welding information includes two-dimensional signals and one-dimensional signals.

[0012] As a preferred embodiment of the present invention, the two-dimensional signal is acquired by capturing dynamic images of the weld pool in real time using an industrial camera.

[0013] As a preferred embodiment of the present invention, the one-dimensional signal is acquired by obtaining the instantaneous values ​​of current and voltage during the welding process through a sensor; and the actual motion parameters of the welding torch are acquired through the internal data bus of the controller or an encoder.

[0014] As a preferred embodiment of the present invention, the weld size difference includes width difference, reinforcement height difference, depth difference, and area difference.

[0015] As a preferred embodiment of the present invention, during the process of outputting process parameters one in the pre-welding model and storing pre-welding fusion information, the process parameters include control parameters and motion parameters. The control parameters are one or more of the following parameters: welding current and welding voltage; the motion parameters are one or more of the following parameters: welding speed, welding torch oscillation amplitude, oscillation frequency, left / right dwell time, lateral tilt angle, forward tilt angle, lateral offset, and height offset.

[0016] A parameter analysis system suitable for welding control systems, comprising: The process parameter calculation module is configured to output process parameter one from the pre-welding model and store pre-welding fusion information. The welding module is configured to initiate automatic welding based on the process parameter one. The difference calculation module is configured to acquire welding information and pre-welding fusion information during the automatic welding process from the welding model, output the weld prediction value through positive prediction calculation, and determine the weld size difference based on the weld prediction value and the weld target value. The adjustment calculation module is configured to determine the process parameter adjustment amount based on the weld size difference and the pre-weld fusion information in the welding model; The adjustment module is configured to adjust the process parameter amount in conjunction with process parameter one to form process parameter two.

[0017] An electronic device includes a processor and a memory; the processor is connected to the memory. The memory is used to store executable program code; the processor reads the executable program code stored in the memory to run a program corresponding to the executable program code, so as to execute the method described above.

[0018] By leveraging the above three technological advantages, the number of repeated and ineffective pre-weld process tests has been significantly reduced. Furthermore, the inclusion of welding condition factors allows the process model to handle more complex scenarios, resulting in a substantial improvement in model generalization performance. The model architecture encompasses full-cycle quality management from pre-weld to during-weld, ensuring that the predicted process parameters before and during welding are nearly identical under the same conditions. The factors involved during welding are more comprehensive, allowing for appropriate adjustments to the predicted process parameters before welding, making the predicted parameters more closely match the requirements of the welding process and achieving high-precision, high-quality welding.

[0019] In summary, the present invention has the following beneficial effects: 1. Through a unique model architecture design, the pre-welding model and the in-welding model are organically integrated into a collaborative intelligent agent. This constructs a bridge for information flow between pre-welding and in-welding processes, ensuring that every real-time decision during welding is based on a full understanding of the strategic intent of the pre-welding plan, thus achieving consistency and collaborative optimization of process parameter predictions.

[0020] 2. The generalization performance and intelligence level of the model have been improved. When faced with changing welding conditions and complex working conditions, the model has the ability to make comprehensive judgments. Its decision-making is both forward-looking and adaptable, which significantly broadens the application scenarios.

[0021] 3. It achieves proactive and precise control over welding quality. By integrating global information for real-time adjustment, the system can effectively prevent various common welding defects, thereby consistently producing high-quality welds in complex dynamic environments and reducing the reliance on manual experience in production. Attached Figure Description

[0022] Figure 1 This is a schematic diagram of the architecture of Example 1; Figure 2 This is a diagram illustrating the classification of welding conditions; Figure 3 This is a schematic diagram of the welding application process. Detailed Implementation

[0023] The present invention will now be described in further detail.

[0024] This specific embodiment is merely an explanation of the present invention and is not intended to limit the invention. After reading this specification, those skilled in the art can make modifications to this embodiment without contributing any inventive step, but such modifications are protected by patent law as long as they are within the scope of the claims of the present invention.

[0025] Example 1: As mentioned above, in the prior art, there are two models: pre-weld calculation and in-weld calculation. Calculating welding process parameters separately based on these two models each has its own problems. Even if both models are introduced into a single technical solution, resulting in two sets of welding process parameters which are then combined, technical defects still exist.

[0026] Specifically, because the input sources and decision objectives of the two models differ, the process parameter instructions generated by them often show significant deviations. In this situation, if the parameters predicted during welding are used, the bevel condition information is lost; while using the parameters predicted and planned before welding results in discrepancies with the actual welding conditions, making a choice difficult. The process parameters generated before and during welding cannot form an effective synergy, often making it difficult to achieve good welding quality.

[0027] In this case, the two models were combined, and instead of generating two separate sets of process parameters for mechanical combination or selection, they were deeply integrated.

[0028] Specifically, S01, bevel feature extraction.

[0029] In this step, sensing technology is used to accurately measure and digitally convert the geometry of the position to be welded, thus obtaining the bevel characteristics.

[0030] Specifically, a laser vision contour scanning sensor is used to scan at a uniform speed along the axis of the bevel to be welded. The laser stripes projected by the sensor form a deformed light band on the bevel surface, which is captured by an industrial camera to obtain continuous contour point cloud data of the bevel cross-section. This process uniformly collects n sets of contour samples along the weld length. Using existing contour recognition algorithms, key dimensional features characterizing the geometry of the bevel are accurately extracted from these samples.

[0031] For example, for root pass welds, these characteristics typically include: Groove width: the maximum distance between the upper opening of the groove. Groove depth: the vertical distance from the root of the groove to the workpiece surface. Bottom clearance: the width of the gap between the workpieces on both sides of the groove root. Misalignment: the misalignment value of the workpieces on both sides of the groove in the height direction.

[0032] S02, Input welding conditions.

[0033] The core of this step is to transform the various qualitative and quantitative conditions that affect welding quality into standardized digital codes that can be recognized and processed by computer models, thereby achieving a comprehensive digital description of the welding environment and using it as input data for subsequent steps.

[0034] The system has a pre-defined structured welding condition library that comprehensively covers all types of conditions involved in the welding process.

[0035] For example, such as Figure 2 As shown, there are five categories and sixteen specific conditions: Welding attributes: welding method, welding machine model, power supply polarity, and shielding gas composition.

[0036] Welding material characteristics: base material grade, welding wire grade, welding wire diameter.

[0037] Workpiece specifications: joint type, blunt edge size / wall thickness, workpiece volume.

[0038] Operation type: welding position, weld type, oscillation mode.

[0039] Environmental factors: preheating temperature, ambient temperature, and relative humidity.

[0040] Operators selectively configure conditions from a condition library based on actual operating conditions. Instead of simply recording text, the system uses word embedding technology to map each selected condition into a distributed vector representation in a high-dimensional space. This encoding method captures the process correlations between different conditions at a deeper level.

[0041] The words corresponding to the above sixteen conditions are embedded and encoded, then concatenated and normalized according to their category order to form a fixed-length, dense welding condition encoding vector. This vector, as a whole, encapsulates all global static constraint information of the current welding task and is one of the key inputs for the process reasoning model to achieve scenario adaptability.

[0042] S03, Pre-welding model calculation.

[0043] In this step, the pre-welding model needs to calculate process parameter one based on the groove features and welding conditions described above as input data. The groove features and welding conditions are considered as pre-welding fusion information and are stored together in the intermediate layer of the pre-welding model.

[0044] Specifically, the bevel feature size vector generated in S01 is concatenated with the welding condition encoding vector in S02 to form the input data for the pre-weld model. This ensures that the bevel geometry information at each location is processed within the context of the global welding conditions. This input data is then fed into an artificial neural network (ANN) trained on a large amount of process data. This network learns the complex mapping relationship between input features and target process parameters through multiple layers of nonlinear transformations. Figure 1 As shown, the network output layer predicts a set of welding process planning parameters to guide welding execution. This set typically includes ten key parameters: welding speed, welding current, welding voltage, welding torch oscillation amplitude, oscillation frequency, left / right dwell time, lateral tilt angle, forward tilt angle, lateral offset, and height offset.

[0045] A key operation in this step is intermediate layer feature extraction: While performing parameter prediction via forward propagation, the system extracts the activation values ​​of the least number of neurons in the middle layer of the neural network, denoted as the intermediate layer feature. Figure 1 The middle layer 1 in the middle.

[0046] This layer's feature is a highly condensed representation of the original input—the bevel geometry and welding conditions—obtained after abstraction by the network's lower layers. It strips away the specificities directly related to the final output parameters, retaining general, deep-level semantic information about the current welding condition. This feature will be cached, serving as an information bridge connecting pre-weld planning and in-weld control, and is crucial for achieving end-to-end collaborative intelligence.

[0047] S04, Welding begins.

[0048] Welding equipment, such as welding robots, performs welding according to the process parameters calculated in step S03.

[0049] S05, Welding process data acquisition.

[0050] This step is the sensing phase of real-time feedback control, aiming to perform comprehensive, multi-modal data acquisition and feature extraction of the dynamically changing welding process through a multi-sensor system. The data acquisition in this process includes, for example... Figure 1 As shown, it includes two-dimensional and one-dimensional signals, and synchronous acquisition of multimodal sensing data.

[0051] The two-dimensional visual information acquisition process employs a high-speed industrial camera, equipped with a specific wavelength filtering system, to capture dynamic images of the weld pool in real time at a fixed frame rate. These images contain rich two-dimensional spatial information, including the shape, size, and brightness distribution of the weld pool.

[0052] One-dimensional timing signal acquisition utilizes high-precision Hall effect sensors and voltage sensors to acquire the instantaneous values ​​of welding current and arc voltage in real time. Simultaneously, the actual motion parameters of the welding torch, including welding speed, oscillation amplitude, oscillation frequency, and left and right dwell times, are acquired in real time via the robot controller's internal data bus or an external encoder.

[0053] In some embodiments, parallel preprocessing and feature extraction of multimodal data are performed: One-dimensional signal processing: After preprocessing the acquired one-dimensional time-series signals such as current and voltage by filtering, denoising, and normalization, a one-dimensional feature vector is directly formed, ready to be input into the fully connected network of the B model.

[0054] Two-dimensional image processing: After preprocessing the acquired molten pool image (such as noise reduction, contrast enhancement, and ROI extraction), it is input into a convolutional neural network (CNN) module. This CNN module is responsible for automatically extracting the deep spatial features of the molten pool image and flattening it into a two-dimensional feature vector.

[0055] The output of this step is a one-dimensional signal feature vector and a two-dimensional visual feature vector that represent the instantaneous state of the welding process after preliminary processing. Together, they form the dynamic data basis for the welding model to perform real-time inference.

[0056] S06, Positive Prediction This step is one of the core steps of the present invention, and it also embodies the fusion of the pre-welding model and the in-welding model.

[0057] This step is based on the data of groove features and welding conditions stored in the intermediate layer 1 of the pre-welding model in S03, which are fused with the welding process data collected in S05. The predicted weld formation size is then calculated using the in-welding model.

[0058] Specifically, multi-source feature fusion involves taking the one-dimensional signal feature vector and the two-dimensional visual feature vector generated in S05, performing dimensionality upscaling or downscaling through their respective fully connected layers, and mapping them to a unified feature space. Then, these two processed dynamic feature vectors are concatenated with the static intermediate layer feature vector of the pre-welding model described in S03, i.e., the feature vectors in intermediate layer 1 of the pre-welding model are concatenated to form a fused feature vector that combines global prior and local transient characteristics.

[0059] Subsequently, calculations are performed using the forward model of the welding process model. This fused feature vector is input into the core fully connected network of the forward part of the welding process model for forward propagation. Through the learned complex mapping function, the network ultimately provides predicted values ​​for the key dimensions of weld formation in the current state at the output layer. These dimensions typically include quantitative indicators such as weld width, reinforcement height, penetration depth, and cross-sectional area.

[0060] In this step, the prediction is not based solely on instantaneous sensor data, but is guided and constrained by a global understanding of the pre-welding operating conditions, i.e., the intermediate layer of the pre-welding model. This makes the prediction results more reflective of the trend of process changes, rather than simple instantaneous fluctuations, significantly improving the robustness and accuracy of the prediction, and providing a reliable basis for the next step of reverse control.

[0061] S07, Difference Calculation.

[0062] In this step, the difference between the predicted weld value obtained in S06 and the target weld value is calculated to obtain the weld size difference.

[0063] Specifically, the system presets or the operator inputs target weld values ​​based on process standards and quality requirements. This target value is a vector that typically includes key indicators such as ideal weld width, reinforcement height, penetration depth, and weld area.

[0064] Subsequently, the predicted weld value output from the positive portion of the welding model in S06 is subtracted element-wise from the aforementioned target weld value. This calculation generates a weld size difference value, where each component represents the algebraic difference between the predicted and target values ​​in a specific dimension, for example... Figure 1 As shown, the differences are: width, height, depth, and area.

[0065] In some embodiments, the deviation vector can also be standardized. Standardizing or normalizing the calculated size difference vector eliminates the numerical magnitude differences caused by different physical units, such as millimeters and square millimeters, making it suitable for subsequent reverse inference of the model. This step transforms the macroscopic "mass gap" into a precise mathematical expression that the model can handle, laying the foundation for accurate reverse inference of control quantities.

[0066] S08, Reverse calculation of the welding model.

[0067] This step is also one of the core steps of the present invention. It is the intelligent decision-making link of the entire adaptive control system. Its key lies in using the difference in weld size to back-calculate the required adjustment amount of process parameters.

[0068] Specifically, the weld formation size difference vector calculated in S07 is concatenated and fused with the intermediate layer feature vector of the pre-weld model in S03. This is the second crucial information fusion operation in the model. This fusion ensures that the reverse reasoning process not only clearly understands the gap between the current result and the target, but also deeply understands the global working condition background that produces this gap.

[0069] like Figure 1 and Figure 3 As shown, the fused feature vectors are input into a deep fully connected network in the inverse part of the welding model. This network, after pre-training, is able to learn the complex inverse mapping relationship between weld formation deviations and process parameter adjustments.

[0070] Through forward propagation of the network, a set of welding process parameter control quantities are finally generated at the output layer. This control quantity is a vector, and each component corresponds to a process parameter that needs to be adjusted, such as current, voltage, speed, etc. The required correction value is positive or negative, indicating increase or decrease.

[0071] Because it incorporates prior knowledge from the pre-welding model, the control quantities derived through reverse reasoning are no longer simple corrections to eliminate gaps, but rather suggestions under global operating condition constraints. This effectively prevents control commands from exceeding equipment capabilities or physical limits, ensuring the safety and effectiveness of control behavior.

[0072] S09, Correction and Execution.

[0073] The process parameter control vector generated in the previous step, along with corresponding parameter identifiers, timestamps, and other information, is encapsulated into a data packet conforming to a communication protocol. This packet is then transmitted to the welding robot control system with extremely low latency via a high-speed industrial bus, such as EtherCAT or real-time Ethernet. The welding robot control system receives and parses this data packet, superimposing the control values ​​within it with the currently executing baseline process parameters to generate a new round of real-time setpoints, which is process parameter two. The system then drives the actuators, such as the welding power supply and servo motors, to dynamically adjust the output, thereby precisely changing the energy input, material deposition, and welding torch movement during the welding process.

[0074] This concludes the entire plan.

[0075] In this invention, we innovatively propose and implement an intelligent decision-making scheme for welding process parameters based on deep information fusion. The core of this scheme lies in its unique model architecture design, which organically integrates the pre-welding model and the in-welding model into a collaborative intelligent agent.

[0076] Specifically, this solution simultaneously embeds the intermediate layer feature vectors containing global operating condition information generated during the pre-welding model inference process into the two key stages of forward prediction and reverse inference in the welding model. This design builds a bridge for information flow between pre-welding and welding, ensuring that every real-time decision during welding is based on a full understanding of the strategic intent of the pre-welding planning.

[0077] Based on this approach, the present invention overcomes the bottleneck of model fragmentation. It fundamentally solves the technical problems of the separation between pre-weld planning and in-weld control, and the conflict of decision commands in traditional methods, enabling consistent and coordinated optimization of process parameter predictions before and after welding.

[0078] Secondly, it also improves the model's generalization performance and intelligence level. When faced with changing welding conditions and complex working conditions, the model has the ability to make comprehensive judgments. Its decision-making is both forward-looking and adaptable, significantly broadening the application scenarios.

[0079] Finally, it also enables proactive and precise control of welding quality. By integrating global information for real-time adjustment, the system can effectively prevent a variety of common welding defects, thereby consistently producing high-quality welds in complex dynamic environments and reducing the reliance on human experience in production.

Claims

1. A parameter resolution method suitable for use in a welding control system, characterized by, The method comprises: a welding-before model outputs a first process parameter and stores welding-before fusion information; a welding device starts automatic welding based on the first process parameter; a welding-in-process model acquires welding-in-process information and the welding-before fusion information during automatic welding, calculates a welding seam prediction value through forward prediction, and determines a welding seam size difference value according to the welding seam prediction value and a welding seam target value; the welding-in-process model determines a process parameter adjustment value according to the welding seam size difference value and the welding-before fusion information; the process parameter adjustment value is used in cooperation with the first process parameter to form a second process parameter.

2. The method of claim 1, wherein: During the process of outputting the first process parameter by the welding-before model and storing the welding-before fusion information, input data input into the welding-before model includes groove feature data and welding condition input data.

3. A method for parameter resolution for a welding control system according to claim 2, wherein: The welding-before fusion information is stored in an intermediate layer of the welding-before model.

4. The method of claim 1, wherein: During the process of acquiring the welding-in-process information and the welding-before fusion information by the welding-in-process model, the acquisition of the welding-in-process information is multi-modal acquisition, and the welding-in-process information includes two-dimensional signals and one-dimensional signals.

5. A method for parameter resolution for a welding control system according to claim 4, wherein: The acquisition of the two-dimensional signals is real-time capture of dynamic images of a welding molten pool through an industrial camera.

6. A method for parameter resolution for a welding control system as defined in claim 4, wherein: The acquisition of the one-dimensional signals is acquisition of instantaneous values of current and voltage in the welding process through a sensor, and acquisition of actual motion parameters of a welding torch through an internal data bus or an encoder of a controller.

7. The method of claim 1, wherein: The welding seam size difference value includes a width difference value, a reinforcement difference value, a depth difference value, and an area difference value.

8. The method of claim 1, wherein: During the process of outputting the first process parameter by the welding-before model and storing the welding-before fusion information, the process parameter includes control parameters and motion parameters, the control parameters are one or more of the following parameters: welding current and welding voltage, and the motion parameters are one or more of the following parameters: welding speed, welding torch swing amplitude, swing frequency, left / right side dwell time, transverse inclination angle, forward inclination angle, transverse offset, and height offset.

9. A parameter resolution system suitable for use in a welding control system, characterized by, The method comprises: a process parameter one calculation module configured to make a welding-before model output a first process parameter and store welding-before fusion information; a welding module configured to make a welding device start automatic welding based on the first process parameter; a difference value calculation module configured to make a welding-in-process model acquire welding-in-process information and the welding-before fusion information during automatic welding, calculate a welding seam prediction value through forward prediction, and determine a welding seam size difference value according to the welding seam prediction value and a welding seam target value; an adjustment value calculation module configured to make the welding-in-process model determine a process parameter adjustment value according to the welding seam size difference value and the welding-before fusion information; an adjustment module configured to make the process parameter adjustment value be used in cooperation with the first process parameter to form a second process parameter. 10.An electronic device, comprising a processor and a memory; the processor is connected with the memory; the memory is used to store executable program codes; the processor runs a program corresponding to the executable program codes by reading the executable program codes stored in the memory, so as to execute the method according to any one of claims 1-8.

Images