Electric arc dynamic self-adaptive control method based on multi-mode sensing and sensing system

Abstract

The invention discloses an electric arc dynamic self-adaptive control method and sensing system based on multi-modal sensing, and relates to the technical field of electric arc control. A multi-modal sensor integrated on a welding gun is used for synchronously collecting an electric arc form image, a molten pool area image, a molten pool oscillation sound signal and current and voltage multi-dimensional data; the collected data are preprocessed, and real-time parameters in the welding process are extracted; calculating and outputting an optimal welding control parameter in real time by combining a depth Q model and taking the welding quality as an optimization target; the control parameters are transmitted to a welding actuator, real-time adjustment of the welding process is completed, and real-time control over the welding arc is achieved; the control method can be used for enhancing the real-time self-adaption of a welding control system to complex welding components and the flexible manufacturing capacity of replacing a product line; the technical problems that in an existing automatic electric arc welding control method, electric arc welding information collection is single, control response is delayed, the flexible degree is low, and welding feature extraction errors are large are solved.

Description

Technical Field

[0001] The present invention relates to the technical field of arc control, and in particular to an arc dynamic adaptive control method and a sensing system based on multi-modal sensing. Background Art

[0002] Arc welding is a core process in modern manufacturing, and its welding quality control directly affects the service performance of components.

[0003] Traditional arc welding control systems primarily rely on single sensors or simple signal feedback for parameter adjustment. They lack comprehensive monitoring of multi-dimensional information during the welding process. This leads to the following technical deficiencies in complex working conditions, such as sudden changes in base material thickness and ambient airflow disturbances:

[0004] (1) Single sensing dimension: Existing technologies mainly rely on current / voltage signal feedback and are unable to perceive the dynamic characteristics of the molten pool, resulting in inaccurate judgment of the penetration state. When welding thin-walled aerospace components, traditional methods are prone to incomplete fusion or burn-through defects due to the lack of molten pool oscillation frequency monitoring.

[0005] (2) Control response hysteresis: Existing arc welding technology also has a parameter setting control method based on empirical formulas, but it is difficult to achieve millisecond-level dynamic adjustment. When the welding gun is in the vertical welding position, it is easy to cause a sudden change in the penetration depth due to the gravity of the molten pool. The response time of existing control methods and systems generally exceeds 50ms, which can easily cause uneven weld formation.

[0006] (3) Low flexibility: Existing intelligent control methods mostly use fixed rules. When switching welding material types or shielding gas ratios, parameters need to be recalibrated and tested. The equipment debugging cycle is as long as more than 72 hours, which seriously affects the flexibility of the production line.

[0007] (4) Severe noise interference: When collecting AE signals from the molten pool, they are contaminated by arc electromagnetic noise. The existing bandpass filter method will cause some of the collected effective frequency band information to be lost, resulting in high-error welding feature extraction.

[0008] Among them, the existing disclosed patent technology proposes a robot arc welding process control and multi-sensor signal synchronous acquisition system. By collecting arc spectra, molten pool images, welding sounds, current and voltage signals, it realizes panoramic monitoring of the welding process, providing a hardware foundation for online monitoring and control of welding quality.

[0009] However, this system mainly focuses on the synchronous acquisition and monitoring of signals and lacks the ability to adaptively adjust welding parameters under complex working conditions.

[0010] In summary, it is found that the existing technology has at least the following technical problems:

[0011] The existing automated arc welding control methods have technical problems such as single arc welding information collection, delayed control response, low flexibility and large welding feature extraction errors, resulting in low adaptive adjustment capability of welding parameters. Summary of the Invention

[0012] The purpose of the present invention is to provide an arc dynamic adaptive control method and sensing system based on multimodal sensing to solve the technical problems in the existing automated arc welding control method, such as single arc welding information collection, delayed control response, low flexibility and large welding feature extraction error.

[0013] The various technical effects that can be produced by the preferred technical solutions among the various technical solutions provided by the present invention are described in detail below.

[0014] In order to solve the above technical problems, the present invention provides the following technical solutions:

[0015] The present invention provides an arc dynamic adaptive control method and sensing system based on multimodal sensing, comprising the following steps: data acquisition: acquiring multidimensional data of the welding process through a multimodal sensor integrated in the welding gun, including: arc morphology image, molten pool area image, molten pool oscillation acoustic signal, welding current and welding voltage; feature extraction: pre-processing the acquired multidimensional data to obtain arc morphology characteristics and molten pool oscillation main frequency characteristics, and calculating characteristic parameters through post-processing; decision control: pre-training to form a deep Q model, and based on the extracted characteristic parameters, calculating and outputting the optimal welding control parameters in real time, including wire feeding speed, pulse waveform, pulse frequency and shielding gas flow; execution control: sending the output welding control parameters to the welding executor, adjusting the welding process of the welding gun in real time, and realizing dynamic adaptive control of the arc.

[0016] In one embodiment, in data acquisition, a multimodal sensor is set up including a high-speed camera, an acoustic emission sensor, a Hall current sensor and a shunt; wherein, the high-speed camera and the acoustic emission sensor are detected toward the welding point, and the Hall current sensor and the shunt are set on the welding circuit; the arc shape image and the molten pool area image are collected by the high-speed camera; the molten pool oscillation acoustic emission data are collected by the acoustic emission sensor; the end welding current and welding voltage data are collected respectively by the Hall current sensor and the shunt.

[0017] In one embodiment, in the feature extraction step: the collected arc morphology image is preprocessed by a convolutional neural network, the arc morphology features are filtered to obtain the arc morphology features, the arc morphology features are analyzed, the arc geometric morphology data is obtained, and then the arc length variation coefficient is calculated, and the arc length variation coefficient data is output; in the decision control step: the arc length variation coefficient is input into the deep Q model, the arc stability is judged according to the arc length variation coefficient, and the parameters of the wire feeding speed, pulse waveform, pulse frequency and shielding gas flow are controlled and adjusted to realize real-time adjustment of the arc during the welding process and maintain stable arc output.

[0018] In one of the embodiments, in the feature extraction step: the collected molten pool area image is preprocessed by an image processing algorithm, the image of the spatter metal generated during the welding process is extracted, the number of spatter metal particles is analyzed and counted, the real-time spatter particle amount is obtained, and the spatter particle amount data is output; in the decision control step: the spatter particle amount is input into the deep Q model, the molten pool state and the arc state are judged according to the spatter particle amount, and the parameters of the pulse frequency and the shielding gas flow are controlled and adjusted.

[0019] In one of the embodiments, in the feature extraction step: the collected molten pool oscillation acoustic signal is pre-processed for noise reduction through the wavelet conversion algorithm, the main frequency characteristics of the molten pool oscillation are filtered to obtain the molten pool penetration value, and then the molten pool penetration standard deviation is calculated; in the decision control step: the penetration standard deviation is input into the depth Q model, the consistency of the weld penetration is quantified according to the penetration standard deviation, the stability of the welding process is judged, and the parameters of the wire feeding speed, pulse frequency and shielding gas flow rate are controlled and adjusted.

[0020] In one embodiment, in the feature extraction step: by collecting the welding current and welding voltage, the instantaneous arc energy input is characterized by the product of the welding current and welding voltage, and then the cumulative heat input data is calculated by the integration algorithm; in the decision control step: the cumulative heat input data is input into the deep Q model, and it is determined whether the preset parameters are reached based on the cumulative heat input data, and the parameters of the pulse waveform and pulse frequency are controlled and adjusted.

[0021] In one embodiment, the deep Q model constructs a reward function through the arc length variation coefficient, the standard deviation of the penetration depth, the amount of spatter particles and the cumulative heat input, and adjusts the wire feed speed, pulse waveform, pulse frequency and shielding gas flow rate.

[0022] In one embodiment, the following steps are also included: library construction: according to different welding materials and base materials and process requirements, a weight library and a threshold library are established; before starting welding, the actual welding materials and base materials parameters are input to perform virtual feature extraction, and the deep Q model automatically calls the corresponding welding weight values ​​and welding threshold values ​​from the weight library and threshold library, and uses the welding threshold value as the boundary parameter of the welding control and the called welding weight value as the initial welding parameter to generate a virtual welding instruction, introduces an external welding simulator and uses the welding threshold value as the boundary parameter of the welding control to perform welding simulation, verifies the welding effect of the virtual welding instruction, and determines whether the welding effect is qualified; if the welding effect is qualified, generates a welding instruction, outputs the welding control parameters after conversion, and the welding executor receives the welding control parameters to perform welding; if the welding effect is unqualified, automatically expands and selects similar base materials and welding materials in the left and right domains of the input actual welding materials and base materials parameters, and re-enters the weight library to retrieve the welding weight numerical value, generate virtual welding instructions, introduce an external welding simulator to perform welding simulation according to the welding threshold values ​​of the actual welding material and base material parameters, and verify the welding effect of the virtual welding instruction; in the execution of welding: after the data acquisition step is realized by monitoring the welding process through multi-modal sensors to obtain multi-dimensional data, the feature extraction step is entered, and the extracted parameters are input into the deep Q model to implement the decision control step. The deep Q model optimizes the welding weight value through the reward function, adjusts the welding parameters, regenerates the welding instructions, forms an instruction set, and outputs the welding control parameters after conversion. The welding actuator receives the welding control parameters and executes welding; when executing welding, the data acquisition step is run in parallel, and the multi-modal sensor monitoring is continued. The change of the welding effect is judged by the feature extraction step. If the welding effect changes, the decision control step is entered, the welding parameters are adjusted and output to the welding actuator, and the welding process of the welding gun is adjusted in real time. In this way, dynamic adaptive control of the arc is realized in a cycle.

[0023] A sensing system is also provided, comprising a welding gun, an embedded signal processing unit, a welding power supply and an edge computing unit and the above-mentioned arc dynamic adaptive control method based on multimodal sensing, wherein the welding actuator comprises the welding power supply and the welding gun; the welding gun is electrically connected to the welding power supply for receiving the output of the welding power supply and forming an output arc; the embedded signal processing unit is electrically connected to the welding gun, the welding power supply and the multimodal sensor for collecting and arranging data on the real-time wire feeding speed and real-time shielding gas flow of the welding gun, the pulse waveform, pulse frequency, output current and output voltage output by the welding power supply, and the arc shape data collected in real time by the multimodal sensor. The edge computing unit is electrically connected to the embedded signal processing unit and the welding actuator, and is used to receive the data collected and sorted by the embedded signal processing unit, extract the characteristic parameters through the built-in module, calculate the arc length variation coefficient, the standard deviation of the penetration depth, the amount of spatter particles and the cumulative heat input, and input them into the deep Q model module, and then calculate and output the welding control parameters to the welding actuator through the deep Q model module.

[0024] In one of the embodiments, the edge computing unit is provided with a simulator interface for introducing an external welding simulator.

[0025] The beneficial effects of the present invention are as follows:

[0026] The present invention provides an arc dynamic adaptive control method based on multimodal sensing, which has the following beneficial effects compared with the prior art:

[0027] 1. Improve the comprehensiveness and accuracy of information acquisition: By installing a multi-modal sensor on the welding gun that integrates arc shape images, molten pool area images, molten pool oscillation sound signals, welding current and welding voltage monitoring, and simultaneously obtains multiple key physical quantities during the arc welding process, the accuracy and comprehensiveness of welding status perception are greatly improved, overcoming the defect that traditional single sensors are difficult to reflect complex welding conditions, and effectively improving the diversity and reliability of data acquisition.

[0028] 2. Higher feature extraction accuracy: Multimodal data fusion technology is used to jointly process and extract features from images, acoustic signals, and electrical signals, which improves the robustness of arc morphology and molten pool oscillation feature recognition and reduces the probability of misjudgment due to interference from single-source data.

[0029] 3. Enhance the adaptability and real-time performance of welding control: Deploy and pre-train deep learning models, and use deep Q models to optimize and control welding quality. Based on the feature parameters extracted in real time, after analysis, the system can automatically adjust the wire feed speed, pulse waveform, pulse frequency, shielding gas flow rate and other control variables that can affect welding quality according to the direction of adjustment needed. In response to real-time changing welding conditions, the welding control parameters can be flexibly adjusted, the arc output can be controlled in real time, and a dynamic response can be achieved to complex welding conditions. It can adapt to complex working conditions, thereby significantly enhancing the flexibility and adaptability of the welding control system using this method.

[0030] Fast response speed and precise control: The control model built based on the deep Q learning algorithm, the deep Q model, and the use of multimodal sensors to monitor welding in real time can realize real-time analysis of welding characteristic parameters and dynamic adjustment of welding strategies. Compared with the traditional static parameter setting method, the use of the deep learning model enables the welding control system using this method to have self-learning and self-adaptation capabilities, effectively shortening arc control delay and improving welding quality consistency.

[0031] 4. Improve welding quality and production efficiency: Through the combination of multimodal fusion perception and intelligent control, the welding control system using this method can maintain arc stability, uniform penetration, and reduced spatter particles under conditions of complex welding components, such as curved welds and changes in welding material thickness. It can adjust the output welding current and welding voltage in time according to changes in welding conditions, thereby improving the welding qualification rate of complex components and reducing energy consumption. It is especially suitable for high-standard thin-walled structure welding, such as thin-walled and complex-shaped components in the aerospace field. BRIEF DESCRIPTION OF THE DRAWINGS

[0032] In order to more clearly illustrate the technical solution of the present invention, the following is a brief introduction to the drawings required for use in the implementation. Obviously, the drawings described below are only some implementation methods of the present invention. For ordinary technicians in this field, other drawings can be obtained based on these drawings without paying any creative work.

[0033] Figure 1 This is one of the flow charts of the control method of the present invention;

[0034] Figure 2 This is the second flow chart of the control method of the present invention. DETAILED DESCRIPTION

[0035] The technical solutions in the embodiments of the present invention will be described clearly and completely below with reference to the accompanying drawings in the embodiments of the present invention.

[0036] In a specific embodiment, a dynamic adaptive control method for arc based on multimodal sensing is provided. The control method utilizes a multimodal sensor integrated in the welding gun to synchronously collect arc morphology images, molten pool area images, molten pool oscillation sound signals and current and voltage multi-dimensional data; pre-processes the collected data to extract real-time parameters of the welding process; combines the deep Q model, takes welding quality as the optimization goal, and calculates and outputs the optimal welding control parameters in real time; transmits the control parameters to the welding actuator to complete the real-time adjustment of the welding process and realize real-time control of the welding arc; the application of this control method can be used to enhance the welding control system, the real-time adaptation of complex welding components and the flexible manufacturing capability of replacing product lines; effectively solves the technical problems of existing automated arc welding control methods, such as single arc welding information collection, delayed control response, low flexibility and large welding feature extraction errors.

[0037] The first implementation of the arc dynamic adaptive control method is as follows Figure 1 The method includes the following steps: data acquisition: obtaining multi-dimensional data of the welding process through a multi-modal sensor integrated in the welding gun, including: arc shape image, molten pool area image, molten pool oscillation acoustic signal, welding current and welding voltage; feature extraction: pre-processing the collected multi-dimensional data to obtain arc shape characteristics and molten pool oscillation main frequency characteristics, and calculating the characteristic parameters through post-processing; decision control: pre-training to form a deep Q model, and based on the extracted characteristic parameters, calculating and outputting the optimal welding control parameters in real time, including wire feeding speed, pulse waveform, pulse frequency and shielding gas flow; execution control: sending the output welding control parameters to the welding actuator, adjusting the welding process of the welding gun in real time, and realizing dynamic adaptive control of the arc.

[0038] Among them, in data acquisition, the multimodal sensor setting includes a high-speed camera, an acoustic emission sensor, a Hall current sensor and a shunt; among them, the high-speed camera and the acoustic emission sensor are detected toward the welding point, and the Hall current sensor and the shunt are set on the welding circuit; the arc shape image and the molten pool area image are collected by the high-speed camera; the molten pool oscillation acoustic data are collected by the acoustic emission sensor; the end welding current and welding voltage data are collected respectively by the Hall current sensor and the shunt.

[0039] Specifically, the shooting speed of the high-speed camera reaches more than 2000fps.

[0040] During application, the Hall current sensor and shunt for collecting real-time welding current and welding voltage can be directly integrated into the welding power supply output end or welding gun cable, and the welding circuit to directly measure the real-time welding current and welding voltage, ensuring that the welding current and welding voltage that intuitively reflect the output welding energy can be accurately measured and obtained.

[0041] By applying the arc dynamic adaptive control method based on multimodal sensing provided by the technical solution of the present invention, the welding control system modified has the following beneficial effects compared with the existing technology:

[0042] Improve the comprehensiveness and accuracy of information acquisition: By installing a multi-modal sensor on the welding gun that integrates arc shape images, molten pool area images, molten pool oscillation sound signals, welding current and welding voltage monitoring, and simultaneously acquiring multiple key physical quantities in the arc welding process, the accuracy and comprehensiveness of welding status perception are greatly improved, overcoming the defect that traditional single sensors are difficult to reflect complex welding conditions, and effectively improving the diversity and reliability of data acquisition.

[0043] Higher feature extraction accuracy: Multimodal data fusion technology is used to jointly process and extract features from images, acoustic signals, and electrical signals, which improves the robustness of arc morphology and molten pool oscillation feature recognition and reduces the probability of misjudgment due to interference from single-source data.

[0044] Enhance the adaptability and real-time performance of welding control: deploy and pre-train deep learning models, and use deep Q models to optimize and control welding quality. Based on the feature parameters extracted in real time, after analysis, the system can automatically adjust the wire feed speed, pulse waveform, pulse frequency, shielding gas flow rate and other control variables that can affect welding quality according to the direction of adjustment needed. In response to real-time changing welding conditions, the welding control parameters can be flexibly adjusted, the arc output can be controlled in real time, and a dynamic response can be achieved to complex welding conditions, thereby adapting to complex working conditions, thereby significantly enhancing the flexibility and adaptability of the welding control system using this method.

[0045] Fast response speed and precise control: The control model built based on the deep Q learning algorithm, the deep Q model, and the use of multimodal sensors to monitor welding in real time can realize real-time analysis of welding characteristic parameters and dynamic adjustment of welding strategies. Compared with the traditional static parameter setting method, the use of the deep learning model enables the welding control system using this method to have self-learning and self-adaptation capabilities, effectively shortening arc control delay and improving welding quality consistency.

[0046] Improve welding quality and production efficiency: Through the combination of multimodal fusion perception and intelligent control, the welding control system using this method can maintain arc stability, uniform penetration, and reduced spatter particles under conditions of complex welding components, such as curved welds and changes in welding material thickness. It can adjust the output welding current and welding voltage in time according to changes in welding conditions, thereby improving the welding qualification rate of complex components and reducing energy consumption. It is especially suitable for welding high-standard thin-walled structures and complex-shaped components.

[0047] In contrast, the technical solution of the present invention is based on the application of the arc dynamic adaptive control method of multimodal sensing in the welding control system. Through multimodal sensing, it integrates multiple real-time monitored welding data, which can improve the information perception ability of the welding control system. Combined with the deep Q model, the collected welding data is judged and the output welding control parameters are automatically forced to change according to the real-time changing welding conditions, thereby realizing dynamic adaptive control of the arc welding process, improving the control accuracy and response speed of the welding arc, and enabling welding to maintain stable welding quality under complex working conditions, improve welding qualification rate, and reduce energy consumption. It is suitable for high-requirement and high-variability welding scenarios, especially for the welding needs of aerospace thin-walled components, and has good engineering application prospects.

[0048] As one of the optional implementations

[0049] Regarding the specific settings in the above-mentioned feature extraction and decision-making control steps of the arc morphology, in the feature extraction step: the collected arc morphology image is preprocessed by a convolutional neural network, the arc morphology features are filtered to obtain the arc morphology features, the arc morphology features are analyzed, the geometric morphology data of the arc is obtained, and then the arc length variation coefficient is calculated, and the arc length variation coefficient data is output; in the decision-making control step: the arc length variation coefficient is input into the deep Q model, the arc stability is judged according to the arc length variation coefficient, and the parameters of the wire feeding speed, pulse waveform, pulse frequency and shielding gas flow are controlled and adjusted to realize real-time adjustment of the arc during the welding process and maintain stable arc output.

[0050] When applying, the smaller the arc length variation coefficient, the more stable the arc, and the larger the arc length variation coefficient, the more unstable the arc; a sudden increase in arc length may indicate that the parent material is not fully melted, while arc contraction may cause spatter; through the preliminary analysis of the arc length variation coefficient, the deep Q model can be used to determine whether the current arc length is stable. If it is unstable, the impact of the arc length change can be further analyzed; in the component reward function, the increase in the arc length variation coefficient can be limited as a penalty item. When the arc length variation coefficient increases, the lower the reward function value, the deeper the Q model is forced to adjust the welding control parameters. By adjusting the wire feeding speed, the melting amount of the welding material is increased, the distance between the welding material and the parent material, the pulse waveform, the pulse frequency and the parameters of the shielding gas flow rate are shortened, and the stable arc length is optimized, thereby realizing real-time adjustment of the welding process and improving the welding quality.

[0051] Regarding the specific settings in the above-mentioned feature extraction and decision control steps of the molten pool area image, in the feature extraction step: the collected molten pool area image is preprocessed by the image processing algorithm, the image of the spatter metal generated during the welding process is extracted, the number of spatter metal particles is analyzed and counted, the real-time spatter particle amount is obtained, and the spatter particle amount data is output; in the decision control step: the spatter particle amount is input into the deep Q model, the molten pool state and the arc state are judged according to the spatter particle amount, and the parameters of the pulse frequency and the shielding gas flow are controlled and adjusted.

[0052] When applied, the number of spatter metal particles generated during welding directly reflects the process stability; the fewer the number of spatter metal particles during welding, the higher and more sufficient the welding energy utilization rate; when the number of spatter metal particles increases, the welding energy utilization rate decreases, and the welding current, welding voltage, arc output, shielding gas flow rate, etc. need to be adjusted; when constructing the reward function of the deep Q model, the amount of spatter particles can be limited as a penalty item. The larger the amount of spatter particles, the lower the reward function value, forcing the deep Q model to optimize the welding control parameters to reduce the number of spatter metal particles.

[0053] Regarding the specific settings in the above-mentioned feature extraction and decision control steps of the molten pool oscillation acoustic signal, in the feature extraction step: the collected molten pool oscillation acoustic signal is pre-processed to reduce noise using a wavelet conversion algorithm, filtered to obtain the molten pool oscillation main frequency feature, and then the molten pool penetration value is calculated, and the penetration standard deviation is calculated;

[0054] In the decision-making and control step: the standard deviation of penetration is input into the depth Q model, the consistency of weld penetration is quantified according to the standard deviation of penetration, the stability of the welding process is judged, and the parameters of wire feed speed, pulse frequency and shielding gas flow are controlled and adjusted.

[0055] When applied, the standard deviation of penetration depth can be used to judge the dynamic stability of penetration depth. The smaller the standard deviation of penetration depth, the more stable the fusion of the weld zone, the appropriate cumulative heat input, and the higher the welding quality. The larger the standard deviation of penetration depth, the more unstable the fusion of the weld zone, the cumulative heat input is too large or too small, the welding quality is lower, and the anti-interference ability of the welding process is lower. When the thickness of the base material suddenly becomes smaller or thicker, or the airflow disturbance changes, the penetration depth value will fluctuate greatly. In this way, when constructing the reward function of the deep Q model, the negative penalty mechanism can be used to force the deep Q model to adjust the welding control parameters to reduce penetration fluctuations.

[0056] Regarding the specific settings in the above-mentioned feature extraction and decision control steps for welding current and welding voltage, in the feature extraction step: by collecting welding current and welding voltage, the instantaneous arc energy input is represented by the product of welding current and welding voltage, and then the cumulative heat input data is calculated by an integration algorithm;

[0057] In the decision-making control step: the accumulated heat input data is input into the deep Q model, and it is determined whether the preset parameters are reached based on the accumulated heat input data, and the parameters of the pulse waveform and pulse frequency are controlled and adjusted.

[0058] When applied, the cumulative heat input data can assist the depth Q model, coordinate with the standard deviation of penetration depth, the number of spattering metal particles and the coefficient of variation of arc length, judge the real-time status of the welding zone, and enable the depth Q model to make optimal adjustments to the welding control parameters.

[0059] Regarding the specific method of adjusting the output welding control parameters in real time according to the real-time monitoring welding parameters to improve the welding quality of complex welding paths and complex components, this embodiment is as follows: Figure 2 As shown, the following steps are also included: Library construction: according to different welding materials and base materials as well as process requirements, a weight library and a threshold library are established.

[0060] Before starting welding, the actual welding material and base material parameters are input to perform virtual feature extraction. The deep Q model automatically retrieves the corresponding welding weight values ​​and welding threshold values ​​from the weight library and threshold library, and uses the welding threshold value as the boundary parameter of welding control and the called welding weight value as the initial welding parameter to generate a virtual welding instruction. An external welding simulator is introduced to perform welding simulation according to the welding threshold value as the boundary parameter of welding control, verify the welding effect of the virtual welding instruction, and determine whether the welding effect is qualified.

[0061] If the welding effect is qualified, a welding instruction is generated, and welding control parameters are output after conversion. The welding executor receives the welding control parameters and executes welding.

[0062] If the welding effect is unqualified, similar parent materials and welding materials in the left and right domains of the actual welding material and parent material parameters will be automatically expanded and selected, and the weight library will be re-entered to retrieve the welding weight value, generate a virtual welding instruction, and introduce an external welding simulator to perform welding simulation according to the welding threshold value of the actual welding material and parent material parameters to verify the welding effect of the virtual welding instruction.

[0063] Specifically, during welding: after the data acquisition step is completed by monitoring the welding process through multi-modal sensors to obtain multi-dimensional data, the feature extraction step is entered, and the extracted parameters are input into the deep Q model to implement the decision control step. The deep Q model optimizes the welding weight value through the reward function, adjusts the welding parameters, regenerates the welding instructions, forms an instruction set, and outputs the welding control parameters after conversion. The welding actuator receives the welding control parameters and executes welding; when performing welding, the data acquisition step is run in parallel, and monitoring through multi-modal sensors continues. The changes in the welding effect are judged through the feature extraction step. If the welding effect changes, the decision control step is entered, the welding parameters are adjusted and output to the welding actuator, and the welding process of the welding gun is adjusted in real time. In this way, dynamic adaptive control of the arc is achieved in a cycle.

[0064] When applying, input the actual welding material and base material parameters, deep Q model, and automatically retrieve the corresponding welding weight value and welding threshold value from the weight library and threshold library. This step is the step of importing the welding indicators and parameters of the product; use the welding threshold value as the boundary parameter of welding control and the called welding weight value as the initial welding parameter to generate a virtual welding instruction, introduce an external welding simulator and use the welding threshold value as the boundary parameter of welding control to perform welding simulation and verify the welding effect of the virtual welding instruction. This step is the pre-verification step of the welding plan; after the data acquisition step is realized by monitoring the welding process with multi-dimensional data through multi-modal sensors, enter the feature extraction step, and extract the extracted data. The parameters are input into the deep Q model to implement the decision-making control step. The deep Q model optimizes the welding weight value through the reward function, adjusts the welding parameters, regenerates the welding instructions, and outputs the welding control parameters after conversion. The welding actuator receives the welding control parameters and executes welding; when performing welding, the data acquisition step runs in parallel, continues to be monitored by the multimodal sensor, and judges the changes in the welding effect through the feature extraction step. If the welding effect changes, the decision-making control step is entered, and the welding parameters are adjusted to be output to the welding actuator, and the welding process of the welding gun is adjusted in real time. In this way, dynamic adaptive control of the arc is realized in a cycle. This step is the actual verification step of the welding plan and the real-time adjustment step during the welding process.

[0065] Through the above-mentioned steps of importing product welding indicators, pre-verification of welding plans, actual verification of welding plans, and real-time adjustment during the welding process, the welding process production line debugging cycle can be shortened and manual intervention can be reduced when changing product lines. By introducing an external welding simulator to pre-verify the welding plan in a virtual environment, the time and labor cost of repeated experiments in actual welding are reduced. At the same time, it lays the foundation for the actual verification of the welding plan and real-time adjustment when encountering fluctuations in welding conditions after the welding plan is finalized.

[0066] Regarding the specific settings of the above-mentioned deep Q model, the deep Q model constructs a reward function through the arc length variation coefficient, the standard deviation of the penetration depth, the amount of spatter particles and the cumulative heat input, and adjusts the wire feeding speed, pulse waveform, pulse frequency and shielding gas flow rate.

[0067] During application, the reward function converts the abstract "welding quality" into a computable optimization target through the quantitative coupling of the arc length variation coefficient, the standard deviation of the penetration depth, the amount of spatter particles, and the cumulative heat input. Then, the priority of multi-objective optimization can be balanced by selecting the weight value. Combined with the constraint of the welding threshold value on the welding control parameters, Pareto optimality is achieved in the adjustment of the welding control parameters. The dynamic adjustment of the weight value further enhances the system's adaptability to complex working conditions, enhances the welding quality stability of the welding control system, and solves the pain point of multi-objective coordination in traditional welding control methods.

[0068] Based on the above embodiment of the arc dynamic adaptive control method, a sensing system is provided, including a welding gun, an embedded signal processing unit, a welding power supply and an edge computing unit, and the above-mentioned arc dynamic adaptive control method based on multimodal sensing, wherein the welding actuator includes a welding power supply and a welding gun; the welding gun is electrically connected to the welding power supply, and is used to receive the output of the welding power supply and form an output arc; the embedded signal processing unit is electrically connected to the welding gun, the welding power supply and the multimodal sensor, and is used to collect and organize data on the real-time wire feeding speed and real-time shielding gas flow of the welding gun, the pulse waveform, pulse frequency, output current and output voltage data output by the welding power supply, and the arc morphology image, molten pool area image, molten pool oscillation sound data, welding current and welding voltage data collected in real time by the multimodal sensor; the edge computing unit is electrically connected to the embedded signal processing unit and the welding actuator, and is used to receive the data collected and organized by the embedded signal processing unit, extract feature parameters through the built-in module, calculate the arc length variation coefficient, the standard deviation of the penetration depth, the amount of spatter particles and the cumulative heat input, and input them into the deep Q model module, and then calculate and output the welding control parameters to the welding actuator through the deep Q model module.

[0069] Among them, in order to realize the corresponding calculations, the edge computing unit is deployed with a deep Q model module, a convolutional neural network module, an image processing algorithm module, a wavelet conversion algorithm module and a welding electrical parameter calculation module.

[0070] During application, an edge computing unit is used for data processing and welding control parameter output. The response time of the welding system can reach 10ms. With the improvement of response speed, this technical solution controls the arc of the welding process in real time by monitoring-extracting features-analyzing real-time parameters-outputting real-time welding parameters. This method can enhance the control speed of the welding control system over the welding process, further accelerate the welding adjustment performance of the welding system in the face of complex and changing welding conditions, thereby further improving the flexibility of the welding system and the quality of the welded components.

[0071] Regarding the above-mentioned welding pre-verification step of introducing an external welding simulator to perform welding simulation according to the welding threshold value as the boundary parameter of welding control to verify the welding effect of the virtual welding instruction, the edge computing unit is provided with a simulator interface for introducing an external welding simulator.

[0072] When applying, an external welding simulator, such as a finite element analysis system or a CAD\CAM simulation system, is called through a simulator interface to perform virtual simulation of welding, which can enhance the flexible welding performance of the welding system to which this technical solution is applied. When changing product lines, virtual parameters can be generated based on the changed or imported parameters to perform welding simulation and verify the welding effect. The welding plan can be verified in advance for the real-time simulation of the welding product, thereby improving the efficiency of changing product lines and improving the capacity and efficiency of the welding production line.

[0073] The second embodiment of the sensing system differs from the first embodiment in that the welding gun is an integrated welding gun, in which the multimodal sensor is directly integrated into the welding gun of the welding control system. The welding gun has the function of adjusting the wire feeding speed and the shielding gas flow at the welding gun mouth, which serves as the structural basis for real-time adjustment of welding control parameters.

[0074] The technical features of the above embodiments may be combined arbitrarily. To simplify the description, not all possible combinations of the technical features in the above embodiments are described.

Claims

1. Arc dynamic adaptive control method based on multimodal sensing, characterized in that: The following steps are involved: Data acquisition: Multi-dimensional data of the welding process is acquired through a multi-modal sensor integrated in the welding gun, including: arc morphology image, weld pool area image, weld pool oscillation acoustic signal, welding current and welding voltage; Feature extraction: Pre-process the collected multi-dimensional data to obtain arc morphological characteristics and molten pool oscillation main frequency characteristics, and calculate the characteristic parameters through post-processing; Decision-making control: Pre-training forms a deep Q model, which calculates and outputs the optimal welding control parameters in real time based on the extracted feature parameters, including wire feed speed, pulse waveform, pulse frequency, and shielding gas flow rate; Execution control: The output welding control parameters are sent to the welding actuator to adjust the welding process of the welding gun in real time to achieve dynamic adaptive control of the arc.

2. The arc dynamic adaptive control method based on multimodal sensing according to claim 1, characterized in that: In data acquisition, the multimodal sensor setup includes a high-speed camera, acoustic emission sensor, Hall current sensor, and shunt. The high-speed camera and acoustic emission sensor are oriented toward the weld, while the Hall current sensor and shunt are located on the weld loop. The arc shape image and molten pool area image are collected by a high-speed camera; the molten pool oscillation acoustic emission data are collected by an acoustic emission sensor; the end welding current and welding voltage data are collected by a Hall current sensor and a shunt respectively.

3. The arc dynamic adaptive control method based on multimodal sensing according to claim 2, characterized in that: In the feature extraction step: the collected arc morphology image is preprocessed by a convolutional neural network, arc morphology features are filtered and obtained, the arc morphology features are analyzed, the arc geometric data is obtained, and then the arc length variation coefficient is calculated and the arc length variation coefficient data is output; In the decision-making and control steps: the arc length variation coefficient is input into the deep Q model, the arc stability is judged according to the arc length variation coefficient, and the parameters of wire feeding speed, pulse waveform, pulse frequency and shielding gas flow are controlled and adjusted to achieve real-time adjustment of the arc during the welding process and maintain stable arc output.

4. The arc dynamic adaptive control method based on multimodal sensing according to claim 3, characterized in that: In the feature extraction step: The collected images of the molten pool area are preprocessed by image processing algorithms to extract images of the spatter metal generated during the welding process, analyze and count the number of spatter metal particles, obtain the real-time spatter particle amount, and output the spatter particle amount data; In the decision-making control step: the amount of spatter particles is input into the depth Q model, the molten pool state and arc state are judged according to the amount of spatter particles, and the parameters of the pulse frequency and shielding gas flow are controlled and adjusted.

5. The arc dynamic adaptive control method based on multimodal sensing according to claim 4, characterized in that: In the feature extraction step: The collected molten pool oscillation acoustic signal is pre-processed to reduce noise using the wavelet conversion algorithm, and the main frequency characteristics of the molten pool oscillation are obtained by filtering. The molten pool penetration value is then deduced, and the standard deviation of the penetration depth is calculated. In the decision-making and control step: the standard deviation of penetration is input into the depth Q model, the consistency of weld penetration is quantified according to the standard deviation of penetration, the stability of the welding process is judged, and the parameters of wire feed speed, pulse frequency and shielding gas flow are controlled and adjusted.

6. The arc dynamic adaptive control method based on multimodal sensing according to claim 5, characterized in that: In the feature extraction step: By collecting welding current and welding voltage, the instantaneous arc energy input is represented by the product of welding current and welding voltage, and the cumulative heat input data is calculated by the integration algorithm; In the decision-making control step: the accumulated heat input data is input into the deep Q model, and it is determined whether the preset parameters are reached based on the accumulated heat input data, and the parameters of the pulse waveform and pulse frequency are controlled and adjusted.

7. The arc dynamic adaptive control method based on multimodal sensing according to claim 6, characterized in that: The deep Q model constructs a reward function through the arc length variation coefficient, penetration standard deviation, spatter particle amount and cumulative heat input, and adjusts the wire feeding speed, pulse waveform, pulse frequency and shielding gas flow rate.

8. The arc dynamic adaptive control method based on multimodal sensing according to claim 7, characterized in that: The following steps are also included: Library construction: according to different welding materials, base materials and process requirements, establish weight library and threshold library; Before starting welding, the actual welding material and base material parameters are input to perform virtual feature extraction. The deep Q model automatically retrieves the corresponding welding weight values ​​and welding threshold values ​​from the weight library and threshold library, uses the welding threshold value as the boundary parameter of welding control, and uses the called welding weight value as the initial welding parameter to generate a virtual welding instruction. An external welding simulator is introduced to perform welding simulation according to the welding threshold value as the boundary parameter of welding control, verify the welding effect of the virtual welding instruction, and judge whether the welding effect is qualified. If the welding effect is qualified, a welding instruction is generated, and welding control parameters are output after conversion. The welding executor receives the welding control parameters and performs welding; If the welding effect is unqualified, similar parent materials and welding materials in the left and right domains of the actual welding material and parent material parameters are automatically expanded and selected, and the welding weight value is retrieved from the weight library again to generate a virtual welding instruction. An external welding simulator is introduced to perform welding simulation according to the welding threshold value of the actual welding material and parent material parameters to verify the welding effect of the virtual welding instruction; During welding: after the data acquisition step is completed by monitoring the welding process with multi-dimensional data through multi-modal sensors, the feature extraction step is entered, and the extracted parameters are input into the deep Q model to implement the decision control step. The deep Q model optimizes the welding weight value through the reward function, adjusts the welding parameters, regenerates the welding instructions, forms an instruction set, and outputs the welding control parameters after conversion. The welding actuator receives the welding control parameters and executes welding; when executing welding, the data acquisition step is run in parallel, and monitoring continues through multi-modal sensors. The changes in the welding effect are judged through the feature extraction step. If the welding effect changes, the decision control step is entered, the welding parameters are adjusted and output to the welding actuator, and the welding process of the welding gun is adjusted in real time. In this way, dynamic adaptive control of the arc is achieved in a cycle.

9. A sensing system, characterized in that: The invention comprises a welding gun, an embedded signal processing unit, a welding power source and an edge computing unit and the arc dynamic adaptive control method based on multimodal sensing according to any one of claims 1 to 8, The welding actuator includes the welding power source and the welding gun; The welding gun is electrically connected to the welding power source, and is used to receive the output of the welding power source and form an output arc; The embedded signal processing unit is electrically connected to the welding gun, the welding power supply and the multimodal sensor, and is used to collect and organize data on the real-time wire feeding speed and real-time shielding gas flow of the welding gun, data on the pulse waveform, pulse frequency, output current and output voltage output by the welding power supply, and data on arc morphology, molten pool area, molten pool oscillation sound emission, welding current and welding voltage collected in real time by the multimodal sensor; The edge computing unit is deployed with a deep Q model module, a convolutional neural network module, an image processing algorithm module, a wavelet conversion algorithm module and a welding electrical parameter calculation module; The edge computing unit is electrically connected to the embedded signal processing unit and the welding actuator, and is used to receive data collected and organized by the embedded signal processing unit, extract characteristic parameters through the built-in module, calculate the arc length variation coefficient, penetration standard deviation, spatter particle amount and cumulative heat input and input them into the deep Q model module, and then calculate and output welding control parameters to the welding actuator through the deep Q model module.

10. The sensing system according to claim 9, characterized in that The edge computing unit is provided with a simulator interface for introducing an external welding simulator.

Images