A deep learning-based welding process parameter self-adaptive optimization method

Abstract

This invention relates to the field of intelligent manufacturing technology and provides a deep learning-based adaptive optimization method for welding process parameters. The method includes: acquiring the weld path and quality target of the parts to be welded; dividing the weld path into a set of weld segment units; collecting pre-welding geometric data and in-welding process data for each weld segment unit to obtain segment welding state characteristics; constructing a local weldability risk field to obtain segment risk field results; generating a segment parameter candidate set based on the welding parameter safety working window as a constraint and the segment risk field results; inputting the segment welding state characteristics, segment risk field results, and segment parameter candidate set into a pre-constructed physical constraint deep learning model to obtain segment quality prediction results; and performing rolling parameter optimization based on the segment quality prediction results to obtain a segment welding parameter sequence and sending it to the welding execution end.

Description

Technical Field

[0001] This invention relates to the field of intelligent manufacturing technology, specifically a deep learning-based adaptive optimization method for welding process parameters. Background Technology

[0002] In the robotic welding of long weld seams in aluminum alloy battery trays or underbody frames of new energy vehicles, batch differences in aluminum profiles, fluctuations in clamping gaps, changes in misalignment, changes in local heat dissipation boundaries, and heat accumulation in previous weld seams will cause different requirements for laser power, welding speed, wire feed speed, oscillation amplitude, and shielding gas flow rate at different locations of the same weld seam. Therefore, it is necessary to generate segmented welding process parameters in real time based on pre-weld geometric deviations, the state of the molten pool during welding, and the state of heat accumulation along the weld seam, without relying on post-weld destructive testing or significantly reducing the production line cycle time. This will ensure that the weld seam meets the requirements of stable penetration depth, controlled porosity and crack risks, no excessive heat input, and controllable post-weld deformation.

[0003] Current technologies typically begin by determining a set of welding process parameters suitable for a specific type of component and weld through process experiments, empirical parameter tables, response surface methodology, orthogonal experiments, or neural network models. Then, cameras, infrared sensors, arc voltage and current signals, or post-weld visual inspection are used to determine if the weld is abnormal. When the molten pool morphology, weld width, spatter, porosity, or appearance defects exceed allowable limits, the system issues an alarm, or the operator adjusts the parameters based on experience.

[0004] However, in continuous welding scenarios of long welds, existing technologies mostly use the entire weld or a fixed process window as the parameter setting object, without identifying the changes in local gaps, misalignments, heat dissipation paths, and heat accumulation according to the continuous position of the weld. This makes it difficult for the same set of parameters to simultaneously adapt to the different needs of the front, middle, end, and corner areas. When the local gap increases or heat dissipation is enhanced, fixed parameters can easily lead to insufficient penetration. When heat accumulates or heat dissipation deteriorates, fixed parameters can easily lead to overheating, cracking, or increased deformation. Therefore, existing technologies lack the ability to identify segmented states and adaptively generate segmented parameters for continuous changes in the local state of long welds. Summary of the Invention

[0005] In the robotic welding of long weld seams in aluminum alloy battery trays or underbody frames of new energy vehicles, the local gaps, misalignments, heat dissipation boundaries, and heat accumulation states of the weld seam change continuously along the path. Existing fixed process windows are difficult to adapt to the penetration depth, heat input, and deformation control requirements of different segments. At the same time, existing quality prediction and parameter recommendation lack joint constraints on quality objectives, heat input limitations, deformation risks, and parameter smoothness, making it difficult to stably transform the prediction results into an executable sequence of segmented welding parameters.

[0006] To at least partially solve the above problems, this invention proposes an adaptive optimization method for welding process parameters based on deep learning, comprising: Obtain the weld path and quality target of the parts to be welded, and divide the weld path into a set of weld segment units along the robot's running direction; Collect pre-welding geometric data and in-welding process data for each weld segment unit to obtain segment welding state characteristics; Construct the local weldability risk field for each weld segment unit to obtain the segment risk field results; A pre-constructed welding parameter safety working window is used as a constraint, and a segmented parameter candidate set is generated in combination with the segmented risk field results. The segmented welding state characteristics, segmented risk field results, and segmented parameter candidate set are input into a pre-built physical constraint deep learning model to obtain the segmented quality prediction results of the quality target. Based on the segmented quality prediction results, rolling parameter optimization is performed on each weld segment unit to obtain the segmented welding parameter sequence, which is then sent to the welding execution end.

[0007] In a preferred embodiment, the weld path is divided into a set of weld segment units along the robot's running direction, including: Taking the weld path coordinate sequence corresponding to the weld path as the object, multiple weld segmentation units are divided according to the robot's running direction; the path length corresponding to a single frame is determined based on the robot's travel speed and sensor sampling frequency, and the minimum sampling segment length is determined in combination with the lower limit of the frame number; a segment length range including a lower limit and an upper limit of the segment length is constructed based on the minimum sampling segment length; the straight stable region and the corner region are determined by the angle between the direction vectors of adjacent path points and a preset corner threshold; the straight stable region is segmented using the upper limit of the segment length, and the corner region is segmented using the lower limit of the segment length; a preset local region is segmented using the lower limit of the segment length to obtain the set of weld segmentation units; wherein, the preset local region includes the weld path end region, the fixture obstruction edge region, and the historical defect high-incidence region.

[0008] As a preferred implementation, pre-welding geometric data and in-welding process data of each weld segment are collected to obtain segment welding state characteristics, including: Pre-welding scanning is performed on the weld area of ​​each weld segment unit to obtain pre-welding geometric data. The pre-welding geometric data is then compared with the theoretical weld model to obtain pre-welding geometric deviation characteristics. During the welding process, welding process data of each weld segment unit is collected, and feature extraction is performed on the welding process data to obtain welding process stability characteristics. The pre-welding geometric deviation characteristics and welding process stability characteristics are aligned and fused according to the weld segment units to obtain the segmented welding state characteristics.

[0009] As a preferred implementation, a local weldability risk field is constructed for each weld segment unit to obtain the segment risk field results, including: The heat accumulation state along the weld is obtained by accumulating the temperature peak and cooling rate of the completed weld segment units along the path direction. Using the segment welding state characteristics and the heat accumulation state along the weld as fusion inputs, a local weldability risk field is constructed for each weld segment unit. The local weldability risk field includes geometric disturbance risk value, molten pool instability risk value, heat accumulation risk value, and deformation accumulation risk value. The above four types of risk values ​​are weighted and merged to obtain the comprehensive risk value of each weld segment unit. The comprehensive risk value and the four types of risk values ​​together constitute the segment risk field result.

[0010] As a preferred embodiment, the geometric disturbance risk value, the molten pool instability risk value, the heat accumulation risk value, and the deformation accumulation risk value include: The geometric disturbance risk value is determined by comparing the gap value, misalignment amount, and weld center offset of each weld segment unit with the corresponding upper limit of the threshold. The molten pool instability risk value is determined by normalizing the molten pool size deviation, molten pool brightness fluctuation, light intensity fluctuation, acoustic emission mutation, voltage fluctuation, and current fluctuation. The heat accumulation risk value is determined by the heat accumulation state value along the weld, the temperature peak value of adjacent welded segment units, the cooling rate attenuation, and the heat dissipation effect corresponding to the adjacent distance of the fixture. The deformation accumulation risk value is determined by weighting the position of the current weld segment unit along the weld path, the heat input accumulation of the previous weld segment unit, and the assembly tolerance sensitive area identifier.

[0011] As a preferred implementation, the physical constraint deep learning model includes: A multimodal feature extraction and fusion architecture is adopted to extract the molten pool morphology feature vector from the coaxial molten pool image sequence in the segmented welding state features, extract the process signal feature vector from the continuous process signal in the segmented welding state features, concatenate the comprehensive risk value and four types of risk components in the segmented risk field results into a risk state vector, and encode the current parameter candidate combination to be evaluated into a parameter vector; input the molten pool morphology feature vector, the process signal feature vector, the risk state vector and the parameter vector into a fully connected fusion layer, and output the segmented quality prediction result corresponding to the current parameter candidate combination.

[0012] As a preferred embodiment, after obtaining the segmented quality prediction results of the quality target, the method further includes: Using the lower limit of the penetration depth corresponding to the quality target, the upper limit of the heat input corresponding to the component to be welded, and the upper limit of the allowable deformation corresponding to the assembly tolerance in the quality target as screening constraints, the feasibility of the parameter candidate combinations in the segment parameter candidate set is screened; when the segment quality prediction result does not meet any of the screening constraints of the lower limit of penetration depth, upper limit of heat input, or upper limit of allowable deformation, the corresponding parameter candidate combination is removed from the segment parameter candidate set; retaining the parameter candidate combinations that meet all screening constraints, the feasible candidate set of each weld segment unit is obtained.

[0013] As a preferred embodiment, rolling parameter optimization is performed on each weld segment unit to obtain a segmented welding parameter sequence, including: Rolling parameter optimization is performed sequentially on each weld segment unit according to the weld path order. The optimal parameter combination is selected from the feasible candidate set of each weld segment unit to obtain the segmented welding parameter sequence. The rolling parameter optimization takes quality deviation, heat input risk, deformation accumulation risk and parameter jump penalty as optimization objectives. After the parameters of the current weld segment unit are determined, they are used as the benchmark parameters for the optimization of the next weld segment unit. The optimization is carried out segment by segment along the weld path until the parameters of all weld segment units have been selected. The segmented welding parameter sequence is converted into execution instructions that can be recognized by the welding execution end and sent to the robot controller and welding machine controller.

[0014] As a preferred implementation, the optimal parameter combination is selected from the feasible candidate set of each weld segment unit, including: The deviation of the segmented quality prediction result from the quality target is used as the quality deviation term; the relationship between the heat input prediction information and the heat input upper limit in the segmented quality prediction result is used as the heat input risk term; the relationship between the deformation accumulation prediction information and the allowable deformation upper limit in the segmented quality prediction result is used as the deformation accumulation risk term; and the change between the current candidate parameter combination and the actual executed parameter combination of the previous weld segment unit is used as the parameter jump penalty term. Based on the weighted result of the quality deviation term, heat input risk term, deformation accumulation risk term, and parameter jump penalty term, the optimal parameter combination of the current weld segment unit is determined.

[0015] In a preferred embodiment, after obtaining the segmented welding parameter sequence, the method further includes: The minimum executable segment length is determined based on the response capability of the welding execution end. For continuous weld segment units with a cumulative path length less than the minimum executable segment length, a preset merging threshold is used as the judgment boundary. Adjacent weld segment units whose comprehensive risk value difference and parameter difference do not exceed the corresponding judgment boundary are merged into a parameter execution interval. For adjacent weld segment units whose comprehensive risk value difference or parameter difference exceeds the corresponding judgment boundary, a parameter transition segment is set at the junction. After the above processing, a smooth segmented welding parameter sequence is obtained, and the smooth segmented welding parameter sequence is used to replace the segmented welding parameter sequence and sent to the welding execution end.

[0016] Compared with the prior art, the present invention has the following advantages: 1. By acquiring the weld path and quality target of the parts to be welded, and dividing the weld path into a set of weld segment units along the robot's running direction, the long weld is transformed from a unified control object into multiple independently identifiable and optimizable segment objects. This solves the problem that the existing fixed process window is difficult to adapt to the continuous changes in local gaps, misalignments, heat dissipation boundaries, and heat accumulation states of long welds.

[0017] 2. By collecting pre-welding geometric data and in-welding process data of each weld segment unit, the segment welding state characteristics are obtained, and the local weldability risk field of each weld segment unit is constructed. This enables the system to identify the risks of geometric disturbance, molten pool instability, heat accumulation and deformation accumulation in different segments before parameter output, thereby avoiding passive adjustments only after post-weld detection or in-weld abnormal alarm.

[0018] 3. By inputting the segmented welding state characteristics, segmented risk field results, and segmented parameter candidate set into a pre-built physical constraint deep learning model, the segmented quality prediction results corresponding to the quality objectives are obtained. This enables parameter evaluation to move beyond simple defect identification or acceptance judgment, and instead predict the quality consequences of different candidate parameters based on factors such as weld penetration, heat input, deformation, and defect risk.

[0019] 4. By performing rolling parameter optimization on each weld segment unit based on the segment quality prediction results, the segment welding parameter sequence is obtained and issued, enabling the welding parameters to be adaptively adjusted along the weld path according to local state changes. At the same time, it takes into account quality deviation, heat input risk, deformation accumulation risk and parameter jump control, thereby improving the penetration stability of long welds, reducing the risks of porosity, cracks, overheating and post-weld deformation, and improving the continuity and executability of robot welding parameter output. Attached Figure Description

[0020] To more clearly illustrate the technical solutions of the embodiments of this application, the accompanying drawings used in the embodiments will be briefly introduced below. It should be understood that the following drawings only show some embodiments of this application and should not be regarded as a limitation on the scope of this application.

[0021] Figure 1 This is an exemplary flowchart of the adaptive optimization method for welding process parameters provided in an embodiment of the present invention; Figure 2 This is a schematic diagram illustrating the effect of the adaptive optimization method for welding process parameters provided in this embodiment of the invention compared with the prior art, wherein gray bars represent the prior art and black bars represent the present invention. Detailed Implementation

[0022] To make the technical means, creative features, objectives, and effects of this invention easier to understand, the invention is further described below with reference to specific embodiments. However, the following embodiments are merely preferred embodiments of this invention and not all of them. Other embodiments obtained by those skilled in the art based on the embodiments described herein without creative effort are all within the protection scope of this invention.

[0023] Example 1:

[0024] To achieve the above objectives, please refer to Figures 1 to 2 This invention provides an adaptive optimization method for welding process parameters based on deep learning, the method comprising the following steps: Welding task segmentation: Obtain the weld path and quality target of the parts to be welded, and divide the weld path into a set of weld segment units along the robot's running direction; Segmented State Awareness: Collect pre-welding geometric data and in-welding process data of each weld segment unit to obtain segmented welding state characteristics; Weldability risk modeling: Construct the local weldability risk field for each weld segment unit and obtain the segment risk field results; Generate candidate parameters: Preconstruct a safe working window for welding parameters, and generate a candidate set of segmented parameters by using the safe working window for welding parameters as a constraint and combining the results of the segmented risk field; Predicting segment quality: Input the segment welding state characteristics, segment risk field results and segment parameter candidate set into a pre-built physical constraint deep learning model to obtain the segment quality prediction results of the quality target; Rolling parameter optimization: Based on the segmented quality prediction results, rolling parameter optimization is performed on each weld segment unit to obtain the segmented welding parameter sequence, which is then sent to the welding execution end.

[0025] In this embodiment, the specific content of the welding task segmentation includes: Retrieve the manufacturing execution system work order to obtain the part number, material grade, plate thickness or profile wall thickness, weld type, weld path coordinate sequence, target penetration depth, sealing grade, strength grade, and assembly tolerance of the components to be welded, forming welding task data. Among them, the target penetration depth, sealing grade, strength grade, and assembly tolerance constitute the quality target constraints for subsequent parameter optimization, while the material grade, plate thickness or profile wall thickness, weld type, and weld path coordinate sequence constitute the process basis constraints for subsequent parameter optimization. If there is no manufacturing execution system work order, the above data are obtained sequentially from the product drawings, welding procedure qualification documents, and component dimensional chain requirements.

[0026] Further, taking the weld path coordinate sequence as the object, it is divided into multiple weld segment units according to the robot's running direction; the path length corresponding to a single frame is obtained by dividing the robot's travel speed by the sensor sampling frequency, and then multiplied by the lower limit of the frame number to obtain the minimum sampling segment length, so that each weld segment unit corresponds to at least 3 frames of pre-welding geometric image acquisition windows and 3 sets of welding process signal acquisition windows; a segment length range is constructed based on the minimum sampling segment length, the segment length range including a lower limit of the segment length and an upper limit of the segment length; in one embodiment, the lower limit of the segment length is 5mm, and the upper limit of the segment length is 20mm; the angle between the direction vectors of adjacent path points in the weld path coordinate sequence is used as the determination criterion, and a preset turning angle threshold is used as the determination boundary to perform region type determination on each path segment; in one embodiment, the preset turning angle threshold is 15°; when the angle does not exceed the preset turning angle threshold, the path segment is determined to be a straight stable region, and the region is selected as the region. The upper limit of the segment length; when the included angle exceeds the preset corner threshold of the path segment, it is determined to be a corner area, and the lower limit of the segment length is taken; the overlapping end is determined within a range of 10mm from the beginning and end of the weld path coordinate sequence, and the clamp obstruction edge is determined by extending 5mm to both sides based on the clamp position coordinates recorded in the robot tooling file; for the overlapping end and the clamp obstruction edge, the lower limit of the segment length is used for segmentation; the historical defect high-incidence area is obtained by statistical analysis of defect coordinates in historical welding records; when the historical welding records have not reached the preset sample number, segmentation is performed according to the unified segment length of the entire path; when the historical welding records have reached the preset sample number, the historical defect high-incidence area is updated according to the defect coordinate distribution. In one embodiment, the preset sample number is 30; all weld segmentation units together constitute a weld segmentation unit set, which serves as a unified index object for subsequent segmented welding state feature acquisition, segmented risk field construction, and segmented parameter optimization.

[0027] In this embodiment, the specific content of segmented state perception includes: Taking each weld segment unit in the weld segment unit set as the object, the weld area of ​​each weld segment unit is scanned before welding to obtain a three-dimensional point cloud of the weld surface; the three-dimensional point cloud is compared with the theoretical weld model to extract the gap value, misalignment amount, weld center offset, lap width, weld curvature and fixture proximity distance.

[0028] Specifically, the theoretical weld model is constructed from the weld path coordinate sequence. Taking the three-dimensional coordinates of each path point in the weld path coordinate sequence as the center line, and combining the target gap value, target bevel width, and target overlap width specified in the process qualification document, it expands to both sides along the path normal direction to generate the theoretical geometric envelope surface corresponding to each weld segment unit. The theoretical geometric envelope surface is the theoretical weld model, which represents the standard geometric shape of the weld area under the conditions of no deviation of the parts and no clamping error. If the process qualification document does not specify the target gap value or bevel width, the average value of the corresponding geometric quantity of qualified samples in the trial production stage is used instead.

[0029] The gap value is calculated by extracting cross-sectional point cloud slices of each segment unit along the weld seam normal in the 3D point cloud, identifying the inner edge points of the point groups on both sides of the base material surface in the slices, and taking the horizontal distance between the inner edge points on both sides as the gap value of the current segment unit; the misalignment is calculated by the difference in the average height of the point groups on both sides of the base material surface in the same cross-sectional point cloud slice, with the height direction taken as the normal component perpendicular to the base material surface; the weld seam center offset is obtained by extracting the actual weld seam center position from the cross-sectional slices of each weld seam segment unit and comparing the actual weld seam center position with the robot's theoretical path; for bevel welds, the point groups on both sides of the bevel are fitted, and the intersection of the extended lines of the two bevels or the center line of the bevel opening is taken as the actual weld seam center position; for lap welds, the lap edge positions of the upper and lower plates are extracted, and the center position of the effective lap width is taken as the actual weld seam center position; the gap value is calculated by extracting the cross-sectional point cloud slices of each segment unit along the weld seam normal in the 3D point cloud slices, and taking the intersection of the extended lines of the two bevels or the center line of the bevel opening as the actual weld seam center position; the gap value is calculated by extracting the cross-sectional point cloud slices of each segment unit along the weld seam normal in the 3D point cloud slices, identifying the inner edge points of the point groups on both sides of the bevel, and taking the center position of the effective lap width as the actual weld seam center position; the gap value is calculated by extracting the cross-sectional point cloud slices of each segment unit along the weld seam normal in the 3D point cloud slices, and taking the intersection of the extended lines of the two bevels or the center line of the bevel opening as the actual weld seam center position; the gap value is calculated by extracting the cross-sectional point cloud slices of each segment unit along the weld seam normal in the 3D The actual weld center position is projected onto the straight line or curve of the robot's theoretical path, and the vertical distance between the projection point and the actual weld center position is taken as the weld center offset. The overlap width is obtained by detecting the step difference edge points of the point groups on the surfaces of the upper and lower plates along the normal direction in the cross-sectional slice, and taking the horizontal distance between the inner step difference edge points of the upper plate and the inner step difference edge points of the lower plate as the actual overlap width. The weld curvature is obtained by fitting a second-order polynomial to the point cloud centerline point sequence in each weld segment unit, and taking the ratio of the second derivative to the first derivative of the fitted curve at the center of the segment unit. The fixture proximity distance is obtained by calculating the minimum Euclidean distance between the point group belonging to the fixture contour in the three-dimensional point cloud and the center coordinate point of the current weld segment unit, wherein the fixture contour point group is obtained by registering the point cloud with the three-dimensional fixture model recorded in the robot tooling file and then removing the parent material point group.

[0030] Furthermore, during the welding process, the following multimodal process signals are simultaneously acquired for each weld segment unit: Specifically, coaxial molten pool images of the welding area are acquired, and the molten pool width, length, and brightness fluctuations are extracted to characterize the penetration trend and wetting stability; infrared radiation signals of the welding area are acquired, and temperature peaks and cooling rates are extracted to characterize the local heat input level; plasma light intensity signals of the welding area are acquired, and light intensity fluctuations are extracted; acoustic emission signals generated during welding are acquired, and acoustic emission mutations are extracted; arc voltage and welding current signals of the welding circuit are acquired, and voltage and current fluctuations are extracted; the aforementioned light intensity fluctuations, acoustic emission mutations, voltage fluctuations, and current fluctuations are combined to characterize the risks of porosity, collapse, and unstable keyholes; and robot joint pose data are acquired, and the difference between the actual operating speed and the commanded speed of each weld segment unit is calculated to obtain the robot speed deviation, which is used to eliminate false abnormal signals caused by motion execution errors.

[0031] For the weld segment unit to be optimized, the welding process data includes the process data collected within the current welding window, or the process data of adjacent welded weld segment units; the welding process data is used to update the segment welding status characteristics of weld segment units after the current welding window, avoiding the use of future process data of weld segment units that have not yet been welded to generate the current welding parameters.

[0032] The geometric deviation obtained from the pre-weld 3D scan is aligned with the modal process signals obtained during welding according to the weld segment units and then merged to obtain the segmented welding state characteristics, which serve as the input for subsequent local weldability risk field construction and physical constraint deep learning model.

[0033] In this embodiment, the specific content of weldability risk modeling includes: Using the segmented welding state characteristics and the heat accumulation state along the weld as fusion inputs, a local weldability risk field is constructed for each weld segment unit; wherein, the heat accumulation state along the weld is calculated by accumulating the temperature peak and cooling rate of the weld segment unit that has been welded along the path direction, as shown in the formula: , in Let be the thermal accumulation state value of the i-th weld segment unit, characterizing the residual heat transferred to the i-th weld segment unit from the preceding welding process. The heat decay factor characterizes the residual heat along the path from the first... The degree of attenuation during the transmission from the first weld segment unit to the i-th weld segment unit; The thermal attenuation coefficient is determined by the material grade and plate thickness, and is derived from the thermophysical parameters of the corresponding material in the process qualification document. In one embodiment, for aluminum alloy materials, the thermal attenuation coefficient is... The value is 0.05 ; For the first The path length of each weld segment unit is derived from the coordinate sequence length of the corresponding segment unit in the weld segment unit set; For the first The actual heat input of each weld segment unit; The heat input conversion coefficient characterizes the proportion of contribution of a unit heat input to the value of the thermal accumulation state, and is determined by the material's heat capacity and plate thickness. In one embodiment, the heat input conversion coefficient... The value is 0.1; for the first weld segment unit in the weld path, the value is... It is equal to the difference between the ambient temperature before welding and the initial temperature of the component. In one embodiment, the initial value is 0.

[0034] The local weldability risk field uses each weld segment unit as the index object and includes four types of risk components: geometric disturbance risk value. 1. Molten pool instability risk value Heat accumulation risk value and cumulative risk value of deformation .

[0035] The geometric disturbance risk value The value is obtained by averaging the gap value, misalignment amount, and weld center offset by their respective upper threshold values. These upper threshold values ​​are derived from the process qualification document; if not specified in the document, the 95th percentile of the geometric deviation corresponding to a qualified sample from the trial production stage is used. The molten pool instability risk value... The heat accumulation risk value is obtained by averaging the normalized values ​​of molten pool width deviation, molten pool length deviation, molten pool brightness fluctuation, light intensity fluctuation, acoustic emission abrupt change, voltage fluctuation, and current fluctuation. The normalization benchmark is the upper limit of the statistical range of the corresponding feature values ​​of qualified samples. The temperature is determined by the ratio of the peak temperature of adjacent welded seam segments to the upper temperature limit calculated from the allowable heat input of the material, the attenuation of the cooling rate of adjacent welded seam segments, and the heat dissipation effect corresponding to the proximity distance of the fixture; wherein, the peak temperature and cooling rate are derived from the segmented welding state characteristics, the residual heat information is derived from the heat accumulation state along the weld, and the proximity distance of the fixture is derived from the segmented welding state characteristics; the deformation accumulation risk value The value is calculated by weighting the position index of the current weld segment unit in the weld path, the cumulative heat input of the previous weld segment unit, and the assembly tolerance sensitive area identifier. The assembly tolerance sensitive area identifier is obtained by mapping the assembly tolerance and the weld path coordinate sequence according to the position correspondence.

[0036] The above four risk components are combined into a comprehensive risk value using the following formula. : , in, The comprehensive risk value for the i-th weld segment unit; , , , These are weighting coefficients for geometric disturbance risk, molten pool instability risk, thermal accumulation risk, and deformation accumulation risk, respectively; in one embodiment, the weighting coefficients... , , , The initial value of each is 0.25.

[0037] During the trial production phase, once the number of historical welding records reaches the preset sample size, the weighting coefficients are updated based on the frequency proportion of defects in manually marked samples, such as insufficient penetration, porosity, cracks, and deformation exceeding tolerance. , , , This gives higher weight to risk components that are more likely to lead to component scrapping; in one embodiment, the preset sample size is 30 samples; the comprehensive risk value of each weld segment unit. Together with the four types of risk components, they constitute the segmented risk field results.

[0038] Specifically, the weighting coefficients , , , The update method is as follows: Defect types are categorized according to their dominant risk components: insufficient penetration is categorized into geometric disturbance risk and molten pool instability risk; porosity and cracks are categorized into molten pool instability risk; and deformation exceeding tolerance is categorized into heat accumulation risk and deformation accumulation risk. For defect types involving multiple risk components, they are proportionally allocated to their respective risk components. The proportion of defect frequency corresponding to each risk component is calculated to obtain the defect frequency percentage for each risk component. The corresponding weighting coefficient is directly replaced by the defect frequency percentage. In one implementation, if insufficient penetration accounts for 30%, porosity and cracks account for 40%, and deformation exceeding tolerance accounts for 30% in historical samples, and insufficient penetration is proportionally allocated to… and Deformation deviations are distributed proportionally to and The weighting coefficients are updated on a rolling basis according to a preset update cycle as historical welding records are continuously accumulated. In one embodiment, the update cycle is performed once every 30 new samples.

[0039] In this embodiment, the specific content of generating candidate parameters includes: Based on the welding task data, and combined with the welding machine capability boundary and the process qualification range, a safe working window for welding parameters is constructed. The safe working window for welding parameters is a set of allowable value ranges for each welding process parameter, including the laser power range, welding speed range, wire feed speed range, oscillation frequency range, oscillation amplitude range, defocusing amount range, and shielding gas flow rate range.

[0040] Specifically, the sources of the parameter ranges are as follows: the laser power range is derived from the upper limit of the rated capacity of the welding machine and the qualified range of the process qualification; the welding speed range is derived from the lower limit of the production line cycle time requirement and the statistical results of the penetration depth of qualified samples; the wire feed speed range is derived from the filler wire diameter, the target weld reinforcement height, and the wettability requirements of the weld pool; the oscillation amplitude and oscillation frequency ranges are derived from the weld gap compensation requirements and the target weld width; the shielding gas flow rate range is derived from the oxidation sensitivity corresponding to the material grade and the statistical results of historical porosity defects; the defocusing amount range is derived from the weld type, plate thickness, and laser spot diameter; the above parameter ranges together constitute the safe working window for welding parameters, which represents the executable boundary of the parameters under the conditions of not damaging parts, not exceeding the equipment capacity, and not violating the constraints of the process qualification.

[0041] Furthermore, using the welding parameter safety working window as the initial constraint range, and combining the segmented risk field results, a candidate set of segmented parameters is generated for each weld segment unit; the comprehensive risk value of each weld segment unit is then used as the initial constraint. Based on the four risk categories, the range of each parameter in the welding parameter safety working window is dynamically narrowed or expanded; when When the preset geometric risk threshold is exceeded, the upper limit of the wire feeding speed range and the upper limit of the oscillation amplitude range are expanded to compensate for insufficient filling caused by increased gap and misalignment; when When the preset molten pool risk threshold is exceeded, the upper limit of the laser power range is narrowed to reduce molten pool instability and porosity risk; when When the preset heat accumulation risk threshold is exceeded, the upper limit of the laser power range and the lower limit of the welding speed range are simultaneously narrowed to suppress further heat input accumulation; when When the preset deformation risk threshold is exceeded, the upper limit of the laser power range is narrowed to limit the total local heat input; in one embodiment, the preset geometric risk threshold, preset molten pool risk threshold, preset heat accumulation risk threshold and preset deformation risk threshold are all 0.6.

[0042] The parameter candidate set for each segment is subject to a constraint on the variation of parameters in adjacent segments. Specifically, the actual execution parameters of the previous weld segment unit are used as a reference, and the difference between the candidate parameters of the current weld segment unit and the reference value is limited to no more than a preset variation range. In one embodiment, the difference between the candidate laser power value and the reference value is no more than 10% of the reference value, and the difference between the candidate welding speed value and the reference value is no more than 15% of the reference value. The preset variation range is determined by the welding machine power ramp-up speed, the robot controller response cycle, and stable welding samples during the trial production stage. After the welding parameter safety working window is dynamically narrowed or expanded and the parameter variation constraint for adjacent segments is applied, discrete sampling is performed within each parameter range at a preset step size to generate the parameter candidate combination for the current weld segment unit. The parameter candidate combinations corresponding to all weld segment units together constitute the segment parameter candidate set, which serves as the input for the subsequent physical constraint deep learning model and the constraint range for rolling parameter optimization.

[0043] In this embodiment, the specific content of predicting segment quality includes: After aligning the segmented welding state characteristics, segmented risk field results, and segmented parameter candidate set according to the weld segment unit index, the data is input into a pre-built physical constraint deep learning model. Using the upper limit of heat input, the lower limit of penetration depth, and the upper limit of deformation accumulation as physical constraints, the model predicts the quality consequences of each parameter candidate combination in the segmented parameter candidate set, thus obtaining the segmented quality prediction results. The segmented quality prediction results include predicted penetration depth, predicted weld width, predicted porosity risk value, predicted crack risk value, predicted heat input, predicted deformation accumulation, and predicted cycle time influence.

[0044] Specifically, the physical constraint deep learning model was trained offline before deployment. The training samples were derived from historical welding records accumulated during the trial production phase. These historical welding records included the segmented welding state features corresponding to each weld segment unit, actual execution parameter combinations, post-weld inspection results, and manual defect annotations. A multimodal feature extraction and fusion architecture was adopted. For the coaxial molten pool image sequence in the segmented welding state features, a convolutional neural network was used to extract spatial features, outputting a molten pool morphology feature vector. For the continuous process signals in the segmented welding state features, including temperature peak sequences, cooling rate sequences, voltage fluctuation sequences, and current fluctuation sequences, a time-series network was used to extract temporal features, outputting a process signal feature vector. The comprehensive risk value in the segmented risk field results was then used to... The four types of risk components are concatenated into a risk state vector, which, together with the aforementioned molten pool morphology feature vector and process signal feature vector, is input into the fully connected fusion layer. The current candidate combination of parameters to be evaluated is encoded into a parameter vector and then synchronously input into the fully connected fusion layer. The fully connected fusion layer outputs the predicted melt depth, predicted weld width, predicted porosity risk value, predicted crack risk value, predicted heat input, predicted cumulative deformation, and predicted cycle time effect corresponding to the current candidate combination of parameters. In one embodiment, the convolutional neural network adopts a three-layer convolutional plus two-layer fully connected structure, and the temporal network adopts a two-layer lightweight gated recurrent unit structure.

[0045] Furthermore, if the predicted melt depth output by the model is lower than the lower limit of the target melt depth, the candidate combination of the parameter is marked as infeasible and removed from the candidate set; if the predicted thermal input output by the model exceeds the upper limit of the thermal input corresponding to the material grade, it is also marked as infeasible and removed; if the predicted cumulative deformation output by the model exceeds the upper limit of the allowable deformation corresponding to the assembly tolerance, it is marked as infeasible and removed; the upper limit of the thermal input is derived from the upper limit of the qualified thermal input range corresponding to the material grade and plate thickness in the process evaluation document, and the upper limit of the allowable deformation is calculated by the assembly tolerance according to the material elastic modulus and the weld distribution position.

[0046] After the above hard constraint screening, the parameter candidate combinations that satisfy all physical constraints and their corresponding segmented quality prediction results are retained to form a feasible candidate set, which is then passed to the subsequent rolling parameter optimization step. If the feasible candidate set is empty, the upper limit of heat input and the upper limit of deformation accumulation are each relaxed by 5% and the constraint screening is re-executed. The current segmented unit is marked as a constraint relaxation region in the segmented quality prediction results for subsequent process traceability. In one embodiment, the relaxation ratio is 5%.

[0047] In this embodiment, the specific content of the scrolling parameter optimization includes: Using the feasible candidate set obtained after physical constraint screening and the corresponding segment quality prediction results as input, rolling parameter optimization is performed sequentially on each weld segment unit according to the weld path order. The optimal parameter combination is selected from the feasible candidate set of each weld segment unit to obtain the segment welding parameter sequence. The rolling parameter optimization takes quality deviation, heat input risk, deformation accumulation risk and parameter jump penalty as optimization objectives. After the parameters of the current weld segment unit are determined, they are used as the benchmark parameters for the optimization of the next weld segment unit. The optimization is carried out segment by segment along the weld path until all weld segment units have completed parameter selection. The segment welding parameter sequence uses the weld segment unit index as the key and the corresponding optimal parameter combination as the value, including the laser power, welding speed, wire feed speed, oscillation frequency, oscillation amplitude, defocusing amount and shielding gas flow rate of each segment unit.

[0048] Furthermore, for the i-th weld segment element, the optimal parameter combination is selected from the feasible candidate set according to the following objective function: , in, The optimal parameter combination for the i-th weld segment unit is... Let argmin be the set of feasible candidates for the i-th weld segment element after physical constraint screening, and let argmin represent the set of feasible candidates. The search term is for the combination of parameters within the parentheses that minimizes the objective function. From The combination of candidate parameters traversed in the middle. The quality deviation term is obtained by averaging the normalized values ​​of the absolute values ​​of the differences between the predicted and target weld penetration depths and the predicted and target weld widths. This is the heat input risk item, calculated as the ratio of predicted heat input to the upper limit of heat input; For the risk items of porosity and cracking, the average of the predicted porosity risk value and the predicted cracking risk value is taken; For the cumulative deformation risk item, take the ratio of the predicted cumulative deformation to the upper limit of allowable deformation; The penalty term for parameter jumps between adjacent segments is calculated from the normalized distance of the difference between the current candidate parameter combination and the actual executed parameter combination of the previous weld segment unit. For the first weld segment unit, take... ; to The weighting coefficients are the weighting factors for each optimization objective; in one implementation, the weighting coefficients... , , , , The values ​​are 0.35, 0.20, 0.20, 0.15 and 0.10 respectively.

[0049] After selecting parameters for all weld segment units, the obtained segmented welding parameter sequence is converted into execution instructions recognizable by the welding execution end and sent to the robot controller and welding machine controller. The execution instructions use the start and end coordinates of the path of each weld segment unit as the trigger boundary, and activate the corresponding parameter combination when the welding torch reaches the start coordinate of the corresponding weld segment unit. If the controller does not support continuous parameter updates according to the length of the weld segment unit, then weld segment units with adjacent comprehensive risk values ​​that do not exceed a preset merging threshold are merged into a parameter execution interval to meet the minimum response cycle requirement of the controller. In one embodiment, the preset merging threshold is 0.05. After the transmission is completed, the welding execution end completes the welding of each weld segment unit in sequence according to the segmented welding parameter sequence to obtain the welding execution result.

[0050] like Figure 2 This is a schematic diagram illustrating the effect of a deep learning-based adaptive optimization method for welding process parameters. The horizontal axis lists key performance indicators, and the vertical axis represents the exemplified performance scores, ranging from 0 to 100%. Higher values ​​indicate better performance. The aim is to visually demonstrate the expected improvement of the invention in key capabilities compared to typical prior art.

[0051] Example 2:

[0052] In Example 1, by acquiring the weld path and quality target of the parts to be welded, the weld path is divided into a set of weld segment units; by collecting the pre-weld geometric data and in-weld process data of each weld segment unit, the segment welding state characteristics are obtained; by constructing a local weldability risk field, the segment risk field results are obtained; and under the constraint of the welding parameter safety working window, a set of segment parameter candidates is generated, and then the segment quality prediction results are obtained through a physical constraint deep learning model. Finally, the segment welding parameter sequence is obtained and issued through rolling parameter optimization.

[0053] In this embodiment, the execution boundary merging, parameter switching transition, and execution end response matching process of the segmented welding parameter sequence are defined, so that the segmented welding parameter sequence can be converted into a parameter execution range and parameter gradient segment that can be stably executed by the welding execution end, thereby reducing the risk of molten pool fluctuation, heat input deviation, and robot execution instability caused by excessively fine segmentation or parameter abrupt changes.

[0054] Specifically, after obtaining the segmented welding parameter sequence, the minimum parameter update cycle, welding machine power ramp-up speed, robot speed response cycle, wire feeding mechanism response cycle, and shielding gas flow rate adjustment response cycle of the welding execution end are obtained. The parameter with the largest corresponding path length is determined as the minimum executable segment length. When the cumulative path length of multiple consecutive weld segment units is less than the minimum executable segment length, the multiple consecutive weld segment units are determined as a segment group to be merged.

[0055] Further, a merging determination is performed on the segments to be merged; when the difference in the comprehensive risk value of adjacent weld segment units does not exceed a preset merging threshold, and the difference in parameters between the optimal parameter combinations corresponding to adjacent weld segment units does not exceed the preset change range in step four, the adjacent weld segment units are merged into a parameter execution interval. The parameter combination with the lowest predicted penetration depth and the lowest predicted porosity risk value and predicted crack risk value among the segment quality prediction results of each weld segment unit before merging is taken as the execution parameter of the merged parameter execution interval; in one embodiment, the preset merging threshold is 0.05; when the difference in the comprehensive risk value of adjacent weld segment units exceeds the preset merging threshold, or the difference in parameters between the optimal parameter combinations corresponding to adjacent weld segment units exceeds the preset change range, the independent parameter boundaries of the adjacent weld segment units are retained, and a parameter transition segment is set at the corresponding parameter switching position.

[0056] Furthermore, the parameter transition segment extends forward and backward along the robot's running direction, centered on the boundary between two adjacent parameter execution intervals, by a preset transition length. The preset transition length is calculated by dividing the parameter switching difference with the largest change amplitude in the adjacent parameter execution intervals by the corresponding device response rate, and taking the maximum value among the transition lengths corresponding to each parameter. In one embodiment, the preset transition length is between 3mm and 8mm. Within the parameter transition segment, the laser power, welding speed, wire feed speed, oscillation frequency, oscillation amplitude, defocusing amount, and shielding gas flow rate are smoothly transitioned from the execution parameters of the previous parameter execution interval to the execution parameters of the next parameter execution interval using a linear interpolation method, so that the welding execution end completes the parameter switching according to its own response capability. After the above merging process and transition segment setting, a smooth segmented welding parameter sequence is obtained. This smooth segmented welding parameter sequence replaces the segmented welding parameter sequence in Embodiment 1 and is sent to the welding execution end to obtain the welding execution result.

[0057] The embodiments of the present invention described above are subject to modification and change of method by those skilled in the art without departing from the embodiments and broader aspects of the present invention. The appended claims are intended to include all such modifications and changes of method that do not depart from the present invention.

Claims

1. A deep learning-based welding process parameter self-adaptive optimization method, characterized in that, include: Obtain the weld path and quality target of the parts to be welded, and divide the weld path into a set of weld segment units along the robot's running direction; Collect pre-welding geometric data and in-welding process data for each weld segment unit to obtain segment welding state characteristics; Construct the local weldability risk field for each weld segment unit to obtain the segment risk field results; A pre-constructed welding parameter safety working window is used as a constraint, and a segmented parameter candidate set is generated in combination with the segmented risk field results. The segmented welding state characteristics, segmented risk field results, and segmented parameter candidate set are input into a pre-built physical constraint deep learning model to obtain the segmented quality prediction results of the quality target. Based on the segmented quality prediction results, rolling parameter optimization is performed on each weld segment unit to obtain the segmented welding parameter sequence, which is then sent to the welding execution end.

2. The self-adaptive optimization method of claim 1, wherein, The weld path is divided into a set of weld segment units along the robot's running direction, including: Taking the weld path coordinate sequence corresponding to the weld path as the object, multiple weld segmentation units are divided according to the robot's running direction; the path length corresponding to a single frame is determined based on the robot's travel speed and sensor sampling frequency, and the minimum sampling segment length is determined in combination with the lower limit of the frame number; a segment length range including a lower limit and an upper limit of the segment length is constructed based on the minimum sampling segment length; the straight stable region and the corner region are determined by the angle between the direction vectors of adjacent path points and a preset corner threshold; the straight stable region is segmented using the upper limit of the segment length, and the corner region is segmented using the lower limit of the segment length; a preset local region is segmented using the lower limit of the segment length to obtain the set of weld segmentation units; wherein, the preset local region includes the weld path end region, the fixture obstruction edge region, and the historical defect high-incidence region.

3. The self-adaptive optimization method of claim 1, wherein, Collect pre-welding geometric data and in-welding process data for each weld segment to obtain segment welding state characteristics, including: Pre-welding scanning is performed on the weld area of ​​each weld segment unit to obtain pre-welding geometric data. The pre-welding geometric data is then compared with the theoretical weld model to obtain pre-welding geometric deviation characteristics. During the welding process, welding process data of each weld segment unit is collected, and feature extraction is performed on the welding process data to obtain welding process stability characteristics. The pre-welding geometric deviation characteristics and welding process stability characteristics are aligned and fused according to the weld segment units to obtain the segmented welding state characteristics.

4. The self-adaptive optimization method of claim 1, wherein, Construct the local weldability risk field for each weld segment unit to obtain the segment risk field results, including: The heat accumulation state along the weld is obtained by accumulating the temperature peak and cooling rate of the completed weld segment units along the path direction. Using the segment welding state characteristics and the heat accumulation state along the weld as fusion inputs, a local weldability risk field is constructed for each weld segment unit. The local weldability risk field includes geometric disturbance risk value, molten pool instability risk value, heat accumulation risk value, and deformation accumulation risk value. The above four types of risk values ​​are weighted and merged to obtain the comprehensive risk value of each weld segment unit. The comprehensive risk value and the four types of risk values ​​together constitute the segment risk field result.

5. The self-adaptive optimization method of claim 4, wherein, Geometric disturbance risk value, molten pool instability risk value, heat accumulation risk value, and deformation accumulation risk value include: The geometric disturbance risk value is determined by comparing the gap value, misalignment amount, and weld center offset of each weld segment unit with the corresponding upper limit of the threshold. The molten pool instability risk value is determined by normalizing the molten pool size deviation, molten pool brightness fluctuation, light intensity fluctuation, acoustic emission mutation, voltage fluctuation, and current fluctuation. The heat accumulation risk value is determined by the heat accumulation state value along the weld, the temperature peak value of adjacent welded segment units, the cooling rate attenuation, and the heat dissipation effect corresponding to the adjacent distance of the fixture. The deformation accumulation risk value is determined by weighting the position of the current weld segment unit along the weld path, the heat input accumulation of the previous weld segment unit, and the assembly tolerance sensitive area identifier.

6. The self-adaptive optimization method of claim 1, wherein, Physically constrained deep learning models, including: A multimodal feature extraction and fusion architecture is adopted to extract the molten pool morphology feature vector from the coaxial molten pool image sequence in the segmented welding state features, extract the process signal feature vector from the continuous process signal in the segmented welding state features, concatenate the comprehensive risk value and four types of risk components in the segmented risk field results into a risk state vector, and encode the current parameter candidate combination to be evaluated into a parameter vector; input the molten pool morphology feature vector, the process signal feature vector, the risk state vector and the parameter vector into a fully connected fusion layer, and output the segmented quality prediction result corresponding to the current parameter candidate combination.

7. The self-adaptive optimization method of claim 1, wherein, After obtaining the segmented quality prediction results of the quality target, the method further includes: Using the lower limit of the penetration depth corresponding to the quality target, the upper limit of the heat input corresponding to the component to be welded, and the upper limit of the allowable deformation corresponding to the assembly tolerance in the quality target as screening constraints, the feasibility of the parameter candidate combinations in the segment parameter candidate set is screened; when the segment quality prediction result does not meet any of the screening constraints of the lower limit of penetration depth, upper limit of heat input, or upper limit of allowable deformation, the corresponding parameter candidate combination is removed from the segment parameter candidate set; retaining the parameter candidate combinations that meet all screening constraints, the feasible candidate set of each weld segment unit is obtained.

8. The self-adaptive optimization method of claim 1, wherein, Rolling parameter optimization was performed on each weld segment unit to obtain a segmented welding parameter sequence, including: Rolling parameter optimization is performed sequentially on each weld segment unit according to the weld path order. The optimal parameter combination is selected from the feasible candidate set of each weld segment unit to obtain the segmented welding parameter sequence. The rolling parameter optimization takes quality deviation, heat input risk, deformation accumulation risk and parameter jump penalty as optimization objectives. After the parameters of the current weld segment unit are determined, they are used as the benchmark parameters for the optimization of the next weld segment unit. The optimization is carried out segment by segment along the weld path until the parameters of all weld segment units have been selected. The segmented welding parameter sequence is converted into execution instructions that can be recognized by the welding execution end and sent to the robot controller and welding machine controller.

9. The self-adaptive optimization method of claim 8, wherein, The optimal parameter combination is selected from the feasible candidate set of each weld segment unit, including: The deviation of the segmented quality prediction result from the quality target is used as the quality deviation term; the relationship between the heat input prediction information and the heat input upper limit in the segmented quality prediction result is used as the heat input risk term; the relationship between the deformation accumulation prediction information and the allowable deformation upper limit in the segmented quality prediction result is used as the deformation accumulation risk term; and the change between the current candidate parameter combination and the actual executed parameter combination of the previous weld segment unit is used as the parameter jump penalty term. Based on the weighted result of the quality deviation term, heat input risk term, deformation accumulation risk term, and parameter jump penalty term, the optimal parameter combination of the current weld segment unit is determined.

10. The self-adaptive optimization method of claim 1, wherein, After obtaining the segmented welding parameter sequence, it also includes: The minimum executable segment length is determined based on the response capability of the welding execution end. For continuous weld segment units with a cumulative path length less than the minimum executable segment length, a preset merging threshold is used as the judgment boundary. Adjacent weld segment units whose comprehensive risk value difference and parameter difference do not exceed the corresponding judgment boundary are merged into a parameter execution interval. For adjacent weld segment units whose comprehensive risk value difference or parameter difference exceeds the corresponding judgment boundary, a parameter transition segment is set at the junction. After the above processing, a smooth segmented welding parameter sequence is obtained, and the smooth segmented welding parameter sequence is used to replace the segmented welding parameter sequence and sent to the welding execution end.

Images