Precise resistance spot welding spot position calibration system based on machine vision

Abstract

The invention relates to the technical field of precision welding machine vision calibration, and discloses a precision resistance spot welding spot position calibration system based on machine vision. According to the system, a workpiece image is acquired and segmented through an image acquisition and processing module; the feature analysis and coordinate generation module generates initial positioning coordinates in each calibration area according to the geometrical relationship between the microstructure of the surface of the workpiece and a preset theoretical contour, and the initial positioning coordinates are summarized into a table by the position mapping integration module. The electric signal acquisition and analysis module acquires a welding test current waveform and extracts a characteristic frequency band energy distribution spectrum of the welding test current waveform. The spectrum comparison and calibration judgment module judges whether coordinate mapping needs to be calibrated or not by calculating the deviation degree of the energy spectrum and the reference spectrum. According to the system, the positioning robustness under a complex working condition is enhanced by using micro-topography geometric analysis, and an intelligent calibration decision based on process state feedback is realized through real-time analysis of an electric signal spectrum.

Description

Technical Field

[0001] This invention relates to the field of precision welding machine vision calibration technology, specifically a precision resistance spot welding weld point position calibration system based on machine vision. Background Technology

[0002] In precision resistance spot welding, accurate positioning of the weld points is the primary prerequisite for ensuring welding quality and product consistency. Existing technologies generally employ machine vision-based positioning methods, which acquire images by scanning the workpiece with a camera and use algorithms such as template matching or edge detection to identify feature locations, thereby generating a coordinate mapping table for the weld points. These methods perform well under ideal conditions where the workpiece surface is clean and has a regular and consistent shape.

[0003] In real-world production environments, especially when dealing with precision workpieces with complex micro-textures, machining marks, or slight deformations, conventional vision positioning technology exhibits limitations. Its positioning accuracy heavily relies on the clarity and consistency of macroscopic image features. When there are microscopic morphological differences on the workpiece surface, matching methods based on fixed templates or simple contours are prone to positioning errors, or even recognition failures. Existing systems often rely on fixed time periods, welding cycles, or manual intervention to trigger the calibration of weld point coordinate mapping tables. This approach lacks awareness and response to the real-time status of the welding process, failing to proactively and promptly trigger calibration when implicit drifts in welding process parameters occur but are not yet apparent in macroscopic defects, resulting in a delayed quality risk assessment.

[0004] A solution is needed that can more accurately adapt to changes in the micro-morphology of the workpiece for positioning, and can intelligently determine calibration requirements based on the real-time physical state of the welding process, in order to improve welding accuracy and system autonomy under complex working conditions. Summary of the Invention

[0005] The purpose of this invention is to provide a precision resistance spot welding solder joint position calibration system based on machine vision, so as to solve the problems mentioned in the background art.

[0006] To achieve the above objectives, the present invention provides a precision resistance spot welding joint position calibration system based on machine vision, the system comprising:

[0007] The image acquisition and processing module is used to acquire the original image set generated by the vision camera scanning the surface of the workpiece on the welding platform, and to separate the local image subsets corresponding to several preset calibration areas from the original image set.

[0008] The feature parsing and coordinate generation module is used to analyze the geometric relationship between the micro-morphological features of the workpiece surface and the preset theoretical contour of the weld point for each local image subset, and generate an initial positioning coordinate in the corresponding calibration area based on the geometric relationship.

[0009] The position mapping integration module is used to integrate the initial positioning coordinates of all calibration areas to form an initial position mapping table describing the current weld point distribution of the workpiece.

[0010] The electrical signal acquisition and analysis module is used to acquire the real-time electrical signal waveform of the welding power supply device under a preset test current, and extract the energy distribution spectrum in the characteristic frequency band from the real-time electrical signal waveform;

[0011] The spectral comparison and calibration judgment module is used to compare the energy distribution spectrum with the reference energy distribution spectrum under standard operating conditions, calculate the spectral deviation between the two in multiple characteristic dimensions, and determine whether the coordinate data in the initial position mapping table needs to be calibrated and corrected based on the magnitude of the spectral deviation.

[0012] Preferably, the step of generating an initial positioning coordinate within the corresponding calibration area based on the geometric relationship includes:

[0013] Multi-level decomposition of local image subsets yields image pyramid levels that reflect surface features at different scales;

[0014] In each image pyramid level, the ridges, edge intersections, and curvature abrupt change points on the workpiece surface are located, and these points are marked as candidate feature points.

[0015] Spatial clustering and matching are performed on all candidate feature points in the image pyramid levels to remove isolated points and form a stable feature point group.

[0016] Calculate the spatial centroid of the feature point group, and project the spatial centroid along the normal vector direction of the workpiece surface onto the working plane of the welding platform. Use the coordinates of the projected point as the initial positioning coordinates of the calibration area.

[0017] Preferably, the step of determining whether the coordinate data in the initial position mapping table needs to be calibrated and corrected includes:

[0018] Set a trigger threshold and an ignore threshold for the spectral deviation, where the trigger threshold is greater than the ignore threshold;

[0019] When the spectral deviation is greater than or equal to the trigger threshold, it is determined that a comprehensive calibration and correction process must be performed on the initial position mapping table.

[0020] When the spectral deviation is less than the trigger threshold but greater than the ignore threshold, the determination only needs to selectively calibrate and correct the local coordinates associated with the specific dimension that contributes the most to the spectral deviation in the initial position mapping table.

[0021] When the spectral deviation is less than or equal to the ignore threshold, the current initial position mapping table is deemed valid and no calibration correction is required.

[0022] Preferably, the comprehensive calibration and correction process for the initial position mapping table includes:

[0023] After determining that a comprehensive calibration and correction is needed, the vision camera is controlled to re-acquire images of each calibration area at a higher resolution to obtain a fine image set.

[0024] In a fine image set, the deformation of the theoretical weld point profile caused by workpiece assembly stress, thermal deformation or platform vibration is identified and quantified, including profile translation, rotation angle and local distortion.

[0025] The contour translation, rotation angle, and local distortion are combined into a set of spatial transformation parameters;

[0026] The spatial transformation parameter set is applied to each initial positioning coordinate in the initial position mapping table. Through affine transformation calculation, the updated calibration positioning coordinates are output, thereby generating the corrected position mapping table.

[0027] Preferably, selective calibration and correction of the local coordinates associated with the specific dimension that contributes the most to the spectral deviation in the initial position mapping table includes:

[0028] Identify the key feature dimensions that cause the spectral deviation to be in the middle range, and establish a correlation between the key feature dimensions and specific mechanical axes of the welding platform or specific electrical parameters of the welding power supply device.

[0029] Only the initial positioning coordinates of the calibration areas in the initial position mapping table that are covered by the specific mechanical axis movement range or directly affected by the specific electrical parameters are re-evaluated.

[0030] During the reassessment, the initial positioning coordinates of the target are interpolated and compensated by combining the coordinate data of successful weld points under similar working conditions in the historical welding logs to obtain optimized coordinates.

[0031] The optimized coordinates replace the original coordinates in the initial position mapping table, while the unaffected coordinates are retained, thereby generating a locally corrected position mapping table.

[0032] Preferably, after generating the corrected location mapping table, the process also includes an update step for the solder joint location prediction model:

[0033] The spatial transformation parameter set generated this time, the characteristic frequency band energy distribution spectrum of the real-time electrical signal waveform, and the finally generated corrected position mapping table are used together as a data sample;

[0034] The data samples are added to the system's historical calibration database, and the historical calibration database is used to incrementally train the built-in solder joint location prediction model.

[0035] The solder joint position prediction model is used to pre-calibrate the initial position mapping table based on the acquired preliminary electrical signal waveform characteristics during the next system startup.

[0036] Preferably, when performing interpolation compensation calculations on the initial positioning coordinates of the target, an environmental parameter fusion step is also included:

[0037] The current temperature field distribution data and vibration spectrum data of the welding platform are obtained by additional environmental sensors;

[0038] Retrieve historical successful welding records from historical welding logs that match the current temperature field distribution data and vibration spectrum data;

[0039] Extract the weld point coordinate compensation vector from the historical successful welding records, and then superimpose the weld point coordinate compensation vector with the preliminary optimized coordinates obtained by interpolation.

[0040] The result of vector superposition is used as the final optimized coordinates to replace the original coordinates in the initial position mapping table.

[0041] Preferably, the incremental training steps of the solder joint location prediction model specifically include:

[0042] Principal component analysis was performed on all data samples in the historical calibration database to extract several core feature factors that affect the deviation of the solder joint position.

[0043] Based on core feature factors, a nonlinear mapping network is constructed to describe the deviation from electrical signal features and visual features to actual coordinates.

[0044] When new data samples are added, a rolling time window is used to iteratively fine-tune the connection weights of the nonlinear mapping network with a fixed number of the latest data samples.

[0045] After each iteration of fine-tuning, the prediction accuracy of the nonlinear mapping network for coordinate deviations in recent data samples is verified, and the learning rate is dynamically adjusted based on the verification results.

[0046] Preferably, when retrieving historical successful welding records from historical welding logs that match the current temperature field distribution data and vibration spectrum data, a multi-level matching strategy is adopted:

[0047] The first-level matching uses the overall gradient direction of the temperature field distribution data as the primary matching key;

[0048] The second-level matching is performed in the result set of the first-level matching, using the main peak frequency of the vibration spectrum data as the secondary matching key for filtering.

[0049] The third-level matching is finalized based on the material type and thickness of the current welded workpiece, using the result set of the second-level matching as the specific matching key.

[0050] The final output of the historical successful welding records is a set of records that simultaneously meet the three-level matching bond conditions.

[0051] Preferably, after generating the initial position mapping table, the system performs a virtual welding simulation to verify the effectiveness of the mapping table:

[0052] Based on the initial or revised position mapping table, the welding torch of the automatic welding machine is driven to perform current-free path planning and motion simulation in virtual space.

[0053] During the motion simulation, the system continuously monitors whether there is interference between the virtual path of the welding torch and the 3D model of the workpiece, and whether the posture of the welding torch exceeds its physical motion limits.

[0054] If interference or exceeding limits is detected, the coordinates of the interference area or the welding torch attitude parameters that caused the exceeding limits are marked as anomalies and fed back to the coordinate generation step. This guides the system to recalculate the initial positioning coordinates of the relevant calibration area and generate a new position mapping table after eliminating the anomalies.

[0055] Beneficial effects include:

[0056] The feature parsing and coordinate generation module generates initial positioning coordinates by analyzing the geometric relationship between the microscopic morphological features of the workpiece surface and the preset theoretical contour of the weld point. This technology does not rely on the strict consistency of macroscopic surface features, but rather on positioning by analyzing the spatial relationship between inherent features such as microscopic textures and undulations and the ideal model. This enables the system to overcome the instability of macroscopic image features caused by machining marks, oxidation, or slight plastic deformation on the workpiece surface, elevating the robustness of positioning from the image pixel level to the geometric model adaptation level. Thus, even when there are microscopic differences in the workpiece surface condition, it can still obtain coordinate data with high theoretical consistency and stability.

[0057] The electrical signal acquisition and analysis module, along with the spectrum comparison and calibration judgment module, constitutes an independent process feedback loop. By acquiring and analyzing real-time electrical signal waveforms under welding test current, the energy distribution spectrum of characteristic frequency bands is extracted and compared with a standard spectrum in multiple dimensions to calculate the spectral deviation. This technology uses the frequency domain characteristics of electrical signals reflecting the core physical processes of welding as an intrinsic basis for evaluating process stability. Based on the quantified spectral deviation, it determines whether coordinate calibration should be initiated, realizing a shift from "timed or manual calibration" to "calibration based on the necessity of process conditions." This effectively captures hidden process drift caused by electrode wear, power fluctuations, etc., proactively triggering maintenance before detectable macroscopic degradation of welding quality occurs, improving the proactiveness of quality control and the system's autonomous decision-making capability. Attached Figure Description

[0058] Figure 1This is a schematic diagram illustrating the working principle of the precision resistance spot welding weld point position calibration system based on machine vision as described in this invention.

[0059] Figure 2 A flowchart for generating initial positioning coordinates;

[0060] Figure 3 A flowchart for comprehensive calibration correction;

[0061] Figure 4 A comparison chart showing the number of successful retrieval records for different workpiece materials under a multi-level matching strategy;

[0062] Figure 5 This is a graph showing the relationship between the welding torch attitude angle change and the over-limit detection curve in a virtual welding simulation. Detailed Implementation

[0063] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0064] Please see Figure 1This invention provides a precision resistance spot welding weld point position calibration system based on machine vision. The system includes: an image acquisition and processing module that first controls a vision camera to scan the surface of a workpiece placed on a welding platform, acquiring an original image set covering the entire area to be welded; this module, based on the spatial coordinate range of several preset calibration areas, cropping and separating local image subsets corresponding to each calibration area from the original image set; a feature analysis and coordinate generation module that sequentially processes each local image subset, analyzing the microscopic morphological features of the workpiece surface, such as texture direction, pits or protrusions, and their geometric positional relationship with the theoretical weld point contour pre-input into the system, such as relative distance and angle; and calculating an initial positioning coordinate in the image space of each calibration area based on this geometric relationship; and a position mapping integration module that arranges and integrates the initial positioning coordinates from all calibration areas according to their corresponding physical positions, thereby forming an initial position mapping table describing the current position status of all predetermined weld points on the workpiece. The electrical signal acquisition and analysis module acquires real-time electrical signal waveforms flowing through the workpiece and electrodes under a preset test current from the welding power supply device. It performs time-frequency analysis on these waveforms and extracts the energy distribution spectrum within the characteristic frequency band reflecting the contact state. The spectrum comparison and calibration judgment module compares the real-time acquired energy distribution spectrum with a reference energy distribution spectrum stored in the system and measured under standard operating conditions. It calculates the comprehensive difference value between the two in multiple predefined characteristic dimensions, i.e., the spectrum deviation. Based on the calculated spectrum deviation value, the system automatically determines whether the initial position mapping table generated by the vision system needs calibration correction and the required correction degree.

[0065] Example 1: See Figure 2In the feature parsing and coordinate generation module, the step of generating initial positioning coordinates within the corresponding calibration area based on geometric relationships proceeds as follows: First, Gaussian pyramid construction or multi-resolution wavelet transform is performed on the acquired local image subset to complete multi-level decomposition, obtaining image pyramid levels reflecting surface features at different scales. In each image pyramid level, edge detection and ridge tracking algorithms are used to locate the ridges, edge intersections, and curvature abrupt change points on the workpiece surface, and these points are marked as candidate feature points for that level. Subsequently, a spatial clustering algorithm, such as DBSCAN based on Euclidean distance, is used to perform cross-level matching and aggregation of candidate feature points in all image pyramid levels, eliminating isolated noise points in spatial distribution to form a stable and consistent feature point group across multiple scales. The average spatial coordinates of all points in this feature point group are calculated to obtain the spatial centroid, which is then projected onto the two-dimensional working plane coordinate system of the welding platform along the direction of the workpiece surface normal vector determined by the surface normal estimation operator. The coordinates of this projected point are then determined as the initial positioning coordinates of the calibration area. The spectral comparison and calibration judgment module determines whether calibration correction is needed for the coordinate data in the initial position mapping table based on the spectral deviation: the system presets a trigger threshold and an ignore threshold for the spectral deviation, with the trigger threshold being greater than the ignore threshold. When the calculated spectral deviation is greater than or equal to the trigger threshold, the system determines that a full calibration correction process must be performed on the initial position mapping table. When the spectral deviation is less than the trigger threshold but greater than the ignore threshold, the system determines that selective calibration correction is only needed for the local coordinates associated with the specific dimensions in the initial position mapping table that contribute the most to the overall spectral deviation. When the spectral deviation is less than or equal to the ignore threshold, the system determines that the current initial position mapping table is valid and no calibration correction operation is required.

[0066] In practical implementation, the image acquisition and processing module obtains a set of original images of a specific workpiece on the welding platform and then performs image segmentation based on a preset calibration region. The feature parsing and coordinate generation module receives the segmented local image subsets and performs a Gaussian pyramid construction operation on these subsets. Through continuous Gaussian smoothing and downsampling, an image pyramid hierarchy reflecting surface features at different scales is obtained. The image pyramid hierarchy contains multiple image layers ranging from high resolution to low resolution. In each image pyramid hierarchy, a ridge detection algorithm based on the Hessian matrix is ​​used to locate the microscopic ridges on the workpiece surface. Simultaneously, the Canny edge detection operator is used to locate edge intersections, and curvature abrupt change points are located using an image curvature map. These located points are marked as candidate feature points for the current image pyramid hierarchy. It can be understood that all candidate feature points in all image pyramid hierarchy levels are aggregated into a common spatial coordinate system.

[0067] In practical implementation, a density-based spatial clustering method is used for all collected candidate feature points. Clustering analysis is performed using a preset neighborhood radius and a minimum point count threshold. Spatially adjacent candidate feature points are grouped into the same feature point cluster, while isolated points that do not meet the density requirements are removed, thus forming stable feature point clusters. Each feature point cluster contains multiple spatially clustered candidate feature points. Optionally, the spatial centroid of the feature point cluster is calculated. The spatial centroid is the arithmetic mean of the coordinates of all points in the feature point cluster in the image coordinate system. The spatial centroid coordinates are calculated using the following formula:

[0068]

[0069] Where: character G represents the calculated spatial centroid coordinates, character N represents the total number of points in the feature point group, and character X... i and Y i Let x and y represent the x and y coordinates of the i-th feature point, respectively.

[0070] In some embodiments, the calculated spatial centroid coordinates are projected onto the coordinate system of the two-dimensional working plane of the welding platform along the direction of the workpiece surface normal vector obtained by the surface normal estimation operator from a local image subset. The projection transformation process involves the rotation and translation of the coordinate system, converting the image coordinates into world coordinates. The final projection point coordinates are determined as the initial positioning coordinates of the corresponding calibration area. It can be understood that the electrical signal acquisition and analysis module synchronously acquires the real-time electrical signal waveform of the welding power supply device under a preset test current and extracts the energy distribution spectrum within the characteristic frequency band through a fast Fourier transform. The spectrum comparison and calibration judgment module compares the real-time energy distribution spectrum with the stored reference energy distribution spectrum and calculates the spectral deviation in multiple characteristic dimensions. The spectral deviation is a comprehensive scalar value used to quantify the overall difference between the two energy distribution spectra.

[0071] In practical implementation, the spectral comparison and calibration judgment module internally presets a trigger threshold and an ignore threshold, with the trigger threshold set to a value greater than the ignore threshold. When the calculated spectral deviation is greater than or equal to the trigger threshold, the spectral comparison and calibration judgment module determines that a comprehensive calibration correction process must be performed on the initial position mapping table. When the calculated spectral deviation is less than the trigger threshold but greater than the ignore threshold, the spectral comparison and calibration judgment module determines that only selective calibration correction is needed for the local coordinates associated with the specific dimension in the initial position mapping table that contributes the most to the spectral deviation. Optionally, when the calculated spectral deviation is less than or equal to the ignore threshold, the spectral comparison and calibration judgment module determines that the current initial position mapping table is valid, and no calibration correction operation needs to be initiated.

[0072] Example 2: See Figure 3After determining that a comprehensive calibration correction is needed, the system executes a comprehensive calibration correction process. The position mapping integration module controls the vision camera to re-acquire images of each calibration area with higher optical resolution, resulting in a set of more detailed, refined images. In this refined image set, image registration and deformation field analysis algorithms are used to identify and quantify the deformation of the theoretical weld point profile relative to the standard template caused by workpiece assembly stress, thermal deformation, or platform vibration. These deformation variables include the overall translation of the profile, the rotation angle around the normal, and the local nonlinear distortion of the profile itself. The quantified profile translation, rotation angle, and local distortion variables are combined into a complete set of spatial transformation parameters. This set of spatial transformation parameters is applied to each initial positioning coordinate in the initial position mapping table. Through affine transformation calculations, each initial coordinate is transformed to the corrected position, thereby outputting the updated calibration positioning coordinates and generating the final corrected position mapping table. After generating the corrected position mapping table, the system executes the update step of the weld point position prediction model. The system encapsulates the spatial transformation parameter set generated during this calibration process, the characteristic frequency band energy distribution spectrum of the real-time acquired electrical signal waveform, and the finally generated corrected position mapping table into a timestamped data sample. This data sample is added to the system's historical calibration database, and the system uses the updated historical calibration database to incrementally train the built-in weld point position prediction model. This weld point position prediction model is used to pre-calibrate and adjust the initial position mapping table generated by the vision system based on the preliminary electrical signal waveform characteristics acquired in real time when the system starts up again or processes a new workpiece.

[0073] In practical implementation, when the spectral comparison and calibration judgment module determines that the spectral deviation is greater than or equal to the trigger threshold, the system initiates a comprehensive calibration and correction process for the initial position mapping table. The position mapping integration module sends instructions to the image acquisition and processing module to control the vision camera to re-acquire images of each calibration area at a higher optical resolution, resulting in a fine image set containing richer surface details. In some embodiments, the vision camera acquires the fine image set by adjusting the lens focal length and increasing the single-point illumination intensity; the resolution of the fine image set is higher than that of the initially acquired original image set. In the acquired fine image set, the system uses a feature point-based image registration algorithm to compare the actual contours of the weld points in the fine image set with the pre-stored theoretical weld point contour standard template in the system. By calculating the affine transformation matrix of the feature points between the two images, the system identifies and quantifies the contour deformation caused by workpiece assembly stress, thermal deformation, or platform vibration. The quantified deformation parameters include the overall translation of the theoretical weld point contour in the image plane, the rotation angle of the theoretical weld point contour about its normal direction, and the local nonlinear distortion of the edge of the theoretical weld point contour. It is understandable that local nonlinear distortions are obtained by calculating the B-spline deformation field between the corresponding point sets of the standard template profile and the actual profile.

[0074] Optionally, the system combines the contour translation, rotation angle, and local distortions into a spatial transformation parameter set. This spatial transformation parameter set, as a data structure, contains the translation vector, rotation matrix, and coordinates and displacement vectors of the distortion field control points. The position mapping integration module applies this spatial transformation parameter set to each initial positioning coordinate in the initial position mapping table. For each initial positioning coordinate P in the initial position mapping table... init (x, y) is calculated using the following formula through an affine transformation, outputting the updated calibration and positioning coordinates P. corr (u,v):

[0075]

[0076] Where: characters Indicates the initial positioning coordinates, character This represents the calculated updated calibration positioning coordinates. The character M represents the 3x3 homogeneous affine transformation matrix consisting of translation and rotation angles, and the character D(x,y) represents the two-dimensional displacement compensation vector at coordinates (x,y) determined by local nonlinear distortions. The calculation process iterates through all coordinates in the initial position mapping table to generate the corrected position mapping table.

[0077] In some embodiments, after generating the corrected position mapping table, the system performs an update step for the solder joint position prediction model. The system encapsulates the set of spatial transformation parameters generated during the calibration process, the characteristic frequency band energy distribution spectrum of the real-time electrical signal waveform acquired by the electrical signal acquisition and analysis module, and the finally generated corrected position mapping table into a data sample with a timestamp and operating condition label. This data sample is added to the system's historical calibration database, which is a relational database or time-series database used to store data from previous calibration processes. In a specific implementation, the system uses the updated historical calibration database to incrementally train the built-in solder joint position prediction model. The solder joint position prediction model is a neural network model. The incremental training process uses the backpropagation algorithm, combining newly added data samples with some recent samples from the historical calibration database as the training set to fine-tune the connection weights of the neural network model. Optionally, when the system is started again or a new workpiece is processed, the electrical signal acquisition and analysis module acquires preliminary electrical signal waveform features during the welding test phase. Based on the input preliminary electrical signal waveform features, the weld point position prediction model outputs a pre-calibration offset to pre-calibrate and adjust the initial position mapping table that will be generated by the image acquisition and processing module and the feature analysis and coordinate generation module.

[0078] Example 3: When the spectral deviation is between the trigger threshold and the ignore threshold, the system performs selective calibration and correction on the local coordinates associated with the specific dimension that contributes the most to the spectral deviation in the initial position mapping table. The system first analyzes the energy distribution spectrum, identifies the key characteristic dimension that causes the spectral deviation to be in this intermediate range, such as energy anomalies in a specific frequency band, and establishes an association mapping between this key characteristic dimension and the motion error of a specific mechanical axis of the welding platform or specific electrical parameters of the welding power supply device, such as fluctuations in the current rise time. The system only re-evaluates the initial positioning coordinates of the corresponding areas in the initial position mapping table that are covered by the motion range of the specific mechanical axis or directly affected by the specific electrical parameter. During re-evaluation, the system accesses historical welding logs, retrieves the coordinate data of successful weld points under similar working conditions, combines them with the coordinates of the current target area, and uses an interpolation algorithm to perform compensation calculations to obtain an optimized coordinate. This optimized coordinate is used to replace the original coordinates in the initial position mapping table, while the remaining coordinates unaffected by the associated dimension are retained, thus generating a locally corrected position mapping table. During the interpolation compensation calculation of the target initial positioning coordinates, the system also includes an environmental parameter fusion step. The system acquires the current temperature field distribution data and vibration spectrum data of the welding platform through additional temperature and vibration sensors. It retrieves historical successful welding records from the historical welding log that match the current temperature field distribution data and vibration spectrum data. The system extracts the applied weld point coordinate compensation vector from the successfully matched records and superimposes this historical compensation vector with the preliminary optimized coordinates calculated through interpolation under the current operating conditions. The result of this vector superposition is used as the final optimized coordinates, taking into account the influence of environmental factors, to replace the original coordinates in the initial position mapping table.

[0079] In specific implementation, when the spectrum comparison and calibration judgment module determines that the spectrum deviation is less than the trigger threshold but greater than the ignore threshold, the system performs selective calibration correction on the local coordinates associated with the specific dimension that contributes the most to the spectrum deviation in the initial position mapping table. The spectrum comparison and calibration judgment module analyzes the energy distribution spectrum and identifies the key feature dimension that causes the spectrum deviation to be in the middle range. The key feature dimension is the frequency band or energy peak in the energy distribution spectrum that differs most significantly from the standard reference spectrum. The system establishes an association mapping between the identified key feature dimension and the motion error range of a specific mechanical axis of the welding platform, or an association mapping between it and a specific electrical parameter of the welding power supply device, such as the current pulse rise time. In some embodiments, based on the association mapping results, the system re-evaluates only the initial positioning coordinates of the calibration area physically covered by the motion range of the specific mechanical axis in the initial position mapping table, or only the initial positioning coordinates of the calibration area directly affected by the specific electrical parameter. It is understood that the re-evaluation process requires accessing the historical welding log and retrieving historical successful welding records with similar welding current parameters, workpiece clamping posture, and current working conditions from the historical welding log. The system extracts the coordinate data of successful weld points from historical successful welding records, compares this data with the initial positioning coordinates to be re-evaluated, and performs compensation calculations using linear interpolation or spline interpolation algorithms to obtain optimized coordinates. The interpolation compensation calculation is performed using the following formula:

[0080] Q opt =Q init +α(Q hist_near -Q hist_far )

[0081] Among them: character Q opt The Q character represents the calculated optimized coordinates. init This represents the initial positioning coordinates to be re-evaluated in the initial position mapping table, represented by the character Q. hist_near and Q hist_far These represent the coordinates of the nearest and second nearest successful weld points retrieved from historical welding logs in the feature space relative to the current coordinates, respectively. The character α represents a coordinate based on the current coordinates and Q. hist_near and Q hist_far The weighting coefficients for calculating the distance relationships between them are used. Optimized coordinates are used to replace the original initial positioning coordinates in the initial position mapping table, while the remaining coordinates that are not affected by the key feature dimensions are retained, thereby generating a locally corrected position mapping table.

[0082] In practical implementation, when performing interpolation compensation calculations for the initial target positioning coordinates, the system simultaneously executes an environmental parameter fusion step. The system acquires current temperature field distribution data and vibration spectrum data of the welding platform through additional environmental sensors. The temperature field distribution data is collected by a thermocouple array deployed at multiple temperature measurement points on the welding platform, and the vibration spectrum data is collected by an accelerometer mounted on the platform base. The system retrieves historical successful welding records from the historical welding logs that match the current temperature field distribution data and vibration spectrum data. Optionally, the matching process is based on the similarity between the temperature gradient and the dominant vibration frequency.

[0083] From the successfully matched historical welding records, the system extracts the weld point coordinate compensation vector that was applied and recorded in the log at the time of the historical successful welding. The extracted historical weld point coordinate compensation vector is then superimposed with the preliminary optimized coordinates obtained through interpolation. The result of this vector superposition serves as the final optimized coordinates, incorporating current environmental parameters, and is used to replace the original coordinates in the initial position mapping table. In some embodiments, the system appends the associated mapping relationships, interpolation weight coefficients, and environmental parameter data generated during this selective calibration correction process as a new log entry to the historical welding log for future retrieval.

[0084] Example 4: The incremental training steps of the weld point location prediction model are as follows. The system performs principal component analysis on all data samples stored in the historical calibration database, extracting several core feature factors affecting weld point location deviation from high-dimensional electrical signal features and visual deformation features. Based on these core feature factors, the system constructs a nonlinear mapping network describing the complex relationship between electrical signal features, visual features, and actual coordinate deviation. This network can be a deep neural network or a support vector regression machine. When new data samples are added to the historical calibration database, the system uses a rolling time window to select a fixed number of the latest data samples as the training set, iteratively fine-tuning the connection weights of the nonlinear mapping network. After each iteration of fine-tuning, the system uses a recent independent validation sample set to verify the prediction accuracy of the nonlinear mapping network for coordinate deviation, and dynamically adjusts the learning rate of the model training according to the changing trend of the validation accuracy to prevent overfitting or underfitting. When retrieving historical successful welding records from historical welding logs that match the current temperature field distribution data and vibration spectrum data, the system employs a multi-level matching strategy to improve retrieval accuracy and efficiency. The first level of matching uses the overall gradient direction and gradient amplitude range of the temperature field distribution data as the primary matching key for initial screening. The second level of matching further filters the results from the first level using the main peak frequency and harmonic components of the vibration spectrum data as secondary matching keys. The third level of matching finalizes the results from the second level using the material type and thickness parameters of the current welded workpiece as the specific matching key. The final output of historical successful welding records is a set of records that simultaneously meet the conditions of the above three levels of matching keys.

[0085] In practical implementation, the incremental training steps of the solder joint position prediction model are performed as follows: the system performs principal component analysis on all data samples stored in the historical calibration database. Principal component analysis extracts several core feature factors affecting solder joint position deviation from the high-dimensional electrical signal features, visual deformation features, and coordinate deviation vectors contained in each data sample. These core feature factors are new feature dimensions obtained by linear transformation of the original features, ordered by variance contribution. In some embodiments, based on the core feature factors, the system constructs a nonlinear mapping network describing the complex relationship between electrical signal features, visual features, and actual coordinate deviation. The nonlinear mapping network adopts a multilayer perceptron structure, where the number of input layer nodes equals the number of core feature factors, and the number of output layer nodes equals the dimension of the coordinate deviation.

[0086] When new data samples are added to the historical calibration database, the system uses a rolling time window to select training data. The rolling time window uses a fixed number of the latest data samples as the current training set to iteratively fine-tune the connection weights of the nonlinear mapping network. The iterative fine-tuning process uses the gradient descent algorithm to update the network weights by minimizing the loss function between the predicted coordinate deviation and the actual coordinate deviation. The weight update can be understood as following the formula:

[0087]

[0088] Among them: character W τ The character W represents the connection weight matrix of the nonlinear mapping network at the τth iteration. τ+1 Let η represent the connection weight matrix after the (τ+1)th iteration update. τ This represents the learning rate used in the τth iteration, represented by the character... The loss function L is expressed with respect to the weights W. τ The gradient is calculated. After each iteration of fine-tuning, the system uses a subset of data samples that have not recently participated in training as a validation set to verify the prediction accuracy of the nonlinear mapping network for coordinate deviations in the validation set. The system dynamically adjusts the learning rate η based on the changing trend of the validation accuracy. τ If the verification accuracy does not improve after multiple consecutive iterations, then reduce the learning rate η. τ If the verification accuracy is significantly improved, then the learning rate η is maintained or slightly increased. τ .

[0089] In practical implementation, when retrieving historical successful welding records from historical welding logs that match the current temperature field distribution data and vibration spectrum data, the system employs a multi-level matching strategy. This strategy comprises three sequentially executed matching levels, each filtering records in the historical welding logs based on a matching key. The first-level matching uses the overall gradient direction of the temperature field distribution data as the primary matching key, defined by calculating the direction of the maximum temperature change in the two-dimensional plane. The system filters all records from the historical welding logs where the angle between the overall gradient direction of the temperature field and the current data direction is less than a preset threshold, forming the first-level matching result set. Optionally, the second-level matching filters the first-level matching result set using the main peak frequency of the vibration spectrum data as a secondary matching key. The system calculates the absolute difference between the current vibration spectrum main peak frequency and the main peak frequency stored in each record in the first-level matching result set, filtering records where the difference is less than a preset tolerance, forming the second-level matching result set.

[0090] In some embodiments, the third-level matching is finalized based on the material type and thickness of the current welding workpiece within the result set of the second-level matching. The system precisely compares the working material type and workpiece thickness fields of each record in the second-level matching result set with the corresponding parameters of the current workpiece, and outputs a set of completely matching records as the final historical successful welding records. Refer to Table 1 for the filtering conditions of the multi-level matching strategy.

[0091] Table 1: Matching Keys and Filtering Criteria for Multi-Level Matching Strategies

[0092] Matching level Match key Filtering criteria Level 1 Overall gradient direction of temperature field Directional angle < 5 degrees Level 2 Vibration spectrum main peak frequency Frequency difference <2Hz Level 3 Workpiece material type and thickness The model number and thickness value are completely consistent.

[0093] It is understood that the final output of historical successful welding records is a set of records that simultaneously meet the three-level matching conditions. The system extracts the required weld point coordinate compensation vector or other process parameters from the set of historical successful welding records for subsequent coordinate calibration or optimization steps. Optionally, if the multi-level matching strategy fails to find a record that meets all conditions, the system can relax the accuracy requirements for the thickness parameter in the third-level matching and instead use the record with the closest thickness, or initiate model-based interpolation compensation calculation.

[0094] See Figure 4 The statistical analysis of the search results using the multi-level matching strategy reveals the distribution of successful records for four types of workpiece materials—copper alloy, aluminum alloy, stainless steel, and carbon steel—under different matching levels (Level 1, Level 2, and Level 3) and with relaxed thickness matching. Specifically, Level 1 matching had the highest number of successful records for all materials: over 80 for copper alloy, nearly 80 for aluminum alloy, 70 for stainless steel, and approximately 65 for carbon steel. Level 2 matching had the second highest number, generally between 50 and 60. Level 3 matching significantly decreased, typically between 20 and 30. The number of records with relaxed thickness matching fell between Level 2 and Level 3, exhibiting varying degrees of variation depending on the material type. This distribution reflects that in the multi-level matching strategy, the number of successful records decreases as the matching conditions become more refined (from the temperature gradient direction to the main peak frequency of the vibration spectrum, and then to the material type and thickness). Relaxing thickness constraints can, to some extent, increase the number of search results, providing a supplementary solution for scenarios where matching fails. At the parameter level, combined with the screening conditions of the matching strategy (such as the temperature field gradient direction angle <5 degrees, the difference in the main peak frequency of the vibration spectrum <2Hz, etc.), the difference in the number of records for different materials can help evaluate the adaptability of the matching bond to various workpieces.

[0095] Example 5: After the position mapping integration module generates the initial position mapping table, or after the calibration correction process generates the corrected position mapping table, the system performs a virtual welding simulation to verify the effectiveness of the mapping table. Based on the generated position mapping table, the system drives the welding torch of the automatic welding machine to perform path planning and motion simulation in a three-dimensional virtual space without welding current. During the motion simulation, the system continuously detects whether there is spatial interference between the virtual motion path of the welding torch and the imported three-dimensional model of the workpiece, and simultaneously monitors whether the simulated welding torch posture angle exceeds the motion limits of its physical mechanism. If path interference or posture exceeding limits is detected, the system marks the virtual space coordinate points where interference occurs or the welding torch posture parameters that cause the exceeding limits as abnormal points. This abnormal information is fed back to the feature analysis and coordinate generation module. Based on the feedback abnormal point information, the feature analysis and coordinate generation module recalculates the initial positioning coordinates of the relevant calibration area, for example, adjusting the welding torch approach angle or avoiding the interference area, and generates a new, verified, and effective position mapping table after eliminating all abnormalities.

[0096] In practical implementation, after the position mapping integration module generates the initial position mapping table, or after the calibration correction process generates the corrected position mapping table, the system performs a virtual welding simulation to verify the validity of the position mapping table. Based on the calibration positioning coordinate sequence contained in the generated position mapping table, the system plans a continuous motion path without welding current for the welding torch of the automatic welding machine in a three-dimensional virtual space. The motion path connects the coordinate points in the position mapping table and drives the welding torch model to perform motion simulation along this path. In some embodiments, the motion path planning employs linear interpolation or spline curve interpolation algorithms to ensure that the tip of the welding torch can smoothly pass through the spatial position corresponding to each calibration positioning coordinate.

[0097] During the motion simulation of virtual welding, the system continuously detects whether there is spatial interference between the virtual path of the welding torch and the welding torch model itself and the pre-imported 3D model of the workpiece. Interference detection is achieved by calculating the minimum distance between the outer envelope geometry of the welding torch model and the triangular facets of the workpiece 3D model. When the minimum distance is less than or equal to zero, interference is determined to exist. The system synchronously monitors the simulated welding torch attitude angles, including the pitch and yaw angles of the welding torch axis relative to the welding platform coordinate system. The system determines whether the welding torch attitude angles exceed the motion limit range defined by the physical mechanism hardware of the automatic welding machine. The attitude over-limit judgment follows the formula:

[0098] Φ violate =max(|θ pitch |-Θ pitch_max ,|θ yaw |-Θ yaw_max ,0)

[0099] Where: character Φ violateIndicates the amount of posture violation, character θ pitch and θ yaw These represent the calculated pitch and yaw angles of the welding torch's virtual attitude, respectively, represented by the character Θ. pitch_max and Θ yaw_max These represent the absolute values ​​of the maximum pitch angle and the maximum yaw angle allowed by the physical mechanism of the automatic welding machine, respectively. It can be understood that when Φ... violate When the value is greater than 0, the welding torch posture is determined to have exceeded the physical motion limit.

[0100] If spatial interference is detected between the virtual path of the welding torch and the 3D model of the workpiece, or if the welding torch posture exceeds the physical motion limits, the system marks the set of spatial coordinate points where the interference occurs, or the set of welding torch posture angle parameters that cause the exceedance, as an anomalous point set. This anomalous point set is associated with the original data used to generate the current verification position mapping table. This anomalous point information is sent to the feature analysis and coordinate generation module through the system's internal feedback channel. After receiving the feedback anomalous point information, the feature analysis and coordinate generation module analyzes the calibration area corresponding to the anomalous point and guides the system to recalculate the initial positioning coordinates of the relevant calibration area. Optionally, the recalculation process may consider avoiding the interference area or adjusting the welding torch approach angle. After excluding all identified anomalies, the system generates a new position mapping table that has been verified as valid through virtual welding simulation. In some embodiments, all detection logs recorded during the virtual welding simulation process, including interference distance, posture angle history, and final verification conclusions, are stored as an additional information block along with the generated position mapping table for process traceability and analysis.

[0101] See Figure 5 In the attitude and interference detection analysis of virtual welding simulation, the dynamic changes of the welding torch pitch angle (red curve) and yaw angle (blue curve) with the simulation steps are visually presented. The maximum allowable pitch angle (25°) and maximum yaw angle (20°) of the automatic welding machine's physical mechanism are marked with dashed lines, and the attitude over-limit areas are indicated by color blocks. Specifically, during the simulation, the pitch angle gradually increases from an initial 15° to 30° and then decreases. The pitch angle exceeds the maximum allowable value of 25° in the range of steps 20 to 40, forming the pitch angle over-limit area marked in red. Similarly, the yaw angle gradually increases from 5° to 25° and then decreases. The yaw angle exceeds the maximum allowable value of 20° in the range of steps 40 to 60, forming the yaw angle over-limit area marked in blue. This figure, by quantifying the attitude angle changes and comparing them with the over-limit boundaries, can directly provide anomaly feedback for the feature analysis and coordinate generation modules, assisting in readjusting the calibration area coordinates to avoid attitude over-limit problems.

Claims

1. A precision resistance spot welding weld point position calibration system based on machine vision, characterized in that, The system includes: The image acquisition and processing module is used to acquire the original image set generated by the vision camera scanning the surface of the workpiece on the welding platform, and to separate the local image subsets corresponding to several preset calibration areas from the original image set. The feature parsing and coordinate generation module is used to analyze the geometric relationship between the micro-morphological features of the workpiece surface and the preset theoretical contour of the weld point for each local image subset, and generate an initial positioning coordinate in the corresponding calibration area based on the geometric relationship. The position mapping integration module is used to integrate the initial positioning coordinates of all calibration areas to form an initial position mapping table describing the current weld point distribution of the workpiece. The electrical signal acquisition and analysis module is used to acquire the real-time electrical signal waveform of the welding power supply device under a preset test current, and extract the energy distribution spectrum in the characteristic frequency band from the real-time electrical signal waveform; The spectral comparison and calibration judgment module is used to compare the energy distribution spectrum with the reference energy distribution spectrum under standard operating conditions, calculate the spectral deviation between the two in multiple characteristic dimensions, and determine whether the coordinate data in the initial position mapping table needs to be calibrated and corrected based on the magnitude of the spectral deviation.

2. The precision resistance spot welding weld point position calibration system based on machine vision according to claim 1, characterized in that, The step of generating an initial positioning coordinate within the corresponding calibration area based on the geometric relationship includes: Multi-level decomposition of local image subsets yields image pyramid levels that reflect surface features at different scales; In each image pyramid level, the ridges, edge intersections, and curvature abrupt change points on the workpiece surface are located, and these points are marked as candidate feature points. Spatial clustering and matching are performed on all candidate feature points in the image pyramid levels to remove isolated points and form a stable feature point group. Calculate the spatial centroid of the feature point group, and project the spatial centroid along the normal vector direction of the workpiece surface onto the working plane of the welding platform. Use the coordinates of the projected point as the initial positioning coordinates of the calibration area.

3. The precision resistance spot welding weld point position calibration system based on machine vision according to claim 1, characterized in that, The steps to determine whether the coordinate data in the initial position mapping table needs to be calibrated and corrected include: Set a trigger threshold and an ignore threshold for the spectral deviation, where the trigger threshold is greater than the ignore threshold; When the spectral deviation is greater than or equal to the trigger threshold, it is determined that a comprehensive calibration and correction process must be performed on the initial position mapping table. When the spectral deviation is less than the trigger threshold but greater than the ignore threshold, the determination only needs to selectively calibrate and correct the local coordinates associated with the specific dimension that contributes the most to the spectral deviation in the initial position mapping table. When the spectral deviation is less than or equal to the ignore threshold, the current initial position mapping table is deemed valid and no calibration correction is required.

4. The precision resistance spot welding weld point position calibration system based on machine vision according to claim 3, characterized in that, Performing a comprehensive calibration and correction process on the initial position mapping table includes: After determining that a comprehensive calibration and correction is needed, the vision camera is controlled to re-acquire images of each calibration area at a higher resolution to obtain a fine image set. In a fine image set, the deformation of the theoretical weld point profile caused by workpiece assembly stress, thermal deformation or platform vibration is identified and quantified, including profile translation, rotation angle and local distortion. The contour translation, rotation angle, and local distortion are combined into a set of spatial transformation parameters; The spatial transformation parameter set is applied to each initial positioning coordinate in the initial position mapping table. Through affine transformation calculation, the updated calibration positioning coordinates are output, thereby generating the corrected position mapping table.

5. The precision resistance spot welding weld point position calibration system based on machine vision according to claim 3, characterized in that, Selective calibration and correction of the local coordinates associated with the specific dimension that contributes the most to the spectral deviation in the initial position mapping table includes: Identify the key feature dimensions that cause the spectral deviation to be in the middle range, and establish a correlation between the key feature dimensions and specific mechanical axes of the welding platform or specific electrical parameters of the welding power supply device. Only the initial positioning coordinates of the calibration areas in the initial position mapping table that are covered by the specific mechanical axis movement range or directly affected by the specific electrical parameters are re-evaluated. During the reassessment, the initial positioning coordinates of the target are interpolated and compensated by combining the coordinate data of successful weld points under similar working conditions in the historical welding logs to obtain optimized coordinates. The optimized coordinates replace the original coordinates in the initial position mapping table, while the unaffected coordinates are retained, thereby generating a locally corrected position mapping table.

6. The precision resistance spot welding weld point position calibration system based on machine vision according to claim 4, characterized in that, After generating the corrected location mapping table, the process also includes updating the solder joint location prediction model: The spatial transformation parameter set generated this time, the characteristic frequency band energy distribution spectrum of the real-time electrical signal waveform, and the finally generated corrected position mapping table are used together as a data sample; The data samples are added to the system's historical calibration database, and the historical calibration database is used to incrementally train the built-in solder joint location prediction model. The solder joint position prediction model is used to pre-calibrate the initial position mapping table based on the acquired preliminary electrical signal waveform characteristics during the next system startup.

7. The precision resistance spot welding weld point position calibration system based on machine vision according to claim 5, characterized in that, The interpolation compensation calculation for the initial positioning coordinates of the target also includes an environmental parameter fusion step: The current temperature field distribution data and vibration spectrum data of the welding platform are obtained by additional environmental sensors; Retrieve historical successful welding records from historical welding logs that match the current temperature field distribution data and vibration spectrum data; Extract the weld point coordinate compensation vector from the historical successful welding records, and then superimpose the weld point coordinate compensation vector with the preliminary optimized coordinates obtained by interpolation. The result of vector superposition is used as the final optimized coordinates to replace the original coordinates in the initial position mapping table.

8. The precision resistance spot welding weld point position calibration system based on machine vision according to claim 6, characterized in that, The incremental training steps for the solder joint location prediction model specifically include: Principal component analysis was performed on all data samples in the historical calibration database to extract several core feature factors that affect the deviation of the solder joint position. Based on core feature factors, a nonlinear mapping network is constructed to describe the deviation from electrical signal features and visual features to actual coordinates. When new data samples are added, a rolling time window is used to iteratively fine-tune the connection weights of the nonlinear mapping network with a fixed number of the latest data samples. After each iteration of fine-tuning, the prediction accuracy of the nonlinear mapping network for coordinate deviations in recent data samples is verified, and the learning rate is dynamically adjusted based on the verification results.

9. The precision resistance spot welding weld point position calibration system based on machine vision according to claim 7, characterized in that, When retrieving historical successful welding records from historical welding logs that match the current temperature field distribution data and vibration spectrum data, a multi-level matching strategy is employed: The first-level matching uses the overall gradient direction of the temperature field distribution data as the primary matching key; The second-level matching is performed in the result set of the first-level matching, using the main peak frequency of the vibration spectrum data as the secondary matching key for filtering. The third-level matching is finalized based on the material type and thickness of the current welded workpiece, using the result set of the second-level matching as the specific matching key. The final output of the historical successful welding records is a set of records that simultaneously meet the three-level matching bond conditions.

10. The precision resistance spot welding weld point position calibration system based on machine vision according to claim 1, characterized in that, After generating the initial position mapping table, the system performs a virtual welding simulation to verify the effectiveness of the mapping table: Based on the initial or revised position mapping table, the welding torch of the automatic welding machine is driven to perform current-free path planning and motion simulation in virtual space. During the motion simulation, the system continuously monitors whether there is interference between the virtual path of the welding torch and the 3D model of the workpiece. Does the welding torch's posture exceed its physical limits? If interference or exceeding limits is detected, the coordinates of the interference area or the welding torch attitude parameters that caused the exceeding limits are marked as anomalies and fed back to the coordinate generation step. This guides the system to recalculate the initial positioning coordinates of the relevant calibration area and generate a new position mapping table after eliminating the anomalies.

Images