Robot spot welding controller

Abstract

The invention discloses a robot spot welding controller, and relates to the technical field of robot spot welding control, and the robot spot welding controller comprises a sensing layer which is used for collecting multi-source heterogeneous data in a welding process in real time, and carrying out the preprocessing of the multi-source heterogeneous data to form a plurality of paths of data streams; the core processing unit is connected with the sensing layer, and is used for adaptively switching control modes of the welding tongs and outputting welding parameters according to the multi-source heterogeneous data in a mode of combining prediction and real-time adjustment, and issuing corresponding welding control instructions subjected to self-healing compensation optimization to each module; the decision-making layer interacts with the core processing unit and the sensing layer and is used for carrying out three-level quality monitoring based on the multi-source heterogeneous data and the welding result and feeding back a quality monitoring result to the core processing unit so as to optimize and obtain a welding control instruction; and the execution layer is connected with the core processing unit and used for executing the optimized welding control instruction. According to the method, by constructing a layered closed-loop intelligent control framework, the autonomous decision-making and optimizing capacity of the spot welding process is remarkably improved.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of robot spot welding control, and in particular to a robot spot welding controller. BACKGROUND

[0002] In the field of industrial robot automatic welding, especially in large-scale pipeline production such as automobile body manufacturing, robot spot welding as a key connection process directly determines the safety and reliability of the whole vehicle structure.

[0003] At present, with the acceleration of product iteration and the trend of material diversification, welding production is facing the challenge of increasingly complex and variable working conditions such as workpiece material, plate thickness combination, assembly gap, etc. The traditional spot welding control system mainly adopts the combination of preset parameters and fixed control mode, and adjusts the welding current through the basic PID controller to ensure the stable output of welding energy to a certain extent. In recent years, with the development of sensing technology, some advanced systems have begun to integrate current and voltage sensors, and introduce rule-based quality monitoring, such as current upper and lower limit alarm, in an attempt to improve process controllability.

[0004] However, the existing technology gradually exposes some limitations in practice: 1. The parameter setting is seriously dependent on the experience of process experts, and it is difficult to realize adaptive adjustment for dynamic changes in working conditions, resulting in poor welding quality consistency; 2. The control mode is single and fixed, and cannot be intelligently switched to the optimal strategy according to the material characteristics or abnormal state, lacking the flexibility of multi-mode cooperation; 3. The quality evaluation is mainly offline sampling after welding, lacking real-time, comprehensive evaluation and prediction feedback mechanism for the welding process, and cannot intervene at the moment of defect generation, only realizing post-tracing rather than pre-prevention. 4. The information island phenomenon between system modules is prominent, and the sensing data, control instructions and quality information cannot form an effective closed loop, restricting the improvement of overall efficiency.

[0005] The above problems have become the key bottleneck restricting the further breakthrough of welding quality under the background of intelligent manufacturing, and an intelligent control system capable of integrating multi-source information and having autonomous decision-making and continuous optimization ability is urgently needed. SUMMARY

[0006] Therefore, it is necessary to provide a robot spot welding controller aiming at the above technical problems.

[0007] The robot spot welding controller provided by the present application comprises: A perception layer is used to collect multi-source heterogeneous data of the welding process in real time, and form multi-channel data streams after preprocessing, so as to realize multi-dimensional dynamic monitoring of the welding process. a core processing unit connected with the perception layer, configured to adaptively switch the control mode of the welding gun and output the welding parameters in combination of prediction and real-time adjustment based on the multi-source heterogeneous data, and issue corresponding welding control instructions with self-healing compensation optimization to each module; a decision layer interacting with the core processing unit and the perception layer, configured to perform three-level quality monitoring based on the multi-source heterogeneous data and the welding result, and feed back the quality monitoring result to the core processing unit to obtain the welding control instructions; an execution layer connected with the core processing unit, configured to execute the optimized welding control instructions; wherein the perception layer, the core processing unit, the decision layer and the execution layer form a closed-loop control circuit to realize adaptive control of the robot welding gun performing the welding task.

[0008] Further, the core processing unit comprises: a parameter prediction module configured to extract the current working condition information of the workpiece to be welded and the welding gun based on the real-time received multi-channel data stream, and perform working condition feature mining and matching in combination with a historical welding database to obtain the welding parameters for performing the welding task and form an initial welding parameter set; a parameter adjustment module configured to extract the multi-source heterogeneous data in real time based on a preset frequency, compare the initial welding parameter set, calculate the parameter adjustment result in real time, and generate a real-time parameter adjustment instruction set through instruction conversion to compensate for the instantaneous interference existing in the welding process; a mode management module configured to automatically select and switch the optimal control mode based on the material properties, working condition changes and equipment state information, and establish parameter mapping conversion rules to realize seamless conversion of the welding parameters when switching between different control modes; a self-healing optimization module configured to diagnose abnormal working conditions in the welding process in real time based on the real-time parameter adjustment instruction set and the feedback quality evaluation result, trigger a self-healing adjustment strategy, and generate self-healing compensation optimized welding control instructions to eliminate existing abnormal phenomena and realize online self-healing.

[0009] Further, based on the real-time received multi-channel data stream, the current working condition information of the workpiece to be welded and the welding gun is extracted, and working condition feature mining and matching are performed in combination with a historical welding database to obtain the welding parameters for performing the welding task and form an initial welding parameter set, which comprises: a Pearson correlation coefficient analysis method is used to calculate the correlation between all variables in the multi-channel data stream, and for the variable group with a correlation coefficient greater than a preset correlation threshold, representative variables are retained to screen a key variable set representing the current working condition information; The key variables in the key variable set are normalized, the data space is divided into uniform grids according to the dimensions of the normalized key variables, and a plurality of actual data points are randomly extracted in each grid as characteristic working condition points to obtain a working condition data set representing welding characteristics; Based on the historical welding database, the distribution of each to-be-predicted welding parameter in the historical success record is counted, a probability density function is fitted by using a normal distribution, and a reasonable search space of each to-be-predicted welding parameter is set in combination with a prior parameter specification set of a device manufacturer to generate a prior probability model. According to the prior probability model, weights are assigned to the historical data, the working condition data set is taken as input, and the welding parameters are taken as output to construct a weighted welding parameter prediction model. The real-time collected key variable set is input into the weighted welding parameter prediction model, the initial value of the welding parameter is calculated, and based on the model residual and the prior distribution, the confidence interval of the predicted value is calculated to obtain an initial welding parameter set.

[0010] Further, according to the prior probability model, weights are assigned to the historical data, the working condition data set is taken as input, and the welding parameters are taken as output to construct a weighted welding parameter prediction model, which comprises: A Gaussian kernel function similarity measurement method is used to calculate the matching degree weight of each sample in the working condition data set and the prior distribution to obtain a weighted welding characteristic working condition data set with a weight coefficient. Based on the weighted characteristic working condition data set as input, a full feature regression model containing interaction terms and quadratic terms is constructed to obtain a multivariate nonlinear regression model mathematical structure; and a weighted least squares method is used for regression coefficient estimation to establish an objective function, and a regression coefficient matrix and a covariance matrix are obtained. A welding verification data set is constructed by using the historical welding database, the model performance of the multivariate nonlinear regression model is evaluated by using a cross-validation method, and an optimal hyperparameter combination is selected. Based on the combination of the regression coefficient matrix and the covariance matrix, the posterior distribution of the model prediction value and the prediction variance of the new sample prediction value are calculated by using Bayesian linear regression theory, and the complete weighted welding parameter prediction model is obtained in combination with the hyperparameter combination. The trained model parameters, feature encoders and standardizers are serialized and stored to obtain a versioned weighted welding parameter prediction model file and corresponding metadata records.

[0011] Further, based on the preset frequency, the multi-source heterogeneous data is extracted in real time, compared with the initial welding parameter set, the parameter adjustment result is calculated in real time, and the real-time parameter adjustment instruction set is generated through instruction conversion. A unified timestamp mechanism is established to cache and align the multi-path data streams of the multi-source heterogeneous data based on a preset frequency, and to match the multi-source heterogeneous data with the expected values in the initial welding parameter set on the time axis, so as to obtain a set of reference expected values synchronized in time sequence; Real-time characteristic quantities of the multi-source heterogeneous data are extracted, and the real-time characteristic quantities are compared with the reference expected values to generate a multi-dimensional deviation feature vector of the deviation amount and the change trend of each welding parameter; The size of the particle swarm is set, the position of each particle represents a set of parameter adjustment amounts, and the search space and the fitness function are set based on the allowed fluctuation range of the initial welding parameter set. After iteration, the optimal parameter adjustment amount candidate set in the current control period is obtained. According to the deviation size and change trend, a preset fuzzy rule base is called for fuzzy reasoning to dynamically adjust the proportional, integral and derivative coefficients of the PID controller, generate a set of real-time parameter adjustment instructions for compensation, and perform instruction feasibility verification when the verification is passed. Instruction packaging and sending.

[0012] Further, based on material properties, working condition changes and equipment state information, the optimal control mode is automatically selected and switched, and parameter mapping conversion rules are established to realize seamless conversion of welding parameters when switching between different control modes, including: The initial welding parameter set, the real-time parameter adjustment instruction set and the real-time monitored welding gun loop state identifier are quantized, and based on the feature fusion method of the preset rule, the characteristic quantities are mapped to the preset working condition category to form a multi-dimensional working condition feature vector; The accurate numerical value in the multi-dimensional working condition feature vector is converted into a modulus language variable, the preset fuzzy rule base is activated and executed, the fuzzy recommended results output by multiple rules are combined into a control mode recommended value through fuzzy solving, and the recommended result with the highest weight is selected as the preliminary control decision; According to the target control mode of the preliminary control decision, the corresponding parameter mapping function is called, and the parameter values in the initial welding parameter set and the real-time parameter adjustment instruction set are converted into the parameter space of the target control mode using the mapping function, to generate a mode welding parameter set conforming to the mode specification; Before triggering the control mode switching, it is checked whether the welding equipment meets the safe switching condition. If the condition is met, the control mode switching instruction is issued to the power inverter unit, and the mode welding parameter set is sent to the execution layer.

[0013] Further, the control mode includes a constant secondary current mode, a constant primary peak current mode, a constant pulse width mode and a hybrid control mode.

[0014] Further, based on the real-time parameter adjustment instruction set and the quality evaluation result of the feedback, the abnormal working condition in the welding process is diagnosed in real time and the self-healing adjustment strategy is triggered to generate the self-healing compensation optimized welding control instruction, which includes: Based on the real-time parameter adjustment instruction set and the quality evaluation result, an abnormal feature vector containing time domain fluctuation, frequency domain energy distribution and quality deviation degree is constructed; and the abnormal feature vector is identified for automatic division into three abnormal type identifiers: abnormal, sudden abnormal and gradual abnormal; Based on the identified abnormal type identifier, the historical self-healing strategy case library is queried to query similar historical cases. If there is a historical case that meets the similarity standard, the historical adjustment strategy is directly adopted. If not, a new adjustment strategy is generated by using the reinforcement learning strategy network; The selected adjustment strategy after determination is simulated, and the current adjustment strategy is adjusted according to the simulation feedback to generate the final welding control instruction with self-healing compensation.

[0015] Further, the decision layer includes: A pre-inspection module before welding is used to systematically check the workpiece assembly state, welding electrode mechanism and basic parameters before the welding task is executed to generate a pre-inspection report before welding; A welding monitoring module is used to monitor the welding dynamic characteristics in real time during the welding execution process through multi-sensor fusion technology to realize abnormal instantaneous diagnosis; A post-welding analysis module is used to comprehensively detect the weld after welding, establish a welding quality full-parameter traceability file in a standardized storage format, and support defect tracing and visual display; A prediction feedback module is used to build a quality evaluation model, take the multi-source heterogeneous data collected during the welding process as input, evaluate the welding completion degree of the current welding point, and generate a quality evaluation result fed back to the core processing unit.

[0016] Further, a quality prediction model is built to take the process parameters of the continuous welding point as input to predict the quality trend of the future welding point, and generate a quality evaluation result according to the quality trend prediction result, which includes: The real-time parameter adjustment instruction set output by the core processing unit and the multi-source heterogeneous data collected during the welding process by the perception layer are obtained, and the electrical parameter features, thermodynamic features and mechanical features in the multi-source heterogeneous data are extracted respectively to form a quality evaluation feature set; Based on the pre-defined quality evaluation rule library containing static threshold rules and dynamic range rules, each feature in the quality evaluation feature set is evaluated, the compliance with the quality evaluation rules is calculated, and it is mapped to the corresponding quality level; The quality evaluation results of various types are packaged as structured data, and the welding point ID, timestamp and equipment identifier are attached, and are fed back to the core processing unit through a standard communication protocol.

[0017] The beneficial effects of the present application are: 1. By constructing a hierarchical closed-loop intelligent control architecture, the autonomous decision-making and optimization capability of the spot welding process is significantly improved; based on multi-source heterogeneous data fusion analysis, adaptive prediction and real-time dynamic adjustment of welding parameters are realized, which can actively respond to changes in working conditions caused by plate thickness fluctuations, electrode wear, shunt effect, etc., effectively overcoming the limitations of traditional control systems that rely on fixed parameters and are difficult to dynamically respond to external disturbances; in addition, experience-driven manual debugging is transformed into data-driven intelligent decision-making, thereby continuously outputting stable and reliable welding quality in a variable production environment.

[0018] 2. By establishing a three-level quality assurance mechanism covering the whole process before, during and after welding, the transition from defect post-detection to triple prevention before is realized; before welding, the assembly state and equipment health degree are accurately checked through digital pre-inspection, during welding, abnormalities are diagnosed and intervened in real time using multi-modal monitoring, and after welding, comprehensive quality analysis and data archiving are performed to form a complete and traceable quality file; the quality control mode throughout the entire process fundamentally moves the quality control point forward, realizes early identification and active defense of potential defects, and significantly reduces the risk of batch quality accidents.

[0019] 3. It has the ability of online self-learning and continuous optimization, and can accumulate knowledge and evolve iteratively through running data; after handling abnormal conditions, effective adjustment strategies are stored as new cases in the knowledge base, and the internal prediction model parameters are updated using incremental learning technology, so that the system performance is continuously improved with the extension of use time; the self-improving feature makes the system not only adapt to the initial setting conditions, but also actively follow the long-term changes such as production line rhythm and equipment aging, showing excellent long-term running stability and life cycle adaptability, providing a solid foundation for the continuous optimization of production systems in the context of intelligent manufacturing. BRIEF DESCRIPTION OF DRAWINGS

[0020] The drawings described herein are used to provide further understanding of the present application, and form a part of the present application. The illustrative embodiments of the present application and their descriptions are used to explain the present application, and do not constitute an improper limitation on the present application. In the drawings: Figure 1 is a system principle block diagram of a robot spot welding controller according to an embodiment of the present application. DETAILED DESCRIPTION

[0021] In order to make the purpose, technical scheme and advantages of the present application more clear, the present application is further described in detail below in combination with the drawings and examples. It should be understood that the specific examples described herein are only used to explain the present application and not to limit the present application.

[0022] Referring to Figure 1 The present application provides a robot spot welding controller adopting a layered closed-loop control architecture, including a perception layer, a decision layer, an execution layer and a core processing unit.

[0023] The core of the present application is a layered closed-loop control architecture, which simulates the decision-making process of human experts: perceiving the environment through "senses", analyzing and deciding through "brain", and executing actions through "hands and feet", and constantly adjusting according to the results. The perception layer is equivalent to the "senses" of the system, responsible for synchronously collecting multi-source heterogeneous data such as current, voltage, resistance, pressure, etc. in the welding process through multiple sensors, providing real-time and comprehensive data basis for the upper layer decision. The core processing unit is equivalent to the "brain" of the system, and its core is the core processing unit. This module has four functions of prediction, regulation, mode management and self-healing optimization, can dynamically predict the initial parameters according to the perception data, real-time adjust the welding process, intelligently switch the control mode, and automatically repair when an exception occurs. The decision layer builds a three-level quality assurance mechanism of "pre-welding inspection, welding monitoring, post-welding analysis", and the prediction feedback module inside can evaluate the welding completion degree of the current welding spot and feed back the quality evaluation result to the core processing unit in real time, forming a closed loop of learning and optimization. The execution layer is equivalent to the "hands and feet" of the system, including pressure control, waveform control and other modules, responsible for accurately executing the instructions issued by the core processing unit.

[0024] The operation process of the system can be briefly summarized as follows: the perception layer collects data, the core processing unit makes decisions and optimization, the execution layer executes actions, the decision layer evaluates quality and feedback, and the core processing unit performs the next round of optimization, thus forming an intelligent closed loop that continuously improves itself.

[0025] The perception layer is used for real-time collection of multi-source heterogeneous data of the welding process, and after preprocessing, multi-channel data streams are formed to realize multi-dimensional dynamic monitoring of the welding process.

[0026] Specifically, by integrating multiple high-precision sensors, the key parameters of the whole process of robot spot welding are synchronously collected, including current sensors, voltage sensors, welding pressure sensors, laser displacement sensors and temperature sensors, etc. for real-time capture of core process parameters such as welding current, electrode voltage, applied pressure, deformation amount of the welding surface and welding spot temperature.

[0027] To further comprehensively evaluate the welding environment and the weld shape, a laser vision sensor is introduced to scan the weld profile, and an environmental sensor module including temperature, humidity, and even gas composition monitoring is provided. To achieve efficient synchronization of data, the module uses a unified timestamp mechanism and a dedicated protocol conversion interface to align sensor data of different types and different sampling frequencies, effectively solving the data misplacement problem caused by communication protocol differences or acquisition time differences. The collected multi-source heterogeneous data will undergo preliminary filtering, denoising, and standardization preprocessing steps, and finally form a time-consistent, specification-unified multi-channel data stream, providing a reliable data foundation for subsequent intelligent decision-making and control.

[0028] The core processing unit is connected with the perception layer and is configured to adaptively switch the control mode of the welding gun and output welding parameters by combining prediction and real-time adjustment based on the multi-source heterogeneous data, and issue corresponding welding control instructions with self-healing compensation optimization to each module.

[0029] In the description of the present application, the core processing unit includes a parameter prediction module, a parameter adjustment module, a mode management module, and a self-healing optimization module.

[0030] The parameter prediction module is configured to extract current working condition information of the workpiece to be welded and the welding gun based on the real-time received multi-channel data stream, and combine a historical welding database to perform working condition feature mining and matching to obtain welding parameters for executing the welding task and form an initial welding parameter set.

[0031] In the description of the present application, based on the real-time received multi-channel data stream, the current working condition information of the workpiece to be welded and the welding gun is extracted, and a historical welding database is combined to perform working condition feature mining and matching to obtain welding parameters for executing the welding task and form an initial welding parameter set, which includes: S11, using Pearson correlation coefficient analysis method, the correlation between all variables in the multi-channel data stream is calculated, and for the variable group with correlation coefficient greater than the preset correlation threshold, the representative variable is retained, and the key variable set representing the current working condition information is screened.

[0032] Specifically, the pre-check loop impedance value from the perception layer, the workpiece material type code, the estimated plate thickness data fed back by the servo mechanism, and the full-dimensional operation data such as welding current setting value, welding time setting value, and electrode pressure setting value in the historical database are obtained.

[0033] The correlation between all variables is calculated using the Pearson correlation coefficient analysis method. For example, the correlation strength of the welding current set value and the pre-check loop impedance value, the workpiece material type code, and other variables is analyzed. For variable groups with a correlation coefficient higher than 0.75, such as the welding current set value and the welding power value, one representative variable is retained, and the redundant variable is discarded.

[0034] Finally, the workpiece material type code, the estimated plate thickness data, and the pre-check loop impedance value are selected as core variables as key feature dimensions.

[0035] S12, each key variable in the key variable set is normalized, the data space is divided into uniform grids according to the normalized key variable dimension, and multiple actual data points are randomly extracted in each grid as feature working condition points to obtain a working condition data set representing the welding features.

[0036] Specifically, the numerical range of each key variable is linearly mapped to the interval [0, 1] to eliminate the dimension effect. According to the normalized key variable dimension, the data space is divided into uniform grids, for example, each dimension is divided by 0.2. Up to 5 actual data points are randomly extracted in each grid as feature working condition points to ensure that the sampling points cover all typical working condition regions, such as high / medium / low load combinations.

[0037] For data-sparse grids, all original data points are retained to avoid synthesizing false working conditions.

[0038] S13, based on the historical welding database, the distribution of each to-be-predicted welding parameter in the historical success record is counted, the probability density function is fitted using the normal distribution, and the reasonable search space of each to-be-predicted welding parameter is set in combination with the prior parameter specification set of the equipment manufacturer to generate a prior probability model.

[0039] Specifically, the successful welding record set in the historical welding database is obtained, including the welding current set value, the welding time set value, the electrode pressure set value, and the corresponding welding quality evaluation result; the prior parameter specification set from the equipment manufacturer, such as the recommended parameter range for different material-plate thickness combinations.

[0040] For each to-be-predicted parameter, its distribution in the historical success record is counted, and the probability density function is fitted using the normal distribution or Gaussian mixture model. For example, for low-carbon steel 1.5mm+1.5mm lap welding, the welding current set value may be concentrated in the range of 6.0±0.5kA.

[0041] In combination with the manufacturer's specifications and historical distribution, the reasonable search interval (such as the ±3σ range) of each parameter is determined to avoid excessive interval and increase the optimization burden.

[0042] S14, according to the prior probability model, the historical data is weighted, and the working condition data set is input, and the welding parameter is output, and a weighted welding parameter prediction model is constructed.

[0043] In the description of the application, according to the prior probability model, the historical data is weighted, and the working condition data set is input, and the welding parameter is output, and a weighted welding parameter prediction model is constructed. S141, the matching degree weight of each sample in the working condition data set and the prior distribution is calculated by using the Gaussian kernel function similarity measurement method, and a weighted welding feature working condition data set with weight coefficients is obtained.

[0044] Specifically, for each sample in the representative welding feature working condition data set, the matching degree weight of the sample and the prior distribution is calculated. The Gaussian kernel function similarity measurement method is used, and the specific calculation is as follows: ; Among them, The weight coefficient of the i-th sample is represented; The value of the sample in the j-th parameter dimension is represented; The expected value of the j-th parameter in the prior distribution is represented, such as the mean value; The kernel function bandwidth parameter is adaptively adjusted according to the parameter variance.

[0045] The core idea of weight calculation is that the closer the sample parameter value is to the expected value of the prior distribution, the higher the weight obtained, and the greater the influence in subsequent regression training.

[0046] S142, based on the weighted feature working condition data set as input, a full feature regression model containing interaction term and quadratic term is constructed, and a multivariate nonlinear regression model mathematical structure is obtained. And the regression coefficient is estimated by using the weighted least square method, the objective function is established, and the regression coefficient matrix and the covariance matrix are obtained.

[0047] Specifically, a full feature regression model containing interaction term and quadratic term is constructed. The mathematical expression of the model is: ; In the formula, The predicted target is represented, the welding current set value, the welding time set value or the electrode pressure set value; The workpiece material type code is represented, and the discrete variable is one-hot coded; The estimated plate thickness data is represented; The pre-check loop impedance value is represented; The regression coefficient to be estimated is represented. The model specially retains the quadratic term and the interaction term to capture the nonlinear effects existing in the welding process, such as the coupling effect between the plate thickness and the material.

[0048] The regression coefficient is estimated by using the weighted least square method, and the objective function is: ; In the formula, is the sample weight calculated in step S141; is the measured parameter value; is the model prediction value.

[0049] S143, a welding verification data set is constructed using a historical welding database, a cross-validation method is used to evaluate the model performance of the multivariate nonlinear regression model, and the optimal hyperparameter combination is selected.

[0050] Specifically, the cross-validation method is used to evaluate the model prediction performance. Two hyperparameters are mainly adjusted: the regularization strength λ (controls the model complexity to prevent overfitting) and the Gaussian kernel bandwidth γ, which affects the concentration degree of the prior weight distribution.

[0051] The Bayesian optimization method is used to automatically search for hyperparameters in the preset range (λ∈[0.01, 1], γ∈[0.001, 0.1]) to minimize the mean absolute percentage error on the validation set as the optimization objective.

[0052] S144, based on the combination of the regression coefficient matrix and the covariance matrix, the posterior distribution of the model prediction value is calculated using the Bayesian linear regression theory, and the prediction variance of the new sample prediction value is calculated, and the complete weighted welding parameter prediction model is obtained combined with the hyperparameter combination.

[0053] S145, the trained model parameters, feature encoders, and standardizers are serialized and stored to obtain a versioned weighted welding parameter prediction model file and corresponding metadata records.

[0054] Specifically, the trained model parameters, feature encoders, standardizers, and other related components are serialized and stored, and a model version management system is established. Each model version records the following metadata: training data time range, key variable statistical characteristics, validation set performance indicators, prior distribution version number, etc. The model rolling update mechanism is supported, and the model retraining is automatically triggered when the system detects that the working condition distribution drifts.

[0055] S15, input the real-time collected key variable set into the weighted welding parameter prediction model, calculate the initial value of the welding parameter, and based on the model residual and the prior distribution, calculate the confidence interval of the prediction value to obtain the initial welding parameter set.

[0056] Specifically, the real-time key variables are input into the prediction model to calculate the initial values of the predicted welding current set value, the predicted welding time sequence value, and the predicted electrode pressure set value. Based on the model residual and the prior distribution, the confidence interval of the prediction value is calculated to provide an error tolerance reference for the subsequent dynamic adjustment module.

[0057] a parameter adjustment module configured to extract multi-source heterogeneous data in real time based on a preset frequency, compare the multi-source heterogeneous data with an initial welding parameter set, calculate a parameter adjustment result in real time, and generate a real-time parameter adjustment instruction set through instruction conversion to compensate for transient interference existing in the welding process.

[0058] In the description of the present application, the real-time extraction of multi-source heterogeneous data based on a preset frequency, comparison with an initial welding parameter set, real-time calculation of parameter adjustment results, and generation of a real-time parameter adjustment instruction set through instruction conversion include: S21, a unified timestamp mechanism is established to buffer and align the multi-channel data stream of the multi-source heterogeneous data based on a preset frequency, and to match the multi-source heterogeneous data with the expected values in the initial welding parameter set on the time axis to obtain a set of reference expected values synchronized in time sequence.

[0059] Specifically, real-time secondary current waveforms, real-time inter-electrode voltage waveforms, real-time dynamic resistance curves, and real-time electrode pressure readings from the perception layer are obtained, as well as an initial welding parameter set from the welding parameter prediction submodule, including predicted welding current set values, predicted welding time sequence values, and predicted electrode pressure set values.

[0060] A unified timestamp mechanism is established to buffer and align the multi-channel data stream of the multi-source heterogeneous data based on a preset frequency, and to match the multi-source heterogeneous data with the expected values in the initial welding parameter set on the time axis to obtain a set of reference expected values synchronized in time sequence.

[0061] S22, real-time feature quantities of the multi-source heterogeneous data are extracted, and the real-time feature quantities are compared with the reference expected values to generate a multi-dimensional deviation feature vector of the deviation quantity and the change trend of each welding parameter.

[0062] Specifically, real-time feature quantities are calculated from the synchronized data stream, including but not limited to: 1, the effective value, peak value, and rise rate (di / dt) of the real-time secondary current waveform; 2, the mean value and fluctuation amplitude of the real-time inter-electrode voltage waveform; 3, the instantaneous value and change trend of the real-time dynamic resistance curve; 4, the stability index of the real-time electrode pressure reading.

[0063] The above real-time feature quantities are compared with the reference expected values to generate real-time welding current deviation values, real-time welding voltage deviation values, real-time dynamic resistance deviation values, and real-time electrode pressure deviation values. The deviation calculation adopts a combination of percentage and absolute value to adapt to different dimensional parameters.

[0064] S23, set the particle swarm size, the position of each particle represents a set of parameter adjustment amount, and set the search space and fitness function with the allowed fluctuation range of the initial welding parameter set, and the optimal parameter adjustment amount candidate set in the current control period is obtained after iteration by particle updating and convergence.

[0065] Specifically, first, the particle swarm size is set, and the position of each particle represents a set of possible current adjustment instructions, time adjustment coefficients and pressure compensation values. The range of the search space is determined by the allowed fluctuation range of the initial welding parameter set.

[0066] Fitness function: ; Where, ΔI, ΔR, ΔP are the deviations of current, resistance and pressure respectively, and w1, w2 and w3 are weight coefficients, which are dynamically adjusted according to the welding stage, such as paying more attention to current stability in the slow rise stage and paying more attention to resistance consistency in the welding stage.

[0067] The application introduces inertia weight adaptive mechanism and quantum behavior improvement to enhance local search ability. In each iteration, the particle updates its state according to the individual optimal and global optimal positions. When the fitness function changes less than the threshold value or reaches the maximum number of iterations, the global optimal solution is output.

[0068] S24, according to the size and change trend of the deviation, a preset fuzzy rule base is called for fuzzy reasoning, the proportional, integral and differential coefficients of the PID controller are dynamically adjusted, the real-time parameter adjustment instruction set is generated, and the instruction feasibility is checked, and when the check is passed, the instruction is packaged and sent.

[0069] Specifically, according to the size and change trend of the deviation, a preset fuzzy rule base is called. For example, IF the real-time dynamic resistance deviation value is "negative large" and the change rate is "positive large", which indicates that it may splash, THEN output a larger current adjustment instruction, that is, negative adjustment.

[0070] Based on the fuzzy reasoning result, the proportional, integral and differential coefficients of the PID controller are dynamically adjusted. Gradient descent method is used to optimize the PID parameters online, and the integral square error of the deviation is minimized. Finally, the adjustment amounts of current, time and pressure are weighted and fused to ensure that the adjustment of a single parameter will not cause sharp shock of electrical or mechanical characteristics. The weight is dynamically allocated according to the welding stage, such as preferentially ensuring current stability in the middle of welding and paying attention to pressure maintenance in the end.

[0071] The adjustment commands are compared with the system's allowed upper and lower limits to ensure their feasibility. For example, current adjustment commands must not exceed the welding machine's maximum output capacity, and pressure compensation values ​​must not exceed the cylinder's working range. Sudden changes in command values ​​between adjacent cycles are smoothed to prevent actuators from failing to respond. A slope limiting algorithm is used to ensure smooth command changes. The final verified command set is encapsulated according to a predetermined communication protocol and sent to the waveform control module and pressure control module at the execution layer.

[0072] The mode management module is used to automatically select and switch the optimal control mode based on material properties, changes in working conditions and equipment status information, and to establish parameter mapping and conversion rules to achieve seamless conversion of welding parameters when switching between different control modes.

[0073] In the description of this invention, based on material properties, changes in operating conditions, and equipment status information, the optimal control mode is automatically selected and switched, and parameter mapping and conversion rules are established to achieve seamless conversion of welding parameters when switching between different control modes, including: S31. The initial welding parameter set, real-time parameter adjustment instruction set and real-time monitored welding clamp circuit status identifier are quantified by features, and the feature quantities are mapped to the preset working condition categories based on the feature fusion method with preset rules to form a multi-dimensional working condition feature vector.

[0074] Specifically, it receives an initial welding parameter set from the welding parameter prediction submodule, including the predicted welding current set value, the predicted welding time series value, and the predicted electrode pressure set value; welding clamp circuit status identifiers and grid voltage fluctuation monitoring data from the sensing layer; and a real-time parameter adjustment instruction set from the dynamic parameter adjustment submodule, such as current adjustment instructions and time adjustment coefficients.

[0075] First, the input data is quantized. For example, the welding clamp circuit status identifier is converted into a specific circuit impedance value (unit: mΩ); the grid voltage fluctuation monitoring data is converted into fluctuation amplitude (ΔU) and fluctuation duration (t). fluctuation Two characteristic quantities are used: the mean and variance of recent current adjustment commands are extracted from the real-time parameter adjustment command set as indirect indicators of system stability.

[0076] Feature fusion and operating condition classification: This invention employs a rule-based feature fusion method to map the aforementioned feature quantities to predefined operating condition categories. For example: 1. IF loop impedance > threshold Z max OR loop impedance value < threshold Z min THEN Condition Category = "Circuit Abnormal"; 2. If the amplitude of the grid voltage fluctuation ΔU > the threshold U max AND fluctuation duration t fluctuation Threshold tmax THEN operating condition category = "severe grid fluctuation"; 3、IF current adjustment instruction variance > threshold value Var max THEN operating condition category = "process instability".

[0077] S32, convert the precise numerical values in the multi-dimensional operating condition feature vector into fuzzy language variables, activate and execute the pre-set fuzzy rule base, combine the fuzzy recommendation results output by multiple rules into a control mode recommendation value through fuzzy solving, and select the recommendation result with the highest weight as the preliminary control decision.

[0078] Specifically, the precise values in the feature vector, such as the loop impedance value, are converted into fuzzy language variables, such as "normal impedance" and "high impedance", and a membership function is defined for each language variable to determine its degree of belonging to different fuzzy sets.

[0079] Fuzzy rule reasoning: activate and execute the built-in fuzzy rule base. The rule base is established based on welding domain knowledge, and a typical rule form is: 1、IF operating condition category IS "abnormal loop" AND current adjustment instruction mean IS "large positive value" THEN recommended control mode IS "fixed secondary current mode", weight 0.8; 2、IF operating condition category IS "severe grid fluctuation" THEN recommended control mode IS "constant pulse width mode", weight 0.9; 3、IF operating condition category IS "process instability" AND workpiece material type code IS "high-strength steel" THEN recommended control mode IS "hybrid control mode", weight 0.7.

[0080] Defuzzification uses defuzzification algorithms such as Centroid to combine the fuzzy recommendation results output by multiple rules into a precise control mode recommendation value. Finally, the mode with the highest weight is selected as the preliminary control mode decision.

[0081] S33, according to the target control mode of the preliminary control decision, call the corresponding parameter mapping function, and use the mapping function to convert the initial welding parameter set and the parameter values in the real-time parameter adjustment instruction set to the parameter space of the target control mode, to generate a mode-based welding parameter set that meets the mode specification.

[0082] Specifically, the parameter mapping function call needs to be made according to the target mode of the preliminary control mode decision, and the corresponding parameter mapping function is called. This function is pre-established based on the physical relationship under different control modes. For example, when switching from "fixed primary current mode" to "fixed secondary current mode", the mapping function needs to consider the transformer ratio and loop loss, and map the primary current set value (I primary ) to the secondary current expected value (Isecondary ).

[0083] The parameter conversion calculation converts the parameter values in the initial welding parameter set and the real-time parameter adjustment instruction set into the parameter space of the target control mode by using a mapping function. The welding energy is ensured to be continuous before and after the mode switching, and energy mutation caused by different parameter systems is avoided.

[0084] S34, before triggering the control mode switching, it is checked whether the welding equipment meets the safe switching condition, if the condition is met, a control mode switching instruction is issued to the power inverter unit, and the mode welding parameter set is sent to the execution layer.

[0085] Specifically, before triggering the mode switching, it is checked whether the system meets the safe switching condition. For example, it is checked whether the current is near the zero point of the welding output, and the welding transformer primary current zero-crossing signal is detected to avoid arc instability or equipment impact caused by switching at the current peak.

[0086] The switching instruction is issued and the state is synchronized: if the safe condition is met, a control mode switching instruction is issued to the power inverter unit, and the mode welding parameter set is sent to the waveform control module of the execution layer. At the same time, the active control mode identifier, i.e. the final control mode code, is synchronized to the decision layer of the decision layer and other related parts of the system, to ensure that the entire system is monitored and decided based on the same mode.

[0087] Wait for the power unit to return the confirmation signal of "mode switching completed". If the confirmation signal is not received within a specified time, or the system monitors an abnormality (such as an output abnormality), a safety protection process is triggered immediately, and an attempt is made to return to the last stable mode or a safe mode.

[0088] The self-healing optimization module is used to diagnose abnormal working conditions in the welding process in real time based on the real-time parameter adjustment instruction set and the feedback quality evaluation result, trigger a self-healing regulation strategy, generate a self-healing compensation optimized welding control instruction, eliminate existing abnormal phenomena, and realize online self-healing.

[0089] In the description of the present application, the control mode includes a constant secondary current mode, a constant primary peak current mode, a constant pulse width mode, and a hybrid control mode.

[0090] In the description of the present application, based on the real-time parameter adjustment instruction set and the feedback quality evaluation result, the abnormal working conditions in the welding process are diagnosed in real time, and a self-healing regulation strategy is triggered to generate a self-healing compensation optimized welding control instruction. S41, based on real-time parameter adjustment instruction set and quality evaluation results, construct an abnormal feature vector containing time domain volatility, frequency domain energy distribution and quality deviation degree. And abnormal identification is carried out on the abnormal feature vector, and it is automatically divided into three abnormal type identifications of abnormality, mutation abnormality and gradual abnormality.

[0091] Specifically, real-time parameter adjustment instruction set and intermediate algorithm data need to be collected, including differential signal of real-time dynamic resistance calculation data, real-time secondary current waveform fluctuation coefficient, real-time electrode pressure reading variation; And welding quality evaluation results, such as spot nugget diameter measurement value, tensile shear strength test value, welding spatter grade identification.

[0092] Firstly, the differential signal of real-time dynamic resistance calculation data is subjected to fast Fourier transform, and the high-frequency fluctuation energy proportion characteristics are extracted, such as the ratio of 0.5-2kHz frequency band energy to total energy, which is used to identify the instability signs of molten pool. At the same time, the real-time secondary current waveform fluctuation coefficient (ratio of current standard deviation to mean value in current period) is calculated as a time domain stability index. Combined with the deviation of spot nugget diameter measurement value and standard value, a multi-dimensional abnormal feature vector containing time domain volatility, frequency domain energy distribution and quality deviation degree is constructed.

[0093] Isolation forest algorithm is used for unsupervised anomaly detection of multi-dimensional abnormal feature vector. This algorithm divides the data space by randomly selecting features, calculates the anomaly score of each sample, and the shorter the path length, the higher the possibility of anomaly. Set dynamic threshold, automatically identify three abnormal modes of point anomaly (instantaneous peak), mutation anomaly (continuous offset) and gradual anomaly (slow drift).

[0094] S42, based on the identified abnormal type identification, query the historical self-healing strategy case library, query similar historical cases, if there is a historical case that meets the similarity standard, directly use the historical adjustment strategy, if not, enable the reinforcement learning strategy network to generate a new adjustment strategy.

[0095] Specifically, by querying the historical self-healing strategy case library, the library stores the characteristics of past abnormal events, the compensation measures taken and the effect evaluation. K nearest neighbor algorithm is used, and multi-dimensional abnormal feature vector is used as input to find the most similar K historical cases of the current abnormality.

[0096] If there is a high similarity case in the case library, that is, greater than the preset similarity threshold, its verified effective compensation strategy is directly used. Otherwise, the reinforcement learning strategy network (using Actor-Critic architecture) is started to generate a new strategy. The network outputs action compensation according to the abnormal feature vector of the current state.

[0097] According to the type of the abnormality, a strategy is customized: for a sudden abnormality, the strategy is "stepwise increase in current by 8% for 3 cycles"; for a gradual abnormality, the strategy is "linear increase in electrode pressure by 5%, with an increase of 1% per cycle"; and the strategy is finally encapsulated as a self-recovery strategy instruction draft.

[0098] S43, simulate the selected adjustment strategy after the determination, and adaptively adjust the current adjustment strategy according to the simulation feedback to generate a final welding control instruction with self-recovery compensation.

[0099] Specifically, the current working condition parameters and the self-recovery strategy instruction draft are input into a welding process digital twin model, and through coupling of electric, thermal and mechanical fields, the molten pool dynamics, thermal cycle curve and final weld appearance after execution of the compensation strategy can be simulated. Through running simulation, key result indicators are predicted, including predicted nugget diameter, predicted heat-affected zone size and predicted spatter index. If the prediction results meet the requirements, such as the nugget diameter meeting the requirements and the spatter index being lower than the threshold, the strategy is verified. At the same time, Monte Carlo simulation is performed to inject parameter noise and test the robustness of the strategy. If the prediction shows a risk of burn-through or incomplete fusion, the strategy is adjusted in step S42.

[0100] Finally, the amplitude and rate of the compensation curve are finely adjusted according to the simulation feedback to generate a final safe and reliable welding control instruction.

[0101] The decision layer interacts with the core processing unit and the perception layer, and is used for three-level quality monitoring based on the multi-source heterogeneous data and the welding result, and feeds back the quality monitoring result to the core processing unit to obtain the welding control instruction.

[0102] In the description of the present application, the decision layer includes a pre-inspection module before welding, a monitoring module during welding, an analysis module after welding and a prediction feedback module.

[0103] The pre-inspection module before welding is used for systematic verification of the workpiece assembly state, the welding torch mechanism and the basic parameters before the welding task is performed, to generate a pre-inspection report before welding.

[0104] Specifically, the input data includes three-dimensional point cloud data of the workpiece joint gap from the perception layer, welding torch electrode centering detection data, and basic parameter pre-inspection results such as cooling water flow and gas source pressure; and welding process specification requirements such as standard gap range and centering accuracy requirements from the management system.

[0105] The pre-inspection items before welding include the following aspects: 1. Assembly quality evaluation with 3D vision guidance: High-speed scanning of the workpiece joint by a 3D line laser profilometer generates 3D point cloud data of the workpiece joint gap. Real-time scanning data is compared with the standard CAD model using a point cloud registration algorithm to calculate parameters such as actual assembly gap and edge misalignment. If the gap is detected to be out of tolerance, an abnormality alert is generated.

[0106] 2. Multi-dimensional verification of welding gun mechanism: Laser displacement sensors detect the perpendicularity of the electrode to the workpiece surface, and simultaneously verify whether the electrode pressure build-up curve is normal; a logistic regression model is used to analyze historical data to predict electrode wear trends, and if the predicted remaining life is below a threshold, a replacement warning is triggered. 3. Process parameter compliance verification: The current parameters read by the controller are compared with the welding process specification requirements item by item, and a rule engine is used for compliance judgment. All verification results generate a pre-welding inspection report, including pass / fail status and detailed data records.

[0107] The in-process monitoring module is used to monitor the dynamic characteristics of welding in real time during the welding execution process through multi-sensor fusion technology to achieve instantaneous abnormality diagnosis.

[0108] Specifically, the input data includes real-time secondary current waveform, electrode voltage waveform, dynamic resistance curve, electrode pressure reading from the perception layer, and molten pool image sequence, arc morphology video stream from the vision system.

[0109] The items of in-process monitoring include the following aspects: 1. Synchronous acquisition and alignment of multiple physical quantities: A hardware trigger mechanism is used to ensure synchronization of the time stamps of current and voltage sensors, high-speed cameras and other devices; multi-source data is filtered and time-aligned using a sliding window weighted average method to form a unified multi-modal data stream of the welding process; 2. Feature extraction and abnormal pattern recognition: Key features are extracted from the synchronized data stream, including effective value and peak value of current waveform, fluctuation amplitude of voltage waveform, change rate of dynamic resistance curve, width-to-length ratio and oscillation frequency of molten pool image. The Isolation Forest algorithm is used to calculate the abnormal scores of each feature, and when multiple features are abnormal at the same time, a composite abnormality alarm is triggered; 3. Adaptive threshold dynamic adjustment: Based on historical qualified weld data, a statistical process control model is established to dynamically update the warning limit and control limit of each monitoring parameter. For example, as the electrode wears out, the resistance reference value will slowly drift, and the system automatically adjusts the qualified range of resistance to avoid false alarms.

[0110] The post-weld analysis module is used to conduct comprehensive detection of the weld after welding, and a standardized storage format is used to establish a full-parameter traceability file of welding quality, supporting defect tracing and visual display.

[0111] Specifically, the input data includes three-dimensional topography data of the weld from a visual sensor, surface defect images, and actual execution records of welding parameters from a control system.

[0112] The items for post-weld analysis include the following aspects: 1. Accurate reconstruction of three-dimensional topography of the weld: a 3D line laser profiler is used to scan the cooled weld to obtain three-dimensional topography data of the weld. Key dimensions such as fusion width, reinforcement height, and undercut depth are accurately calculated through point cloud processing algorithms. The measurement results are compared with standard values to automatically determine the size eligibility. 2. Intelligent identification of surface defects: a deep learning model is used to analyze the weld surface images to detect defects such as porosity, undercut, and spatter. 3. Generation of full-parameter traceability archives: pre-inspection data before welding, welding process monitoring data, and post-weld detection results are associated with a unique weld point ID to generate a full-parameter traceability archive of welding quality. The archive is stored in a standardized format and supports multi-dimensional queries and analysis based on time, equipment, and operators.

[0113] The prediction feedback module is used to build a quality evaluation model, taking multi-source heterogeneous data collected during the welding process as input, to evaluate the welding completion degree of the current weld point and generate a quality evaluation result, which is fed back to the core processing unit.

[0114] In the description of the present application, a quality prediction model is built, taking process parameters of consecutive weld points as input, to predict the quality trend of future weld points, and generate a quality evaluation result based on the quality trend prediction result, which includes: S51, obtain the real-time parameter adjustment instruction set output by the core processing unit and the multi-source heterogeneous data collected during the welding process by the perception layer, respectively extract the electrical parameter features, thermodynamic features and mechanical features in the multi-source heterogeneous data to form a quality evaluation feature set.

[0115] Specifically, the electrical parameter feature extraction includes: energy-related features, calculating the total welding energy within a single weld cycle; stability features, calculating the standard deviation and coefficient of variation (standard deviation / mean) of the dynamic resistance curve to quantify the welding process stability; waveform features, extracting the peak value, rise time, and effective value of the current-voltage curve through real-time UI profile analysis.

[0116] Thermodynamic feature extraction includes: temperature field distribution features, segmenting the welding area from thermal imaging data, calculating the maximum temperature, average temperature, and temperature gradient, i.e., the maximum temperature difference between regions; heat accumulation features, integrating the temperature-time curve to obtain the cumulative heat input.

[0117] Mechanical feature extraction calculates the electrode pressure stability index from pressure sensor data, which is the rolling standard deviation of the pressure value.

[0118] The above features are combined into a real-time feature vector according to a time window, and the dimensions include energy, resistance stability, temperature gradient, pressure stability, etc., and each welding point period corresponds to a feature vector.

[0119] S52, based on a pre-defined quality evaluation rule library, including static threshold rules and dynamic range rules, each feature in the quality evaluation feature set is evaluated, the compliance with the quality evaluation rules is calculated, and is mapped to the corresponding quality level.

[0120] Specifically, the pre-defined quality evaluation rule library includes static threshold rules and dynamic range rules: 1. Static threshold rule: for example, if the total welding energy E is not within the standard range [90J, 110J], it is marked as energy abnormal; if the temperature gradient is > 50°C / mm, it is marked as uneven heat distribution.

[0121] 2. Dynamic range rule: dynamically adjust the threshold value based on historical qualified welding point data. For example, the qualified range of the resistance coefficient of variation is dynamically updated according to the mean value ± 2σ of the recent 100 welding points.

[0122] Rule matching and scoring, for each feature, calculate its compliance with the rule library. For example, if the resistance coefficient of variation is 0.05 and the qualified range is [0.02, 0.08], the compliance is (0.08-0.05) / (0.08-0.02)=0.5.

[0123] Weighted compliance aggregation is adopted, each feature is assigned a weight, such as energy weight 0.3, temperature gradient weight 0.4, based on feature importance, total compliance , wherein w i is the weight, and c i is the single feature compliance.

[0124] Finally, the total compliance S is mapped to the quality level, for example: S≥0.9: excellent quality (all features within tolerance); 0.7≤S<0.9: qualified quality (key features meet the requirements, minor deviations in secondary features); S<0.7: unqualified quality (key features exceed the limit).

[0125] S53, encapsulate various quality evaluation results and quality levels as structured data, add welding point ID, timestamp and device identifier, and feedback to the core processing unit through a standard communication protocol.

[0126] The execution layer includes a pressure control module, a waveform control module, a data processing module, and an electrode management module, which respectively receive and respond to the feedback optimized control instructions output by the core processing unit, forming the control of the robot spot welding.

[0127] Specifically, the core task of the pressure control module is to ensure the precise and stable electrode pressure applied to the welding point, which is composed of high-response electro-pneumatic proportional valve or servo electric cylinder, high-precision piezoelectric force sensor and dedicated control unit. The working principle is to form a closed-loop control system: the controller receives the target pressure trajectory instruction from the core processing unit, at the same time, the force sensor monitors the actual electrode pressure in real time and feeds back to the controller; the controller calculates the pressure deviation in real time through high-speed PID algorithm, and drives the actuator to make accurate adjustment, so as to resist the interference caused by the uneven workpiece surface or cylinder friction, and provide a solid mechanical foundation for forming qualified welding points.

[0128] The waveform control module is the precise commander of welding energy, responsible for generating and controlling the current waveform flowing through the workpiece. The core of this module is powerful power switching devices and digital waveform synthesizer based on programmable logic devices. It receives the activation control mode identifier and specific mode welding parameter set from the upper layer. Inside the module, direct digital frequency synthesis or advanced pulse width modulation (PWM) technology is used to generate the required fundamental frequency waveform and modulation waveform. Through full-bridge circuit driving output transformer, high-frequency welding current is generated. During the entire welding process, the module will continuously collect the secondary current waveform through the Hall sensor and compare it with the set value, dynamically adjust the width and density of the PWM pulse to compensate for power grid fluctuations or loop impedance changes, and achieve nanometer-level precision energy control.

[0129] The data processing module acts as the "information hub" inside the execution layer, responsible for receiving, coordinating and distributing all control instructions and feedback data. It has multiple communication interfaces and can receive multiple instruction streams from the core processing unit and massive data streams from sensors at the same time. The core functions of this module include timestamp alignment, which aligns different asynchronous and different frequency data to the same time reference through high-precision clock and interpolation algorithm; data caching and packaging, which packages and distributes processed instructions and data according to the communication protocol recognized by each module inside the execution layer; at the same time, it also feeds back the device status of the execution layer to the upper layer system in real time, ensuring the smooth flow of information in the entire control loop.

[0130] The electrode management module focuses on the long-term health and performance maintenance of the welding tongs electrode, a key consumable, and its core is a strategy based on preventive maintenance. The module has a welding point counter built-in, which automatically accumulates the number of uses of each electrode, and comprehensively analyzes data such as dynamic resistance curves in the welding process, combines a preset electrode wear mathematical model, and calculates the estimated wear and remaining life of the electrode in real time. When the wear reaches a certain threshold, the module will automatically send a current step compensation instruction to the waveform control module, and moderately increase the welding current to compensate for the decrease in current density caused by the deformation of the electrode tip. When the electrode life is about to be exhausted or needs to be ground, it actively issues a warning or grinding request, thereby avoiding the decline in weld quality or production interruption caused by excessive electrode wear, and achieving predictive maintenance.

[0131] In summary, by means of the above technical solutions of the present application, by constructing a hierarchical closed-loop intelligent control architecture, the autonomous decision-making and optimization capability of the spot welding process is significantly improved; based on multi-source heterogeneous data fusion analysis, adaptive prediction and real-time dynamic adjustment of welding parameters are realized, which can actively respond to changes in working conditions caused by fluctuations in sheet thickness, electrode wear, and shunt effect, effectively overcoming the limitations of traditional control systems that rely on fixed parameters and are difficult to dynamically respond to external disturbances; in addition, experience-driven manual debugging is transformed into data-driven intelligent decision-making, thereby continuously outputting stable and reliable welding quality in a variable production environment. By establishing a three-level quality assurance mechanism covering the whole process before, during and after welding, the transition from defect after-detection to triple prevention before-hand is realized; before welding, the digital pre-inspection accurately checks the assembly state and equipment health, during welding, multi-modal monitoring is used to diagnose abnormalities in real time and intervene instantaneously, and after welding, comprehensive quality analysis and data archiving are performed to form a complete and traceable quality archive; the quality control mode throughout the entire process fundamentally moves the quality control point forward, realizes early identification and active defense of potential defects, and significantly reduces the risk of batch quality accidents. With online self-learning and continuous optimization capability, it can continuously accumulate knowledge and evolve through running data; after handling abnormal working conditions, it will store effective adjustment strategies as new cases in the knowledge base, and update the internal prediction model parameters using incremental learning technology, so that the system performance improves continuously with the extension of the use time; the self-improving characteristics make the system not only adapt to the initial setting conditions, but also actively follow the long-term changes such as production line rhythm and equipment aging, showing excellent long-term running stability and life cycle adaptability, providing a solid foundation for the continuous optimization of production systems in the context of intelligent manufacturing.

[0132] It should be understood that although the steps in the flowcharts of the drawings are shown in sequential order, such that each step depends on completion of the previous step before execution of the next step, the steps are not necessarily performed in the order indicated by the arrows. Unless specifically stated otherwise, the steps can be performed in other orders. Moreover, at least some of the steps in the flowcharts of the drawings can include multiple sub-steps or stages, which are not necessarily performed at the same time, but can be performed at different times, and which are not necessarily performed sequentially, but can be performed in rotation or alternation with other steps or sub-steps or stages of other steps.

Claims

1. A robotic spot welding controller characterized by, The application relates to a welding process adaptive control system based on a closed-loop control method. The welding process adaptive control system comprises a perception layer, a core processing unit, a decision layer and an execution layer. The perception layer is used for collecting multi-source heterogeneous data of a welding process in real time, and forms a multi-path data stream after preprocessing, so that multi-dimensional dynamic monitoring of the welding process is realized. The core processing unit is connected with the perception layer, and is used for adaptively switching the control mode of a welding clamp and outputting welding parameters by combining prediction with real-time adjustment according to the multi-source heterogeneous data, and issuing corresponding welding control instructions with self-recovery compensation optimization to each module. The decision layer is connected with the core processing unit and the perception layer, and is used for performing three-level quality monitoring based on the multi-source heterogeneous data and welding results, and feeding back the quality monitoring results to the core processing unit to obtain the welding control instructions. The execution layer is connected with the core processing unit, and is used for executing the optimized welding control instructions.

2. The robotic spot welding controller of claim 1, wherein, The perception layer, the core processing unit, the decision layer and the execution layer form a closed-loop control loop, so that adaptive control of a robot welding clamp performing a welding task is realized. The core processing unit comprises a parameter prediction module, a parameter adjustment module, a mode management module and a self-recovery optimization module. The parameter prediction module is used for extracting current working condition information of a workpiece to be welded and a welding clamp based on the multi-path data stream received in real time, and performing working condition feature mining and matching in combination with a historical welding database, so as to obtain welding parameters for executing the welding task and form an initial welding parameter set. The parameter adjustment module is used for extracting multi-source heterogeneous data in real time based on a preset frequency, comparing the initial welding parameter set, calculating a parameter adjustment result in real time, and generating a real-time parameter adjustment instruction set through instruction conversion to compensate for instantaneous interference existing in the welding process. The mode management module is used for automatically selecting and switching an optimal control mode based on material characteristics, working condition changes and equipment state information, establishing parameter mapping conversion rules, and realizing seamless conversion of welding parameters when different control modes are switched.

3. The robotic spot welding controller of claim 2, wherein, The self-recovery optimization module is used for diagnosing abnormal working conditions in the welding process in real time based on the real-time parameter adjustment instruction set and the feedback quality evaluation result, triggering a self-recovery adjustment strategy, generating a self-recovery compensation optimized welding control instruction, eliminating existing abnormal phenomena, and realizing online self-recovery. The method for extracting current working condition information of a workpiece to be welded and a welding clamp based on the multi-path data stream received in real time, and performing working condition feature mining and matching in combination with a historical welding database, so as to obtain welding parameters for executing the welding task and form an initial welding parameter set comprises the following steps. Pearson correlation coefficient analysis is adopted to calculate the correlation between all variables in the multi-path data stream, representative variables are reserved in a variable group with a correlation coefficient greater than a preset correlation threshold, and a key variable set representing current working condition information is screened. Normalization processing is performed on each key variable in the key variable set, the data space is divided into uniform grids according to the normalized key variable dimensions, a plurality of actual data points in each grid are randomly extracted as feature working condition points, and a working condition data set representing welding features is obtained. Based on the historical welding database, the distribution of each to-be-predicted welding parameter in the historical success record is counted, a probability density function is fitted by using a normal distribution, and a reasonable search space of each to-be-predicted welding parameter is set by combining the prior parameter specification set of the equipment manufacturer, so as to generate a prior probability model; According to the prior probability model, weights are assigned to the historical data, and a weighted welding parameter prediction model is constructed by taking the working condition data set as input and taking the welding parameter as output; The initial value of the welding parameter is calculated by inputting the real-time collected key variable set into the weighted welding parameter prediction model, and the confidence interval of the predicted value is calculated based on the model residual and the prior distribution, so as to obtain the initial welding parameter set.

4. The robotic spot welding controller of claim 3, wherein, The method according to the prior probability model includes: A Gaussian kernel function similarity measurement method is used to calculate the matching degree weight of each sample in the working condition data set and the prior distribution, so as to obtain a weighted welding feature working condition data set with a weight coefficient; Based on the weighted feature working condition data set as input, a full feature regression model containing interaction terms and quadratic terms is constructed, a multivariate nonlinear regression model mathematical structure is obtained, and a weighted least squares method is used for regression coefficient estimation to establish a target function, so as to obtain a regression coefficient matrix and a covariance matrix; A welding verification data set is constructed by using the historical welding database, the model performance of the multivariate nonlinear regression model is evaluated by using a cross-validation method, and the optimal hyperparameter combination is selected; Based on the combination of the regression coefficient matrix and the covariance matrix, the posterior distribution of the model predicted value and the prediction variance of the new sample predicted value are calculated by using the Bayesian linear regression theory, and the complete weighted welding parameter prediction model is obtained in combination with the hyperparameter combination; The trained model parameters, feature encoders and standardizers are serialized and stored to obtain a versioned weighted welding parameter prediction model file and corresponding metadata records.

5. The robotic spot welding controller of claim 2, wherein, The method of real-time extraction of multi-source heterogeneous data based on a preset frequency, comparison with the initial welding parameter set, real-time calculation of parameter adjustment results, and generation of real-time parameter adjustment instruction set through instruction conversion includes: A unified timestamp mechanism is established to cache and align the multi-path data stream of the multi-source heterogeneous data based on the preset frequency, and the multi-source heterogeneous data and the expected values in the initial welding parameter set are matched on the time axis to obtain a set of time-synchronized baseline reference expected values; Real-time feature quantities of the multi-source heterogeneous data are extracted, and the real-time feature quantities are compared with the baseline expected values to generate a multi-dimensional deviation feature vector of the deviation quantity and the change trend of each welding parameter; The size of the particle swarm is set, the position of each particle represents a set of parameter adjustment amounts, and the search space and the fitness function are set based on the allowed fluctuation range of the initial welding parameter set, and the optimal parameter adjustment amount candidate set in the current control period is obtained after iteration by particle updating and convergence. According to the deviation size and the change trend, a preset fuzzy rule base is called to perform fuzzy reasoning, the proportional, integral and differential coefficients of the PID controller are dynamically adjusted, a real-time parameter adjustment instruction set is generated, and the instruction feasibility is checked.

6. The robotic spot welding controller of claim 2, wherein, The optimal control mode is automatically selected and switched based on the material characteristics, working condition changes and equipment state information, and parameter mapping conversion rules are established to realize seamless conversion of the welding parameters during switching of different control modes, including: The initial welding parameter set, the real-time parameter adjustment instruction set and the real-time monitored welding gun loop state identifier are feature quantized, and based on a feature fusion method of a preset rule, the feature quantities are mapped to a preset working condition category to form a multi-dimensional working condition feature vector; The accurate numerical values in the multi-dimensional working condition feature vector are converted into analog-digital language variables, a preset fuzzy rule base is activated and executed, fuzzy recommended results output by multiple rules are combined into a control mode recommended value through fuzzy solving, and the recommended result with the highest weight is selected as a preliminary control decision; According to the target control mode of the preliminary control decision, a corresponding parameter mapping function is called, and the parameter values in the initial welding parameter set and the real-time parameter adjustment instruction set are converted into the parameter space of the target control mode by using the mapping function to generate a mode-based welding parameter set conforming to the mode specification; Before triggering the control mode switching, it is checked whether the welding equipment meets the safe switching conditions, if the conditions are met, the control mode switching instruction is sent to the power inverter unit, and the mode-based welding parameter set is sent to the execution layer.

7. The robotic spot welding controller of claim 6, wherein, The control mode includes a constant secondary current mode, a constant primary peak current mode, a constant pulse width mode and a hybrid control mode.

8. The robotic spot welding controller of claim 2, wherein, Based on the real-time parameter adjustment instruction set and the feedback quality evaluation result, abnormal working conditions in the welding process are diagnosed in real time and a self-healing adjustment strategy is triggered to generate a self-healing compensation optimized welding control instruction, including: Based on the real-time parameter adjustment instruction set and the quality evaluation result, an abnormal feature vector including time domain fluctuation, frequency domain energy distribution and quality deviation degree is constructed, and the abnormal feature vector is identified to automatically divide into three abnormal type identifiers, namely abnormal, sudden abnormal and gradual abnormal; Based on the identified abnormal type identifier, a historical self-healing strategy case library is queried to query similar historical cases, if there is a historical case meeting the similarity standard, the historical adjustment strategy is directly used, if not, a reinforcement learning strategy network is enabled to generate a new adjustment strategy; The selected adjustment strategy after determination is simulated, and the current adjustment strategy is adjusted according to the simulation feedback to generate a final welding control instruction with self-healing compensation.

9. The robotic spot welding controller of claim 1, wherein, The decision layer includes: A pre-inspection module before welding is used to systematically check the workpiece assembly state, welding gun mechanism and basic parameters before welding task execution to generate a pre-inspection report before welding; A monitoring module during welding is used to monitor the welding dynamic characteristics in real time during welding execution by using a multi-sensor fusion technology to realize abnormal instantaneous diagnosis. A post-weld analysis module is configured to comprehensively detect the weld after welding, establish a full-parameter traceability file of welding quality in a standardized storage format, and support defect tracing and visual display; A prediction feedback module is configured to build a quality evaluation model, take multi-source heterogeneous data collected during the welding process as input, evaluate the welding completion degree of the current welding point, and generate a quality evaluation result, which is fed back to the core processing unit.

10. The robotic spot welding controller of claim 9, wherein, The quality prediction model is built by taking process parameters of consecutive welding points as input, predicting the quality trend of future welding points, and generating a quality evaluation result based on the quality trend prediction result, which includes: The real-time parameter adjustment instruction set output by the core processing unit and the multi-source heterogeneous data collected during the welding process by the perception layer are obtained, and the electrical parameter features, thermodynamic features, and mechanical features in the multi-source heterogeneous data are extracted to form a quality evaluation feature set; Based on a pre-defined quality evaluation rule library containing static threshold rules and dynamic range rules, each feature in the quality evaluation feature set is evaluated, the compliance with the quality evaluation rules is calculated, and the result is mapped to the corresponding quality level; Each type of quality evaluation result and quality level is packaged as structured data, with the welding point ID, timestamp, and device identifier attached, and is fed back to the core processing unit through a standard communication protocol.

Images