A method and apparatus for controlling welding parameters of composite materials

Abstract

This invention discloses a method and apparatus for controlling welding parameters of composite materials, belonging to the field of intelligent welding technology. The method includes: constructing a process coupling model, setting a critical positioning method, generating a parameter bifurcation map, constructing a chaotic boundary probability cloud map, associating image feature differences between different materials, and locating the critical region of parameters; acquiring a dynamic dataset, setting a critical warning method, extracting key features, and triggering an alarm once an abnormal change is detected by analyzing abnormal changes in image features; calculating a safety offset, setting a dynamic reconstruction method, introducing a time-varying damping factor, dynamically adjusting the collaborative output, and reconstructing the optimal parameter combination by combining the welding depth and interface fusion state fed back by real-time image feedback; collecting batch welding data, setting an iterative optimization method, updating the parameters of the process coupling model, dynamically expanding the critical region identification range, and correcting the bifurcation boundary conditions in real time to enhance the adaptability to parameter critical point drift.

Description

Technical Field

[0001] This invention relates to a method and apparatus for controlling welding parameters of composite materials, belonging to the field of intelligent welding technology. Background Technology

[0002] Ultrasonic welding uses high-frequency mechanical vibration to generate frictional heat at the material contact surface, achieving molecular-level bonding. Its welding quality is closely related to the welding parameters.

[0003] However, existing technologies do not consider the critical point drift problem of nonlinear processes, especially the welding quality risks caused by chaotic effects when welding parameters approach the critical values ​​of nonlinear dynamics. Specifically, in ultrasonic welding processes operated by robotic arms, the motion control of the robotic arm, along with pressure, ultrasonic frequency, welding depth, and ultrasonic amplitude adjustment, constitute a nonlinear welding process mode. When relevant parameters approach the critical threshold, a chaotic state is easily triggered. Taking ultrasonic frequency as an example, when the parameter approaches the resonant critical point of the transducer, amplitude transformer, and welding head combination, heat energy is generated during actual operation, leading to a decrease in amplitude, affecting welding strength, and reducing the mechanical properties of the product. In addition, subtle changes in welding head stiffness caused by changes in the robotic arm load can be amplified through nonlinear acoustic coupling, causing the actual vibration frequency to exceed the critical value. This leads to unpredictable and drastic fluctuations in key welding parameters such as heat input. Such abnormal parameter fluctuations cause overheating and carbonization or lack of fusion defects at the welding interface: the molecular chains of the material in the overheated area break, resulting in a significant decrease in strength; the lack of fusion forms weak points with interfacial peel strength lower than the standard value, affecting the stability of welding quality. Summary of the Invention

[0004] To address the shortcomings of existing technologies, the present invention aims to provide a method and apparatus for controlling welding parameters of composite materials. Through deep fusion of data and image features, it achieves accurate identification and dynamic control of nonlinear process critical drift, effectively suppresses heat input fluctuations, reduces defects such as overheating and carbonization, and improves welding quality stability.

[0005] To achieve the above objectives, the present invention provides the following technical solution:

[0006] A method for controlling welding parameters of composite materials, comprising:

[0007] A process coupling model is constructed, a critical location method is set, a parameter bifurcation map is generated, and a chaotic boundary probability cloud map is constructed. The differences in image features of different materials are correlated to locate the critical region of parameters.

[0008] Acquire multispectral sequences, set critical early warning methods, extract key features, and trigger early warnings once abnormal changes are detected by analyzing abnormal changes in image features.

[0009] Calculate the safety offset, set the dynamic reconstruction method, introduce the time-varying damping factor, dynamically adjust the collaborative output, and combine the welding depth and interface fusion state fed back by real-time image feedback to reconstruct the optimal parameter combination.

[0010] Collect batch welding data, set iterative optimization methods, update process coupling model parameters, dynamically expand the critical region identification range, and correct bifurcation boundary conditions in real time.

[0011] Specifically, the critical location method includes:

[0012] Historical parameter data under dynamic operating conditions are collected and processed in layers. The image sequences are aligned between frames and Gaussian filtered to generate a layered dataset.

[0013] Cross-scale image feature parameter extraction is performed, super-resolution reconstruction is carried out on visible light images of material surfaces, surface texture entropy is extracted, surface micro-defects are segmented and the density and distribution entropy of defects are calculated, temperature field inversion is performed on infrared thermal imaging images based on the heat conduction equation, the highest temperature and temperature gradient slope of the interface are extracted, and feature parameters are clustered according to material type to generate feature parameter sets.

[0014] Theoretical constraints between parameters are established based on the ultrasonic vibration energy transfer equation and the composite material fusion dynamics model. Cross-scale image feature parameters are introduced into the model as soft constraints to construct a process coupling model. The weight coefficients of image features are adjusted using Bayesian optimization algorithm based on historical welding data to train the process coupling model.

[0015] A basic critical threshold is set, image features and quality fluctuation data during the trial welding process are extracted, sensitivity is calculated, and the basic critical threshold is dynamically corrected to generate critical positioning rules.

[0016] Specifically, the critical location method further includes:

[0017] Construct a three-dimensional coordinate axis, generate multiple sets of parameter combinations within a preset parameter range, combine with the process coupling model, mark the bifurcation points, and divide the region into stable region, subcritical region, and chaotic region to generate a parameter bifurcation map.

[0018] For the bifurcation points in the parameter bifurcation map, multiple sets of disturbance parameters are generated. The baseline probability distribution is calculated through the process coupling model, and a probability cloud map is plotted in the three-dimensional coordinate axis.

[0019] For new materials, probability cloud maps of similar materials are retrieved from historical databases, and a domain-adaptive transfer learning algorithm is used to transfer the probability distribution of similar materials to new materials. Specifically, by extracting the feature parameter set of new materials and calculating the Euclidean distance between it and the feature vectors of materials in historical databases, materials with an Euclidean distance less than the similarity difference threshold are defined as similar materials.

[0020] Microscopic images of wear on the robotic arm joints are collected and correlation coefficients are calculated. These images are then incorporated into a probability cloud map to form a chaotic boundary probability cloud map.

[0021] Based on the chaotic boundary probability cloud map, the area is divided into low-risk zone, early warning zone, and high-risk zone according to the chaotic probability value. The distribution pattern of image feature parameters in each zone is statistically analyzed, and an association matrix is ​​established.

[0022] The deviation between the actual chaotic probability value and the predicted value is calculated. When the deviation is greater than the correction threshold, the region boundary is adjusted to form a critical region parameter set.

[0023] Specifically, the critical warning method includes:

[0024] Real-time acquisition of image data and extraction of cross-scale image feature parameters to generate real-time feature sequences;

[0025] Based on the critical region parameter set, a stable critical threshold is obtained. If any feature in the real-time feature sequence exceeds the stable critical threshold, it is marked as a potential abnormal feature; otherwise, it is determined to be without abnormality, and continuous abnormality monitoring is performed.

[0026] For the potential abnormal features, the single feature deviation rate is calculated and weighted to obtain the comprehensive abnormality of the real-time feature sequence.

[0027] Specifically, the critical warning method further includes:

[0028] By combining the process coupling model, the parameter combination is reversed, and the region is located in the parameter bifurcation map. At the same time, the chaotic boundary probability cloud map is queried to obtain the chaotic probability change rate.

[0029] If the overall anomaly degree is greater than the anomaly threshold and the rate of change of the chaotic probability is greater than the anomaly rate threshold, then it is determined to be a critical drift.

[0030] A graded early warning is issued for critical drift. If the comprehensive anomaly degree is between the anomaly threshold and the secondary anomaly threshold and the parameter combination is located in the subcritical region, a first-level early warning is triggered and an anomaly feature time series curve is generated; otherwise, a second-level early warning is triggered.

[0031] Specifically, the dynamic reconstruction method includes:

[0032] Upon receiving a secondary early warning signal, the location of the combination of early warning parameters in the parameter bifurcation map is determined, defined as a risk area, and the stable critical threshold of the risk area is obtained, and the basic offset is calculated.

[0033] Based on the aforementioned correlation matrix, a coupling coefficient is introduced to correct the basic offset. Simultaneously, a material correction coefficient is invoked to further correct the basic offset.

[0034] Calculate the drift of real-time feature parameters, weighted by sensitivity, and combine with load drift to calculate the overall drift.

[0035] The drift acceleration is obtained, and the time-varying damping factor is calculated by combining the basic coefficient and dynamic coefficient, and the initial pressure and initial amplitude are corrected accordingly.

[0036] The system obtains the real-time welding depth, calls the target welding depth, calculates the target welding deviation, and once the target welding deviation exceeds the deviation threshold, it adjusts the control parameters based on the PID algorithm to generate a compensation parameter set.

[0037] Specifically, the dynamic reconstruction method further includes:

[0038] Set the rolling period, construct the dynamic objective function, and set the constraints.

[0039] A particle swarm optimization algorithm with mutation operators is used to search for the optimal parameter combination within the safe zone and then adjust the parameters.

[0040] Collect feature parameters of 3 frames of images after adjustment and calculate the average deviation rate;

[0041] If the average deviation rate is less than the deviation verification threshold, the adjustment is deemed effective; otherwise, the adjustment is deemed invalid.

[0042] Continuously monitor multiple rolling cycles and calculate the standard deviation. If the standard deviation is less than the standard verification threshold, it is determined that the risk has been eliminated; otherwise, it is determined that the risk has not been eliminated.

[0043] If any verification fails, repeat the parameter adjustment and compensation process until all indicators meet the steady-state requirements and the optimal parameter combination is generated.

[0044] Specifically, the iterative optimization method includes:

[0045] Summarize all kinds of key data in the welding process and perform structured processing on the key data;

[0046] The real-time image features during the welding process and the interface image features after welding are temporally aligned, the similarity is calculated, and matching samples with similarity exceeding the association threshold are selected.

[0047] For the matched samples, calculate the linear and non-linear correlation between image feature drift and parameter drift, and simultaneously calculate the quality influence weight to generate an association rule base.

[0048] The process coupling model is updated hierarchically, incremental learning is performed using a federated learning framework, sub-models are trained in blocks according to welding batches, and the sub-model parameters are fused by weighted averaging. A validation set is then constructed to output the incrementally optimized process coupling model.

[0049] Specifically, the iterative optimization method further includes:

[0050] Collect relevant parameters of new materials, compare them with historical material feature databases, and construct feature difference vectors;

[0051] If the magnitude of the feature difference vector is greater than the filing threshold, it is determined to be a high-difference material. A trial welding group is constructed, trial welding parameters are collected, and the identification parameters are reversed. Based on the identification parameters, the process coupling model and the feature difference vector are updated to correct the bifurcation boundary.

[0052] Otherwise, it is determined to be an approximately matching material, the initial scheme is invoked, and the bifurcation boundary is fine-tuned.

[0053] A composite material welding parameter control device includes: a critical positioning module, a critical early warning module, a dynamic reconstruction module, and an iterative optimization module;

[0054] The critical positioning module is used to construct a process coupling model, generate a parameter bifurcation map, construct a chaotic boundary probability cloud map, associate the image feature differences of different materials, and locate the parameter critical region.

[0055] The critical warning module is used to acquire multispectral sequences, extract key features, and trigger an alarm once abnormal changes are detected by analyzing abnormal changes in image features.

[0056] The dynamic reconstruction module is used to calculate the safety offset, dynamically adjust the collaborative output, and reconstruct the optimal parameter combination by combining the welding depth and interface fusion state fed back by real-time image feedback.

[0057] The iterative optimization module is used to collect batch welding data, update process coupling model parameters, dynamically expand the critical region identification range, and correct bifurcation boundary conditions in real time.

[0058] The beneficial effects of this invention are:

[0059] Based on the coupling mechanism of robotic arm motion, ultrasonic vibration, and material response, a process coupling model is established by combining cross-scale image features. By using parameter bifurcation maps and chaotic boundary probability cloud maps, the critical region of parameters is accurately located, overcoming the limitation of traditional methods in identifying critical point drift. This allows for early detection of abnormal parameter fluctuations caused by ultrasonic frequencies approaching the resonance critical point and changes in robotic arm load. With high-resolution visual monitoring and multi-dimensional feature analysis, nonlinear anomalies in image features are captured in real time, enabling early warning of chaotic states and preventing problems such as sudden amplitude drops and heat input fluctuations after parameters exceed critical values. Furthermore, this approach targets warning signals... By reconstructing dynamic parameters, calculating safety offsets, and introducing time-varying damping factors, pressure, amplitude, and welding depth are coordinated to counteract nonlinear amplification effects, stabilize heat input, and effectively suppress overheating, carbonization, or lack of fusion defects. Through iterative optimization of the entire process data, parameter drift, image features, and quality feedback are integrated to dynamically update the model and critical region boundaries, enhancing the adaptability to critical point drift under scenarios such as new materials and equipment wear. Ultimately, this achieves precise control over chaotic effects in nonlinear processes, significantly improves the stability of welding quality, and avoids problems such as strength reduction and interface defects caused by abnormal parameter fluctuations. Attached Figure Description

[0060] Figure 1 This is a schematic diagram of a method for controlling welding parameters of composite materials;

[0061] Figure 2 This is a flowchart of the critical positioning method in this invention;

[0062] Figure 3 This is a flowchart of the critical early warning method in this invention;

[0063] Figure 4 This is a flowchart of the dynamic reconstruction method in this invention;

[0064] Figure 5 This is a flowchart of the iterative optimization method in this invention. Detailed Implementation

[0065] The technical solution of the present invention will be described in detail below with reference to the accompanying drawings and specific embodiments. It should be understood that the embodiments of the present invention and the specific features in the embodiments are detailed descriptions of the technical solution of the present invention, rather than limitations thereof. In the absence of conflict, the embodiments of the present invention and the technical features in the embodiments can be combined with each other.

[0066] Example 1:

[0067] refer to Figures 1 to 5 As shown in the figure, this embodiment introduces a method for adjusting welding parameters of composite materials, including the following steps:

[0068] Step S1: Based on the physical mechanism of the robotic arm's motion characteristics, ultrasonic vibration transmission path, and material response, and combined with the image features of the welding interface and material surface, a multi-parameter coupled process coupling model is established, including pressure, welding depth, vibration frequency, ultrasonic amplitude, and image feature parameters. A critical positioning method is set, and a parameter bifurcation map is generated through simulation analysis. Combined with the differences in image features of different materials, a chaotic boundary probability cloud map of the combination of robotic arm speed and vibration frequency is constructed, and the differences in image features of different materials are correlated to clarify the chaotic boundary in the parameter space and locate the parameter critical region.

[0069] Step S2: Utilize a high-resolution vision acquisition device to continuously acquire dynamic datasets, including images of the motion trajectory of the welding head at the end of the robotic arm, real-time images of the welding interface, and images of the microscopic features of the material surface. Set up a critical warning method, combine edge detection and texture entropy analysis, extract key features from the images, and capture nonlinear features before parameters approach critical values ​​by analyzing abnormal changes in image features. Once abnormal changes are identified, trigger an early warning to achieve early warning of chaotic states.

[0070] Step S3: After triggering the pre-alarm, based on the parameter bifurcation map, calculate the safe offset of the current parameter point from the subcritical region, synchronously adjust the robot arm movement speed and ultrasonic amplitude, migrate the parameter combination to the subcritical range, set a dynamic reconstruction method, introduce a time-varying damping factor related to the image feature drift, and dynamically adjust the coordinated output of pressure and welding depth through a fuzzy logic algorithm to offset the nonlinear amplification effect caused by amplitude attenuation. Combined with the welding depth and interface fusion state fed back by real-time image, the optimal parameter combination of pressure, welding depth, and ultrasonic amplitude is reconstructed within the rolling optimization cycle to ensure the stability of heat input.

[0071] Step S4: Collect parameter drift data, post-weld interface images, and quality inspection feedback for each batch of welding. Set up iterative optimization methods, analyze the correlation between image features and parameter drift based on image feature matching, update process coupling model parameters using incremental learning, dynamically expand the critical region identification range, and for new materials or process fluctuation scenarios, start the online parameter identification process by comparing the differences between the surface image features of new materials and historical data, correct bifurcation boundary conditions in real time, and form a closed-loop iterative system that includes image monitoring, parameter control, and model optimization, gradually enhancing the adaptability to parameter critical point drift.

[0072] Specifically, the critical location method includes the following steps:

[0073] Historical parameter data under dynamic working conditions is collected, including basic data and dynamic images. The parameter data is then processed in layers according to static basic values ​​and dynamic fluctuations. For example, the robot arm speed is decomposed into rated speed and real-time fluctuation value to facilitate subsequent analysis of the stability and fluctuation characteristics of the parameters. The image sequence is aligned between frames to eliminate image displacement deviations caused by robot arm movement. Then, Gaussian filtering is used to remove noise from the infrared image and retain key temperature gradient features, thereby generating a layered dataset. The basic parameters include, but are not limited to, pitch angle, end effector speed / acceleration, welding physical parameters, vibration frequency, and ultrasonic amplitude.

[0074] Cross-scale image feature parameter extraction is performed. For visible light images of material surfaces, super-resolution reconstruction is carried out to improve image accuracy and clearly observe microstructures. The LBP algorithm is used to extract surface texture entropy to reflect the uniformity of surface roughness. The watershed algorithm is used to segment surface micro-defects and calculate the density and distribution entropy of defects to describe the degree of defect aggregation. For infrared thermal imaging images, temperature field inversion is performed based on the heat conduction equation to extract the highest temperature and temperature gradient slope of the interface. The infrared temperature field and visible light images are then registered at the pixel level to establish a spatiotemporal mapping relationship including temperature gradient and texture entropy, thereby realizing the association of features at different scales. The feature parameters are clustered according to material type to generate a feature parameter set containing the correlation weights between features.

[0075] At the physical mechanism layer, theoretical constraints between parameters are established based on the ultrasonic vibration energy transfer equation and the composite material fusion dynamics model. At the data-driven layer, cross-scale image feature parameters are introduced into the model as soft constraints. A physical information neural network architecture is adopted to construct a process coupling model to predict welding quality indicators. During the model training phase, historical welding data containing qualified and unqualified samples are used, and a Bayesian optimization algorithm is employed to adjust the weight coefficients of image features to improve the sensitivity of the process coupling model to sudden changes in image features. This ensures that the process coupling model can reflect physical laws and accurately capture the correlation between image features and welding quality. The process coupling model includes an input layer, a hidden layer, and an output layer. The input layer receives various datasets, the hidden layer embeds physical mechanism constraints, and the output layer is used to predict welding quality indicators.

[0076] A basic critical threshold is set, and three sets of verification samples are collected for trial welding for different batches of materials. Image features and quality fluctuation data during the trial welding process are extracted. Through variance analysis, the sensitivity of the characteristics and quality of the batch materials is calculated. The basic critical threshold is dynamically adjusted based on the sensitivity. If the sensitivity is high, the basic critical threshold is lowered, thereby generating critical positioning rules, which include batch-specific critical thresholds and three-dimensional judgment rules. For example, a batch of materials must meet the parameter combination requirement of being less than the critical threshold in order to be judged as being far from the critical state.

[0077] Using robotic arm speed, vibration frequency, and pressure as three-dimensional coordinate axes, multiple parameter combinations are generated within a preset parameter range using the Latin hypercube sampling method, covering both normal and extreme working conditions to ensure comprehensive sampling. Each parameter combination is input into the process coupling model to calculate the corresponding welding quality indicators and image feature parameters. Combined with critical thresholds, bifurcation points from stable to chaotic are marked and annotated on the three-dimensional coordinate axis. Stable, subcritical, and chaotic regions are marked with different colors. Simultaneously, the critical threshold surface of image features is superimposed to form a parameter bifurcation map that integrates parameters and image features. Among them, the parameter combinations in the stable region are all less than the stable critical threshold. The parameter combinations in the subcritical region are between the stable critical threshold and the chaotic critical threshold. That is, the parameter combinations in the subcritical region are close to the chaotic critical threshold but have not entered the chaotic region. The welding quality begins to fluctuate slightly, and the image features show a transitional state with abnormal precursors. The parameter combinations in the chaotic region are greater than the chaotic critical threshold, indicating that the critical threshold has been exceeded and the nonlinear chaotic region has been entered. The welding quality fluctuates drastically and there is a serious risk of defects.

[0078] For the bifurcation points in the parameter bifurcation map, multiple sets of perturbation parameter combinations are generated using Monte Carlo simulation within the preset parameter fluctuation neighborhood. Through the process coupling model, the corresponding welding quality indicators are predicted, and the critical threshold is used to determine whether the current parameter combination is in a chaotic state. At the same time, the same parameter combination is simulated multiple times to calculate the probability of each parameter combination entering a chaotic state, obtain the baseline probability distribution, and then draw a probability cloud map in the three-dimensional coordinate axis.

[0079] For the new material, probability cloud maps of similar materials are retrieved from the historical database. A domain-adaptive transfer learning algorithm is used to transfer the probability distribution of similar materials to the new material, reducing the sampling amount and improving efficiency. Simultaneously, microscopic images of wear on the robotic arm joints are collected, the correlation coefficient between the wear degree and the chaotic probability is calculated, and this is incorporated into the probability cloud map to form a chaotic boundary probability cloud map that simultaneously correlates material type and equipment state. Furthermore, by extracting the feature parameter set of the new material and calculating the Euclidean distance between it and the feature vectors of materials in the historical database, materials with an Euclidean distance less than the similarity difference threshold are defined as similar materials.

[0080] Based on the chaotic boundary probability cloud map, three levels of chaotic regions are divided according to the chaotic probability value, including low-risk zone, early warning zone, and high-risk zone. Each level of region corresponds to a different image feature threshold. The distribution pattern of image feature parameters in each region is statistically analyzed, and an association matrix containing parameter range, image features, and chaotic probability is established. The test welding data is substituted into the association matrix to calculate the deviation between the actual chaotic probability value and the predicted value. Once the deviation is greater than the correction threshold, the region boundary is fine-tuned to form a critical region parameter set, so as to clarify the parameter range, corresponding image features, and chaotic probability of different risk regions.

[0081] Specifically, the specific methods of the critical early warning method include:

[0082] The high-resolution multispectral vision acquisition device is activated to continuously acquire image data at a frequency synchronized with the welding cycle, including images of the welding head motion trajectory at the end of the robotic arm, real-time images of the welding interface, and images of the microscopic features of the material surface. Based on the inter-frame alignment algorithm, the image displacement deviation caused by the movement or vibration of the robotic arm is corrected in real time by identifying fixed marker points in the image, ensuring pixel-level alignment of images at different times. At the same time, real-time noise suppression is performed on the infrared image to preserve temperature gradient features and generate a multispectral sequence.

[0083] Multi-scale image feature parameter extraction is performed on multispectral sequences to generate real-time feature sequences;

[0084] Based on the critical region parameter set, the stable critical threshold corresponding to the current material batch is obtained, which is the boundary value between the stable region and the subcritical region. The real-time feature parameters are compared with the stable critical threshold. If any feature exceeds the stable critical threshold, it is marked as a potential abnormal feature. If all features are less than the stable critical threshold, it is determined that there is no abnormality in the current welding, and the real-time feature sequence of the next interval is continuously collected to continue abnormal monitoring.

[0085] For potential anomalous features, the single feature deviation rate is calculated based on the ratio of the difference between the real-time feature parameters and the stable critical threshold to the stable critical threshold. Combined with the feature association weight, a weighted calculation is performed to obtain the comprehensive anomalousness of the real-time feature sequence.

[0086] By combining the process coupling model, the parameter combination of the real-time feature sequence is deduced, and the region is located in the parameter bifurcation map. At the same time, the chaotic boundary probability cloud map is queried to obtain the chaotic probability and the rate of change of the chaotic probability of the parameter combination.

[0087] If the overall anomaly degree is greater than the anomaly threshold and the rate of change of chaotic probability is greater than the anomaly rate threshold, it is determined to be a critical drift, and a graded warning is issued for the critical drift. If the overall anomaly degree is between the anomaly threshold and the secondary anomaly threshold and the parameter combination is located in the subcritical region, a first-level warning is triggered, prompting enhanced monitoring and generating anomaly characteristic time series curves. Otherwise, a second-level warning is triggered, and parameter reconstruction is immediately initiated.

[0088] Specifically, the steps of the dynamic reconstruction method include:

[0089] Receive the secondary warning signal and diagnostic data, perform multi-level analysis, extract the potential abnormal features and time series curves of abnormal features contained in the warning, combine with the parameter bifurcation map, accurately locate the position of the current warning parameter combination in the parameter bifurcation map, and define it as a risk area, clarifying the area it is located in, such as a high-risk area or a subcritical area.

[0090] The critical region parameter set is called to query the safe zone boundary parameters corresponding to the risk region, which is the stable critical threshold. Then, the root cause of the anomaly is deduced through the process coupling model.

[0091] With the safety zone boundary as the target, the basic offset of each parameter in the parameter combination from the current value to the safety zone boundary is calculated. Since there is a coupling relationship between the parameters, such as the need to increase the pressure synchronously to maintain energy input when the vibration frequency decreases, the coupling coefficient is introduced in combination with the correlation matrix to correct the basic offset, ensuring that the adjusted parameter combination can still meet the energy balance. At the same time, the corresponding material correction coefficient is called in combination with the material properties to correct the offset again, so as to avoid the adjusted parameters still being at the critical edge. The smoothness of the robotic arm's motion trajectory is monitored at the same time to avoid mechanical impact caused by sudden parameter changes.

[0092] The drift of real-time characteristic parameters is calculated by the difference between real-time values ​​and stable critical thresholds. The overall drift is calculated by weighting sensitivity and combining it with load drift. A base coefficient is set for the overall drift. Drift acceleration is acquired synchronously and a dynamic coefficient is set to calculate the time-varying damping factor.

[0093] The initial pressure and initial amplitude are obtained, and the pressure and amplitude are coordinated and corrected by combining the time-varying damping factor. Simultaneously, the welding depth is monitored in real time through the vision device, and the target welding depth is called to calculate the target welding deviation. Once the target welding deviation is greater than the deviation threshold, the control parameters are adjusted based on the PID algorithm to finally generate a set of compensation parameters.

[0094] Based on the image acquisition frequency, a rolling cycle is set. A dynamic objective function is constructed by weighting the temperature gradient fluctuation value, the defect density growth rate, and the deviation of the fusion zone area ratio. The weights are the feature correlation weights. The thermal input stability is the optimization objective. At the same time, constraints are set, such as the parameter adjustment range within a single cycle does not exceed the safe range to avoid drastic fluctuations.

[0095] The particle swarm optimization algorithm with mutation operator is used to search for the optimal parameter combination within the safe zone. Initial parameters are generated based on the compensation parameter set. Random mutation is introduced in each iteration to avoid local optima, thereby generating the optimal parameter combination for the current period. When adjusting, linear interpolation is used to achieve smooth parameter transition and avoid new fluctuations caused by sudden changes.

[0096] The feature parameters of the three frames of images after adjustment are collected, and the single feature deviation rate of each parameter in each frame is calculated to obtain the average deviation rate of the stable critical threshold in the three frames. If the average deviation rate is less than the deviation verification threshold, the adjustment is determined to be initially effective; otherwise, the adjustment is determined to be ineffective.

[0097] Continuously monitor multiple rolling cycles and calculate the standard deviation of the parameter combination during the monitoring process. If the standard deviation is less than the standard verification threshold, it indicates that the trend is stable and there are no drastic fluctuations, and it is judged as having escaped the risk; otherwise, it is judged as not having escaped the risk.

[0098] If any verification fails, it means the adjustment is deemed invalid or has not escaped the risk. The above parameter adjustment and compensation process is repeated until all indicators meet the requirements of a stable state, thereby generating the optimal parameter combination and adjustment record.

[0099] Specifically, the iterative optimization method includes the following steps:

[0100] The system automatically aggregates various key data during the welding process, performs structured processing on the key data, and classifies it into three levels of tags based on welding batch, material type, and equipment status to ensure traceability. It converts parameter data into dimensionless indicators, such as the percentage of the current value relative to the boundary of the stable zone, and converts quality data into deviation rates relative to process standard values, ultimately forming a full-dimensional database. Key data includes optimal parameter combinations, real-time drift curves during parameter adjustment, anomaly feature time-series data and key image frames when warnings are triggered, post-weld inspection data, and historical benchmark data.

[0101] The real-time image features during the welding process and the interface image features after welding are time-series aligned. The similarity is calculated using a dynamic time warping algorithm. Samples with similarity exceeding the association threshold are selected and defined as matching samples to ensure the consistency between process features and result features.

[0102] Based on the Pearson correlation coefficient, for matched samples, the linear correlation between image feature drift and parameter drift is calculated to analyze the linear relationship between parameter drift and image feature drift. Based on mutual information entropy, the nonlinear correlation between image feature drift and parameter drift is calculated. At the same time, the influence weight of the combination of each parameter and image feature on the quality score is calculated through the random forest algorithm. The correlation is sorted based on the correlation to generate an association rule base, including linear correlation coefficient, nonlinear correlation coefficient, quality influence weight, and regulation guidance rules.

[0103] By combining the association rule base, the process coupling model is updated in layers. In the physical layer, the material coefficients in the ultrasonic vibration energy transfer equation are corrected according to the linear association rules between parameters and energy transfer. In the data-driven layer, the weights of image features in the neural network are adjusted based on the nonlinear association degree, and the three rules with the highest weights in terms of quality influence are transformed into hard constraints of the model.

[0104] Incremental learning is carried out using a federated learning framework. Sub-models are trained in blocks according to welding batches. Each batch of data is trained separately, and the model parameters are retained. The sub-model parameters are fused by weighted averaging to avoid forgetting historical data. The prediction error of the new model for the most recent 3 batches of data must be less than the learning ratio of the original model; otherwise, the weights are adjusted retrospectively.

[0105] Historical data of different materials and equipment states are selected as the validation set to ensure the prediction accuracy of the new model. After the validation is passed, the optimization is confirmed to be effective, and the incrementally optimized process coupling model is output.

[0106] Collect surface microscopic images and thermodynamic parameters of new materials, compare them with historical material feature databases, and construct feature difference vectors based on the difference between real-time feature values ​​and historical mean values, and then divide by historical standard deviations.

[0107] If the magnitude of the feature difference vector is greater than the filing threshold, it is determined to be a high-difference material. Three sets of test welding are performed on the high-difference material, and parameter drift and image feature data are collected during the test welding process. Bayesian filtering is used to back-infer the identification parameters of the process coupling model, such as thermodynamic parameters, fusion kinetic parameters, and energy matching parameters. The identified parameters are substituted into the physical layer equation of the process coupling model to correct the temperature field prediction logic of the model. Key parameters of the model are identified in real time. Combining the identification parameters and the feature difference vector, the bifurcation boundary is corrected. The identification parameters are used to replace the corresponding basic parameter components in the initial feature difference vector, the vector magnitude is recalculated, and the equipment wear coefficient is obtained. The boundary correction coefficient is calculated to correct the bifurcation boundary. The corrected bifurcation boundary adapts to both material properties and equipment status to ensure the dynamic adaptability of the boundary.

[0108] Otherwise, it is determined to be an approximately matching material. The bifurcation boundary and parameters of the most similar material in the historical material feature library are reused as the initial scheme. One set of test welding is carried out. The slight differences between the test welding and the historical data are compared. The bifurcation boundary is fine-tuned according to the feature difference vector and the correlation weight. If it is verified to be qualified, it is solidified as the new material boundary. If it is not qualified, the fine-tuning is repeated to balance efficiency and adaptability.

[0109] Example 2:

[0110] Another embodiment of the present invention provides: a composite material welding parameter control device, comprising: a critical positioning module, a critical early warning module, a dynamic reconstruction module, and an iterative optimization module;

[0111] The critical positioning module is used to establish a process coupling model based on the physical mechanism of robotic arm motion characteristics, ultrasonic vibration transmission path and material response, combined with the image features of welding interface and material surface, including pressure, welding depth, vibration frequency, ultrasonic amplitude and image feature parameters; by collecting historical parameter data and processing it in layers, extracting cross-scale image features, setting critical thresholds, generating parameter bifurcation map, and constructing chaotic boundary probability cloud map by combining material image feature differences, the critical region of parameters is located.

[0112] The critical warning module is used to continuously acquire image data of the robotic arm end-effector welding head movement trajectory, welding interface state, and material surface micro-features through a high-resolution multispectral vision acquisition device. After correction and noise reduction, real-time features are extracted. Based on the critical region parameter set, abnormal changes in features are analyzed, nonlinear precursors of parameters approaching critical values ​​are captured, and graded early warnings are triggered.

[0113] The dynamic reconstruction module is used to calculate the safe offset of the current parameters from the subcritical zone based on the parameter bifurcation map after the pre-alarm is triggered, and correct the offset by combining the parameter coupling relationship; it introduces a time-varying damping factor related to the image feature drift, and dynamically adjusts the pressure-welding depth coordinated output through fuzzy logic algorithm to counteract the nonlinear amplification effect; it reconstructs the optimal parameter combination within the rolling optimization cycle to ensure the stability of heat input;

[0114] The iterative optimization module collects parameter drift, interface images, and quality inspection data for each batch of welding. After structured processing, it mines the correlation between image features and parameter drift. It uses incremental learning to update the process coupling model. For new materials or process fluctuations, it corrects bifurcation boundaries in real time by comparing feature differences and identifying online parameters, forming a closed-loop system of image monitoring, parameter control, and model optimization, thereby enhancing the adaptability to parameter critical point drift.

[0115] Working principle and effects:

[0116] Based on the robotic arm's motion characteristics, ultrasonic vibration transmission laws, and material response mechanisms, a multi-parameter process coupling model is established by integrating cross-scale image features of the welding interface and surface. Simulation generates parameter bifurcation maps and chaotic boundary probability cloud maps, clearly defining the parameter boundaries of the stable, subcritical, and chaotic regions. High-resolution vision devices are used to acquire image data in real time, extract key features, and analyze abnormal changes, capturing nonlinear precursors of parameters approaching critical values ​​and triggering tiered early warnings. After an early warning, by calculating a safety offset and introducing a time-varying damping factor, parameters such as robotic arm speed, ultrasonic amplitude, and pressure are dynamically adjusted to migrate the parameter combination to the safe zone. It counteracts nonlinear amplification effects; simultaneously collects data from the entire process, mines the correlation between image features and parameter drift, updates the model through incremental learning, dynamically corrects critical boundaries for fluctuations in new materials or processes, and forms an adaptive closed loop. This effectively solves the critical point drift problem that traditional methods struggle to address. By accurately identifying chaotic precursors, dynamically adjusting parameters, and continuously optimizing the model, it avoids drastic fluctuations in heat input caused by parameters exceeding critical values, suppresses defects such as overheating carbonization and incomplete fusion, significantly improves the stability of welding quality, ensures that the strength of the weld interface meets standards, and effectively avoids the risk of decreased product mechanical properties due to chaotic effects.

[0117] The above description is merely a preferred embodiment of the present invention. The scope of protection of the present invention is not limited to the above embodiments. All technical solutions falling within the scope of the present invention's concept are within the scope of protection of the present invention. It should be noted that for those skilled in the art, any improvements and modifications made without departing from the principles of the present invention should also be considered within the scope of protection of the present invention.

Claims

1. A method for controlling welding parameters of composite materials, characterized in that, include: Based on the physical mechanisms of robotic arm motion characteristics, ultrasonic vibration transmission path and material response, and combined with the image features of welding interface and material surface, a multi-parameter coupled process coupling model is established, including pressure, welding depth, vibration frequency, ultrasonic amplitude and image feature parameters. A critical positioning method is set, and a parameter bifurcation map is generated through simulation analysis. Combined with the differences in image features of different materials, a chaotic boundary probability cloud map of the combination of robotic arm speed and vibration frequency is constructed, and the differences in image features of different materials are correlated to locate the critical region of parameters. Using a high-resolution visual acquisition device, dynamic datasets are continuously acquired, including images of the motion trajectory of the welding head at the end of the robotic arm, real-time images of the welding interface, and images of the microscopic features of the material surface. A critical warning method is set up, and key features are extracted from the images by combining edge detection and texture entropy analysis. By analyzing abnormal changes in image features, an early warning is triggered once an abnormal change is identified. After triggering the pre-alarm, based on the parameter bifurcation map, the safe offset of the current parameter point from the subcritical zone is calculated. The robot arm movement speed and ultrasonic amplitude are adjusted synchronously to migrate the parameter combination to the subcritical zone. A dynamic reconstruction method is set up, introducing a time-varying damping factor related to the image feature drift. Through a fuzzy logic algorithm, the coordinated output of pressure and welding depth is dynamically adjusted. Combined with the welding depth and interface fusion state fed back by real-time image, the optimal parameter combination of pressure, welding depth, and ultrasonic amplitude is reconstructed within the rolling optimization cycle. Collect parameter drift data, post-weld interface images, and quality inspection feedback for each batch of welding. Set up iterative optimization methods, analyze the correlation between image features and parameter drift based on image feature matching, update process coupling model parameters using incremental learning, dynamically expand the critical region identification range, and initiate online parameter identification process by comparing the differences between surface image features of new materials and historical data for new material or process fluctuation scenarios, and correct bifurcation boundary conditions in real time.

2. The method for adjusting welding parameters of composite materials according to claim 1, characterized in that, The critical location method includes: Historical parameter data under dynamic operating conditions are collected and processed in layers. The image sequences are aligned between frames and Gaussian filtered to generate a layered dataset. Cross-scale image feature parameter extraction is performed, super-resolution reconstruction is carried out on visible light images of material surfaces, surface texture entropy is extracted, surface micro-defects are segmented and the density and distribution entropy of defects are calculated, temperature field inversion is performed on infrared thermal imaging images based on the heat conduction equation, the highest temperature and temperature gradient slope of the interface are extracted, and feature parameters are clustered according to material type to generate feature parameter sets. Theoretical constraints between parameters are established based on the ultrasonic vibration energy transfer equation and the composite material fusion dynamics model. Cross-scale image feature parameters are introduced into the model as soft constraints to construct a process coupling model. The weight coefficients of image features are adjusted using Bayesian optimization algorithm based on historical welding data to train the process coupling model. A basic critical threshold is set, image features and quality fluctuation data during the trial welding process are extracted, sensitivity is calculated, and the basic critical threshold is dynamically corrected to generate critical positioning rules.

3. The method for adjusting welding parameters of composite materials according to claim 2, characterized in that, The critical location method further includes: Construct a three-dimensional coordinate axis, generate multiple sets of parameter combinations within a preset parameter range, combine with the process coupling model, mark the bifurcation points, and divide the region into stable region, subcritical region, and chaotic region to generate a parameter bifurcation map. For the bifurcation points in the parameter bifurcation map, multiple sets of disturbance parameters are generated. The baseline probability distribution is calculated through the process coupling model, and a probability cloud map is plotted in the three-dimensional coordinate axis. For new materials, probability cloud maps of similar materials are retrieved from historical databases, and a domain-adaptive transfer learning algorithm is used to transfer the probability distribution of similar materials to new materials. Specifically, by extracting the feature parameter set of new materials and calculating the Euclidean distance between it and the feature vectors of materials in historical databases, materials with an Euclidean distance less than the similarity difference threshold are defined as similar materials. Microscopic images of wear on the robotic arm joints are collected and correlation coefficients are calculated. These images are then incorporated into a probability cloud map to form a chaotic boundary probability cloud map. Based on the chaotic boundary probability cloud map, the area is divided into low-risk zone, early warning zone, and high-risk zone according to the chaotic probability value. The distribution pattern of image feature parameters in each zone is statistically analyzed, and an association matrix is ​​established. The deviation between the actual chaotic probability value and the predicted value is calculated. When the deviation is greater than the correction threshold, the region boundary is adjusted to form a critical region parameter set.

4. The method for adjusting welding parameters of composite materials according to claim 3, characterized in that, The critical early warning method includes: Real-time acquisition of image data and extraction of cross-scale image feature parameters to generate real-time feature sequences; Based on the critical region parameter set, a stable critical threshold is obtained. If any feature in the real-time feature sequence exceeds the stable critical threshold, it is marked as a potential abnormal feature; otherwise, it is determined to be without abnormality, and continuous abnormality monitoring is performed. For the potential abnormal features, the single feature deviation rate is calculated and weighted to obtain the comprehensive abnormality of the real-time feature sequence.

5. The method for adjusting welding parameters of composite materials according to claim 4, characterized in that, The critical early warning method also includes: By combining the process coupling model, the parameter combination is deduced, and the region is located in the parameter bifurcation map. At the same time, the chaotic boundary probability cloud map is queried to obtain the chaotic probability change rate. If the overall anomaly degree is greater than the anomaly threshold and the rate of change of the chaotic probability is greater than the anomaly rate threshold, then it is determined to be a critical drift. A graded early warning is issued for critical drift. If the comprehensive anomaly degree is between the anomaly threshold and the secondary anomaly threshold and the parameter combination is located in the subcritical region, a first-level early warning is triggered and an anomaly feature time series curve is generated; otherwise, a second-level early warning is triggered.

6. The method for adjusting welding parameters of composite materials according to claim 5, characterized in that, The dynamic reconstruction method includes: Upon receiving a secondary early warning signal, the location of the combination of early warning parameters in the parameter bifurcation map is determined, defined as a risk area, and the stable critical threshold of the risk area is obtained, and the basic offset is calculated. Based on the aforementioned correlation matrix, a coupling coefficient is introduced to correct the basic offset. Simultaneously, a material correction coefficient is invoked to further correct the basic offset. Calculate the drift of real-time feature parameters, weighted by sensitivity, and combine with load drift to calculate the overall drift. The drift acceleration is obtained, and the time-varying damping factor is calculated by combining the basic coefficient and dynamic coefficient, and the initial pressure and initial amplitude are corrected accordingly. The system obtains the real-time welding depth, calls the target welding depth, calculates the target welding deviation, and once the target welding deviation exceeds the deviation threshold, it adjusts the control parameters based on the PID algorithm to generate a compensation parameter set.

7. The method for adjusting welding parameters of composite materials according to claim 6, characterized in that, The dynamic reconstruction method further includes: Set the rolling period, construct the dynamic objective function, and set the constraints. A particle swarm optimization algorithm with mutation operators is used to search for the optimal parameter combination within the safe zone and then adjust the parameters. Collect feature parameters of 3 frames of images after adjustment and calculate the average deviation rate; If the average deviation rate is less than the deviation verification threshold, the adjustment is deemed effective; otherwise, the adjustment is deemed invalid. Continuously monitor multiple rolling cycles and calculate the standard deviation. If the standard deviation is less than the standard verification threshold, it is determined that the risk has been eliminated; otherwise, it is determined that the risk has not been eliminated. If any verification fails, repeat the parameter adjustment and compensation process until all indicators meet the steady-state requirements and the optimal parameter combination is generated.

8. The method for adjusting welding parameters of composite materials according to claim 7, characterized in that, The iterative optimization method includes: Summarize all kinds of key data in the welding process and perform structured processing on the key data; The real-time image features during the welding process and the interface image features after welding are temporally aligned, the similarity is calculated, and matching samples with similarity exceeding the association threshold are selected. For the matched samples, calculate the linear and non-linear correlation between image feature drift and parameter drift, and simultaneously calculate the quality influence weight to generate an association rule base. The process coupling model is updated hierarchically, incremental learning is performed using a federated learning framework, sub-models are trained in blocks according to welding batches, and the sub-model parameters are fused by weighted averaging. A validation set is then constructed to output the incrementally optimized process coupling model.

9. A method for adjusting welding parameters of composite materials according to claim 8, characterized in that, The iterative optimization method further includes: Collect relevant parameters of new materials, compare them with historical material feature databases, and construct feature difference vectors; If the magnitude of the feature difference vector is greater than the filing threshold, it is determined to be a high-difference material. A trial welding group is constructed, trial welding parameters are collected, and the identification parameters are reversed. Based on the identification parameters, the process coupling model and the feature difference vector are updated to correct the bifurcation boundary. Otherwise, it is determined to be an approximately matching material, the initial scheme is invoked, and the bifurcation boundary is fine-tuned.

10. A composite material welding parameter control device, used to implement a composite material welding parameter control method as described in any one of claims 1-9, characterized in that, include: The system includes a critical location module, a critical early warning module, a dynamic reconstruction module, and an iterative optimization module. The critical positioning module is used to construct a process coupling model, generate a parameter bifurcation map, construct a chaotic boundary probability cloud map, associate the image feature differences of different materials, and locate the parameter critical region. The critical warning module is used to acquire multispectral sequences, extract key features, and trigger an alarm once abnormal changes are detected by analyzing abnormal changes in image features. The dynamic reconstruction module is used to calculate the safety offset, dynamically adjust the collaborative output, and reconstruct the optimal parameter combination by combining the welding depth and interface fusion state fed back by real-time image feedback. The iterative optimization module is used to collect batch welding data, update process coupling model parameters, dynamically expand the critical region identification range, and correct bifurcation boundary conditions in real time.

Images