Intelligent control methods and systems for industrial robot welding

Abstract

This invention relates to the field of industrial automation welding technology, and discloses an intelligent control method and system for industrial robot welding. The method includes: after the welding task is started, calling an initial welding process parameter model based on the current workpiece specifications to generate a first welding control command sequence; controlling the welding torch to perform welding operations according to the first welding control command sequence while simultaneously acquiring multi-source sensor data streams of the welding process; performing fusion feature extraction on the real-time sensor data streams to generate a current welding status feature vector; when there is a deviation between the current welding status feature vector and the reference welding status feature vector, obtaining an adaptive welding process parameter model through online incremental learning; and generating an optimized welding control command sequence based on the adaptive welding process parameter model and sending it to the robot controller and welding torch to execute the adaptive welding operation. This invention can improve the stability of welding quality and production efficiency, and reduce equipment maintenance costs.

Description

Technical Field

[0001] This invention relates to the field of industrial automated welding technology, and more specifically, to an intelligent control method and system for industrial robot welding. Background Technology

[0002] In the field of industrial robot welding, as the manufacturing industry's requirements for welding quality and efficiency continue to increase, traditional welding control methods are gradually revealing their limitations. Traditional methods typically rely on preset, fixed welding process parameters, which are difficult to adjust precisely in response to factors such as different workpiece specifications, changes in the welding environment, and equipment aging. For example, during welding, key indicators such as arc stability, molten pool morphology, and spatter distribution are affected by various factors, and traditional methods cannot perceive these changes in real time and make dynamic adjustments. Furthermore, the monitoring and evaluation of the health status of welding equipment in existing technologies is relatively lagging, making it difficult to detect potential equipment failure risks in a timely manner, thus affecting welding quality and production efficiency.

[0003] The existing technology has at least the following problems: traditional welding control methods usually cannot perform real-time fusion and comprehensive analysis of multi-source sensor data during the welding process, thus making it difficult to accurately and comprehensively perceive the dynamic changes in the welding status; existing solutions fail to organically combine the optimization process of welding process parameters with the health status assessment of key equipment components, making it difficult to implicitly characterize and provide early warning of equipment health status. Summary of the Invention

[0004] This invention provides an intelligent control method and system for industrial robot welding.

[0005] In a first aspect, the present invention provides an intelligent control method for industrial robot welding, comprising:

[0006] After the welding task is started, the initial welding process parameter model is called based on the current workpiece specifications to generate the first welding control command sequence;

[0007] The welding torch is controlled to perform welding operations according to the first welding control command sequence, and multi-source sensor data streams of the welding process are collected simultaneously to obtain real-time sensor data streams.

[0008] The real-time sensor data stream is fused and its features are extracted to generate a feature vector of the current welding status.

[0009] When it is determined that there is a deviation between the current welding status feature vector and the benchmark welding status feature vector corresponding to the initial welding process parameter model, online incremental learning is performed on the initial welding process parameter model based on the current welding status feature vector to obtain an adaptive welding process parameter model. The online incremental learning of the initial welding process parameter model is used to realize the closed-loop adaptive adjustment of the welding process and the implicit representation of the health status of key components.

[0010] When a weld segment switching command or a process parameter readjustment command is received, an optimized welding control command sequence is generated using an adaptive welding process parameter model.

[0011] The optimized welding control command sequence is sent to the robot controller and welding torch to perform adaptive welding operations.

[0012] Furthermore, the method also includes:

[0013] When it is determined that there is no deviation between the current welding status feature vector and the benchmark welding status feature vector corresponding to the initial welding process parameter model, the current welding status feature vector is stored in the historical normal working condition feature library.

[0014] Based on all feature vectors in the historical normal working condition feature library, the statistical distribution parameters of the baseline welding situation feature vector are dynamically updated to optimize the representation range of the baseline welding situation feature vector for fluctuations in normal working conditions.

[0015] Upon receiving a weld segment switching command or a process parameter readjustment command, a welding control command sequence is generated based on the updated baseline welding status feature vector and the initial welding process parameter model.

[0016] Furthermore, the method also includes:

[0017] The current welding status feature vector is input into the process quality assessment teacher model in the pre-constructed dual-path collaborative knowledge distillation network to generate process quality knowledge labels. The dual-path collaborative knowledge distillation network includes a process quality assessment teacher model and an equipment health assessment teacher model. The process quality assessment teacher model is obtained by adversarial generative training based on historical high-quality weld process data, and is used to assess the degree of conformity of the current welding process with the high-quality process standard.

[0018] The current welding status feature vector is input into the equipment health assessment teacher model to generate equipment health knowledge tags. The equipment health assessment teacher model is obtained by time-series prediction training based on the failure mechanism of key components and multi-sensor fusion data, and is used to assess the comprehensive health status of the conductive nozzle, protective nozzle and gas supply system.

[0019] Calculate the first loss gradient between the process quality knowledge tag and the current welding status feature vector;

[0020] Calculate the second loss gradient between the equipment health knowledge tags and the current welding status feature vector;

[0021] Determine the vector projection relationship and the included angle between the first loss gradient and the second loss gradient, where the vector projection relationship and the included angle characterize the coupling contribution between process quality degradation and equipment health loss;

[0022] When it is determined that the projection component of the first loss gradient onto the second loss gradient is greater than or equal to zero, and the magnitude of the first loss gradient is less than the preset process fluctuation threshold, and the included angle is less than the preset coupling threshold, it is determined that there is no situation deviation between the current welding situation feature vector and the reference welding situation feature vector. Here, the absence of situation deviation indicates that the welding process fluctuation is mainly caused by acceptable random disturbances.

[0023] When it is determined that the projection component of the first loss gradient onto the second loss gradient is less than zero, or the magnitude of the first loss gradient is greater than or equal to the preset process fluctuation threshold, or the included angle is greater than or equal to the preset coupling threshold, it is determined that there is a situation deviation between the current welding situation feature vector and the reference welding situation feature vector. The existence of situation deviation indicates that there is a systematic deviation in the welding process caused by equipment performance degradation or abnormal working conditions.

[0024] Furthermore, based on the current welding status feature vector, the initial welding process parameter model is subjected to online incremental learning to obtain an adaptive welding process parameter model, including:

[0025] Based on the ratio of the magnitudes of the first loss gradient to the second loss gradient, determine the process optimization dominance coefficient and the health compensation dominance coefficient.

[0026] The process optimization loss components are obtained by weighting the first loss gradient based on the process optimization dominance coefficient.

[0027] The second loss gradient is weighted based on the health compensation dominance coefficient to obtain the health compensation loss component.

[0028] The process optimization loss component and the health compensation loss component are vector-synthesized to obtain the fusion optimization direction vector;

[0029] Based on the fusion optimization direction vector, the weight parameters of the initial welding process parameter model are updated using the backpropagation algorithm to obtain the adaptive welding process parameter model.

[0030] The methods also include:

[0031] Based on the historical data of health compensation loss components accumulated during the learning process of the adaptive welding process parameter model, a prediction model for the remaining service life of key components is constructed.

[0032] When the predicted remaining service life is lower than a preset threshold, a visual early warning report containing specific component identification, recommended replacement time and expected risks is generated and pushed to the production management system and maintenance terminal.

[0033] When a maintenance completion confirmation signal is received from the production management system, the health status characterization parameters related to the replaced parts are reset, and the weight parameters of the adaptive welding process parameter model are partially rolled back to the general baseline state.

[0034] Furthermore, based on the ratio of the magnitudes of the first loss gradient to the second loss gradient, the process optimization dominance coefficient and the health compensation dominance coefficient are determined, including:

[0035] Obtain the first magnitude of the first loss gradient and the second magnitude of the second loss gradient;

[0036] Calculate the first ratio and the second ratio of the first modulus to the second modulus;

[0037] The first ratio and the second ratio are input into the Sigmoid activation function for normalization and smoothing to obtain the process optimization dominance coefficient and the health compensation dominance coefficient.

[0038] Furthermore, the current welding status feature vector is input into the process quality assessment teacher model in a pre-constructed dual-path collaborative knowledge distillation network to generate process quality knowledge labels, including:

[0039] The geometric morphology index of the molten pool is input into the first sub-network, and the deep features of the molten pool image are extracted by the convolutional neural network and the weld formation quality prediction score is obtained by regression.

[0040] The arc stability quantification index and spatter distribution density index are input into the second sub-network, and the dynamic process features are extracted through the temporal neural network and the welding process stability score is obtained by regression.

[0041] Based on the weld formation quality prediction score and the welding process stability score, a process quality knowledge label is generated by weighted summation and mapping with the Softmax function.

[0042] Furthermore, the current welding status feature vector is input into the equipment health assessment teacher model to generate equipment health knowledge labels, including:

[0043] Based on the dynamic response index of wire feeding and the preset mapping relationship between wire feeding resistance and wear, the real-time equivalent wear of the conductive tip is calculated.

[0044] Based on the splash distribution density index and the simulation model of the protective airflow field, the clogging probability of the protective nozzle and the degree of local flow field distortion are evaluated.

[0045] Based on the energy spectrum characteristics of high-frequency noise in the quantitative index of arc stability, the potential risk level of insufficient protective gas flow or leakage is identified.

[0046] By integrating real-time equivalent wear, blockage probability, local flow field distortion, and potential risk level into a multi-objective decision-making process, equipment health knowledge tags that characterize the overall health status of the equipment are generated.

[0047] Furthermore, the step of calculating the real-time equivalent wear of the conductive tip adopts the following wear evolution model:

[0048] ;

[0049] in, for Real-time equivalent wear at any given moment. This is the initial wear amount. for The wire feed current at any given moment, The nominal wire feed current, for Arc voltage at any given moment This is the wear rate coefficient. This is the index of the nonlinear influence of current.

[0050] Furthermore, using an adaptive welding process parameter model, an optimized welding control command sequence is generated, including:

[0051] The current welding status feature vector and the specification features of the target weld segment are input together into the adaptive welding process parameter model;

[0052] The optimized set of welding process parameters is obtained by mapping the multilayer perceptron network in the adaptive welding process parameter model. The set of welding process parameters includes welding current, welding voltage, welding speed, wire feed speed, welding torch attitude angle and oscillation parameters.

[0053] Based on the robot's kinematics model and welding timing constraints, the set of welding process parameters is solved into an optimized welding control command sequence that can be executed by the robot controller, containing spatial path points and process parameter points.

[0054] Secondly, the present invention provides an intelligent control system for industrial robot welding, comprising:

[0055] The module includes a command response and initial control module, a welding execution and data acquisition module, a multi-source feature fusion and extraction module, a situation deviation judgment and model learning module, an adaptive command generation module, and a control command issuance module.

[0056] The embodiments of the present invention have at least the following beneficial effects:

[0057] 1. By acquiring multi-source sensor data in real time during the welding process and extracting fusion features, it is possible to accurately perceive changes in the welding status, including key indicators such as arc stability, wire feeding dynamic response, molten pool geometry, and spatter distribution density. This provides an accurate basis for adaptive adjustment of the welding process and solves the problem of unstable welding quality caused by the inability of traditional welding control methods to perceive changes in welding status in real time.

[0058] 2. Based on the deviation between the current welding status feature vector and the reference welding status feature vector, the initial welding process parameter model is incrementally learned online to generate an adaptive welding process parameter model. Based on this, the welding control command sequence is optimized, realizing closed-loop adaptive adjustment of the welding process. This enables the welding process parameters to dynamically adapt to different workpiece specifications, changes in the welding environment, and equipment aging, thereby improving welding quality and production efficiency and solving the problem that traditional fixed parameter welding methods are difficult to cope with complex working conditions. Attached Figure Description

[0059] The above and other objects, features, and advantages of exemplary embodiments of the present invention will become readily apparent from the following detailed description taken in conjunction with the accompanying drawings. Several embodiments of the invention are illustrated in the drawings by way of example and not limitation, wherein:

[0060] Figure 1 This is a flowchart illustrating an intelligent control method for industrial robot welding according to an embodiment of the present invention.

[0061] Figure 2 This is a schematic diagram of the structure of an intelligent control system for industrial robot welding provided in an embodiment of the present invention. Detailed Implementation

[0062] The technical solutions of this application will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of this application, and not all embodiments.

[0063] like Figure 1 As shown, this application proposes an intelligent control method for industrial robot welding, including:

[0064] S1. After the welding task is started, the initial welding process parameter model is called based on the current workpiece specifications to generate the first welding control instruction sequence.

[0065] S2. Control the welding torch to perform welding operations according to the first welding control command sequence, and simultaneously collect multi-source sensor data streams of the welding process to obtain real-time sensor data streams;

[0066] S3. Perform fusion feature extraction on the real-time sensor data stream to generate a feature vector of the current welding status.

[0067] S4. When a deviation is found between the current welding status feature vector and the baseline welding status feature vector corresponding to the initial welding process parameter model, online incremental learning is performed on the initial welding process parameter model based on the current welding status feature vector to obtain an adaptive welding process parameter model.

[0068] S5. Using the adaptive welding process parameter model, generate an optimized welding control command sequence and send the optimized welding control command sequence to the robot controller and welding torch to execute the adaptive welding operation.

[0069] Upon receiving the start command, the system can search for and load an initial welding process parameter model from a pre-set process parameter database based on the workpiece material type and thickness input by the operator. Subsequently, this model will generate a first welding control command sequence containing parameters such as welding current, voltage, and speed, based on the geometric characteristics of the workpiece, through interpolation or table lookup.

[0070] Specifically, the initial welding process parameter model can employ a pre-trained Multilayer Perceptron (MLP) model, comprising an input layer, hidden layers, and an output layer. The input layer receives the current welding status feature vector and the specification feature vector of the target weld segment, such as material type, plate thickness, and joint type. The hidden layer consists of three fully connected layers, each with 128 neurons, and uses ReLU as the activation function. The output layer outputs a set of welding process parameters, including welding current, welding voltage, welding speed, wire feed speed, and welding torch attitude angle. The loss function is shown in the following equation;

[0071] ;

[0072] in, and These are the process optimization dominance coefficient and the health compensation dominance coefficient, used to balance the contributions of the two losses. For process quality loss items, based on process quality knowledge tags With model output parameters get, Equipment health loss items, based on equipment health knowledge tags Feature vector of the current welding situation calculate.

[0073] The process quality loss term quantifies the deviation between the current welding process parameters and the ideal process standard. It transforms the set of welding process parameters generated by the model based on real-time conditions into a predicted process quality score vector through a mapping relationship. The difference between this predicted vector and the knowledge label vector generated by the process quality assessment teacher model can be calculated using methods such as mean squared error (MSE). The lower the loss value of the process quality loss term, the closer the model's output process parameters are to a high-quality weld.

[0074] The equipment health loss term quantifies the deviation between the equipment status reflected by the current welding process sensor data and the health baseline. By evaluating the difference between the real-time extracted welding status feature vector and the statistical baseline representing the normal operating condition of the equipment, methods such as Mahalanobis distance or its derivatives can be used to measure the degree to which the current features deviate from the normal operating condition distribution; alternatively, a regressor can be used to directly predict the health status vector from the current features and then compare it with the health baseline label given by the equipment health assessment teacher model. An increase in the equipment health loss term indicates equipment performance degradation or an increased probability of potential failure.

[0075] Upon receiving the command sequence, the welding torch is driven to move along a preset path and welding begins. During this process, current and voltage sensors can be deployed to collect arc data, while a camera captures images of the molten pool. After being acquired, this sensor data is transmitted to the data processing unit via its respective data interface, forming a real-time sensor data stream.

[0076] The real-time sensor data stream is fused and feature extracted to generate a feature vector of the current welding status. Arc sensing data sub-stream and visual sensing data sub-stream are separated from the real-time sensor data stream. For example, the data processing unit can separate the arc current and voltage data in the real-time sensor data stream into an arc sensing data sub-stream, and the image frame data into a visual sensing data sub-stream, based on the source address or data type identifier of the data packet.

[0077] Short-time Fourier transforms were performed on the separated arc current and voltage signals to analyze their spectral distribution, and energy fluctuations within a specific frequency range were calculated as a quantitative indicator of arc stability. Simultaneously, by analyzing the slopes of the rising and falling edges of the current waveform and combining this with encoder data from the wire feed motor, the wire feeding response speed was estimated as a dynamic response indicator.

[0078] The visual sensing data substream is segmented into molten pool regions and its features are analyzed to extract molten pool geometric morphology indicators and spatter distribution density indicators. For example, an image processing algorithm based on a brightness threshold can be used to identify high-brightness regions in the visual image as molten pools, and their area and perimeter can be calculated as molten pool geometric morphology indicators. For spatter, the number and size of discrete bright spots in the image can be used to statistically analyze their distribution, which can then be used as a spatter distribution density indicator.

[0079] The arc stability quantification index, wire feed dynamic response index, molten pool geometry index, and spatter distribution density index are spatiotemporally aligned and vectorized, forming a feature vector representing the current welding status. Specifically, since data acquisition from different sensors may have time differences, linear interpolation or nearest neighbor matching can be used to align these indices along the time axis. Subsequently, these aligned quantification indices are stacked sequentially to form a multidimensional feature vector, which is the current welding status feature vector.

[0080] Online incremental learning is performed on the initial welding process parameter model to achieve closed-loop adaptive adjustment of the welding process and implicit representation of the health status of key components. For example, the system can preset a threshold range for a baseline welding status feature vector. When any dimension of the current welding status feature vector exceeds this range, a status deviation is determined. At this point, a gradient descent algorithm can be used to adjust the weight parameters in the initial welding process parameter model based on the difference between the current status feature vector and the baseline vector, aligning it with the current operating conditions to obtain an adaptive welding process parameter model. This learning process aims to enable the model to adjust based on real-time feedback to maintain welding quality levels.

[0081] When a deviation from the target welding status is detected, the total loss is calculated based on the current process quality knowledge tags, the specification feature vector of the target weld segment, the equipment health knowledge tags, and the current welding status feature vector. The weight parameters of the initial model are then updated using the backpropagation algorithm. The updated model is the adaptive welding process parameter model.

[0082] The baseline welding status feature vector is a multi-dimensional feature vector obtained based on historical high-quality welding process data. It is used to characterize the ideal welding state and serves as a key benchmark for determining whether model adaptive adjustments are necessary. The construction process of the baseline welding status feature vector is as follows: From the historical welding process database, welding task records rated as high-quality by experts are selected, and their corresponding multi-source sensor data streams are extracted, including arc current, voltage timing data, molten pool image sequences, wire feed speed, etc. For each historical data point, feature extraction is performed to generate a set of historical welding status feature vectors. The mean vector and covariance matrix of this set are calculated, as shown in the following formula;

[0083] ;

[0084] ;

[0085] in, As the baseline welding situation feature vector, Let covariance matrix be the variance matrix. The total number of historical data. For the first The feature vector corresponding to each historical data point.

[0086] The initial baseline welding status feature vector is represented by this mean vector, and its fluctuation range is defined by the covariance matrix. For the current welding status feature vector extracted in real time, its Mahalanobis distance with the baseline vector is calculated as shown in the following formula;

[0087] ;

[0088] in The Mahalanobis distance, This is the feature vector of the current welding status extracted in real time.

[0089] When the distance exceeds a certain confidence level, such as 95%, a situational deviation is determined to exist. The current feature vectors that are determined not to have a situational deviation are stored in the historical normal operating condition feature library and used to periodically and dynamically update the mean vector and covariance matrix.

[0090] When the operator enters a new weld segment number, or when the system detects that the current weld segment has been completed, an instruction will be generated. At this time, the adaptive welding process parameter model will output a set of optimized welding current, voltage, speed, and other parameters based on its currently learned parameters and the new weld segment information.

[0091] The generated optimized instruction sequence is transmitted to the robot controller via a communication protocol. The robot controller parses these instructions and adjusts the welding torch's trajectory and the welding equipment's output parameters in real time to execute the optimized welding operation.

[0092] This application further proposes that when it is determined that there is no deviation between the current welding status feature vector and the reference welding status feature vector corresponding to the initial welding process parameter model, the current welding status feature vector is stored in the historical normal working condition feature library; based on all feature vectors in the historical normal working condition feature library, the statistical distribution parameters of the reference welding status feature vector are dynamically updated to optimize the representation range of the reference welding status feature vector for normal working condition fluctuations; after receiving a weld segment switching command or a process parameter readjustment command, a welding control command sequence is generated based on the updated reference welding status feature vector through the initial welding process parameter model.

[0093] The historical normal operating condition feature library is used to store the current welding status feature vectors that have been determined to have no status deviation during the welding process. The feature library can be a relational database or a time-series database to store feature vectors with timestamps; it can also be a circular buffer in memory to store normal operating condition feature vectors within a recent period.

[0094] The statistical distribution parameters of the dynamically updated baseline welding condition feature vector are designed to adjust the statistical description of the baseline welding condition feature vector in real time based on data from the historical normal operating condition feature database. Statistical distribution parameters may include the mean, variance, and covariance matrix. The dynamic update is continuously adjusted as new normal operating condition data is added, using methods such as moving average, exponentially weighted average, or Kalman filtering to ensure that the baseline model can adapt to long-term, slow changes in the welding environment and more accurately represent the fluctuation range under normal operating conditions.

[0095] Optimizing the representation range of normal operating condition fluctuations by the baseline welding situation feature vector, through dynamically updating statistical distribution parameters, allows the baseline welding situation feature vector to more accurately capture the permissible random fluctuations during normal welding. For example, by adjusting the variance or covariance matrix, the boundary of the normal situation can be expanded or narrowed, thereby avoiding misjudging normal, acceptable fluctuations as situation deviations and improving the accuracy and robustness of situation deviation judgment. Based on the updated baseline welding situation feature vector, a welding control command sequence is generated through the initial welding process parameter model. When a weld segment switching command or a process parameter readjustment command is received, the system utilizes the dynamically updated baseline welding situation feature vector. The updated baseline vector can more accurately reflect the welding characteristics under the current normal operating conditions, thereby guiding the initial welding process parameter model to generate a welding control command sequence that better reflects the actual situation.

[0096] When the Euclidean distance between the current welding status feature vector and the reference welding status feature vector is less than a preset threshold, indicating no status deviation, the current welding status feature vector is added to a circular buffer based on the first-in-first-out (FIFO) principle. This buffer serves as the historical normal operating condition feature library. The circular buffer can store the most recent 1000 normal operating condition feature vectors.

[0097] Periodically, for example, every 10 seconds, after processing each weld segment, or after each new feature vector is added, the mean vector and covariance matrix of all feature vectors in the historical normal operating condition feature library are recalculated. These mean vectors and covariance matrices are the statistical distribution parameters of the baseline welding condition feature vectors. In this way, the mean of the baseline welding condition feature vectors smoothly tracks the central trend of normal operating conditions, while the covariance matrix reflects the range and correlation of normal fluctuations across various feature dimensions, thereby optimizing its representation range of normal operating condition fluctuations.

[0098] When an operator issues a weld segment switching command or a process parameter readjustment command, the system retrieves the mean and covariance information of the latest dynamically updated baseline welding status feature vector. When generating a new welding control command sequence, the initial welding process parameter model references these updated statistical parameters. For example, during parameter optimization, the updated baseline mean is used as the target center, and the covariance matrix is ​​used to guide the parameter exploration range, ensuring that the generated command sequence meets the new task requirements.

[0099] Furthermore, the dual-path collaborative knowledge distillation network is a machine learning model architecture that can consist of two or more independent neural networks. For example, it can be a sequence model based on the Transformer architecture for processing time-series data; or it can be a hybrid model combining convolutional neural networks and recurrent neural networks for processing image and time-series features simultaneously.

[0100] Vector projection relationships represent the magnitude of one gradient component in the direction of another gradient, indicating the degree of influence of one factor on another. Angles, on the other hand, quantify the similarity or difference between two gradient directions. For example, an angle close to 0 degrees indicates that the two directions are the same, an angle close to 90 degrees indicates that they are orthogonal, and an angle close to 180 degrees indicates that they are opposite.

[0101] The preset coupling threshold is a pre-defined value used to determine whether the angle between the first and second loss gradients is small enough, thereby identifying a significant coupling relationship between process quality degradation and equipment health loss. This threshold can be determined based on historical data analysis, expert experience, or experimental verification. The preset process fluctuation threshold is a boundary value used to determine whether process quality loss is within an acceptable range of random fluctuations. It can be set based on the statistical distribution of the magnitude of the first loss gradient under historical normal operating conditions, such as using the 95th percentile.

[0102] In implementation, the process quality assessment teacher model can employ a dual-branch hybrid network. The first branch is a convolutional neural network (CNN, such as ResNet-18), which takes the ROI image of the molten pool as input and outputs deep visual features of the molten pool. The second branch is a long short-term memory network (LSTM), which takes the time-series sequence of arc stability and spatter density as input and outputs dynamic process features. The features from the two branches are concatenated and passed through a fully connected layer to obtain process quality knowledge labels.

[0103] When training the process quality assessment teacher model, multi-source sensor data corresponding to historical weld process archives are used as input. The output label is a multi-dimensional vector, such as a four-dimensional vector obtained by normalizing the percentage scores given by experts based on weld formation quality and process stability. An adversarial generative training method is adopted, which introduces a discriminator network to distinguish the labels predicted by the model from the actual expert-annotated labels, thereby enabling the process quality assessment teacher model to learn and generate process quality knowledge labels that are more in line with the distribution of process standards.

[0104] The equipment health assessment teacher model can employ a time-series prediction model based on a Transformer encoder. The input consists of multi-sensor time-series data within a sliding time window, such as wire feeding current, arc voltage, and protective gas pressure. A self-attention mechanism captures dependencies within the sequence, ultimately outputting equipment health knowledge labels. The model's training data comes from equipment maintenance logs and synchronously collected sensor data. When a critical component is replaced or a system failure occurs, the corresponding sensor data before and after the event is labeled with a specific health state. Its output label is also a multi-dimensional vector, for example, a three-dimensional vector representing the wear index of the conductive nozzle, the probability of blockage of the protective nozzle, and the risk level of the gas supply system. The values ​​of each dimension are mapped to the [0,1] interval based on the remaining service life or failure probability of the component. Through time-series prediction training, the model learns to predict or assess the evolution trend of the overall equipment health status over a future period from historical sensor sequence data, thereby obtaining the equipment health knowledge labels.

[0105] After obtaining the process quality knowledge tag and the equipment health knowledge tag, the system calculates the loss gradient between these two knowledge tags and the current welding status feature vector.

[0106] For example, the L2 norm can be used as the loss function to calculate the first and second loss gradients. Vector algebra operations are then used to calculate the angle and projection components between these two gradient vectors. Assume a preset coupling threshold of 40 degrees. If the calculated angle is less than 40 degrees, and the projection component of the first loss gradient onto the second loss gradient is greater than or equal to zero, and the magnitude of the first loss gradient is less than 0.1, then the fluctuation in the current welding process is considered an acceptable random disturbance, and online incremental learning is not required. However, if the calculated angle is greater than or equal to 40 degrees, or the projection component of the first loss gradient onto the second loss gradient is less than zero, or the magnitude of the first loss gradient is greater than or equal to 0.1, then a situational deviation is identified. For example, when the process quality knowledge tag shows a significant decrease in weld formation quality, while the equipment health knowledge tag shows severe wear on the contact tip, and the directions of the two loss gradients are highly consistent, the system will identify this as process quality degradation caused by equipment health issues. In this case, the system will trigger a subsequent online incremental learning process to adjust the initial welding process parameter model, thereby achieving adaptive adjustment to systemic deviations caused by equipment performance degradation or abnormal operating conditions.

[0107] Furthermore, after obtaining the first magnitude of the first loss gradient and the second magnitude of the second loss gradient, a first ratio can be calculated, for example, the first magnitude divided by the second magnitude, and a second ratio, for example, the second magnitude divided by the first magnitude. These two ratios are then input into the Sigmoid activation function for processing. For example, when the first magnitude is much larger than the second magnitude, the first ratio will be larger, and after processing with the Sigmoid function, the resulting process optimization dominance coefficient will be close to 1. Conversely, the second ratio will be smaller, and after processing with the Sigmoid function, the resulting health compensation dominance coefficient will be close to 0.5. This indicates that the current process quality problem is more prominent, and the model update will focus more on process optimization.

[0108] Conversely, when the second module length is much greater than the first module length, the health compensation dominance coefficient will approach 1, while the process optimization dominance coefficient will approach 0.5. In this case, the model update will focus more on equipment health compensation. When the two module lengths are similar, both ratios will approach 1. After processing with the Sigmoid function, both dominance coefficients will approach 0.73, indicating that process optimization and health compensation will play a relatively balanced role in model updates.

[0109] Furthermore, the molten pool geometry index is input into the first sub-network, and deep features of the molten pool image are extracted through a convolutional neural network, and regression is used to obtain a weld formation quality prediction score. This step aims to use the geometric information of the molten pool to evaluate the weld formation quality. The molten pool geometry index is data reflecting the shape, size, and other characteristics of the molten metal pool during the welding process, such as the width, length, depth, and area of ​​the molten pool.

[0110] The first sub-network can be a deep convolutional network containing multiple convolutional layers, pooling layers, and fully connected layers. Its input can be a preprocessed molten pool contour image or a feature map composed of molten pool geometric parameters, and its output is a continuous score value between 0 and 100, representing the percentage of forming quality. Alternatively, the first sub-network can employ a more complex CNN architecture such as ResNet or Inception to enhance feature extraction capabilities and incorporate an attention mechanism, allowing the network to focus more on regions in the molten pool that significantly affect quality, thereby improving the accuracy of the predicted score.

[0111] The arc stability quantification index and the spatter distribution density index are input into the second sub-network. Dynamic process features are extracted through a temporal neural network, and the welding process stability score is obtained through regression. The arc stability quantification index reflects the smoothness of arc combustion, such as the fluctuation amplitude and frequency of arc voltage or current; the spatter distribution density index quantifies the number, size, and spatial distribution of spatter particles during welding, and is an important characterization of welding process instability.

[0112] The second sub-network can be a recurrent neural network containing multiple layers of LSTM or GRU units. Its input is sequential data of the arc stability quantification index and the spatter distribution density index within a time window, and its output is a continuous score value between 0 and 100, representing the percentage of process stability. Alternatively, the second sub-network can employ a Transformer architecture, utilizing its self-attention mechanism to capture more complex dependencies between different time steps, or it can combine one-dimensional convolutional layers to extract local temporal features, thereby more comprehensively evaluating the dynamic stability of the welding process.

[0113] Based on the weld formation quality prediction score and the welding process stability score, a process quality knowledge label is generated through weighted summation and mapping using the Softmax function. This step aims to comprehensively consider the final weld formation quality and the dynamic stability of the welding process to generate a comprehensive process quality knowledge label. The weld formation quality prediction score and the welding process stability score reflect welding quality from two dimensions: static results and dynamic processes, respectively. The comprehensive quality assessment result is mapped to different quality levels or confidence levels, such as excellent, good, average, and poor, or a normalized vector representing the overall quality level is directly output. This process quality knowledge label ultimately serves as the quality assessment result of the current welding situation feature vector, guiding subsequent situation deviation judgments and model learning. The weighted summation can be expressed as...

[0114] Quality Score =w1 * forming quality score + w2 * process stability score;

[0115] Here, w1 and w2 are preset weights, and w1 + w2 = 1. The Softmax function can measure the quality... Score Mapped to a vector representing the probabilities of different quality levels, for example, [P]. 优 ,P 良 ,P 中 ,P 差 ] .

[0116] Furthermore, estimating the real-time equivalent wear of the conductive tip refers to quantifying the degree of wear of the conductive tip during the current welding process by analyzing the dynamic response index of wire feeding and combining the physical characteristics and wear mechanism of the conductive tip.

[0117] Assessing the probability of nozzle blockage and the degree of local flow field distortion involves analyzing the spatter distribution density index in conjunction with a simulation model of the shielding gas flow field to determine whether the nozzle is blocked or deformed, and the resulting anomalies in the shielding gas flow field. Blockage or deformation of the nozzle severely affects the shielding gas coverage, leading to defects such as weld oxidation and porosity. The assessment can be performed using the spatter distribution density index extracted from visual sensor data sub-streams, combined with a pre-established simulation model of the shielding gas flow field using computational fluid dynamics (CFD) methods. For example, abnormally concentrated or deviated spatter distribution may indicate disturbance to the shielding gas flow field. By comparing the actual spatter distribution with the spatter distribution predicted by the simulation model under normal flow field conditions, the probability of nozzle blockage and the degree of flow field distortion can be quantified. Furthermore, analyzing the gas pressure or velocity distribution at the nozzle outlet can also aid in the assessment.

[0118] Identifying the potential risk level of insufficient protective gas flow or leakage involves analyzing the energy spectrum characteristics of high-frequency noise in the arc stability quantification index to determine whether there is a problem with insufficient flow or leakage in the protective gas supply system. For example, when the protective gas flow is insufficient or leakage occurs, the plasma environment around the arc changes, resulting in specific high-frequency noise patterns in the arc voltage or current signal. By analyzing the energy spectrum of these high-frequency noises, characteristics related to the protective gas problem can be identified, and the potential risk level can be assessed accordingly.

[0119] This method integrates multiple independent health indicators, including the calculated real-time equivalent wear of the conductive nozzle, the assessed probability of protective nozzle blockage and the degree of local flow field distortion, and the identified potential risk level of insufficient protective gas flow or leakage, to generate a knowledge label that comprehensively reflects the overall health status of the welding equipment. The fusion can be achieved by assigning weights to each indicator based on expert experience and performing a weighted sum; alternatively, a fuzzy logic reasoning system can be used to fuse the fuzzy values ​​of each indicator; or machine learning algorithms, such as support vector machines or neural networks, can be used to train on historical fault data to learn how to map multiple indicators to a comprehensive health status label.

[0120] Furthermore, the step of calculating the real-time equivalent wear of the conductive tip adopts the following wear evolution model: .in, for Real-time equivalent wear at any given moment. This represents the initial wear amount, which is typically set to 0 for brand-new parts; when initialized based on test data, the initial wear amount is the measured value. for The wire feed current at any given moment, The nominal wire feed current is set according to the specific welding process specifications and is a known constant. for The arc voltage at any given time is obtained through sensor measurement. The wear rate coefficient represents the amount of wear caused by the product of current and voltage per unit time. Its value typically ranges from 0.1 to 2.0 μm and can be obtained through experimental calibration or by measuring the actual wear amount after operating under specific conditions (current and voltage) for a period of time. The current nonlinearity influence index represents the degree of nonlinearity in the effect of wire feed current fluctuations on wear. When When =1, the effect is linear; When the value is greater than 1, the positive deviation (excessive current) has a stronger accelerating effect on wear. Its value range is usually 1.0~2.5, which is obtained by fitting experimental data through wear experiments under different current fluctuation amplitudes.

[0121] The wear evolution model based on physical mechanisms estimates the initial wear amount of the conductive tip. As a benchmark, and based on this, by adjusting the wire feeding current during the welding process With nominal wire feed current The deviation between them, and the arc voltage Real-time monitoring and integral accumulation are performed to dynamically calculate the increase in wear. The deviation of the wire feed current is normalized by its ratio to the nominal current and then processed through a nonlinear influence index. To characterize its effect on accelerating or slowing down wear;

[0122] Wear rate coefficient These dynamic influencing factors are then converted into actual wear accumulation.

[0123] As an example, suppose that during a multi-segment weld welding task on a complex workpiece, the robot controller receives a weld segment switching command. It obtains the current welding status feature vector; for example, the arc stability quantification index obtained through sensor fusion is 0.88, the molten pool width is 5.2 mm, and the spatter distribution density is 0.005.

[0124] The target weld segment's specifications are obtained from the task planning module; for example, the weld segment is a fillet weld, the material is stainless steel, and the plate thickness is 4mm. This information is input into an adaptive welding process parameter model obtained through online incremental learning. The multilayer perceptron network within this model, such as a neural network with three hidden layers, receives these inputs, performs calculations, and outputs a set of optimized welding process parameters, such as: welding current 200A, welding voltage 22V, welding speed 7mm / s, wire feed speed 9m / min, welding torch tilt angle 10°, oscillation amplitude 1.5mm, and oscillation frequency 4Hz. Using the industrial robot's kinematic model, such as a kinematic model based on Denavit-Hartenberg parameters, and preset welding timing constraints, such as 0.5 seconds of pre-blowing shielding gas before arc ignition and 1 second of post-blowing shielding gas after arc termination, this set of optimized process parameters is solved along with the spatial path points of the target weld segment. Finally, a series of instructions executable by the robot controller are generated.

[0125] like Figure 2 As shown, in some other embodiments, this application proposes an intelligent control system for industrial robot welding, comprising:

[0126] The instruction response and initial control module is used to call the initial welding process parameter model based on the current workpiece specifications and generate the first welding control instruction sequence after the welding task is started.

[0127] The welding execution and data acquisition module is used to control the welding torch to perform welding operations according to the first welding control command sequence, and to simultaneously acquire multi-source sensor data streams of the welding process to obtain real-time sensor data streams.

[0128] The multi-source feature fusion extraction module is used to extract fusion features from the real-time sensor data stream to generate a feature vector of the current welding status. This module includes: a data stream separation unit, used to separate the arc sensing data sub-stream and the visual sensing data sub-stream from the real-time sensor data stream; an arc feature analysis unit, used to perform time-frequency domain analysis on the arc sensing data sub-stream to extract arc stability quantification indicators and wire feeding dynamic response indicators; a visual feature parsing unit, used to perform molten pool region segmentation and feature parsing on the visual sensing data sub-stream to extract molten pool geometric morphology indicators and spatter distribution density indicators; and a feature splicing unit, used to perform spatiotemporal alignment and vectorization splicing of the arc stability quantification indicators, wire feeding dynamic response indicators, molten pool geometric morphology indicators, and spatter distribution density indicators as the current welding status feature vector.

[0129] The situation deviation judgment and model learning module is used to determine that there is a situation deviation between the current welding situation feature vector and the benchmark welding situation feature vector corresponding to the initial welding process parameter model. Based on the current welding situation feature vector, the module performs online incremental learning on the initial welding process parameter model to obtain an adaptive welding process parameter model. The online incremental learning on the initial welding process parameter model is used to realize the closed-loop adaptive adjustment of the welding process and the implicit representation of the health status of key components.

[0130] The control module is used to generate an optimized welding control command sequence based on the adaptive welding process parameter model, and then send the optimized welding control command sequence to the robot controller and welding torch to execute the adaptive welding operation.

[0131] The above are merely embodiments of this application and are not intended to limit the scope of protection of this application. Various modifications and variations can be made to this application by those skilled in the art. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of this application should be included within the scope of protection of this application.

Claims

1. A method for intelligent control of industrial robot welding, characterized in that, The method comprises the following steps: After the welding task is started, an initial welding process parameter model is called based on the current workpiece specification to generate a first welding control instruction sequence; The welding gun is controlled to perform welding operation according to the first welding control instruction sequence, and a multi-source sensing data stream of the welding process is synchronously collected to obtain a real-time sensing data stream; Fusion feature extraction is performed on the real-time sensing data stream to generate a current welding situation feature vector, including: separating an arc sensing data substream and a visual sensing data substream from the real-time sensing data stream; performing time-frequency domain analysis on the arc sensing data substream to extract an arc stability quantitative index and a wire feeding dynamic response index; performing molten pool region segmentation and feature analysis on the visual sensing data substream to extract a molten pool geometric shape index and a spatter distribution density index; and performing spatiotemporal alignment and vectorization splicing on the arc stability quantitative index, the wire feeding dynamic response index, the molten pool geometric shape index and the spatter distribution density index to obtain the current welding situation feature vector; When there is a situation deviation between the current welding situation feature vector and a benchmark welding situation feature vector corresponding to the initial welding process parameter model, online incremental learning is performed on the initial welding process parameter model based on the current welding situation feature vector to obtain an adaptive welding process parameter model, wherein the online incremental learning on the initial welding process parameter model is used to realize closed-loop adaptive adjustment of the welding process and implicit representation of the health status of the key components; Based on the adaptive welding process parameter model, an optimized welding control instruction sequence is generated, and the optimized welding control instruction sequence is sent to the robot controller and the welding gun to perform adaptive welding operation.

2. The intelligent control method of industrial robot welding according to claim 1, characterized in that, The method further comprises the following steps: When it is determined that there is no situation deviation between the current welding situation feature vector and the benchmark welding situation feature vector corresponding to the initial welding process parameter model, the current welding situation feature vector is stored in a historical normal working condition feature library; Based on all the feature vectors in the historical normal working condition feature library, the statistical distribution parameters of the benchmark welding situation feature vector are dynamically updated to optimize the representation range of the benchmark welding situation feature vector to normal working condition fluctuations; After receiving a welding seam section switching instruction or a process parameter re-adjustment instruction, a welding control instruction sequence is generated based on the updated benchmark welding situation feature vector through the initial welding process parameter model.

3. The intelligent control method of industrial robot welding as claimed in claim 1, wherein, The method further comprises the following steps: The current welding situation feature vector is input into a process quality evaluation teacher model in a pre-constructed dual-path collaborative knowledge distillation network to generate a process quality knowledge label, wherein the dual-path collaborative knowledge distillation network comprises the process quality evaluation teacher model and an equipment health evaluation teacher model, and the process quality evaluation teacher model is obtained through adversarial generation training based on historical high-quality welding seam process data and is used to evaluate the compliance of the current welding process with the high-quality process standard; The current welding situation feature vector is input into the equipment health evaluation teacher model to generate an equipment health knowledge label, wherein the equipment health evaluation teacher model is obtained through time series prediction training based on the failure mechanism of the key components and the multi-sensor fusion data and is used to evaluate the comprehensive health status of the electrode, the shielding nozzle and the gas supply system. calculating a first loss gradient between the process quality knowledge label and the current welding situation feature vector; calculating a second loss gradient between the equipment health knowledge label and the current welding situation feature vector; determining a vector projection relationship and an included angle between the first loss gradient and the second loss gradient, wherein the vector projection relationship and the included angle represent a coupling contribution degree between process quality degradation and equipment health loss; when it is determined that a projection component of the first loss gradient on the second loss gradient is greater than or equal to zero, a module length of the first loss gradient is less than a preset process fluctuation threshold, and the included angle is less than a preset coupling threshold, it is determined that there is no situation deviation between the current welding situation feature vector and the reference welding situation feature vector, wherein the absence of the situation deviation represents that welding process fluctuation is mainly caused by acceptable random disturbance; when it is determined that the projection component of the first loss gradient on the second loss gradient is less than zero, or the module length of the first loss gradient is greater than or equal to the preset process fluctuation threshold, or the included angle is greater than or equal to the preset coupling threshold, it is determined that there is a situation deviation between the current welding situation feature vector and the reference welding situation feature vector, wherein the existence of the situation deviation represents that the welding process exists systematic deviation caused by equipment performance degradation or abnormal working condition.

4. The intelligent control method of industrial robot welding as claimed in claim 3, characterized in that, based on the current welding situation feature vector, performing online incremental learning on the initial welding process parameter model to obtain an adaptive welding process parameter model, including: determining a process optimization dominance coefficient and a health compensation dominance coefficient according to a module length ratio of the first loss gradient and the second loss gradient; weighting the first loss gradient based on the process optimization dominance coefficient to obtain a process optimization loss component; weighting the second loss gradient based on the health compensation dominance coefficient to obtain a health compensation loss component; performing vector synthesis on the process optimization loss component and the health compensation loss component to obtain a fusion optimization direction vector; updating a weight parameter of the initial welding process parameter model by using a back propagation algorithm according to the fusion optimization direction vector to obtain the adaptive welding process parameter model; wherein the method further includes: based on historical data of the health compensation loss component accumulated in the learning process of the adaptive welding process parameter model, constructing a key component residual service life prediction model; when the predicted residual service life is lower than a preset threshold, generating a visual warning report containing specific component identification, recommended replacement time and expected risk, and pushing it to a production management system and a maintenance terminal; when a maintenance completion confirmation signal is received from the production management system, resetting the health state representation parameters related to the replaced components and partially rolling back the weight parameters of the adaptive welding process parameter model to a general reference state.

5. The intelligent control method of industrial robot welding as claimed in claim 4, characterized in that, determining a process optimization dominance coefficient and a health compensation dominance coefficient according to a module length ratio of the first loss gradient and the second loss gradient, including: obtaining a first module length of the first loss gradient and a second module length of the second loss gradient; calculating a first ratio and a second ratio of the first module length and the second module length; inputting the first ratio and the second ratio into a Sigmoid activation function for normalization and smoothing processing to obtain the process optimization dominance coefficient and the health compensation dominance coefficient.

6. The intelligent control method of industrial robot welding as claimed in claim 3, characterized in that, The current welding situation feature vector is input into a process quality evaluation teacher model in a pre-constructed two-way collaborative knowledge distillation network to generate a process quality knowledge label, including: The molten pool geometry index is input into the first sub-network, the deep features of the molten pool image are extracted through the convolutional neural network, and the weld forming quality prediction score is obtained by regression; The arc stability quantitative index and the spatter distribution density index are input into the second sub-network, the dynamic process features are extracted through the time series neural network, and the welding process stability score is obtained by regression; According to the weld forming quality prediction score and the welding process stability score, the process quality knowledge label is generated by weighted summation and Softmax function mapping.

7. The intelligent control method of industrial robot welding as claimed in claim 3, characterized in that, The current welding situation feature vector is input into the equipment health evaluation teacher model to generate the equipment health knowledge label, including: Based on the wire feeding dynamic response index and the preset wire feeding resistance-wear amount mapping relationship, the real-time equivalent wear amount of the electrode nozzle is calculated; Based on the spatter distribution density index and the simulation model of the protective gas flow field, the clogging probability and the local flow field distortion degree of the protective nozzle are evaluated; Based on the energy spectrum characteristics of the high-frequency noise in the arc stability quantitative index, the potential risk level of insufficient or leakage of protective gas flow is identified; The real-time equivalent wear amount, the clogging probability and the local flow field distortion degree, and the potential risk level are fused by multi-objective decision to generate the equipment health knowledge label representing the comprehensive health state.

8. The intelligent control method of industrial robot welding as claimed in claim 7, characterized in that, The step of calculating the real-time equivalent wear amount of the electrode nozzle adopts the following wear evolution model: ； wherein, is the real-time equivalent wear amount at the time instant, is the initial wear amount, is the wire feed current at the time instant, is the nominal wire feed current, is the arc voltage at the time instant, is the wear rate coefficient, is the current non-linearity influence index.

9. The intelligent control method of industrial robot welding as claimed in claim 1, characterized in that, An adaptive welding process parameter model is used to generate an optimized welding control instruction sequence, including: The current welding situation feature vector and the specification characteristics of the target weld section are jointly input into the adaptive welding process parameter model; Through the multi-layer perception network in the adaptive welding process parameter model, the optimized welding process parameter set is mapped, including welding current, welding voltage, welding speed, wire feeding speed, welding gun attitude angle and swing parameters; According to the robot kinematics model and the welding time sequence constraint, the welding process parameter set is calculated into an optimized welding control instruction sequence executable by the robot controller, including spatial path points and process parameter points.

10. An intelligent control system for industrial robot welding for implementing the method according to any one of the preceding claims 1 - 9, characterized in that, The system includes: An instruction response and initial control module for generating a first welding control instruction sequence based on the current workpiece specification after the welding task is started; A welding execution and data acquisition module for controlling the welding gun to perform welding operation according to the first welding control instruction sequence and synchronously collecting multi-source sensing data stream of the welding process to obtain real-time sensing data stream; A multi-source feature fusion extraction module for fusion feature extraction of the real-time sensing data stream to generate the current welding situation feature vector, The situation deviation determination and model learning module is configured to, when it is determined that there is a situation deviation between the current welding situation feature vector and the reference welding situation feature vector corresponding to the initial welding process parameter model, perform online incremental learning on the initial welding process parameter model based on the current welding situation feature vector to obtain an adaptive welding process parameter model. The online incremental learning on the initial welding process parameter model is configured to realize closed-loop adaptive adjustment of the welding process and implicit representation of the health state of the key component. The control module is configured to generate an optimized welding control instruction sequence based on the adaptive welding process parameter model, and send the optimized welding control instruction sequence to the robot controller and the welding torch to perform adaptive welding operation.

Images