Welding deviation compensation control method for mechanical arm based on real-time tracking of weld

Abstract

This invention relates to a robotic arm welding deviation compensation control method based on real-time weld seam tracking. By collecting real-time welding process parameters and weld seam tracking deviations, multi-source data is fused, standardized, and feature extracted to construct a parameter semantic graph reflecting the mechanism of the process. A sparse subgraph is dynamically generated based on working condition information. Combining graph structure analysis and attention propagation, dominant parameter nodes and their linkage patterns are identified. Behavioral templates are matched to obtain the inter-parameter compensation control logic, generating compensation quantities with semantic consistency constraints to drive the robotic arm to achieve posture correction and complete closed-loop feedback. This method achieves multi-parameter welding coordination, transparent control logic, and highly adaptive quality control, improving weld seam formation consistency and the intelligence level of the welding process.

Description

Technical Field

[0001] This invention relates to the field of intelligent control technology for welding processes, and in particular to a robotic arm welding deviation compensation control method based on real-time weld seam tracking. Background Technology

[0002] Real-time weld seam tracking and deviation compensation control technology in robotic arm welding processes has become a key aspect of improving intelligent manufacturing and ensuring stable welding quality. Existing weld seam tracking and compensation schemes mostly use the positional deviation of the robotic arm's end effector as input, directly mapping the deviation characteristics collected by sensors into displacement or angle correction commands to control the robotic arm to perform trajectory correction, thereby achieving real-time compensation for weld seam tracking. These traditional methods have the advantages of simplicity and fast response, and have been widely used in automated production lines such as robotic arc welding and laser welding.

[0003] With the development of intelligent welding processes, the industry is gradually incorporating typical welding process parameters such as current, voltage, speed, and wire feed rate into the design of weld tracking control and compensation strategies. For example, by introducing PID adaptive controllers based on molten pool images or using multimodal signal fusion methods, some solutions can achieve auxiliary adjustment of compensation amounts by process parameters. However, most mainstream technologies in the industry still treat welding process parameters as background variables independent of the compensation decision-making process, failing to truly establish a native coupling relationship between compensation strategies and process parameters.

[0004] Existing patents and papers typically employ the following characteristics in their technical solutions: First, they widely use single-channel signal feedback (such as weld centerline offset and molten pool width) as the basis for compensation calculation, lacking the ability to model multi-dimensional collaborative processes. Second, even when some solutions introduce parameter optimization algorithms, the parameter tuning process often uses fuzzy rule bases, static weighted coefficients, LSTM, and other time-series predictors, limiting the relationships between process parameters to weight adjustment or time-series prediction, making it difficult to accurately capture the inherent physical collaborative laws of parameters such as current, speed, and voltage under specific welding scenarios. These methods are suitable for welding scenarios with stable parameters and simple operating conditions, but in intelligent welding systems with diverse materials, strong environmental disturbances, and high real-time performance requirements, they exhibit problems such as response lag and insufficient compensation accuracy.

[0005] Overall, the current field of real-time weld tracking and compensation control has the following key shortcomings:

[0006] Poor coordination between compensation strategies and process parameters: Mainstream compensation control schemes often design trajectory correction actions separately from welding process parameters, lacking consideration for dynamic response mechanisms of parameters such as current, voltage, speed, and wire feed rate. In actual system operation, the lack of physical self-consistency between compensation amount and parameter adjustment easily leads to weld quality fluctuations and makes it difficult to achieve highly intelligent and automated welding process control.

[0007] The complex coupling relationship between process parameters is difficult to model: In complex welding scenarios, there is nonlinear synergy among multiple parameters such as the heat input dominated by current and speed, and the stability of the molten pool and voltage coordination. Existing methods are unable to effectively capture the implicit correlation and response path between parameters, resulting in a lack of targeted and adaptive compensation mechanisms.

[0008] The lack of a dynamic semantic association-based intelligent compensation generation mechanism: the existing compensation quantity generation methods are mainly based on numerical calculations or single mappings, lacking in-depth integration of high-level information such as parameter semantic relationships and process mechanisms, making it difficult to form an integrated intelligent control link of "parameter-compensation".

[0009] The balance between intelligence and interpretability is difficult to achieve: Some technologies based on deep learning and weighted multimodal fusion have improved compensation effects, but the models are complex and have poor interpretability, posing an obstacle to practical industrial application. Traditional rule-based solutions, while helpful for understanding, rely heavily on human experience and are difficult to dynamically adapt to new working conditions.

[0010] Based on the above analysis, the industry urgently needs a novel welding compensation control technology that can overcome the disconnect between "parameters and compensation," achieve dynamic and native synergy between compensation strategies and process parameters, and possess high interpretability and adaptability, in order to significantly improve welding quality and system intelligence under complex working conditions. This invention addresses the core technological shortcoming of existing real-time weld tracking compensation control—the poor synergy between compensation strategies and welding process parameters—and, considering the new demands of dynamic working condition changes on the compensation strategy response mechanism, proposes a dynamic coupling generation method for compensation quantities driven by a semantic graph of process parameters. This provides a higher quality and more reliable technological foundation for industrial intelligent welding. Summary of the Invention

[0011] This application provides a robotic arm welding deviation compensation control method based on real-time weld seam tracking, which aims to solve one of the problems or issues of the prior art mentioned in the background section.

[0012] The robotic arm welding deviation compensation and control method based on real-time weld seam tracking provided in this application specifically includes:

[0013] S1: Obtain the process parameter flow and weld tracking deviation feature vector during the welding process. The process parameter flow includes welding current, voltage, speed, wire feed rate, oscillation amplitude, and shielding gas flow rate.

[0014] S2: Based on the attribute database, weld quality dataset and welding process knowledge base, each discrete or continuous parameter in the process parameter flow is modeled as a graph node, and semantic edges containing the coupling relationship between nodes including heat input, molten pool stability, spatter suppression and heat-affected zone are defined to construct a process parameter semantic graph.

[0015] S3: Based on the material thickness, ambient temperature, and weld curvature, prune weak connection semantic edges with confidence levels below the threshold in the semantic graph of the process parameters to generate a working condition adaptation sparse subgraph.

[0016] S4: Search the sparse subgraph of the working condition adaptation for the minimum closed-loop structure that covers all active parameter nodes and satisfies semantic connectivity constraints to obtain the parameter cooperative subgraph.

[0017] S5: Perform graph attention propagation calculation on the parameter collaborative subgraph to identify the parameter nodes that play a dominant regulatory role and their semantic neighborhoods, and extract implicit collaborative patterns.

[0018] S6: Based on the implicit collaborative pattern, match the corresponding behavior template from the pre-set compensation knowledge base. The behavior template defines the proportional constraint relationship and response timing difference between the dominant compensation channel and the auxiliary compensation channel.

[0019] S7: Convert the behavior template into a structured instruction, and perform compensation calculation based on the structured instruction. The compensation amount includes a velocity component and a current component with deterministic proportional constraints.

[0020] S8: Control the robotic arm to perform posture correction actions according to the process semantic compensation amount, and collect new weld tracking deviation feature vectors after execution to update the process parameter flow, forming closed-loop feedback control.

[0021] The robotic arm welding deviation compensation control method based on real-time weld seam tracking provided in this application has the following beneficial effects:

[0022] (1) By constructing a semantic initialization module for process parameters and introducing a time-series graph structure based on domain knowledge, this scheme effectively overcomes the problem of lack of synergy caused by isolated parameter modeling in traditional welding compensation methods. Existing technologies often treat parameters such as current, voltage, and speed as independent control variables, relying on empirical rules or data-driven models for numerical mapping compensation, which makes it difficult to reflect their dynamic correlation under real thermo-mechanical coupling. This invention models each process parameter as a graph node with physical meaning, and establishes an interpretable relational topology through non-numerical semantic edges such as "heat input dominance" and "molten pool stability constraint", so that the interaction mechanism between parameters can be explicitly expressed. During operation, the system combines real-time context (such as material thickness, ambient temperature, and weld curvature) to perform dynamic pruning and subgraph activation operations on the graph, retaining only the smallest closed-loop subgraph with strong semantic connectivity and high confidence under the current operating conditions, thereby accurately identifying the most representative parameter synergy mode of the process. This mechanism significantly improves the consistency between the compensation strategy and the actual welding process, avoiding overcompensation or undercompensation caused by insufficient model generalization ability in complex working conditions under traditional methods, and realizing the transformation from "passive correction" to "feedforward reasoning".

[0023] (2) By leveraging the triple graph operation logic in the lightweight graph neural network encoder—dynamic pruning, subgraph activation, and logic extraction—this solution achieves a balance between high response sensitivity and strong interpretability, overcoming the inherent contradiction between real-time performance and decision transparency in existing compensation systems. Unlike technical paths that rely on multimodal sensor fusion, LSTM sequence prediction, or online reinforcement learning training, this invention does not perform end-to-end regression or introduce external perception redundancy. Instead, based on a pre-built compensation logic knowledge base, it automatically identifies the dominant control nodes and their semantic neighborhood structures through attention propagation analysis of activated subgraphs, thereby matching behavioral templates with clear process meanings. These templates embed proportional constraints and response timing differences between parameters, and the generated compensation quantities naturally possess coupling logic consistency. For example, the collaborative mechanism of speed fine-tuning as the main component and current-assisted following makes the actions of the robotic arm execution layer naturally conform to the laws of welding thermodynamics. The entire inference process does not require complex parameter tuning or rely on large-scale training data, possessing good engineering deployment adaptability. While ensuring a response speed of hundreds of milliseconds, it provides a clear decision tracing path, significantly enhancing the trust of on-site operators and the maintainability of the system.

[0024] (3) This scheme achieves a fundamental shift in the compensation decision-making paradigm by upgrading process parameters from "control variables" to "semantic entities," that is, from the traditional "numerical mapping" to "relational reasoning," thereby constructing a statically interpretable and dynamically evolving closed-loop adaptive system. The semantic graph not only carries the fusion expression of material thermophysical properties, historical high-quality weld data, and welding mechanism knowledge, but also supports incremental updates and weight decay adjustments based on newly accumulated process experience, possessing continuous optimization capabilities. Compared with methods based on PID adaptive adjustment, fuzzy rule bases, or parameter coupling degree matrix calculations, this technical route breaks away from the dependence on fixed mathematical models or manual experience thresholds, and can maintain stable reasoning performance under different material combinations, joint forms, and spatial locations. Especially when facing non-standard working conditions or process disturbances, the system can quickly locate the adaptation strategy through semantic distance constraints and minimum closed-loop subgraph retrieval mechanism, significantly reducing the debugging cycle and scrap rate. The overall architecture has good modularity and expansion potential, and can be seamlessly integrated into intelligent manufacturing systems, supporting the accumulation, reuse, and cross-scenario migration of welding process knowledge.

[0025] The combined effect of the above-mentioned technical means enables this solution to achieve deep coupling between the compensation strategy and the essential characteristics of the welding process without increasing hardware costs or relying on external sensor fusion. This significantly improves the stability of welding quality, the intelligence of system response, and the interpretability of engineering applications, providing a brand-new technical paradigm for high-precision automated welding. Attached Figure Description

[0026] Figure 1 This is the main flowchart of a robotic arm welding deviation compensation control method based on real-time weld seam tracking.

[0027] Figure 2 This is a sub-flowchart of a robotic arm welding deviation compensation control method based on real-time weld seam tracking.

[0028] Figure 3 This is another sub-flowchart of the robotic arm welding deviation compensation control method based on real-time weld seam tracking. Detailed Implementation

[0029] Embodiments of the present invention are described in detail below, examples of which are shown in the accompanying drawings, wherein the same or similar reference numerals denote the same or similar elements or elements having the same or similar functions throughout. The embodiments described below with reference to the accompanying drawings are exemplary and are only used to explain the present invention, and should not be construed as limiting the present invention.

[0030] The following disclosure provides many different embodiments or examples for implementing different structures of the invention. To simplify the disclosure, specific examples of components and arrangements are described below. Of course, these are merely examples and are not intended to limit the invention. Furthermore, reference numerals and / or letters may be repeated in different examples; such repetition is for simplification and clarity and does not in itself indicate a relationship between the various embodiments and / or arrangements discussed.

[0031] like Figure 1 As shown, this application provides a robotic arm welding deviation compensation and control method based on real-time weld seam tracking, specifically including:

[0032] S1: Obtain the process parameter flow and weld tracking deviation feature vector during the welding process. The process parameter flow includes welding current, voltage, speed, wire feed rate, oscillation amplitude, and shielding gas flow rate.

[0033] S2: Based on the attribute database, weld quality dataset and welding process knowledge base, each discrete or continuous parameter in the process parameter flow is modeled as a graph node, and semantic edges containing the coupling relationship between nodes including heat input, molten pool stability, spatter suppression and heat-affected zone are defined to construct a process parameter semantic graph.

[0034] S3: Based on the material thickness, ambient temperature, and weld curvature, prune weak connection semantic edges with confidence levels below the threshold in the semantic graph of the process parameters to generate a working condition adaptation sparse subgraph.

[0035] S4: Search the sparse subgraph of the working condition adaptation for the minimum closed-loop structure that covers all active parameter nodes and satisfies semantic connectivity constraints to obtain the parameter cooperative subgraph.

[0036] S5: Perform graph attention propagation calculation on the parameter collaborative subgraph to identify the parameter nodes that play a dominant regulatory role and their semantic neighborhoods, and extract implicit collaborative patterns.

[0037] S6: Based on the implicit collaborative pattern, match the corresponding behavior template from the pre-set compensation knowledge base. The behavior template defines the proportional constraint relationship and response timing difference between the dominant compensation channel and the auxiliary compensation channel.

[0038] S7: Convert the behavior template into a structured instruction, and perform compensation calculation based on the structured instruction. The compensation amount includes a velocity component and a current component with deterministic proportional constraints.

[0039] S8: Control the robotic arm to perform posture correction actions according to the process semantic compensation amount, and collect new weld tracking deviation feature vectors after execution to update the process parameter flow, forming closed-loop feedback control.

[0040] Step S1: Obtain the process parameter flow and weld tracking deviation feature vector during the welding process. The process parameter flow includes welding current, voltage, speed, wire feed rate, oscillation amplitude, and shielding gas flow rate. Specifically, it includes:

[0041] S1.1: Perform high-frequency sampling and protocol parsing processing on the analog signals and digital communication messages output by the welding power controller to extract the original electrical parameter stream containing the instantaneous value of welding current, the fluctuation value of arc voltage and the set value of wire feed rate, and obtain a standardized continuous process parameter set.

[0042] It should be noted that the real-time process parameter stream refers to a multi-dimensional process parameter data sequence continuously acquired during the welding process at a sampling frequency of no less than 100Hz and synchronized with timestamps. This data sequence includes continuous values ​​of parameters such as welding current, arc voltage, welding speed, wire feed rate, oscillation amplitude, and shielding gas flow rate on the time axis. Each parameter value is bound to a global timestamp at the sampling time, used to characterize the dynamic changes of the welding process on a millisecond-level time scale. The weld tracking deviation feature vector refers to a three-dimensional numerical vector obtained by vectorizing and encapsulating the lateral offset, height deviation, and attitude angle error of the weld centerline relative to the theoretical trajectory. This feature vector is used to quantify the degree of spatial deviation between the current welding torch posture and the ideal weld trajectory.

[0043] A high-frequency synchronous sampling channel is established for the output interface of the welding power controller to be connected in parallel to its analog signal output terminal and digital communication port. The sampling channel frequency is configured to cover the sampling rate of the transient change bandwidth of the arc based on the shortest dynamic response period of the welding process.

[0044] The acquired analog current and voltage signals are processed by front-end conditioning, low-noise operational amplifiers and active filters are used to suppress high-frequency interference components, and continuous analog waveforms are converted into discrete digital sequences through a high-precision analog-to-digital converter module.

[0045] The digital communication messages are parsed frame by frame according to the communication protocol provided by the welding power supply manufacturer. The protocol stack decoding module is called to extract control parameters such as wire feed rate setpoint. The parsing results are then timestamped together with the instantaneous waveform data obtained from the aforementioned analog-to-digital conversion.

[0046] The sliding window extreme value detection method is used to identify the peak and valley values ​​in each sampling period of the current waveform data, and the instantaneous current value is calculated accordingly; short-time mean filtering is applied to the arc voltage waveform to obtain the voltage fluctuation value that reflects the stability of the arc.

[0047] The instantaneous current value, arc voltage fluctuation value, and wire feed rate setting value are normalized and mapped according to the set unified dimensions and numerical precision to form a continuous set of process parameters in ternary vector format, and a sampling timestamp and signal source label are bound to each vector.

[0048] Through the above-mentioned multi-source signal acquisition, protocol parsing, feature calculation and normalization processing, the original output of the power controller is transformed into a standardized continuous set of process parameters that are consistent in timing and dimension and can be directly used in the construction of subsequent graph nodes, thereby achieving high-precision and low-latency acquisition of electrical process parameters.

[0049] For example, in a pulse welding power supply application scenario with a rated output current of 500A, the analog signal sampling channel is set to a sampling rate of 50kHz, the analog-to-digital conversion resolution is 16bit, and the front-end filter cutoff frequency is 5kHz; the digital communication interface adopts the CANopen protocol, with a baud rate of 1Mbps and a message parsing interval of 1ms. The instantaneous current value obtained by peak-valley detection of the filtered current waveform is 320A, and the value after normalization mapping is 0.64; the voltage fluctuation value is 1.2V after short-time average filtering, and the normalized result is 0.24; the wire feed rate setpoint is parsed to 12.0m / min, and the value after dimensional conversion and normalization is 0.48. The ternary continuous process parameter vector is represented as [0.64, 0.24, 0.48], which is used for subsequent S1.2 steps to align with the robot arm motion control parameters with timestamps, realizing the synchronous fusion of electrical and motion parameters. Under these conditions, the extraction delay of instantaneous current value has been verified to be less than 0.5ms, and the calculation window width of short-time average voltage fluctuation value is 2ms. This can significantly improve the response speed of the welding deviation compensation link to the dynamic changes of the arc, and realize the efficient coordination of electrical parameters and mechanical motion control signals in the welding process.

[0050] S1.2: Based on the continuous process parameter set, the pulse commands fed back by the robotic arm motion controller and the flow meter sensor readings are timestamped and quantized to integrate the welding speed vector, the discrete level of the swing amplitude and the real-time reading of the protective gas flow rate to obtain a complete multi-dimensional process parameter flow.

[0051] Based on the continuous set of process parameters obtained in the previous step, including the instantaneous value of welding current, the fluctuation value of arc voltage, and the set value of wire feed rate, a unified time synchronization reference is established using the multi-channel pulse command signal fed back by the robotic arm motion controller and the real-time sampling reading of the protective gas flow meter sensor as inputs. The pulse command channel from the robotic arm control bus undergoes high-precision timestamp extraction processing, and the pulse frequency of the command is quantitatively decoded. Based on a preset step coefficient, the frequency value is converted into a linear velocity vector component, forming a preliminary welding speed vector dataset. The discrete gear command values ​​of the oscillating mechanism are processed through protocol parsing and quantization encoding, mapping different gears to corresponding oscillation amplitude value ranges, generating an oscillation amplitude parameter sequence and establishing a correlation with the velocity vector. The analog signal or digital communication data output by the flow meter sensor is corrected using a multi-point calibration formula; its conversion relationship can be defined by the following formula: ,in, To protect the real-time air flow rate, This is the original output of the flow meter. The scaling factor is obtained from the calibration process. Timestamp alignment is performed simultaneously with the generation of various parameters to ensure that the velocity vector, oscillation amplitude, and protective gas flow rate have consistent time indices at the same sampling time. The velocity, oscillation amplitude, and gas flow rate data corrected with the unified time reference are fused and encapsulated with the aforementioned continuous process parameters to form a complete process parameter stream with multiple parameter dimensions. Through a combination of timestamp alignment, quantization encoding, and physical quantity calculation, the original pulse and flow signals are transformed into standardized numerical parameters that can directly participate in multi-dimensional operating condition modeling, achieving consistency of the process parameter stream in both spatial components and time reference.

[0052] For example, in a six-axis welding robotic arm system, the end effector linear axis pulse frequency transmitted back by the robotic arm motion controller is 5000Hz, with a preset step factor of 0.02mm / pulse. This is converted to a welding speed vector magnitude of 100mm / s. The oscillating mechanism resolves to a stop value of 3, corresponding to an amplitude of 2.5mm. The original digital value of the protective gas flow meter is 250 count units. Based on the calibration factor K = 0.04L / (min·counts), the formula is applied... The calculated protective gas flow rate was 10 L / min. Under this operating condition, the timestamp synchronization module performed microsecond-level alignment on the three types of data, and the data encapsulation module combined the previous current, voltage, and wire feed rate data to construct a six-dimensional real-time process parameter vector {100, 2.5, 10, I, U, Vf}, where I, U, and Vf are the synchronized current, voltage, and wire feed rate values, respectively. Verification results show that the synchronization delay of the multi-dimensional process parameter stream on the system data bus does not exceed 2 ms, significantly improving the time consistency and parameter association accuracy of subsequent operating condition semantic graph construction.

[0053] S1.3: Using the weld pool image sequence captured by the laser structured light sensor or vision camera, sub-pixel level extraction and coordinate transformation are performed on the feature edges in the image to calculate the lateral offset, height deviation and attitude angle error of the weld centerline relative to the theoretical trajectory, and obtain a high-precision weld tracking deviation feature vector.

[0054] Based on the multi-dimensional process parameter stream after timestamp alignment and quantization encoding, the synchronous acquisition interface of the laser structured light sensor or high-resolution industrial vision camera is called to capture a continuous sequence of weld pool images during the robotic arm's execution of the welding trajectory.

[0055] The acquired original image sequence is input into a preset multi-scale edge detection algorithm based on gradient orientation histogram and Canny operator. The detected weld boundary is then refined at the sub-pixel level by combining wavelet multi-resolution feature decomposition to obtain a feature edge point set with higher geometric resolution than the original pixel grid.

[0056] A two-dimensional to three-dimensional coordinate system transformation operation defined by the calibration matrix is ​​performed on the feature edge point set. The spatial position vector of the edge point in the coordinate system of the welding workpiece is calculated by the camera intrinsic and extrinsic parameter calibration matrix and the laser structured light plane equation, and a real-time geometric model of the weld surface is established.

[0057] Based on the established real-time geometric model and the theoretical weld trajectory model, registration and matching are performed, and the least squares method is used to calculate the offset of the weld centerline relative to the theoretical trajectory in the lateral direction. Deviation in the height direction and attitude angle error The formula for calculating the lateral offset is: The formula for calculating the height deviation is: The formula for calculating attitude angle error is: , in the formula , , These are the measured horizontal, vertical, and attitude angle coordinates of the centerline point, respectively. , , These are the coordinates corresponding to the center line of the theoretical trajectory. This represents the number of sampling points.

[0058] The lateral offset, height deviation, and attitude angle error are combined and encapsulated into a three-dimensional deviation feature vector, which is then input into the subsequent spatiotemporal synchronization fusion module.

[0059] By using subpixel-level feature extraction, spatial coordinate transformation, and trajectory registration calculation, the multi-dimensional process parameter stream data from the previous step is transformed into precise quantitative features reflecting weld geometry and attitude deviations, enabling high-precision real-time acquisition of weld tracking deviations.

[0060] For example, in a welding scenario of Q235 steel with a plate thickness of 8mm and a bevel angle of 30 degrees, a laser line structured light 3D sensor was selected, with a working distance of 300mm, a resolution of 0.05mm, and a sampling frequency of 200Hz. This was combined with an industrial camera with a resolution of 1920×1200 pixels and a frame rate of 120fps to simultaneously acquire image data of the weld pool area. In the edge detection stage, the Canny operator's high and low thresholds were set to 80 and 160 respectively, the wavelet decomposition layer was 3 layers, and cubic spline interpolation was used for sub-pixel interpolation. In the coordinate transformation stage, the sensor calibration matrix was calibrated using the Tsai dual-plane method, with plane equation parameters a=0.0023, b=−0.0015, and c=1. During a certain welding time period, the lateral offset Δx = 0.12mm, the height deviation Δz = 0.08mm, and the attitude angle error Δθ = 0.5° were calculated for n=50 sampling points. After the output deviation feature vector [0.12, 0.08, 0.5] is fused by S1.4, the real-time performance and compensation accuracy of the robotic arm posture correction are greatly improved, and the overshoot phenomenon in the weld oscillation section is effectively suppressed in the subsequent closed-loop control.

[0061] S1.4: According to the preset sampling period window, the multidimensional process parameter stream and the weld tracking deviation feature vector are spatiotemporally synchronized and filtered for outliers to eliminate sensor noise interference and unify the data timing reference, thereby obtaining a working condition perception data frame with time synchronization mark.

[0062] S1.5: Based on the working condition perception data frame, normalize and structure the various parameter data to generate standardized data objects that meet the input format requirements of graph nodes, and obtain the process parameter flow and weld tracking deviation feature vector used to drive the subsequent construction of the process parameter semantic graph.

[0063] Step S2: Based on the attribute database, weld quality dataset, and welding process knowledge base, each discrete or continuous parameter in the process parameter flow is modeled as a graph node, and semantic edges are defined between nodes, including coupling relationships such as heat input, weld pool stability, spatter suppression, and heat-affected zone, to construct a process parameter semantic graph. Specifically, this includes:

[0064] S2.1: Obtain welding current, voltage, speed, wire feed rate, oscillation amplitude, and shielding gas flow rate from the process parameter stream. Based on the attribute database, identify the data type and standardize the dimensions of various parameters to generate a standardized set of process parameters that meets the input format requirements of the graph node.

[0065] Based on the process parameter stream generated by S1.5 and its structured and encapsulated standardized data objects, data type and physical attribute discrimination processing is performed on process data such as welding current, voltage, speed, wire feed rate, oscillation amplitude and shielding gas flow rate. The data type or discrete data category of each parameter is identified, and the parameter is matched and verified with the physical quantity category marked in the attribute database to ensure that the physical semantics of the parameter are consistent with the database entries.

[0066] Using the data type identification results as index keys, the corresponding dimensional reference values ​​and physical unit systems of the attribute database are retrieved, and the unit conversion coefficients and zero-point reference parameters used for standardization are extracted, laying a mathematical foundation for subsequent dimensional unification.

[0067] Based on the retrieved unit conversion factors, a linear proportional conversion process is performed on the continuous parameter values ​​to map the original parameter values ​​to the consistent dimensions of the International System of Units (SI). The formula is as follows:

[0068]

[0069] in, These are the parameter values ​​after dimensional standardization. To convert the proportionality factor, These are the original parameter values. These are the zero-point reference parameters.

[0070] For discrete parameters, a status code mapping is performed based on the status code table defined in the attribute database to convert the original gear value or switch status value into the corresponding physical quantity numerical expression. For example, the swing amplitude gear is mapped to the actual angle amplitude, and the gas flow valve position level is mapped to the standard gas volume flow rate per unit time.

[0071] Numerical range verification is performed on all parameters after dimensional unification. Outliers that exceed the physical limits of the working conditions defined in the attribute database are marked and removed to eliminate spurious data that may be caused by sensor drift or abnormal impact, thus ensuring the effectiveness of the standardization results.

[0072] By identifying data types, standardizing and mapping dimensions, converting discrete parameters to physical values, and removing outliers, the process parameter stream from the previous step is transformed into a standardized set of process parameters that meets the input format requirements of the graph nodes and has a unified physical unit and a valid numerical range, thus achieving consistency and comparability of input data in the graph node construction stage.

[0073] S2.2: Based on the standardized set of process parameters, each process parameter is instantiated into an independent graph node object using the statistical regularity of the weld quality dataset, and a node feature vector containing parameter name, real-time value and physical attribute label is assigned to it to construct a process parameter graph node set.

[0074] S2.3: Call the rule entries in the welding process knowledge base regarding heat input calculation, molten pool dynamics and spatter formation mechanism, and logically deduce the physical coupling relationship between any two nodes in the process parameter graph node set to generate a candidate semantic edge list containing heat input dominance, molten pool stability constraint, spatter suppression correlation and heat-affected zone superposition coupling type.

[0075] Based on the process parameter graph node set obtained in step S2.2 and the interface for calling the welding process knowledge base, physical attribute labels are parsed for each pair of process parameter nodes to clarify the range and dimensional characteristics of the physical quantities corresponding to the nodes. For the parsed node physical attributes, matching rule entries in the welding process knowledge base are retrieved. These rule entries cover heat input calculation formulas, molten pool dynamics models, and spatter formation mechanisms, with the node physical attribute matching results used as the rule filtering index. According to the heat input calculation rules, node pairs containing heat supply dominance relationships are input into the heat input evaluation function. By calculating the heat energy input per unit time and analyzing its deviation from the baseline value, it is determined whether the node pair meets the generation conditions for heat input dominance semantic edges. This evaluation can be based on the formula:

[0076]

[0077] in This refers to the heat input per unit length of weld. The coefficient of performance is the thermal efficiency. For welding current, Arc voltage For welding speed, based on the rules of molten pool dynamics, molten pool stability conditions are deduced for node pairs involving current, wire feed rate, and oscillation amplitude. The relationship between material surface tension, gravity, and electromagnetic contraction force is coupled with the real-time values ​​of the nodes to determine whether the node pair constitutes a molten pool stability constraint semantic edge under the current working conditions. Based on the rules of spatter formation mechanism, droplet transition mode is identified for node pairs involving current and voltage fluctuations, and whether they are spatter suppression-related semantic edge generation objects are determined based on the calculation results of the arc pressure to surface tension ratio. For node pairs containing heat-affected zone control, the dynamic changes of the depth and width of the heat-affected zone are solved using the heat diffusion equation and compared with the safety process curves stored in the knowledge base to determine whether they meet the generation conditions of the heat-affected zone superimposed coupling type. Node pairs that meet at least one type of physical coupling relationship obtained through the above multi-dimensional condition derivation are marked as candidate relationship objects, and their corresponding coupling type labels are combined to output a candidate semantic edge list.

[0078] By using rule-driven sequential logic deduction and physical quantity calculation, the combination relationship of process parameter nodes is formalized into candidate semantic edges with physical interpretation, ensuring physical consistency and interpretability in the semantic graph construction stage, and achieving input integrity and accuracy for subsequent weight assignment and topology assembly.

[0079] For example, under the conditions of a single-layer carbon steel plate thickness of 8mm, an ambient temperature of 25℃, a real-time welding current of 230A, a real-time arc voltage of 26V, a welding speed of 0.35m / min, and a thermal efficiency coefficient set to 0.85, the current and voltage nodes are input into the heat input formula. The calculated result is 14613.43 J / mm², which is far higher than the penetration reference heat input threshold of 12000 J / mm² for this material. According to the knowledge base rules, the heat input level corresponding to this value plays a dominant role in weld formation, thus generating a heat input-dominant semantic edge. Further, based on the molten pool dynamics rules, the current node and wire feed rate node are combined and input into the molten pool stability model to calculate the Reynolds number and surface tension coefficient of the instantaneous metal droplet transition. The obtained molten pool surface disturbance frequency is below 5 Hz, meeting the stability constraint conditions, thus generating a molten pool stability constraint semantic edge. The current and voltage fluctuation nodes are input into the spatter discrimination rules, and the droplet transition is determined to be a jet transition mode, significantly reducing the probability of spatter formation, thus generating a spatter suppression-related semantic edge. Therefore, under this operating condition, a candidate semantic edge list containing three edges with different physical attribute labels is output, serving as the basis for subsequent semantic edge weight initialization and graph assembly.

[0080] S2.4: Based on the candidate semantic edge list, bind interpretable domain rule labels to each semantic edge and initialize weight decay function parameters. Generate a complete set of semantic edges by quantitatively evaluating the confidence scores of each physical coupling relationship under different working conditions.

[0081] The initial input dataset based on the candidate semantic edge list contains multiple physical coupling relationship identifiers generated by step S2.3 and their corresponding relationship type information. For this candidate semantic edge list, label binding is performed on each semantic edge according to a predefined domain rule label library. Combining knowledge entries from the welding process knowledge base regarding heat input calculation, molten pool stability, spatter formation, and the evolution of the heat-affected zone, each semantic edge is precisely mapped to a rule label that reflects its physical essence, achieving a unique mapping relationship between edges and rules. For semantic edges with bound rule labels, initial weight decay function parameters are set according to the working condition adaptability evaluation model. This weight decay function describes the decay trend of the confidence level of the semantic edge under different working condition variables. The decay function type can be exponential, piecewise linear, or logistic. Taking the exponential decay function as an example, the confidence level calculation formula is:

[0082]

[0083] in, Let the initial confidence level of the semantic edge be given under standard operating conditions. The attenuation coefficient is... The characteristic value of the current operating condition. This represents the optimal operating condition feature value corresponding to the semantic edge. Based on this decay function, the dynamic confidence score of each semantic edge is calculated under a given set of operating conditions. The calculated dynamic confidence scores are then subjected to quantitative evaluation processing. A weight normalization strategy based on the analytic hierarchy process (AHP) is used to eliminate confidence biases of different dimensions and relation types, making the confidence scores of semantic edges with different physical relations comparable. Based on the quantitative evaluation, the confidence scores, along with the bound rule labels and decay function parameters, are encapsulated into a semantic edge metadata structure, forming a complete set of semantic edges with dynamic weight attributes and rule metadata. This processing method transforms the candidate semantic edge results from step S2.3 into multi-dimensional semantic edge data that supports subsequent graph topology assembly and can adaptively adjust under dynamic operating conditions, achieving stable physical interpretation capabilities of the process parameter semantic graph under varying operating conditions.

[0084] S2.5: Integrate the set of process parameter graph nodes with the set of complete semantic edges, perform graph topology assembly and connectivity verification processing, and output a process parameter semantic graph containing all active parameter nodes and their multidimensional semantic connection relationships.

[0085] The process parameter graph node set and complete semantic edge set output from step S2.4 are loaded into the graph topology assembly module as parallel inputs. Node index matching and semantic edge endpoint mapping are performed on both types of data objects to ensure that the start and end nodes of each semantic edge can be correctly indexed and matched in the node set. A topology matrix construction operation is performed on the mapped node and edge sets to explicitly represent the multidimensional semantic connections between parameter nodes in the form of an adjacency matrix. Dynamic weight values ​​and rule label indices of semantic edges are written into the matrix elements. Based on this adjacency matrix, the node degree distribution and edge connectivity index are calculated. For local regions with isolated nodes or broken edges, a completion algorithm is triggered, calling inference rules from the knowledge base to generate virtual bridging edges based on physical rationality to restore necessary connectivity. The corrected adjacency matrix is ​​converted into a graph data structure storage object, and an index mapping table of node feature tensors and edge attribute tensors is generated for fast access by the downstream dynamic graph processing module. A connectivity verification algorithm is invoked to perform a depth-first traversal of all active nodes within the graph data structure, determining whether all nodes are within a single connected component and verifying that each connected path can be constituted by at least one valid semantic relation. Through this assembly and verification process, the discrete set of nodes and edges from the previous step is transformed into a topologically complete, effectively connected, and semantically consistent semantic graph of process parameters, providing a unique and highly reliable data source for the subsequent generation of sparse subgraphs for adapting to different operating conditions.

[0086] For example, in a dual-pulse gas metal arc welding process, the parameter graph node set contains 6 nodes, corresponding to current (180A), voltage (24V), welding speed (0.35m / min), wire feed rate (8.5m / min), oscillation amplitude (5mm), and shielding gas flow rate (18L / min), respectively. The complete semantic edge set contains 12 edges, with an initial weight of 0.92 for the heat input dominant edge, 0.88 for the molten pool stability constraint edge, 0.81 for the spatter suppression associated edge, and 0.76 for the heat-affected zone superposition coupling edge. After performing node index mapping, a 6×6 adjacency matrix is ​​established, and the weights are filled in according to row and column positions. Using connectivity retrieval, a broken link is found between the speed node and the shielding gas flow rate node. A virtual edge with a confidence of 0.69 is derived by using the rule "shielding gas flow rate affects molten pool stability" in the knowledge base to complete the connection. In the newly generated adjacency matrix after correction, all nodes have a degree value ≥ 2, and depth-first traversal shows that all 6 nodes are in the same connected component. The semantic path contains at least one hot input dominant relationship. The final output process parameter semantic graph maintains complete physical interpretability and multi-parameter collaborative relationship description capability during downstream sparsification processing, ensuring that the minimum closed-loop structure can still be extracted after dynamic pruning and supporting the generation of subsequent compensation strategies.

[0087] like Figure 2As shown, step S3 involves pruning weakly connected semantic edges with confidence levels below a threshold in the process parameter semantic graph based on material thickness, ambient temperature, and weld curvature, generating a sparse subgraph for working condition adaptation. Specifically, this includes:

[0088] S3.1: Obtain material thickness data, ambient temperature data, and weld curvature data during the current welding process as working condition context feature vectors, and based on a preset working condition feature mapping rule base, convert the working condition context feature vectors into a set of boundary constraint conditions for evaluating the effectiveness of semantic edges, so as to establish the physical basis for dynamic pruning operations.

[0089] The system acquires raw signal streams from thickness sensors, thermocouple temperature acquisition units, and laser displacement sensors, which correspond to continuous monitoring data of material thickness, ambient temperature, and weld curvature, respectively. These data are then decoded via analog-to-digital conversion and digital communication bus to form a set of raw operating condition data with a unified timestamp.

[0090] The original working condition data set is subjected to dimensional unification and physical unit standardization processing. The thickness data is converted into a metric linear scale, the temperature data is converted into a Celsius absolute temperature value, and the curvature data is converted into a real quantitative value reflecting the radius of curvature, ensuring that parameters from different sources are comparable in the same numerical space.

[0091] The data parser in the working condition feature mapping rule base is called, and the standardized values ​​of thickness, temperature and curvature are mapped to the corresponding physical influence indicators one by one according to the material thermal conductivity, specific heat capacity and arc heat input empirical thresholds, including thermal diffusivity coefficient influence factor, cooling rate influence factor and trajectory deflection potential coefficient.

[0092] The above physical influence indicators are assembled into a working condition context feature vector according to parameter categories. A normalization function is then used to perform interval compression mapping on each dimension of the indicator. A commonly used normalization formula is:

[0093]

[0094] in, The original physical impact index, and These are the minimum and maximum physical boundary values ​​defined in the rule base for this indicator. These are the dimensionless eigenvalues ​​after normalization.

[0095] The normalized operating condition context feature vector is input into the condition parsing module of the rule base. By comparing the matching degree of each indicator with the preset critical interval, a set of boundary constraint conditions containing boundary values, symbol thresholds and operating condition levels is generated, which is used to evaluate the effectiveness of semantic edges in the dynamic pruning process.

[0096] By standardizing, mapping, and generating conditions, the original thickness, temperature, and curvature signals from the previous step are transformed into a set of boundary constraints that can be directly used as the basis for semantic edge confidence correction, thus providing physical parameterization support for dynamic pruning operations.

[0097] For example, on a high dynamic response welding production line, the thickness sensor outputs a plate thickness of 0.0125 meters, the thermocouple outputs an ambient temperature of 38 degrees Celsius, and the laser displacement sensor calculates a weld curvature radius of 0.45 meters. After data standardization, the resulting values ​​are: thickness 0.0125 meters, temperature 311.15 K, and curvature calculated to be 2.222 meters. -1 The rule base defines the thickness boundary as 0.005 meters to 0.02 meters, the temperature boundary as 293.15 K to 323.15 K, and the curvature boundary as 1 meter. -1 1 to 5m -1 The corresponding physical influence indicators, after normalization using formulas, yield a thickness normalized value of 0.5, a temperature normalized value of 0.6, and a curvature normalized value of 0.3055. The condition parsing module combines these three values ​​to form a working condition context feature vector [0.5, 0.6, 0.3055]. It determines that the thickness and temperature indicators are within the stable working condition range set by the rule base, while the curvature indicator is in the high range. The output set of boundary constraints includes two boundary conditions for pruning: "effective thermal input dominant relationship" and "weakened melt pool stability constraint." These conditions can be used to prune low-confidence weak connections in melt pool stability in the semantic graph, thereby improving the matching degree and interpretability of the pruning results.

[0098] S3.2: Based on the set of boundary constraints, the semantic edge weight decay function is called to perform real-time correction calculation on the initial confidence scores of all thermal input dominant relationship edges, molten pool stability constraint relationship edges, splash suppression relationship edges, and thermally affected zone superimposed coupling relationship edges in the semantic graph of process parameters, so as to generate a dynamic confidence score sequence that reflects the adaptability of the current working condition.

[0099] The set of boundary constraints generated based on the feature mapping of the operating context is input into the semantic edge weight decay function call interface, and the semantic graph of process parameters is specified as the object to be corrected, so as to ensure that the correction process can cover the initial confidence values ​​corresponding to all heat input dominant relationship edges, molten pool stability constraint relationship edges, splash suppression relationship edges, and heat-affected zone superimposed coupling relationship edges.

[0100] For each type of semantic edge, based on the matching results of the physical quantities and boundary constraints involved in the domain rule labels it is bound to, the corresponding weight decay function parameter set γ is extracted. This parameter set is used to characterize the decay curve shape and rate change law of the physical coupling relationship strength under the current working condition.

[0101] Initial confidence score for each semantic edge Perform attenuation calculation, the formula is as follows:

[0102]

[0103] in, This is the corrected dynamic confidence score. Let the semantic edge weight decay function be used. For the set of parameters of the decay function, For the current set of boundary constraints, through the function Output the effective confidence level under the current operating conditions.

[0104] When performing edge-by-edge operations on the decay function f, the physical response mechanism is matched according to the semantic edge type. For example, the edge dominated by heat input uses a logarithmic decay function that is inversely proportional to thermal conductivity and thickness, while the edge constrained by molten pool stability uses an exponential decay function that is proportional to the rate of change of temperature gradient, so as to ensure the consistency between the physical mechanism and the mathematical model.

[0105] All correction results are sequentially written into the dynamic confidence score buffer. The resulting dynamic confidence score sequence maintains a one-to-one correspondence with the semantic edge indices in the original process parameter semantic graph, providing direct input for subsequent identification of weak connections below the threshold.

[0106] Through the above-mentioned attenuation correction process, the set of boundary constraints generated in the previous step is transformed into a dynamic confidence score sequence that can quantify the adaptability of the current working condition, thereby realizing real-time updating of semantic edge strength and maintaining physical interpretability.

[0107] For example, under the conditions of welding a 12mm thick high-strength steel plate, an ambient temperature of 45℃, and an average weld curvature radius of 150mm, the forced attenuation coefficient γ of the edge triggered by the thickness condition in the boundary constraint set is 0.65, the attenuation coefficient γ of the edge triggered by the temperature condition for molten pool stability constraint is 0.72, and the curvature condition has a relatively small impact on the edge related to spatter suppression, with an attenuation coefficient γ of 0.92. For an initial confidence level... The dominant edge with a heat input of 0.88 has a decay function in logarithmic form:

[0108]

[0109] in This is a composite influence factor calculated by combining thickness and thermal conductivity. In this example... Substituting 0.85 into the calculation, the attenuation function output value is 0.644, resulting in a dynamic confidence level of 0.566. This value will be marked as a potential weak connection in the threshold comparison stage. Similarly, the corresponding functions and parameters are calculated for the edges related to melt pool stability constraints and splash suppression correlations to generate a complete sequence. Verification shows that the dynamic confidence level distribution under this condition significantly distinguishes between strong and weak connection types, ensuring the accuracy and adaptability of subsequent pruning decisions.

[0110] S3.3: Compare each dynamic confidence score in the dynamic confidence score sequence with a preset working condition adaptability threshold to identify all weak connection semantic edge identifiers that are lower than the preset working condition adaptability threshold, so as to form a list of invalid topology connections to be removed.

[0111] Based on the dynamic confidence score sequence output by S3.2 and the preset working condition adaptability threshold as input conditions, a threshold comparison operation is performed on each item in the score sequence to establish an adaptive state label set for each semantic edge. In the comparison operation, a threshold determination function is called to calculate and symbolize the difference between the dynamic confidence score and the preset threshold to parse the strength attribute of the current connection relationship. Semantic edges judged to be below the threshold are labeled "invalid connection," and their edge identifiers and corresponding domain rule metadata are packaged and written into the candidate elimination buffer. Using a buffer traversal mechanism, invalid connection records are scanned sequentially according to the unique identifier index of the semantic edge, eliminating duplicate edge indices that have already been marked and removed in the historical pruning steps to ensure list uniqueness and eliminate processing redundancy. The deduplicated invalid connection records are grouped and encoded according to semantic edge category to form a list of invalid topological connections to be eliminated, with a physical coupling type field. Through the above chain-like determination and grouping output, the dynamic confidence evaluation results are transformed into the basis for eliminating the physical validity of semantic edges, realizing the basic data preparation for working condition adaptation sparsity.

[0112] For example, the dynamic confidence score sequence generated by S3.2 in a certain welding scenario is set to [0.92, 0.48, 0.73, 0.35], where the scores correspond to the heat input dominant relationship edge E1, the molten pool stability constraint relationship edge E2, the spatter suppression correlation relationship edge E3, and the heat-affected zone superposition coupling relationship edge E4, respectively. The preset working condition adaptability threshold is set to 0.5. A comparison operation is performed through the threshold determination function:

[0113]

[0114] in, For the first The confidence difference of semantic edges, This is a dynamic confidence score. The threshold for adaptability is set. The difference between E1 and E4 is calculated to be 0.42, -0.02, 0.23, and -0.15, respectively. Based on the sign of the difference, E2 and E4 are determined to be weak connections below the threshold and are marked as invalid connections. The edge identifiers and semantic category labels (molten pool stability constraint and thermally affected zone superimposed coupling, respectively) of these two connections are written into the candidate rejection buffer. After deduplication and grouping by category, a list of invalid topological connections to be rejected, containing E2 and E4, is output. In this example, the final output list is used in S3.4 to prune the semantic graph of process parameters, removing the edges corresponding to E2 and E4, significantly improving the matching accuracy of subsequent sparse subgraphs with the current operating condition and reducing low-correlation topological interference.

[0115] S3.4: Based on the list of invalid topological connections to be removed, perform a topological pruning operation on the semantic graph of the process parameters, remove all weak connection semantic edges and their associated domain rule labels from the list, so as to reconstruct a sparse subgraph of the working condition adaptation that retains only high-confidence strong connections.

[0116] Based on the list of invalid topological connections to be removed, the list is compared one by one with the topological data structure of the current process parameter semantic graph, matching node pairs and edge indices to generate a corresponding topological element location index table to ensure the accuracy of the removal operation. Based on the topological element location index table, the graph structure editing module is invoked to perform edge-level deletion operations, removing all weakly connected semantic edges below the operating condition adaptability threshold from the graph edge set. Simultaneously, the in-degree and out-degree counts of graph nodes are maintained to prevent the generation of isolated nodes. For each deleted semantic edge, its attached domain rule label and weight decay function reference are simultaneously removed, and the relevant rule object is deregistered from the rule metadata mapping table to eliminate logical misjudgments caused by residual rule labels in subsequent calculations. The adjacency list or adjacency matrix of node attributes is reconstructed using the remaining edge set, the index sequence is reallocated, and the edge weight matrix is ​​updated to ensure that the adjusted data structure has consistency and readability in memory. Semantic type statistics and high-confidence judgment operations are performed on the reconstructed topology. Only strong connections with a confidence level higher than the working condition adaptability threshold and satisfying edge type legality are retained, forming a working condition-adaptive sparse subgraph that conforms to the current material thickness, ambient temperature, and weld curvature conditions. Through the above topology pruning and reconstruction processing, the invalid connection identification results of the previous step are transformed into a simplified graph structure that retains only strong semantic coupling relationships, realizing the structural optimization and redundancy elimination of the graph under dynamic working conditions.

[0117] For example, in an actual automated welding production line, the process parameter semantic graph contains 6 nodes, representing current, voltage, welding speed, wire feed rate, oscillation amplitude, and shielding gas flow rate, respectively, and 12 edges. The edge types cover heat input dominance, weld pool stability constraint, spatter suppression correlation, and heat-affected zone superposition coupling. The current operating condition context features are a material thickness of 8mm, an ambient temperature of 45℃, and a weld curvature radius of 250mm. The list of invalid topology connections to be removed, extracted by S3.3, contains 4 weakly connected semantic edges with dynamic confidence scores of 0.18, 0.22, 0.25, and 0.28, respectively, while the preset operating condition adaptability threshold is 0.3. When performing topology pruning, the corresponding edges are deleted from the adjacency matrix, and the labels "spatter suppression - low-speed failure" and "heat-affected zone coupling - high-temperature failure" in the rule label index table are simultaneously deleted. The adjacency matrix was updated from the original 6×6 to a 6×6 sparse structure, the edge density decreased from 0.33 to 0.22, and the average weight of high-confidence edges increased to 0.74. After pruning, the heat input dominant relationship loop of the working condition-adapted sparse subgraph was completely preserved, and the molten pool stability constraint edges were reduced to 2. This ensures that the subgraph only describes topological relationships with clear physical meaning and high synergy between parameters in the current 8mm thick high-temperature medium curvature welding scenario, providing a simplified and efficient basic structure for the connectivity verification in S3.5 and the minimum closed-loop structure retrieval in S4.

[0118] S3.5: Perform connectivity verification on the working condition adaptation sparse subgraph to verify whether the remaining nodes meet the semantic connectivity constraints, so as to output the final working condition adaptation sparse subgraph data structure with the explanatory power of the current working condition, for subsequent use by the minimum closed-loop structure retrieval module.

[0119] The current node set and semantic edge set of the sparse subgraph for work condition adaptation are initialized through graph traversal. The unique identifiers of all active parameter nodes and the connection matrix of their retained strong semantic edges are extracted. An adjacency list data structure is constructed based on this connection matrix, and a connectivity analysis algorithm based on depth-first search or breadth-first search is called to sequentially perform a global reachability determination for each active parameter node, recording the complete set of reachable nodes. The reachable node set of each node is matched against the entire set of active nodes. If any active node has an unreachable target, the corresponding subgraph is marked as an invalid subgraph with unsatisfactory connectivity, and a list of broken paths is generated. For the terminal nodes of broken paths in the list, substitute high-confidence edges that are within the allowable range of the current work condition boundary constraints are retrieved from the semantic edge backup library and temporarily inserted into the connection matrix to repair local connectivity, and the adjacency list is updated. The repaired connectivity matrix undergoes another global reachability check until at least one valid path conforming to semantic type constraints exists between all active nodes, and each edge on the path satisfies any of the following valid semantic relationships: heat input dominance, molten pool stability constraint, splash suppression association, or thermally affected zone superposition coupling. Through the above connectivity verification and repair processes, the dynamically pruned sparse topology is transformed into a final-state adaptable sparse subgraph data structure that meets the physical interpretability requirements of the current operating condition parameters, thus providing structurally complete and semantically valid input conditions for minimum closed-loop structure retrieval.

[0120] For example, in a high-speed welding scenario, the condition-adapted sparse subgraph contains 6 active parameter nodes (welding current, voltage, speed, wire feed rate, oscillation amplitude, and shielding gas flow rate). After initial pruning, only 7 high-confidence semantic edges are retained. Connectivity analysis reveals no semantic path between the speed node and the shielding gas flow rate node, failing to meet connectivity requirements. The broken path set identifies the speed node as the end of the broken path. The system retrieves a heat-input-dominant high-confidence edge (speed-wire feed rate, weight 0.82) from the backup database and temporarily inserts it into the connection matrix. After updating the adjacency list, global reachability analysis is performed, confirming that all nodes are connected via at least one valid semantic path. A path search algorithm verifies that the path between any two nodes is included in the four types of semantic relationships mentioned above, ensuring subgraph connectivity meets requirements. This final-state sparse subgraph data structure is completely output to the minimum closed-loop structure retrieval module, ensuring that the subsequently activated parameter cooperative subgraph has sufficient physical interpretability and reliable control command generation under high-speed welding conditions.

[0121] like Figure 3 As shown, step S4 involves retrieving a minimal closed-loop structure that covers all active parameter nodes and satisfies semantic connectivity constraints from the sparse subgraph of the working condition adaptation, thereby obtaining a parameter cooperative subgraph. Specifically, this includes:

[0122] S4.1: Based on the set of active parameter nodes and the pre-defined semantic edge connection relationships in the sparse subgraph of the working condition adaptation, perform full path connectivity traversal processing to generate a set of candidate connected components containing all possible node combinations, ensuring that the subsequent retrieval scope covers all effective process parameter associations under the current welding scenario.

[0123] It should be noted that the active parameter nodes refer to all graph nodes whose corresponding process parameters (such as welding current, voltage, speed, wire feed rate, oscillation amplitude, shielding gas flow rate, etc.) are within the preset physical effective range under the current welding conditions, have not been marked or removed by sensor failure or data loss, and participate in the current graph topology calculation. Nodes that do not meet any of the above conditions are defined as "inactive nodes" and are ignored during pruning and loop closure retrieval.

[0124] It should be noted that the semantic connectivity constraint refers to the existence of at least one unbroken connected path between any two parameter nodes in the process parameter semantic graph or its subgraphs, consisting of effective semantic edges (including edges representing heat input dominance, molten pool stability constraints, splash suppression, or heat-affected zone superposition coupling). Furthermore, the dynamic confidence of each semantic edge on this path is not lower than a preset threshold for the current operating condition. A graph structure satisfying this constraint is considered to have complete process parameter semantic connectivity, capable of supporting subsequent closed-loop retrieval and compensation calculation.

[0125] S4.2: For each connected component in the candidate connected component set, perform the closed-loop structure determination algorithm to filter out the initial closed-loop structure list with overlapping first and last nodes and no internal breaks, thereby eliminating open-loop paths and locking parameter interaction loops with feedback adjustment potential.

[0126] S4.3: Based on the number of nodes and the sum of edge weights in the initial closed-loop structure list, perform a minimum cost evaluation calculation to select the candidate minimum closed-loop structure that covers all active parameter nodes and has the lowest topological complexity from multiple initial closed-loop structures, thus realizing the simplest expression of the parameter cooperative path.

[0127] S4.4: Perform semantic connectivity constraint verification on the candidate minimum closed-loop structure to verify whether there is a valid semantic path between any two nodes in the structure that meets the definition of heat input dominance or melt pool stability constraint, and then generate a compliant minimum closed-loop structure that passes the verification to ensure the physical interpretability of the logical relationship between parameters.

[0128] Based on the node attribute set and semantic edge feature parameters of the candidate minimum closed-loop structure list, the domain rule definition containing thermal input dominance and melt pool stability constraints is read to establish an effective semantic path determination matrix for path verification.

[0129] Map the node identifiers of any two nodes in the candidate minimum closed loop structure to the semantic edge decision matrix, search for the existence of a path composed of continuous valid semantic edges, and compare the consistency of all edge types in the path with the domain rule labels to eliminate path segments with type mismatch or label conflict.

[0130] Based on the path consistency comparison results, a semantic path completeness detection algorithm based on depth-first traversal is invoked to calculate the path coverage between each pair of nodes. When insufficient path coverage or the presence of isolated nodes is detected, the closed-loop structure is marked as non-compliant and added to the elimination queue.

[0131] For the remaining candidate minimum closed-loop structures that have not been eliminated, the weighted product of the effective edges of each path is calculated. When the weighted product is lower than the preset validity coefficient lower limit, the structure is eliminated.

[0132] The closed-loop structures that pass path consistency comparison and weighted product threshold test are aggregated to generate a set of "compliant minimum closed-loop structures". Each structure is then bound with its physical semantic interpretation metadata to ensure that the logical relationship between parameters meets the physical interpretability of the mechanical welding process.

[0133] Through the semantic connectivity constraint verification process described above, the candidate minimum closed-loop structure obtained in the previous step is transformed into a compliant minimum closed-loop data unit that contains only valid semantic paths and has high interpretability, so as to realize the accurate utilization of the coordination of process parameters in the subsequent compensation calculation.

[0134] S4.5: Based on the compliant minimum closed-loop structure and its contained node and edge attributes, perform subgraph activation instantiation processing to generate the most explanatory parameter collaboration subgraph representing the current working condition, which serves as the sole data carrier for subsequent graph attention propagation calculation and implicit collaboration pattern extraction.

[0135] Based on the compliant minimum closed-loop structure and its contained node and edge attributes, the graph topology instantiation engine is invoked to perform a deep copy of the attributes of all active parameter nodes within the structure. The physical attribute labels, real-time numerical fields, and dimensional information of the nodes are indexed and bound according to the unique identifier of the nodes, ensuring that the instantiated node objects can maintain uniqueness and traceability in subsequent calculations.

[0136] For each semantic edge in the compliant minimum closed-loop structure, attribute inheritance mapping is performed. Domain rule labels and dynamic weight values ​​are synchronously loaded according to the latest confidence coefficient of the sparse subgraph adapted to the current working condition. The connection direction, thermophysical relationship type, and timeliness flag of the edge are assembled into a complete metadata structure for the edge.

[0137] Based on the complete metadata of the node set and edge set, the subgraph data structure construction operation is performed, and the nodes and their adjacent edges are mapped in memory space according to the sparse matrix storage format to generate an active topological data model with efficient indexing and traversal capabilities.

[0138] After the subgraph data structure is constructed, the weight distribution of the edge set is normalized to ensure that the participation of different types of physical relationships in the subsequent attention propagation calculation remains consistent. The normalization result is linearly compressed to the [0,1] interval to ensure numerical stability.

[0139] Based on the normalized node attribute matrix and edge weight matrix, the subgraph object is encapsulated as a parametric collaborative subgraph instance. A timestamp label, working condition context identifier, and structure version number are appended to the instance metadata. This allows for direct invocation and avoids ambiguity in the subsequent graph attention propagation calculation stage. Through this instantiation process, the compliant minimum closed-loop structure from the previous step is transformed into a parametric collaborative subgraph data carrier that can be directly input into deep graph algorithms, achieving high-fidelity semantic expression of multi-parameter relationships under the current working condition.

[0140] For example, in an automated welding scenario with a steel plate thickness of 12mm, an ambient temperature of 35℃, and a weld curvature radius of 250mm, the compliant minimum closed-loop structure includes 6 active parameter nodes (current: 280A, voltage: 32V, welding speed: 380mm / min, wire feed rate: 6.5m / min, oscillation amplitude levels: 3, shielding gas flow rate: 18L / min) and 8 high-confidence semantic edges (3 of which are heat input-dominant relationships, 2 are molten pool stability constraint relationships, 2 are spatter suppression relationships, and 1 is a heat-affected zone superposition coupling relationship). After calling the instantiation engine, attributes such as the current node are encapsulated as... The structure {“Label”:“Current (A)”,“Real-time Value”:280,“Dimension”:“A”} encapsulates the thermal input dominant edge attribute into {“Type”:“Temperature Input Dominant”,“Weight”:0.88,“Rule”:“IVS Coupling Law”}. The node set and edge set are then stored as an adjacency list. After weight normalization, the weight of the thermal input dominant edge is 0.88, the weight of the melt pool stability constraint edge is 0.75, the weight of the splash suppression associated edge is 0.69, and the weight of the thermally affected zone superimposed coupling edge is 0.81. The generated parameter co-graph object carries the operating condition identifier {T12mm,E35C,R250mm} and the version number v1.0. Subsequently, in step S5, this object is directly used as input for multi-head graph attention propagation calculation, which can significantly improve the stability of key node identification and the interpretability of process logic.

[0141] Step S5: Perform graph attention propagation calculation on the parameter cooperative subgraph to identify the parameter nodes that play a dominant regulatory role and their semantic neighborhoods, and extract implicit cooperative patterns. Specifically, this includes:

[0142] S5.1: Based on the node attribute vectors and semantic edge connection relationships in the parameter collaborative subgraph, perform multi-head self-attention mechanism initialization processing on all active parameter nodes in the graph to generate an initial attention feature set containing query matrix, key matrix and value matrix, and establish the basic mapping space for information interaction between nodes.

[0143] It should be noted that the graph attention propagation computation mentioned above refers to an iterative computation process based on the Graph Attention Network (GAT) framework, performing a multi-head self-attention mechanism on the feature vectors of nodes in the parametric collaborative subgraph. This process calculates the attention coefficients between neighboring nodes, weights and aggregates the feature information of each node along semantic edges to its neighboring nodes, and performs nonlinear transformations. After multiple iterations, the hidden state representation of each node is integrated with the contextual information of its multi-hop neighborhood, thereby achieving quantitative evaluation and propagation enhancement of the importance of nodes and the implicit dependencies between parameters in the parametric collaborative subgraph.

[0144] It should be noted that the implicit collaborative pattern refers to a data set used to characterize the joint response logic between the dominant control parameter node and the auxiliary control parameter node. This data set includes at least the dominant node identifier, the neighboring node identifier, the edge weights, and the associated semantic tags, and is stored in a computer-readable format for subsequent steps to call.

[0145] S5.2: Using the query matrix and key matrix in the initial attention feature set, perform dot product operation and scaling normalization on the semantic correlation between any two nodes in the parameter collaboration subgraph to calculate the original attention score matrix that reflects the dynamic dependence strength between parameters under the current working condition.

[0146] S Based on the original attention score matrix and the preset masking rules, negative infinity masking and soft maximum function normalization are performed on the score items corresponding to non-connected node pairs and low-confidence semantic edges to generate a standardized attention weight matrix that satisfies probability distribution constraints and retains only valid topological connections, thereby eliminating information interference from invalid paths.

[0147] Based on the original attention score matrix and preset masking rules, the attention score index positions of corresponding non-connected node pairs in the matrix are mapped to the masking flags in the masking rules. A binary masking matrix is ​​used to replace the value at this position with a negative infinity constant, ensuring that the weight of this path tends to zero during subsequent normalization. For score items corresponding to low-confidence semantic edges in the matrix, the dynamic weight threshold judgment module is invoked. After determining them as low-efficiency connections, the same negative infinity masking process is performed to block the information transmission of low-quality connections. After completing all masking operations, an exponential mapping operation is performed on each row of the masked attention score matrix, calculating the exponential value of each element to amplify the differences between high-scoring items and suppress the contribution of low-scoring items. The exponential values ​​of each row are summed row by row to obtain the normalization factor matrix corresponding to each query node. The division operation module is invoked to divide the exponential matrix element by the corresponding normalization factor, achieving soft maximum function normalization that satisfies probability distribution constraints, ensuring that the sum of the weights in each row is 1. This normalization design generates a standardized attention weight matrix containing only valid topological connections, providing an optimized weight distribution for subsequent weighted aggregation using the value matrix. The formula is as follows:

[0148] in, The original attention score matrix, It is a mask matrix containing a masking value of negative infinity. To standardize the attention weight matrix, which controls the contribution ratio of adjacent nodes in subsequent information aggregation, a masking and soft maximum normalization process is used to transform the original dynamic dependency strength calculated in the previous step into monotonic and normalized probabilistic weight data. This achieves the technical effect of filtering out invalid paths and enhancing high-confidence topological connections.

[0149] For example, in a certain welding scenario, the parameter co-graph contains three nodes: welding current, arc voltage, and welding speed. The original attention score matrix is ​​[[0.8,0.1,0.3],[0.2,0.7,0.4],[0.5,0.2,0.6]]. The masking rule determines that the connection from welding speed to arc voltage is invalid and fills the corresponding position with negative infinity. After performing an exponential transformation on the masked matrix, the exponential value of high-scoring items, such as the current-voltage path, is significantly higher than that of other elements. The weight of the normalized attention weight matrix obtained by row normalization on this path is close to 1, while the weights of other masked paths are close to 0. Inputting this weight matrix into the subsequent value matrix weighted aggregation can significantly improve the context feature extraction effect dominated by the current-voltage path, and maintain stable weld seam tracking compensation decision accuracy under different material thicknesses and ambient temperature changes.

[0150] S5.4: Based on the standardized attention weight matrix and the value matrix in the initial attention feature set, perform weighted aggregation and linear projection transformation on the parameter cooperative subgraph to update the hidden state representation of each parameter node and generate an enhanced node embedding vector that integrates neighborhood context information.

[0151] Based on the standardized attention weight matrix and the value matrix in the initial attention feature set, the value vector corresponding to each active parameter node in the parameter collaboration subgraph is weighted and accumulated with the value vectors of its neighboring nodes according to the weight coefficients, generating a linear combination vector containing neighborhood semantic information. The formal calculation of the above weighted accumulation operation is expressed as:

[0152]

[0153] in, This represents the hidden state of the node after the update. This represents the set of indices of the neighboring nodes of the current node. This represents the weights from node i to node j in the standardized attention weight matrix. Let represent the vector corresponding to node j in the value matrix. Based on the above aggregation results, a linear transformation operator W is introduced into the weighted neighborhood vector of each node, and matrix multiplication is performed to adjust the dimension and distribution of the feature space. This linear transformation process is formally represented as:

[0154] ,in A trainable weight matrix specific to the parameter co-graph is determined through prior training or adaptive iteration. A nonlinear activation function, such as LeakyReLU, is applied to the linear transformation output to enhance the ability of the node embeddings to express nonlinear dependencies between parameters. After this nonlinear activation, it can be expressed as:

[0155] ,in The activation function is represented. Normalization is performed on the node representation vectors after nonlinear mapping to ensure consistency in numerical scale across node embedding vectors with different parameters, avoiding bias effects caused by numerical scale differences in subsequent pattern recognition processes. The normalized node representation set is encapsulated into an enhanced node embedding vector set, providing high signal-to-noise ratio semantic input for dominant control node recognition and implicit cooperative pattern extraction. Through weighted aggregation, linear projection, nonlinear activation, and normalization, the attention weight distribution from the previous step is transformed into a node embedding expression containing complete neighborhood context information, achieving deep propagation and enhancement of local semantic features along the parameter cooperative subgraph.

[0156] For example, in a highly dynamic welding scenario, the parameter co-graph contains three types of active parameter nodes: welding current, arc voltage, and welding speed. In the standardized attention weight matrix, the weight of the current node pointing to the speed node is 0.45, the weight of the voltage node pointing to the speed node is 0.35, and the weight of the speed node pointing to itself is 0.20. The corresponding vector dimension of the three nodes in the value matrix is ​​64. Taking the speed node as an example, its updated hidden state representation is calculated as follows:

[0157] , among which , , The vector values ​​originate from the 64-dimensional floating-point features of the initial matrix. Projected onto a 128-dimensional space by matrix W, and subjected to a nonlinear transformation using LeakyReLU (with a negative slope parameter of 0.01), the final enhanced embedding vector is obtained through L2 norm normalization. This enhanced embedding vector successfully highlighted the sensitivity of welding speed nodes to thermal input dynamics under current operating conditions in subsequent dominant node recognition tests, significantly improving the accuracy of implicit cooperative pattern extraction and verifying the role of aggregation and projection processing in preserving and enhancing node semantic information.

[0158] S5.5: Based on the enhanced node embedding vector, sorting and threshold truncation are performed on the feature norms of each node in the parameter collaboration subgraph to identify the dominant control parameter node with a significantly higher attention weight than the average level and its directly connected semantic neighborhood, and extract the implicit collaboration pattern.

[0159] Based on the enhanced node embedding vector, the feature norms of all nodes in the parametric co-graph are precisely calculated and formed into a feature norm sequence in descending order, so as to construct a basic sorted index of node importance.

[0160] The feature norm sequence is mapped one-to-one with the node labels and their weight values ​​in the standardized attention weight matrix to construct a mapping table containing the correspondence between node numerical representations and attention weights, which is used for subsequent weight threshold determination.

[0161] Based on statistical distribution characteristics, the attention weight threshold is dynamically set using the mean-multiplicative deviation coefficient method. The threshold is calculated by calling the following formula. ,in The average of the attention weights of all nodes. Standard deviation, These are coefficient parameters selected based on empirical sensitivity to operating conditions.

[0162] The attention weight in the mapping table is higher than The nodes are labeled as dominant control parameter nodes, and the directly connected semantic neighborhood nodes of each dominant control node in the parameter coordination subgraph are further retrieved to form a dominant-neighborhood node subset set.

[0163] Perform connectivity and physical interpretability verification on the dominant-neighborhood node subset set, remove node pairs that do not meet domain logic rules such as heat input dominance and molten pool stability constraints, and retain node connection structures that conform to the multi-parameter physical response mechanism.

[0164] The validated subset of dominant-neighbor nodes, combined with their edge attributes in the parametric collaborative subgraph, is encapsulated into an implicit collaborative pattern data structure, including dominant node identifier, neighbor node identifier, edge weight, and associated semantic label.

[0165] Through the chain-like processing described above, the standardized attention weight matrix and the enhanced node embedding vector are transformed into an implicit collaborative pattern that accurately characterizes the dominant regulatory mechanism, thereby enabling the structured extraction and interpretable output of multi-parameter joint response logic.

[0166] For example, in a six-axis robotic arm welding system for high-precision thin plate welding, the dimension of the enhanced node embedding vector is 64. The node feature norm obtained after graph attention propagation ranges from 0.45 to 1.32, the mean attention weight is 0.54, the standard deviation is 0.12, and the coefficient k is set to 1.2. Then the weight threshold T is calculated as 0.54 + 1.2 × 0.12, which is 0.684. After comparison and screening, welding current node (weight 0.91), welding speed node (weight 0.87), and wire feed rate node (weight 0.73) are selected into the set of dominant control parameter nodes. When searching their first-order semantic neighborhood, it is found that there is a heat input dominant relationship edge between welding current node and arc voltage node, a molten pool stability constraint relationship edge between welding speed node and oscillation amplitude node, and a spatter suppression correlation edge between wire feed rate node and shielding gas flow rate node. After physical interpretability verification, all satisfy the defined domain logic rules, and the finally generated implicit cooperative pattern contains three sets of dominant-neighborhood pairings and their edge weight attributes. In the subsequent control strategy generation stage, this mode ensures that the compensation strategy can significantly improve the accuracy of response to the coordinated changes in current and speed, and maintain the stability of weld formation in multiple dynamic thermal input disturbance scenarios.

[0167] Step S6: Based on the implicit collaborative pattern, a corresponding behavioral template is matched from a pre-set compensation knowledge base. The behavioral template defines the proportional constraint relationship and response timing difference between the dominant compensation channel and the auxiliary compensation channel. Specifically, this includes:

[0168] S6.1: Based on the dominant control parameter nodes and their semantic neighborhood features in the implicit collaborative mode, perform semantic matching processing on the index key values ​​in the preset compensation knowledge base to lock the target behavior template set that is highly consistent with the current multi-parameter joint response logic.

[0169] S6.2: Using the process mechanism rules defined in the target behavior template set, the functional roles of the dominant compensation channel and the auxiliary compensation channel are dynamically allocated to clarify the primary and secondary control status of the current component or speed component under specific operating conditions.

[0170] Based on the process mechanism rules defined in the target behavior template set, logical entries concerning the parameter coupling relationships caused by changes in heat input, molten pool stability, and spatter suppression are extracted and used as the rule basis for dynamic allocation decisions.

[0171] The attribute vectors of the dominant and auxiliary control parameter nodes in the implicit collaborative mode are compared one by one with the above rule entries to calculate the response sensitivity coefficient and stability coefficient of each node under different process mechanisms.

[0172] Based on the weighted results of the response sensitivity coefficient and the stability coefficient, a comprehensive impact factor for each parameter node under the current operating conditions is generated, and the impact factors are sorted to form a priority sequence to be assigned.

[0173] The channel allocation matrix for different working conditions is called from the rule base. The channel roles are mapped by combining the priority sequence. The node with the highest influence factor is marked as the dominant compensation channel, and the other nodes that meet a certain threshold are marked as auxiliary compensation channels.

[0174] When both current and velocity components are candidates for the dominant compensation channel, the instantaneous heat input contribution values ​​of the two are compared according to the heat input dominance threshold in the process mechanism rules:

[0175]

[0176] in, This is the real-time value of the welding current. This is the real-time value of the arc voltage. The thermal efficiency coefficient is determined by the following: when the Q value corresponding to the current component exceeds the threshold thermal input corresponding to the velocity component, the current is retained as the dominant compensation channel; otherwise, the velocity is used as the dominant compensation channel.

[0177] This dynamic allocation method integrates the target behavior template from the previous step with the parameter characteristics of the actual working conditions, thereby achieving the optimal functional role division of the main compensation channel and the auxiliary compensation channel under specific working conditions. This ensures that the subsequent proportional constraints and timing difference calculations have physical consistency and process adaptability.

[0178] For example, in a multi-pass welding scenario of a high-strength steel plate, the sensitivity coefficient of the current node output by the implicit collaborative mode is 0.82, the stability coefficient is 0.88, the sensitivity coefficient of the velocity node is 0.79, and the stability coefficient is 0.91. The dominant threshold for heat input in the target behavior template set is 4.5kW, and the thermal efficiency coefficient η is set to 0.85 according to the material database. The measured current I is 220A, the voltage U is 24V, and the calculated heat input of the current component is Q=220×24×0.85=4488W. The heat input of the velocity component is converted to 4120W. Comparing the two, the input of the current component exceeds that of the velocity component and is higher than the threshold. In the dynamic allocation, the current is the dominant compensation channel, and the velocity is the auxiliary compensation channel. Under this allocation strategy, the subsequent proportional constraint coefficient calculation is stable, the molten pool fluctuation amplitude is reduced during the attitude correction process, and the weld formation uniformity is significantly improved, verifying the robustness and adaptability of the allocation logic under high dynamic heat input.

[0179] S6.3: Based on the allocation results of the functional roles and the statistical laws in the weld quality dataset, the numerical ratio coefficient between the dominant compensation channel and the auxiliary compensation channel is calculated to generate channel coupling parameters with deterministic ratio constraints.

[0180] S6.4: Based on the dynamic characteristics of thermal input and the stability constraints of the molten pool corresponding to the channel coupling parameters, the action trigger timestamps of the dominant compensation channel and the auxiliary compensation channel are time-aligned to construct a timing control sequence containing precise response timing differences.

[0181] Based on the numerical output and physical property identifiers of the channel coupling parameters, a set of characteristic parameters reflecting the dynamic characteristics of the heat input and the stability constraints of the molten pool is obtained, which serves as the basis for triggering the timestamp calibration operation.

[0182] A time event vector containing a dominant compensation channel and an auxiliary compensation channel is constructed. Each element in the time event vector corresponds to the original trigger time of the channel action, and a mapping relationship is established between it and the rate of change of heat input, the heat accumulation threshold, and the molten pool morphology stability criterion.

[0183] By invoking the thermal input dynamic model, and based on the influence coefficients of the instantaneous power of the dominant compensation channel and the displacement velocity of the auxiliary compensation channel on the time response curve, the optimal trigger time difference between the two channels under the condition of satisfying the thermal stability range is calculated. .

[0184] Time difference correction formula The triggering time of the auxiliary compensation channel involved in the time event vector is linearly shifted so that the main and auxiliary channels can achieve phase synchronization or optimal phase difference under the dual constraints of heat input and molten pool stability.

[0185] By combining the corrected time event vectors, a timing control matrix containing the precise trigger times of each channel is generated. Based on the time sequence of different behavioral units in the matrix, a logical link verification is performed on the timing to ensure that there are no trigger conflict paths that violate physical constraints.

[0186] By using the timing alignment processing method based on the dual model of heat input and molten pool stability, the channel coupling parameters of the previous step are transformed into a precise timing control sequence that can be directly analyzed by the actuator, thereby achieving high consistency and stability of the compensation action in the time domain.

[0187] For example, in a stainless steel sheet welding scenario, the primary compensation channel is the current component, with a real-time value of 125A, corresponding to an arc voltage of 24V. The auxiliary compensation channel is the welding torch transverse speed, with a real-time value of 0.45m / min. The material thickness is 2mm, and the molten pool stability criterion is set to a molten pool interface fluctuation amplitude of no more than 0.1mm. This is based on the heat input formula. Where η is the efficiency coefficient of 0.8, the calculated result for the dominant compensation channel is approximately 5333.33 J / mm. Based on the fitting of the dynamic curve of the heat input and the velocity response curve of the auxiliary compensation channel, the optimal triggering time difference under the condition of stable melt pool threshold is obtained. It takes 0.12 seconds. Using... The triggering time sequence of the auxiliary compensation channel was shifted. The adjusted timing control matrix shows that in each welding cycle, the auxiliary compensation channel starts 0.12s later than the dominant compensation channel. Thus, in multiple rounds of repeated experiments, the consistency of weld formation boundary and the spatter were significantly improved, verifying the compensation stability and process robustness of this timing alignment strategy under high dynamic welding conditions.

[0188] S6.5: Integrate the channel coupling parameters and timing control sequence, instantiate and encapsulate the target behavior template to output a structured behavior template containing complete proportional constraint relationships and response timing differences, which serves as the direct instruction input for driving compensation calculation.

[0189] Based on the channel coupling parameters and timing control sequence output from step S6.4, the behavior template instantiation engine is called to establish the input data parsing channel. Data type verification and unit system normalization are performed on the proportional coefficient matrix of the channel coupling parameters and the time offset vector of the timing control sequence, respectively, to ensure that the proportional constraints and timing differences have mathematical consistency that can be directly combined within the same physical dimension system.

[0190] The normalized proportional coefficient matrix is ​​used as a weight tensor and is subjected to positional operations in the channel identifier mapping table in the behavior template metadata structure to generate a template instance parameter block containing proportional constraints binding between the dominant compensation channel and the auxiliary compensation channel.

[0191] By utilizing the trigger timestamp information of each channel in the time offset vector, a time control field is appended to the template instance parameter block, and this is achieved through construction. The timing adjustment formula is in the form of a time series formula, which records the start and end phases of the action for each compensation channel. Define the time for the original behavior template. The phase difference is calculated from the timing control sequence.

[0192] The template instance parameter block containing both proportional and time domains is encapsulated in a structured manner, integrating physical constraint metadata, proportional calculation logic, and timing triggering logic into a unified data exchange format, such as based on hierarchical JSON or a binary protocol stack, to enable the compensation strategy to be directly invoked in subsequent modules.

[0193] The encapsulated structured behavior template data object is version-identified and signed to ensure the traceability and consistency of compensation decisions in different execution cycles. The result of the previous step is converted into a highly coupled control instruction that can be directly parsed by the compensation amount calculation module through template instantiation, so as to achieve precise synchronous execution of compensation strategy and process parameters.

[0194] For example, in the scenario of welding a longitudinal seam in a vehicle frame, the ratio of the velocity component to the current component in the channel coupling parameter matrix is ​​set to 1:0.85, indicating that under the current conditions of a material thickness of 6mm, stainless steel, and an ambient temperature of 15℃, the velocity adjustment range is slightly higher than the current adjustment range. In the timing control sequence, the velocity component trigger time leads the current component by 35 milliseconds to ensure that the heat input is first stabilized by the motion control end to form the weld, and then the current fine-tunes the heat distribution. The above ratio and time difference are converted into a standardized coefficient matrix and a phase vector, embedded into the proportional and time domain fields of the behavior template, and after data type validation, unit conversion, and physical consistency verification, a structured template object is generated, which includes T. speed =0ms and T current =35ms is precisely defined. When the compensation calculation module is called, this template directly generates speed and current compensation signals that meet the proportional and timing constraints. After being executed by the robotic arm, the actual weld trajectory offset is significantly reduced and the stability of the molten pool is greatly improved.

[0195] Step S7: Convert the behavior template into structured instructions, and perform compensation calculation based on the structured instructions. The compensation amount includes a velocity component and a current component with deterministic proportional constraints. Specifically, this includes:

[0196] S7.1: Obtain the implicit collaborative pattern output by the behavior template and parameter collaborative subgraph matched in the preset compensation knowledge base. Based on the dominant compensation channel identifier and auxiliary compensation channel identifier defined in the behavior template, parse the node weights in the implicit collaborative pattern to generate a set of channel mapping relationships containing dominant control parameter nodes and auxiliary control parameter nodes.

[0197] S7.2: Based on the set of channel mapping relationships and the preset proportional constraint coefficients in the behavior template, perform coupled proportional calculation processing on the real-time values ​​of the dominant control parameter node and the auxiliary control parameter node to generate a speed component reference value and a current component reference value with deterministic mathematical correlation.

[0198] Based on the set of channel mapping relationships and the preset proportional constraint coefficients in the behavior template, the physical dimensions and data types of the real-time values ​​from the dominant control parameter node and the auxiliary control parameter node are identified and confirmed respectively, so as to ensure that the input values ​​meet the unit consistency and accuracy requirements of the coupled calculation.

[0199] The real-time values ​​of the dominant control parameter nodes are registered with proportional coefficients. The pre-set proportional constraint coefficients in the behavior template are multiplied by the real-time values ​​to obtain the intermediate value of the dominant compensation channel amplified by the proportional factor. This operation is used to reflect the priority adjustment weight of the dominant compensation channel in the compensation strategy.

[0200] Normalization correction is performed on the real-time values ​​of the auxiliary control parameter nodes, mapping them to the corresponding dimensional standard range and keeping them consistent with the dimensions of the dominant control parameter nodes, so as to perform cross-channel numerical coupling calculations.

[0201] A weighted summation operation is used to fuse the intermediate values ​​of the proportionally registered dominant compensation channel with the corrected real-time values ​​of the auxiliary compensation channel to generate reference values ​​for the velocity component and the current component. The coupling relationship between the two satisfies the following mathematical constraints:

[0202]

[0203]

[0204] in, As the reference value for the velocity component, The reference value for the current component. The proportional constraint coefficient defined in the behavior template. To control the real-time values ​​of the parameter nodes. To assist in adjusting the real-time values ​​of parameter nodes.

[0205] Numerical boundary constraint checks are performed on the above calculation results to ensure that the reference values ​​of the velocity component and the current component are both within the preset process safety range. If they are exceeded, saturation limit processing is performed to prevent nonlinear instability effects from occurring during actual execution.

[0206] By using proportional coupling operations and boundary constraint processing, the set of channel mapping relationships and proportional constraint coefficients are transformed into reference values ​​for velocity components and current components with deterministic mathematical relationships, thereby achieving dual consistency of the compensation quantity at both the numerical and physical logic levels.

[0207] For example, in a working condition with a high-strength steel plate thickness of 10mm, a real-time welding speed of 480mm / min, and a real-time current of 210A, the proportional constraint coefficient k set in the behavior template is 0.15. Through channel mapping relationship set analysis, the dominant control parameter node is determined to be the velocity component, and the auxiliary control parameter node to be the current component. Substituting the velocity and current values ​​into the formula, the velocity component reference value is obtained, calculated to be 282. Substituting the current component reference value into the formula, the current component reference value is calculated to be 511.5. Within the safe range [250, 550], both the velocity component reference value 282 and the current component reference value 511.5 are within the range, requiring no saturation limitation treatment. In multiple batches of welding tests, the reference value generated by this proportional coupling calculation can maintain stable weld formation dimensions under conditions of slight fluctuations in material composition and ambient temperature changes of ±20℃, significantly improving weld quality consistency and the adaptability of compensation to changes in working conditions.

[0208] S7.3: Obtain the response timing difference parameter defined in the behavior template and the timestamp information of the current welding cycle. Based on the response timing difference parameter, perform timing offset alignment processing on the velocity component reference value and the current component reference value to generate a velocity component timing sequence and a current component timing sequence with time phase markers.

[0209] Using the speed and current component reference values ​​output from S7.2 as inputs for this sub-step, the predefined response timing difference parameter dataset in the behavior template is called, and the global timestamp information corresponding to the current welding cycle is extracted from the process parameter stream to establish a time calibration relationship between the two types of input parameters. The timing difference parameters are then used... and The desired trigger delays for the velocity and current components are respectively calculated using the formula.

[0210]

[0211] and

[0212]

[0213] global timestamp The trigger time is corrected to the independent execution time of each component. Based on the corrected trigger time, the velocity component reference value is adjusted. Reference value of current component Time-off interpolation is performed to expand the original scalar values ​​into time-varying sequence data. Linear interpolation or piecewise spline interpolation is used to generate corresponding component values ​​at the corrected time nodes, constructing a time-continuous velocity and current reference curve. The generated time series is phase-labeled according to the corrected trigger time difference, and a corresponding phase information tag is bound to each time point in the data structure to explicitly express the synchronization and delay relationships of different components in global execution. Through timestamp correction, interpolation generation, and phase labeling of velocity and current components, the static reference values ​​under proportional constraints are transformed into dynamically executable time-series sequences with phase labels, achieving coordinated and precise control of compensation actions in the time domain.

[0214] For example, in a high-dynamic-response welding environment, the velocity component reference value is configured as 200 mm / s, the current component reference value is configured as 180 A, the response timing difference defined by the behavior template is 15 ms advance for velocity and 10 ms delay for current, and the global timestamp of the current welding cycle is 5230 ms. The velocity correction time is calculated to be 5245 ms, and the current correction time is also 5240 ms. During the interpolation stage, the velocity vector is linearly interpolated within the interval [5230, 5245] ms to generate transition points, and the current vector is also linearly interpolated within the interval [5230, 5240] ms to generate transition points, forming a smooth transition curve. The velocity phase is marked as +15 ms, and the current phase as +10 ms in the sequence. After this processing, the robotic arm control system can simultaneously adjust the velocity and current with precise phase difference, achieving a significant improvement in molten pool stability and heat input control under high-speed welding conditions. Verification results show that the weld formation quality and deviation compensation stability are significantly improved.

[0215] S7.4: Based on the velocity component time sequence and the current component time sequence, a multivariate joint coding algorithm is used to perform structured encapsulation processing on the velocity component time sequence with time phase mark and the current component time sequence with time phase mark to generate a structured control instruction package carrying parameter coupling logic.

[0216] S7.5: Obtain the structured control instruction package and input it into the compensation amount calculation module. Based on the multivariate joint encoding data in the structured control instruction package, perform interpretation and dimension conversion processing to generate the final process semantic compensation amount with process semantic consistency. This process semantic compensation amount is directly used to drive the robotic arm to perform posture correction actions.

[0217] The system acquires a structured control instruction package encapsulated by a multivariate joint coding algorithm as input and calls the data interpretation unit of the compensation calculation module to perform field-level parsing on the instruction package, extracting the encoded data frames and time phase markers of the velocity component time series and current component time series. The extracted encoded data frames are then mapped from symbols to physical quantities according to a pre-defined interpretation mapping table, converting the encoded symbols into corresponding physical measurement value sequences, including velocity component reference value sequences and current component reference value sequences. The physical measurement value sequences are then converted to units based on the dimensional information embedded in the instruction package, unifying the data to the international unit system (SI) predefined by the compensation calculation module, with velocity components in meters per second and current components in amperes. Phase reconstruction is then performed on the dimensionally converted velocity and current sequences according to the time phase markers, using linear interpolation or spline interpolation to achieve subdivision and alignment of the time axis, ensuring data correspondence between the two components at the same time sampling point. Finally, an element-level coupling calculation is performed on the aligned velocity and current sequences using a proportional constraint coefficient matrix to construct a bivariate compensation vector with explicit proportional constraints. The output timing is adjusted based on the response timing difference coefficient, enabling the generated compensation quantity to achieve synchronous control at the execution layer that is semantically consistent with the process parameters. Through interpretation, dimension conversion, phase reconstruction, and proportional coupling operation, the structured control instruction package generated in the previous step is transformed into a final process semantic compensation quantity with process semantic consistency, realizing the native coupling drive of process parameters in the robotic arm posture correction stage.

[0218] For example, in a real high dynamic response welding scenario, the acquired structured control instruction package contains a velocity component time sequence and a current component time sequence recorded in symbolic encoding form, with a sampling frequency of 500Hz and a time phase marking accuracy of 0.1ms. The decoding unit identifies the code "1010" as a velocity of 0.25 in meters per second and the code "1100" as a current of 150 in amperes, obtaining the original physical value sequence after decoding. The dimension conversion module detects that some velocity data are in millimeters per second and converts them to meters per second; the current data contains extreme values ​​expressed in kiloamperes, which are converted according to the rule 1kA=1000A. The phase reconstruction process uses cubic spline interpolation to fill in the sampling points when the phase marking difference is less than 0.1ms, so that the velocity and current are completely synchronized in time. The proportional coupling operation uses the proportional constraint coefficient matrix [0.8,0.2;0.3,0.7] to perform element-wise operations on the bivariate sequence, outputting a compensation vector ( The system outputs the current component with a 5ms delay based on the response timing difference, and correspondingly advances the velocity component, thereby achieving coordinated control of heat input and molten pool stability. Tests on an automated welding verification platform show that the attitude correction response time after compensation is generated is significantly shortened, and weld formation consistency and welding quality stability are greatly improved.

[0219] Step S8: Based on the process semantic compensation amount, control the robotic arm to perform attitude correction actions, and after execution, collect new weld seam tracking deviation feature vectors to update the process parameter flow, forming closed-loop feedback control. Specifically, this includes:

[0220] S8.1: Perform inverse kinematic mapping on the process semantic compensation quantity carrying parameter coupling logic. Based on the current joint angle state of the robotic arm and the pose constraints of the end effector, convert the process semantic compensation quantity containing the velocity component and current component with deterministic proportional constraints into the target displacement increment sequence of each joint axis to generate a set of joint control instructions with physical executableness.

[0221] S8.2: Based on the generated joint control instruction set, the servo drive system of the robotic arm is adjusted by multi-axis collaborative interpolation. The dynamic tracking error is eliminated by using a feedforward compensation algorithm combined with a real-time feedback loop. The robotic arm end effector is driven to complete the posture correction action according to the preset response timing difference, so as to obtain the corrected actual welding posture of the weld.

[0222] S8.3: Multi-dimensional sensor data fusion is performed on the actual welding posture area of ​​the corrected weld. The weld geometry data and molten pool thermal distribution data are acquired simultaneously using a laser vision sensor and an arc spectral analysis module to form an original welding process observation dataset containing spatial location information and thermophysical state information.

[0223] S8.4: Based on the original welding process observation dataset, the deviation feature vector reconstruction calculation is performed. The point cloud registration algorithm is used to compare the real-time weld geometry data with the ideal weld model, and the temperature gradient anomaly value is extracted by combining the molten pool thermal distribution data to output a new weld tracking deviation feature vector that represents the current welding quality status.

[0224] S8.5: Perform timestamp alignment and data stream merging operations on the new weld tracking deviation feature vector and the current process parameter stream. Based on the sliding window mechanism, discard outdated historical data and inject the newly generated deviation feature into the data queue to update the process parameter stream containing the latest operating condition information, thereby forming a complete closed-loop feedback control loop.

[0225] For those skilled in the art, various other corresponding changes and modifications can be made based on the technical solutions and concepts described above, and all such changes and modifications should fall within the protection scope of the claims of this invention.

[0226] Unless otherwise defined, the technical or scientific terms used herein shall have the ordinary meaning as understood by one of ordinary skill in the art to which this application pertains. The terms “first,” “second,” “third,” and similar terms used in this patent application specification and claims do not indicate any order, quantity, or importance, but are merely used to distinguish different components. Similarly, the terms “an” or “a” and similar terms do not indicate a quantity limitation, but rather indicate the presence of at least one. The terms “comprising” or “including” and similar terms mean that the element or object preceding “comprising” or “including” encompasses the element or object listed following “comprising” or “including” and its equivalents, and do not exclude other elements or objects. The “multiple” mentioned in the embodiments of this application refers to two or more. A and / or B indicate three possibilities: A; B; and A and B.

[0227] The above description is merely an exemplary embodiment of this application, but the scope of protection of this application is not limited thereto. Any person skilled in the art can easily conceive of various equivalent modifications or substitutions within the technical scope disclosed in this application, and such modifications or substitutions should all be covered within the scope of protection of this application. Therefore, the scope of protection of this application should be determined by the scope of the claims.

Claims

1. A robotic arm welding deviation compensation and control method based on real-time weld seam tracking, specifically including: S1: Obtain the process parameter flow and weld tracking deviation feature vector during the welding process. The process parameter flow includes welding current, voltage, speed, wire feed rate, oscillation amplitude, and shielding gas flow rate. S2: Based on the attribute database, weld quality dataset and welding process knowledge base, each discrete or continuous parameter in the process parameter flow is modeled as a graph node, and semantic edges containing the coupling relationship between nodes including heat input, molten pool stability, spatter suppression and heat-affected zone are defined to construct a process parameter semantic graph. S3: Based on the material thickness, ambient temperature and weld curvature, prune the weak connection semantic edges with confidence levels below the threshold in the semantic graph of the process parameters to generate a working condition adaptation sparse subgraph. S4: Search the sparse subgraph of the working condition adaptation for the minimum closed-loop structure that covers all active parameter nodes and satisfies the semantic connectivity constraint to obtain the parameter cooperative subgraph. S5: Perform graph attention propagation calculation on the parameter cooperative subgraph to identify the parameter nodes that play a dominant regulatory role and their semantic neighborhoods, and extract implicit cooperative patterns; S6: Based on the implicit collaborative pattern, match the corresponding behavior template from the pre-set compensation knowledge base. The behavior template defines the proportional constraint relationship and response timing difference between the dominant compensation channel and the auxiliary compensation channel. S7: Convert the behavior template into a structured instruction, and perform compensation calculation based on the structured instruction. The compensation amount includes a velocity component and a current component with deterministic proportional constraints.

2. The robotic arm welding deviation compensation control method based on real-time weld seam tracking according to claim 1, characterized in that, Step S7 is followed by step S8, which specifically includes: S8: Control the robotic arm to perform posture correction actions according to the process semantic compensation amount, and collect new weld tracking deviation feature vectors after execution to update the process parameter flow, forming closed-loop feedback control.

3. The robotic arm welding deviation compensation control method based on real-time weld seam tracking according to claim 1, characterized in that, Step S2 specifically includes: The welding current, voltage, speed, wire feed rate, oscillation amplitude and shielding gas flow rate in the process parameter stream are obtained. Based on the attribute database, the data type of various parameters is identified and the dimensions are standardized to generate a set of standardized process parameters that meet the input format requirements of the graph node. Based on the standardized set of process parameters, each process parameter is instantiated into an independent graph node object using the statistical regularity of the weld quality dataset, and a node feature vector containing parameter name, real-time value and physical attribute label is assigned to it to construct a process parameter graph node set. By calling the rule entries in the welding process knowledge base regarding heat input calculation, molten pool dynamics, and spatter formation mechanism, the physical coupling relationship between any two nodes in the process parameter diagram node set is logically deduced, generating a candidate semantic edge list containing heat input dominance, molten pool stability constraints, spatter suppression correlation, and heat-affected zone superposition coupling types; Based on the candidate semantic edge list, an interpretable domain rule label is bound to each semantic edge and the weight decay function parameter is initialized. The confidence score of each physical coupling relationship under different working conditions is evaluated by quantification to generate a complete set of semantic edges. The process parameter graph node set and the complete semantic edge set are integrated, and graph topology assembly and connectivity verification are performed to output a process parameter semantic graph containing all active parameter nodes and their multidimensional semantic connections.

4. The robotic arm welding deviation compensation control method based on real-time weld seam tracking according to claim 1, characterized in that, Step S5 specifically includes: Based on the node attribute vectors and semantic edge connections in the parameter collaborative subgraph, a multi-head self-attention mechanism initialization process is performed on all active parameter nodes in the graph to generate an initial attention feature set containing a query matrix, a key matrix, and a value matrix, and to establish the basic mapping space for information interaction between nodes. Using the query matrix and key matrix in the initial attention feature set, the semantic correlation between any two nodes in the parameter collaboration subgraph is subjected to dot product operation and scaling normalization to calculate the original attention score matrix that reflects the dynamic dependence strength between parameters under the current working condition. Based on the original attention score matrix and the preset masking rules, negative infinity masking and soft maximum function normalization are performed on the score items corresponding to non-connected node pairs and low-confidence semantic edges to generate a standardized attention weight matrix that satisfies probability distribution constraints and retains only valid topological connections, thereby eliminating information interference from invalid paths. Based on the standardized attention weight matrix and the value matrix in the initial attention feature set, weighted aggregation and linear projection transformation are performed on the parameter collaborative subgraph to update the hidden state representation of each parameter node and generate an enhanced node embedding vector that integrates neighborhood context information. Based on the enhanced node embedding vector, the feature norms of each node in the parameter collaboration subgraph are sorted, filtered, and truncated to identify the dominant control parameter nodes with significantly higher attention weights than the average level and their directly connected semantic neighborhoods, thereby extracting the implicit collaboration patterns.

5. The robotic arm welding deviation compensation control method based on real-time weld seam tracking as described in claim 2, characterized in that, Each semantic edge is bound to an interpretable domain rule label and its weight decay function parameters are initialized. The dynamic confidence level is adaptively adjusted based on the parameter deviation using an exponential or logarithmic function for subsequent dynamic pruning.

6. The robotic arm welding deviation compensation control method based on real-time weld seam tracking as described in claim 1, characterized in that, The threshold removal rule for pruning is to use the normalized index of the working condition context and the mapping method of boundary constraints. The heat input relationship edge, the melt pool stability relationship edge, the splash suppression relationship edge and the heat-affected zone edge are all dynamically judged in real time. When the confidence is lower than 0.5 to 0.7, they are automatically removed.

7. The robotic arm welding deviation compensation control method based on real-time weld seam tracking as described in claim 1, characterized in that, The allocation of the dominant compensation channel and the auxiliary compensation channel in the behavior template is determined by weighted evaluation of the sensitivity and stability of the dominant parameter node. When the current thermal input is greater than the speed thermal input, the current is retained as the dominant compensation channel; otherwise, the speed is used as the dominant compensation channel.

8. The robotic arm welding deviation compensation control method based on real-time weld seam tracking as described in claim 1, characterized in that, In S6.4, the timing difference between the actions of the main / auxiliary compensation channels is calculated by the response time criterion based on the dynamics of heat input and the stability constraints of the molten pool. The action of the main channel can be preferably advanced, while the action of the auxiliary compensation channel is delayed, with a timing difference of 0.01 to 0.5 seconds.

9. The robotic arm welding deviation compensation control method based on real-time weld seam tracking as described in claim 1, characterized in that, The mathematical relationship between the velocity component and the current component in the compensation amount is as follows: velocity component V = k × dominant parameter + auxiliary parameter, current component I = dominant parameter + k × auxiliary parameter, k is preferably 0.05-0.5, and automatic saturation limit is applied when the process safety range is exceeded.

10. The robotic arm welding deviation compensation control method based on real-time weld seam tracking as described in claim 1, characterized in that, The node types and channel allocation schemes of the dominant compensation channel, auxiliary compensation channel, and parameter coordination subgraph can be configured according to process requirements to adapt to different types of welding materials, weld morphology, and working condition changes.

Images