A method for adaptive compensation of welding quality

By integrating a multi-source sensor signal acquisition system into the end effector of a welding robot, arc sound, current, voltage, and temperature signals are collected and processed in real time. Welding parameters are adjusted using an online evaluation model, which solves the problem of difficulty in real-time judgment of welding quality in existing technologies and improves the consistency and reliability of welding quality.

CN123103133APending Publication Date: 2026-11-17CHINACOAL BEIJING COAL MINING MACHINERY CO LTD
View PDF 0 Cites 0 Cited by

Patent Information

Application Number
CN202611399825.4
Authority / Receiving Office
CN · China
Patent Type
Applications(China)
Current Assignee / Owner
Filing Date
2026-09-10
Publication Date
2026-11-17

AI Technical Summary

Technical Problem

In the existing technology, the welding quality of workpieces such as hydraulic cylinder connector seats cannot be judged in real time, which makes it difficult to guarantee the consistency and reliability of welding quality. In particular, when the gap between workpieces fluctuates or the heat dissipation conditions change, the welding process parameters cannot be adjusted in time.

Method used

By integrating a multi-source sensor signal acquisition system at the end of the welding robot, the system collects arc sound signals, welding current values, welding voltage values, and temperature field signals in the weld area in real time, forming a multimodal feature sequence. This sequence is then input into a pre-trained online welding quality assessment model to judge the welding quality in real time, calculate the scoring deviation and rate of change, and adjust the welding process parameters through fuzzy inference to achieve adaptive compensation.

Benefits of technology

It enables real-time quality assessment and timely parameter adjustment during the welding process, improving the consistency and reliability of welding quality and ensuring the stability of the welding process.

✦ Generated by Eureka AI based on patent content.

Smart Images

  • Figure CN123103133A_ABST
    Figure CN123103133A_ABST
Patent Text Reader

Abstract

The present application relates to a kind of welding quality self-adaptive compensation method, it relates to intelligent welding manufacturing technical field, the method includes: the average of the welding quality score of the continuous multiple time windows before current time window is calculated, and the welding quality score of current time window is compared with average, when the welding quality score of current time window is lower than average and its difference exceeds preset amplitude, the score deviation between the welding quality score of current time window and preset target score and score deviation change rate are calculated, and score deviation and score deviation change rate are input, and welding process parameter correction amount is obtained by fuzzy inference, and current welding process parameter is revised according to welding process parameter correction amount, and the welding process parameter after revision is obtained, and continue to execute welding with the welding process parameter after revision, so that welding process parameter is adjusted in time when welding quality appears decline, realize the self-adaptive compensation of welding quality.
Need to check novelty before this filing date? Find Prior Art

Description

Technical Field

[0001] This invention relates to the field of intelligent welding manufacturing technology, and in particular to an adaptive compensation method for welding quality. Background Technology

[0002] Currently, the welding process for workpieces such as hydraulic cylinder connector seats is typically performed according to preset welding process parameters. Welding current, welding voltage, welding speed, and oscillation amplitude are set before welding begins and remain constant throughout the process. Weld quality evaluation mainly relies on post-weld inspection, i.e., after welding is completed and cooled, the weld is judged for its quality through visual inspection or non-destructive testing. This method cannot assess the current welding quality in real time during the welding process, nor can it adjust the welding process parameters promptly when the welding quality deteriorates. When interference factors such as fluctuations in workpiece assembly gaps or changes in heat dissipation conditions occur during welding, the actual welding quality may deviate from expectations. Existing methods lack real-time sensing and dynamic adjustment capabilities, making it difficult to effectively guarantee the consistency and reliability of welding quality. Summary of the Invention

[0003] (a) Technical problems to be solved In view of the above-mentioned shortcomings and deficiencies of the prior art, the present invention provides a welding quality adaptive compensation method, which solves the technical problem that the consistency and reliability of welding quality are difficult to be effectively guaranteed in the prior art.

[0004] (II) Technical Solution To achieve the above objectives, the main technical solutions adopted by the present invention include: In a first aspect, embodiments of the present invention provide a welding quality adaptive compensation method, comprising: A multi-source sensor signal acquisition system integrated into the end effector of a welding robot synchronously acquires arc acoustic signals, welding current values, welding voltage values, and weld zone temperature field signals during the welding process. The multi-source sensor signal acquisition system includes an arc acoustic sensor for acquiring arc acoustic signals, a current acquisition module for acquiring welding current values, a voltage acquisition module for acquiring welding voltage values, and an infrared thermal imager for acquiring weld zone temperature field signals. The acquired arc acoustic signals, welding current values, welding voltage values, and weld zone temperature field signals are preprocessed to extract arc acoustic features, current features, voltage features, and temperature field features, forming a multi-modal feature sequence. This multi-modal feature sequence is then input into a pre-trained welding quality control system. The linear evaluation model obtains the welding quality score for the current time window; it calculates the average welding quality score of multiple consecutive time windows preceding the current time window and compares the welding quality score of the current time window with the average. When the welding quality score of the current time window is lower than the average and the difference exceeds a preset range, it calculates the score deviation and the rate of change of the score deviation between the welding quality score of the current time window and the preset target score. Using the score deviation and the rate of change of the score deviation as input, it obtains the welding process parameter correction amount through fuzzy inference, and corrects the current welding process parameters according to the welding process parameter correction amount to obtain the corrected welding process parameters. Welding continues to be performed with the corrected welding process parameters to achieve adaptive compensation.

[0005] Optionally, the online welding quality assessment model includes: The physical prior weighted fusion layer is used to determine the physical correlation weights between each modal feature based on the physical generation mechanism of the arc sound feature, current feature, voltage feature and temperature field feature and the corresponding sensor installation pose. Based on the physical correlation weights, the arc sound feature, current feature, voltage feature and temperature field feature in the multimodal feature sequence are weighted and fused to obtain the multimodal feature tensor calibrated by physical prior. The multi-scale temporal convolutional fusion layer contains multiple temporal convolutional branches with different dilation rates. Each temporal convolutional branch performs temporal convolution processing on the multimodal feature tensor in parallel to obtain the temporal feature representation corresponding to each temporal convolutional branch. The temporal feature representations corresponding to each temporal convolutional branch are then adaptively fused to obtain multi-scale temporal features. The temporal causal self-attention layer is used to encode the position of multi-scale temporal features along the time dimension to obtain the position-encoded temporal features. The position-encoded temporal features are then processed by a multi-head self-attention mechanism with causal masking to obtain context-aware features. A cross-modal attention gating fusion layer is used to perform multimodal cross-attention interaction on context-aware features to obtain multimodal features after interaction. The layer also dynamically calculates the confidence weight of each modality based on the current welding condition and adaptively filters and fuses the multimodal features after interaction based on the confidence weight to obtain fused features. The prediction output layer is used to map the fused features to the welding quality score for the current time window.

[0006] Optionally, the physical prior weighted fusion layer determines the physical association weights between each modal feature, specifically including: obtaining the spatial pose parameters of the welding torch relative to the workpiece under the current welding condition; wherein, the spatial pose parameters include the welding torch tilt angle, the welding torch pointing angle, and the distance between the welding torch tip and the workpiece surface; based on the installation pose of the arc acoustic sensor and the spatial pose parameters of the welding torch relative to the workpiece, calculating the propagation path distance of the arc acoustic signal radiated from the molten pool to the arc acoustic sensor and the angle between the propagation direction and the sensor axis, and based on the propagation path distance and the angle, calculating the sound wave attenuation compensation coefficient according to the law that the sound pressure attenuates inversely with distance, and using it as the first physical association weight of the arc acoustic feature; based on the installation pose of the infrared thermal imager and the spatial pose parameters of the welding torch relative to the workpiece, calculating the heat conduction path length and the heat diffusion time constant of the temperature field in the weld area from the center of the molten pool to the boundary of the heat-affected zone, and based on the heat conduction path length and the heat diffusion time constant, calculating the sound wave attenuation compensation coefficient according to the law that the temperature field response lags with the heat conduction... Based on the increasing relationship of the square root of the depth, the temperature field hysteresis compensation coefficient is calculated and used as the second physical correlation weight of the temperature field feature. According to the physical coupling relationship between welding current, welding voltage and arc acoustic power, a theoretical mapping function is established with the effective values ​​of welding current and welding voltage as inputs and arc acoustic power as output. The deviation between the theoretical acoustic power and the measured acoustic power corresponding to the actual collected welding current and welding voltage is calculated, and the weight adjustment coefficient is calculated according to the deviation and the preset negative correlation mapping function, and used as the third physical correlation weight of the current-voltage feature. The larger the deviation, the smaller the weight adjustment coefficient. After normalizing the first, second and third physical correlation weights, the first physical correlation weight is applied to the arc acoustic feature, the second physical correlation weight is applied to the temperature field feature, and the third physical correlation weight is applied to both the current and voltage features to obtain a multimodal feature tensor calibrated by physical prior.

[0007] Optionally, the collected arc sound signal, welding current value, welding voltage value, and weld area temperature field signal are preprocessed to extract arc sound features, current features, voltage features, and temperature field features, respectively. This includes: performing bandpass filtering and frame-by-frame windowing on the arc sound signal in sequence to extract arc sound features including Mel frequency cepstral coefficients, short-time energy, and spectral centroid; performing median filtering and moving average filtering on the welding current value and welding voltage value respectively to extract current features including mean, standard deviation, and peak factor; and performing spatial filtering on the weld area temperature field signal to extract temperature field features including the highest temperature of the molten pool, the average temperature of the heat-affected zone, and the temperature gradient.

[0008] Optionally, the welding process parameter correction amount is obtained by taking the scoring deviation and the rate of change of scoring deviation as inputs and through fuzzy inference, including: mapping the scoring deviation and the rate of change of scoring deviation to a preset fuzzy domain to obtain the deviation fuzzy amount and the deviation rate of change fuzzy amount; inputting the deviation fuzzy amount and the deviation rate of change fuzzy amount into a preset fuzzy inference rule base to obtain the fuzzy output amount; and defuzzifying the fuzzy output amount to obtain the welding process parameter correction amount.

[0009] Optionally, before inputting the deviation fuzzy value and the deviation change rate fuzzy value into the preset fuzzy inference rule base, the method further includes: obtaining the current welding process parameter combination; the current welding process parameter combination includes the welding current setting value, welding voltage setting value, welding speed setting value, and oscillation amplitude setting value; according to the current welding process parameter combination, querying the corresponding membership function shape parameter in the preset process parameter-membership parameter mapping table; and dynamically constructing the deviation membership function and deviation change rate membership function within the current control cycle based on the queried membership function shape parameter. Furthermore, the scoring deviation and the rate of change of scoring deviation are mapped to a preset fuzzy domain to obtain the fuzzy quantity of deviation and the fuzzy quantity of the rate of change of deviation, specifically including: fuzzifying the scoring deviation input into the deviation membership function to obtain the fuzzy quantity of deviation; and fuzzifying the rate of change of scoring deviation input into the rate of change of deviation membership function to obtain the fuzzy quantity of the rate of change of deviation.

[0010] Optionally, after continuing welding with the corrected welding process parameters, the method further includes: obtaining the welding quality scores for multiple consecutive time windows after correction; when the welding quality scores for multiple consecutive time windows after correction are not lower than a preset recovery threshold, multiplying the welding process parameter correction amount by a preset attenuation coefficient to obtain a reduced welding process parameter correction amount, and correcting the current welding process parameters according to the reduced welding process parameter correction amount; and triggering a manual intervention alarm when the welding quality scores for multiple consecutive time windows after correction are still lower than the preset recovery threshold.

[0011] Optionally, calculating the scoring deviation and the rate of change of the scoring deviation between the welding quality score of the current time window and the preset target score includes: calculating the difference between the welding quality score of the current time window and the preset target score to obtain the scoring deviation; obtaining the welding quality scores of multiple consecutive time windows before the current time window; and performing linear fitting on the welding quality scores of multiple consecutive time windows before the current time window, and using the slope obtained from the fitting as the rate of change of the scoring deviation.

[0012] Optionally, after continuing welding with the corrected welding process parameters, the process further includes: constructing a welding quality time-series traceability map; wherein the welding quality time-series traceability map uses the welding time window sequence as the horizontal axis, and synchronously displays the welding quality score, welding process parameter correction amount, and feature values ​​in the multimodal feature sequence on the same visualization interface; responding to the operator's selection of any time window in the welding quality time-series traceability map, retrieving and replaying the synchronously recorded welding process data within a preset time period before and after the time window; wherein the synchronously recorded welding process data includes the arc sound signal audio segment, the waveform segments of the welding current value and the welding voltage value, and the thermal image sequence of the temperature field of the weld area corresponding to the time window; based on the intervention start time window and intervention end time window marked by the operator in the welding quality time-series traceability map, extracting the multimodal feature sequence, welding quality score sequence, and welding process parameter correction amount sequence within the corresponding time interval as typical quality event samples, and storing the typical quality event samples in the welding quality event knowledge base for subsequent adjustment of the rule confidence of the fuzzy inference rule base under similar working conditions and incremental training of the online welding quality evaluation model.

[0013] Optionally, the welding process parameter corrections include welding current corrections, welding voltage corrections, welding speed corrections, and oscillation amplitude corrections.

[0014] In a second aspect, embodiments of the present invention provide an industrial control computer for executing the welding quality adaptive compensation method as described in the first aspect.

[0015] (III) Beneficial Effects The beneficial effects of this invention are: This invention proposes an adaptive compensation method for welding quality. A multi-source sensor signal acquisition system integrated into the end effector of a welding robot synchronously acquires arc sound signals, welding current values, welding voltage values, and temperature field signals in the weld area during the welding process. The acquired arc sound signals, welding current values, welding voltage values, and temperature field signals are preprocessed, and arc sound features, current features, voltage features, and temperature field features are extracted to form a multimodal feature sequence. This multimodal feature sequence is then input into a pre-trained online welding quality assessment model to obtain the welding quality score for the current time window, thereby enabling real-time judgment of the current welding quality during the welding process. When the welding quality score of the current time window is lower than the average of the welding quality scores of the previous multiple consecutive time windows and the difference exceeds a preset range, the scoring deviation and the rate of change of the scoring deviation between the welding quality score of the current time window and the preset target score are calculated. The welding process parameter correction amount is obtained by fuzzy inference using the scoring deviation and the rate of change of the scoring deviation as input. The current welding process parameters are corrected according to the welding process parameter correction amount to obtain the corrected welding process parameters. Welding continues to be performed with the corrected welding process parameters. In this way, the welding process parameters are adjusted in a timely manner when the welding quality declines, so as to achieve adaptive compensation of welding quality and improve the consistency and reliability of welding quality. Attached Figure Description

[0016] Figure 1 A flowchart of a welding quality adaptive compensation method provided in an embodiment of this application is shown; Figure 2 A schematic diagram of the structure of an online welding quality evaluation model provided in an embodiment of this application is shown. Detailed Implementation

[0017] To better explain and facilitate understanding of the present invention, the present invention will be described in detail below with reference to the accompanying drawings and specific embodiments.

[0018] This invention proposes an adaptive compensation method for welding quality. A multi-source sensor signal acquisition system integrated into the end effector of a welding robot synchronously acquires arc sound signals, welding current values, welding voltage values, and temperature field signals in the weld area during the welding process. The acquired arc sound signals, welding current values, welding voltage values, and temperature field signals are preprocessed, and arc sound features, current features, voltage features, and temperature field features are extracted to form a multimodal feature sequence. This multimodal feature sequence is then input into a pre-trained online welding quality assessment model to obtain the welding quality score for the current time window, thereby enabling real-time judgment of the current welding quality during the welding process. When the welding quality score of the current time window is lower than the average of the welding quality scores of the previous multiple consecutive time windows and the difference exceeds a preset range, the scoring deviation and the rate of change of the scoring deviation between the welding quality score of the current time window and the preset target score are calculated. The welding process parameter correction amount is obtained by fuzzy inference using the scoring deviation and the rate of change of the scoring deviation as input. The current welding process parameters are corrected according to the welding process parameter correction amount to obtain the corrected welding process parameters. Welding continues to be performed with the corrected welding process parameters. In this way, the welding process parameters are adjusted in a timely manner when the welding quality declines, so as to achieve adaptive compensation of welding quality and improve the consistency and reliability of welding quality.

[0019] To better understand the above technical solutions, exemplary embodiments of the present invention will be described in more detail below with reference to the accompanying drawings. Although exemplary embodiments of the present invention are shown in the drawings, it should be understood that the present invention can be implemented in various forms and should not be limited to the embodiments set forth herein. Rather, these embodiments are provided so that the present invention can be understood more clearly and thoroughly, and that the scope of the present invention can be fully conveyed to those skilled in the art.

[0020] Please see Figure 1 , Figure 1 A flowchart of a welding quality adaptive compensation method provided in an embodiment of this application is shown. Specifically, the method can be executed by an industrial control computer; the method includes: Step S110 involves simultaneously acquiring arc sound signals, welding current values, welding voltage values, and weld area temperature field signals during the welding process using a multi-source sensor signal acquisition system integrated at the end effector of the welding robot. This multi-source sensor signal acquisition system includes an arc sound sensor for acquiring arc sound signals during welding, a current acquisition module for acquiring welding current values, a voltage acquisition module for acquiring welding voltage values, and an infrared thermal imager for acquiring weld area temperature field signals.

[0021] Specifically, a multi-source sensor signal acquisition system is integrated at the end effector of the welding robot. This system includes an arc sound sensor for acquiring arc sound signals during welding, a current acquisition module for acquiring welding current values, a voltage acquisition module for acquiring welding voltage values, and an infrared thermal imager for acquiring temperature field signals in the weld area. Each component is installed at the end effector of the welding robot near the welding torch and moves with the torch to ensure that multi-source sensor signals corresponding to the current welding position can be acquired in real time during the welding process.

[0022] The arc acoustic sensor is used to collect the acoustic signals radiated from the molten pool and arc area during welding. This arc acoustic sensor can be an electret condenser microphone or a piezoelectric acoustic emission sensor, and is installed 50-80mm from the center axis of the welding torch nozzle, with the sensor pointing towards the molten pool. A wind shield and a metal spatter protection mesh are installed at the front end of the arc acoustic sensor. The wind shield reduces interference from ambient airflow on the acoustic signal acquisition, and the metal spatter protection mesh prevents damage to the sensor from welding spatter. The sampling frequency of the arc acoustic sensor is no less than 20kHz to ensure complete capture of the high-frequency components of the arc acoustic signal during welding. For example, in practical implementation, an electret condenser microphone with a frequency response range of 50Hz-20kHz can be selected, and the sampling frequency can be set to 25kHz to effectively acquire the arc acoustic signal.

[0023] Furthermore, current and voltage acquisition modules are used to acquire welding current and welding voltage values ​​during the welding process, respectively. Both modules are installed between the welding torch cable and the welding power source to directly obtain the electrical parameters in the welding circuit. The current sampling accuracy of the current acquisition module is no greater than ±1A, and the voltage sampling accuracy of the voltage acquisition module is no greater than ±0.1V. Both modules have a sampling frequency of no less than 10kHz. Through these high sampling frequencies, the current and voltage acquisition modules can accurately reflect the real-time changes in welding current and welding voltage values, including transient fluctuations that may occur during pulse welding. For example, a Hall current sensor can be used as the core component of the current acquisition module, and a Hall voltage sensor or voltage transmitter can be used as the core component of the voltage acquisition module. Both modules can have their sampling frequency set to 12.5kHz to ensure the accuracy and real-time performance of the electrical signal acquisition.

[0024] Additionally, an infrared thermal imager is used to acquire temperature field signals in the weld area. This infrared thermal imager is installed at the end effector of the welding robot near the welding torch. Its temperature measurement range is 0-650℃, with a measurement accuracy of ±1%, and a spectral response range of 8-14. The spatial resolution should be no less than 320×240 pixels. The field of view of the infrared thermal imager can be selected according to the actual size of the welded workpiece. For example, an infrared thermal imager with a field of view of 24°×18° can be selected to acquire the temperature field distribution data of the weld and its surrounding 75mm range in real time. The frame rate of the infrared thermal imager can be set as needed, such as 30Hz, in order to capture the dynamic changes in the temperature field of the molten pool and heat-affected zone during the welding process.

[0025] To achieve synchronous data acquisition from the arc sound sensor, current acquisition module, voltage acquisition module, and infrared thermal imager, all sensors and modules are connected to the industrial control computer via a data acquisition card. The data acquisition card has a sampling accuracy of at least 16 bits and features multi-channel synchronous sampling, enabling it to simultaneously receive multiple sensor signals. The industrial control computer is equipped with a synchronous trigger acquisition module, which sends synchronous trigger signals to the data acquisition card and each sensor to ensure strict synchronization in time between the arc sound signal acquired by the arc sound sensor, the welding current value acquired by the current acquisition module, the welding voltage value acquired by the voltage acquisition module, and the temperature field signal of the weld area acquired by the infrared thermal imager, with a time synchronization accuracy of no more than 1ms. This ensures that the obtained arc sound features, current features, voltage features, and temperature field features are accurately aligned in time during subsequent preprocessing and feature extraction of the multi-source signals, providing a reliable data foundation for forming a multimodal feature sequence.

[0026] In actual welding, synchronous data acquisition can be performed as follows: an arc sound sensor acquires arc sound signals near the molten pool at a sampling frequency of 25 kHz; current and voltage acquisition modules acquire welding current and welding voltage values ​​at sampling frequencies of 12.5 kHz, respectively; and an infrared thermal imager acquires a sequence of thermal images of the weld area temperature field at a frame rate of 30 Hz. These four signals are initiated by the same synchronous trigger signal and stored in the industrial control computer according to a unified timestamp, ensuring that the arc sound signal, welding current value, welding voltage value, and weld area temperature field signal within each time window correspond to the same welding moment. Through the arrangement and synchronous acquisition method of the above multi-source sensor signal acquisition system, time-aligned multimodal sensor data can be provided for online welding quality assessment.

[0027] The sampling frequency and frame rate mentioned above are for illustrative purposes only. In practical applications, they can be adjusted appropriately based on welding process requirements, sensor performance, and the processing capabilities of the data acquisition system. The four signals are initiated by the same synchronous trigger signal and stored in the industrial control computer according to a unified timestamp, ensuring that the arc sound signal, welding current value, welding voltage value, and weld area temperature field signal within each time window correspond to the same welding moment. Through the arrangement and synchronous acquisition method of the multi-source sensor signal acquisition system, time-aligned multimodal sensor data can be provided for online welding quality assessment.

[0028] Step S120: The collected arc sound signal, welding current value, welding voltage value and weld area temperature field signal are preprocessed to extract arc sound features, current features, voltage features and temperature field features to form a multimodal feature sequence.

[0029] Specifically, during the welding process, after the industrial control computer receives the arc sound signal collected by the arc sound sensor, the welding current value collected by the current acquisition module, the welding voltage value collected by the voltage acquisition module, and the temperature field signal of the weld area collected by the infrared thermal imager, it needs to preprocess and extract features from these signals to obtain arc sound features, current features, voltage features, and temperature field features that can characterize the welding state. Furthermore, these features are organized into a multimodal feature sequence according to the time sequence, which is used as the input of the online welding quality evaluation model.

[0030] For the arc sound signal, bandpass filtering is performed first. Since there is low-frequency environmental noise at the welding site and high-frequency electrical noise that may be introduced by the sensor itself, bandpass filtering can retain the effective frequency band sensitive to welding quality while suppressing out-of-band interference. For example, the passband frequency range can be set from 500Hz to 20kHz, implemented using a fourth-order Butterworth filter. After filtering, the arc sound signal is processed by framing and windowing. Framing divides the continuous sound signal into short-time frames to perform feature calculations under the assumption of stationarity within each frame. For example, the frame length can be set to 25ms, the frame shift to 10ms, and a Hamming window can be used as the window function to reduce spectral leakage at frame boundaries. For each frame, arc sound features such as Mel-frequency cepstral coefficients, short-time energy, and spectral centroid are extracted. Among them, the Mel frequency cepstral coefficients reflect the spectral envelope characteristics of the acoustic signal at the Mel frequency scale. Their calculation can employ a 24-Meler filter bank, covering a frequency range from 200Hz to 20kHz, with the first 13 dimensions of coefficients used as features. Short-time energy reflects the energy magnitude of the acoustic signal in that frame; the spectral centroid reflects the concentrated frequency position of the acoustic signal's spectral energy. These features collectively constitute the arc sound feature vector for the corresponding time frame.

[0031] Furthermore, median filtering and moving average filtering are applied to the welding current and welding voltage values, respectively. Median filtering removes spike pulse interference caused by momentary short circuits or spatter in the arc, and its filtering window length can be set to 5 to 7 points. Moving average filtering smooths high-frequency random fluctuations in the welding electrical signal, and its window length can be set to 10 to 20 points. After filtering, current and voltage characteristics, including mean, standard deviation, and peak value, are extracted from the welding current and welding voltage values, respectively. The mean reflects the average level of the welding current or welding voltage value within the corresponding time window, the standard deviation reflects its fluctuation degree, and the peak value reflects the ratio of the signal peak value to the effective value, which can be used to characterize arc stability. In practical implementation, the above statistics can be calculated for the welding current and welding voltage values ​​within the same time window to obtain the current feature vector and voltage feature vector.

[0032] Furthermore, regarding the temperature field signal in the weld area, since the infrared thermal imager outputs a sequence of temperature field thermal images, each frame corresponds to a two-dimensional temperature distribution at a given moment. To extract temperature field features, spatial filtering is first applied to each frame of the thermal image, for example, using a 3×3 window median filter to suppress isolated dead pixels and random noise in the thermal image. Then, temperature field features, including the highest temperature of the molten pool, the average temperature of the heat-affected zone (HAZ), and the temperature gradient, are extracted from the filtered thermal images. The highest temperature of the molten pool is the highest temperature value in the thermal image; the average temperature of the HAZ can be taken as the average temperature of the area around the weld where the temperature exceeds a certain threshold, such as an area with a temperature not lower than 100℃; the temperature gradient can be taken as the maximum rate of temperature change from the center of the molten pool to the edge of the HAZ. These features reflect changes in welding heat input and heat dissipation conditions.

[0033] After extracting the modal features described above, the arc sound features, current features, voltage features, and temperature field features obtained within the same time window are concatenated into a single feature vector, which serves as the multimodal feature vector corresponding to that time window. As the welding process continues, the multimodal feature vectors from multiple consecutive time windows are arranged in chronological order, forming a multimodal feature sequence. This multimodal feature sequence can simultaneously carry welding process information across three physical dimensions: arc sound, electrical, and thermal, providing an input basis for the online welding quality assessment model. For example, in actual welding, a time window can be generated every 0.5 seconds, and the aforementioned modal features can be extracted within each time window and stored sequentially to form a multimodal feature sequence for model processing.

[0034] Step S130: Input the multimodal feature sequence into the pre-trained online welding quality evaluation model to obtain the welding quality score for the current time window.

[0035] It should be understood that the specific model and structure of the online welding quality assessment model can be set according to actual needs, and the embodiments of this application are not limited thereto.

[0036] Optionally, please see Figure 2 , Figure 2 A schematic diagram of the structure of an online welding quality evaluation model provided in an embodiment of this application is shown. Figure 2 As shown, the model includes a physical prior weighted fusion layer, a multi-scale temporal convolutional fusion layer, a temporal causal self-attention layer, a cross-modal cross-attention gating fusion layer, and a prediction output layer connected in sequence.

[0037] Specifically, this online welding quality assessment model first establishes a physical prior weighted fusion layer. During the welding process, the arc sound signal collected by the arc sound sensor, the welding current and welding voltage values ​​obtained by the current and voltage acquisition modules, and the temperature field signal of the weld area collected by the infrared thermal imager, although reflecting the welding state from different physical dimensions, undergo different physical propagation paths and varying degrees of attenuation or lag as they travel from the molten pool to being received by the sensors. For example, during the process of the arc sound signal radiating from the molten pool to the sensor, the sound pressure attenuates with the increase of propagation distance, and the directivity of the sensor also affects the receiving sensitivity; the temperature field signal takes a certain amount of time to propagate from the center of the molten pool to the boundary of the heat-affected zone, so the temperature field collected by the infrared thermal imager at a certain moment actually reflects the thermal diffusion state at several moments prior; although the welding current and welding voltage values ​​have a relatively fast response speed as electrical parameters, their measurement reliability is also affected when there is a large amount of welding spatter or strong electromagnetic interference. If the modal features are treated as equals and fused without considering the differences in physical propagation and attenuation, the quality of the model input data will be reduced, which will affect the accuracy of subsequent evaluations.

[0038] Based on this, after receiving the multimodal feature sequence formed after the aforementioned preprocessing, the physical prior weighted fusion layer first determines the physical association weight of each modal feature according to the physical generation mechanism of each type of signal and the corresponding sensor mounting pose. When determining this weight, the physical prior weighted fusion layer needs to obtain the spatial pose parameters of the welding torch relative to the workpiece under the current welding condition. These parameters specifically include the welding torch tilt angle, the welding torch pointing angle, and the distance between the welding torch tip and the workpiece surface. These spatial pose parameters can be provided in real time by the welding robot's controller and change with the movement of the robot's joints during the welding process.

[0039] Based on this, for the first physical association weight corresponding to the arc sound characteristics, the physical prior weighted fusion layer calculates the propagation path distance of the arc sound signal radiating from the molten pool to the arc sound sensor, as well as the angle between this propagation direction and the sensor axis, according to the installation posture of the arc sound sensor and the aforementioned spatial posture parameters of the welding torch relative to the workpiece. The longer the propagation path distance, the greater the attenuation of the sound wave during propagation in the air, and the smaller the actual sound pressure received by the sensor; the larger the angle, the farther the sound wave deviates from the sensor's sensitive axis, the smaller the effective receiving area of ​​the sensor, and the correspondingly reduced receiving sensitivity. The physical prior weighted fusion layer comprehensively considers the above propagation path distance and angle factors, and calculates the sound wave attenuation compensation coefficient according to the basic physical law that sound pressure attenuates inversely with distance, and uses this coefficient as the first physical association weight of the arc sound characteristics. The specific meaning of the first physical correlation weight is as follows: when the sensor is far from the molten pool or the deflection angle is large, the measured arc sound signal deviates from the true sound radiation state of the molten pool due to physical attenuation, and its reliability is relatively reduced. Therefore, the weight value is reduced accordingly. Conversely, when the sensor is close to the molten pool and faces the direction of the molten pool, the weight value is larger.

[0040] For example, during the welding of a hydraulic cylinder connector, the welding robot controller provides the welding torch tip spatial coordinates as (100, 200, 50) mm (the origin of the coordinates is a fixed reference point on the welding station). The arc acoustic sensor's installation offset on the robot's end flange is (30, 0, -20) mm, meaning the sensor is offset by 30 mm horizontally and 20 mm vertically relative to the welding torch tip. The center of the molten pool is approximately taken as the projection point of the welding torch tip onto the workpiece surface, i.e., (100, 200, 0) mm. Therefore, the propagation path distance of the sound wave from the center of the molten pool to the arc acoustic sensor can be calculated to be approximately 76 mm. Furthermore, based on the sensor's installation direction, its sensitive axis points perpendicularly to the molten pool, while the angle between the actual propagation direction of the sound wave and the sensor's sensitive axis is approximately 15 degrees. In this case, since the propagation distance is only 76 mm, the attenuation of the sound wave in the air is relatively limited, and the included angle is small and the sensor directivity loss is small. Therefore, the calculated sound wave attenuation compensation coefficient is relatively large, about 0.92, which indicates that the currently acquired arc sound signal has high reliability. This weight is used to enhance the arc sound feature channel so that it maintains a high contribution ratio in the subsequent fusion.

[0041] For the second physical correlation weight corresponding to the temperature field characteristics, the physical prior weighted fusion layer calculates the heat conduction path length from the center of the molten pool to the boundary of the heat-affected zone in the weld area based on the installation pose of the infrared thermal imager and the aforementioned spatial pose parameters of the welding torch relative to the workpiece. It also determines the heat diffusion time constant by combining this with the material thermal diffusion characteristics of the current welded workpiece. Since heat conduction in the base metal takes time, the further away from the center of the molten pool, the more significant the lag in temperature response relative to the temperature change at the center. That is, the temperature value of a region far from the molten pool collected by the infrared thermal imager at a certain moment actually reflects the thermal state of the molten pool several milliseconds or even seconds prior, rather than the real-time temperature at the current moment. Based on the physical relationship that the temperature field response lag increases with the square root of the heat conduction depth, the physical prior weighted fusion layer calculates the temperature field lag compensation coefficient according to the aforementioned heat conduction path length and heat diffusion time constant, and uses this coefficient as the second physical correlation weight for the temperature field characteristics. The specific meaning of the second physical correlation weight is as follows: when the heat conduction path is long and the heat diffusion time constant is large, the lag effect of the temperature field signal is significant, and the correspondence between the temperature field characteristics collected at the current moment and the actual state of the molten pool is weak. Therefore, the weight value is reduced accordingly. When the heat conduction path is short and the heat diffusion time constant is small, the temperature field signal can reflect the changes in the thermal state of the molten pool more promptly, and the weight value is increased accordingly.

[0042] For example, in a hydraulic cylinder connector welding process, the base material of the connector to be welded is 27SiMn steel, which has a thermal diffusivity of approximately 5 mm² / s within the welding temperature range. An infrared thermal imager is mounted on the end effector of the welding robot, maintaining a relatively fixed position relative to the welding torch, with its optical axis pointing towards the molten pool and its surrounding area. Once the welding process reaches a stable stage, the system tracks the temperature distribution pixel-by-pixel from the center of the molten pool outwards along the radial direction based on the real-time temperature field image acquired by the infrared thermal imager, identifying the isotherm where the temperature drops to 100 degrees Celsius as the boundary of the heat-affected zone (HAZ). Assuming that at the current moment, the system detects an equivalent heat conduction path length of 8 mm between the center of the molten pool and this 100-degree Celsius isotherm boundary, this value reflects the thickness of the base material that heat must traverse from the molten pool to the HAZ boundary. Based on this, and combined with the aforementioned material thermal diffusivity, the system calculates a thermal diffusion time constant of approximately 3.2 seconds. This time constant means that the temperature change at the HAZ boundary lags behind the temperature change at the center of the molten pool by approximately 3.2 seconds. In other words, the temperature value at the boundary of the heat-affected zone acquired by the infrared thermal imager at the current moment actually reflects the thermal state of the molten pool center approximately 3.2 seconds ago. Under these circumstances, the temperature field lag effect is significant, and the calculated temperature field lag compensation coefficient is approximately 2.5. This indicates a large deviation between the currently acquired temperature field characteristics and the actual state of the molten pool. While this weight is relatively large, it signifies a significant lag, and the contribution of the temperature field characteristic channel needs to be appropriately suppressed to prevent the model from overly relying on the current temperature field data to determine the real-time state of the molten pool.

[0043] For the third physical correlation weights corresponding to current and voltage characteristics, the physical prior weighted fusion layer determines them based on the physical coupling relationship between welding current, welding voltage, and arc acoustic power. Specifically, the arc energy input during welding is jointly determined by welding current and welding voltage, and changes in arc energy directly affect the radiation intensity of arc sound. Therefore, under ideal welding conditions, there is a definite theoretical mapping relationship between welding current and welding voltage values ​​and arc acoustic power. The physical prior weighted fusion layer pre-establishes a theoretical mapping function with the effective values ​​of welding current and welding voltage as inputs and arc acoustic power as output. This function reflects the correspondence between electrical and acoustic parameters under normal welding conditions. During actual welding, the physical prior weighted fusion layer substitutes the currently acquired welding current and welding voltage values ​​into this theoretical mapping function to calculate the theoretical acoustic power corresponding to the current electrical parameters. Then, it compares this theoretical acoustic power with the acoustic power actually acquired by the arc acoustic sensor and calculates the deviation between the two. When the deviation is small, it indicates that the current, voltage, and arc sound characteristics at the current moment corroborate each other, and both have high reliability. When the deviation is large, it indicates that at least one of the electrical or acoustic parameters has been disturbed, and the reliability of both has decreased. Therefore, the physical prior weighted fusion layer calculates a weight adjustment coefficient based on this deviation, according to the negative correlation mapping relationship where the larger the deviation, the smaller the weight. This coefficient is then used as the third physical correlation weight for the current-voltage characteristics. The specific meaning of this third physical correlation weight is: when the deviation between the theoretical sound power and the measured sound power is small, the contribution weight of the current and voltage characteristics in subsequent fusion remains at a high level; when the deviation is large, the contribution weight of the current and voltage characteristics is reduced accordingly to prevent unreliable electrical characteristics from misleading the quality assessment.

[0044] Regarding the aforementioned theoretical mapping function, in the gas metal arc welding (GMAW) process, the generation mechanism of arc sound involves multiple physical processes, including the oscillation of the arc plasma, the vaporization of the molten metal, and the noise from the jetting of shielding gas and metal vapor. However, overall, there is a definite nonlinear positive correlation between the radiated power of the arc sound and the arc input power. Under standard welding conditions, with stable welding process parameters, sufficient shielding gas flow, reasonable wire extension, and appropriate welding torch angle, the relationship between welding current and welding voltage and arc sound power can be established through calibration experiments. Specifically, this theoretical mapping function can be constructed as the product of the power of the effective value of the welding current and the power of the effective value of the welding voltage, multiplied by a proportionality coefficient. Where the power of the welding current is greater than 1, it reflects the high sensitivity of the arc sound power to current changes; the power of the welding voltage is less than 1, reflecting the relatively mild impact of voltage changes on the sound power. The values ​​of the proportionality coefficient and the two powers are determined as follows: Under standard welding conditions, multiple sets of calibration welding experiments with different combinations of welding current and welding voltage are conducted. Welding current values, welding voltage values, and arc sound signals are simultaneously collected from each set of experiments. Then, the least squares method is used to fit the calibration data, minimizing the prediction error of the theoretical mapping function on the calibration experimental data, thereby determining the parameters of the function. After calibration, in subsequent actual welding processes, the system can substitute the welding current and welding voltage values ​​collected at any time into the theoretical mapping function to quickly calculate the corresponding theoretical sound power value.

[0045] It should be noted that the above theoretical mapping function only needs to be established through a one-time calibration experiment when the welding system is first put into use, or by recalibrating after changing the welding wire grade or shielding gas type. After calibration, the mapping function is fixed in the system as a reference relationship between electrical and acoustic parameters.

[0046] For example, suppose that after the aforementioned calibration experiment, the specific numerical relationship of the theoretical mapping function under the current welding process conditions is determined. In a certain actual welding process, the system collects the effective value of the welding current within the current time window as 220 amperes and the effective value of the welding voltage as 24 volts. Substituting these two values ​​into the pre-calibrated theoretical mapping function, the theoretical acoustic power value corresponding to the current electrical parameters is calculated to be approximately 240 (in relative units). At the same time, the system squares and averages the time-domain amplitude of the acoustic signal collected by the arc acoustic sensor within the current time window, obtaining a measured acoustic power value of approximately 205. Comparing the theoretical acoustic power value with the measured acoustic power value, the deviation between the two is calculated. This deviation reflects the degree of consistency between the arc state characterized by the electrical parameters and the arc state characterized by the acoustic parameters. In this example, the theoretical value is 240, the measured value is 205, and the relative deviation between the two is approximately 15%. This deviation is at a moderate level, and according to the negative correlation mapping relationship where the larger the deviation, the smaller the weight, the corresponding weight adjustment coefficient is approximately 0.72. The weights were applied to both the current and voltage characteristic channels simultaneously. This means that the contributions of both the current and voltage characteristics were moderately reduced to about 72 percent, but not completely suppressed. This indicates that there is a certain degree of difference between the current electrical and acoustic parameters (possibly due to slight fluctuations in the protective gas flow rate or minor changes in the wire feed speed), but it is still within a controllable range. The current and voltage characteristics still retain some reference value.

[0047] After obtaining the first, second, and third physical correlation weights, the physical prior weighted fusion layer normalizes these three weights, unifying their numerical ranges to the same scale. After normalization, the first physical correlation weight is applied to the feature channel corresponding to the arc sound feature, the second physical correlation weight to the feature channel corresponding to the temperature field feature, and the third physical correlation weight to the feature channels corresponding to both the current and voltage features. Each weight is then multiplied element-wise with its corresponding feature channel, achieving physical calibration at the feature level. After this weighted fusion process, the physical prior weighted fusion layer outputs a physical prior calibrated multimodal feature tensor. Each feature channel in this tensor has been rationally adjusted according to physical propagation laws and sensor mounting pose, providing high-quality input guided by physical mechanisms for subsequent feature extraction in deep networks.

[0048] For example, when the welding torch tilt angle is large, causing the arc acoustic sensor to be away from the molten pool, the distance the sound wave travels from the molten pool to the sensor increases and deviates from the sensor's axis. Consequently, the first physical correlation weight decreases, and the contribution of the arc acoustic characteristic channel is moderately suppressed. When the infrared thermal imager is far from the molten pool or the workpiece material has slow thermal diffusion, the second physical correlation weight decreases, and the contribution of the temperature field characteristic channel is moderately suppressed. When welding spatter is large, causing a significant deviation between the measured and theoretical sound power, the third physical correlation weight decreases, and the contributions of the current and voltage characteristic channels are simultaneously suppressed. Through the differentiated processing of the above-mentioned physical prior weighted fusion layer, even if the quality of a certain mode's sensing signal decreases under the current welding condition, its corresponding characteristic channel weight will automatically decrease, thereby avoiding the negative impact of low-quality signals on subsequent evaluations.

[0049] Following the physical prior weighted fusion layer, this online welding quality assessment model employs a multi-scale temporal convolutional fusion layer. Since changes in welding quality are not determined solely by the sensor signal at a single moment, but rather by the signal's evolution over time, and different types of welding defects correspond to different signal anomalies at different time scales, the model needs to simultaneously focus on feature changes at different time scales. For example, transient short-circuit fluctuations in welding current typically occur within millisecond time scales, representing local detail features; while the gradual increase in temperature in the heat-affected zone or the progressive change in weld penetration occurs within several seconds time scales, representing global trend features. To simultaneously capture these feature changes at different time scales, the multi-scale temporal convolutional fusion layer contains multiple temporal convolutional branches with different expansion rates. Each branch performs temporal convolution processing on the aforementioned physically prior-calibrated multimodal feature tensors in parallel. Among them, convolutional branches with smaller dilation rates have smaller receptive fields, mainly capturing local abrupt changes between adjacent time steps, and are more sensitive to short-term events such as short-circuit transition anomalies and spatter impacts. Convolutional branches with larger dilation rates have larger receptive fields, covering a longer historical time range, mainly capturing slow drifts and trend changes in the signal, and are more sensitive to long-term evolution processes such as thermal accumulation and gradual degradation of the protective atmosphere. After each branch completes temporal convolution processing, it outputs the temporal feature representation under its corresponding receptive field. The multi-scale temporal convolutional fusion layer then adaptively fuses the temporal feature representations of these different branches, enabling the model to comprehensively utilize feature information at different time scales to obtain multi-scale temporal features. This adaptive fusion preferably employs an attention mechanism, allowing the model to dynamically adjust the degree of attention to features at different time scales according to the specific state of the current welding condition.

[0050] Following the multi-scale temporal convolutional fusion layer, this online welding quality evaluation model incorporates a temporal causal self-attention layer. The welding process exhibits a strict causal order in time; that is, the welding quality at the current moment depends only on the process state at the current moment and its preceding moments, and is unrelated to the state at future moments. Therefore, when performing temporal modeling, it is crucial to avoid interference from future information with current predictions; otherwise, the model's performance will drastically degrade during online inference due to the inability to access future data, resulting in information leakage. This temporal causal self-attention layer first encodes the aforementioned multi-scale temporal features along the time dimension, explicitly embedding the temporal order information of each feature point into the feature representation, enabling the model to distinguish the sequential relationship between different moments. Then, this layer employs a multi-head self-attention mechanism with causal masking to process the positionally encoded temporal features. The causal masking means that during the self-attention calculation process, all positions after the current moment are forcibly masked, ensuring that the output at each moment can only focus on itself and its preceding historical moments, preventing access to future moment information. Through the aforementioned causal masking mechanism, the temporal causal self-attention layer ensures the causality of the model in the time dimension and the real-time performance of online inference, effectively preventing illegal interference from future data to the current prediction. After processing by this layer, the model outputs context-aware features, which not only contain the multi-scale information extracted by the aforementioned multi-scale temporal convolutional fusion layer, but also establish the dependencies between different time points in the global time dimension.

[0051] Following the temporal causal self-attention layer, this online welding quality assessment model employs a cross-modal cross-attention gating fusion layer. Because the signals from different sensing modalities are subject to varying degrees of external interference during welding, and these interferences are dynamic—for example, when welding spatter is large, the arc acoustic sensor may be affected by noise from spatter particles; when the arc light is intense, the imaging quality of the infrared thermal imager may decrease; and when the power grid fluctuates or the wire feed speed is unstable, current and voltage values ​​may become the most reliable signal sources reflecting the welding status—the reliability of each modal feature changes in real time, and fixed fusion weights cannot adapt to this dynamic change. This cross-modal cross-attention gating fusion layer first performs multimodal cross-attention interaction on the context-aware features output from the aforementioned temporal causal self-attention layer. This involves pairwise interaction calculations between features of different modalities, allowing each modal feature to fully capture the intermodal correlation information through mutual reference. Examples include the temporal correspondence between current changes and arc acoustic spectrum changes, and the causal relationship between temperature field distribution changes and voltage fluctuations, thus obtaining the interacted multimodal features. Building upon this, this layer further dynamically calculates the confidence weights of each modality based on the current welding conditions. These confidence weights reflect the signal-to-noise ratio and reliability of each modal signal under the current welding state. The calculation is based on, but is not limited to, the energy fluctuation amplitude of each modal signal at the current moment, its deviation from historical statistical values, and its consistency with other modal signals. Subsequently, the cross-modal attention-gated fusion layer adaptively filters and fuses the interacting multimodal features based on the dynamically calculated confidence weights. Modal features with high confidence dominate the fusion process, while those with low confidence are suppressed, resulting in the final fused features. Through this dynamic gating mechanism, when a modal signal is subjected to strong interference, its confidence weight automatically decreases, and the model shifts its decision focus to other modalities with higher confidence, thus ensuring the robustness of the evaluation results.

[0052] Following the cross-modal attention-gated fusion layer, the online welding quality assessment model includes a prediction output layer. This layer maps the aforementioned fused features to a welding quality score corresponding to the current time window via a fully connected network. This score is output numerically and serves as the quantitative basis for subsequent adaptive compensation decisions. Thus, the online welding quality assessment model completes the entire forward computation process from multimodal feature sequence input to welding quality score output.

[0053] When constructing an online welding quality assessment model, it is necessary to utilize multi-source feature sequence samples collected during historical welding processes and corresponding welding quality scoring labels to build a training sample set. Each training sample set contains multimodal feature sequences across multiple consecutive time windows during a welding process, along with a corresponding welding quality scoring label. This welding quality scoring label can be obtained by performing quality inspection and scoring on the weld seam after welding is completed, and aligned with the corresponding multimodal feature sequences along the time axis. By constructing a training sample set covering different welding conditions and quality levels, a sufficient data foundation can be provided for subsequent model training.

[0054] To ensure the accuracy of online welding quality assessment while meeting the real-time requirements of actual welding sites, a training method combining teacher and student networks can be adopted. The teacher network is a full-parameter version of the online welding quality assessment model, containing a physical prior weighted fusion layer, a multi-scale temporal convolutional fusion layer, a temporal causal self-attention layer, a cross-modal cross-attention gating fusion layer, and a prediction output layer, possessing strong feature representation capabilities and high assessment accuracy. The student network is a lightweight version obtained by channel pruning and quantization compression of the teacher network. Channel pruning refers to reducing the number of channels in some convolutional layers to reduce computational cost; quantization compression refers to converting floating-point parameters in the network into low-bit-width representations to reduce storage footprint and accelerate inference speed. The student network retains the same layer structure as the teacher network, and the output dimensions of the distilled intermediate layers are aligned with those of the teacher network to facilitate the distillation learning of intermediate layer features during training.

[0055] After constructing the teacher and student networks, a progressive distillation training strategy was adopted to train the student network. This allowed the lightweight student network to inherit as much knowledge as possible from the teacher network, thus maintaining high accuracy in welding quality assessment under real-time inference conditions. The progressive distillation training was implemented in three stages, specifically: In the first stage, all parameters of the teacher network are frozen, and no further parameter updates are performed. Multi-source feature sequence samples from the training sample set are input into the teacher network, which then performs forward propagation. During forward propagation, the intermediate layer outputs of the multi-scale temporal convolutional fusion layer, the intermediate layer outputs of the temporal causal self-attention layer, and the final output of the prediction output layer are extracted from the teacher network. These three outputs are used as the first, second, and third soft labels, respectively. The first and second soft labels contain the multi-scale temporal features and context-aware feature information extracted by the teacher network in the intermediate layers, while the third soft label contains the welding quality score output by the teacher network. By extracting soft labels at different depths, multi-level knowledge supervision can be provided for the intermediate and output layers of the subsequent student network.

[0056] In the second stage, the same training sample set is simultaneously input into both the teacher network and the student network. The student network performs forward propagation on the input multi-source feature sequence samples and outputs the intermediate layer outputs of its multi-scale temporal convolutional fusion layer, the intermediate layer outputs of its temporal causal self-attention layer, and the final output of its prediction output layer. Then, the first distillation loss between the intermediate layer output of the multi-scale temporal convolutional fusion layer and the first soft label is calculated; the second distillation loss between the intermediate layer output of the temporal causal self-attention layer and the second soft label is calculated; and the third distillation loss between the final output of the prediction output layer and the third soft label is calculated. These three distillation losses are used to constrain the student network's output at the corresponding layer to maintain consistency with the teacher network. Simultaneously, the hard label loss between the final output of the prediction output layer and the welding quality score labels in the training sample set is calculated. This hard label loss is used to constrain the student network's final output to be consistent with the true quality score. The first distillation loss, second distillation loss, third distillation loss, and hard label loss are weighted and summed according to preset weights to obtain the total loss. Using the total loss as the objective, the parameters of the student network are updated through backpropagation. By simultaneously optimizing the intermediate layer distillation loss, the output layer distillation loss, and the hard label loss, the student network can approximate the evaluation performance of the teacher network as closely as possible while maintaining its lightweight nature.

[0057] In the third stage, the student network trained in the second stage is deployed to an embedded inference platform for pre-deployment inference speed testing. The embedded inference platform can be an industrial control computer, edge computing device, or embedded AI inference module used in actual welding operations. If the tested inference speed is lower than a preset real-time threshold, meaning it cannot complete online inference for welding quality scoring within the set time window, an adaptive layer fusion operation is performed on the student network. Adaptive layer fusion refers to merging the parameters of adjacent convolutional layers and batch normalization layers in the student network that meet the fusion conditions, fusing their parameters into the parameters of a single convolutional layer. This reduces the computational layers and memory accesses during inference, improving inference speed. After layer fusion, the fused student network is fine-tuned to recover any accuracy loss that might have been caused by parameter merging. After fusion and fine-tuning, the inference speed is tested again until it meets the preset real-time threshold. The resulting student network is the online welding quality evaluation model for deployment.

[0058] Through the aforementioned progressive distillation training process using a teacher-student network, a lightweight model that meets real-time requirements and possesses high evaluation accuracy can be obtained without altering the overall structure of the online welding quality assessment model. Once trained, the online welding quality assessment model is deployed on an industrial control computer, and during actual welding operations, it can output the welding quality score for the current time window based on real-time acquired multimodal feature sequences.

[0059] Furthermore, during the actual welding process, the industrial control computer inputs the real-time extracted multimodal feature sequences into a pre-trained online welding quality assessment model according to a preset time window. This online welding quality assessment model sequentially performs physical prior weighted fusion, multi-scale temporal convolutional fusion, temporal causal self-attention processing, cross-modal cross-attention gating fusion, and prediction output mapping on the input multimodal feature sequences, ultimately outputting a welding quality score for the current time window. This welding quality score reflects the current welding quality level in real time, providing a basis for subsequent quality status judgment and adaptive compensation of process parameters.

[0060] Step S140: Calculate the average welding quality score of multiple consecutive time windows before the current time window, and compare the welding quality score of the current time window with the average value. When the welding quality score of the current time window is lower than the average value and the difference exceeds a preset range, calculate the score deviation and score deviation change rate between the welding quality score of the current time window and the preset target score. Using the score deviation and score deviation change rate as input, obtain the welding process parameter correction amount through fuzzy inference, and correct the current welding process parameters according to the welding process parameter correction amount to obtain the corrected welding process parameters. Continue welding with the corrected welding process parameters to achieve adaptive compensation.

[0061] Specifically, after the online welding quality assessment model outputs the welding quality score for the current time window, the industrial control computer stores this score in its internal storage for subsequent quality trend analysis and adaptive compensation. As the welding process continues, the industrial control computer continuously acquires new welding quality scores, forming a welding quality score sequence in chronological order. For the current time window, the industrial control computer obtains the welding quality scores from several consecutive time windows prior to it, calculates the average of these historical scores, and uses this average as a reference benchmark for the current welding quality. For example, the welding quality scores from five consecutive time windows prior to the current time window can be taken, summed, and divided by 5 to obtain the historical average score. Using the average of multiple consecutive time windows as a comparison benchmark avoids misjudgments caused by occasional fluctuations in a single time window, making the judgment of quality decline more robust.

[0062] After obtaining the historical average score, the industrial control computer compares the welding quality score of the current time window with this average. When the welding quality score of the current time window is lower than this average, and the difference between the two exceeds a preset range, it indicates a significant downward trend in welding quality, triggering the adaptive compensation process. The preset range can be set according to the actual welding process requirements, for example, 5 points. For instance, if the average welding quality score of the previous 5 consecutive time windows is 82 points, and the welding quality score of the current time window is 75 points, the difference is 7 points, exceeding the preset range of 5 points, then it is determined that the welding quality has significantly declined, and the adaptive compensation phase begins. Conversely, if the difference does not exceed the preset range, the current welding process parameters are maintained, and the score changes in subsequent time windows are continuously monitored.

[0063] When adaptive compensation is triggered, the industrial control computer first calculates the score deviation between the welding quality score of the current time window and the preset target score. The preset target score is the expected quality score of the welding process, for example, it can be set to 80 points. The score deviation is the welding quality score of the current time window minus the preset target score, and this deviation can be positive, negative, or zero. For example, if the welding quality score of the current time window is 72 points and the preset target score is 80 points, the score deviation is -8 points, indicating that the current quality is below the target level. Simultaneously, the industrial control computer also needs to calculate the rate of change of the score deviation to reflect the dynamic trend of the quality score. The rate of change of the score deviation is obtained by linearly fitting the welding quality scores of several consecutive time windows preceding the current time window, and the slope obtained from the fitting is used as the rate of change of the score deviation. For example, taking the welding quality scores from the five consecutive time windows preceding the current time window, and performing a univariate linear fit with the time window number as the independent variable and the welding quality score as the dependent variable, if the slope of the fitted result is -2.5 points / time window, then the rate of change of the score deviation is -2.5 points / time window, indicating that the welding quality score is deteriorating at a rate of 2.5 points per time window. If the time unit is seconds, and each time window corresponds to 0.5 seconds, then the rate of change of the score deviation can be converted to -5 points / second. By simultaneously considering the score deviation and the rate of change of the score deviation, a more comprehensive assessment of the current degree of deviation and trend in welding quality can be made, providing more accurate input information for subsequent fuzzy inference.

[0064] After obtaining the scoring deviation and the rate of change of the scoring deviation, the industrial control computer first dynamically constructs the deviation membership function and the rate of change of the deviation membership function for the current control cycle to adapt to the differences in the fluctuation characteristics of welding quality scoring under different combinations of welding process parameters. The deviation membership function maps the scoring deviation to a preset fuzzy universe and outputs the corresponding membership value; its shape determines the degree of membership of the scoring deviation in different fuzzy sets. The rate of change of the deviation membership function maps the rate of change of the scoring deviation to a preset fuzzy universe and outputs the corresponding membership value; its shape determines the degree of membership of the rate of change of the scoring deviation in different fuzzy sets. The shapes of both functions are determined by the membership function shape parameters, such as the base width, vertex position, and side slope of a triangular membership function, or the starting and ending points of a trapezoidal membership function. By adjusting the membership function shape parameters, the membership distribution of the scoring deviation and the rate of change of the scoring deviation in different fuzzy sets can be changed, thereby affecting the fuzzy inference results.

[0065] When constructing the deviation membership function and the deviation change rate membership function, the industrial control computer obtains the current welding process parameter combination, which includes the welding current setting, welding voltage setting, welding speed setting, and oscillation amplitude setting. Then, based on this current welding process parameter combination, the corresponding membership function shape parameter is looked up in the preset process parameter-membership parameter mapping table. The preset process parameter-membership parameter mapping table is obtained in advance through experimental calibration; it establishes the correspondence between different welding process parameter combinations and the membership function shape parameters. For example, for a set of process parameters with a welding current setting of 230A, a welding voltage setting of 26V, a welding speed setting of 40cm / min, and an oscillation amplitude setting of 4mm, the corresponding deviation membership function can be found in the mapping table with a vertex position of -5 minutes and a base half-width of 10 minutes, and a deviation change rate membership function with a vertex position of -2 minutes / time window and a base half-width of 5 minutes / time window. Based on the queried membership function shape parameters, the industrial control computer dynamically constructs the deviation membership function and the deviation change rate membership function within the current control cycle, so that the shape of the membership function is adaptively adjusted according to the change of the current welding process parameter combination, thereby improving the pertinence and accuracy of fuzzification processing.

[0066] After constructing the membership functions, the industrial control computer fuzzifies the scoring deviation input into the deviation membership function to obtain the fuzzy value of the deviation; it also fuzzifies the scoring deviation change rate input into the deviation change rate membership function to obtain the fuzzy value of the deviation change rate. The fuzzy universe can be preset according to actual needs. For example, the fuzzy universe of the scoring deviation is [-20, 40] points, and the fuzzy universe of the scoring deviation change rate is [-30, 30] points / time window. During the fuzzification process, the scoring deviation and the scoring deviation change rate are mapped to the corresponding fuzzy sets. The fuzzy sets can include seven levels: negative large, negative medium, negative small, zero, positive small, positive medium, and positive large. For example, if the scoring deviation is -8 points, after mapping by the deviation membership function, it may simultaneously belong to the "negative medium" and "negative small" fuzzy sets, and obtain the corresponding membership values; if the scoring deviation change rate is -5 points / time window, it may belong to the "negative large" and "negative medium" fuzzy sets. The output of fuzzification processing is the deviation fuzzy quantity and the deviation change rate fuzzy quantity, which represent the fuzzy state of the input quantity in the form of fuzzy sets and membership degrees.

[0067] Subsequently, the industrial control computer inputs the deviation fuzzy quantity and the deviation change rate fuzzy quantity into a preset fuzzy inference rule base. This preset fuzzy inference rule base contains multiple rules expressed in the form of "if the deviation fuzzy quantity is a certain fuzzy set and the deviation change rate fuzzy quantity is a certain fuzzy set, then the output fuzzy quantity is a certain fuzzy set." For example, one rule could be: if both the deviation fuzzy quantity and the deviation change rate fuzzy quantity are negatively large, then the output fuzzy quantity is positively large, indicating that the welding quality is severely below the target and is still deteriorating, requiring a significant increase in heat input for compensation; another rule could be: if both the deviation fuzzy quantity and the deviation change rate fuzzy quantity are negatively small, then the output fuzzy quantity is zero, indicating that although the welding quality is slightly below the target, it has shown an upward trend, and no adjustment is needed for now. The preset fuzzy inference rule base can cover all possible combinations of deviation fuzzy quantity and deviation change rate fuzzy quantity, for example, 7×7 rules totaling 49 rules. Through fuzzy inference, the rule base outputs one or more fuzzy output quantities, which characterize the fuzziness level of the welding process parameter correction.

[0068] After obtaining the fuzzy output, the industrial control computer defuzzifies it, converting it into precise welding process parameter corrections. Defuzzification can be performed using the centroid method, which calculates the centroid of the area enclosed by the membership functions of the fuzzy output and uses the value corresponding to that centroid as the precise output. The welding process parameter corrections obtained after defuzzification include welding current corrections, welding voltage corrections, welding speed corrections, and oscillation amplitude corrections. For example, for the case where the scoring deviation is -8 points and the scoring deviation change rate is -5 points / time window, the defuzzified values ​​might be +15A for welding current, +0.5V for welding voltage, -4cm / min for welding speed, and +1mm for oscillation amplitude. Increasing the welding current and welding voltage corrections improves the welding heat input, decreasing the welding speed correction extends the heating time per unit length of weld, and increasing the oscillation amplitude correction improves weld width and fusion, thus collectively providing compensation. The sign and magnitude of the above corrections can be either positive or negative, depending on the fuzzy inference rules and the defuzzification result.

[0069] After obtaining the welding process parameter correction amount, the industrial control computer corrects the current welding process parameters based on this correction amount. The current welding process parameters refer to the welding current setting, welding voltage setting, welding speed setting, and oscillation amplitude setting values ​​that the welding power supply and robot controller were executing before the correction. During correction, the current welding current setting value is added to the welding current correction amount to obtain the corrected welding current setting value; the current welding voltage setting value is added to the welding voltage correction amount to obtain the corrected welding voltage setting value; the current welding speed setting value is added to the welding speed correction amount to obtain the corrected welding speed setting value; and the current oscillation amplitude setting value is added to the oscillation amplitude correction amount to obtain the corrected oscillation amplitude setting value. For example, if the current welding current setting is 230A, the welding voltage setting is 26V, the welding speed setting is 40cm / min, and the oscillation amplitude setting is 4mm, and the calculated welding current correction is +15A, the welding voltage correction is +0.5V, the welding speed correction is -4cm / min, and the oscillation amplitude correction is +1mm, then the corrected welding current setting is 245A, the welding voltage setting is 26.5V, the welding speed setting is 36cm / min, and the oscillation amplitude setting is 5mm. The industrial control computer sends the corrected welding process parameters to the welding power supply and robot controller in real time via the industrial Ethernet communication protocol, enabling the welding power supply and robot controller to continue welding according to the corrected welding process parameters. At the same time, the industrial control computer ensures that the corrected parameter values ​​do not exceed the preset safety range, for example, the welding current setting does not exceed 180A to 320A, the welding voltage setting does not exceed 20V to 35V, the welding speed setting does not exceed 30cm / min to 60cm / min, and the oscillation amplitude setting does not exceed 0mm to 8mm. If the corrected parameter exceeds the safe range, then limit it to the boundary values ​​of the safe range.

[0070] Through the above process, when the welding quality score drops significantly, the industrial control computer can obtain an appropriate correction amount for the welding process parameters based on the score deviation and the rate of change of the score deviation through fuzzy inference. It then performs online correction of the current welding process parameters and continues welding with the corrected parameters, thereby achieving adaptive compensation for welding quality. In subsequent welding processes, the industrial control computer continues to acquire the welding quality score according to a preset time window and repeats the above steps of quality comparison, score deviation calculation, fuzzy inference, and parameter correction until the welding quality score recovers to a reasonable level or the welding task ends.

[0071] After continuing welding with the corrected welding process parameters, the industrial control computer (ICC) continues to acquire welding quality scores according to preset time windows to evaluate the actual effect of adaptive compensation. The ICC acquires welding quality scores for multiple consecutive time windows after correction, such as scores for five consecutive time windows. If the welding quality scores for these multiple consecutive time windows are not lower than the preset recovery threshold, it indicates that the compensation has restored the welding quality to an acceptable level. At this point, the ICC enters fine-tuning mode, gradually reducing the welding process parameter correction amount to avoid introducing new quality fluctuations due to excessive correction. Specifically, the ICC multiplies the welding process parameter correction amount by a preset attenuation coefficient to obtain a reduced welding process parameter correction amount, and then corrects the current welding process parameters based on this reduced correction amount. The preset attenuation coefficient can be set according to the welding process's requirements for adjustment stability, for example, set to 0.5 or 0.7. For example, if the welding current correction in a compensation is +20A, and the welding quality scores for five consecutive time windows after the correction are 78, 79, 81, 82, and 83 respectively, all above the preset recovery threshold of 75, then the welding quality has significantly improved. In this case, if compensation is triggered again later, the industrial control computer can multiply the welding current correction by a preset attenuation coefficient of 0.5, using +10A as the new correction, and continue to adjust the welding current accordingly. By gradually reducing the correction amount, the risk of over-adjustment can be reduced while maintaining stable welding quality.

[0072] If the welding quality scores for multiple consecutive time windows after correction remain below the preset recovery threshold, it indicates that the compensation has failed to effectively improve the welding quality, and the current welding conditions may be abnormal. In this case, the industrial control computer triggers a manual intervention alarm to prompt the operator to intervene and check. The manual intervention alarm can be triggered by sound and light, and the current welding quality score, the amount of welding process parameter correction already performed, and suggested inspection items will be displayed on the human-machine interface. Upon receiving the alarm, the operator can pause welding and check the shielding gas flow rate, wire feeding status, bevel assembly, etc., to eliminate abnormal factors. This avoids batch defects caused by continuing welding under abnormal conditions and provides the operator with a clear opportunity and basis for handling the situation.

[0073] Furthermore, during and after the adaptive compensation process, the industrial control computer can construct a welding quality time-series traceability map for visual traceability and analysis of the welding process. This welding quality time-series traceability map uses the welding time window sequence as the horizontal axis, synchronously displaying the welding quality score, welding process parameter correction, and feature values ​​in the multimodal feature sequence within the same visualization interface. Since the welding quality score, welding process parameter correction, and feature values ​​in the multimodal feature sequence have different dimensions and numerical ranges, a multi-vertical axis partitioning display method can be used within the same visualization interface. This means setting multiple vertical regions under the same horizontal axis, each corresponding to a different feature value in the welding quality score, welding process parameter correction, and multimodal feature sequence. For example, the upper region of the visualization interface displays the welding quality score as a function of the time window using a line graph; the middle region displays the welding process parameter correction corresponding to each time window using a bar chart; and the lower region displays selected feature values ​​from the multimodal feature sequence using multiple lines, such as short-time energy in the arc sound characteristics, the mean value in the current characteristics, and the highest temperature of the molten pool in the temperature field characteristics. Through synchronous display, operators can intuitively observe the changes in modal characteristic values ​​before and after the decrease in welding quality score, as well as the effect of parameter correction on improving welding quality.

[0074] Operators can select any time window in the welding quality time-series traceability map. In response to this selection, the industrial control computer retrieves and replays synchronously recorded welding process data within a preset time period before and after that window. The preset time period can be set according to traceability needs, such as 2 seconds or 5 seconds before and after. The synchronously recorded welding process data includes audio clips of the arc sound signal corresponding to that time window, waveform clips of the welding current and welding voltage values, and a sequence of thermal images of the weld area temperature field. For example, when the operator clicks on a time window in the map where a score begins to decline, the industrial control computer simultaneously plays the arc sound signal audio within 5 seconds before and after that time window, displays the change curves of the welding current and welding voltage values ​​in waveform form during that period, and simultaneously replays the thermal images of the weld area temperature field in image sequence form during that period. Through the playback of multi-source synchronously recorded data, operators can trace the welding quality decline process from multiple perspectives—acoustic, electrical, and thermal—providing a basis for analyzing the causes of quality anomalies.

[0075] Operators can also mark the intervention start time window and intervention end time window in the welding quality time sequence traceability map. The intervention start time window and intervention end time window represent the start and end points of the time interval that the operator believes is related to the quality anomaly event, respectively. Based on the intervention start time window and intervention end time window marked by the operator, the industrial control computer extracts the multimodal feature sequence, welding quality score sequence, and welding process parameter correction sequence within the time interval as typical quality event samples. The typical quality event sample records the complete sensing characteristics before and after a certain type of quality anomaly occurs, the process of quality score change, and the process parameter correction measures taken. The industrial control computer stores the extracted typical quality event samples in the welding quality event knowledge base. This welding quality event knowledge base is used for subsequent adjustment of the rule confidence of the fuzzy inference rule base under similar working conditions and incremental training of the online welding quality assessment model. For example, when a multimodal feature pattern similar to a typical quality event sample appears during subsequent welding processes, the industrial control computer can increase or decrease the confidence of the corresponding fuzzy inference rule based on the correction effect of the sample, allowing the fuzzy inference rule base to be continuously optimized over long-term operation. Simultaneously, this sample can also be used as new training data for incremental training of the online welding quality assessment model, improving the model's adaptability to complex working conditions. Through the construction of the aforementioned welding quality time-series traceability map and welding quality event knowledge base, this application, based on achieving adaptive compensation for welding quality, further provides a mechanism for visual traceability of the welding process and accumulation of quality event knowledge, supporting the continuous improvement of welding quality control levels.

[0076] In summary, by employing the above technical solution, this application has the following advantages: First, this application synchronously acquires arc acoustic signals, welding current values, welding voltage values, and weld area temperature field signals during the welding process through a multi-source sensor signal acquisition system integrated at the end of the welding robot. These signals are preprocessed to extract arc acoustic features, current features, voltage features, and temperature field features, forming a multi-modal feature sequence. This sequence can comprehensively reflect the welding state from multiple physical dimensions, including acoustic, electrical, and thermal aspects, thereby improving the richness and reliability of the information used for welding quality assessment.

[0077] Second, this application obtains the welding quality score for the current time window by inputting multimodal feature sequences into a pre-trained online welding quality evaluation model, thereby realizing real-time online evaluation of welding quality during the welding process, allowing operators to understand the current welding quality level without waiting for post-weld inspection.

[0078] Third, this application calculates the average welding quality score of multiple consecutive time windows before the current time window and compares the welding quality score of the current time window with the average value. When the welding quality score of the current time window is lower than the average value and the difference exceeds a preset range, the subsequent correction process is triggered. This can promptly capture obvious downward trends in welding quality and provide a basis for online compensation.

[0079] Fourth, this application calculates the score deviation and score deviation change rate between the welding quality score and the preset target score in the current time window when the welding quality score drops significantly, and obtains the welding process parameter correction amount through fuzzy reasoning by inputting the score deviation and score deviation change rate. It can automatically determine the appropriate correction amount according to the actual deviation and change trend of the welding quality.

[0080] Fifth, this application achieves online adaptive compensation of welding process parameters by correcting the current welding process parameters according to the correction amount, obtaining the corrected welding process parameters, and continuing to perform welding with the corrected welding process parameters. This allows for dynamic adjustment of the welding process without machine shutdown or manual intervention, thereby effectively suppressing welding quality fluctuations, improving the consistency and reliability of welding quality, and reducing the risk of batch defects caused by welding quality deviating from expectations.

[0081] It should be understood that the above-described adaptive compensation method for welding quality is merely exemplary, and those skilled in the art can make various modifications based on the above content. Such modified solutions also fall within the protection scope of this application.

[0082] This application also provides an industrial control computer for executing the above-described adaptive compensation method for welding quality.

[0083] Those skilled in the art will understand that embodiments of the present invention can be provided as methods, systems, or computer program products. Therefore, the present invention can take the form of a completely hardware embodiment, a completely software embodiment, or an embodiment combining software and hardware aspects. Furthermore, the present invention can take the form of a computer program product embodied on one or more computer-usable storage media (including, but not limited to, disk storage, CD-ROM, optical storage, etc.) containing computer-usable program code.

[0084] This invention is described with reference to flowchart illustrations and / or block diagrams of methods, apparatus (systems), and computer program products according to embodiments of the invention. It will be understood that each block of the flowchart illustrations and / or block diagrams, as well as combinations of blocks in the flowchart illustrations and / or block diagrams, can be implemented by computer program instructions.

[0085] It should be noted that any reference numerals placed between parentheses in the claims should not be construed as limiting the claims. The word "comprising" does not exclude the presence of components or steps not listed in the claims. The word "a" or "an" preceding a component does not exclude the presence of a plurality of such components. The invention can be implemented by means of hardware comprising several different components and by means of a suitably programmed computer. In claims that enumerate several means, several of these means may be embodied by the same hardware. The use of the terms first, second, third, etc., is merely for convenience of expression and does not indicate any order. These terms can be understood as part of the component names.

[0086] Furthermore, it should be noted that in the description of this specification, the terms "one embodiment," "some embodiments," "embodiment," "example," "specific example," or "some examples," etc., refer to specific features, structures, materials, or characteristics described in connection with that embodiment or example, which are included in at least one embodiment or example of the present invention. In this specification, the illustrative expressions of the above terms do not necessarily refer to the same embodiment or example. Moreover, the specific features, structures, materials, or characteristics described may be combined in any suitable manner in one or more embodiments or examples. Furthermore, without contradiction, those skilled in the art can combine and integrate the different embodiments or examples described in this specification, as well as the features of different embodiments or examples.

[0087] Although preferred embodiments of the invention have been described, those skilled in the art, upon learning the basic inventive concept, can make other changes and modifications to these embodiments. Therefore, the claims should be interpreted to include both the preferred embodiments and all changes and modifications falling within the scope of the invention.

[0088] Obviously, those skilled in the art can make various modifications and variations to this invention without departing from its spirit and scope. Therefore, if these modifications and variations fall within the scope of the claims of this invention and their equivalents, then this invention should also include these modifications and variations.

Claims

1. A method of adaptive compensation of welding quality, characterized in that, include: The multi-source sensor signal acquisition system integrated at the end of the welding robot synchronously acquires arc sound signals, welding current values, welding voltage values, and temperature field signals in the weld area during the welding process. The multi-source sensor signal acquisition system includes an arc sound sensor for acquiring the arc sound signals during the welding process, a current acquisition module for acquiring the welding current values, a voltage acquisition module for acquiring the welding voltage values, and an infrared thermal imager for acquiring the temperature field signals in the weld area. The collected arc sound signal, welding current value, welding voltage value and temperature field signal of the weld area are preprocessed to extract arc sound features, current features, voltage features and temperature field features respectively to form a multimodal feature sequence; The multimodal feature sequence is input into a pre-trained online welding quality evaluation model to obtain the welding quality score for the current time window; The average welding quality score of multiple consecutive time windows preceding the current time window is calculated, and the welding quality score of the current time window is compared with the average value. When the welding quality score of the current time window is lower than the average value and the difference exceeds a preset range, the scoring deviation and the rate of change of the scoring deviation between the welding quality score of the current time window and the preset target score are calculated. Using the scoring deviation and the rate of change of the scoring deviation as input, the welding process parameter correction amount is obtained through fuzzy inference. The current welding process parameters are then corrected according to the welding process parameter correction amount to obtain the corrected welding process parameters. Welding is then continued to be performed with the corrected welding process parameters to achieve adaptive compensation.

2. The method of claim 1, wherein, The online welding quality assessment model includes: The physical prior weighted fusion layer is used to determine the physical correlation weights between each modal feature based on the physical generation mechanism of the arc sound feature, the current feature, the voltage feature and the temperature field feature and the corresponding sensor installation pose, and to perform weighted fusion of the arc sound feature, current feature, voltage feature and temperature field feature in the multimodal feature sequence based on the physical correlation weights to obtain a multimodal feature tensor calibrated by physical prior. The multi-scale temporal convolutional fusion layer contains multiple temporal convolutional branches with different dilation rates. Each of the temporal convolutional branches performs temporal convolution processing on the multimodal feature tensor in parallel to obtain the temporal feature representation corresponding to each of the temporal convolutional branches. The temporal feature representations corresponding to each of the temporal convolutional branches are then adaptively fused to obtain multi-scale temporal features. The temporal causal self-attention layer is used to encode the multi-scale temporal features along the time dimension to obtain the position-encoded temporal features, and then process the position-encoded temporal features using a multi-head self-attention mechanism with causal masking to obtain context-aware features. A cross-modal attention gating fusion layer is used to perform multimodal cross-attention interaction on the context-aware features to obtain the multimodal features after interaction, and to dynamically calculate the confidence weight of each modality according to the current welding condition, and to adaptively filter and fuse the multimodal features after interaction based on the confidence weight to obtain fused features. A prediction output layer is used to map the fused features to a welding quality score for the current time window.

3. The method of claim 2, wherein, The physical prior weighted fusion layer determines the physical association weights between each modality feature, specifically including: Obtain the spatial pose parameters of the welding torch relative to the workpiece under the current welding condition; wherein, the spatial pose parameters include the welding torch tilt angle, the welding torch pointing angle, and the distance between the welding torch tip and the workpiece surface; Based on the installation posture of the arc acoustic sensor and the spatial posture parameters of the welding torch relative to the workpiece, the propagation path distance of the arc acoustic signal radiated from the molten pool to the arc acoustic sensor and the angle between the propagation direction and the sensor axis are calculated. Based on the propagation path distance and the angle, and according to the law that the sound pressure decreases inversely with distance, the sound wave attenuation compensation coefficient is calculated and used as the first physical correlation weight of the arc acoustic feature. Based on the installation orientation of the infrared thermal imager and the spatial orientation parameters of the welding torch relative to the workpiece, the heat conduction path length and heat diffusion time constant of the temperature field in the weld area from the center of the molten pool to the boundary of the heat-affected zone are calculated. Based on the heat conduction path length and the heat diffusion time constant, and according to the relationship between the temperature field response lag and the square root of the heat conduction depth, the temperature field lag compensation coefficient is calculated and used as the second physical correlation weight of the temperature field characteristics. Based on the physical coupling relationship between the welding current, the welding voltage, and the arc acoustic power, a theoretical mapping function is established with the effective values ​​of the welding current and the welding voltage as inputs and the arc acoustic power as the output. The deviation between the theoretical acoustic power and the measured acoustic power corresponding to the actual collected welding current and welding voltage is calculated, and a weight adjustment coefficient is calculated according to the deviation and a preset negative correlation mapping function. This weight is then used as the third physical correlation weight of the current-voltage characteristic. The larger the deviation, the smaller the weight adjustment coefficient. After normalizing the first physical association weight, the second physical association weight, and the third physical association weight, the first physical association weight is applied to the arc sound feature, the second physical association weight is applied to the temperature field feature, and the third physical association weight is applied to both the current feature and the voltage feature to obtain a physical prior-calibrated multimodal feature tensor.

4. The method of claim 1, wherein, The preprocessing of the collected arc acoustic signal, welding current value, welding voltage value, and weld area temperature field signal to extract arc acoustic features, current features, voltage features, and temperature field features includes: The arc sound signal is sequentially subjected to bandpass filtering and frame windowing to extract the arc sound features, including Mel frequency cepstral coefficients, short-time energy, and spectral centroid. The welding current value and the welding voltage value are respectively processed by median filtering and moving average filtering to extract the current feature and the voltage feature, including mean, standard deviation and peak value. Spatial filtering is performed on the temperature field signal of the weld area to extract the temperature field features, including the highest temperature of the molten pool, the average temperature of the heat-affected zone, and the temperature gradient.

5. The method according to claim 1, characterized in that, The step of obtaining welding process parameter corrections through fuzzy inference, using the scoring deviation and the rate of change of the scoring deviation as inputs, includes: The scoring deviation and the rate of change of the scoring deviation are respectively mapped to a preset fuzzy universe of discourse to obtain the fuzzy quantity of the deviation and the fuzzy quantity of the rate of change of the deviation. The deviation fuzzy quantity and the deviation change rate fuzzy quantity are input into a preset fuzzy inference rule base to obtain the fuzzy output quantity; The fuzzy output is defuzzified to obtain the welding process parameter correction.

6. The method of claim 5, wherein, Before inputting the deviation fuzzy amount and the deviation change rate fuzzy amount into the preset fuzzy inference rule base, the method further includes: Obtain the current welding process parameter combination; the current welding process parameter combination includes welding current setting value, welding voltage setting value, welding speed setting value, and oscillation amplitude setting value; Based on the current welding process parameter combination, the corresponding membership function shape parameter is queried in the preset process parameter-membership parameter mapping table; Based on the queried membership function shape parameters, dynamically construct the deviation membership function and deviation change rate membership function within the current control cycle; Furthermore, the step of mapping the scoring deviation and the rate of change of the scoring deviation to a preset fuzzy universe of discourse to obtain the deviation fuzzy quantity and the deviation change rate fuzzy quantity specifically includes: The scoring deviation is input into the deviation membership function for fuzzification to obtain the deviation fuzziness quantity; The rate of change of the scoring deviation is input into the membership function of the rate of change of deviation for fuzzification to obtain the fuzzy quantity of the rate of change of deviation.

7. The method of claim 1, wherein, After continuing welding with the modified welding process parameters, the process further includes: Obtain the corrected welding quality scores for multiple consecutive time windows; When the welding quality scores of multiple consecutive time windows after correction are not lower than the preset recovery threshold, the welding process parameter correction amount is multiplied by the preset attenuation coefficient to obtain the reduced welding process parameter correction amount, and the current welding process parameters are corrected according to the reduced welding process parameter correction amount. If the welding quality score for multiple consecutive time windows after correction is still lower than the preset recovery threshold, a manual intervention alarm is triggered.

8. The method of claim 1, wherein, The calculation of the scoring deviation and the rate of change of the scoring deviation between the welding quality score of the current time window and the preset target score includes: The difference between the welding quality score of the current time window and the preset target score is calculated to obtain the score deviation; The welding quality scores of multiple consecutive time windows preceding the current time window are obtained, and a linear fit is performed on the welding quality scores of multiple consecutive time windows preceding the current time window. The slope obtained from the fit is used as the rate of change of the score deviation.

9. The method of claim 1, wherein, After continuing welding with the modified welding process parameters, the process further includes: Construct a welding quality time-series traceability map; wherein, the welding quality time-series traceability map uses the welding time window sequence as the horizontal axis, and synchronously displays the welding quality score, welding process parameter correction amount, and feature values ​​in the multimodal feature sequence in the same visualization interface; In response to the operator's selection of any time window in the welding quality time sequence traceability map, the synchronous recording data of the welding process within a preset time period before and after the time window is retrieved and played back; wherein, the synchronous recording data of the welding process includes the arc sound signal audio segment, the waveform segments of the welding current value and the welding voltage value corresponding to the time window, and the thermal image sequence of the temperature field of the weld area. Based on the intervention start time window and intervention end time window marked by the operator in the welding quality time-series traceability map, the multimodal feature sequence, welding quality score sequence and welding process parameter correction sequence within the corresponding time interval are extracted as typical quality event samples, and the typical quality event samples are stored in the welding quality event knowledge base for subsequent adjustment of the rule confidence of the fuzzy inference rule base under similar working conditions and incremental training of the online welding quality evaluation model.

10. The method of claim 1, wherein, The welding process parameter corrections include welding current corrections, welding voltage corrections, welding speed corrections, and oscillation amplitude corrections.