A real-time monitoring and parameter adaptive control method for laser welding penetration

Abstract

The application discloses a kind of real-time monitoring and parameter self-adaptive control method of laser welding penetration, and the multi-dimensional fusion feature vector is extracted by the multimodal sensing data of welding process, the multi-dimensional fusion feature vector is input into the penetration soft measurement model formed by the weighted fusion of mechanism analysis submodel and data-driven submodel, to generate real-time penetration estimate value;According to the deviation of real-time penetration estimate value and target penetration value, generate feedback control correction, and according to the heat accumulation transfer model, the heat accumulation state of each position of welding track is predicted, to generate feedforward compensation correction;Feedback control correction and feedforward compensation correction are superimposed, to generate laser power set value and welding speed set value.The application realizes the real-time accurate control of penetration, effectively suppresses the penetration fluctuation caused by heat accumulation, and improves welding quality consistency.

Description

Technical Field

[0001] This invention relates to the field of intelligent control technology for welding processes, specifically to a method for real-time monitoring and adaptive parameter control of laser welding penetration depth. Background Technology

[0002] Laser welding is widely used in precision manufacturing due to its advantages such as high energy density, small heat-affected zone, and ease of automation. However, when welding high thermal conductivity materials such as pure copper, the rapid heat transfer to the base material causes drastic dynamic changes in the molten pool temperature field, making the weld penetration extremely sensitive to process parameters and thermal boundary conditions. Existing weld penetration control methods mainly rely on offline sampling after welding or indirect inference from a single sensor signal. The former cannot provide real-time feedback adjustment of the welding process, while the latter is prone to strong interference from the welding process, making it difficult to guarantee inference accuracy. Furthermore, existing control strategies generally do not fully consider the differences in material thermal properties and the heat accumulation effect along the welding trajectory, resulting in insufficient weld penetration control accuracy during continuous welding or under fluctuating operating conditions. Summary of the Invention

[0003] In view of the above problems, the present invention provides a real-time monitoring and parameter adaptive control method for laser welding penetration depth. By integrating multi-source sensor information and combining heat accumulation prediction and feedback adjustment, the real-time and precise control of penetration depth can be achieved.

[0004] To achieve the above objectives, this application provides a method for real-time monitoring and adaptive parameter control of laser welding penetration depth, comprising: Acquire multimodal sensing data of the welding process, including plasma light intensity data, infrared thermal imaging temperature field data, laser back reflection light intensity data, and molten pool visual image data; Perform time-series synchronization and feature extraction on multimodal sensing data to generate multidimensional fused feature vectors; The multidimensional fusion feature vector is input into the soft penetration measurement model to generate a real-time penetration estimate. The soft penetration measurement model includes a mechanism analysis sub-model and a data-driven sub-model. The mechanism analysis sub-model is established based on the welding heat conduction mechanism, and the data-driven sub-model is trained based on welding test data. The real-time penetration estimate is generated by weighted fusion of the output of the mechanism analysis sub-model and the output of the data-driven model. Based on the deviation between the real-time melt depth estimate and the target melt depth, a feedback control correction is generated; The heat accumulation state at each position of the welding trajectory is predicted based on the heat accumulation and transfer model, and a feedforward compensation correction is generated. The heat accumulation and transfer model discretizes the welding trajectory into multiple micro-segments and calculates the heat accumulation of the current micro-segment based on the heat conduction relationship between each micro-segment. The feedback control correction and the feedforward compensation correction are superimposed to generate the laser power setting and welding speed setting.

[0005] In some embodiments, time-series synchronization and feature extraction are performed on multimodal sensing data to generate a multidimensional fused feature vector, including: Using the rising edge of the laser emission signal as a unified trigger reference, the plasma light intensity data, infrared thermal image temperature field data, laser back reflection light intensity data, and molten pool visual image data are time-aligned. The time-aligned plasma intensity data are subjected to moving average filtering to extract normalized plasma intensity features. Spatial and temporal filtering are performed on the time-aligned infrared thermal image temperature field data to extract the maximum temperature of the molten pool region, the average temperature of the molten pool region, the area of ​​the molten pool, the aspect ratio of the molten pool, and the temperature gradient features in front of the molten pool. Normalize the time-aligned laser back reflection intensity data and extract the normalized back reflection coefficient feature. Distortion correction and pixel-physical scale calibration are performed on the temporally aligned visual image data of the molten pool, and the features of molten pool length, molten pool width, molten pool trailing length and visual molten pool area are extracted. The normalized plasma light intensity characteristic, the highest temperature of the molten pool region, the average temperature of the molten pool region, the molten pool area, the aspect ratio of the molten pool, the temperature gradient in front of the molten pool, the normalized back reflection coefficient, the molten pool length, the molten pool width, the molten pool tail length, and the visual molten pool area characteristic are combined to generate a multi-dimensional fused feature vector.

[0006] In some embodiments, the multidimensional fused feature vector is input into the melt depth soft measurement model to generate a real-time melt depth estimate, including: Based on the welding heat conduction mechanism, a mapping relationship between the characteristic quantities of the molten pool temperature field and the weld depth is established. The reference values ​​of the weld depth under different combinations of process parameters are calculated by numerical simulation. A mechanism database is constructed. Each record in the mechanism database contains the characteristic quantities of the molten pool temperature field and the characteristic quantities of the molten pool morphology under the corresponding working conditions. The characteristic quantities of the molten pool temperature field and the characteristic quantities of the molten pool morphology are used for similarity matching with the multidimensional fused feature vector. The multidimensional fusion feature vector is matched with each record in the mechanism database for similarity. Based on the matching results, the melting depth reference value is interpolated and estimated to generate the output value of the mechanism analysis sub-model. Using a multi-dimensional fused feature vector as input and the measured weld penetration value from welding experiments as the training objective, a data-driven sub-model is generated by training a gradient boosting tree algorithm. The multi-dimensional fused feature vector is then input into the data-driven sub-model to generate the output value of the data-driven sub-model. Calculate the similarity between the multidimensional fusion feature vector and the most recent record in the mechanism database. Determine the fusion weight between the output value of the mechanism analysis sub-model and the output value of the data-driven sub-model based on the similarity. When the similarity is higher than a preset threshold, the output value of the mechanism analysis sub-model has a larger weight. When the similarity is lower than the preset threshold, the output value of the data-driven sub-model has a larger weight. The output values ​​of the mechanism analysis sub-model and the data-driven sub-model are weighted and summed according to the fusion weights to generate a real-time melt depth estimate.

[0007] In some embodiments, the multidimensional fused feature vector is matched with each record in the mechanism database for similarity, and the melting depth reference value is interpolated and estimated based on the matching results to generate the output value of the mechanism analysis sub-model, including: Calculate the similarity measure between the multidimensional fusion feature vector and the feature vector of each record in the mechanism database, and select the record with the highest similarity measure as the benchmark working condition record. The benchmark working condition record includes the benchmark feature vector and the corresponding benchmark melting depth reference value. Calculate the feature difference vector between the multidimensional fused feature vector and the baseline feature vector. The feature difference vector includes the temperature field feature difference component and the melt pool morphology feature difference component. Based on the mapping relationship of the welding heat conduction mechanism on which the mechanism analysis sub-model is based, the corresponding penetration correction coefficients for the temperature field characteristic difference component and the molten pool morphology characteristic difference component are determined respectively. Based on the sign and magnitude of each component in the feature difference vector, and combined with their respective melt depth correction coefficients, the reference melt depth value is corrected positively or negatively to generate the output value of the mechanism analysis sub-model.

[0008] In some embodiments, a data-driven sub-model is generated by training a gradient boosting tree algorithm using a multi-dimensional fused feature vector as input and the measured weld penetration value from a welding experiment as the training objective. The multi-dimensional fused feature vector is then input into the data-driven sub-model to generate the output value of the data-driven sub-model, including: The mechanism sample data generated by numerical simulation is combined with the test sample data collected by welding test to construct a training dataset. The mechanism sample data includes feature vectors and corresponding penetration depth values ​​under simulated working conditions, while the test sample data includes feature vectors recorded synchronously during welding test and measured penetration depth values ​​after welding. The training dataset is divided into a training subset and a validation subset according to a preset ratio; The hyperparameters of the gradient boosting tree algorithm are optimized by cross-validation to generate optimized hyperparameters; Based on optimized hyperparameters, the model is iteratively trained to generate a data-driven sub-model by taking the feature vectors in the training subset as input and the measured melt depth in the training subset as the training objective. The validation subset is input into the trained data-driven sub-model, and the error index between the predicted melt depth value and the measured melt depth value is calculated. When the error index meets the preset accuracy requirements, the training of the data-driven sub-model is determined to be complete. The multidimensional fused feature vector is input into the trained data-driven sub-model to generate the output value of the data-driven sub-model.

[0009] In some embodiments, a feedback control correction is generated based on the deviation between the real-time melt depth estimate and the target melt depth value, including: Calculate the deviation and the rate of change of the deviation between the real-time melt depth estimate and the target melt depth; The baseline parameters of the PID controller are determined based on the current welding conditions. The mapping relationship between the baseline parameters and the nominal laser power value and the nominal welding speed value is pre-calibrated through a step response test. The deviation and the rate of change of deviation are input into the fuzzy inference engine. The fuzzy inference engine outputs the PID controller parameter correction amount according to the preset fuzzy rule library. The fuzzy rule library contains special rules for welding thermal inertia conditions. The special rules increase the proportional parameter when the deviation and the rate of change of deviation increase in the same direction, and increase the derivative parameter when the deviation and the rate of change of deviation change in opposite directions. The baseline parameters are superimposed with the PID controller parameter corrections to generate the actual PID controller parameters. Based on the actual parameters of the PID controller, the deviation is calculated using proportional-integral-derivative operations to generate the laser power feedback correction amount. When the absolute value of the laser power feedback correction exceeds the preset threshold, a welding speed feedback correction is generated based on the laser power feedback correction. The welding speed feedback correction is in the same direction as the laser power feedback correction and its amplitude is proportional.

[0010] In some embodiments, the deviation and the rate of change of deviation are input to the fuzzy inference engine, and the fuzzy inference engine outputs the PID controller parameter correction amount according to a preset fuzzy rule base, including: The deviation and the rate of change of deviation are fuzzified to generate fuzzy variables of deviation and fuzzy variables of the rate of change of deviation. Each of the fuzzy variables of deviation and the fuzzy variables of the rate of change of deviation corresponds to multiple fuzzy linguistic values. The deviation fuzzy variable and the deviation change rate fuzzy variable are matched with the antecedents of each fuzzy rule in the preset fuzzy rule library. Fuzzy rules whose antecedents meet the conditions are activated. The preset fuzzy rule library contains multiple fuzzy rules. The antecedent of each fuzzy rule is a combination of the deviation fuzzy linguistic value and the deviation change rate fuzzy linguistic value. The consequent of each fuzzy rule is a combination of the proportional parameter correction fuzzy linguistic value, the integral parameter correction fuzzy linguistic value, and the differential parameter correction fuzzy linguistic value. Based on the matching degree of the antecedent of the activated fuzzy rule, the consequent of the activated fuzzy rule is weighted and aggregated to generate fuzzy outputs of proportional parameter correction, integral parameter correction, and differential parameter correction. The fuzzy outputs of proportional parameter correction, integral parameter correction, and derivative parameter correction are defuzzified to generate PID controller parameter correction values.

[0011] In some embodiments, the heat accumulation state at each position of the welding trajectory is predicted based on the heat accumulation transfer model to generate a feedforward compensation correction, including: The welding trajectory is discretized into multiple equally spaced welding micro-segments; For each welding micro-segment, based on the one-dimensional unsteady heat conduction analytical relationship, the sum of the residual heat of the heat input of each previous welding micro-segment at the current welding micro-segment is calculated to generate the heat accumulation of the current welding micro-segment. The residual heat is positively correlated with the heat input of each previous welding micro-segment and negatively correlated with the distance and time interval between each previous welding micro-segment and the current welding micro-segment. The heat accumulation is normalized based on the nominal heat input of the current welding micro-segment to generate a relative heat accumulation coefficient. When the relative thermal accumulation coefficient is greater than the first preset threshold, it is determined that the current welding micro-segment is in a strong thermal accumulation state, and a negative laser power feedforward compensation correction amount is generated according to the degree of deviation between the relative thermal accumulation coefficient and the reference value. When the relative heat accumulation coefficient is less than the second preset threshold, it is determined that the current welding micro-segment is in a weak heat accumulation state, and a positive laser power feedforward compensation correction amount is generated according to the degree of deviation between the relative heat accumulation coefficient and the reference value. When the relative thermal accumulation coefficient is between the second preset threshold and the first preset threshold, a zero-value laser power feedforward compensation correction is generated. The laser power feedforward compensation correction is calculated based on the feedforward compensation gain coefficient, the nominal laser power value, and the degree of deviation between the relative thermal accumulation coefficient and the reference value. The feedforward compensation gain coefficient is predetermined through welding calibration tests.

[0012] In some embodiments, for each welding micro-segment, based on a one-dimensional unsteady-state heat conduction analytical relationship, the sum of the residual heat of the heat input of each preceding welding micro-segment at the current welding micro-segment is calculated to generate the heat accumulation of the current welding micro-segment, including: The laser power and welding duration of each preceding welding micro-segment are obtained, and the heat input of each preceding welding micro-segment is generated based on the product of the laser power and the welding duration. Based on the one-dimensional unsteady heat conduction analytical relationship, the thermal diffusion attenuation coefficient from each of the previous welding micro-segments to the current welding micro-segment is calculated. The thermal diffusion attenuation coefficient is negatively correlated with the square of the distance and time interval between each of the previous welding micro-segments to the current welding micro-segment, and positively correlated with the thermal diffusivity of the workpiece material. The thermal diffusivity is predetermined based on the workpiece material. The heat input of each preceding welding micro-segment is multiplied by the corresponding thermal diffusion attenuation coefficient to generate the residual heat of each preceding welding micro-segment at the current welding micro-segment. The residual heat from each of the preceding welded micro-segments is added together to generate the heat accumulation for the current welded micro-segment.

[0013] In some embodiments, the feedback control correction amount and the feedforward compensation correction amount are superimposed to generate the laser power setting value and the welding speed setting value, including: Obtain the nominal laser power value and the nominal welding speed value; The laser power feedback correction and the laser power feedforward compensation correction are superimposed to generate the total power correction. The nominal laser power value is added to the total power correction amount to generate the laser power setting value; Calculate the ratio of the absolute value of the feedback control correction to the nominal laser power value. When the ratio exceeds the preset linkage threshold, generate a speed correction coefficient based on the ratio of the feedback control correction to the nominal laser power value. The speed correction coefficient is in the same direction as the feedback control correction. Multiply the nominal welding speed value by the speed correction factor to generate the welding speed setting value; When the ratio does not exceed the preset linkage threshold, the nominal welding speed value will be directly used as the welding speed setting value.

[0014] Unlike existing technologies, the above-mentioned technical solution provides a real-time monitoring and adaptive parameter control method for laser welding penetration depth. It extracts multi-dimensional fusion feature vectors from multi-modal sensor data during the welding process, inputs these feature vectors into a soft-measurement model for penetration depth composed of a weighted fusion of a mechanism analysis sub-model and a data-driven sub-model, and generates a real-time penetration depth estimate. Based on the deviation between the real-time penetration depth estimate and the target penetration depth, a feedback control correction is generated. Furthermore, based on a heat accumulation transfer model, the heat accumulation state at each position on the welding trajectory is predicted, generating a feedforward compensation correction. The feedback control correction and the feedforward compensation correction are then superimposed to generate the laser power setpoint and welding speed setpoint. This invention achieves real-time and precise penetration depth control, effectively suppresses penetration depth fluctuations caused by heat accumulation, and improves the consistency of welding quality.

[0015] The above description of the invention is merely an overview of the technical solution of this application. In order to enable those skilled in the art to better understand the technical solution of this application and to implement it based on the description and drawings, and to make the above-mentioned objectives and other objectives, features and advantages of this application easier to understand, the following description is provided in conjunction with the specific embodiments and drawings of this application. Attached Figure Description

[0016] The accompanying drawings are only used to illustrate the principles, implementation methods, applications, features, and effects of specific embodiments of the present invention and other related contents, and should not be considered as limitations on this application.

[0017] In the accompanying drawings of the instruction manual: Figure 1 This is a schematic diagram illustrating steps S101 to S106 of the method described in the specific implementation embodiment; Figure 2 This is a schematic diagram illustrating steps S201 to S205 of the method described in a specific implementation. Figure 3 The diagram illustrates steps S301 to S306 of the method described in the specific implementation. Detailed Implementation

[0018] To illustrate the possible application scenarios, technical principles, implementable specific solutions, and achievable objectives and effects of this application in detail, the following description, in conjunction with the listed specific embodiments and accompanying drawings, provides a detailed explanation. The embodiments described herein are merely illustrative of the technical solutions of this application and are therefore intended to limit the scope of protection of this application.

[0019] In this document, the term "embodiment" means that a specific feature, structure, or characteristic described in connection with an embodiment may be included in at least one embodiment of this application. The term "embodiment" appearing in various places throughout the specification does not necessarily refer to the same embodiment, nor does it specifically limit its independence or connection with other embodiments. In principle, in this application, as long as there are no technical contradictions or conflicts, the technical features mentioned in each embodiment can be combined in any way to form corresponding implementable technical solutions.

[0020] Unless otherwise defined, the technical terms used herein have the same meaning as commonly understood by one of ordinary skill in the art to which this application pertains; the use of related terms herein is merely for the purpose of describing particular embodiments and is not intended to limit this application.

[0021] In the description of this application, the term "and / or" is used to describe the logical relationship between objects, indicating that three relationships can exist. For example, A and / or B means: A exists, B exists, and A and B exist simultaneously. Additionally, the character " / " in this document generally indicates that the preceding and following objects have an "or" logical relationship.

[0022] In this application, terms such as “first” and “second” are used only to distinguish one entity or operation from another, and do not necessarily require or imply any actual quantity, hierarchy or order relationship between these entities or operations.

[0023] Without further limitations, the use of terms such as “comprising,” “including,” “having,” or other similar open-ended expressions in this application is intended to cover non-exclusive inclusion, which does not exclude the presence of additional elements in a process, method, or product that includes the stated elements, such that a process, method, or product that includes a list of elements may include not only those defined elements but also other elements not expressly listed, or elements inherent to such a process, method, or product.

[0024] As understood in the Examination Guidelines, in this application, expressions such as "greater than," "less than," and "exceeding" are understood to exclude the stated number; expressions such as "above," "below," and "within" are understood to include the stated number. Furthermore, in the description of the embodiments in this application, "multiple" means two or more (including two), and similar expressions related to "multiple" are also understood in this way, such as "multiple groups" and "multiple times," unless otherwise explicitly specified.

[0025] In the description of the embodiments of this application, the space-related expressions used, such as "center," "longitudinal," "lateral," "length," "width," "thickness," "upper," "lower," "front," "rear," "left," "right," "vertical," "horizontal," "vertical," "top," "bottom," "inner," "outer," "clockwise," "counterclockwise," "axial," "radial," and "circumferential," indicate the orientation or positional relationship based on the orientation or positional relationship shown in the specific embodiments or drawings. They are only for the purpose of describing the specific embodiments of this application or for the reader's understanding, and do not indicate or imply that the device or component referred to must have a specific position, a specific orientation, or be constructed or operated in a specific orientation. Therefore, they should not be construed as limitations on the embodiments of this application.

[0026] The processor described in the embodiments of this application can be implemented by hardware, firmware, software, or a combination thereof. It can be a circuit, one or more of an application-specific integrated circuit (ASIC), a digital signal processor (DSP), a digital signal processing device (DSPD), a programmable logic device (PLD), a field-programmable gate array (FPGA), a central processing unit (CPU), a controller, a microcontroller, or a microprocessor. It also includes other physical, biological, or chemical structures that can implement the same or equivalent functions as the processors listed above, such as biological neurons, quantum computing units, DNA computing units, etc., so that the processor can execute some or all of the steps in the computer program or method involved in the various embodiments of this application, or any combination of the steps mentioned therein.

[0027] The computer program involved in the embodiments can be stored in a computer device readable storage medium, which includes, but is not limited to, disks, magnetic tapes, magnetic cards, floppy disks, flash memory, optical disks, optical cards, read-only memory (ROM), random access memory (RAM), erasable programmable ROM (EPROM), and electrically erasable programmable ROM (EEPROM), etc., and also includes other biological, physical, or chemical structures that can achieve the same or equivalent functions as the storage media listed above, such as DNA, RNA, proteins, and other units with information storage capabilities. In specific embodiments, the storage medium involved can be one of the above-mentioned media types, or a combination of the above-mentioned media types. In different embodiments, the computer program involved in the embodiments can be centrally stored in a single medium, or distributed and stored in multiple media. The memory containing the computer device readable storage medium can be non-volatile memory or random access memory. These computer device readable storage media can be built into the device, or can be connected to the device involved in the embodiments as an external device or part of an external device. In some embodiments, the memory having a computer device readable storage medium is deployed locally; in other embodiments, the memory may be deployed remotely from the processor, for example, as a network-attached memory accessed via RF circuitry or an external port and a communication network, wherein the communication network may be the Internet, one or more intranets, a local area network (LAN), a wide area network (WLAN), a storage area network (SAN), or a suitable combination thereof, as long as computer device access to the memory is enabled. Furthermore, the computer program involved in the embodiments may be stored in plaintext / ciphertext form, or it may be designed as training data, integrated and recombined through model training and implicitly stored in the parameter states of a deep neural network or other machine learning model.

[0028] Please see Figure 1 This embodiment provides a method for real-time monitoring and adaptive parameter control of laser welding penetration depth, including: S101. Acquire multimodal sensing data of the welding process, including plasma light intensity data, infrared thermal imaging temperature field data, laser back reflection light intensity data, and molten pool visual image data. S102. Perform time-series synchronization and feature extraction on multimodal sensing data to generate a multidimensional fusion feature vector; S103. Input the multi-dimensional fusion feature vector into the soft measurement model of weld depth to generate a real-time weld depth estimate. The soft measurement model of weld depth includes a mechanism analysis sub-model and a data-driven sub-model. The mechanism analysis sub-model is established based on the welding heat conduction mechanism, and the data-driven sub-model is trained based on welding test data. The real-time weld depth estimate is generated by weighted fusion of the output of the mechanism analysis sub-model and the output of the data-driven model. S104. Generate feedback control correction amount based on the deviation between the real-time melt depth estimate and the target melt depth value; S105. Predict the heat accumulation state at each position of the welding trajectory according to the heat accumulation and transfer model, and generate feedforward compensation correction. The heat accumulation and transfer model discretizes the welding trajectory into multiple micro-segments, and calculates the heat accumulation of the current micro-segment based on the heat conduction relationship between each micro-segment. S106. The feedback control correction amount and the feedforward compensation correction amount are superimposed to generate the laser power setting value and the welding speed setting value.

[0029] In step S101, plasma light intensity data is obtained by acquiring plasma radiation signals above the molten pool through a photoelectric sensor and a narrowband filter, characterizing the intensity variation of the plasma; infrared thermal image temperature field data is obtained by acquiring the two-dimensional temperature distribution of the molten pool and heat-affected zone through an infrared thermal imager, reflecting the thermal state during the welding process; laser back reflection light intensity data is obtained by separating the laser signal reflected from the molten pool from the laser output optical path through a beam splitter and converting it through a photoelectric detector, characterizing the changes in the absorption and reflection characteristics of the laser on the workpiece surface; molten pool visual image data is obtained by acquiring images of the molten pool area through a high-speed camera and a filter and illumination assembly, providing geometric morphology information of the molten pool.

[0030] In step S102, timing synchronization aligns the acquisition times of each sensor channel using a unified hardware trigger signal, establishing a consistent correspondence between sensor data with different sampling frequencies and transmission delays on the time axis. Feature extraction performs corresponding signal processing and quantization analysis based on the characteristics of each modal data, extracting feature quantities that characterize the welding process state from each modal data. The multidimensional fusion feature vector is a vector form composed of multiple feature quantities obtained after timing synchronization and feature extraction of each modal data, combined in a preset order, comprehensively carrying multidimensional state information of the welding process.

[0031] In step S103, the mechanism analysis sub-model constructs a mapping relationship between the thermal state of the molten pool and the weld depth based on heat conduction theory, reflecting the physical laws of the welding process. The data-driven sub-model learns and fits the statistical relationship between the feature vectors collected in the welding experiment and the measured weld depth value through machine learning algorithms. By assigning weights, the outputs of the mechanism analysis sub-model and the data-driven sub-model complement each other, jointly constraining the weld depth estimation results at both the physical law and data statistics levels, thereby improving the accuracy and adaptability of the weld depth estimation.

[0032] In step S104, the target penetration depth is preset according to the welding process requirements. The feedback control correction is calculated and generated by a closed-loop control algorithm based on the deviation and its changing trend. It is used to adjust the process parameters in real time during the welding process, so that the real-time penetration depth estimate gradually approaches the target penetration depth.

[0033] In step S105, the heat accumulation and transfer model is used to describe the heat transfer and accumulation law along the welding trajectory during the welding process. It divides the continuous welding trajectory into multiple discrete micro-segments at preset intervals, allowing the thermal state at each location to be quantified and calculated point-by-point. For the heat accumulation amount of the current micro-segment, the model is determined by accumulating the residual heat formed at the current micro-segment through heat conduction from the heat input during the welding of previous micro-segments. This residual heat decreases as the distance between the previous and current micro-segments increases, thus reflecting the superimposed influence of the previous welding process on the initial thermal state of the current welding position. The feedforward compensation correction amount is calculated based on the deviation of the current micro-segment's heat accumulation amount from the reference state. It is used to adjust the process parameters in advance before welding the current micro-segment to counteract the penetration drift caused by the heat accumulation effect.

[0034] In step S106, the feedback control correction and the feedforward compensation correction are algebraically summed to generate a comprehensive correction. The laser power setpoint is generated after correcting the reference value of the laser power based on the comprehensive correction, and the welding speed setpoint is generated after correcting the reference value of the welding speed based on the comprehensive correction. By superimposing the feedback correction and the feedforward correction, real-time deviations in the welding process can be corrected, and predictable disturbances such as heat accumulation can be compensated in advance, achieving coordinated adjustment of laser power and welding speed.

[0035] This embodiment comprehensively characterizes the welding process state from multiple dimensions, including light radiation, heat distribution, energy coupling, and geometric morphology, through the acquisition and fusion of multimodal sensing data, providing a rich information foundation for weld penetration estimation. It employs a soft-measurement model for weld penetration that combines mechanistic analysis with data-driven approaches, taking into account both the physical laws and statistical characteristics of the welding process, thus improving the accuracy and adaptability of weld penetration estimation. By combining feedback control with feedforward control based on heat accumulation prediction, it can correct real-time deviations and compensate for predictable disturbances such as heat accumulation in advance, achieving coordinated adaptive adjustment of laser power and welding speed. This effectively improves the control accuracy of weld penetration and the stability of welding quality.

[0036] In some embodiments, time-series synchronization and feature extraction are performed on multimodal sensing data to generate a multidimensional fused feature vector, including: Using the rising edge of the laser emission signal as a unified trigger reference, the plasma light intensity data, infrared thermal image temperature field data, laser back reflection light intensity data, and molten pool visual image data are time-aligned. The time-aligned plasma intensity data are subjected to moving average filtering to extract normalized plasma intensity features. Spatial and temporal filtering are performed on the time-aligned infrared thermal image temperature field data to extract the maximum temperature of the molten pool region, the average temperature of the molten pool region, the area of ​​the molten pool, the aspect ratio of the molten pool, and the temperature gradient features in front of the molten pool. Normalize the time-aligned laser back reflection intensity data and extract the normalized back reflection coefficient feature. Distortion correction and pixel-physical scale calibration are performed on the temporally aligned visual image data of the molten pool, and the features of molten pool length, molten pool width, molten pool trailing length and visual molten pool area are extracted. The normalized plasma light intensity characteristic, the highest temperature of the molten pool region, the average temperature of the molten pool region, the molten pool area, the aspect ratio of the molten pool, the temperature gradient in front of the molten pool, the normalized back reflection coefficient, the molten pool length, the molten pool width, the molten pool tail length, and the visual molten pool area characteristic are combined to generate a multi-dimensional fused feature vector.

[0037] In this embodiment, the instant the laser begins to emit light is taken as the common zero point of the acquisition start time for each sensor channel. Timing alignment is achieved by simultaneously connecting a unified trigger reference to the acquisition control terminal of each sensor channel through a hardware trigger circuit. This ensures that each channel starts acquisition immediately upon receiving the trigger signal. For sensor data with different sampling frequencies or transmission delays, a common time coordinate with the unified trigger reference as the origin is established on the time axis. This ensures that the modal data acquired at the same time reflect the state of the same spatial position during the welding process.

[0038] The moving average filtering smooths high-frequency fluctuations in the signal by taking the arithmetic mean of the time-aligned plasma intensity data within a preset time window. The window length is determined based on the fluctuation characteristics of the plasma intensity signal. The normalized plasma intensity characteristic is the dimensionless ratio obtained by dividing the filtered plasma intensity value by the calibrated reference intensity value. The calibrated reference intensity value is determined by the stable plasma intensity collected before welding under set process parameters. The normalization process eliminates the influence of sensor sensitivity differences and optical path attenuation on the absolute value of the intensity, making plasma intensity data from different batches and under different operating conditions comparable.

[0039] Spatial filtering suppresses spatial noise by weighted averaging of neighboring pixels in the infrared thermal image temperature field data. Temporal filtering filters the temperature values ​​of the same pixel across multiple consecutive frames to suppress random fluctuations in the temporal dimension. The molten pool region is defined by a temperature threshold; the continuous pixel region with a temperature higher than the threshold constitutes the molten pool region. The temperature threshold can be determined based on the melting point of the workpiece material. The highest temperature in the molten pool region is the maximum temperature value within the region. The average temperature is the average temperature of all pixels within the region. The molten pool area is the area obtained by converting the number of pixels covered by the molten pool region to pixel-to-physical scale. The aspect ratio is the ratio of the length of the molten pool region in the welding direction to its width perpendicular to the welding direction. The temperature gradient in front of the molten pool characterizes the rate of temperature change of the heat-affected zone in front of the welding direction along the welding direction.

[0040] Normalization involves dividing the time-aligned laser backreflection intensity data by the baseline backreflection intensity value in the molten pool-free state to generate a normalized backreflection coefficient. The baseline backreflection intensity value is collected during the pre-emission stage before formal welding, when a molten pool has not yet formed on the workpiece surface; at this time, the backreflection intensity reflects the reflection characteristics of the solid surface to the laser. The normalized backreflection coefficient characterizes the degree of change in the reflection characteristics of the molten pool surface to the laser relative to the solid surface at the current moment, eliminating the influence of laser output power fluctuations and optical path attenuation on the absolute value of the backreflection intensity.

[0041] Distortion correction obtains the lens distortion coefficient through calibration and remaps the pixel positions in the molten pool visual image data, eliminating image geometric distortion caused by radial and tangential lens distortion. Pixel-physical scale calibration establishes the conversion relationship between image pixel coordinates and physical coordinates by photographing a calibration board with known physical dimensions, giving the subsequently extracted geometric features a clear physical unit. The molten pool length represents the extension dimension of the molten pool in the welding direction, the molten pool width represents the extension dimension of the molten pool perpendicular to the welding direction, the molten pool tail length represents the extension dimension of the molten pool tail in the opposite welding direction, and the visual molten pool area is the area value obtained by converting the number of pixels enclosed by the molten pool outline to pixel-physical scale.

[0042] This embodiment achieves precise temporal alignment of multimodal sensing data through a unified triggering benchmark, ensuring consistency of each modal data in the time dimension. Preprocessing techniques such as filtering, normalization, and distortion correction are employed to address the characteristics of each modal data, eliminating the influence of sensor differences and the measurement environment on the raw data. Multiple feature quantities extracted from each modal data characterize the welding process state from different physical dimensions, and the resulting multidimensional fusion feature vector provides a complete and low-redundancy input for the soft penetration measurement model, laying a data foundation for accurate penetration estimation.

[0043] Please see Figure 2In some embodiments, a multi-dimensional fused feature vector is input into the melt depth soft measurement model to generate a real-time melt depth estimate, including: S201. Based on the welding heat conduction mechanism, establish the mapping relationship between the characteristic quantity of the molten pool temperature field and the molten depth. Calculate the reference value of the molten depth under different combinations of process parameters through numerical simulation, and construct a mechanism database. Each record in the mechanism database contains the characteristic quantity of the molten pool temperature field and the characteristic quantity of the molten pool morphology under the corresponding working condition. The characteristic quantity of the molten pool temperature field and the characteristic quantity of the molten pool morphology are used for similarity matching with the multi-dimensional fused feature vector. S202. Perform similarity matching between the multidimensional fusion feature vector and each record in the mechanism database, and perform interpolation estimation of the melting depth reference value based on the matching results to generate the output value of the mechanism analysis sub-model. S203. Using the multi-dimensional fused feature vector as input and the measured weld penetration value from the welding test as the training target, a data-driven sub-model is generated by training the gradient boosting tree algorithm. The multi-dimensional fused feature vector is then input into the data-driven sub-model to generate the output value of the data-driven sub-model. S204. Calculate the similarity between the multidimensional fusion feature vector and the most recent record in the mechanism database. Determine the fusion weight between the output value of the mechanism analysis sub-model and the output value of the data-driven sub-model based on the similarity. When the similarity is higher than a preset threshold, the output value of the mechanism analysis sub-model has a larger weight. When the similarity is lower than the preset threshold, the output value of the data-driven sub-model has a larger weight. S205. Based on the fusion weights, the output values ​​of the mechanism analysis sub-model and the data-driven sub-model are weighted and summed to generate a real-time melt depth estimate.

[0044] In step S201, the welding heat conduction mechanism refers to the physical law that after laser energy acts on the workpiece surface, heat is conducted inside the workpiece to form a molten pool. The molten pool temperature field characteristic quantity characterizes the temperature distribution state of the molten pool region, and the molten pool morphology characteristic quantity characterizes the geometry and size of the molten pool. Numerical simulation uses the finite element method to perform thermo-mechanical coupling simulation calculations of the welding process. The process parameter combinations cover the main process variables affecting the weld penetration, such as laser power, welding speed, and defocusing amount. Each parameter is taken within its feasible range according to a preset step size and cross-combined to form multiple working conditions. The mechanism database stores the process parameters, molten pool temperature field characteristic quantities, molten pool morphology characteristic quantities, and corresponding weld penetration reference values ​​for each working condition in a structured form. Among them, the molten pool temperature field characteristic quantities and molten pool morphology characteristic quantities serve as the index basis for subsequent similarity matching.

[0045] In step S202, similarity matching is achieved by calculating the distance metric between the multidimensional fused feature vector and the molten pool temperature field characteristics and molten pool morphology characteristics of each record in the mechanism database. The smaller the distance metric, the closer the current welding state is to the working condition corresponding to that record. Interpolation estimation is based on several records with the highest similarity. The corresponding melt depth reference values ​​are weighted and synthesized according to the closeness between the multidimensional fused feature vector and the feature quantities of these records to generate the output value of the mechanism analysis sub-model. The number of records participating in the estimation can be set according to the coverage density of the mechanism database.

[0046] In step S203, the measured weld penetration value is obtained through destructive testing of the welded sample. The multidimensional fusion feature vector recorded synchronously during the welding test is paired with the corresponding measured weld penetration value to form a training sample set. The gradient boosting tree algorithm uses the multidimensional fusion feature vector as input and the measured weld penetration value as the regression target. Through multiple iterations, a decision tree is generated to fit the prediction residual. The newly generated decision tree is accumulated into the existing model to gradually reduce the prediction residual. Finally, the model outputs data based on the input multidimensional fusion feature vector to drive the sub-model output value.

[0047] In steps S204 and S205, the similarity between the multidimensional fusion feature vector and the most recent record in the mechanism database is quantified by the distance metric between them. This distance metric is the value with the smallest distance to the multidimensional fusion feature vector among all records in the mechanism database. A preset threshold is determined in advance based on the coverage of the mechanism database and the required accuracy of the weld penetration estimation. When the similarity is higher than the preset threshold, it indicates that the current welding state is relatively close to the working condition in the mechanism database, and the reliability of the mechanism analysis sub-model output is high, thus assigning it a larger weight. When the similarity is lower than the preset threshold, it indicates that the current welding state exceeds the coverage of the mechanism database, thus assigning a larger weight to the data-driven sub-model output value. The final weighted summation of the real-time weld penetration estimate is updated in each control cycle as the welding process progresses.

[0048] This embodiment constructs a mechanism database covering multiple sets of process parameters through numerical simulation, enabling the mechanism analysis sub-model to infer the penetration depth based on physical laws. At the same time, the training data of welding test data drives the sub-model, enabling it to infer the penetration depth based on statistical laws. The similarity matching mechanism adaptively adjusts the fusion weight of the outputs of the two sub-models according to the degree of similarity between the current welding state and the mechanism database. In the working condition area where the mechanism database is fully covered, physical laws take the lead, while in the working condition area outside the coverage area, data statistics take the lead. This ensures that the real-time penetration depth estimate can maintain high accuracy and reliability under different working conditions.

[0049] In some embodiments, the multidimensional fused feature vector is matched with each record in the mechanism database for similarity, and the melting depth reference value is interpolated and estimated based on the matching results to generate the output value of the mechanism analysis sub-model, including: Calculate the similarity measure between the multidimensional fusion feature vector and the feature vector of each record in the mechanism database, and select the record with the highest similarity measure as the benchmark working condition record. The benchmark working condition record includes the benchmark feature vector and the corresponding benchmark melting depth reference value. Calculate the feature difference vector between the multidimensional fused feature vector and the baseline feature vector. The feature difference vector includes the temperature field feature difference component and the melt pool morphology feature difference component. Based on the mapping relationship of the welding heat conduction mechanism on which the mechanism analysis sub-model is based, the corresponding penetration correction coefficients for the temperature field characteristic difference component and the molten pool morphology characteristic difference component are determined respectively. Based on the sign and magnitude of each component in the feature difference vector, and combined with their respective melt depth correction coefficients, the reference melt depth value is corrected positively or negatively to generate the output value of the mechanism analysis sub-model.

[0050] In this embodiment, the multi-dimensional fused feature vector and the feature vector of each record in the mechanism database both contain temperature field-related feature components and melt pool morphology-related feature components. Each feature component is assigned a different weighting coefficient based on its dimensional differences and sensitivity to melt depth, ensuring that feature components with larger dimensions or weaker influence on melt depth do not dominate the distance calculation result. The smaller the weighted distance, the closer the current welding state is to the working condition corresponding to that record, and the higher the similarity metric. The baseline working condition record is the record with the smallest weighted distance to the multi-dimensional fused feature vector among all records; its baseline feature vector and baseline melt depth reference value serve as the benchmark for subsequent interpolation correction.

[0051] The feature difference vector is obtained by subtracting each component of the multi-dimensional fused feature vector from the corresponding component of the reference feature vector. The temperature field feature difference component encompasses differences in temperature-related features such as the highest temperature in the molten pool region, the average temperature in the molten pool region, the molten pool area, the aspect ratio of the molten pool, and the temperature gradient in front of the molten pool. The molten pool morphology feature difference component encompasses differences in the geometric morphology features of the molten pool, such as the molten pool length, the molten pool width, the length of the molten pool trail, and the visual molten pool area. The feature difference vector reflects the direction and degree of deviation of the current welding state from the corresponding condition recorded in the reference working condition.

[0052] The melt depth correction coefficients corresponding to the temperature field characteristic difference components are determined based on the physical correspondence between the molten pool temperature field characteristic quantities and the melt depth. The melt depth correction coefficients corresponding to the molten pool morphology characteristic difference components are determined based on the correlation between the molten pool geometry and the melt depth. The correction coefficients corresponding to each component reflect the change in melt depth caused by a unit change in that characteristic component, and their values ​​can be pre-calibrated through regression analysis of multiple sets of records in the mechanism database.

[0053] When making positive or negative corrections to the reference weld penetration value, the sign of each component in the characteristic difference vector determines the correction direction, and the magnitude, combined with the corresponding weld penetration correction coefficient, determines the contribution of that component to the weld penetration correction amount. The correction amounts corresponding to each component are accumulated and then superimposed onto the reference weld penetration value; the result is the output value of the mechanism analysis sub-model. When a component in the characteristic difference vector is positive, it indicates that the current welding condition is higher than the reference condition in that characteristic dimension, and a positive or negative correction is made based on the correlation between this characteristic and the weld penetration; similarly, a negative correction is made when the component is negative.

[0054] This embodiment uses weighted distance calculation to measure the similarity between multidimensional fusion feature vectors and mechanism database records, avoiding the imbalance of matching results caused by feature components with different dimensions. The feature difference vector describes the deviation between the current welding state and the reference working condition from two dimensions: temperature field and weld pool morphology. Combined with the penetration correction coefficient determined according to the welding heat conduction mechanism mapping relationship, the difference at the feature level is quantitatively converted into the correction amount at the penetration level. This allows the output value of the mechanism analysis sub-model to be finely adjusted according to the deviation of the current welding state from the reference working condition, improving the adaptability and output accuracy of the mechanism analysis sub-model under different working conditions.

[0055] In some embodiments, a data-driven sub-model is generated by training a gradient boosting tree algorithm using a multi-dimensional fused feature vector as input and the measured weld penetration value from a welding experiment as the training objective. The multi-dimensional fused feature vector is then input into the data-driven sub-model to generate the output value of the data-driven sub-model, including: The mechanism sample data generated by numerical simulation is combined with the test sample data collected by welding test to construct a training dataset. The mechanism sample data includes feature vectors and corresponding penetration depth values ​​under simulated working conditions, while the test sample data includes feature vectors recorded synchronously during welding test and measured penetration depth values ​​after welding. The training dataset is divided into a training subset and a validation subset according to a preset ratio; The hyperparameters of the gradient boosting tree algorithm are optimized by cross-validation to generate optimized hyperparameters; Based on optimized hyperparameters, the model is iteratively trained to generate a data-driven sub-model by taking the feature vectors in the training subset as input and the measured melt depth in the training subset as the training objective. The validation subset is input into the trained data-driven sub-model, and the error index between the predicted melt depth value and the measured melt depth value is calculated. When the error index meets the preset accuracy requirements, the training of the data-driven sub-model is determined to be complete. The multidimensional fused feature vector is input into the trained data-driven sub-model. After being accumulated and calculated by the multi-layer decision tree inside the model, the corresponding melt depth prediction value is output, which is the output value of the data-driven sub-model.

[0056] In this embodiment, the mechanism sample data is taken from numerical simulation calculations, and the experimental sample data comes from actual welding experiments. Before merging the two types of data, the format is unified and normalized to ensure that the dimensional order and units of the feature vectors are consistent. The merged training dataset covers a wide range of sample spaces from simulation conditions to actual conditions, providing diverse training samples for data-driven sub-models.

[0057] The preset ratio is determined based on the total number of training samples and the complexity of the model. The training subset is used to learn the model parameters, and the validation subset is used to evaluate the generalization ability of the model. Random sampling is used when splitting the subset to ensure the consistency of the sample distribution between the training subset and the validation subset.

[0058] Cross-validation involves dividing the training subset into multiple complementary subsets, alternating between using one subset as validation data and the remaining subsets as training data for multiple training and validation runs. The average of the multiple validation results is used as the model performance evaluation value for that set of hyperparameters. Hyperparameter optimization involves performing cross-validation on different hyperparameter combinations within a predefined range of candidate hyperparameters, selecting the hyperparameter combination with the best performance evaluation value as the optimized hyperparameters. The hyperparameters of the gradient boosting tree algorithm cover key parameters affecting model capacity and generalization ability, such as the number of trees, the maximum tree depth, the learning rate, and the subsampling ratio.

[0059] During iterative training, the gradient boosting tree algorithm generates a decision tree in each iteration to fit the prediction residual of the current model. The newly generated decision tree is then added to the existing model with a certain learning rate. Through multiple iterations, the prediction residual is gradually reduced until the preset number of iterations is reached or the residual change becomes gradual.

[0060] The error index is calculated based on the deviation between the predicted and measured melt depth values. The preset accuracy requirement is pre-set according to the melt depth control accuracy requirements. When the error index meets the preset accuracy requirement, it indicates that the prediction accuracy of the data-driven sub-model has reached a usable level, and training is considered complete.

[0061] This embodiment combines samples from both numerical simulation and welding experiments to construct a training dataset, expanding the coverage of the training samples and enabling the data-driven sub-model to learn a wider range of working conditions. Cross-validation is used to optimize the hyperparameters of the gradient boosting tree algorithm, avoiding underfitting or overfitting problems caused by improper hyperparameter settings. By dividing the training and validation subsets and evaluating error indices, the prediction accuracy and generalization ability of the data-driven sub-model are guaranteed, enabling the output values ​​of the data-driven sub-model to provide reliable data-driven basis for weld penetration estimation in actual welding processes.

[0062] Please see Figure 3 In some embodiments, a feedback control correction is generated based on the deviation between the real-time melt depth estimate and the target melt depth value, including: S301. Calculate the deviation and the rate of change of the deviation between the real-time estimated melting depth and the target melting depth; S302. Determine the reference parameters of the PID controller based on the current welding conditions. The mapping relationship between the reference parameters and the nominal laser power value and the nominal welding speed value is pre-calibrated through a step response test. S303. Input the deviation and the rate of change of deviation into the fuzzy inference engine. The fuzzy inference engine outputs the PID controller parameter correction amount according to the preset fuzzy rule library. The fuzzy rule library contains special rules for welding thermal inertia conditions. The special rules increase the proportional parameter when the deviation and the rate of change of deviation increase in the same direction, and increase the derivative parameter when the deviation and the rate of change of deviation change in opposite directions. S304. Superimpose the reference parameters with the PID controller parameter correction to generate the actual PID controller parameters; S305. Perform proportional-integral-derivative calculations on the deviation based on the actual parameters of the PID controller to generate the laser power feedback correction amount. S306. When the absolute value of the laser power feedback correction exceeds the preset threshold, a welding speed feedback correction is generated based on the laser power feedback correction. The welding speed feedback correction is in the same direction as the laser power feedback correction and its amplitude is proportional.

[0063] In step S301, the deviation between the real-time melt depth estimate and the target melt depth is obtained. A positive value indicates that the current melt depth is less than the target melt depth, while a negative value indicates that the current melt depth is greater than the target melt depth. The deviation change rate is obtained by dividing the change in deviation within adjacent control cycles by the control cycle duration, reflecting the trend of deviation change over time. A positive deviation change rate indicates that the deviation is increasing, while a negative rate indicates that the deviation is decreasing.

[0064] In step S302, the current welding condition is characterized by process setting parameters such as nominal laser power and nominal welding speed. A step response test is conducted, applying power step excitation under different combinations of nominal laser power and nominal welding speed, recording the penetration depth response curve. Based on the dynamic characteristic indicators such as overshoot, settling time, and steady-state error of the response curve, the benchmark parameters that optimize control performance under this condition are determined, thereby establishing the mapping relationship between the benchmark parameters and the nominal laser power and nominal welding speed.

[0065] In step S303, the fuzzy inference engine receives the deviation and the rate of change of deviation as input. Through fuzzification processing, it maps the precise input values ​​to a fuzzy set. Fuzzy inference is then performed according to rules in a preset fuzzy rule base, and the inference result is defuzzified and output as a precise value. The fuzzy rule base is pre-defined based on control experience and process characteristics during welding. The specific rules for welding thermal inertia conditions are designed based on the characteristics of high thermal inertia and delayed penetration response during the welding of high thermal conductivity materials. When the deviation and the rate of change of deviation increase in the same direction, it indicates that the penetration depth is deviating more significantly from the target; in this case, the proportional parameter is increased to accelerate the response speed. When the deviation and the rate of change of deviation change in opposite directions, it indicates that the penetration depth is approaching the target; in this case, the differential parameter is increased to suppress overshoot.

[0066] In steps S304 to S305, the reference parameters provide the basic control parameters under the current operating conditions. The PID controller parameter correction amount dynamically adjusts the basic control parameters according to the real-time deviation and the rate of change of the deviation, so that the control parameters can adapt to the state changes during the welding process. The proportional-integral-derivative operation amplifies the deviation proportionally, accumulates integrally, and predicts differentially before summing. The proportional term responds to the current deviation in real time, the integral term eliminates steady-state error, and the derivative term applies adjustment in advance according to the trend of deviation change. The calculation result is the laser power feedback correction amount.

[0067] In step S306, the preset threshold is predetermined based on the sensitivity of laser power adjustment to the control of weld penetration and the adjustable range of welding speed. When the absolute value of the laser power feedback correction exceeds the preset threshold, it indicates that adjusting the laser power alone is insufficient to meet the weld penetration control requirements. At this time, a welding speed feedback correction is generated based on the laser power feedback correction. The welding speed feedback correction is in the same direction as the laser power feedback correction, and its amplitude is scaled according to a preset proportional coefficient, so that the laser power and welding speed are adjusted in synergy to enhance the control effect.

[0068] This embodiment quantifies the degree and trend of the weld penetration deviation from the target in real time by calculating the deviation and the rate of change of deviation. Based on the mapping relationship calibrated by the step response test, it determines the reference parameters that are suitable for different working conditions. It uses a fuzzy inference engine to dynamically adjust the PID controller parameters according to the deviation and the rate of change of deviation, so that the control parameters can adapt to the response characteristics under the welding thermal inertia condition. It generates the laser power feedback correction through proportional-integral-derivative operations, and adjusts the welding speed in conjunction when the power adjustment amplitude is large, thereby realizing closed-loop feedback control of the weld penetration and improving the response speed and stability of the weld penetration control.

[0069] In some embodiments, the deviation and the rate of change of deviation are input to the fuzzy inference engine, and the fuzzy inference engine outputs the PID controller parameter correction amount according to a preset fuzzy rule base, including: The deviation and the rate of change of deviation are fuzzified to generate fuzzy variables of deviation and fuzzy variables of the rate of change of deviation. Each of the fuzzy variables of deviation and the fuzzy variables of the rate of change of deviation corresponds to multiple fuzzy linguistic values. The deviation fuzzy variable and the deviation change rate fuzzy variable are matched with the antecedents of each fuzzy rule in the preset fuzzy rule library. Fuzzy rules whose antecedents meet the conditions are activated. The preset fuzzy rule library contains multiple fuzzy rules. The antecedent of each fuzzy rule is a combination of the deviation fuzzy linguistic value and the deviation change rate fuzzy linguistic value. The consequent of each fuzzy rule is a combination of the proportional parameter correction fuzzy linguistic value, the integral parameter correction fuzzy linguistic value, and the differential parameter correction fuzzy linguistic value. Based on the matching degree of the antecedent of the activated fuzzy rule, the consequent of the activated fuzzy rule is weighted and aggregated to generate fuzzy outputs of proportional parameter correction, integral parameter correction, and differential parameter correction. The fuzzy outputs of proportional parameter correction, integral parameter correction, and derivative parameter correction are defuzzified to generate PID controller parameter correction values.

[0070] In this embodiment, fuzzification maps the precise values ​​of the deviation and the rate of change of deviation to fuzzy sets within their respective universes of discourse. The mapping process can be completed according to a pre-defined membership function, which describes the numerical range corresponding to each fuzzy linguistic value and the degree to which each value within the range belongs to that fuzzy linguistic value. The multiple fuzzy linguistic values ​​corresponding to the deviation fuzzy variable and the deviation rate of change of deviation fuzzy variable cover the entire range of their universes of discourse. There is an overlap between the membership functions of adjacent fuzzy linguistic values, allowing the deviation and the rate of change of deviation to transition smoothly at the boundaries of the universes of discourse.

[0071] Each fuzzy rule in the preset fuzzy rule base uses a combination of the deviation fuzzy linguistic value and the deviation change rate fuzzy linguistic value as the antecedent, and a combination of the proportional parameter correction fuzzy linguistic values, the integral parameter correction fuzzy linguistic values, and the differential parameter correction fuzzy linguistic values ​​as the consequent. Antecedent matching is achieved by calculating the membership degree of each of the deviation and deviation change rate fuzzy variables to each fuzzy linguistic value under the current input, and then combining these membership degrees to obtain the antecedent matching degree of each fuzzy rule. Fuzzy rules with an antecedent matching degree greater than zero are activated.

[0072] Weighted aggregation combines the consequents of each activated fuzzy rule according to their respective antecedent matching degree. The fuzzy rule with a higher antecedent matching degree contributes more to the aggregation result. Aggregation is performed on the fuzzy linguistic values ​​of the proportional parameter correction, the integral parameter correction, and the differential parameter correction respectively.

[0073] Defuzzing converts fuzzy outputs into precise values, which can be achieved using the centroid method or the maximum membership method. For example, the centroid method can be used to calculate the centroid of the area enclosed by the membership functions of the fuzzy output. The value of the universe of discourse corresponding to this centroid position is then used as the defuzzification result. Defuzzing is then performed on the fuzzy outputs of proportional parameter correction, integral parameter correction, and derivative parameter correction respectively to generate PID controller parameter correction values.

[0074] This embodiment converts the precise values ​​of deviation and deviation change rate into fuzzy language values ​​through fuzzification processing, enabling control experience to be expressed in the form of fuzzy rules. Fuzzy rules related to the current state are filtered out through antecedent matching and activation mechanism. Weighted aggregation is used to fuse the inference results of multiple rules according to the matching degree. Finally, after defuzzification processing, accurate PID controller parameter correction values ​​are generated, realizing the adaptive adjustment of control parameters according to the deviation state.

[0075] In some embodiments, the heat accumulation state at each position of the welding trajectory is predicted based on the heat accumulation transfer model to generate a feedforward compensation correction, including: The welding trajectory is discretized into multiple equally spaced welding micro-segments; For each welding micro-segment, based on the one-dimensional unsteady heat conduction analytical relationship, the sum of the residual heat of the heat input of each previous welding micro-segment at the current welding micro-segment is calculated to generate the heat accumulation of the current welding micro-segment. The residual heat is positively correlated with the heat input of each previous welding micro-segment and negatively correlated with the distance and time interval between each previous welding micro-segment and the current welding micro-segment. The heat accumulation is normalized based on the nominal heat input of the current welding micro-segment to generate a relative heat accumulation coefficient. When the relative thermal accumulation coefficient is greater than the first preset threshold, it is determined that the current welding micro-segment is in a strong thermal accumulation state, and a negative laser power feedforward compensation correction amount is generated according to the degree of deviation between the relative thermal accumulation coefficient and the reference value. When the relative heat accumulation coefficient is less than the second preset threshold, it is determined that the current welding micro-segment is in a weak heat accumulation state, and a positive laser power feedforward compensation correction amount is generated according to the degree of deviation between the relative heat accumulation coefficient and the reference value. When the relative thermal accumulation coefficient is between the second preset threshold and the first preset threshold, a zero-value laser power feedforward compensation correction is generated. The laser power feedforward compensation correction is calculated based on the feedforward compensation gain coefficient, the nominal laser power value, and the degree of deviation between the relative thermal accumulation coefficient and the reference value. The feedforward compensation gain coefficient is predetermined through welding calibration tests.

[0076] In this embodiment, the spacing of the welding trajectory discretization is determined based on the matching relationship between the welding speed and the control cycle and the thermal diffusion characteristics of the workpiece material. Specifically, the product of the welding speed and the control cycle can be used as the reference value of the spacing, and then the reference value is corrected according to the heat conduction range in the workpiece, so that the spacing can reflect the advance distance of the welding position in adjacent control cycles, and ensure that the spatial resolution of heat accumulation calculation is compatible with the physical scale of heat conduction.

[0077] One-dimensional unsteady-state heat conduction analytical relationships describe the physical laws governing the conduction and attenuation of heat within the workpiece along the welding trajectory. The heat input during the welding of each preceding micro-segment diffuses outwards from the workpiece via heat conduction after welding is completed. A portion of this heat remains within the workpiece when the current micro-segment is reached. The magnitude of this residual heat is influenced by the heat input of each preceding micro-segment, the distance between each preceding micro-segment and the current micro-segment, and the time interval between the completion of each preceding micro-segment and the start of the current micro-segment. A larger heat input results in more residual heat, while greater distances or longer time intervals result in less residual heat. The sum of the residual heat from each preceding micro-segment represents the heat accumulation in the current micro-segment.

[0078] The nominal heat input is determined based on the nominal laser power and nominal welding speed, reflecting the heat input level of the current welding micro-segment under standard process conditions. Normalization divides the accumulated heat by the nominal heat input, making the relative heat accumulation coefficient a dimensionless ratio, thus eliminating the influence of differences in the absolute amount of heat input under different operating conditions on the determination of the heat accumulation state.

[0079] The first and second preset thresholds are pre-set based on the critical degree of the impact of heat accumulation on the weld penetration, with the first preset threshold being higher than the second. The baseline value is determined based on the nominal state without heat accumulation. The degree to which the relative heat accumulation coefficient deviates from the baseline value reflects the potential impact of heat accumulation on the current weld penetration of the micro-segment. The feedforward compensation gain coefficient is pre-determined through welding calibration tests. These tests measure the change in weld penetration under different heat accumulation conditions, and the feedforward compensation gain coefficient is determined based on the relationship between the change in weld penetration and the degree of deviation of the relative heat accumulation coefficient.

[0080] This embodiment transforms the continuous welding process into the thermal state calculation of discrete micro-segments by discretizing the welding trajectory. It uses a one-dimensional unsteady heat conduction analytical relationship to quantify the influence of the preceding welding process on the heat accumulation at the current welding position. Normalization is used to eliminate the interference of working condition differences on heat accumulation determination. Based on the comparison result of the relative heat accumulation coefficient and the preset threshold, a corresponding laser power feedforward compensation correction amount is generated, so that the laser power can be pre-adjusted according to the heat accumulation state before the welding reaches the current micro-segment, thereby offsetting the penetration drift caused by the heat accumulation effect.

[0081] In some embodiments, for each welding micro-segment, based on a one-dimensional unsteady-state heat conduction analytical relationship, the sum of the residual heat of the heat input of each preceding welding micro-segment at the current welding micro-segment is calculated to generate the heat accumulation of the current welding micro-segment, including: The laser power and welding duration of each preceding welding micro-segment are obtained, and the heat input of each preceding welding micro-segment is generated based on the product of the laser power and the welding duration. Based on the one-dimensional unsteady heat conduction analytical relationship, the thermal diffusion attenuation coefficient from each of the previous welding micro-segments to the current welding micro-segment is calculated. The thermal diffusion attenuation coefficient is negatively correlated with the square of the distance and time interval between each of the previous welding micro-segments to the current welding micro-segment, and positively correlated with the thermal diffusivity of the workpiece material. The thermal diffusivity is predetermined based on the workpiece material. The heat input of each preceding welding micro-segment is multiplied by the corresponding thermal diffusion attenuation coefficient to generate the residual heat of each preceding welding micro-segment at the current welding micro-segment. The residual heat from each of the preceding welded micro-segments is added together to generate the heat accumulation for the current welded micro-segment.

[0082] In this embodiment, the laser power and welding duration of each preceding welding micro-segment are obtained from the real-time control data of the welding process. The real-time control data records the laser power setting value and the duration of welding for each micro-segment. The range of each preceding welding micro-segment is determined based on the heat conduction distance in the workpiece. The residual heat of micro-segments that are more than the heat-affected zone of the current welding micro-segment can be ignored.

[0083] The thermal diffusivity attenuation coefficient is determined by the distance and time interval between each preceding welding micro-segment and the current welding micro-segment, based on the segmentation position and welding sequence. The distance is the arc length between each preceding welding micro-segment and the current welding micro-segment, and the time interval is the duration between the completion time of each preceding welding micro-segment and the start time of the current welding micro-segment. The thermal diffusivity is predetermined based on the workpiece material; different materials have different thermal diffusivity. A higher thermal diffusivity indicates faster heat conduction within the workpiece, resulting in a higher thermal diffusivity attenuation coefficient. The thermal diffusivity attenuation coefficient decreases with increasing distance and time interval, but increases with increasing thermal diffusivity. The residual heat at the current welding micro-segment reflects the proportion of heat input from each preceding welding micro-segment that remains at the current welding micro-segment location after conduction attenuation.

[0084] This embodiment calculates the heat input by acquiring the laser power and welding duration of each preceding welding micro-segment, calculates the heat diffusion attenuation coefficient using the one-dimensional unsteady heat conduction analytical relationship, multiplies the heat input by the heat diffusion attenuation coefficient to obtain the residual heat of each preceding micro-segment, and accumulates them to generate the heat accumulation amount. This realizes the quantitative calculation of the heat accumulation effect during the welding process and provides a quantitative basis for the generation of feedforward compensation correction amount.

[0085] In some embodiments, the feedback control correction amount and the feedforward compensation correction amount are superimposed to generate the laser power setting value and the welding speed setting value, including: Obtain the nominal laser power value and the nominal welding speed value; The laser power feedback correction and the laser power feedforward compensation correction are superimposed to generate the total power correction. The nominal laser power value is added to the total power correction value to generate the laser power setting value. This setting value is updated and sent to the laser for execution in each control cycle. Calculate the ratio of the absolute value of the feedback control correction to the nominal laser power value. When the ratio exceeds the preset linkage threshold, generate a speed correction coefficient based on the ratio of the feedback control correction to the nominal laser power value. The speed correction coefficient is in the same direction as the feedback control correction. Multiply the nominal welding speed value by the speed correction factor to generate the welding speed setting value; When the ratio does not exceed the preset linkage threshold, the nominal welding speed value will be directly used as the welding speed setting value.

[0086] In this embodiment, the nominal laser power value and the nominal welding speed value are preset by the welding process program and serve as the standard process parameters for the current welding task. The feedback correction amount is generated based on the real-time penetration deviation, and the feedforward compensation correction amount is generated based on the predicted thermal accumulation state. When the two are in the same direction, the amplitude of the total power correction amount increases, and when they are in opposite directions, the amplitude of the total power correction amount decreases.

[0087] The preset linkage threshold is pre-set based on the sensitivity of power adjustment to the penetration depth control and the adjustable range of welding speed. When the ratio of the absolute value of the feedback control correction to the nominal laser power value exceeds the preset linkage threshold, it indicates that the power correction amplitude is large. Adjusting the laser power alone may cause excessive changes in the weld pool width. In this case, a speed correction coefficient is generated based on the ratio of the feedback control correction to the nominal laser power value. The speed correction coefficient is in the same direction as the feedback control correction, that is, the speed increases synchronously when the power increases and decreases synchronously when the power decreases. When the ratio does not exceed the preset linkage threshold, the power correction amplitude is small, and the welding speed can meet the control requirements by maintaining the nominal value. The nominal welding speed value is directly used as the welding speed setpoint.

[0088] This embodiment generates a total power correction by superimposing the laser power feedback correction and the laser power feedforward compensation correction, so that the power setting value takes into account both real-time deviation correction and heat accumulation compensation. The speed linkage is triggered by comparing the ratio with the preset linkage threshold. When the power correction amplitude is large, the welding speed is adjusted synchronously so that the laser power and welding speed change in synergy, which maintains the weld formation quality while ensuring the accuracy of the penetration control.

[0089] In some embodiments, the real-time monitoring and adaptive parameter control method for laser welding penetration depth further includes: Destructive testing of the weld after welding is carried out according to a set ratio to obtain the measured penetration depth and the corresponding multi-dimensional fusion feature vector. The measured melt depth values ​​and corresponding multidimensional fusion feature vectors were added to the training dataset and mechanism database. When the amount of new data reaches a preset proportion of the original dataset size, incremental retraining of the data-driven sub-model and updating of the mechanism database are triggered.

[0090] In this embodiment, the set ratio is predetermined based on production batch and quality control requirements. Destructive testing extracts a corresponding proportion of samples from the post-weld weld for metallographic measurement to obtain the measured penetration depth value. Data corresponding to the sample is retrieved from the multi-dimensional fusion feature vector recorded synchronously during the welding process. The measured penetration depth value and the corresponding multi-dimensional fusion feature vector are added as new samples to the training dataset and the mechanism database, respectively. This allows the data-driven sub-model to learn the mapping rules under new working conditions, and the mechanism database to cover a wider range of processes. The preset ratio is predetermined based on the model accuracy decay rate and data accumulation efficiency. When the amount of new data reaches a preset ratio of the original dataset size, incremental retraining of the data-driven sub-model and updating of the mechanism database are triggered, enabling the model to adapt to process drift caused by factors such as material batch fluctuations and equipment status changes. This embodiment, through continuous feedback of post-weld inspection data and incremental model updates, ensures that the penetration depth soft measurement model maintains predictive accuracy in long-term production.

[0091] In some embodiments, the real-time monitoring and adaptive parameter control method for laser welding penetration depth further includes: During the pre-light emission stage before formal welding, the baseline intensity of the laser back reflection light in the state without molten pool is collected; The surface reflectivity index is generated based on the ratio of the laser back reflection intensity baseline to the standard back reflection intensity baseline. The surface condition of the incoming material is determined based on the surface reflectivity index. When the surface reflectivity index exceeds the first surface threshold, it is determined that the surface of the incoming material is oxidized or contaminated, and the nominal laser power value is automatically increased. When the surface reflectivity index is lower than the second surface threshold, it is determined that the surface of the incoming material is abnormal, welding is suspended and an alarm is triggered. During the welding process, the assembly gap index is generated based on the ratio of the molten pool length feature in the molten pool visual image data to the molten pool length reference value under the nominal gap condition. The assembly gap status is determined based on the assembly gap index. When the assembly gap index exceeds the first gap threshold, the welding speed setting is reduced and the laser power setting is increased; when the assembly gap index is lower than the second gap threshold, the welding speed setting is increased.

[0092] In this embodiment, the laser emits light at low power during the pre-emission stage. At this time, a molten pool has not yet formed on the workpiece surface. The collected laser back-reflection intensity baseline reflects the reflection characteristics of the solid surface to the laser. The standard back-reflection intensity baseline is obtained by pre-emission of a brand-new polished standard sample under the same conditions. The ratio of the laser back-reflection intensity baseline to the standard back-reflection intensity baseline is the surface reflectivity index. The degree to which this index deviates from the reference value characterizes the oxidation or contamination state of the incoming material surface.

[0093] The first and second surface thresholds are preset based on the critical degree of the influence of surface condition on laser absorption rate. When the surface reflectivity index exceeds the first surface threshold, it indicates that surface oxidation or contamination has led to an increase in reflectivity, and the nominal laser power value is automatically increased to compensate for the decrease in energy coupling efficiency. When the surface reflectivity index is lower than the second surface threshold, it indicates that the surface condition is abnormal, welding is suspended and an alarm is triggered.

[0094] The baseline value of the weld pool length under nominal gap conditions is pre-calibrated through welding tests when the assembly gap is within the standard range. The ratio of the weld pool length characteristic value to the baseline value is the assembly gap index. The first and second gap thresholds are pre-set based on the critical degree of influence of assembly gap changes on weld penetration. When the assembly gap index exceeds the first gap threshold, it indicates that the gap is too large; the welding speed setting is reduced and the laser power setting is increased to increase the weld pool filling volume. When the assembly gap index is below the second gap threshold, it indicates that the gap is too small; the welding speed setting is increased to prevent overheating. This embodiment achieves adaptive compensation of process parameters under fluctuating incoming material conditions through pre-weld surface condition recognition and in-weld gap condition recognition.

[0095] In some embodiments, the real-time monitoring and adaptive parameter control method for laser welding penetration depth further includes: When the real-time melt depth estimate is lower than the non-fusion threshold for multiple consecutive control cycles, it is determined to be an non-fusion anomaly, triggering the laser shut-off protection. When the real-time melt depth estimate is higher than the burn-through threshold, it is judged as a burn-through risk and triggers the laser shut-off protection. When the plasma light intensity data shows a precipitous drop lasting for more than a preset duration and the drop exceeds a preset ratio, it is determined to be an abnormal keyhole collapse. The laser power is reduced to a preset ratio and an alarm is triggered within a set time to prevent the keyhole collapse from causing the weld quality to deteriorate.

[0096] In this embodiment, the non-fusion threshold is determined based on the minimum weld penetration required for weld strength, while the burn-through threshold is determined based on the workpiece wall thickness and the upper limit of the weld penetration. If the real-time weld penetration estimate is lower than the non-fusion threshold for multiple consecutive control cycles, it indicates insufficient weld penetration, triggering laser shut-off protection to prevent defective welds. If the real-time weld penetration estimate is higher than the burn-through threshold, it indicates a risk of burn-through, immediately triggering laser shut-off protection. The preset duration and preset ratio are predetermined based on the normal fluctuation range of the plasma intensity signal and the characteristic law of keyhole instability. This embodiment achieves safety protection during the welding process through dual anomaly monitoring of the weld penetration estimate and the plasma intensity signal.

[0097] In some embodiments, determining the reference parameters of the PID controller based on the current welding conditions includes: Establish a mapping relationship between the benchmark parameters and the nominal laser power value and the nominal welding speed value. The mapping relationship is pre-calibrated through step response tests under different nominal working conditions. Store the mapping relationship in the parameter lookup table; When switching to a welding program for a different product, the corresponding reference parameters are automatically retrieved from the parameter lookup table based on the nominal laser power and nominal welding speed of the current product.

[0098] In this embodiment, the parameter lookup table stores the reference parameters calibrated for each nominal operating condition, indexed by the nominal laser power value and the nominal welding speed value. Different products have their own nominal laser power value and nominal welding speed value in their welding programs. When switching welding programs, the matching reference parameters are automatically retrieved from the parameter lookup table based on the current product's nominal laser power value and nominal welding speed value. This embodiment achieves rapid adaptation of the PID controller's reference parameters when switching welding programs for different products using a lookup table approach.

[0099] By adopting the above technical solutions, this invention differs from existing technologies and has the following beneficial effects: Through the temporal synchronization and feature extraction of multimodal sensing data, the welding process state is characterized from multiple dimensions such as light radiation, heat distribution, energy coupling, and geometric morphology. The generated multidimensional fused feature vector serves as the input to the soft-measurement model of weld penetration. The mechanism analysis sub-model and the data-driven sub-model adaptively weight and fuse the model based on the similarity between the current working condition and the mechanism database, outputting a real-time weld penetration estimate that takes into account both physical laws and statistical data characteristics. Feedback control generates a correction amount based on the deviation between the real-time weld penetration estimate and the target weld penetration value, and feedforward compensation... Based on the heat accumulation and transfer model, the heat accumulation state at each position of the welding trajectory is predicted and a correction value is generated in advance. The two are superimposed to generate the laser power setting value and the welding speed setting value. While correcting the real-time deviation, the penetration depth drift caused by heat accumulation is offset, and the laser power and welding speed are coordinated and adjusted. The data-driven sub-model and mechanism database are incrementally updated by the post-weld inspection data, and the process parameters are adaptively adjusted according to the surface condition of the incoming material and the assembly gap condition. With the real-time protection of penetration depth abnormality and keyhole collapse, the penetration depth control accuracy is maintained in long-term production, and the stability and consistency of welding quality are improved.

[0100] Finally, it should be noted that although the above embodiments have been described in the text and drawings of this application, this should not limit the scope of patent protection of this application. Any technical solutions that are based on the essential concept of this application and utilize the content described in the text and drawings of this application, resulting in equivalent structural or procedural substitutions or modifications, as well as the direct or indirect application of the technical solutions of the above embodiments to other related technical fields, are all included within the scope of patent protection of this application.

Claims

1. A method for real-time monitoring and adaptive parameter control of laser welding penetration depth, characterized in that, include: Acquire multimodal sensing data of the welding process, including plasma light intensity data, infrared thermal imaging temperature field data, laser back reflection light intensity data, and molten pool visual image data; The multimodal sensing data is time-series synchronized and feature extracted to generate a multidimensional fusion feature vector; The multidimensional fused feature vector is input into the soft penetration measurement model to generate a real-time penetration estimate. The soft penetration measurement model includes a mechanism analysis sub-model and a data-driven sub-model. The mechanism analysis sub-model is established based on the welding heat conduction mechanism, and the data-driven sub-model is trained based on welding test data. The real-time penetration estimate is generated by weighted fusion of the output of the mechanism analysis sub-model and the output of the data-driven model. Based on the deviation between the real-time melt depth estimate and the target melt depth, a feedback control correction amount is generated; The heat accumulation state at each position of the welding trajectory is predicted based on the heat accumulation and transfer model, and a feedforward compensation correction is generated. The heat accumulation and transfer model discretizes the welding trajectory into multiple micro-segments and calculates the heat accumulation of the current micro-segment based on the heat conduction relationship between each micro-segment. The feedback control correction amount is superimposed with the feedforward compensation correction amount to generate the laser power setting value and the welding speed setting value.

2. The method for real-time monitoring and adaptive parameter control of laser welding penetration depth according to claim 1, characterized in that, The multimodal sensing data is time-synchronized and feature extracted to generate a multidimensional fused feature vector, including: Using the rising edge of the laser emission signal as a unified trigger reference, the plasma light intensity data, infrared thermal image temperature field data, laser back reflection light intensity data, and molten pool visual image data are time-aligned. The time-aligned plasma intensity data are subjected to moving average filtering to extract normalized plasma intensity features. Spatial and temporal filtering are performed on the time-aligned infrared thermal image temperature field data to extract the maximum temperature of the molten pool region, the average temperature of the molten pool region, the area of ​​the molten pool, the aspect ratio of the molten pool, and the temperature gradient features in front of the molten pool. Normalize the time-aligned laser back reflection intensity data and extract the normalized back reflection coefficient feature. Distortion correction and pixel-physical scale calibration are performed on the temporally aligned visual image data of the molten pool, and the features of molten pool length, molten pool width, molten pool trailing length and visual molten pool area are extracted. The normalized plasma light intensity characteristic, the highest temperature of the molten pool region, the average temperature of the molten pool region, the molten pool area, the molten pool aspect ratio, the temperature gradient in front of the molten pool, the normalized back reflection coefficient, the molten pool length, the molten pool width, the molten pool tail length, and the visual molten pool area characteristic are combined to generate the multidimensional fused feature vector.

3. The method for real-time monitoring and adaptive parameter control of laser welding penetration depth according to claim 1, characterized in that, The multidimensional fused feature vector is input into the melt depth soft measurement model to generate a real-time melt depth estimate, including: Based on the welding heat conduction mechanism, a mapping relationship between the molten pool temperature field characteristics and the weld depth is established. The reference values ​​of the weld depth under different combinations of process parameters are calculated by numerical simulation, and a mechanism database is constructed. Each record in the mechanism database contains the molten pool temperature field characteristics and molten pool morphology characteristics under the corresponding working conditions. The molten pool temperature field characteristics and molten pool morphology characteristics are used for similarity matching with the multidimensional fused feature vector. The multidimensional fusion feature vector is matched with each record in the mechanism database for similarity. Based on the matching results, the melting depth reference value is interpolated and estimated to generate the output value of the mechanism analysis sub-model. Using the multidimensional fused feature vector as input and the measured weld penetration value as the training target, a data-driven sub-model is generated by training the gradient boosting tree algorithm. The multidimensional fused feature vector is then input into the data-driven sub-model to generate the output value of the data-driven sub-model. Calculate the similarity between the multidimensional fusion feature vector and the most recent record in the mechanism database, and determine the fusion weight between the output value of the mechanism analysis sub-model and the output value of the data-driven sub-model based on the similarity. Wherein, the output value of the mechanism analysis sub-model has a larger weight when the similarity is higher than a preset threshold, and the output value of the data-driven sub-model has a larger weight when the similarity is lower than the preset threshold. The output values ​​of the mechanism analysis sub-model and the data-driven sub-model are weighted and summed according to the fusion weights to generate the real-time melt depth estimate.

4. The method for real-time monitoring and adaptive parameter control of laser welding penetration depth according to claim 3, characterized in that, The multidimensional fused feature vector is matched with each record in the mechanism database for similarity. Based on the matching results, the melting depth reference value is interpolated and estimated to generate the output value of the mechanism analysis sub-model, including: Calculate the similarity metric between the multidimensional fusion feature vector and the feature vector of each record in the mechanism database, and select the record with the highest similarity metric as the benchmark working condition record. The benchmark working condition record includes the benchmark feature vector and the corresponding benchmark melting depth reference value. Calculate the feature difference vector between the multidimensional fused feature vector and the baseline feature vector, wherein the feature difference vector includes temperature field feature difference components and melt pool morphology feature difference components; Based on the welding heat conduction mechanism mapping relationship on which the mechanism analysis sub-model is based, the corresponding weld depth correction coefficients for the temperature field characteristic difference component and the weld pool morphology characteristic difference component are determined respectively. Based on the sign and magnitude of each component in the feature difference vector, and combined with their respective melt depth correction coefficients, the reference melt depth value is positively or negatively corrected to generate the output value of the mechanism analysis sub-model.

5. The method for real-time monitoring and adaptive parameter control of laser welding penetration depth according to claim 3, characterized in that, Using the multidimensional fused feature vector as input and the measured weld penetration value from welding experiments as the training objective, a data-driven sub-model is generated through gradient boosting tree algorithm training. The multidimensional fused feature vector is then input into the data-driven sub-model to generate the data-driven sub-model output value, including: The mechanism sample data generated by numerical simulation is combined with the test sample data collected by welding test to construct a training dataset. The mechanism sample data includes feature vectors and corresponding penetration depth values ​​under simulated working conditions, and the test sample data includes feature vectors recorded synchronously in the welding test and measured penetration depth values ​​after welding. The training dataset is divided into a training subset and a validation subset according to a preset ratio; The hyperparameters of the gradient boosting tree algorithm are optimized by cross-validation to generate optimized hyperparameters; Based on the optimized hyperparameters, the data-driven sub-model is generated by iteratively training, using the feature vectors in the training subset as input and the measured melt depth in the training subset as the training objective. The validation subset is input into the trained data-driven sub-model, and the error index between the predicted melt depth value and the measured melt depth value is calculated. When the error index meets the preset accuracy requirement, the training of the data-driven sub-model is determined to be complete. The multidimensional fused feature vector is input into the trained data-driven sub-model to generate the output value of the data-driven sub-model.

6. The method for real-time monitoring and adaptive parameter control of laser welding penetration depth according to claim 1, characterized in that, Based on the deviation between the real-time melt depth estimate and the target melt depth value, a feedback control correction is generated, including: Calculate the deviation and the rate of change of the deviation between the real-time melt depth estimate and the target melt depth value; The reference parameters of the PID controller are determined based on the current welding conditions. The mapping relationship between the reference parameters and the nominal laser power value and the nominal welding speed value is pre-calibrated through a step response test. The deviation and the rate of change of the deviation are input to the fuzzy inference engine. The fuzzy inference engine outputs the PID controller parameter correction amount according to the preset fuzzy rule library. The fuzzy rule library contains special rules for welding thermal inertia conditions. The special rules increase the proportional parameter when the deviation and the rate of change of the deviation increase in the same direction, and increase the derivative parameter when the deviation and the rate of change of the deviation change in opposite directions. The baseline parameters are superimposed with the PID controller parameter correction values ​​to generate the actual PID controller parameters. Based on the actual parameters of the PID controller, a proportional-integral-derivative operation is performed on the deviation to generate a laser power feedback correction amount. When the absolute value of the laser power feedback correction exceeds a preset threshold, a welding speed feedback correction is generated based on the laser power feedback correction. The welding speed feedback correction is in the same direction as the laser power feedback correction and its amplitude is proportional.

7. The method for real-time monitoring and adaptive parameter control of laser welding penetration depth according to claim 6, characterized in that, The deviation and the rate of change of the deviation are input to the fuzzy inference engine, which outputs PID controller parameter corrections based on a preset fuzzy rule base, including: The deviation and the rate of change of the deviation are respectively fuzzified to generate a fuzzy variable of deviation and a fuzzy variable of the rate of change of deviation. Each of the fuzzy variable of deviation and the fuzzy variable of the rate of change of deviation corresponds to multiple fuzzy language values. The deviation fuzzy variable and the deviation change rate fuzzy variable are matched with the antecedents of each fuzzy rule in the preset fuzzy rule library, and the fuzzy rule whose antecedent meets the condition is activated. The preset fuzzy rule library contains multiple fuzzy rules. The antecedent of each fuzzy rule is a combination of deviation fuzzy linguistic value and deviation change rate fuzzy linguistic value, and the consequent of each fuzzy rule is a combination of proportional parameter correction fuzzy linguistic value, integral parameter correction fuzzy linguistic value and differential parameter correction fuzzy linguistic value. Based on the matching degree of the antecedent of the activated fuzzy rule, the consequent of the activated fuzzy rule is weighted and aggregated to generate fuzzy output of proportional parameter correction, fuzzy output of integral parameter correction, and fuzzy output of differential parameter correction. The fuzzy outputs of the proportional parameter correction, integral parameter correction, and derivative parameter correction are defuzzified to generate the PID controller parameter correction values.

8. The method for real-time monitoring and adaptive parameter control of laser welding penetration depth according to claim 1, characterized in that, Based on the heat accumulation and transfer model, the heat accumulation state at each position of the welding trajectory is predicted, and feedforward compensation corrections are generated, including: The welding trajectory is discretized into multiple equally spaced welding micro-segments; For each welding micro-segment, based on the one-dimensional unsteady heat conduction analytical relationship, the sum of the residual heat at the current welding micro-segment is calculated from the heat input of each preceding welding micro-segment to generate the heat accumulation of the current welding micro-segment. The residual heat is positively correlated with the heat input of each preceding welding micro-segment and negatively correlated with the distance and time interval between each preceding welding micro-segment and the current welding micro-segment. The heat accumulation is normalized based on the nominal heat input of the current welding micro-segment to generate a relative heat accumulation coefficient. When the relative thermal accumulation coefficient is greater than the first preset threshold, it is determined that the current welding micro-segment is in a strong thermal accumulation state, and a negative laser power feedforward compensation correction amount is generated according to the degree of deviation between the relative thermal accumulation coefficient and the reference value. When the relative thermal accumulation coefficient is less than the second preset threshold, it is determined that the current welding micro-segment is in a weak thermal accumulation state, and a positive laser power feedforward compensation correction amount is generated according to the degree of deviation between the relative thermal accumulation coefficient and the reference value. When the relative thermal accumulation coefficient is between the second preset threshold and the first preset threshold, a zero-value laser power feedforward compensation correction amount is generated. The laser power feedforward compensation correction amount is calculated based on the feedforward compensation gain coefficient, the nominal laser power value, and the degree of deviation of the relative thermal accumulation coefficient from the reference value. The feedforward compensation gain coefficient is predetermined through welding calibration tests.

9. The method for real-time monitoring and adaptive parameter control of laser welding penetration depth according to claim 8, characterized in that, For each welding micro-segment, based on the one-dimensional unsteady-state heat conduction analytical relationship, the sum of the residual heat of the heat input of each preceding welding micro-segment at the current welding micro-segment is calculated to generate the heat accumulation of the current welding micro-segment, including: The laser power and welding duration of each preceding welding micro-segment are obtained, and the heat input of each preceding welding micro-segment is generated based on the product of the laser power and the welding duration. Based on the one-dimensional unsteady-state heat conduction analytical relationship, the heat diffusion attenuation coefficient from each of the preceding welding micro-segments to the current welding micro-segment is calculated. The heat diffusion attenuation coefficient is negatively correlated with the square of the distance and time interval between each of the preceding welding micro-segments to the current welding micro-segment, and positively correlated with the heat diffusivity of the workpiece material. The heat diffusivity is predetermined based on the workpiece material. The heat input of each preceding welding micro-segment is multiplied by the corresponding heat diffusion attenuation coefficient to generate the residual heat of each preceding welding micro-segment at the current welding micro-segment. The residual heat from each of the preceding welded micro-segments is added together to generate the heat accumulation amount of the current welded micro-segment.

10. The method for real-time monitoring and adaptive parameter control of laser welding penetration depth according to claim 1, characterized in that, The feedback control correction amount is superimposed with the feedforward compensation correction amount to generate the laser power setting value and the welding speed setting value, including: Obtain the nominal laser power value and the nominal welding speed value; The laser power feedback correction amount is superimposed with the laser power feedforward compensation correction amount to generate a total power correction amount; The nominal laser power value is added to the total power correction amount to generate the laser power setting value; Calculate the ratio of the absolute value of the feedback control correction amount to the nominal laser power value. When the ratio exceeds a preset linkage threshold, generate a speed correction coefficient based on the ratio of the feedback control correction amount to the nominal laser power value. The speed correction coefficient is in the same direction as the feedback control correction amount. The nominal welding speed value is multiplied by the speed correction coefficient to generate the welding speed setting value; When the ratio does not exceed the preset linkage threshold, the nominal welding speed value is directly used as the welding speed setting value.

Images