A laser welding process control system and control method

Abstract

The present application relates to the technical field of laser welding, and particularly relates to a laser welding process control system and a control method, comprising the following steps: S1: synchronously acquiring a multi-source physical signal sequence in a laser welding process, comprising: coaxial visual keyhole morphological features, coaxial OCT absolute depth features, and coaxial photoelectric specific waveband radiation intensity; S2: inputting the physical signal sequence into a preset multi-modal defect prediction network; when a jump amplitude of the radiation intensity exceeds a first set threshold, the fusion weight of the keyhole morphological features is reduced, and the fusion weights of the absolute depth features and the radiation intensity are synchronously increased. Through hardware-level multi-source clock synchronization and a dynamic feature fusion network triggered based on physical laws, in the extreme working condition of visual instantaneous blindness caused by plasma explosion, the fusion weights of the OCT absolute depth and the specific photoelectric waveband features can be adaptively increased by milliseconds, so that the high robustness of defect early prediction is ensured.

Description

Technical Field

[0001] This invention relates to the field of laser welding technology, and specifically to a laser welding process control system and control method. Background Technology

[0002] In cutting-edge manufacturing fields such as new energy vehicle power batteries and aerospace, high-speed deep-penetration laser welding of highly reflective materials (such as copper and aluminum alloys) is becoming increasingly widespread. However, due to the nonlinear abrupt change in the laser absorption rate and the extremely high thermal conductivity of these materials, the hydrodynamic state inside the weld pool and keyhole is extremely unstable, which can easily lead to metal droplet splashing, keyhole collapse, and the resulting process porosity defects, severely restricting the yield of high-end manufacturing.

[0003] Existing laser welding online monitoring and closed-loop control technologies mainly suffer from the following extremely challenging and difficult-to-overcome technical bottlenecks: First, at the level of multimodal feature extraction and fusion, the network is highly susceptible to "collapse" caused by plasma bursts.

[0004] While existing technologies incorporate various sensors, including visual and photoelectric sensors, they typically employ asynchronous software communication, resulting in microsecond to millisecond-level spatiotemporal misalignments that distort the underlying data fusion. More critically, the deep-penetration welding of highly reflective materials easily generates high-frequency, intense plasma plumes. These plumes can instantly overexpose and blind the imaging target surface of a coaxial visual camera. Furthermore, existing AI prediction models use static, fixed feature fusion weights. When visual sensors capture these highly glossy, noisy, distorted images, the static network blindly trusts the visual features, ultimately causing the confidence level of the entire defect prediction chain to collapse instantly.

[0005] Second, at the closed-loop intervention level, there is a technical bias that "overall power reduction leads to non-fusion" and the problem of secondary oscillation.

[0006] The current control logic in the industry is extremely crude: once the monitoring system detects signs of instability or spatter, the conventional approach is to directly reduce the overall output power of the laser proportionally. However, this operation will instantly and drastically reduce the total heat input of the molten pool, causing the molten pool to cool rapidly, which can easily lead to more fatal defects such as incomplete fusion or cold shuts inside the weld. In addition, existing technologies generally adopt a step-like power switching strategy after the alarm is cleared. This instantaneous change in energy not only cannot be smoothly transitioned, but will also trigger a violent thermodynamic reverse impact (secondary oscillation) of the molten metal, arousing even more serious secondary spatter.

[0007] In summary, there is an urgent need in this field for a novel closed-loop control system that can maintain extremely high predictive robustness under strong plasma interference and resolve the defect crisis from the fundamental level of microfluidics without sacrificing overall heat input. Summary of the Invention

[0008] To achieve the above objectives, the present invention provides the following technical solution: This invention provides a laser welding process control method, comprising the following steps: S1: Simultaneously acquire multi-source physical signal sequences during the laser welding process, including: coaxial visual keyhole morphology features, coaxial OCT absolute depth features, and coaxial photoelectric specific band radiation intensity; S2: Input the physical signal sequence into a preset multimodal defect prediction network; when the increase in radiation intensity exceeds a first set threshold, reduce the fusion weight of the keyhole morphology feature and simultaneously increase the fusion weight of the absolute depth feature and the radiation intensity to output the probability of defect occurrence within a set time period in the future. S3: When the probability of the defect occurring is greater than the second set threshold, under the premise of keeping the total power of the laser basically constant, the power of the center beam is reduced by a preset ratio, and the power of the ring beam is simultaneously increased by an equal amount to realize the spatial transfer of the spot energy. S4: When the physical signal sequence detects that the keyhole has stabilized, a preset attenuation function is used to control the center beam and the ring beam to smoothly recover to the initial power ratio within a set time window.

[0009] Furthermore, the specific band radiation intensity of the coaxial optoelectronic system includes the first band radiation intensity and the second band radiation intensity acquired synchronously. The first band is the ultraviolet band of 350nm~450nm; the second band is the infrared band that includes the center wavelength of the laser output. The steps for obtaining the coaxial visual keyhole morphological features include: A sequence of molten pool images is acquired using a narrowband filter that includes the center wavelength of the laser output in the cutoff band and the first band. Edge extraction is performed on the molten pool image sequence to obtain a geometric feature sequence, which includes the two-dimensional area of ​​the keyhole opening, the aspect ratio, and the curvature gradient of the keyhole back wall edge.

[0010] Furthermore, the steps for obtaining the absolute depth features of the coaxial OCT include: It emits a probe light coaxially toward the bottom of the keyhole and receives the reflected signal light. The signal light is interfered with the reference light to form an interference spectrum; The interference spectrum is solved in the frequency domain to extract the absolute depth coordinates of the bottom of the keyhole and construct a depth coordinate sequence. The step of synchronously acquiring the multi-source physical signal sequence during the laser welding process specifically includes: Use a field-programmable gate array or a main control microprocessor as a unified clock source; The unified clock source sends microsecond-level synchronous trigger pulses to the coaxial vision sensor, coaxial OCT sensor, and coaxial photoelectric sensor. The collected keyhole morphology features, absolute depth features, and specific band radiation intensity are assigned global timestamps and spliced ​​together to construct a spatiotemporal input tensor.

[0011] Furthermore, the multimodal defect prediction network includes a spatiotemporal feature extraction module and a dynamic feature fusion module; The step of inputting the physical signal sequence into a preset multimodal defect prediction network includes: The spatiotemporal feature extraction module extracts the keyhole morphology features, the absolute depth features, and the temporal dependence features of the specific band radiation intensity, respectively. The dynamic feature fusion module performs weighted concatenation of each temporally dependent feature according to the real-time allocated fusion weights to generate a multimodal fusion feature vector.

[0012] Furthermore, the criterion for determining that the increase in radiation intensity exceeds the first preset threshold is: Extract the real-time derivative value of the radiation intensity of the specific band within a sliding time window of 1 microsecond to 2 microseconds; When the real-time derivative value is greater than 150% of the average radiation intensity within a set reference time period, it is determined that the jump exceeds the first set threshold.

[0013] Furthermore, the step of reducing the fusion weight of the keyhole morphology feature and simultaneously increasing the fusion weight of the absolute depth feature and the radiation intensity specifically includes: The initial fusion weight of the keyhole morphological features is reduced to 0.1 or below; The fusion weights of the absolute depth feature and the radiation intensity are simultaneously increased, and the sum of the increased fusion weights of the absolute depth feature and the radiation intensity is limited to be greater than or equal to 0.8. The output of the probability of defect occurrence within a set future time period specifically includes: The multimodal fused feature vector is input into the fully connected classification layer; Output the predicted probability of spatter or non-fusion defects occurring in the molten pool within a time window of 2 to 5 milliseconds.

[0014] Furthermore, the step of reducing the power of the center beam by a preset ratio and simultaneously increasing the power of the ring beam by the same amount while keeping the total power of the laser essentially constant is executed using the following energy space transfer model: in, and These represent the initial power of the center beam and the initial power of the ring beam before triggering modulation, respectively. and These are the modulated target power of the center beam and the target power of the ring beam, respectively. Let be the energy transfer coefficient, and set . During the modulation period, the real-time total power of the laser satisfy: ,and The absolute value of the fluctuation deviation relative to the initial total power is less than or equal to .

[0015] Furthermore, the timing requirements for the execution of the trigger beam spatial energy transfer modulation are as follows: From the trigger moment when the probability of the defect occurring exceeds the second set threshold, the laser is controlled to adjust the center beam power downward within a hardware response time window of 2 to 5 milliseconds. And the power of the ring beam is adjusted upwards to The action.

[0016] Furthermore, the determination criteria for the keyhole to return to stability as detected by the physical signal sequence include: The probability of current defect occurrence output by the multimodal defect prediction network decreases below a safe threshold, and the radiation intensity in the specific band remains continuously stable within the average radiation intensity over a set reference time period. The following lasts for at least 5 milliseconds; In the step of controlling the central beam and the annular beam to smoothly recover to their initial power ratio within a set time window, the preset attenuation function is a first-order exponential attenuation function, and its power recovery trajectory satisfies: in, The recovery phase Real-time power of the center beam and real-time power of the ring beam at any given moment. This represents the actual power difference transferred. , The decay time constant is set for the system, and the length of the set time window is 10 milliseconds to 20 milliseconds.

[0017] A laser welding process control system, comprising: A multi-source sensing subunit is configured to synchronously acquire a multi-source physical signal sequence during the laser welding process. The multi-source sensing subunit includes at least a coaxial vision sensor, a coaxial OCT sensor, and a coaxial photoelectric sensor, which are used to extract the coaxial vision keyhole morphology features, the coaxial OCT absolute depth features, and the coaxial photoelectric specific band radiation intensity, respectively. The edge computing main control subunit is communicatively connected to the multi-source sensing subunit and has a built-in preset multimodal defect prediction network configured to receive the physical signal sequence. When it is determined that the increase in radiation intensity exceeds a first set threshold, the fusion weight of the keyhole shape feature is reduced, and the fusion weight of the absolute depth feature and the radiation intensity is simultaneously increased to calculate and output the probability of defect occurrence within a set time in the future. The beam modulation execution subunit, communicatively connected to the edge computing main control subunit, includes a dual-path laser with an independently adjustable center beam and a ring beam; configured to, when the probability of the defect occurrence is greater than a second preset threshold, reduce the power of the center beam by a preset ratio and simultaneously increase the power of the ring beam by the same amount while keeping the total power of the laser basically constant, thereby realizing the spatial transfer of the spot energy; and, after monitoring that the keyhole has recovered to stability, use a preset attenuation function to control the center beam and the ring beam to smoothly recover to the initial power ratio within a set time window.

[0018] Beneficial effects This invention utilizes hardware-level multi-source clock synchronization and a dynamic feature fusion network triggered by physical laws. Even in extreme conditions where plasma bursts cause instantaneous visual blindness, it can adaptively increase the fusion weights of OCT absolute depth and specific photoelectric band features at the millisecond level, ensuring extremely high robustness in defect prediction. Furthermore, this invention breaks away from the traditional bias that "overall power reduction easily leads to incomplete fusion," proposing a spatial transfer model for spot energy with constant total power, reduced center weighting, and increased ring weighting. By weakening the bottom vaporization backpressure and guiding the upper part of the keyhole to open in a trumpet shape, it collaboratively blocks the spatter and porosity formation chain from the fundamental fluid dynamics. Combined with a subsequent exponential damping smooth recovery strategy, it perfectly suppresses secondary physical oscillations in the molten pool thermodynamics. Ultimately, without reducing welding penetration and total heat input, it achieves high-precision, non-destructive, and zero-defect closed-loop control throughout the entire lifecycle of deep penetration welding of highly reactive materials. Attached Figure Description

[0019] Figure 1 This is a flowchart of the control method of the present invention; Figure 2 This is a system architecture diagram of the present invention. Detailed Implementation

[0020] To enable those skilled in the art to better understand the present invention, the technical solutions of the present invention will be clearly and completely described below with reference to the accompanying drawings of the embodiments of the present invention. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort should fall within the scope of protection of the present invention.

[0021] It should be noted that the terms "first," "second," etc., in the specification, claims, and accompanying drawings of this invention are used to distinguish similar objects and are not necessarily used to describe a specific order or sequence. It should be understood that such data can be interchanged where appropriate so that the embodiments of the invention described herein can be implemented in orders other than those illustrated or described herein. Furthermore, the terms "comprising" and "having," and any variations thereof, are intended to cover a non-exclusive inclusion; for example, a process, method, system, product, or apparatus that comprises a series of steps or units is not necessarily limited to those steps or units explicitly listed, but includes other steps or units not explicitly listed or inherent to such processes, methods, products, or apparatus.

[0022] The present invention will now be described in further detail with reference to the accompanying drawings: Example 1: like Figure 1 As shown, a laser welding process control method includes the following steps: S1: Simultaneously acquire multi-source physical signal sequences during the laser welding process, including: coaxial visual keyhole morphology features, coaxial OCT absolute depth features, and coaxial photoelectric specific band radiation intensity; S2: Input the physical signal sequence into a preset multimodal defect prediction network; when the increase in radiation intensity exceeds a first set threshold, reduce the fusion weight of the keyhole morphology feature and simultaneously increase the fusion weight of the absolute depth feature and the radiation intensity to output the probability of defect occurrence within a set time period in the future. S3: When the probability of the defect occurring is greater than the second set threshold, under the premise of keeping the total power of the laser basically constant, the power of the center beam is reduced by a preset ratio, and the power of the ring beam is simultaneously increased by an equal amount to realize the spatial transfer of the spot energy. S4: When the physical signal sequence detects that the keyhole has stabilized, a preset attenuation function is used to control the center beam and the ring beam to smoothly recover to the initial power ratio within a set time window.

[0023] Furthermore, the specific implementation process of step S1 is as follows: During high-speed laser welding, the morphology and internal state of the keyhole change drastically on a microsecond timescale. Traditional single sensors cannot fully reflect this complex physical metallurgical process. Therefore, this embodiment provides a high-precision, highly synchronized multi-source physical signal acquisition architecture.

[0024] 1. Hardware-level microsecond clock synchronization mechanism: To eliminate timing misalignment errors during multimodal data fusion, this embodiment abandons the traditional software polling triggering method and adopts a field-programmable gate array (FPGA) as a globally unified hardware clock source.

[0025] Specifically, the FPGA integrates a high-frequency crystal oscillator to generate a reference clock signal and sends microsecond-level synchronous trigger pulses in parallel to the coaxial vision sensor, coaxial OCT (Optical Coherence Tomography) sensor, and coaxial photoelectric sensor. The system reference sampling period is set to... (For example (corresponding to a system-level sampling rate of 10kHz), any first... Global timestamp of the subsampling task Strictly meet: Through the above synchronization mechanism, the heterogeneous data collected by each sensor will be assigned an absolutely consistent global timestamp, which fundamentally eliminates the problem of spatiotemporal feature misalignment caused by communication delay, and lays the foundation for data alignment for the subsequent construction of high-confidence multimodal input tensors.

[0026] 2. Precise separation and acquisition of coaxial optoelectronic dual-band signals: Laser welding is accompanied by intense plasma plume and metal vapor radiation. In this embodiment, the coaxial photoelectric sensor integrates two bandpass filter channels and corresponding high-frequency photodiodes. The front end of the first channel is configured with a transmittance range of [missing information]. A narrow-band ultraviolet filter is used to isolate ambient light and thermal radiation, and to precisely capture the ultraviolet radiation intensity signal that characterizes the intensity of plasma plume bursts. The second channel has a transmission center wavelength that is exactly the same as the processing laser. Infrared filters are used to continuously capture back reflection signals from the molten pool surface. These two high-frequency photoelectric signals can reflect macroscopic thermodynamic abrupt changes in the welding process without delay.

[0027] 3. High-frequency morphological feature extraction of keyholes based on strong light-resistant vision: To address the technical bias of existing visual monitoring systems being easily "blinded" by plasma, this embodiment innovatively configures a specific composite narrowband filter at the front end of the industrial camera. Its cutoff band strictly covers the output center wavelength of the laser and the aforementioned... The first band. After acquiring a high-resolution sequence of grayscale images of the molten pool, the system not only extracts the conventional two-dimensional area of ​​the keyhole opening. and aspect ratio More importantly, it introduces the "keyhole rear wall edge curvature gradient" feature, which is highly indicative of keyhole rear wall instability.

[0028] The specific extraction process is as follows: First, a sub-pixel level edge detection algorithm is used to extract the contour curve of the keyhole opening edge, which is then parameterized as follows: ,in Let be the arc length parameter. Calculate the instantaneous curvature at each point on the curve. : Since the keyhole back wall typically experiences high-frequency micro-oscillations and a tendency to collapse before splashing occurs, this embodiment further differentiates the curvature along the arc length to obtain the high-frequency curvature gradient. : The system extracts the maximum curvature gradient of the rear wall region in each frame of the image. As a key dimension of the morphological feature sequence, this feature has a sensitivity to minute physical instability that is nearly an order of magnitude higher than that of traditional area features.

[0029] 4. Interferometric solution of absolute depth features in frequency domain OCT: Visual sensors can only acquire two-dimensional surface information. To obtain the three-dimensional depth of the keyhole, this embodiment employs coaxial OCT technology. After beam splitting, one beam serves as a reference beam entering a reference arm with a fixed optical path, while the other beam serves as a signal beam coaxially incident on the bottom of the high aspect ratio keyhole. The signal beam reflected from the bottom of the keyhole interferes with the reference beam and is received by a spectrometer to form a wavenumber response. Changing interference spectrum : in, This represents the power spectrum of the light source. and The reflectivities of the reference arm and the signal arm are respectively. The optical path difference between the two arms (which contains keyhole depth information) is represented by this value. The acquired discrete interference spectrum signal is transformed from the frequency domain to the spatial depth domain by performing a Fast Fourier Transform. By identifying spatial domain functions The peak coordinates in the graph can be used to calculate the absolute depth of the keyhole bottom after removing environmental interference. .

[0030] 5. Construction of the spatiotemporal input tensor: Finally, in each synchronization clock cycle Next, the main control microprocessor aligns the extracted physical quantities according to the global timestamp and concatenates them to form the feature vector of the current moment. : The system then captures the current moment and history. Eigenvectors of each period Construct a two-dimensional spatiotemporal input tensor for use as input to the AI ​​prediction model. : This tensor not only contains multimodal spatial physical features, but also implies the evolution trend of the time dimension.

[0031] Furthermore, the specific implementation process of step S2 is as follows: After obtaining the high-precision spatiotemporal input tensor, this embodiment uses a multimodal defect prediction network deployed in an edge computing main control subsystem (such as an industrial control computer or FPGA architecture with an AI acceleration engine) to perform microsecond-level analysis and prediction of the welding state. The core of this step is to overcome the lack of physical interpretability in traditional "black box" AI algorithms by innovatively introducing a dynamic weight allocation mechanism constrained by specific physical metallurgical phenomena (plasma bursts).

[0032] 1. Network infrastructure and spatiotemporal feature extraction: The multimodal defect prediction network mainly consists of a "spatiotemporal feature extraction module" and a "dynamic feature fusion module." First, the spatiotemporal input tensor is fed in parallel into the spatiotemporal feature extraction module (e.g., using a lightweight Long Short-Term Memory (LSTM) network or a one-dimensional Temporal Convolutional Network (TCN)). This module does not interfere with the data independence of each sensor, but instead extracts the "temporal evolution pattern" (i.e., time-dependent features) of the visual keyhole morphology, OCT absolute depth, and photoelectric radiation intensity over the past few sampling periods. For example, it can identify the keyhole depth over the past... Does the interior show a gradually shallowing collapse trend?

[0033] 2. Determination of triggering conditions under physical phenomenon constraints (plasma burst identification): In the high-speed deep-penetration welding process of highly reflective materials such as new energy batteries (e.g., copper, aluminum alloys), a sudden change in the material's laser absorption rate can instantly trigger a violent plasma plume. Even under the suppression of narrow-band filters, this high-intensity plume can still cause a brief "overexposure" or "blinding" of the imaging target surface of a visual camera, resulting in complete distortion of the extracted keyhole morphology features. If AI models continue to rely on this distorted visual data, it will lead to serious misjudgments.

[0034] Therefore, this embodiment constructs a physical triggering condition based on the high-frequency derivative of the photoelectric signal. The system in... Real-time calculation of ultraviolet radiation intensity within an extremely short sliding time window. First time derivative Its discretization calculation formula is: The system does not rely solely on the absolute value of radiation intensity, but rather monitors its "sudden jump rate". When this real-time derivative value is greater than 150% of the average radiation intensity during a set reference period (e.g., the first 10 milliseconds of the welding plateau), the system determines that a strong plasma burst has occurred above the current physical molten pool, and the confidence level of the visual sensor is on the verge of collapse.

[0035] 3. Dynamic feature fusion and hard real-time weight redistribution: Traditional multi-sensor fusion algorithms typically use static, fixed weight coefficients, which cannot cope with transient changes during the welding process. This invention adopts a dynamic weight allocation strategy triggered by hardware-level interrupts.

[0036] Under normal, stable welding conditions, visual signals are the most intuitive, and the system assigns the keyhole morphology feature the highest weight (e.g., ), depth and radiation characteristics are used as auxiliary factors.

[0037] However, once the aforementioned "plasma burst" condition is triggered, the system will forcibly intervene in the fusion logic of the AI ​​network, performing the following transient reallocation of spatial weights: Visual feature weighting: Instantly weighting the fusion of visual features Forced attenuation to 0.1 or below to maximize the shielding of interference noise from "blinding" images.

[0038] Depth and Radiation Feature Upweighting: Simultaneously weighting the absolute depth features of OCT Weighting of radiation intensity in a specific band The increase is significant, ensuring that the sum of the weights of the two satisfies... .

[0039] It needs further explanation that because OCT uses highly penetrating near-infrared coherent light and only resolves interference fringes, it is minimally affected by plasma emission and can continuously reflect the true depth of the keyhole bottom; while the photoelectric sensor itself captures the state changes of the plasma. Through this dynamic switching of "abandoning vision, preserving depth and photoelectric," the system can still maintain the stable output of the multimodal network feature layer under extremely severe strong light interference, ensuring the uninterrupted connection of the control link.

[0040] 4. Microsecond-level advance prediction of defect occurrence probability: The multimodal fusion feature vector generated after the above dynamic weighted concatenation is fed into the fully connected classification layer at the end of the network. Unlike the logic of traditional monitoring systems that "alarm only after a defect occurs," this model outputs the future... The predicted probability of defects such as "spatter" or "keyhole collapse causing porosity" occurring in the molten pool within the time window. It utilizes the microsecond-level computing speed of a microprocessor to outpace the time difference in physical defect formation (which typically requires several milliseconds of metal liquid flow and stress instability), reserving sufficient physical response time for the laser to perform millisecond-level beam energy spatial transfer in the subsequent step (S3).

[0041] Furthermore, the specific implementation process of step S3 is as follows: After the multimodal defect prediction network outputs microsecond-level advanced prediction results, this embodiment completes transient intervention on the physical molten pool through a beam modulation execution subsystem. This step completely abandons the crude control bias of "predicting defects and then reducing overall power" in traditional laser welding, and innovatively proposes a "dynamic transfer of spatial energy" model based on constant heat input, which changes the hydrodynamic balance of the keyhole through nanosecond-level electro-optic response.

[0042] 1. Defect probability threshold triggering and ultra-fast response mechanism When the main control microprocessor receives the probability of defect occurrence within a future set time frame from the network output... When this occurs, the system compares it with a preset second threshold (i.e., a safety critical threshold, preferably set to 1). Real-time comparison is performed.

[0043] Once the trigger condition is met The system immediately sends hardware-level modulation commands to the underlying driver board of the dual-path laser (e.g., a composite spot continuous fiber laser with independently tunable core and annular cladding). Because this system avoids complex host computer software polling and communication handshake protocols, the transmission delay of this command and the electro-optical response time of the laser shutter are strictly compressed within a certain timeframe. Within an extremely short time window, this rapid response ensures that the system can intervene in the spatial distribution of laser energy before physical splashing (droplet necking and detachment from the molten pool surface) actually occurs.

[0044] 2. A mathematical model for energy space transfer that breaks through conventional biases In traditional control logic, when keyhole instability is predicted, the usual practice is to proportionally reduce the total output power of the laser. However, this operation instantly reduces the overall heat input of the molten pool, causing rapid cooling of the molten pool, which can easily lead to incomplete fusion or cold shut defects inside the weld.

[0045] To overcome this industry pain point, the beam modulation execution subsystem in this embodiment forcibly locks the real-time total power when modulation is triggered. The fluctuation deviation is within a very small range (absolute fluctuation deviation) And execute the following space energy transfer mathematical model: in, and These represent the initial power of the center beam and the initial power of the ring beam before triggering modulation, respectively. and The modulated target power of the center beam and the target power of the ring beam. This is the energy transfer coefficient. In this embodiment, [the value will be...]. Strictly limited to Within the range.

[0046] Within this interval, the system equally "extracts" energy from the central light spot and compensates for it in the annular light spot, strictly satisfying the total energy conservation relationship within the control period: 3. The synergistic inhibition mechanism of physical metallurgy for spatial energy redistribution: The physical mechanism of the energy transfer model under constant power described above is as follows: Center weight reduction to suppress bottom back pressure: Splashing often originates from the upward ejection of metal vapor generated by violent vaporization at the bottom of the keyhole, "blowing" the liquid metal away. In this embodiment, according to proportion... Reduce center beam power This directly weakens the extremely high laser power density acting on the bottom of the keyhole. This instantly reduces the intensity of vaporization of the liquid metal at the bottom, thereby reducing the upward back pressure of the metal vapor and eliminating the underlying driving force that causes the liquid metal to splash.

[0047] Ring-shaped weighting guides the keyhole to expand in a "trumpet-like" shape: while reducing the center power, the ring beam power is simultaneously and equally increased. This allows the laser energy to be distributed more extensively in the upper opening area of ​​the keyhole and the surrounding shallow molten pool. The increased heat input effectively reduces the surface tension and viscosity of the liquid metal in this area, causing the upper opening of the keyhole to exhibit an outward expanding "trumpet-shaped" shape driven by Marangoni convection.

[0048] Synergistic anti-porosity effect: This "trumpet-shaped" physical expansion not only prevents the collapse and closure of the upper liquid metal wall in the keyhole, but also provides an extremely wide and smooth escape channel for the metal vapor and plasma generated at the bottom of the keyhole.

[0049] Furthermore, the specific implementation process of step S4 is as follows: After performing the aforementioned space energy transfer and successfully suppressing splashing and keyhole collapse, the system must precisely determine the timing and method of "exiting intervention." If a step-like power recovery strategy is adopted after the intervention ends, the instantaneous change in energy ratio can easily trigger a "secondary physical oscillation" due to the thermodynamic hysteresis effect of the molten metal in the pool.

[0050] 1. A stringent mechanism for determining physical steady states: In actual high-speed welding, the keyhole may experience a brief period of "pseudo-calm." To prevent premature system exit from intervention and subsequent defect recurrence, the beam modulation execution subsystem in this embodiment not only relies on the prediction probability of the AI ​​network but also introduces a hard temporal constraint on the physical signal. The system determines that "keyhole stabilization" must simultaneously satisfy the following two joint conditions: The current probability of defect occurrence output by the multimodal defect prediction network. The temperature steadily drops to the preset safety threshold. )the following; The intensity of radiation in a specific wavelength band (ultraviolet) collected by the coaxial photoelectric sensor must continuously decrease to the average intensity of the radiation within a set reference time period. The following condition must be met, and this state must last for at least 5 milliseconds.

[0051] It should be noted that the physical meaning of "lasting 5 milliseconds" here is that it covers multiple fluid oscillation cycles of the molten metal in the acoustic frequency range. Only when the radiation intensity no longer shows any sharp fluctuations within this cycle does it indicate that the vaporization backpressure at the bottom of the keyhole and the surface tension of the molten pool have truly re-established a steady-state fluid dynamic equilibrium.

[0052] 2. A first-order exponential smoothing recovery mathematical model to overcome the second oscillation: After confirming that the keyhole has stabilized, the system begins executing the power recovery procedure. To avoid "thermal shock" to the liquid metal from sudden changes in heat input, the main control microprocessor uses a preset first-order exponential decay function to control the power distribution between the central beam and the ring beam. A smooth callback occurs within a set time window. The dynamic time trajectory of its power recovery satisfies the solution of the following differential equation: in, For the relative time variable of the recovery phase, The recovery phase Real-time power commands are sent to the central beam and the ring beam. These represent the initial power of the center beam and the ring beam before trigger modulation (i.e., under normal welding conditions). The absolute power difference actually transferred in step S3 .

[0053] In particular, The decay time constant set for the system. The system is in a modulation state (center low, ring high); as time progresses, the exponential term... Gradually approaching The beam power ratio exhibits a smooth, overshoot-free asymptote, eventually seamlessly returning to the initial normal processing state.

[0054] 3. Fluid dynamic synergistic damping mechanism for smooth recovery: Under the S3 modulation state, due to the increase in annular energy, a strong Marangoni convection diffuses outward from the surface of the molten pool. If the power is switched back to the center at this moment, the sudden increase in temperature gradient in the central region will cause a violent reversal of the fluid flow direction. This physical oscillation, similar to the "water hammer effect" in fluid mechanics, will agitate the already calmed liquid metal again, forming an extremely troublesome secondary spatter, and even leaving volcano-shaped shrinkage cavities at the weld end.

[0055] The natural exponential decay model used in this embodiment has a rate of change that gradually slows down over time, perfectly simulating the natural physical relaxation process of liquid metal under viscous resistance. This allows the temperature gradient inside the molten pool to be released slowly in stages, the intensity of Marangoni convection to decay gently, and the liquid metal to solidify smoothly without drastic fluctuations.

[0056] Example 2: like Figure 2 As shown, a laser welding process control system includes: A multi-source sensing subunit is configured to synchronously acquire a multi-source physical signal sequence during the laser welding process. The multi-source sensing subunit includes at least a coaxial vision sensor, a coaxial OCT sensor, and a coaxial photoelectric sensor, which are used to extract the coaxial vision keyhole morphology features, the coaxial OCT absolute depth features, and the coaxial photoelectric specific band radiation intensity, respectively. The edge computing main control subunit is communicatively connected to the multi-source sensing subunit and has a built-in preset multimodal defect prediction network configured to receive the physical signal sequence. When it is determined that the increase in radiation intensity exceeds a first set threshold, the fusion weight of the keyhole shape feature is reduced, and the fusion weight of the absolute depth feature and the radiation intensity is simultaneously increased to calculate and output the probability of defect occurrence within a set time in the future. The beam modulation execution subunit, communicatively connected to the edge computing main control subunit, includes a dual-path laser with an independently adjustable center beam and a ring beam; configured to, when the probability of the defect occurrence is greater than a second preset threshold, reduce the power of the center beam by a preset ratio and simultaneously increase the power of the ring beam by the same amount while keeping the total power of the laser basically constant, thereby realizing the spatial transfer of the spot energy; and, after monitoring that the keyhole has recovered to stability, use a preset attenuation function to control the center beam and the ring beam to smoothly recover to the initial power ratio within a set time window.

[0057] The above embodiments are only used to illustrate the technical solutions of the present invention, and are not intended to limit it. Although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some of the technical features. Such modifications or substitutions will not cause the essence of the corresponding technical solutions to deviate from the spirit and scope of the technical solutions of the embodiments of the present invention.

Claims

1. A laser welding process control method, characterized in that, Includes the following steps: S1: Simultaneously acquire multi-source physical signal sequences during the laser welding process, including: coaxial visual keyhole morphology features, coaxial OCT absolute depth features, and coaxial photoelectric specific band radiation intensity; S2: Input the physical signal sequence into a preset multimodal defect prediction network; when the increase in radiation intensity exceeds a first set threshold, reduce the fusion weight of the keyhole morphology feature and simultaneously increase the fusion weight of the absolute depth feature and the radiation intensity to output the probability of defect occurrence within a set time period in the future. S3: When the probability of the defect occurring is greater than the second set threshold, under the premise of keeping the total power of the laser basically constant, the power of the center beam is reduced by a preset ratio, and the power of the ring beam is simultaneously increased by an equal amount to realize the spatial transfer of the spot energy. S4: When the physical signal sequence detects that the keyhole has stabilized, a preset attenuation function is used to control the center beam and the ring beam to smoothly recover to the initial power ratio within a set time window.

2. The laser welding process control method according to claim 1, characterized in that, The specific band radiation intensity of the coaxial optoelectronic system includes the first band radiation intensity and the second band radiation intensity collected synchronously. The first band is the ultraviolet band of 350nm~450nm; the second band is the infrared band that includes the center wavelength of the laser output. The steps for obtaining the coaxial visual keyhole morphological features include: A sequence of molten pool images is acquired using a narrowband filter that includes the center wavelength of the laser output in the cutoff band and the first band. Edge extraction is performed on the molten pool image sequence to obtain a geometric feature sequence, which includes the two-dimensional area of ​​the keyhole opening, the aspect ratio, and the curvature gradient of the keyhole back wall edge.

3. The laser welding process control method according to claim 2, characterized in that, The steps for obtaining the absolute depth features of the coaxial OCT include: It emits a probe light coaxially toward the bottom of the keyhole and receives the reflected signal light. The signal light is interfered with the reference light to form an interference spectrum; The interference spectrum is solved in the frequency domain to extract the absolute depth coordinates of the bottom of the keyhole and construct a depth coordinate sequence. The step of synchronously acquiring the multi-source physical signal sequence during the laser welding process specifically includes: Use a field-programmable gate array or a main control microprocessor as a unified clock source; The unified clock source sends microsecond-level synchronous trigger pulses to the coaxial vision sensor, coaxial OCT sensor, and coaxial photoelectric sensor. The collected keyhole morphology features, absolute depth features, and specific band radiation intensity are assigned global timestamps and spliced ​​together to construct a spatiotemporal input tensor.

4. The laser welding process control method according to claim 3, characterized in that, The multimodal defect prediction network includes a spatiotemporal feature extraction module and a dynamic feature fusion module; The step of inputting the physical signal sequence into a preset multimodal defect prediction network includes: The spatiotemporal feature extraction module extracts the keyhole morphology features, the absolute depth features, and the temporal dependence features of the specific band radiation intensity, respectively. The dynamic feature fusion module performs weighted concatenation of each temporally dependent feature according to the real-time allocated fusion weights to generate a multimodal fusion feature vector.

5. The laser welding process control method according to claim 4, characterized in that, The determination condition for the jump in radiation intensity exceeding the first preset threshold is as follows: Extract the real-time derivative value of the radiation intensity of the specific band within a sliding time window of 1 microsecond to 2 microseconds; When the real-time derivative value is greater than 150% of the average radiation intensity within a set reference time period, it is determined that the jump exceeds the first set threshold.

6. The laser welding process control method according to claim 5, characterized in that, The step of reducing the fusion weight of the keyhole morphology feature and simultaneously increasing the fusion weight of the absolute depth feature and the radiation intensity specifically includes: The initial fusion weight of the keyhole morphological features is reduced to 0.1 or below; The fusion weights of the absolute depth feature and the radiation intensity are simultaneously increased, and the sum of the increased fusion weights of the absolute depth feature and the radiation intensity is limited to be greater than or equal to 0.

8. The output of the probability of defect occurrence within a set future time period specifically includes: The multimodal fused feature vector is input into the fully connected classification layer; Output the predicted probability of spatter or non-fusion defects occurring in the molten pool within a time window of 2 to 5 milliseconds.

7. The laser welding process control method according to claim 6, characterized in that, The step of reducing the power of the center beam by a preset ratio and simultaneously increasing the power of the ring beam by the same amount while keeping the total power of the laser basically constant is performed using the following energy space transfer model: in, and These represent the initial power of the center beam and the initial power of the ring beam before triggering modulation, respectively. and These are the modulated target power of the center beam and the target power of the ring beam, respectively. Let be the energy transfer coefficient, and set . During the modulation period, the real-time total power of the laser satisfy: ,and The absolute value of the fluctuation deviation relative to the initial total power is less than or equal to .

8. The laser welding process control method according to claim 7, characterized in that, The timing requirements for the execution of the trigger beam spatial energy transfer modulation are as follows: From the trigger moment when the probability of the defect occurring exceeds the second set threshold, the laser is controlled to adjust the center beam power downward within a hardware response time window of 2 to 5 milliseconds. And the power of the ring beam is adjusted upwards to The action.

9. A laser welding process control method according to claim 8, characterized in that, The criteria for determining that the keyhole has recovered to a stable state as monitored by the physical signal sequence include: The probability of current defect occurrence output by the multimodal defect prediction network decreases below a safe threshold, and the radiation intensity in the specific band remains continuously stable within the average radiation intensity over a set reference time period. The following lasts for at least 5 milliseconds; In the step of controlling the central beam and the annular beam to smoothly recover to their initial power ratio within a set time window, the preset attenuation function is a first-order exponential attenuation function, and its power recovery trajectory satisfies: in, The recovery phase Real-time power of the center beam and real-time power of the ring beam at any given moment. This represents the actual power difference transferred. , The decay time constant is set for the system, and the length of the set time window is 10 milliseconds to 20 milliseconds.

10. A laser welding process control system, comprising the laser welding process control method according to any one of claims 1-9, characterized in that, include: A multi-source sensing subunit is configured to synchronously acquire a multi-source physical signal sequence during the laser welding process. The multi-source sensing subunit includes at least a coaxial vision sensor, a coaxial OCT sensor, and a coaxial photoelectric sensor, which are used to extract the coaxial vision keyhole morphology features, the coaxial OCT absolute depth features, and the coaxial photoelectric specific band radiation intensity, respectively. The edge computing main control subunit is communicatively connected to the multi-source sensing subunit, has a built-in preset multimodal defect prediction network, and is configured to receive the physical signal sequence; When it is determined that the increase in radiation intensity exceeds a first set threshold, the fusion weight of the keyhole shape feature is reduced, and the fusion weight of the absolute depth feature and the radiation intensity is increased simultaneously, so as to calculate and output the probability of defect occurrence within a set time in the future. The beam modulation execution subunit, communicatively connected to the edge computing main control subunit, includes a dual-path laser with an independently adjustable center beam and a ring beam; configured to, when the probability of the defect occurrence is greater than a second preset threshold, reduce the power of the center beam by a preset ratio and simultaneously increase the power of the ring beam by the same amount while keeping the total power of the laser basically constant, thereby realizing the spatial transfer of the spot energy; and, after monitoring that the keyhole has recovered to stability, use a preset attenuation function to control the center beam and the ring beam to smoothly recover to the initial power ratio within a set time window.

Images