A laser processing process control method and system based on multi-physical field coupling

Abstract

The application relates to the technical field of laser processing process control, and discloses a laser processing process control method and system based on multi-physical field coupling, wherein the method obtains a molten pool surface monitoring video stream, extracts a surface feature vector, constructs a time sequence feature sequence, obtains a temperature field and a flow field matrix through space-time coupling mapping, fuses to generate a physical field state matrix, extracts a time sequence gradient dimension reduction to generate a nonlinear evolution track, judges a slight disturbance and decomposes to obtain a disturbance amplitude and a frequency, combines the physical field state matrix to classify and identify defect initiation and locate coordinates, calculates an energy compensation value to generate a power regulation instruction to adjust laser output power. The method can realize real-time deduction of molten pool heat flow coupling state and early prediction and regulation of defects, effectively improves defect inhibition capability and process quality stability of laser processing, and meets high-precision manufacturing requirements of an intelligent welding system.

Description

Technical Field

[0001] This invention relates to the field of laser processing control technology, and in particular to a laser processing control method and system based on multi-physics coupling. Background Technology

[0002] Laser processing technology is a core process in high-end manufacturing, playing a key role in scenarios such as welding aerospace components and processing precision medical devices. As intelligent welding systems continue to increase their requirements for processing accuracy and consistency, higher demands are being placed on the ability to control the processing in real time.

[0003] Current laser processing mainly relies on open-loop control schemes with preset empirical parameters. Some optimized schemes only calculate the temperature and flow field distribution of the molten pool through offline numerical simulation or perform simple post-processing feedback adjustments based on surface monitoring signals. Because offline numerical simulation takes a long time per calculation, it cannot meet the millisecond-level real-time response requirements of the processing process. Furthermore, surface monitoring signals cannot reflect the true evolution of thermal-fluid coupling within the molten pool. Existing schemes struggle to identify anomalies in a timely manner when minor disturbances occur and to predict and correct process parameters before defects develop. Ultimately, this results in insufficient defect suppression capability in laser processing, and the process quality stability cannot meet the requirements of high-precision manufacturing.

[0004] Therefore, existing technologies have limitations in achieving real-time simulation of the thermal flux coupling state of the molten pool and early prediction and control of defects, resulting in poor defect suppression capability and process quality stability in laser processing, making it difficult to meet the high precision requirements of laser processing. Summary of the Invention

[0005] This invention provides a laser processing control method and system based on multi-physics coupling to solve the problem that existing laser processing solutions cannot sense the multi-physics coupling state inside the molten pool in real time, and it is difficult to predict and control the defect initiation stage in a timely manner, thus failing to meet the high precision requirements of laser processing.

[0006] In a first aspect, to solve the above-mentioned technical problems, the present invention provides a laser processing control method based on multi-physics coupling, comprising: Acquire surface monitoring video streams during laser processing and extract surface feature vectors of the molten pool; Arrange the surface feature vectors of the molten pool in chronological order for multiple consecutive frames to construct a temporal feature sequence. Perform spatiotemporal coupling feature mapping on the temporal feature sequence to obtain the internal temperature field matrix and flow field matrix of the molten pool. Align and fuse the temperature field matrix and flow field matrix according to spatial coordinates to obtain the physical field state matrix. Extract the temporal variation gradient tensor of the physical field state matrix, and generate a multi-physics nonlinear evolution trajectory through manifold dimensionality reduction. Calculate the spatial deviation between the multi-physics nonlinear evolution trajectory and the preset trajectory baseline. Determine the perturbation segment trajectory data based on the spatial deviation, and perform time-frequency decomposition to obtain the corresponding perturbation amplitude and perturbation frequency. By combining the disturbance amplitude, the disturbance frequency, and the physical field state matrix, molten pool fluctuations are classified to obtain molten pool fluctuation classification results, thereby identifying defect initiation events and determining their corresponding spatial coordinates. The energy compensation value of the target area is calculated based on the spatial coordinates, and a power adjustment command is generated to adjust the output power of the laser device.

[0007] Secondly, the present invention provides a laser processing control system based on multi-physics coupling, comprising: a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor executes the program to implement the method described above.

[0008] Thirdly, the present invention provides a computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, implements the method described above.

[0009] Compared with the prior art, the present invention has the following beneficial effects: (1) This invention extracts the surface feature vector of the molten pool by collecting monitoring video streams of the molten pool surface, constructs a temporal feature sequence and performs spatiotemporal coupling mapping to obtain a physical field state matrix that aligns and fuses the temperature field and the flow field. This scheme maps the surface visual morphological features to the internal thermal-fluid coupling state, replacing the time-consuming offline numerical simulation, and can obtain the real-time distribution of multiple physical fields inside the molten pool at the millisecond time scale, providing accurate internal state basis for subsequent disturbance identification and parameter control.

[0010] (2) This invention extracts the temporal variation gradient of the physical field state matrix, generates a multi-physics nonlinear evolution trajectory through manifold dimensionality reduction, compares it with a preset trajectory baseline to determine minor perturbations, and combines random forest to complete the classification of melt pool fluctuations to identify defect initiation events. This scheme quantifies the degree of perturbation by trajectory spatial deviation, and determines the fluctuation category by combining perturbation characteristics and physical field distribution. It can capture the precursors of defect initiation before the defect is fully formed, realize the early prediction of abnormal states, and avoid the quality loss caused by defect expansion.

[0011] (3) This invention calculates the energy compensation value of the target area by locating the spatial coordinates of the defect and combining the material's thermal properties, and generates a power adjustment command to adjust the laser output power in real time. This scheme accurately calculates the required compensation heat based on the defect volume and melting depth, and performs directional energy replenishment by matching the processing position through pulse width modulation. It can repair small defects in the initiation stage, forming a complete closed loop from state perception to process control, effectively improving the defect suppression capability and process stability of laser processing. Attached Figure Description

[0012] Figure 1 This is a schematic diagram of a laser processing control method based on multi-physics coupling provided by the present invention; Figure 2 This is a schematic diagram of the internal temperature field matrix of the molten pool provided by the present invention; Figure 3 This is a schematic diagram of the internal flow field matrix of the molten pool provided by the present invention; Figure 4 This is a schematic diagram of the temporally varying gradient tensor provided by the present invention; Figure 5 This is a schematic diagram of the nonlinear evolution trajectory of the multiphysics field provided by the present invention. Detailed Implementation

[0013] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. 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 are within the scope of protection of the present invention.

[0014] Reference Figure 1 The first embodiment of the present invention provides a laser processing control method based on multi-physics coupling, comprising the following steps: S11, acquire the surface monitoring video stream during laser processing, and extract the surface feature vector of the molten pool; S12, arrange the surface feature vectors of the molten pool in a continuous time sequence to construct a temporal feature sequence, perform spatiotemporal coupling feature mapping on the temporal feature sequence to obtain the internal temperature field matrix and flow field matrix of the molten pool, and fuse the temperature field matrix and the flow field matrix according to spatial coordinates to obtain the physical field state matrix. S13, extract the temporal variation gradient tensor of the physical field state matrix, generate a multi-physics nonlinear evolution trajectory through manifold dimensionality reduction, calculate the spatial deviation between the multi-physics nonlinear evolution trajectory and the preset trajectory baseline, determine the perturbation segment trajectory data based on the spatial deviation, and perform time-frequency decomposition to obtain the corresponding perturbation amplitude and perturbation frequency. S14, Combine the disturbance amplitude, the disturbance frequency and the physical field state matrix to classify the molten pool fluctuations, obtain the molten pool fluctuation classification results, identify the defect initiation event and determine the corresponding spatial coordinates; S15, calculate the energy compensation value of the target area based on the spatial coordinates, and generate a power adjustment command to adjust the output power of the laser device.

[0015] In step S11, acquiring the surface monitoring video stream during laser processing and extracting the surface feature vector of the molten pool includes: The surface monitoring video stream of the laser processing molten pool area is acquired and split into single-frame surface images according to the time series. Extract the molten pool edge from the single-frame surface image to obtain the molten pool contour matrix; Deep spatial feature extraction is performed on the molten pool contour matrix to obtain the molten pool surface feature vector.

[0016] It should be noted that the surface monitoring video stream is acquired using a high-speed industrial camera, mounted coaxially on the laser processing head. A narrow-band filter with a center wavelength matching the processing laser band is installed at the front end, and the sampling frequency is set to 500 frames per second. The narrow-band filter filters out plasma arc light and broad-spectrum stray light during processing, ensuring uniform grayscale distribution in the molten pool area of ​​the acquired image and preventing local overexposure, thus providing high-quality input for subsequent edge extraction. The sampling frequency of 500 frames per second is determined based on the Nyquist sampling theorem. The typical self-excited oscillation frequency of the laser welding molten pool is approximately 200 Hz. A sampling rate of 500 frames per second corresponds to a sampling interval of 2 milliseconds, which can completely cover all effective fluctuation frequencies without signal aliasing, ensuring the capture of all transient changes in the molten pool morphology. The acquired video stream is split frame by frame in chronological order, with each frame image bound to a unique timestamp to ensure the accuracy of subsequent time-series analysis. Each frame of the surface image is in grayscale format, with the brightness value of each pixel corresponding to the light radiation intensity of the processing area, intuitively reflecting the outer boundary and internal brightness distribution of the molten pool.

[0017] The resolution of each single-frame image is uniformly set to 256 by 256 pixels. This resolution can fully cover the molten pool and its surrounding heat-affected zone, ensuring that the main body of the molten pool is always within the effective centered area of ​​the image, without increasing the processing latency of a single frame due to an excessive number of pixels. Actual testing shows that the feature extraction time per frame at this resolution is controlled within 1 millisecond, which can match a real-time processing pace of 500 frames per second and meet the response requirements of millisecond-level control.

[0018] When extracting the molten pool edge, a 5×5 Gaussian smoothing filter is first applied to the single-frame surface image to remove random noise. Then, the magnitude and direction of the brightness gradient are calculated pixel by pixel. Continuous edge segments are retained using Canny dual-threshold filtering. All pixels inside the closed edge are assigned a value of 1, while background pixels outside the edge are assigned a value of 0. Finally, a two-dimensional molten pool contour matrix is ​​output. The number of rows and columns of the matrix is ​​consistent with the resolution of the original image, fully representing the two-dimensional coverage and boundary shape of the molten pool region. The dual thresholds include a high threshold and a low threshold, with the high threshold set to 150 and the low threshold set to 50. The two thresholds are determined based on the statistical difference in grayscale between the molten pool and the background in the low-carbon steel welding scenario. During normal welding, the grayscale of the core area of ​​the molten pool is generally higher than 200, while the grayscale of the background area is generally lower than 80. The high threshold is used to lock the main edge with strong brightness, and the low threshold is used to supplement and connect edge segments with weak brightness. The high and low thresholds are kept in a 3:1 ratio, which can effectively filter out scattered false edge interference such as spatter spots and retain the complete closed physical contour of the molten pool.

[0019] Deep spatial feature extraction is achieved through a pre-trained multi-layer feature extraction network. The network adopts a residual convolutional structure and consists of an input layer, four feature extraction stages, and a global pooling output layer. The input layer receives a 256x256 molten pool contour matrix as input. The first stage contains one convolutional layer and one max-pooling layer. The convolutional layer uses a 3×3 kernel with a stride of 1, a padding width of 1, and 64 output channels. After the convolution operation, a linear rectified function is applied to perform non-linear activation. The max-pooling layer uses a 2×2 pooling window with a stride of 2 to downsample the feature map and compress its spatial size. The second to fourth stages each contain two residual blocks, each consisting of two convolutional layers and a shortcut connection. The number of output channels in the three stages are 128, 256, and 512, respectively. Each stage also ends with a 2×2 max-pooling layer for downsampling. The shortcut connection in the residual structure avoids the gradient degradation problem of deep networks, ensuring stable training even as the number of layers increases. After four stages of feature extraction, the spatial size of the feature map is compressed to 16×16, and the number of channels is increased to 512. Finally, a global average pooling layer is applied to average the spatial positions of each channel, resulting in a one-dimensional molten pool surface feature vector of length 512.

[0020] This network employs a regression task training approach to optimize parameters. The tens of thousands of molten pool image samples used for training are all labeled with molten pool physical field parameters, derived from high-precision multiphysics simulation results under the same process conditions. These samples include three core values: molten pool center temperature, maximum flow rate, and melt depth, serving as supervision for the physical field parameter regression task. During training, the mean squared error loss of the regression task is used as the overall loss function, and all network parameters are iteratively optimized through backpropagation. After training, the regression output header at the end of the network is removed, retaining only the initial feature extraction portion for deployment. At this point, the 512-dimensional vector output by the network simultaneously integrates morphological semantics and physical correlation information, providing high-quality feature input for subsequent internal physical field deduction.

[0021] For example, in the laser welding of low-carbon steel thin plates, a single-frame surface image with a resolution of 256x256 is acquired. After smoothing and gradient calculation, a molten pool contour matrix of the same resolution is extracted. The contour exhibits a typical keyhole welding tail shape. The contour matrix is ​​input into a feature extraction network, and after multi-layer transformation, a molten pool surface feature vector of length 512 is output. The vector contains multi-dimensional morphological information such as the aspect ratio of the molten pool, the degree of tail drag, and the edge smoothness, providing a complete surface morphology input for subsequent internal physical field deduction.

[0022] In step S12, the temporal feature sequence is constructed using the single-frame molten pool surface feature vector as the basic unit. The feature vectors from 20 consecutive sampling moments are arranged chronologically to form a two-dimensional sequence of 20 x 512. The 20-frame time window corresponds to a 40-millisecond time span, determined based on the typical oscillation period of the molten pool. The self-excited oscillation frequency of the laser welding molten pool is approximately 20 to 50 Hz, corresponding to a period of 20 to 50 milliseconds. The 40-millisecond window can completely cover a full oscillation cycle, ensuring that the temporal features can capture the complete evolutionary pattern without introducing excessive processing delay due to an excessively long window, thus meeting the response requirements of real-time control in the processing.

[0023] In one implementation, the time-series feature sequence is subjected to spatiotemporal coupling feature mapping to obtain the temperature field matrix and flow field matrix inside the molten pool. The temperature field matrix and the flow field matrix are then aligned and fused according to spatial coordinates to obtain the physical field state matrix, including: Extract the long and short-term time-series features of the time-series feature sequence to obtain a spatiotemporal coupled feature tensor containing short-term fluctuation features and long-term trend features; The spatiotemporal coupling feature tensor is nonlinearly spatially mapped to output the temperature field matrix and the flow field matrix, respectively. The temperature field matrix and the flow field matrix are numerically fused according to their corresponding spatial nodes to obtain the coupled physical field state matrix.

[0024] It should be noted that when extracting short- and long-term temporal features, a recursive state calculation is performed on the temporal feature sequence, simultaneously preserving feature information from both dimensions. The features from the most recent 3 frames are taken as short-term fluctuation features, corresponding to drastic morphological changes within the last 6 milliseconds; the moving average of all 20 frames is taken as long-term trend features, corresponding to a gradual evolution within a 40-millisecond range. Both types of features are one-dimensional vectors of length 512. After concatenation and integration along the feature dimensions, a spatiotemporal coupled feature tensor of length 1024 is obtained. This tensor simultaneously integrates morphological information from the spatial dimension and evolutionary patterns from the temporal dimension, serving as the fundamental input for subsequent deduction of the internal physical field state.

[0025] Nonlinear spatial mapping is achieved through a multi-layer nonlinear transformation network. The network employs an encoder-decoder structure, with the input being a spatiotemporally coupled feature tensor of length 1024. The encoder consists of two fully connected layers: the first layer expands the 1024-dimensional features to 2048 dimensions, and the second layer expands them to 4096 dimensions. Each layer is followed by a linear rectified activation function to complete the nonlinear transformation. The decoder comprises three upsampling modules, each consisting of a deconvolutional layer and a batch normalization layer, progressively reshaping the one-dimensional features into a two-dimensional spatial feature map. The first upsampling stage outputs a 16x16 feature map, the second stage outputs a 32x32 feature map, and the third stage outputs a 64x64 feature map. Finally, two independent output branches are connected, outputting a 64x64 temperature field matrix and a flow field matrix, respectively.

[0026] The network uses high-precision multiphysics simulation results as training labels, and the training samples are generated using multiphysics coupling simulation software. The simulation model uses a hexahedral structured mesh to divide the molten pool region, with a basic mesh size of 0.05 mm and a mesh size of 0.02 mm in the central region of the molten pool. The time step is set to 0.1 ms. The wall boundaries are set as convective heat transfer boundaries with a heat transfer coefficient of 200 W / m² / Kelvin. The top free surface considers thermal radiation boundaries with a surface emissivity of 0.4. The volumetric fluid dynamics method is used to track the gas-liquid interface, and the surface tension temperature coefficient is -4 × 10⁻⁴. -4The simulation model uses N / (m·K) and employs the fluid volume method to trace the gas-liquid interface. The governing equations of the simulation model include the continuity equation, the Navier-Stokes momentum equation, and the energy equation. The momentum equation considers the buoyancy-driven term under the Boussinesq approximation and the Marangoni effect driven by the surface tension gradient. The energy equation includes a Gaussian heat source term, fully reproducing the thermal-fluid coupling physical processes within the molten pool. The process parameter space uses Latin hypercube sampling, covering typical values ​​for laser power from 1000 to 4000 watts, welding speed from 0.5 to 3.0 meters per minute, plate thickness from 1 to 5 millimeters, shielding gas flow rate from 10 to 25 liters per minute, and defocusing amount from -2 to 2 millimeters. At least 5000 paired samples are generated, each containing surface morphology features and corresponding internal temperature and flow field data.

[0027] During training, the sum of the mean squared errors of the temperature field and the flow field was used as the loss function. The Adam optimizer was employed for iterative optimization, with an initial learning rate of 0.001, a batch size of 32, and a total of 200 training epochs. Early stopping was triggered when the validation set loss no longer decreased for 10 consecutive epochs. The loss weights for the temperature field and the flow field were set to 0.6 and 0.4, respectively. All network parameters were iteratively optimized through backpropagation. After training, the inference accuracy was evaluated using at least 500 independent validation samples. The average relative error point-to-point for the temperature field did not exceed 15%, and the deviation of the highest temperature point did not exceed 0.2 mm; the average relative error point-to-point for the flow field did not exceed 20%, and the deviation of the maximum flow velocity point did not exceed 0.3 mm.

[0028] In the inference phase, the Monte Carlo dropout method is used to quantify uncertainty. A dropout layer with a dropout probability of 0.1 is set in the fully connected layers of the network. The dropout layer remains active during inference, and 20 random forward propagations are performed on the same sample. The mean of the outputs of each node is taken as the inferred value, and the standard deviation is taken as the uncertainty. The output is an estimation matrix containing the inferred value and the uncertainty. Uncertainty reflects the model's confidence level in the inference results. The physical process in the keyhole region at the center of the molten pool is complex, resulting in relatively high uncertainty; the edge heat-affected zone is stable, resulting in relatively low uncertainty. In subsequent perturbation judgment and defect identification stages, uncertainty can be used as a confidence weight to reduce the confidence level of the judgment results for high-uncertainty regions.

[0029] It's worth noting that both matrices are 64 rows and 64 columns, corresponding to uniformly divided spatial grid nodes within the molten pool. At this grid granularity, a single grid corresponds to an actual physical size of approximately 0.1 millimeters, capable of resolving details of temperature gradients and flow velocity distribution within the molten pool. Simultaneously, the total number of grids is controlled at 4096 to ensure real-time performance of subsequent matrix calculations. The value of each element in the temperature field matrix represents the absolute temperature of the corresponding spatial node, in degrees Celsius, covering the high-temperature range above the material's melting point. The value of each element in the flow field matrix represents the flow velocity of the liquid metal at the corresponding node, in meters per second.

[0030] The two matrices are aligned according to spatial coordinates and then merged at the feature level. The origin of the coordinate system is the geometric center of the molten pool, and a single grid corresponds to a physical size of 0.1 mm. The row and column indices of the two matrices correspond one-to-one to the same spatial position. Before merging, the two types of physical quantities are first normalized to dimensionless.

[0031] The temperature field matrix is ​​normalized using the material's liquidus temperature as a reference, with ambient temperature as the lower limit and liquidus temperature as the upper limit, linearly mapping the absolute temperature values ​​to the range of 0 to 1. For low-carbon steel, the liquidus temperature is taken as 1538 degrees Celsius, and the ambient temperature as 25 degrees Celsius. After normalization, the value in the core region of the molten pool is close to 1, while the value in the edge heat-affected zone gradually decays to near 0, eliminating the influence of the absolute temperature dimension.

[0032] The flow field matrix is ​​normalized using the maximum flow velocity in the molten pool at the current moment as a reference. The flow velocity value of each node is divided by the maximum flow velocity and mapped to the interval between 0 and 1, with a value of 1 corresponding to the position with the fastest flow velocity. The maximum flow velocity changes dynamically with the processing conditions, and is approximately 0.5 to 2.0 meters per second under typical low-carbon steel laser welding conditions. The normalization process eliminates the influence of the absolute dimension of the flow velocity.

[0033] After normalization, the temperature field and flow field are concatenated along the channel dimension to obtain a multi-channel physical field state matrix with dimensions of 64 rows and 64 columns and 2 channels. The first channel carries the normalized temperature distribution information, and the second channel carries the normalized flow field distribution information. This fusion method belongs to the feature-level fusion at the data level, aiming to integrate the two types of physical field information into a unified data structure, facilitating subsequent unified gradient calculation and evolution analysis. The true multi-physics coupling relationship is already implicit in the training data. The simulation process has rigorously solved the coupled partial differential equations, and the mapping network learns the mapping relationship from surface morphology to internal coupling state.

[0034] For example, building upon the single-frame features of the aforementioned low-carbon steel welding, a temporal feature sequence is constructed from the surface feature vectors of the molten pool across 20 consecutive frames. After temporal feature extraction and nonlinear mapping, a 64x64 temperature field matrix and a flow field matrix are output, respectively. The temperature field matrix shows the highest values ​​in the central region, gradually decreasing towards the edges to near the melting point. The flow field matrix shows the fastest flow velocity in the center, gradually decreasing towards the edges, perfectly conforming to the physical laws of thermal convection in the molten pool. The two matrices are then concatenated along the channel dimension to obtain a 64x64, two-channel multi-channel physical field matrix.

[0035] Reference Figure 2 and Figure 3 This invention demonstrates the precise perception and quantitative deduction of the physical states of the internal temperature and flow fields of a molten pool in a real-time laser processing control scenario. Through nonlinear spatial mapping, this invention successfully transforms surface video features into internal thermal distribution and liquid metal flow velocity distribution. Figure 2 The bright yellow-white high-temperature distribution in the central area and the dark red low-temperature transition in the outer area clearly demonstrate the temperature gradient pattern of actual laser deep penetration welding. Figure 3 The highlighted purple area in the center precisely indicates the location of the highest flow velocity, while the surrounding cyan and light blue transition clearly reflects the attenuation of the fluid from the center to the edge. This intuitive visualization makes the previously difficult-to-obtain internal thermodynamic state and invisible thermal convection phenomena clearly visible, thus achieving a complete characterization of the thermal-fluid coupling state inside the molten pool. Ultimately, it provides essential physical field coordination for subsequent steps to calculate energy compensation values ​​and perform precise power control.

[0036] In step S13, the extraction of the temporal gradient tensor of the physical field state matrix, the generation of a multi-physics nonlinear evolution trajectory through manifold dimensionality reduction, the calculation of the spatial deviation between the multi-physics nonlinear evolution trajectory and a preset trajectory baseline, the determination of the perturbation segment trajectory data based on the spatial deviation, and the performance of time-frequency decomposition to obtain the corresponding perturbation amplitude and frequency include: Two sets of physical field state matrices at adjacent sampling times are obtained, and the difference between the corresponding matrix elements is calculated to obtain the temporal gradient tensor. The phase space of the temporally varying gradient tensor is reconstructed to obtain a high-dimensional phase space feature sequence; The high-dimensional phase space feature sequence is subjected to manifold dimensionality reduction and mapped to a low-dimensional feature space; Based on the state transition sequence in the low-dimensional feature space, a multi-physics nonlinear evolution trajectory is generated. The vertical distance from each sampling point on the nonlinear evolution trajectory of the multiphysics field to the preset trajectory baseline is calculated point by point to obtain the spatial deviation of each sampling point; If the spatial deviation of a consecutive preset number of sampling points is greater than a preset deviation threshold, it is determined that there is a slight disturbance in the laser processing process; Empirical mode decomposition was performed on the deviation segment trajectory data to obtain multiple sets of intrinsic mode function components; The intrinsic mode function components are subjected to spectral transformation to extract the disturbance amplitude and disturbance frequency.

[0037] It should be noted that the temporally varying gradient tensor is calculated based on two sets of physical field state matrices at adjacent sampling times. The difference between the current and previous time-series matrices is calculated point-by-point to obtain the numerical change at each spatial node, which is then combined to form the temporally varying gradient tensor. The time interval for gradient calculation is consistent with the video sampling interval, which is 2 milliseconds. This ensures that the temporal resolution of the gradient calculation matches the accuracy of the original data acquisition, preventing additional errors introduced by interpolation. The temporally varying gradient tensor simultaneously contains information on the rate of change in the horizontal, vertical, and time axes, fully reflecting the spatiotemporal change rate of the thermal flux coupling state within the molten pool.

[0038] First, the temporally varying gradient tensor undergoes dimensionality reduction preprocessing. The two-channel gradient tensor at each time step is flattened into a one-dimensional vector in row-major order. Orthogonal projection is then performed using a pre-trained principal component transformation matrix, retaining principal components with a cumulative variance contribution rate of at least 95%, resulting in the dimensionality-reduced one-dimensional gradient feature time series. Dimensionality reduction eliminates redundant noise dimensions while preserving the core variation patterns, significantly improving the computational efficiency and stability of subsequent phase space reconstruction.

[0039] Next, phase space reconstruction is performed on the dimensionality-reduced one-dimensional gradient feature time series. Reconstruction is achieved using delayed embedding. Feature values ​​at different times are sequentially extracted from the one-dimensional sequence according to a preset embedding dimension and delay time, and then concatenated into a high-dimensional vector. Specifically, for time t, feature values ​​from times t, t+τ, t+2τ, up to t+m-1τ are taken to form an m-dimensional vector, where τ is the delay time and m is the embedding dimension. Each time point corresponds to an m-dimensional state point, and all state points are arranged in chronological order to obtain the high-dimensional phase space feature sequence. The embedding dimension is adaptively determined using a pseudo-nearest neighbor method. The dimension is gradually increased from a low dimension, and the nearest neighbor of each state point is found in each dimension. If the distance increase of the nearest neighbor after increasing the dimension by one exceeds a preset threshold of 15, it is considered a pseudo-nearest neighbor. When the proportion of pseudo-nearest neighbors drops below 5%, the current dimension is taken as the optimal embedding dimension, typically ranging from 7 to 10. The delay time is determined based on the time when the autocorrelation function of the gradient feature sequence first crosses zero, corresponding to twice the sampling interval. This ensures that each coordinate component in the reconstructed phase space is independent of the others, avoiding information redundancy.

[0040] Subsequently, manifold dimensionality reduction first constructs a k-nearest neighbor graph of high-dimensional feature points, with the number of neighbors k set to 12. Dijkstra's algorithm is used to find the shortest path length between any two points as the geodesic distance, constructing a geodesic distance matrix. Then, metric-based multidimensional scaling maps the distance matrix to a three-dimensional low-dimensional feature space. The core of multidimensional scaling is to maintain the relative distance relationships between points in the high-dimensional space, solving for the three-dimensional coordinates corresponding to each state point, minimizing the error between the Euclidean distance in the low-dimensional space and the geodesic distance in the high-dimensional space. The three-dimensional coordinates of each state point are connected sequentially in chronological order to generate a multiphysics nonlinear evolution trajectory. Choosing three-dimensional space as the dimensionality reduction target not only preserves the main manifold structure of the system's evolution without losing the core dynamic characteristics, but also facilitates intuitive calculation of the spatial distance and curvature of the trajectory, balancing analytical accuracy and computational efficiency.

[0041] It is worth noting that the preset trajectory baseline is derived from historical statistical data of normal and stable processing. The evolution trajectories of thousands of stable processing samples are selected, and time alignment is performed with the molten pool formation time as the zero point. A smooth baseline curve is obtained using cubic B-spline fitting, representing the ideal evolution path of the molten pool state under standard process parameters. The Euclidean distance from each sampling point on the evolution trajectory to the nearest point on the baseline is calculated point by point. All discrete nodes on the baseline are traversed, and the minimum Euclidean distance from each sampling point to each node is taken as the spatial deviation of that sampling point. The deviation is a dimensionless normalized distance in the feature space after dimensionality reduction, and does not correspond to a unit of length in physical space. The deviation threshold is determined adaptively. Under standard operating conditions, the initial threshold is calibrated based on the three-standard-deviation criterion. The mean and standard deviation of the deviation of one thousand stable samples are statistically analyzed, and the mean plus three standard deviations is taken as the initial threshold.

[0042] When the spatial deviation of 10 consecutive sampling points is greater than the deviation threshold, a minor disturbance is determined to exist in the laser processing. The duration of 20 milliseconds corresponding to 10 consecutive sampling points can eliminate instantaneous misjudgments caused by single random noise or a single sputtering light spot, ensuring the reliability of disturbance identification.

[0043] After determining the presence of a disturbance, the trajectory data of the deviation segment is adaptively decomposed to obtain multiple sets of intrinsic mode function (EMF) components in different frequency bands. The decomposition is achieved using an empirical mode selection method. First, all local maxima and minima of the deviation sequence are identified, and upper and lower envelopes are fitted to obtain them respectively. The mean values ​​of the upper and lower envelopes are calculated to obtain mean curves. The mean curve is subtracted from the original sequence to obtain a selected component. This selection process is repeated until the component meets the stopping conditions of the number of extreme points and zero-crossing points differing by no more than one, and the upper and lower envelopes being symmetric about the time axis. This yields an EMF component. After separating all components, the remaining monotonic sequence is the trend term. The components are sorted in descending order of frequency. The energy of the first two high-frequency components accounts for more than 85% of the total fluctuation energy and carries the main fluctuation characteristics of the disturbance. The instantaneous envelope of the two components is calculated, and the mean of the envelope is taken as the disturbance amplitude. The first two high-frequency components are superimposed and then a spectrum transformation is performed. Using a sampling frequency of 500 Hz as the reference, a Fourier transform is performed on a sequence of length 50 to obtain a spectrum distribution sequence. The frequency point with the most concentrated energy is selected as the disturbance frequency.

[0044] For example, following the aforementioned physical field state matrix, the temporal gradient tensor is obtained by point-by-point subtraction of the matrices at two adjacent sampling times. Principal component analysis (PCA) is used to reduce the dimensionality to obtain a one-dimensional gradient feature sequence. The optimal embedding dimension is determined to be 8 using the pseudo-nearest neighbor method, with a delay time of 3 sampling intervals. Delayed embedding reconstruction yields an 8-dimensional phase space feature sequence. The geodesic distance between state points is calculated to construct a distance matrix, which is then mapped to three-dimensional space using multi-dimensional scaling. Connecting the state points yields the multi-physics nonlinear evolution trajectory. The spatial deviation of each sampling point is calculated point-by-point. If the deviation of 12 consecutive sampling points exceeds the adaptive deviation threshold under the current operating condition, a minor disturbance caused by protective airflow fluctuations is identified. After decomposing the deviation segment data, the first two sets of high-frequency components are used to calculate the disturbance amplitude as 0.042. Spectral transformation of the low-frequency components yields a disturbance frequency of 150 Hz, consistent with typical airflow disturbance characteristics.

[0045] Reference Figure 4This invention demonstrates the precise extraction and quantification of the temporal gradient of the physical field within the molten pool during laser processing control. By calculating the difference between the physical field state matrices at adjacent sampling times, this invention constructs a temporal gradient tensor reflecting the dynamic changes in the thermodynamic state. In the image, dark blue and dark red patches correspond to regions of drastic temperature increases and decreases, respectively, while lighter-colored areas represent relatively stable thermodynamic states. This intuitive visualization makes the thermodynamic transient fluctuations, originally hidden in a continuous image, explicit, clearly revealing the minute disturbances of energy imbalance within the molten pool at the current moment. This provides precise dynamic gradient characteristics for reconstructing the high-dimensional phase space feature sequence and determining spatial deviation in subsequent steps, thereby achieving accurate capture of the thermodynamic transient changes during processing. Ultimately, it provides core data reflecting dynamic trends for subsequent defect initiation determination.

[0046] Reference Figure 5 This invention demonstrates the intuitive mapping process of the nonlinear evolution trajectory of multi-physics fields in a low-dimensional feature space under real-time control of laser processing. By reconstructing the phase space and reducing the dimensionality of the manifold of the temporally varying gradient tensor, this invention successfully maps complex high-dimensional physical field state sequences into clear three-dimensional spatial trajectories. The green starting point, blue broken line path, and red ending point in the figure together constitute the dynamic evolution path of the molten pool system state over time. This intuitive visualization makes the previously abstract and difficult-to-understand multi-physics field coupling evolution trend clearly visible. It provides accurate spatial morphological basis for subsequent steps to calculate the spatial deviation between the evolution trajectory and the preset trajectory baseline, and to determine the existence of minor disturbances. This enables the quantification of the disturbance degree through trajectory spatial deviation, the capture of precursors before complete defect formation, and ultimately provides precise geometric basis for judging abnormal state evolution in subsequent defect initiation identification and energy compensation control.

[0047] In step S14, the process of classifying molten pool fluctuations by combining the disturbance amplitude, the disturbance frequency, and the physical field state matrix to obtain molten pool fluctuation classification results, identifying defect initiation events, and determining their corresponding spatial coordinates includes: The disturbance amplitude and the disturbance frequency are dimensionally expanded to obtain the extended disturbance characteristics; The extended perturbation features are concatenated and fused with the physical field state matrix to obtain a fused feature matrix; The fused feature matrix is ​​then subjected to dimensionality reduction and noise reduction to obtain a dimensionality-reduced feature vector; The reduced feature vectors are classified and identified to obtain the classification results of molten pool fluctuations; Calculate the matching degree between the molten pool fluctuation classification results and the preset anomaly categories; If the matching degree is greater than the preset anomaly judgment threshold, a defect initiation event is determined to have occurred, and the corresponding defect initiation timestamp is recorded. Based on the defect spawn timestamp, the corresponding motion trajectory data is matched to obtain the spatial coordinates of the defect.

[0048] It should be noted that, firstly, the perturbation amplitude and frequency are dimensionally expanded to obtain expanded perturbation features. Dimensional expansion transforms the two global scalar features into standardized feature terms that can be concatenated with the physical field features. Each perturbation parameter corresponds to a one-dimensional feature value. The feature dimensions remain consistent after expansion, facilitating subsequent fusion processing.

[0049] Next, the extended perturbation features are concatenated and fused with the physical field matrix to obtain a fused feature matrix. Before concatenation and fusion, global statistical features are extracted from the dual-channel physical field matrix. The temperature field and flow field channels are processed separately, and the mean, standard deviation, mean gradient magnitude, and variance of gradient magnitude are calculated for each channel, resulting in a total of eight statistical features. The mean reflects the overall level of the molten pool physical quantities, the standard deviation reflects the degree of distribution dispersion, and the gradient correlation statistics reflect the severity of spatial changes. These multi-dimensional statistics can comprehensively characterize the overall distribution state of the molten pool. Then, the extended perturbation features and statistical features are concatenated along the feature dimensions to form a one-dimensional fused feature matrix. The method of fusion using global statistical features not only preserves the core laws of the physical field distribution but also avoids the information redundancy caused by uniformly copying global scalars to all spatial nodes, significantly compressing the feature volume and improving the efficiency of subsequent classification.

[0050] Then, dimensionality reduction and denoising are achieved through orthogonal transformation, divided into two stages: offline pre-construction and online projection. In the offline stage, a training set is constructed using no fewer than 1000 sets of fused feature samples from normal and abnormal processes. Normal samples are taken from continuous welding processes with stable process parameters and qualified weld formation, corresponding to stable molten pool fluctuations. Abnormal samples include five typical abnormal conditions: covering stable states, slight oscillations, severe spatter, precursors to porosity initiation, and precursors to crack initiation. The principal component transformation matrix is ​​obtained through eigenvalue decomposition of the covariance matrix, and the number of principal components to be retained is determined by using a cumulative variance contribution rate of 95%. During online processing, the pre-stored principal component transformation matrix is ​​directly used to perform matrix multiplication projection on the fused feature matrix to obtain the dimensionality-reduced feature vector. Dimensionality reduction further eliminates random noise interference, improving feature discriminability and classification speed.

[0051] Subsequently, fusion pool fluctuation classification was achieved using an ensemble classification model with a random forest structure containing 100 independent decision trees. During model training, labeled historical processing datasets were used. Each decision tree was trained by randomly selecting a subset of samples and features. The training data and splitting rules for each tree were independent, reducing the bias of single-judgment logic. The classification covered five typical fusion pool fluctuation morphologies: stable state, slight oscillation, violent splashing, precursors to porosity formation, and precursors to crack formation. During online inference, the dimensionality-reduced feature vectors were synchronously input into all decision trees. Each tree independently output its category prediction result. The results of all trees were aggregated through a majority voting mechanism to obtain the final fusion pool fluctuation classification result.

[0052] Next, defect initiation is determined using a two-level mechanism combining qualitative classification and quantitative matching, based on a pre-established abnormal fluctuation category library. The category library corresponds one-to-one with the categories of melt pool fluctuations, and each category of abnormal fluctuation corresponds to a standard fluctuation feature template. The template is derived from the typical dimensionality-reduced feature vector of the corresponding category when the defect is in the critical initiation state, calculated by taking the average feature values ​​of hundreds of defect initiation samples of the same category. The template dimension is consistent with the dimensionality-reduced feature vector.

[0053] Molten pool fluctuation classification is used to determine the category of the current fluctuation, providing a qualitative judgment of the anomaly type. It only indicates that the current fluctuation feature has the highest relative similarity to that category, and does not necessarily mean that the defect has reached the point of initiation. After obtaining the molten pool fluctuation classification result, the dimensionality-reduced feature vector at the current moment is extracted. The standard template of the corresponding category in the category library is retrieved, and the cosine similarity between the two is calculated. This similarity value is the category matching degree, used to quantify the proximity of the current anomaly state to the defect initiation threshold. If the classification result corresponds to multiple candidate anomaly categories, the similarity between the current feature and all corresponding templates is calculated separately, and the maximum value is taken as the final category matching degree, ensuring that the most severe anomalies are not missed.

[0054] The anomaly detection threshold is set to 80. This threshold is determined based on the matching degree distribution of historical defect samples. The matching degree of samples with actual defect initiation is generally higher than 85%, while the matching degree of samples with normal fluctuations is generally lower than 70. The threshold of 80 can ensure a defect detection rate of over 95% while keeping the false detection rate below 5%. If the category matching degree is greater than the anomaly detection threshold, it indicates that the current fluctuation pattern highly matches the initiation characteristics of the corresponding defect, and a defect initiation event is determined to have occurred. At the same time, the current system clock time is recorded as the defect initiation timestamp. After the defect initiation is determined, the closed region in the physical field matrix with a temperature lower than the solidus line of the material is used as the defect candidate region. The number of grids in the candidate region is multiplied by the physical volume of a single grid to obtain the defect volume. At the same time, the average temperature of all nodes in the candidate region is extracted as the current temperature estimate of the defect region. If the category matching degree is not greater than the anomaly detection threshold, it indicates that the current disturbance has not yet reached the level of defect initiation and only belongs to the category of normal fluctuations. No defect warning is triggered, and real-time monitoring continues.

[0055] Finally, based on the defect initiation timestamp, the corresponding spatial coordinates are retrieved from the motion trajectory data recorded in real time by the processing control system, which is the precise spatial location of the defect. If there is no record in the trajectory data that is precisely aligned with the defect initiation timestamp, the predecessor and successor nodes with the smallest time difference are searched, and linear interpolation is performed according to the time ratio to obtain the precise spatial coordinates of the defect. For example, taking into account the aforementioned disturbance parameters, eight global statistical features are extracted from the dual-channel physical field state matrix, and concatenated with the disturbance amplitude and disturbance frequency to obtain a 10-dimensional fused feature vector. After orthogonal transformation to reduce the dimensionality, a dimensionality-reduced feature vector is obtained. After inputting into the ensemble classification model, the majority vote results in a classification of molten pool fluctuation as violent splashing. The matching degree between this and the porosity initiation template is calculated to be 88, which is higher than the anomaly judgment threshold of 80, and a porosity initiation event is determined to have occurred, with the corresponding initiation timestamp being the 4500th millisecond. The workpiece coordinate system trajectory data of the machining control system is retrieved, and the two trajectory nodes with the smallest time difference are found. After linear interpolation calculation, the spatial coordinates of the defect are obtained as 120 mm horizontally, 45 mm vertically, and 5 mm deep, thus accurately locating the specific position of the defect on the workpiece.

[0056] In step S15, calculating the energy compensation value of the target area based on the spatial coordinates and generating a power adjustment command to adjust the output power of the laser device includes: Based on the spatial coordinates, the defect volume and melting depth parameters of the target area are extracted, and the missing heat input of the target area is calculated in combination with the material thermal property parameters to obtain the energy compensation value. Based on the energy compensation value, match the laser power adjustment parameters and encapsulate to generate a power adjustment command; The power adjustment command is parsed to generate a hardware drive signal, which adjusts the laser output power.

[0057] It should be noted that the energy compensation value is calculated using the spatial coordinates of the defect as the positioning reference to pinpoint the target area requiring compensation. First, two core parameters are extracted from this area: the defect volume and the actual melting depth. These are then combined with material density, specific heat capacity, melting point, and latent heat of phase change thermophysical properties to calculate the static basic heat input. The calculation formula is as follows: ,in For material density, For the defect volume, For the specific heat capacity of the material, The melting point of the material. This is the estimated current temperature of the defect area. This is the latent heat of fusion of the material. This formula fully covers the total sensible and latent heat required to reheat the defective area to a fully molten state.

[0058] Based on this, a dynamic heat dissipation correction factor is introduced. Using a moving heat source model, this factor comprehensively considers heat loss through conduction, convection, and radiation. A typical correction factor value is approximately 2.46. Multiplying the static base heat input by the correction factor yields the total heat demand considering the dynamic heat dissipation effect. Finally, this is divided by the laser absorptivity of the material under the current process conditions (typically 0.35 for low-carbon steel) to obtain the final energy compensation value. The laser absorptivity and heat loss coefficient are determined through standard test plate calibration experiments. Keeping the laser power and scanning speed constant, the actual melt depth and width of the molten pool are measured, and the actual absorptivity and loss coefficient of the material under the current process conditions are deduced to ensure an accurate match between the energy calculation results and the actual heat input demand.

[0059] When a power adjustment command is generated, the target compensation power is calculated based on the energy compensation value and the equivalent action time of the laser beam through the defect area. The power adjustment amount is determined in conjunction with the current base processing power, and the compensation action time window and power rise / fall slope are set simultaneously. The command includes four core parameters: target compensation power, compensation start time, compensation duration, and power gradient slope. These are encapsulated as a structured command and sent to the laser generator's underlying control board. The timing of the power adjustment is strictly aligned with the movement position of the processing head to ensure that the laser beam scans the defect coordinate range precisely within the compensation power range, avoiding any impact on the repair effect due to timing deviations.

[0060] During the power adjustment execution phase, the power adjustment command is first parsed to obtain a time-sequential sequence of adjustment parameters. A corresponding hardware drive signal is then generated through pulse width modulation and output to the pump source control terminal of the laser equipment to adjust the laser output power in real time. The timing of the power adjustment is strictly aligned with the movement position of the processing head to ensure that energy compensation is performed precisely when the laser beam scans to the defect coordinates, avoiding any impact on the repair effect due to timing deviations in compensation.

[0061] After power adjustment, the control effect can be further verified by simultaneously collecting the geometric morphology and surface temperature characteristics of the current molten pool and comparing them with the preset stable state characteristic range. The stable state characteristic range is determined based on statistical data of molten pools corresponding to 500 consecutive qualified welds. The normal range of molten pool width is 2 to 4 mm, and the normal fluctuation range of center temperature is ±50 degrees Celsius. When both indicators fall into the stable range, it indicates that the thermodynamic state inside the molten pool has returned to equilibrium, and the real-time control for defect initiation is considered complete. If both indicators do not fall into the stable range, it indicates that there is a deviation in the compensation amount. The power adjustment amount is calculated using a proportional adjustment method: the power adjustment amount equals the temperature deviation value multiplied by the proportional coefficient, which is 20 watts per degree Celsius. The compensation time is simultaneously corrected according to the molten pool width deviation, and a second power fine-tuning is performed. After the adjustment, the molten pool characteristics are collected again for verification until all indicators fall into the stable range. At the same time, the maximum number of adjustments is set to 3. If the steady state is not reached after exceeding the number of adjustments, it is judged as an abnormal working condition. The current parameters are recorded and reported to the management terminal to avoid unlimited iteration affecting the processing rhythm.

[0062] For example, using the aforementioned porosity defect parameters, the measured porosity defect volume is 15 cubic millimeters, and the estimated current temperature of the defect area is 800 Kelvin. Combining the density, specific heat capacity, melting point, and latent heat of fusion parameters of low-carbon steel, the calculated static basic heat input is 114.6 Joules. Based on the moving heat source model, which integrates heat conduction, convection, and radiation losses, a dynamic heat dissipation correction factor of approximately 2.46 is obtained. After correction using the material's laser absorptivity of 0.35, the final energy compensation value is approximately 805.4 Joules.

[0063] Based on the spot diameter and welding speed, the equivalent action time of the laser beam through the defect area is calculated to be approximately 0.134 seconds, corresponding to a compensation power of approximately 6000 watts. The laser power is linearly increased starting 0.02 seconds before the defect enters the action range, and then linearly reduced back to the original processing power within 0.02 seconds after passing the defect. The total compensation time window is approximately 0.174 seconds, at which point a power adjustment command is generated and issued. This gradual power adjustment avoids keyhole instability and spatter problems caused by power step changes.

[0064] After adjustment, the width of the molten pool was measured to be 3 mm, the center temperature was 1600 degrees Celsius, the temperature uncertainty of the physical field state matrix dropped to within 50 degrees Celsius, and the spatial deviation of the multiphysics nonlinear evolution trajectory dropped to within one standard deviation of the initial threshold. All indicators were within the preset stable state range, indicating that the porosity defects had been remelted and eliminated, the thermodynamic state of the molten pool had returned to stability, and the adjustment was deemed complete.

[0065] In summary, this invention discloses a laser processing control method based on multi-physics coupling, including steps such as surface monitoring video feature extraction, multi-physics state matrix construction, nonlinear evolution trajectory analysis and disturbance identification, defect initiation determination and spatial positioning, energy compensation calculation and real-time power control. This invention infers the internal thermal flux coupling state of the molten pool through surface visual features, predicts defect initiation based on the evolution trajectory, and executes directional energy compensation. This achieves full-process control of real-time multi-physics sensing of the molten pool, early defect warning, and closed-loop power control, effectively improving the quality stability and defect suppression capability of laser processing and meeting the process requirements of high-precision intelligent welding systems.

[0066] The second embodiment of the present invention provides a laser processing control system based on multi-physics coupling, including: a memory, a processor, and a computer program stored in the memory and executable on the processor. When the processor executes the program, it implements the steps of the method described above.

[0067] It should be noted that the laser processing process control system based on multi-physics coupling provided in this embodiment of the invention is used to execute all the process steps of the laser processing process control method based on multi-physics coupling in the above embodiment. The working principles and beneficial effects of the two are one-to-one, so they will not be described again.

[0068] It should be noted that the device embodiments described above are merely illustrative. The units described as separate components may or may not be physically separate, and the components shown as units may or may not be physical units; that is, they may be located in one place or distributed across multiple network units. Some or all of the modules can be selected to achieve the purpose of this embodiment according to actual needs. Furthermore, in the accompanying drawings of the device embodiments provided by this invention, the connection relationships between modules indicate that they have communication connections, which can be specifically implemented as one or more communication buses or signal lines. Those skilled in the art can understand and implement this without any creative effort.

[0069] The specific embodiments described above further illustrate the purpose, technical solution, and beneficial effects of the present invention. It should be understood that the above descriptions are merely specific embodiments of the present invention and are not intended to limit the scope of protection of the present invention. In particular, it should be noted that any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the scope of protection of the present invention for those skilled in the art.

Claims

1. A laser processing control method based on multi-physics coupling, characterized in that, include: Acquire surface monitoring video streams during laser processing and extract surface feature vectors of the molten pool; Arrange the surface feature vectors of the molten pool in chronological order for multiple consecutive frames to construct a temporal feature sequence. Perform spatiotemporal coupling feature mapping on the temporal feature sequence to obtain the internal temperature field matrix and flow field matrix of the molten pool. Align and fuse the temperature field matrix and flow field matrix according to spatial coordinates to obtain the physical field state matrix. Extract the temporal variation gradient tensor of the physical field state matrix, and generate a multi-physics nonlinear evolution trajectory through manifold dimensionality reduction. Calculate the spatial deviation between the multi-physics nonlinear evolution trajectory and the preset trajectory baseline. Determine the perturbation segment trajectory data based on the spatial deviation, and perform time-frequency decomposition to obtain the corresponding perturbation amplitude and perturbation frequency. By combining the disturbance amplitude, the disturbance frequency, and the physical field state matrix, molten pool fluctuations are classified to obtain molten pool fluctuation classification results, thereby identifying defect initiation events and determining their corresponding spatial coordinates. The energy compensation value of the target area is calculated based on the spatial coordinates, and a power adjustment command is generated to adjust the output power of the laser device.

2. The laser processing control method based on multi-physics coupling according to claim 1, characterized in that, The acquisition of surface monitoring video streams during laser processing and extraction of molten pool surface feature vectors include: The surface monitoring video stream of the laser processing molten pool area is acquired and split into single-frame surface images according to the time series. Extract the molten pool edge from the single-frame surface image to obtain the molten pool contour matrix; Deep spatial feature extraction is performed on the molten pool contour matrix to obtain the molten pool surface feature vector.

3. The laser processing control method based on multi-physics coupling according to claim 1, characterized in that, The process involves performing spatiotemporal coupling feature mapping on the time-series feature sequence to obtain the temperature field matrix and flow field matrix inside the molten pool. The temperature field matrix and the flow field matrix are then aligned and fused according to spatial coordinates to obtain the physical field state matrix, including: Extract the long and short-term time-series features of the time-series feature sequence to obtain a spatiotemporal coupled feature tensor containing short-term fluctuation features and long-term trend features; The spatiotemporal coupling feature tensor is nonlinearly spatially mapped to output the temperature field matrix and the flow field matrix, respectively. The temperature field matrix and the flow field matrix are numerically fused according to their corresponding spatial nodes to obtain the coupled physical field state matrix.

4. The laser processing control method based on multi-physics coupling according to claim 1, characterized in that, The step of extracting the temporal variation gradient tensor of the physical field state matrix and generating multi-physics nonlinear evolution trajectories through manifold dimensionality reduction includes: Two sets of physical field state matrices at adjacent sampling times are obtained, and the difference between the corresponding matrix elements is calculated to obtain the temporal gradient tensor. The phase space of the temporally varying gradient tensor is reconstructed to obtain a high-dimensional phase space feature sequence; The high-dimensional phase space feature sequence is subjected to manifold dimensionality reduction and mapped to a low-dimensional feature space; Based on the state transition sequence in the low-dimensional feature space, a multi-physics nonlinear evolution trajectory is generated.

5. The laser processing control method based on multi-physics coupling according to claim 1, characterized in that, The calculation of the spatial deviation between the multiphysics nonlinear evolution trajectory and the preset trajectory baseline, the determination of the perturbation segment trajectory data based on the spatial deviation, and the time-frequency decomposition to obtain the corresponding perturbation amplitude and frequency include: The vertical distance from each sampling point on the nonlinear evolution trajectory of the multiphysics field to the preset trajectory baseline is calculated point by point to obtain the spatial deviation of each sampling point; If the spatial deviation of a consecutive preset number of sampling points is greater than a preset deviation threshold, it is determined that there is a slight disturbance in the laser processing process; Empirical mode decomposition was performed on the deviation segment trajectory data to obtain multiple sets of intrinsic mode function components; The intrinsic mode function components are subjected to spectral transformation to extract the disturbance amplitude and disturbance frequency.

6. The laser processing control method based on multi-physics coupling according to claim 1, characterized in that, The process of classifying molten pool fluctuations by combining the disturbance amplitude, the disturbance frequency, and the physical field state matrix to obtain molten pool fluctuation classification results includes: The disturbance amplitude and the disturbance frequency are dimensionally expanded to obtain the extended disturbance characteristics; The extended perturbation features are concatenated and fused with the physical field state matrix to obtain a fused feature matrix; The fused feature matrix is ​​then subjected to dimensionality reduction and noise reduction to obtain a dimensionality-reduced feature vector; The reduced-dimensional feature vectors are classified and identified to obtain the classification results of molten pool fluctuations.

7. The laser processing control method based on multi-physics coupling according to claim 1, characterized in that, The process of identifying defect initiation events and determining their corresponding spatial coordinates includes: Calculate the matching degree between the molten pool fluctuation classification results and the preset anomaly categories; If the matching degree is greater than the preset anomaly judgment threshold, a defect initiation event is determined to have occurred, and the corresponding defect initiation timestamp is recorded. Based on the defect spawn timestamp, the corresponding motion trajectory data is matched to obtain the spatial coordinates of the defect.

8. The laser processing control method based on multi-physics coupling according to claim 1, characterized in that, The step of calculating the energy compensation value of the target area based on the spatial coordinates and generating a power adjustment command to adjust the output power of the laser device includes: Based on the spatial coordinates, the defect volume and melting depth parameters of the target area are extracted, and the missing heat input of the target area is calculated in combination with the material thermal property parameters to obtain the energy compensation value. Based on the energy compensation value, match the laser power adjustment parameters and encapsulate to generate a power adjustment command; The power adjustment command is parsed to generate a hardware drive signal, which adjusts the laser output power.

9. A laser processing control system based on multiphysics coupling, comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, characterized in that, When the processor executes the program, it implements the method according to any one of claims 1 to 8.

10. A computer-readable storage medium having a computer program stored thereon, characterized in that, When executed by a processor, the program implements the method described in any one of claims 1 to 8.

Images