Laser welding deformation adaptive compensation system and method for lithium battery module busbar

Abstract

The application discloses a laser welding deformation self-adaptive compensation system and method of a lithium battery module busbar, relates to the technical field of lithium battery module welding processing, and inputs current temperature field, current full-field deformation distribution and preset welding parameters of a next welding point into an embedded physical information neural network; target positions are calculated based on welding point offset prediction vectors and welding reference position coordinates, three-dimensional trajectory compensation amounts are obtained in combination with current positions of laser focal points, and are decomposed into galvanometer deflection angle correction amounts and focusing correction amounts. Through construction of a welding reference three-dimensional digital model and in combination with multi-source sensing synchronous acquisition technology, the application can comprehensively acquire welding process temperature field, three-dimensional topography, welding seam offset and full-field deformation data, and realizes omnibearing real-time sensing of a welding state. Full-field deformation information is acquired by adopting a dual-wavelength laser interference measurement mode, environmental temperature interference is effectively avoided, and dynamic deformation detection precision is improved.

Description

Technical Field

[0001] This invention relates to the field of lithium battery module welding technology, and in particular to an adaptive compensation system and method for laser welding deformation of lithium battery module busbars. Background Technology

[0002] Lithium-ion battery modules are core components of new energy power battery systems. Busbars, as the key connection structure for the series and parallel connection of module cells, directly determine the conductivity, structural strength, and service safety of the lithium-ion battery module through their welding precision and stability. Laser welding, with its advantages of a small heat-affected zone, high welding efficiency, and uniform weld formation, has become the mainstream welding process for lithium-ion battery busbars.

[0003] In actual mass welding production, continuous heat input from the laser beam can cause uneven temperature fields in the busbar, leading to thermal expansion and thermoelastic-plastic deformation of the material, resulting in deformation problems such as weld point misalignment and board warping. Welding deformation has a time-cumulative characteristic; deformation of early weld points will continue to affect the welding position accuracy of subsequent weld points. Conventional fixed-track laser welding methods cannot adapt to the dynamic deformation changes during the welding process, which can easily cause welding defects such as weld misalignment, incomplete welds, and missed welds.

[0004] Existing welding compensation technologies mostly rely on single visual inspection or static error calibration methods, which can only passively correct the forming errors after welding. They cannot capture the dynamic deformation of the entire welding process in real time, nor do they combine the welding heat transfer mechanism and material deformation law for predictive compensation. At the same time, traditional neural network prediction models are mostly based on pure data-driven training and lack physical mechanism constraints. Under complex welding conditions, their prediction accuracy is limited, their generalization ability is poor, and they are difficult to achieve dynamic adaptive optimization of welding trajectory. This results in poor welding consistency and low yield of lithium battery busbars, which cannot meet the high-precision and high-stability mass production welding requirements of high-end lithium battery modules. Summary of the Invention

[0005] To address the aforementioned technical problems, this invention provides an adaptive compensation system and method for laser welding deformation of lithium battery module busbars. The technical solution adopted is as follows: An adaptive compensation method for laser welding deformation of lithium battery module busbars includes the following steps: Step 1: Obtain point cloud data of the position coordinates of all busbars to be welded and the surface morphology of the busbars, construct a pre-welding reference 3D digital model, and define reference position coordinates and reference normal vector for each weld point; Step 2: Welding is performed sequentially according to the preset welding point sequence. During the welding of each welding point and within the preset acquisition window after welding, the temperature field distribution data of the welding area, the three-dimensional morphology data of the busbar surface, and the weld position offset data are collected simultaneously. Step 3: Using a dual-wavelength laser interferometry device, two lasers with different wavelengths are emitted onto the busbar surface to obtain the full-field deformation distribution, including out-of-plane and in-plane displacements, in real time. Step 4: Input the current temperature field, the current full-field deformation distribution, and the preset welding parameters of the next weld point into the embedded physical information neural network. The loss function of the embedded physical information neural network includes the physical constraint term of the partial differential equation of heat conduction and the thermo-elastic-plastic temporal consistency constraint term of the deformation field. It predicts the weld point offset prediction vector and prediction confidence interval of the next weld point online. Step 5: Calculate the target position based on the weld point offset prediction vector and the pre-welding reference position coordinates, and obtain the three-dimensional trajectory compensation amount by combining the current position of the laser focus, and decompose it into the galvanometer deflection angle correction amount and the focusing correction amount. Step 6: Perform phase advance compensation on the correction amount, and perform trajectory smoothing transition when the prediction confidence interval exceeds the threshold, and generate control commands to drive the laser welding actuator. Step 7: After each solder joint is completed, the physical information neural network is incrementally trained using the newly added sensor data. This process is repeated until all solder joints are completed.

[0006] Optionally, step 3 includes the following sub-steps: Step 31, the wavelengths are respectively and Two laser beams are projected onto a preset measurement point array on the surface of the busbar through the same optical path, and interference signals are formed by reflection. Step 32: Perform Hilbert transform and phase unwrapping on the interference signal for each wavelength to obtain the phase signal. and The synthesized wavelength and synthesized phase are calculated, and the interference of ambient temperature fluctuations is eliminated by using the synthesized wavelength correction term. Step 33: Calculate the out-of-plane displacement based on the phase signal, and obtain the in-plane displacement by calculating the two-dimensional cross-correlation of the speckle patterns at adjacent time points; Step 34: The displacement of discrete measurement points is extended to the full-field deformation distribution of the welding area by radial basis function interpolation.

[0007] Optionally, the physical information neural network is composed of a deep convolutional neural network encoder and a long short-term memory network cascaded together; the deep convolutional neural network encoder includes a temperature field encoding branch and a deformation field encoding branch, which extract the spatial features of the temperature field and the deformation field respectively; the long short-term memory network receives the spliced ​​vector of the spatial features and the welding parameters of the next weld point, learns the temporal evolution features and outputs the weld point offset prediction vector.

[0008] Optionally, the loss function for the embedded physical information neural network is: ; in It is the total loss value of the neural network. The mean square error between the network-predicted deformation and the measured deformation by dual-wavelength laser interferometry; Physical constraint error loss in the heat conduction equation The loss is the deformation time-series consistency constraint error loss, and α, β, and γ are the dynamic weighting coefficients of the three types of losses, respectively.

[0009] Optionally, in step 5, decomposing the three-dimensional trajectory compensation into galvanometer deflection angle correction and focusing correction includes the following sub-steps: Step 51, calculate the target correction position. ; Step 52, read the current position of the laser focus. The total compensation amount is obtained. ; Step 53: Based on the optical parameters, convert the total compensation amount into the galvanometer deflection angle correction amount through a coordinate transformation function. and and Z-axis focusing correction amount .

[0010] Optionally, in step 6, the transfer function for phase lead compensation is: ;in This is the transfer function for lead compensation in the complex frequency domain. This is the gain coefficient. It is a time constant. Let Laplace be a complex variable. The leading factor is used; after being discretized into difference equations by bilinear transformation, the correction sequence is subjected to recursive filtering.

[0011] Optionally, the incremental training method is as follows: after each weld joint is completed, the weights of the long short-term memory network layer and the output layer in the physical information neural network that are related to the latest data are updated using the newly added temperature field of the weld joint, the measured deformation amount of dual-wavelength laser interferometry, and the welding process data.

[0012] Optionally, the simultaneous acquisition of multi-source sensor data includes: an infrared thermal imager acquiring the temperature field matrix T(x,y,t) at a frame rate of not less than 100Hz; a line laser 3D contour sensor acquiring the three-dimensional topographic point cloud at a line scan frequency of not less than 2kHz; and a coaxial vision camera acquiring weld point images and detecting the actual offset relative to the reference position. All data is timestamped using a unified clock source.

[0013] Optionally, the dual-wavelength laser interferometry device sets a preset measurement point array on the busbar surface, with at least 4 measurement points arranged within the radius of the heat-affected zone, centered on the current welding joint, and each measurement point is distributed in a radial direction.

[0014] An adaptive compensation system for laser welding deformation of lithium battery module busbars is used to realize an adaptive compensation method for laser welding deformation of lithium battery module busbars. The system includes a multi-source sensor acquisition unit, a welding deformation digital twin prediction unit, a dynamic trajectory adaptive compensation unit, and a laser welding execution unit. The multi-source sensing acquisition unit includes an infrared thermal imager, a line laser 3D contour sensor, a coaxial vision camera, and a dual-wavelength laser interferometry device, which are used to simultaneously acquire the temperature field, three-dimensional morphology, weld position offset, and real-time full-field deformation distribution of the welding area, and align them with a unified clock. The welding deformation digital twin prediction unit has an embedded physical information neural network built in it. It receives the time series data of the multi-source sensing acquisition subsystem, introduces heat conduction physical constraints and deformation time series consistency constraints into the loss function, and predicts the weld point offset prediction vector and confidence interval of the next weld point online. The dynamic trajectory adaptive compensation unit receives the weld point offset prediction vector, calculates the three-dimensional trajectory compensation amount and decomposes it into the galvanometer deflection angle correction amount and the focusing correction amount. After phase advance compensation and smooth transition processing, it generates control commands. The laser welding execution unit includes a fiber laser, a galvanometer scanning head, and a Z-axis focusing mechanism, which execute control commands. After each weld point is completed, the welding deformation digital twin prediction unit uses newly added sensor data to incrementally train the physical information neural network, forming a closed-loop iterative optimization.

[0015] In summary, the present invention has at least one of the following beneficial technical effects: This invention provides an adaptive compensation system and method for laser welding deformation of lithium battery module busbars. By constructing a pre-welding reference three-dimensional digital model and combining it with multi-source sensor synchronous acquisition technology, it can comprehensively acquire temperature field, three-dimensional morphology, weld offset, and full-field deformation data during the welding process, achieving all-round real-time perception of the welding status. The use of dual-wavelength laser interferometry to acquire full-field deformation information effectively avoids environmental temperature interference and improves the accuracy of dynamic deformation detection.

[0016] An embedded physical information neural network integrating thermal conduction physical constraints and thermo-elastic-plastic temporal constraints is introduced, abandoning the traditional pure data-driven prediction mode. By combining welding physical mechanisms, the accuracy and rationality of weld point offset prediction are improved, enabling precise prediction of the deformation offset trend of the next weld point. Through three-dimensional trajectory compensation calculation, phase lead compensation, and trajectory smoothing transition processing, the laser welding trajectory can be corrected in real time, offsetting system hysteresis errors and dynamic deformation errors, and significantly improving weld point alignment accuracy.

[0017] By employing incremental training at single weld points, the network model achieves closed-loop iterative optimization, continuously adapting to real-time welding condition changes and effectively improving the model's dynamic prediction capabilities and condition adaptability. The overall technical solution addresses the issues of deformation accumulation, poor trajectory adaptability, and insufficient welding precision inherent in traditional welding processes. It significantly improves the consistency and yield of laser welding for lithium battery busbars, ensuring the conductivity stability and structural reliability of lithium battery modules, and meeting the demands of large-scale, high-precision mass production. Attached Figure Description

[0018] Figure 1 This is a flowchart illustrating the adaptive compensation method for laser welding deformation of the lithium battery module busbar according to the present invention. Figure 2 This is a schematic diagram of the architecture of the laser welding deformation adaptive compensation system for the lithium battery module busbar of the present invention; Detailed Implementation

[0019] The present invention will be further described in detail below with reference to the accompanying drawings.

[0020] This invention discloses an adaptive compensation system and method for laser welding deformation of lithium battery module busbars.

[0021] Reference Figure 1 and Figure 2 Example 1, an adaptive compensation method for laser welding deformation of lithium battery module busbars, includes the following steps: Step 1: Obtain point cloud data of the position coordinates of all busbars to be welded and the surface morphology of the busbars, construct a pre-welding reference 3D digital model, and define reference position coordinates and reference normal vector for each weld point; Step 2: Welding is performed sequentially according to the preset welding point sequence. During the welding of each welding point and within the preset acquisition window after welding, the temperature field distribution data of the welding area, the three-dimensional morphology data of the busbar surface, and the weld position offset data are collected simultaneously. Step 3: Using a dual-wavelength laser interferometry device, two lasers with different wavelengths are emitted onto the busbar surface to obtain the full-field deformation distribution, including out-of-plane and in-plane displacements, in real time. Step 4: Input the current temperature field, the current full-field deformation distribution, and the preset welding parameters of the next weld point into the embedded physical information neural network. The loss function of the embedded physical information neural network includes the physical constraint term of the partial differential equation of heat conduction and the thermo-elastic-plastic temporal consistency constraint term of the deformation field. It predicts the weld point offset prediction vector and prediction confidence interval of the next weld point online. Step 5: Calculate the target position based on the weld point offset prediction vector and the pre-welding reference position coordinates, and obtain the three-dimensional trajectory compensation amount by combining the current position of the laser focus, and decompose it into the galvanometer deflection angle correction amount and the focusing correction amount. Step 6: Perform phase advance compensation on the correction amount, and perform trajectory smoothing transition when the prediction confidence interval exceeds the threshold, and generate control commands to drive the laser welding actuator. Step 7: After each solder joint is completed, the physical information neural network is incrementally trained using the newly added sensor data. This process is repeated until all solder joints are completed.

[0022] By adopting the above technical solution, firstly, before welding begins, a full-area 3D scan is performed on the assembled but not yet welded battery module. The scanning equipment is a line laser 3D contour sensor. The scan acquires the position coordinates of all busbar solder points to be welded, as well as point cloud data of the busbar surface morphology, constructing a pre-welding reference 3D digital model. In the model, reference position coordinates and a reference normal vector are defined for each solder point. The reference position coordinates characterize the spatial position of the solder point in an ideal, undeformed state, and the reference normal vector is used to subsequently determine the laser incident direction.

[0023] Subsequently, following the pre-defined welding point sequence, laser welding operations were performed on each welding point sequentially. During the welding process of each welding point, and within a pre-defined acquisition window after the welding point was completed, the multi-source sensor acquisition subsystem simultaneously acquired three types of data: temperature field distribution data of the welding area, acquired by an infrared thermal imager at a frame rate of no less than 100Hz to form a temperature matrix; three-dimensional topographic data of the busbar surface, acquired by a line laser 3D contour sensor at a line scanning frequency of no less than 2kHz to output a topographic point cloud; and weld position offset data, obtained by a coaxial vision camera capturing images of the welding point position and detecting the actual offset of the current welding point relative to the reference position using image processing algorithms. and All sensor data is timestamped via a unified clock source on the system control bus, ensuring strict time alignment.

[0024] During the welding process, a dual-wavelength laser interferometry device emits two laser beams of different wavelengths onto a pre-set array of measurement points on the busbar surface. By analyzing the interference signal formed by the reflected light, the out-of-plane and in-plane displacements of each measurement point on the busbar surface are extracted in real time using a dual-wavelength interferometry method. Spatial interpolation is then used to generate the full-field deformation distribution of the welding area. This full-field deformation distribution directly reflects the dynamic deformation state of the busbar caused by heat input.

[0025] The current temperature field, the current full-field deformation distribution, and the preset welding parameters for the next weld point, including laser power, welding speed, and weld point coordinates, are input into an embedded physical information neural network. The network's loss function incorporates physical constraint terms from the heat conduction partial differential equation and temporal consistency constraint terms corresponding to the thermo-elastic-plastic constitutive relation of the deformation field. The network outputs online the weld point offset prediction vector and prediction confidence interval for the next weld point at the welding moment. The weld point offset prediction vector contains the expected offset of the weld point relative to its reference position in three-dimensional space.

[0026] The target correction position is obtained by adding the predicted weld point offset vector to the pre-welding reference position coordinates. The current position information of the laser focus in the laser welding actuator is read in real time, and the difference between the target correction position and the current position is used as the three-dimensional trajectory compensation amount. Subsequently, using the coordinate transformation function of the optical system, this three-dimensional trajectory compensation amount is decomposed into a two-dimensional deflection angle correction amount for the galvanometer and a Z-axis focusing correction amount.

[0027] The aforementioned corrections are then subjected to phase lead compensation. The transfer function of the phase lead compensator is used to perform phase lead shaping on the compensation signal to offset the response lag of the actuator. When the prediction confidence interval exceeds a preset threshold, the system automatically reduces the compensation gain and uses cubic spline interpolation to smooth the trajectory transition between the compensation amounts of adjacent weld points, while limiting the rate of change of the galvanometer angular velocity to prevent mechanical shock. Finally, control commands are generated to drive the laser welding actuator to complete the welding of the current weld point.

[0028] After each weld point is completed, the newly acquired temperature field data, measured deformation by dual-wavelength laser interferometry, and other welding process data are used to perform incremental training on the physical information neural network, updating the weights related to the time series in the network. This process does not interrupt the welding cycle, forming a continuous closed-loop optimization that involves welding, measuring, learning, predicting, and compensating simultaneously, until all weld points are completed.

[0029] Example 2, step 3 includes the following sub-steps: Step 31, the wavelengths are respectively and Two laser beams are projected onto a preset measurement point array on the surface of the busbar through the same optical path, and interference signals are formed by reflection. Step 32: Perform Hilbert transform and phase unwrapping on the interference signal for each wavelength to obtain the phase signal. and The synthesized wavelength and synthesized phase are calculated, and the interference of ambient temperature fluctuations is eliminated by using the synthesized wavelength correction term. Step 33: Calculate the out-of-plane displacement based on the phase signal, and obtain the in-plane displacement by calculating the two-dimensional cross-correlation of the speckle patterns at adjacent time points; Step 34: The displacement of discrete measurement points is extended to the full-field deformation distribution of the welding area by radial basis function interpolation.

[0030] By adopting the above technical solution, the dual-wavelength laser interferometry device incorporates two laser sources, each emitting a wavelength of [wavelength value missing]. and wavelength The laser beam, in which The preferred size is 632.8nm. The preferred wavelength is 532 nm. Two laser beams are coupled to the same optical path through a beam combiner and projected onto a preset measurement point array on the surface of the busbar via an objective lens. The laser beam reflected back from the busbar surface passes through the optical path again, superimposed with the reference beam to form an interference signal, which is received by a photodetector and converted into an electrical signal.

[0031] For each wavelength of the interference signal, the data acquisition and processing unit performs a Hilbert transform to convert the real-valued signal into an analytic signal, thereby extracting the instantaneous phase. The extracted entangled phase is then unwrapped to obtain the continuous phase signal. and To eliminate measurement drift caused by slow fluctuations in ambient temperature, a dual-wavelength synthesis technique is introduced: the synthesized wavelength Λ is calculated to be equal to... and The product of the two divided by the absolute value of their difference; the combined phase equals minus By utilizing the large measurement range provided by the synthetic wavelength, a synthetic wavelength correction term is calculated, which is used to correct the periodic ambiguity problem of single-wavelength measurements.

[0032] out-of-plane displacement The calculation formula is: equal to four-π Multiply Compared to the initial time The difference, plus the synthesized wavelength correction term. In-plane displacement. The speckle correlation method is used to obtain the pixel displacement of the in-plane displacement by performing two-dimensional cross-correlation on speckle intensity images acquired at adjacent time points. The displacement corresponding to the cross-correlation peak is converted into physical displacement by the system calibration coefficient.

[0033] Finally, using the out-of-plane and in-plane displacement data at each discrete measurement point as nodes, a continuous full-field deformation distribution map covering the entire welding area is generated using the radial basis function interpolation method. The interpolated full-field deformation distribution can be directly used as the feature input for a physical information neural network.

[0034] Example 3: The physical information neural network is composed of a deep convolutional neural network encoder and a long short-term memory network cascaded together. The deep convolutional neural network encoder includes a temperature field encoding branch and a deformation field encoding branch, which extract the spatial features of the temperature field and the deformation field, respectively. The long short-term memory network receives the spliced ​​vector of the spatial features and the welding parameters of the next weld point, learns the temporal evolution features, and outputs the weld point offset prediction vector.

[0035] By adopting the above technical solution, the embedded physical information neural network employs a cascaded architecture of encoder-temporal predictor. The encoder part is a deep convolutional neural network containing two parallel input branches: a temperature field encoding branch and a deformation field encoding branch. The temperature field encoding branch receives multi-channel temperature matrix images and extracts feature maps of the temperature field at different spatial scales through several layers of convolution, batch normalization, and activation operations. The deformation field encoding branch receives multi-channel deformation images containing out-of-plane and in-plane displacements and extracts the spatial features of the deformation field in the same way. The two branches are concatenated and fused along the channel dimension to form a joint spatial feature map.

[0036] The temporal predictor is a Long Short-Term Memory (LSTM) network. The joint spatial feature map, after being flattened and processed through fully connected layers, is concatenated with the preset welding parameter vector for the next weld point, serving as the input to the LSM network at each time step. Internally, the LSM network maintains hidden states and cell states, learning the evolution of the temperature and deformation fields in the welding sequence to capture the temporal dependencies of heat accumulation and deformation propagation. The network ultimately outputs a weld point offset prediction vector. , , .

[0037] Example 4: The loss function of the embedded physical information neural network is: ; in It is the total loss value of the neural network. The mean square error between the network-predicted deformation and the measured deformation by dual-wavelength laser interferometry; Physical constraint error loss in the heat conduction equation The loss is the deformation time-series consistency constraint error loss, and α, β, and γ are the dynamic weighting coefficients of the three types of losses, respectively.

[0038] By adopting the above technical solution, the hybrid loss function is defined as the total loss. Equals α multiplied by the data fitting loss Add β multiplied by the physical constraint loss Add γ multiplied by the time consistency loss .

[0039] Data fitting loss The calculation is the mean square error between the solder joint offset predicted by the neural network and the solder joint deformation measured by the dual-wavelength laser interferometry device. This loss drives the network output to approximate the actual physical measurement value.

[0040] Physical constraint loss The equation is constructed based on the partial differential equation of heat conduction. According to Fourier's law and energy conservation, the temperature field inside the busbar should satisfy the following condition: the divergence of the heat conduction term plus the internal heat source Q minus the partial derivative of the product of density and specific heat capacity multiplied by temperature with respect to time equals zero. The intermediate temperature field predictions generated during the forward propagation of the neural network are substituted into this equation to calculate the residual, and the mean square value of the residual is used as the mean square value. The temperature field output by this forced network is physically self-consistent.

[0041] Temporal consistency loss Based on the thermo-elastic-plastic constitutive relation, the time-varying rate of change of the deformation field is a function of the current deformation field and temperature field. The deformation rate is calculated using the time difference between adjacent weld points and the predicted deformation increment, and compared with the theoretical deformation rate derived from the constitutive relation. The mean square error of this comparison constitutes the... This loss ensures that the predicted deformation sequence is smooth in the time dimension and conforms to the laws of mechanics.

[0042] The three weighting coefficients α, β, and γ are not fixed values, but are dynamically adjusted according to the welding process. In the early stages of welding, due to the limited training data, the weights of physical constraints are appropriately reduced and the weights of data fitting are increased. As data accumulates, the weights of physical constraints and temporal consistency constraints are gradually increased, causing the model to converge in a physically reasonable direction.

[0043] In Example 5, step 5, decomposing the three-dimensional trajectory compensation amount into the galvanometer deflection angle correction amount and the focusing correction amount includes the following sub-steps: Step 51, calculate the target correction position. ; Step 52, read the current position of the laser focus. The total compensation amount is obtained. ; Step 53: Based on the optical parameters, convert the total compensation amount into the galvanometer deflection angle correction amount through a coordinate transformation function. and and Z-axis focusing correction amount .

[0044] By adopting the above technical solution, the target correction position is first calculated: the coordinates of the reference position of the next weld point in the pre-welding reference model are marked as... The components of the solder joint offset prediction vector output by the physical information neural network are denoted as... Then the target's corrected position coordinates It equals the sum of the reference coordinates and the predicted offset component by component.

[0045] Next, the actual position coordinates of the current laser focus are read in real time from the galvanometer encoder and Z-axis position sensor of the laser welding execution unit. Subtracting the components of the target position from the components of the current position yields the total 3D compensation. .

[0046] Finally, based on the calibrated laser welding optical system parameters, including focal length f and optical path geometry, the total compensation is converted into the control quantity of the actuator through a coordinate transformation function. (Galvanometer X-direction deflection angle correction) By function The calculation, its input is , and focal length parameters ; Correction amount for the Y-direction deflection angle of the galvanometer By function Calculate, input is , and focal length parameters Z-axis focusing correction amount By function The calculation, with the main input being and focal length parameters These transformation functions take into account the nonlinear spatial movement of the focal spot after the laser beam is deflected by the galvanometer.

[0047] In Example 6, step 6, the transfer function for phase lead compensation is: ;in This is the transfer function for lead compensation in the complex frequency domain. This is the gain coefficient. It is a time constant. Let Laplace be a complex variable. The leading factor is used; after being discretized into difference equations by bilinear transformation, the correction sequence is subjected to recursive filtering.

[0048] By adopting the above technical solution, the transfer function used in the phase lead compensation stage is... Defined as: Where Kc is an adjustable compensation gain coefficient, τ is a time constant related to the system response characteristics, α is a lead factor taking values ​​between 0 and 1, and s is a Laplace complex variable. This transfer function provides positive phase to compensate for the low-pass hysteresis characteristics of the laser scanning galvanometer and focusing mechanism.

[0049] In practical digital control systems, it is necessary to convert the continuous-domain transfer function into a discrete-time domain digital filter. The bilinear transform method is used for discretization, resulting in a recursive filter equation of the form y[n] = b0·x[n] + b1·x[n-1] - a1·y[n-1]. The transformed coefficients b0, b1, and a1 are jointly determined by Kc, τ, α, and the sampling period of the control system.

[0050] In each control cycle, the dynamic trajectory adaptive compensation unit uses the calculated galvanometer deflection angle correction sequence and focusing correction sequence as filter inputs. It then calculates the filtered output in real time using the aforementioned recursive formula, generating a phase-leading smooth compensation signal. This signal drives the actuator in advance, ensuring precise synchronization between the actual movement of the laser focus and the true offset of the weld point position.

[0051] Example 7, the incremental training method is as follows: after each weld point is completed, the weights of the long short-term memory network layer and the output layer in the physical information neural network that are related to the latest data are updated using the newly added temperature field of the weld point, the measured deformation amount of dual-wavelength laser interferometry and the welding process data.

[0052] By adopting the above technical solution, the physical information neural network performs incremental training immediately after each weld point is completed, rather than waiting until all weld points are completed before unified training. The input samples for incremental training only contain data associated with the latest weld point, specifically including: the temperature field sequence during the welding process of that weld point, the real-time deformation obtained by dual-wavelength laser interferometry, and welding process parameters. The scope of model parameter updates is limited: the main part of the deep convolutional neural network encoder is frozen, and only the weights of the long short-term memory network and the weights of the final output layer are updated based on gradients. This local parameter update strategy significantly reduces the computational cost of a single training iteration, allowing the time consumed by each incremental training on the embedded controller to be stably controlled within 50ms, thus seamlessly integrating into the welding production cycle without causing additional waiting time. As the welding sequence progresses, the model continuously absorbs new information, and the prediction accuracy of subsequent weld point offsets gradually improves.

[0053] Example 8, simultaneous acquisition of multi-source sensor data includes: an infrared thermal imager acquiring the temperature field matrix T(x,y,t) at a frame rate of not less than 100Hz; a line laser 3D contour sensor acquiring the three-dimensional topographic point cloud at a line scan frequency of not less than 2kHz; and a coaxial vision camera acquiring weld point images and detecting the actual offset relative to the reference position. All data is timestamped using a unified clock source.

[0054] By adopting the above technical solution, an infrared thermal imager is installed on the side of the welding head to observe the welding area from a top-down angle. Its frame rate is set to no less than 100Hz. Each acquired temperature image frame, after non-uniformity correction and radiometric calibration, is converted into a two-dimensional temperature matrix T(x,y,t), with units in degrees Celsius. A line laser 3D contour sensor uses a blue line laser to perform a horizontal scan perpendicular to the busbar surface. The line scan frequency is set to no less than 2kHz. The resulting high-density point cloud is filtered and stitched together to output the three-dimensional morphology of the busbar surface in real time. A coaxial vision camera acquires visible light images of the welding point area through the internal optical path of the welding head. Image processing algorithms such as edge detection and template matching are used to extract welding point features, calculate the deviation between the current welding point image coordinates and the reference image coordinates, and convert this deviation into the actual physical offset using a hand-eye calibration matrix. and .

[0055] The system control and communication bus uses the IEEE 1588 precision time protocol for clock synchronization, assigning a uniform timestamp with nanosecond-level precision to each frame of data. The data acquisition card buffers and aligns all sensor data according to their timestamps, ensuring that the temperature field, morphology, and offset on the same time profile correspond to the same physical moment.

[0056] Example 9: The dual-wavelength laser interferometry device sets a preset measurement point array on the busbar surface. With the current welding joint as the center, at least 4 measurement points are arranged within the radius of the heat-affected zone, and each measurement point is distributed in a radial direction.

[0057] By adopting the above technical solution, the measurement points of the dual-wavelength laser interferometry device do not cover the entire busbar surface, but are instead pre-arrayed at key locations. Using the center of the currently being welded joint as the origin, and estimating the radius of the heat-affected zone based on the pre-set laser power and duration for that joint, at least four measurement points are arranged on the busbar within this radius. These four measurement points are located radially from the joint, for example, on azimuth lines at 0°, 90°, 180°, and 270°, and their distances from the joint center all fall within the heat-affected zone. This arrangement effectively captures the radial and tangential deformations around the joint caused by thermal expansion, providing high-quality boundary and internal constraint points for full-field deformation interpolation.

[0058] Example 10: Adaptive compensation system for laser welding deformation of lithium battery module busbar, used to realize the adaptive compensation method for laser welding deformation of lithium battery module busbar. The system includes a multi-source sensing acquisition unit, a welding deformation digital twin prediction unit, a dynamic trajectory adaptive compensation unit, and a laser welding execution unit. The multi-source sensing acquisition unit includes an infrared thermal imager, a line laser 3D contour sensor, a coaxial vision camera, and a dual-wavelength laser interferometry device, which are used to simultaneously acquire the temperature field, three-dimensional morphology, weld position offset, and real-time full-field deformation distribution of the welding area, and align them with a unified clock. The welding deformation digital twin prediction unit has an embedded physical information neural network built in it. It receives the time series data of the multi-source sensing acquisition subsystem, introduces heat conduction physical constraints and deformation time series consistency constraints into the loss function, and predicts the weld point offset prediction vector and confidence interval of the next weld point online. The dynamic trajectory adaptive compensation unit receives the weld point offset prediction vector, calculates the three-dimensional trajectory compensation amount and decomposes it into the galvanometer deflection angle correction amount and the focusing correction amount. After phase advance compensation and smooth transition processing, it generates control commands. The laser welding execution unit includes a fiber laser, a galvanometer scanning head, and a Z-axis focusing mechanism, which execute control commands. After each weld point is completed, the welding deformation digital twin prediction unit uses newly added sensor data to incrementally train the physical information neural network, forming a closed-loop iterative optimization.

[0059] By adopting the above technical solution, the measurement points of the dual-wavelength laser interferometry device do not cover the entire busbar surface, but are instead pre-arrayed at key locations. Using the center of the currently being welded joint as the origin, and estimating the radius of the heat-affected zone based on the pre-set laser power and duration for that joint, at least four measurement points are arranged on the busbar within this radius. These four measurement points are located radially from the joint, for example, on azimuth lines at 0°, 90°, 180°, and 270°, and their distances from the joint center all fall within the heat-affected zone. This arrangement effectively captures the radial and tangential deformations around the joint caused by thermal expansion, providing high-quality boundary and internal constraint points for full-field deformation interpolation.

[0060] The following specific embodiments illustrate the implementation principle of the present invention: The laser welding process of the aluminum busbar of a square aluminum-cased battery module will be used as an example for explanation.

[0061] A certain power battery module consists of 16 square aluminum-cased cells connected in series. The busbar is made of 1060-O state aluminum alloy sheet with a thickness of 1.5mm. Each busbar requires two weld points to be welded to the cell terminal, for a total of 32 weld points. The welding process parameters are: fiber laser power 1800W, welding speed 90mm / s, defocusing amount 0mm, and shielding gas argon. Before welding, the entire module is scanned using a line laser 3D contour sensor (LMIGocator2530). The scan obtains point cloud data containing the coordinates of all weld points and the surface topography of the busbar, with a resolution of 20μm. In the constructed pre-welding reference 3D digital model, each weld point is assigned reference position coordinates x_ref_i, y_ref_i, z_ref_i and its normal vector. The reference coordinates of the first weld point are (15.20, 8.30, 0.00), in mm.

[0062] The welding sequence follows an S-shaped path, starting from one end of the module and welding sequentially towards the other. During the welding of the k-th weld point, the multi-source sensor acquisition subsystem operates synchronously: the FLIRA655sc infrared thermal imager acquires the temperature field of the welding area at a frame rate of 125Hz, with each frame calibrated and converted into a 240×320 pixel temperature matrix T(x,y,t) with a temperature accuracy of ±1°C; the line laser 3D contour sensor continuously acquires the three-dimensional topography of the busbar surface at a line scan frequency of 2.5kHz, generating a contour point cloud with a spacing of 15μm; the coaxial vision camera acquires images of the weld point area at a frame rate of 100Hz through the internal optical path of the welding head, and calculates the two-dimensional offset δx and δy of the current weld point relative to the reference position using edge detection and template matching algorithms, with an image processing delay of less than 5ms. All sensor data is connected to the system control unit via the EtherCAT bus and stamped with a nanosecond-level timestamp based on the IEEE1588 protocol to achieve data alignment.

[0063] During the welding of the k-th weld point, a dual-wavelength laser interferometry device mounted above the welding head begins operation. This device integrates two frequency-stabilized lasers with center wavelengths of 632.8 nm and 532 nm, respectively. The two beams are combined and projected onto a pre-set array of measurement points on the busbar surface via a beam splitter and objective lens. The measurement point array is centered on the current weld point, with one measurement point at each of the four azimuth angles (0°, 90°, 180°, and 270°) at a radius of 5 mm from the center of the weld point. All four points are located within the heat-affected zone. The reflected light from the busbar surface interferes with the reference light, and the interference is received by a four-quadrant photodetector. The interference signals at wavelengths of 632.8 nm and 532 nm are subjected to Hilbert transforms and unwrapped to obtain continuous phases φ1(t) and φ2(t). The synthesized wavelength Λ is calculated to be approximately 3341 nm (632.8 × 532 / (632.8 - 532)). The synthesized phase Φ(t) is calculated to be φ1(t) - φ2(t), thus eliminating measurement errors caused by slow temperature drift in the environment. The out-of-plane displacement u_n is calculated in real time using the formula u_n = (632.8 nm / 4π) · [φ1(t) - φ1(t0)] + (Λ / 4π) · round[(Φ(t) - Φ(t0)) / 2π]. The in-plane displacement is determined by speckle patterns at adjacent time points. Figure 2 The out-of-plane and in-plane displacements of the four measurement points were obtained by cross-correlation. After interpolation using radial basis functions, a full-field deformation distribution map covering a 10mm×10mm area around the weld point was generated with a spatial resolution of 50μm and a measurement frequency of 500Hz.

[0064] After the k-th solder joint is welded, the system immediately reads the current full-field deformation distribution and the temperature field T(x,y,t) acquired by the infrared thermal imager. This data, along with the preset parameters for the next solder joint k+1 (power 1800W, speed 90mm / s, solder joint coordinates), is input into the physical information neural network deployed on an embedded AI computing board. This board uses NVIDIA Jetson OrinNX, and the network structure is as follows: both the temperature field encoding branch and the deformation field encoding branch consist of four 3×3 convolutional layers with 16, 32, 64, and 128 channels respectively. After batch normalization and ReLU activation, the layers are flattened, concatenated with the welding parameter vector, and input into a single-layer LSTM with 256 hidden units. The network outputs a solder joint offset prediction vector. and the predicted confidence interval. The hybrid loss function for network training is set as follows: ,in The mean square error between the network-predicted offset and the measured values ​​of the two-wavelength interferometry is given. The mean square of the residuals of the heat conduction equation, The residuals represent the relationship between deformation rate and thermoelasticity. When the current welding stage is in the middle, the weighting coefficients α=0.6, β=0.3, and γ=0.1, and are dynamically adjusted. The network forward inference time is approximately 8ms.

[0065] Assuming the reference coordinates of the (k+1)th solder joint are (15.20, 12.50, 0.00), the network predicts the offset vector as follows: =0.12mm, =0.08mm, =-0.05mm, then the target correction position is (15.32, 12.58, -0.05). At this time, the current position of the laser focus read from the galvanometer encoder and the Z-axis position sensor is (15.10, 12.55, 0.02), and the total compensation amount is calculated. =-0.07mm, given focal length =150mm, through the calibrated coordinate transformation function , , The correction amount for the X-angle deflection of the galvanometer is obtained. =1.46mrad, Y-angle correction =0.20mrad, Z-axis focusing correction amount =-0.07mm.

[0066] Before the compensation amount is sent to the actuator, it first undergoes a phase lead compensation stage. The transfer function of the compensator is... =1.3×(0.006s+1) / (0.25×0.006s+1), corresponding to a time constant τ=6ms and a lead factor α=0.25. Discretized by bilinear transformation, it becomes y[n]=b0·x[n]+b1·x[n-1]-a1·y[n-1], with coefficients b0=1.89, b1=-1.72, and a1=-0.35. The filter operates under a 1kHz control period, outputting smooth and phase-leading angle and focus signals. The half-width of the predicted confidence interval for this solder joint is 0.04mm, which does not exceed the set threshold of 0.06mm, therefore gain reduction is not triggered, and only conventional lead compensation is performed. If the confidence interval of a solder joint is too large, the gain Kc is automatically reduced to 0.8, and cubic spline interpolation is performed using the compensation amounts of the first three solder joints to limit the galvanometer angular velocity change rate to no more than 200mrad / s.

[0067] The compensated commands drive the 2D galvanometer and Z-axis focusing motor, ensuring the laser focus precisely lands at the corrected weld point position. The weld width is 1.2mm, the depth is 0.4mm, and the appearance is uniform. Simultaneously, the acquired temperature field sequence and dual-wavelength interference deformation data are cached in memory. When weld point k+1 is completed, the system immediately initiates an incremental training: freezing all layers of the convolutional encoder and performing stochastic gradient descent updates for 5 epochs only on the input gate, forget gate, and output gate weight matrices of the LSTM layer, as well as the fully connected weights of the output layer. The learning rate is 1×10^-4, the batch size is 1, and the total time is approximately 35ms, completely hidden within the interval when the laser moves to the next weld point.

[0068] As welding progressed from the first weld point to the 32nd, the physical information neural network continuously absorbed new data, and the prediction error gradually converged from the initial 0.12mm to within 0.04mm of the final weld point. Post-weld inspection of the entire welding production line showed that the positional deviation of all weld points was controlled within ±0.05mm, with no defects such as incomplete welds or off-center welds, and the welding qualification rate reached 99.8%.

[0069] The above are all preferred embodiments of the present invention and are not intended to limit the scope of protection of the present invention. Therefore, all equivalent changes made in accordance with the structure, shape and principle of the present invention should be covered within the scope of protection of the present invention.

Claims

1. An adaptive compensation method for laser welding deformation of lithium battery module busbars, characterized in that, Includes the following steps: Step 1: Obtain point cloud data of the position coordinates of all busbars to be welded and the surface morphology of the busbars, construct a pre-welding reference 3D digital model, and define reference position coordinates and reference normal vector for each weld point; Step 2: Welding is performed sequentially according to the preset welding point sequence. During the welding of each welding point and within the preset acquisition window after welding, the temperature field distribution data of the welding area, the three-dimensional morphology data of the busbar surface, and the weld position offset data are collected simultaneously. Step 3: Using a dual-wavelength laser interferometry device, two lasers with different wavelengths are emitted onto the busbar surface to obtain the full-field deformation distribution, including out-of-plane and in-plane displacements, in real time. Step 4: Input the current temperature field, the current full-field deformation distribution, and the preset welding parameters of the next weld point into the embedded physical information neural network. The loss function of the embedded physical information neural network includes the physical constraint term of the partial differential equation of heat conduction and the thermo-elastic-plastic temporal consistency constraint term of the deformation field. It predicts the weld point offset prediction vector and prediction confidence interval of the next weld point online. Step 5: Calculate the target position based on the weld point offset prediction vector and the pre-welding reference position coordinates, and obtain the three-dimensional trajectory compensation amount by combining the current position of the laser focus, and decompose it into the galvanometer deflection angle correction amount and the focusing correction amount. Step 6: Perform phase advance compensation on the correction amount, and perform trajectory smoothing transition when the prediction confidence interval exceeds the threshold, and generate control commands to drive the laser welding actuator. Step 7: After each solder joint is completed, the physical information neural network is incrementally trained using the newly added sensor data. This process is repeated until all solder joints are completed.

2. The adaptive compensation method for laser welding deformation of lithium battery module busbars according to claim 1, characterized in that, Step 3 includes the following sub-steps: Step 31, the wavelengths are respectively and Two laser beams are projected onto a preset measurement point array on the surface of the busbar through the same optical path, and interference signals are formed by reflection. Step 32: Perform Hilbert transform and phase unwrapping on the interference signal for each wavelength to obtain the phase signal. and The synthesized wavelength and synthesized phase are calculated, and the interference of ambient temperature fluctuations is eliminated by using the synthesized wavelength correction term. Step 33: Calculate the out-of-plane displacement based on the phase signal, and obtain the in-plane displacement by calculating the two-dimensional cross-correlation of the speckle patterns at adjacent time points; Step 34: The displacement of discrete measurement points is extended to the full-field deformation distribution of the welding area by radial basis function interpolation.

3. The adaptive compensation method for laser welding deformation of lithium battery module busbars according to claim 2, characterized in that, The physical information neural network consists of a deep convolutional neural network encoder and a long short-term memory network cascaded together. The deep convolutional neural network encoder includes a temperature field encoding branch and a deformation field encoding branch, which extract the spatial features of the temperature field and the deformation field, respectively. The long short-term memory network receives the spliced ​​vector of the spatial features and the welding parameters of the next weld point, learns the temporal evolution features, and outputs the weld point offset prediction vector.

4. The adaptive compensation method for laser welding deformation of lithium battery module busbars according to claim 3, characterized in that, The loss function of the embedded physical information neural network is: ； in It is the total loss value of the neural network. The mean square error between the network-predicted deformation and the measured deformation by dual-wavelength laser interferometry; Physical constraint error loss in the heat conduction equation The loss is the deformation time-series consistency constraint error loss, and α, β, and γ are the dynamic weighting coefficients of the three types of losses, respectively.

5. The adaptive compensation method for laser welding deformation of lithium battery module busbars according to claim 4, characterized in that, Step 5, decomposing the three-dimensional trajectory compensation into galvanometer deflection angle correction and focusing correction includes the following sub-steps: Step 51, calculate the target correction position. ; Step 52, read the current position of the laser focus. The total compensation amount is obtained. ; Step 53: Based on the optical parameters, convert the total compensation amount into the galvanometer deflection angle correction amount through a coordinate transformation function. and and Z-axis focusing correction amount .

6. The adaptive compensation method for laser welding deformation of lithium battery module busbars according to claim 5, characterized in that, In step 6, the transfer function for phase lead compensation is: ;in This is the transfer function for lead compensation in the complex frequency domain. This is the gain coefficient. It is a time constant. Let Laplace be a complex variable. It is a leading factor; After being discretized into difference equations by bilinear transformation, the correction sequence is subjected to recursive filtering.

7. The adaptive compensation method for laser welding deformation of lithium battery module busbars according to claim 6, characterized in that, The incremental training method is as follows: after each weld joint is completed, the weights of the long short-term memory network layer and the output layer in the physical information neural network that are related to the latest data are updated using the newly added temperature field of the weld joint, the measured deformation amount of dual-wavelength laser interferometry, and the welding process data.

8. The adaptive compensation method for laser welding deformation of lithium battery module busbars according to claim 7, characterized in that, Simultaneous acquisition of multi-source sensor data includes: an infrared thermal imager acquiring the temperature field matrix T(x,y,t) at a frame rate of no less than 100Hz; a line laser 3D contour sensor acquiring the three-dimensional topographic point cloud at a line scan frequency of no less than 2kHz; and a coaxial vision camera acquiring solder joint images and detecting the actual offset relative to the reference position. All data is timestamped using a unified clock source.

9. The adaptive compensation method for laser welding deformation of lithium battery module busbars according to claim 8, characterized in that, The dual-wavelength laser interferometry device sets up a preset measurement point array on the busbar surface. Centered on the current welding joint, at least four measurement points are arranged within the radius of the heat-affected zone, with each measurement point distributed in a radial direction.

10. An adaptive compensation system for laser welding deformation of lithium battery module busbars, characterized in that, The system for implementing the adaptive compensation method for laser welding deformation of lithium battery module busbar as described in claim 9 includes a multi-source sensor acquisition unit, a welding deformation digital twin prediction unit, a dynamic trajectory adaptive compensation unit, and a laser welding execution unit. The multi-source sensing acquisition unit includes an infrared thermal imager, a line laser 3D contour sensor, a coaxial vision camera, and a dual-wavelength laser interferometry device, which are used to simultaneously acquire the temperature field, three-dimensional morphology, weld position offset, and real-time full-field deformation distribution of the welding area, and align them with a unified clock. The welding deformation digital twin prediction unit has an embedded physical information neural network built in it. It receives the time series data of the multi-source sensing acquisition subsystem, introduces heat conduction physical constraints and deformation time series consistency constraints into the loss function, and predicts the weld point offset prediction vector and confidence interval of the next weld point online. The dynamic trajectory adaptive compensation unit receives the weld point offset prediction vector, calculates the three-dimensional trajectory compensation amount and decomposes it into the galvanometer deflection angle correction amount and the focusing correction amount. After phase advance compensation and smooth transition processing, it generates control commands. The laser welding execution unit includes a fiber laser, a galvanometer scanning head, and a Z-axis focusing mechanism, which execute control commands. After each weld point is completed, the welding deformation digital twin prediction unit uses newly added sensor data to incrementally train the physical information neural network, forming a closed-loop iterative optimization.

Images