Femtosecond adhesive-film-free surface adaptive welding system and welding method

Abstract

The invention discloses a femtosecond adhesive-film-free surface adaptive welding system and method, and relates to the technical field of laser welding. According to the system, through surface pollution multi-dimensional detection, dynamic seam prediction tracking, multi-area welding parameter self-adaptive distribution and welding effect real-time calibration, the problems of poor welding quality, dynamic seam tracking lag and insufficient multi-area welding parameter adaptation caused by surface pollution in existing femtosecond glue-film-free welding are solved. Pollution sensing, dynamic tracking and a parameter self-adaptive algorithm are combined, precise treatment of polluted areas, dynamic seam real-time following and multi-area high-quality welding are achieved, and the welding quality stability, the dynamic welding efficiency and the multi-area adaptability are remarkably improved.

Description

TECHNICAL FIELD

[0001] The application discloses a femtosecond glue-free film surface adaptive welding system and a welding method, and relates to the technical field of laser welding. BACKGROUND

[0002] The existing femtosecond glue-free film welding technology has the following specific problems:

[0003] Surface pollution leads to poor welding quality: the trace oil stain (thickness ≥ 0.5 μm), oxidation layer (thickness ≥ 1 μm) or dust particle (diameter ≥ 5 μm) existing on the film surface can cause the welding energy reflectivity to increase by more than 30%, and the welding strength of the polluted area to decrease by 40% (for example, the peeling strength of the PET film surface oil stain area after welding is only 6 N / cm), and the existing equipment cannot identify such pollution, which is easy to cause batch quality hidden trouble.

[0004] Dynamic joint tracking lag: in the roll-to-roll continuous welding scene (such as flexible display panel film welding), when the joint moves at a speed of more than 50 mm / s with the conveying mechanism, the tracking error of the existing tracking system exceeds ± 15 μm due to the response delay (≥ 20 ms), which causes the welding track to deviate from the joint and appear continuous segment virtual welding (the length of the virtual welding segment is more than 10 mm).

[0005] Insufficient multi-region welding parameter adaptation: for the film with differential structure (such as composite film with reinforcing ribs, local thickened tab film), the existing equipment adopts uniform welding parameters, which cannot adapt to the thickness (difference ≥ 10 μm) and material distribution of different regions, resulting in insufficient welding strength of the thick region (such as the reinforcing rib area peeling strength is less than 8 N / cm) or overheating of the thin region (such as the non-reinforcing rib area deformation amount is more than 3%). SUMMARY

[0006] The application aims to solve the problems of poor welding quality caused by surface pollution, dynamic joint tracking lag and insufficient multi-region welding parameter adaptation in the existing femtosecond glue-free film welding technology.

[0007] The application provides a femtosecond glue-free film surface adaptive welding system and a welding method, which realizes technical breakthrough through the following four innovations:

[0008] Surface pollution multi-dimensional detection and pretreatment module: integrate hyperspectral imager (400-1000 nm) and ultrasonic detector (50 MHz) to collect the spectral reflectivity (pollution area characteristic peak) and ultrasonic echo signal (pollution layer thickness) of the film surface; determine the pollution type, position and thickness through the pollution identification algorithm, and trigger the pretreatment (such as low-energy laser cleaning), and the pollution residual rate after cleaning is ≤ 5%.

[0009] Dynamic seam prediction and tracking algorithm: Based on the transmission speed (v) and acceleration (a) collected by the encoder, combined with the real-time position (x0) of the vision sensor, the seam position at time t is predicted by formula (1), and the tracking trajectory is generated in advance. The tracking response delay is ≤5ms.

[0010]

[0011] (Where, x(t) is the predicted position at time t, k is the correction coefficient, and e(t) is the tracking error at the previous time.)

[0012] Multi-region welding parameter adaptive allocation module: Identifies differentiated regions of the membrane material (such as stiffened / non-stiffened areas) through structured light scanning, and obtains the thickness (h) of each region. i ), material ratio (c i The regional energy coefficient is calculated using formula (2), and the welding energy is dynamically distributed:

[0013] E i =E b ×(1+α×h i +β×c i )

[0014] (where E) i Let E be the energy density of the i-th region. b (Base energy density, α is the thickness correction factor, β is the material correction factor)

[0015] Real-time welding effect calibration system: During the welding process, the temperature field of the area is collected by an infrared thermal imager (temperature difference ≤2℃) and the penetration depth is detected by a laser confocal instrument (accuracy ±1μm). The results are compared with the preset standard, and the energy parameters and tracking trajectory are dynamically corrected to form a closed loop of "contamination detection - dynamic tracking - parameter allocation - effect calibration".

[0016] Beneficial effects:

[0017] The welding quality of surface contaminated areas has been significantly improved, and the strength reduction rate caused by contamination has been controlled within 10%. The peel strength of contaminated areas of materials such as PET film has been improved.

[0018] The accuracy of dynamic seam tracking is improved, the response delay is reduced, the length of the poor weld section in roll-to-roll welding is reduced, and the efficiency of continuous welding is improved.

[0019] The adaptability of multi-area welding is enhanced, the peel strength of the thick area of ​​the reinforcing rib is improved, and the deformation of the thin area is controlled within 1%.

[0020] The closed-loop calibration system enhances the equipment's adaptability to changes in membrane surface condition and movement speed, enabling stable welding of more than 50 types of membrane materials with differentiated structures. Attached image description:

[0021] Figure 1 Operating principle flowchart. Detailed implementation method:

[0022] Example 1: Welding scenario of PET film and conductive cloth in consumer electronics (surface oil treatment)

[0023] Step 1: Place a PET film (25μm thick) with localized oil stains (0.3mm in diameter) and a conductive cloth on the welding platform. The surface contamination detection module collects the spectrum (650nm characteristic peak of oil stains) and ultrasonic signals (0.8μm thickness of oil stains) to identify the contaminated area.

[0024] Step 2: Trigger low-energy laser cleaning (energy 0.3J / cm²) 2 The residual contamination rate after cleaning was 3%.

[0025] Step 3: The dynamic seam tracking module (static scene) locates the seam, and the multi-region parameter module identifies the indifferent area, generating the energy parameter (reference energy 1.2J / cm²) through formula (2). 2 );

[0026] Step 4: Start welding and calibrate the system in real time to monitor the temperature field (up to 120℃) and penetration depth (5μm);

[0027] Step 5: Welding completed. Test the peel strength of the oil-stained area to 16 N / cm, with a difference of ≤5% between the oil-stained area and the clean area.

[0028] Example 2: Roll-to-roll flexible display panel PI film welding scenario (dynamic tracking)

[0029] Step 1: The roll-to-roll conveyor conveys the PI film (20μm thick) at a speed of 80mm / s. The dynamic seam prediction and tracking module obtains the speed (80mm / s) and acceleration (0) through the encoder, and the vision sensor collects the initial position.

[0030] Step 2: Predict the seam position using formula (1) (correction coefficient k = 0.2) and generate the tracking trajectory in advance (tracking error ±4μm);

[0031] Step 3: The multi-region parameter module identifies indifferent regions and generates a reference energy of 1.5 J / cm². 2 ;

[0032] Step 4: During welding, the conveyor speed is briefly increased to 90 mm / s, and the system updates the prediction parameters in real time, with the tracking error maintained within ±5 μm;

[0033] Step 5: After continuous welding for 10m, the total length of the incomplete weld section was found to be 0.3m, which is 92% shorter than that of existing technologies.

[0034] Example 3: Power Battery Composite Film-Electrical Tab Welding Scenario (Multi-Region Adaptation)

[0035] Step 1: Place the composite membrane with reinforcing ribs (thickness 40μm) and the tabs on the platform, and use structured light scanning to identify the reinforcing rib area (accounting for 30%) and the substrate area.

[0036] Step 2: The multi-region parameter module calculates the energy of the stiffener region E1 using formula (2) as follows: E1 = 1.8 J / cm² 2 (Thickness correction α = 0.02), base region E2 = 1.3 J / cm 2 ;

[0037] Step 3: Dynamically track (statically) locate the joint and start welding;

[0038] Step 4: The real-time calibration system detected that the penetration depth in the reinforcing rib area was slightly low, and the energy was corrected to 1.9 J / cm. 2 ;

[0039] Step 5: Welding completed. Peel strength in the reinforcing rib area is 19 N / cm, and deformation in the base area is 0.8%.

[0040] Example 4

[0041] Intelligent sensing algorithms for surface contamination: From "fuzzy detection" to "precision diagnosis"

[0042] 1. Algorithm Principle: Contaminated "Fingerprint Recognition" Based on Multi-Source Feature Fusion

[0043] The core innovation of this algorithm lies in breaking through the limitations of a single detection dimension and constructing a dual-modal recognition model of "spectral features + ultrasound features," much like creating a "fingerprint profile" for contamination, enabling accurate judgment of the type, location, and characteristics of contamination. Its modeling logic originates from the "multi-indicator comprehensive judgment" approach in medical diagnosis, eliminating interference factors and improving recognition accuracy through complementary verification of data from different dimensions.

[0044] The modeling process is divided into three stages: feature extraction, model training, and decision output. In the feature extraction stage, a hyperspectral imager and a high-frequency ultrasonic detector work synchronously: the hyperspectral imager covers the visible to near-infrared bands, capturing the unique spectral responses of different pollutants—oil stains form absorption peaks in specific bands, oxide layers exhibit characteristic reflection peaks, while the spectral curves of clean membrane materials are smooth and regular; the ultrasonic device emits high-frequency sound waves and utilizes the difference in acoustic impedance between the pollutant and the membrane substrate to obtain changes in the echo signal, thereby distinguishing the physical state of the pollutant (liquid oil stains, solid dust). The sampling frequencies of the two devices are kept synchronized to ensure that the spectral and ultrasonic data at each detection point can be accurately matched, forming a "spectral-ultrasonic" feature pair.

[0045] The model training phase employs an architecture combining Convolutional Neural Networks (CNNs) and Support Vector Machines (SVMs) from deep learning. First, CNNs compress the high-dimensional spectral data, automatically extracting key information such as the position, intensity, and shape of spectral peaks, and filtering out environmental interference such as changes in illumination. Simultaneously, traditional signal processing algorithms denoise the ultrasonic data, extracting feature parameters such as echo delay and amplitude. The two feature vectors are then fused and input into the SVM classification model for training. Through a large amount of sample data with different pollution types and levels, the model learns the correspondence between "feature combinations" and "pollution types." For example, when the spectral data shows a 650nm absorption peak and the ultrasonic echo presents a continuous and uniform signal, the model can identify it as liquid oil; when the spectrum has no obvious features but the ultrasonic echo shows discrete pulses, it is identified as solid dust.

[0046] The decision-making output stage achieves a closed loop of "identification-solution-execution". The model can not only determine whether pollution exists, but also output the precise coordinates of the polluted area, the type of pollution, and its adhesion status, thereby triggering targeted pretreatment solutions: for liquid oil stains, low-energy laser scanning cleaning is used to vaporize and evaporate the oil using laser energy; for oxide layers, plasma treatment is used to remove oxidizing substances through active particles; for dust particles, high-pressure air blowing combined with vacuum adsorption is used for removal. This "on-demand" pretreatment logic avoids the blindness of traditional single cleaning methods, ensuring that the membrane surface is not damaged while removing pollution.

[0047] 2. Solution process: Real-time computation from feature matching to solution generation

[0048] The algorithm's solution process is a pipeline operation of "data input - feature fusion - classification decision - solution output". When the membrane material enters the detection area, hyperspectral and ultrasonic equipment simultaneously acquire data. The data first undergoes noise reduction and standardization via a preprocessing module. Then, the spectral data is input into a CNN network, where convolution and pooling layers extract a 512-dimensional feature vector. The ultrasonic data undergoes wavelet transform to extract a 64-dimensional feature vector. The feature fusion module concatenates the two vectors into a 576-dimensional comprehensive feature, which is then input into a trained SVM model. The model maps the features to a high-dimensional space using kernel functions, matches them with existing pollution feature templates, and outputs the classification result and confidence score. If the confidence score is higher than 95%, the corresponding preprocessing scheme is triggered directly; if it is lower than 95%, a secondary detection is initiated to ensure reliable identification results. The entire solution process takes only milliseconds, fully meeting the requirements of pipeline production.

[0049] 3. Efficiency Enhancement Principle: Preventing quality risks caused by pollution at the source.

[0050] The algorithm's synergistic value lies in two aspects: "precise identification to reduce ineffective operations" and "targeted treatment to improve welding quality." In traditional processes, the inability to identify hidden contaminants leads to either insufficient cleaning resulting in decreased welding strength or excessive cleaning causing damage to the membrane material. This algorithm, through precise identification and on-demand treatment, keeps the contaminant residue rate extremely low, restores the reflectivity of laser energy to normal levels, and significantly reduces the difference in welding strength between contaminated and clean areas. From a production efficiency perspective, early contaminant removal avoids rework and scrap due to quality issues, reducing raw material waste. From a quality stability perspective, the welding defect rate caused by contamination is significantly reduced, and the product qualification rate is greatly improved, making it particularly suitable for the production of consumer electronics components with stringent welding quality requirements.

[0051] (II) Dynamic Joint Look-Ahead Control Algorithm: From "Passive Following" to "Active Leading"

[0052] 1. Algorithm Principle: A Collaborative Strategy of Kinematic Prediction and Visual Correction

[0053] To address the challenge of tracking the high-speed movement of seams in roll-to-roll welding, this algorithm completely abandons the traditional passive mode of "real-time acquisition and immediate response," and innovatively adopts an active control logic of "look-ahead prediction + real-time correction." The core of this approach is to construct a seam position prediction model based on the motion state, allowing the laser head to "wait in advance" for the seam to arrive, rather than "chase" it. Its modeling inspiration comes from intelligent navigation in transportation systems, which plans the route by predicting the target location to achieve precise docking.

[0054] The modeling process is based on kinematic principles and incorporates error compensation based on actual industrial scenarios. First, a high-precision encoder collects the motion parameters of the conveyor mechanism in real time, including instantaneous velocity and acceleration. These two parameters directly determine the positional change trend of the seam. Simultaneously, a vision sensor periodically collects the current position of the seam as a baseline reference point for the prediction model. Based on classical kinematic theory, the algorithm can initially calculate the theoretical position of the seam at a future moment. However, interference factors in industrial production, such as mechanical vibration, frictional resistance, and temperature drift, can cause deviations between the theoretical and actual positions. Therefore, a dynamic correction mechanism must be introduced.

[0055] The core of the correction mechanism is to establish an error compensation model and optimize the correction coefficient in real time using machine learning algorithms. During the initial debugging phase, the system collects a large amount of deviation data between the theoretical predicted position and the actual position under different motion states to build a sample library. A neural network model is trained using gradient descent, enabling the model to output the most suitable correction coefficient based on the current velocity and acceleration values. For example, when the transmission speed suddenly increases, mechanical inertia will cause the actual position of the seam to lag behind the theoretical position; the model will automatically increase the correction coefficient to make the predicted position closer to reality. When the speed is stable, the correction coefficient remains within a fixed range to ensure prediction accuracy. In this way, the predicted position output by the algorithm incorporates both kinematic theoretical calculations and error compensation from the actual scene, achieving a precise combination of "theory + practice."

[0056] At the execution level, the algorithm converts the predicted position information into motion trajectory commands for the laser head. A servo motor drives the laser head to move to the target position in advance, completing "position presetting." When the seam arrives according to the predicted trajectory, the laser head is already in precise alignment, and welding immediately begins, completely eliminating the response delay of traditional systems. Simultaneously, the vision sensor continuously collects the actual position and compares it with the predicted position. If a slight deviation occurs, the laser head position is fine-tuned in real time, forming a dynamic closed loop of "prediction-execution-correction," ensuring that tracking accuracy remains stable even under fluctuating speed conditions.

[0057] 2. Solution process: Millisecond-level calculations from parameter acquisition to trajectory generation

[0058] The algorithm's solution process is supported by high-speed data processing and consists of four steps: parameter acquisition, position prediction, trajectory generation, and real-time correction. The encoder acquires velocity and acceleration data at extremely high frequencies to ensure the capture of every subtle motion change; the vision sensor acquires the current position at fixed intervals as a prediction benchmark. During the solution process, the velocity, acceleration, and current position are first input into the kinematic model to calculate the theoretical predicted position; then, the velocity and acceleration parameters are input into the error compensation model to obtain correction coefficients; finally, the theoretical position and correction coefficients are combined to output the final predicted position. Based on the predicted position, the trajectory planning module generates the laser head's motion path, which is then converted into control signals for the servo motor, driving the laser head to move. During the welding process, the vision sensor continuously samples the actual position and compares it with the predicted position. If the deviation exceeds a threshold, a fine-tuning command is triggered. The entire solution and execution process is completed in milliseconds, perfectly matching the requirements of high-speed transmission.

[0059] 3. Efficiency Enhancement Principle: Ensuring Stable Welding by Overcoming Speed ​​Bottlenecks

[0060] The core value of this algorithm lies in transforming the "speed-precision" contradiction in dynamic welding into a "speed-precision" synergy. In traditional tracking systems, increased speed inevitably leads to decreased precision. However, this algorithm, through look-ahead prediction, keeps the relative motion between the laser head and the seam stable, ensuring that tracking precision does not decrease even in high-speed conveying scenarios. This directly brings two production benefits: first, a significant reduction in defects such as incomplete welds and missed welds, a significant increase in the effective length of continuous welding, and a lower product scrap rate; second, it eliminates the need to reduce conveying speed to maintain precision, thus improving production efficiency. It is particularly suitable for large-scale roll-to-roll production scenarios such as flexible display panels, enabling the improvement of product quality while meeting mass production requirements.

[0061] (III) Multi-region parameter biomimetic allocation algorithm: From "one-size-fits-all" to "precision drip irrigation"

[0062] 1. Algorithm Principle: Simulating the energy distribution logic of human blood circulation

[0063] To address the challenges of welding membrane materials with differentiated structures, this algorithm draws inspiration from the biomimetic logic of the human circulatory system's "on-demand blood supply," constructing a dynamic model of "region identification - demand assessment - energy allocation" to achieve precise delivery of welding energy. Its core innovation lies in treating different regions of the membrane material as individuals with varying "energy needs," allocating corresponding welding energy based on the structural characteristics of each region, rather than using a uniform standard.

[0064] The modeling process is divided into two core stages: regional feature extraction and energy demand modeling. In the regional feature extraction stage, structured light 3D scanning technology is used to perform a full-area scan of the membrane material to acquire three-dimensional morphological data. Through image segmentation algorithms, different structural regions such as reinforcing rib areas, substrate areas, and edge areas are automatically identified, and the boundary coordinates of each region are marked. Simultaneously, combined with a material database, the material composition of each region is determined—for example, the reinforcing rib area of ​​the power battery composite membrane is mainly composed of high-strength PET, while the substrate area is a composite structure of PET and metal foil. Different materials exhibit different absorption and conduction characteristics of laser energy, which is an important basis for energy distribution.

[0065] The energy demand modeling stage is the core of the algorithm. By constructing a mapping relationship between "structure-material-energy," it achieves accurate energy calculation. The algorithm uses the welding energy of a clean, uniformly thick membrane material as the baseline value and introduces a thickness correction factor and a material correction factor. The thickness correction factor is positively correlated with the thickness of the region; thicker regions require higher energy to ensure the required weld depth. The material correction factor is determined based on the laser absorption efficiency of the material. Materials with low absorption efficiency require appropriately increased energy, while materials with high absorption efficiency require reduced energy to avoid overheating. For example, the reinforcing rib area has a large thickness and low material absorption efficiency, so both correction factors are greater than 1, resulting in an energy allocation higher than the baseline value. Conversely, the base area has a small thickness and contains metal foil (high absorption efficiency), so the correction factor is less than 1, resulting in an energy allocation lower than the baseline value.

[0066] To ensure model accuracy, extensive orthogonal experiments are conducted during the initial debugging phase. By varying the thickness, material, and welding energy of different regions, welding strength and deformation are tested, establishing a sample library of "parameter combinations - welding effects." A genetic algorithm is then used to optimize the model parameters, enabling it to output optimal energy parameters based on regional characteristics. During welding, as the laser head moves to different regions, the energy modulator adjusts the laser energy in real time according to the parameters output by the algorithm, achieving seamless integration of "region switching - synchronous energy adjustment."

[0067] 2. Solution Process: Dynamic Matching from Region Identification to Energy Output

[0068] The algorithm's solution process is a real-time workflow of "scanning-recognition-calculation-execution". Once the membrane material enters the welding area, the structured light scanning device first completes a full-area scan, inputting the 3D data into the region recognition module. Through threshold segmentation and contour extraction algorithms, the coordinates and boundaries of each differentiated region are automatically marked. The feature extraction module extracts key parameters such as thickness and material proportion from the recognition results for each region. These parameters are input into the energy demand model, combined with a baseline energy value, to calculate the target energy parameters for each region. The energy modulator receives parameter commands and adjusts the energy output in real-time as the laser head moves. Energy increases when the laser head enters the reinforcing rib area and decreases when it enters the base area, ensuring that the energy of each region precisely matches its requirements. The entire solution process is performed synchronously with the welding process, requiring no additional preprocessing time and not affecting production efficiency.

[0069] 3. Enhancement principle: Multi-regional mass balance that takes into account both strength and deformation.

[0070] The algorithm's synergistic value lies in resolving the contradiction that traditional "one-size-fits-all" parameters cannot simultaneously address the quality issues of multiple regions, achieving the dual goals of "high strength in thick areas and low deformation in thin areas." For thick areas such as reinforcing ribs, sufficient energy ensures weld penetration and peel strength, avoiding incomplete welds caused by insufficient energy. For thin areas such as the substrate, precisely controlled low energy prevents overheating deformation and ablation, ensuring the dimensional accuracy of the membrane material. From a production perspective, there is no need for frequent shutdowns to adjust parameters to adapt to different regions, significantly shortening changeover time, making it particularly suitable for small-batch, multi-specification customized production scenarios. From a quality perspective, the welding quality of each region reaches its optimal state, significantly improving product qualification rate and consistency.

[0071] (iv) Welding effect self-learning calibration algorithm: from "single execution" to "continuous optimization"

[0072] 1. Algorithm Principle: The Self-Evolutionary Logic of a Multi-Dimensional Feedback System

[0073] As the "intelligent hub" of the entire welding system, this algorithm dynamically corrects the parameter settings of preceding modules by collecting multi-dimensional data in real time during the welding process, comparing it with preset standards, and simultaneously achieving self-learning optimization of system performance. Its core innovation lies in breaking through the limitations of traditional "open-loop execution" and constructing a fully closed-loop control system of "process monitoring - deviation identification - parameter correction - model optimization," enabling the system to adapt to changes in membrane material condition and production environment, and achieve continuous and stable welding quality.

[0074] The modeling process is based on feedback control theory, combined with the self-learning capabilities of machine learning. The data acquisition layer integrates an infrared thermal imager and a laser confocal microscope to achieve comprehensive monitoring of the welding process: the infrared thermal imager captures the temperature distribution of the welding area in real time; excessively high temperatures indicate excess energy and a risk of deformation, while excessively low temperatures indicate insufficient energy, potentially leading to insufficient strength. The laser confocal microscope precisely detects the weld penetration depth, a direct indicator of weld strength; shallow penetration indicates insufficient strength, while excessive penetration can easily penetrate the membrane material. The monitoring data from both devices is synchronized with the welding process, ensuring that each welding point has corresponding process data support.

[0075] The deviation identification and parameter correction module is the core execution unit of the algorithm. The system presets standard ranges for temperature and melt depth for different film materials and regions, compares real-time monitoring data with standard values, and calculates the deviation. Based on the magnitude and direction of the deviation, a parameter correction command is generated through a PID (proportional-integral-derivative) control algorithm: if the temperature in a certain region is too high and the melt depth is too large, it indicates excess energy, and an energy reduction command is sent to the parameter allocation module; if the temperature is too low and the melt depth is too shallow, an energy increase command is sent; if the temperature distribution is asymmetrical, it indicates a deviation in the welding trajectory, and a trajectory fine-tuning command is sent to the tracking module.

[0076] The self-learning optimization module is an advanced innovation in the algorithm. It builds a historical database by recording the "parameter settings - process data - welding effect" of each weld. Machine learning algorithms analyze this database to uncover potential correlations between parameters and results. When a certain type of membrane material is found to frequently deviate under specific conditions, the system automatically optimizes the baseline parameters of the preceding algorithm. For example, if repeated detections of residual contamination in a batch of PET film leading to abnormal energy absorption indicate this, the system will automatically adjust the re-inspection frequency and cleaning energy parameters of the contamination detection module, continuously improving the system's adaptability as production progresses.

[0077] To verify the actual application effect of the system, this embodiment selects typical scenarios in three core areas—consumer electronics, flexible displays, and power batteries—for testing. By comparing with traditional processes, the efficiency-enhancing value of innovative technologies is quantitatively demonstrated.

[0078] The use of the word "comprising" and variations thereof in this specification and claims does not limit the claimed invention to excluding any variations or additions.

[0079] Modifications and improvements to this invention will be readily apparent to those skilled in the art. Such modifications and improvements are within the scope of this invention.

Claims

1. A femtosecond adhesive-free surface-adaptive welding method, characterized in that, include: (1) Collect the spectral reflectance and ultrasonic echo signals of the membrane surface using a hyperspectral imager and an ultrasonic detector to identify the type, location and thickness of surface contamination; (2) Pre-treat the contaminated area and predict the seam trajectory by combining the motion parameters of the conveying mechanism and the real-time position of the seam. (3) Identify the differentiated areas of the membrane material using formula E. i =E b ×(1+α×h i +β×c i Calculate the energy density of each region; (4) Welding is performed based on the predicted trajectory, and energy parameters are calibrated in real time by detecting the temperature field and penetration depth; (5) Dynamically update the pollution status, joint location and area parameters to complete the welding.

2. The method according to claim 1, characterized in that, The prediction of the seam trajectory in step (2) includes: using the formula Calculate the position at time t, where v is the transmission speed, a is the acceleration, and e(t) is the tracking error at the previous time.

3. The method according to claim 1, characterized in that, The pretreatment mentioned in step (1) includes low-energy laser cleaning (energy 0.1-0.5 J / cm). 2 The residual contamination rate after cleaning is ≤5%.

4. The method according to claim 1, characterized in that, The real-time calibration in step (4) includes: when the detected melting depth is 10% lower than the standard value, increasing the energy density of the corresponding area by 5%-10%.

5. The method according to claim 1, characterized in that, The differentiated regions include regions with a thickness difference of ≥5μm or a material ratio difference of ≥10%.

6. A femtosecond adhesive-free surface-adaptive welding system, characterized in that, include: (1) Pollution detection and pretreatment module: integrates a hyperspectral imager and an ultrasonic detector to identify surface contaminants and perform cleaning; (2) Dynamic tracking module: acquires transmission motion parameters and seam positions using formulas. Predicted trajectory; (3) Multi-region parameter module: Identifies differentiated regions using formula E i =E b ×(1+α×h i +β×c i Calculate the energy in each region; (4) Real-time calibration module: The welding effect is detected by infrared thermal imager and laser confocal instrument, and the energy parameters are corrected; (5) Main control module: coordinates the operation of each module and outputs welding control commands.

7. The system according to claim 6, characterized in that, In the pollution detection preprocessing module, the hyperspectral imager has a spectral resolution of ≤10nm and the ultrasonic detector has a frequency of ≥50MHz.

8. The system according to claim 6, characterized in that, The dynamic tracking module has a response delay of ≤5ms and a tracking accuracy of ±5μm (when the speed is ≤100mm / s).

9. The system according to claim 6, characterized in that, The multi-region parameter module can identify at least eight different structures (including reinforcing ribs, locally thickened areas, etc.).

10. The system according to claim 6, characterized in that, The real-time calibration module has a temperature detection accuracy of ±1℃ and a melt depth detection accuracy of ±1μm.

Images