A machine vision-based cover film peel force detection method and system

Abstract

This invention discloses a machine vision-based method and system for detecting the peel force of a cover film, relating to the fields of flexible material testing and machine vision technology. The method includes: acquiring a visual image sequence of the peeling interface, macroscopic load, and tomographic scan images of the adhesive layer; calculating the depth profile using an internal volume digital image correlation algorithm to obtain the true three-dimensional strain field of the interface; extracting the microfiber topological backbone direction at the microcrack tip based on this strain field, and driving a multi-degree-of-freedom displacement stage for dynamic pose compensation to obtain a force-slip displacement curve; inputting this mechanical curve and the three-dimensional strain field into a pre-trained cross-modal deep neural network to output the corrected peel force. This invention addresses the problem of peel force measurement distortion caused by the invisible microscopic deformation of the underlying layer, dynamic drift of stress vectors, and mechanical interference from equipment in traditional peel testing.

Description

Technical Field

[0001] This invention relates to the fields of flexible material testing and machine vision technology, and more specifically, to a machine vision-based method and system for detecting the peel force of a cover film. Background Technology

[0002] The service reliability of flexible printed circuit boards and high-density interconnect devices is highly dependent on the interfacial adhesion quality of the surface capping film. As a core mechanical indicator characterizing the adhesion strength, the precise testing of capping film peel force under dynamic ultimate loads plays an irreplaceable role in evaluating the intrinsic peel resistance, micromechanical evolution process, and physical stability of front-end manufacturing processes of flexible electronic components.

[0003] Currently, the mainstream technology for testing the peel force of cover films mainly relies on a combination of a universal testing machine and a high-speed global exposure camera. This approach uses the tensile sensor of the testing machine to acquire macroscopic force-displacement curves, while simultaneously using an industrial camera with high-frequency, high-power strobe illumination to capture the apparent contours of the peel front. In terms of material deformation analysis, existing technologies typically employ two-dimensional digital image correlation (2D-DIC), relying on manually spraying speckle patterns onto the flexible film surface to track the macroscopic planar displacement field during the stretching process.

[0004] However, the aforementioned existing technologies have significant limitations in the precise characterization of micro-interfaces. First, high-frequency intense light illumination easily induces localized heat accumulation on the surface of polymer substrates, leading to high-frequency thermal stress distortion and viscoelastic property distortion in the test samples. Second, artificial speckle patterns on the surface cannot reflect the true internal state of the underlying translucent adhesive, and the inherent "shear hysteresis" effect during strain transmission between heterogeneous layers causes severe distortion in the calculation of the micro-interface strain field. Third, traditional rigid testing mechanisms can only maintain static macroscopic geometric peel angles and cannot perform multi-degree-of-freedom transient compensation for the dynamic changes in the microfiber drawing direction at the microcrack tip, which easily leads to the failure of constant micro-stress vector control and drift of mechanical characteristics. Finally, existing testing systems fragment heterogeneous features such as vision and mechanics, lack cross-modal high-dimensional reconstruction capabilities, and cannot construct holographic failure tracing logic, resulting in an insurmountable gap between quality inspection and front-end process correction. Summary of the Invention

[0005] To overcome the aforementioned deficiencies of the prior art, embodiments of the present invention provide a machine vision-based method for detecting the peel force of a cover film. This method calculates the true three-dimensional strain field of the adhesive layer by fusing internal tomographic scanning data, performs dynamic spatial pose compensation by driving a displacement stage based on the topological direction of the microcrack tip, and finally uses a cross-modal deep neural network to jointly decode and correct the mechanical and visual features. This solves the problem of peel force measurement distortion caused by the invisible strain of the underlying layer, the drift of micro-stress vectors, and mechanical interference in traditional macroscopic peel tests.

[0006] To achieve the above objectives, the present invention provides the following technical solution:

[0007] A machine vision-based method for detecting the peel force of a cover film includes the following steps: acquiring a visual image sequence of the peeling interface, a macroscopic load signal, and an internal tomographic scan image of the adhesive layer inside the cover film during the peeling process; using an internal volume digital image correlation algorithm to perform three-dimensional displacement calculation on the depth profile of the adhesive layer to obtain the true three-dimensional strain field of the interface; extracting the microfiber topology backbone direction of the microcrack tip based on the true three-dimensional strain field of the interface, driving a multi-degree-of-freedom displacement stage to perform dynamic spatial pose compensation, and obtaining a force-slip displacement curve; inputting the force-slip displacement curve and the true three-dimensional strain field of the interface into a pre-trained cross-modal deep neural network to output the corrected peel force of the cover film.

[0008] In a preferred embodiment, the step of using an internal volume digital image correlation algorithm to perform three-dimensional displacement calculation on the depth profile of the adhesive layer includes: extracting the morphological features of filler particles and micropores inside the colloid in the internal tomographic scan image as endogenous speckle markers; using the endogenous speckle markers as a three-dimensional voxel tracking reference, and tracking the dynamic three-dimensional deformation of the interface adhesive layer through the internal volume digital image correlation algorithm.

[0009] In a preferred embodiment, the step of tracking the dynamic three-dimensional deformation of the interfacial adhesive layer using an internal volume digital image correlation algorithm includes: monitoring the local strain gradient of the peeling interface during the tracking process; when the local strain gradient exceeds a preset threshold, triggering a reference state update mechanism, using the internal tomographic image voxel of the current frame as the reference benchmark for the next calculation cycle, and calculating the true yield strain of the interfacial colloid.

[0010] In a preferred embodiment, the extraction of the microfiber topological backbone direction at the microcrack tip includes: performing edge detection and three-dimensional morphological refinement on the microfiberized region at the microcrack tip corresponding to the strain gradient extreme region in the real three-dimensional strain field of the interface to extract the topological skeleton reflecting the spatial connectivity of the microfiber group; calculating the weighted average direction vector of each branch structure in the topological skeleton, and determining the weighted average direction vector as the microfiber topological backbone direction.

[0011] In a preferred embodiment, the dynamic spatial pose compensation of the multi-degree-of-freedom displacement stage includes: calculating the angle comparison difference between the microfiber topology trunk direction and a preset macroscopic peeling angle reference value, and obtaining the translational deviation of the root node at the tip of the microcrack; constructing a three-dimensional homogeneous transformation matrix as a spatial compensation matrix based on the angle comparison difference and the translational deviation; performing pose prediction compensation on the spatial compensation matrix using an unscented Kalman filter algorithm, and obtaining the target extension and retraction of each servo actuator of the multi-degree-of-freedom displacement stage through inverse kinematics calculation; and driving each servo actuator to move in tandem according to the target extension and retraction to adjust the spatial pose of the test sample.

[0012] In a preferred embodiment, the pre-trained cross-modal deep neural network is a cross-modal mask autoencoder network. The pre-training process of the cross-modal mask autoencoder network includes: performing a preset ratio of random masking on the visual features of the input real three-dimensional strain field of the interface on the time axis and spatial axis; and synchronously inputting the mechanical features of the force-slip displacement curve and the remaining unmasked visual features into the feature encoder of the cross-modal mask autoencoder network.

[0013] In a preferred embodiment, the pre-training process of the cross-modal masking autoencoder network further includes: using the mechanical features in the non-masked state as cross-modal cue words through the decoder of the cross-modal masking autoencoder network to cross-reconstruct the masked visual spatial features in the latent space; and using the reconstruction error between the reconstructed visual spatial features and the original visual features as the loss function to backpropagate and update the network parameters.

[0014] In a preferred embodiment, obtaining the visual image sequence of the stripped interface includes: using an event camera to capture dynamic changes in light intensity within the microscopic field of view of the stripped interface, and generating an asynchronous visual event stream consisting of pixel coordinates, timestamps, and polarity changes as the visual image sequence.

[0015] In a preferred embodiment, the acquisition of internal tomographic images of the adhesive layer inside the cover film during the peeling process includes: using a near-infrared sweep frequency light source to penetrate the light-transmitting substrate of the flexible printed circuit board, performing depth profile three-dimensional interferometric imaging of the peeling process of the adhesive layer, and generating internal tomographic images.

[0016] This invention provides a machine vision-based capping film peel force detection system, comprising: a data acquisition module for acquiring visual image sequences of the peeling interface, macroscopic load signals, and internal tomographic scan images of the adhesive layer inside the capping film during the peeling process; a strain field construction module for performing three-dimensional displacement calculation on the depth profile of the adhesive layer using an internal volume digital image correlation algorithm to obtain the true three-dimensional strain field of the interface; a pose compensation module for extracting the microfiber topology backbone direction of the microcrack tip based on the true three-dimensional strain field of the interface, driving a multi-degree-of-freedom displacement stage to perform dynamic spatial pose compensation, and obtaining a force-slip displacement curve; and a peel force correction module for inputting the force-slip displacement curve and the true three-dimensional strain field of the interface into a pre-trained cross-modal deep neural network, and outputting the corrected capping film peel force.

[0017] The technical effects and advantages of the machine vision-based capping film peel force detection method of this invention are as follows:

[0018] This invention simultaneously acquires visual image sequences, macroscopic load signals, and tomographic scan images of the adhesive layer. It then utilizes an internal volume digital image correlation algorithm to perform three-dimensional displacement calculation on the depth profile of the adhesive layer. This allows for direct and precise acquisition of the true three-dimensional strain field of the underlying interface, overcoming the limitations of traditional methods that rely solely on surface contour observation. Furthermore, based on this strain field, the invention extracts the microfiber topological backbone direction at the microcrack tip and uses this to drive a multi-degree-of-freedom displacement stage to perform dynamic spatial pose compensation. This ensures the transient and precise alignment of the microscopic force vectors during the peeling process, thereby obtaining a high-fidelity force-slip displacement curve that eliminates spatial drift errors. Finally, this mechanical curve is jointly decoded with a high-dimensional three-dimensional strain field pre-trained deep neural network through cross-modal input, effectively filtering out physical interference and distortion in single mechanical tests. The final output is a high-precision corrected capping film peeling force, achieving a high-precision leap from macroscopic mechanical measurement to microscopic physical quantification in peel strength detection. Attached Figure Description

[0019] Figure 1 This is a schematic diagram of the machine vision-based cover film peel force detection method provided in an embodiment of the present invention.

[0020] Figure 2 This is a dynamic comparison diagram of the ZNCC coefficient value changing with peeling displacement, provided for an embodiment of the present invention.

[0021] Figure 3 This is a schematic diagram comparing the dynamic evolution curves of force-slip displacement before and after the cross-modal network correction provided in an embodiment of the present invention.

[0022] Figure 4 This is a schematic diagram of a machine vision-based cover film peel force detection system provided in an embodiment of the present invention. Detailed Implementation

[0023] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those of ordinary skill in the art without creative effort are within the scope of protection of the present invention.

[0024] Example 1, Figure 1 The present invention provides a machine vision-based method for detecting the peel force of a cover film, comprising the following steps:

[0025] Step S1: Obtain the visual image sequence of the peeling interface, the macroscopic load signal, and the internal tomographic scan image of the adhesive layer inside the cover film during the peeling process.

[0026] In this embodiment, a multi-source heterogeneous dataset is first acquired during the capping film peeling process. Specifically, for the acquisition of macroscopic load signals, a miniature universal testing machine equipped with a high-frequency piezoelectric tensile sensor is used to perform the dynamic peeling action of the capping film. During the peeling test, the normal and tangential force states of the peeling interface are acquired in real time at a sampling frequency of up to 10kHz, and the high-frequency mechanical vibration noise caused by the servo motor drive of the testing machine is filtered out by a front-end anti-aliasing low-pass filter circuit. To satisfy the Nyquist sampling theorem and retain the true high-frequency abrupt change characteristics of the interface peeling force to the greatest extent, the cutoff frequency of the front-end anti-aliasing low-pass filter circuit is set in the range of 2kHz to 4kHz, preferably two-fifths of the 10kHz sampling frequency (i.e., 4kHz), thereby obtaining continuous one-dimensional macroscopic load time series data with high signal-to-noise ratio.

[0027] Furthermore, acquiring the visual image sequence of the peeling interface includes: under cold light source illumination, using an event camera to capture the dynamic changes in light intensity within the microscopic field of view of the peeling interface, generating an asynchronous visual event stream composed of pixel coordinates, timestamps, and polarity changes, and using the asynchronous visual event stream as the visual image sequence. Traditional high-speed industrial cameras based on global exposure typically require extremely high-power stroboscopic illumination sources to achieve extremely short exposure times when capturing micrometer-level fields of view. However, this high-frequency, intense light easily induces a severe "photothermal effect" on the surface of the transparent substrate or adhesive layer of the flexible printed circuit board (FPC), causing localized softening of the polymer and changes in its viscoelastic properties, thereby introducing unpredictable peeling test errors. To overcome this deficiency, this embodiment uses a low-power LED cold light source ring array to provide stable and soft constant background illumination and introduces a biomimetic dynamic visual sensor (i.e., an event camera). This sensor employs an asynchronous addressing mechanism for independent pixels. Instead of outputting traditional absolute brightness image frames, it triggers an electrical signal only when a local pixel experiences a logarithmic change in light intensity exceeding a preset threshold. This completely avoids localized heat accumulation caused by high-intensity continuous illumination. The generated asynchronous visual event stream can be mathematically represented as a set of discrete spatiotemporal events, expressed by the following formula:

[0028] (1)

[0029] in, Represents a set of visual event streams collected within a continuous time period; Indicates the first A single visual event that is triggered; The total number of events; The spatial coordinates of the pixel that triggered the event on the two-dimensional focal plane of the event camera sensor; This is the absolute timestamp when the pixel was triggered and reached microsecond-level resolution; This represents the polarity change of the logarithmic light intensity, where +1 represents an increase in light intensity over time (ON event), and -1 represents a decrease in light intensity over time (OFF event). Furthermore, the preset threshold (contrast threshold) that triggers this polarity change in logarithmic light intensity is not fixed. Instead, it is dynamically determined by calibrating the background illuminance of the current LED cold light source ring array and the reflectivity of the cover film's translucent substrate and adhesive layer. Its setting range is typically between 10% and 15%, thus balancing the sensitivity for capturing microcrack initiation stages with the filtering capability for background noise events.

[0030] Further, the acquisition of internal tomographic images of the adhesive layer inside the cover film during the peeling process includes: using a near-infrared swept-frequency light source to penetrate the transparent substrate of the flexible printed circuit board, performing depth-section three-dimensional interferometric imaging of the adhesive layer peeling process, and generating the internal tomographic image. To achieve high-resolution penetration imaging of the internal structure of opaque or semi-transparent adhesive interfaces, a near-infrared swept-frequency laser source with a center wavelength of 1310 nm and a spectral bandwidth of 100 nm (i.e., SS-OCT swept-frequency optical coherence tomography) is selected. Near-infrared light in this specific wavelength band can penetrate the transparent substrate of flexible cover films such as polyimide (PI) without damage with extremely low optical scattering loss, reaching the bottom adhesive layer. Backscattered light from different depths inside the adhesive layer is collected through the Michelson interferometer optical path and interfered with the reference arm beam. Then, a fast Fourier transform (FFT) is performed on the collected interference spectrum signal to reconstruct an image with micron-level axial resolution (e.g., 5 nm). The volume of a three-dimensional interferometric image with depth profile.

[0031] To verify the effectiveness of the cold light source and event camera combination described in this embodiment in suppressing the "photothermal effect," a continuous 10-minute microscopic field-of-view peeling comparison test was conducted. During the test, the surface of the flexible printed circuit board (PI substrate) was irradiated using a conventional 500W high-frequency stroboscopic light source and the low-power LED cold light source array of this embodiment, respectively. The average temperature of the adhesive layer interface was extracted in real time using a high-precision infrared thermal imager. Specific data are shown in Table 1.

[0032] Table 1

[0033]

[0034] As shown in Table 1, when continuously capturing high frequencies, traditional high-frequency strong light causes the temperature of the test area to rise significantly (up to 58.6℃), which can easily change the viscoelastic properties of the polymer adhesive. However, the cold light source used in this application, combined with the asynchronous addressing mechanism of the event camera, successfully controls the temperature rise of the interface to within 0.6℃, which fundamentally ensures the stability of the physical and thermodynamic boundaries of the mechanical test.

[0035] This step, through the organic combination of high-dynamic visual perception with zero thermodynamic disturbance and high-penetration deep-penetration scanning, not only completely eliminates the interference of material thermal deformation caused by traditional high-intensity light observation, but also breaks through the limitation of traditional machine vision that can only observe the surface state of the peeled sample edge. It lays an absolutely objective and multi-dimensional spatiotemporally synchronized data foundation for the subsequent accurate calculation of the real three-dimensional strain field of the interface and micro-macro mechanical mapping.

[0036] Step S2: The internal volume digital image correlation algorithm is used to perform three-dimensional displacement calculation on the depth profile of the adhesive layer to obtain the true three-dimensional strain field of the interface.

[0037] In this embodiment, the three-dimensional displacement calculation of the depth profile of the adhesive layer using an internal volume digital image correlation algorithm includes: extracting the morphological features of filler particles and micropores inside the colloid in the internal tomographic scan image as endogenous speckle markers; using the endogenous speckle markers as a three-dimensional voxel tracking reference, and tracking the dynamic three-dimensional deformation of the interface adhesive layer through the internal volume digital image correlation algorithm. In traditional two-dimensional digital image correlation (2D-DIC) technology, speckle patterns are usually manually sprayed onto the surface of the flexible film. However, under peeling stress, there is a significant difference in the elastic modulus between the flexible substrate and the underlying adhesive layer. When the surface deformation is transmitted to the interior, a severe "shear hysteresis" effect occurs due to the viscoelastic damping of the material, causing the speckle displacement field on the surface to fail to reflect the actual stress yielding state of the deep colloid at the interface. To completely overcome this measurement boundary effect, this embodiment abandons the external spraying process and directly targets the high-resolution internal tomographic scan image (e.g., voxel resolution of 10 ... Using grayscale volume blocks, and employing 3D morphological high-pass filtering and connected component analysis, the natural optical contrast of the naturally uniformly distributed silica filler particles and the micron-sized pores formed during the coating process were extracted from the adhesive layer and used as markers for endogenous speckle. These endogenous microstructures are completely bonded to the polymer matrix of the adhesive layer without any relative slip and can undergo real synchronous physical deformation with the material. By employing the 3D normalized zero-mean cross-correlation criterion (3D-ZNCC), the optimal matching position of the subset of endogenous speckle voxels was searched between consecutive time-frame scan volumes.

[0038] To balance spatial resolution and matching accuracy, the three-dimensional size of the endogenous speckle voxel subset is set as follows: The analysis domain of the adhesive layer is uniformly meshed in three dimensions with a set voxel step size. After obtaining the initial value of the integer pixel displacement based on the 3D-ZNCC criterion, a three-dimensional first-order inverse combined Gauss-Newton (IC-GN) algorithm is further adopted, combined with three-dimensional cubic B-spline interpolation to reconstruct the continuous gray-level gradient field, and sub-voxel-level nonlinear iterative optimization is performed.

[0039] This enables high-precision dynamic tracking of three-dimensional displacement vectors at the sub-voxel level at various points within the adhesive layer. Here, The set length of the single-sided voxel of the subset (e.g., 31).

[0040] Furthermore, regarding the large deformation characteristics of the flexible membrane, the step of tracking the dynamic three-dimensional deformation of the interfacial adhesive layer using an internal volume digital image correlation algorithm includes: monitoring the local strain gradient of the peeling interface during the tracking process; when the local strain gradient exceeds a preset threshold, triggering a reference state update mechanism, using the voxel of the internal tomographic image of the current frame as the reference benchmark for the next calculation cycle, and calculating the true yield strain of the interfacial colloid. Under the extreme failure condition of capping membrane peeling, the adhesive layer will undergo rapid nonlinear large tensile and microscopic tearing deformation. If the initial unpeeled state is always used as the fixed reference frame (i.e., pure Lagrange tracking), the speckle space topology in the current deformation frame will be severely distorted, resulting in a rapid decrease in the three-dimensional cross-correlation coefficient, which is very likely to cause the algorithm to "decorrelate" and lead to displacement tracking failure. Therefore, an adaptive reference state update mechanism based on local strain gradient monitoring (i.e., incremental Eulerian-Lagrange hybrid description) is introduced.

[0041] The monitoring and calculation logic of the local strain gradient is as follows: First, the Green-Lagrange strain tensor is calculated based on the three-dimensional displacement field obtained by the internal volume digital image correlation algorithm. Then, the principal strain components of the tensor are extracted and their spatial gradient is calculated.

[0042] (2)

[0043] (3)

[0044] In the above formula, denoted as Green-Lagrange strain tensor, which contains nonlinear terms to accurately describe the state of finite large deformation; This represents the calculated three-dimensional displacement vector field. Let be the gradient tensor of the displacement in space; Represents the strain tensor The maximum principal strain component extracted from the eigenvalue decomposition represents the strain value of the voxel point in space in the direction of the most intense tensile deformation. for Gradient vector in three-dimensional space; Represents the sign of partial derivatives; These are the three independent spatial dimensions of the three-dimensional Cartesian coordinate system; The scalar norm representing the calculated local strain gradient is used to quantify the abrupt changes and distortions in the microscopic spatial distribution of strain.

[0045] During real-time calculation, the maximum local strain gradient in the region at the forefront of the microcrack is continuously calculated. .when Exceeding the preset algorithm tolerance threshold (For example, set to) When the deformation approaches the mesh matching limit, the system immediately triggers a reference state update mechanism, severing the direct association with the initial zero-state frame. It then "freezes" the voxel state of the internal tomographic image effectively computed in the current frame and updates it as the initial reference for the next incremental computation cycle. Finally, the deformation gradient tensor within each incremental computation step is extracted. The total deformation gradient tensor relative to the initial zero-state frame is obtained by multiplying tensors together. Its mathematical expression is ,in Update the number of steps for the current total increment. For the first The deformation gradient tensor of the current frame relative to its local reference frame within each computation step. Based on this total deformation gradient tensor The true yield strain of the interfacial colloid throughout the entire large deformation peeling cycle was recalculated, and its calculation formula is as follows:

[0046] (4)

[0047] in, It represents the total Green-Lagrange strain tensor throughout the entire large deformation peeling cycle, which is the true yield strain of the interfacial colloid. Represents the total deformation gradient tensor The transpose tensor; This represents a third-order unit tensor.

[0048] To visually demonstrate the anti-decorrelation effect of the incremental reference state update mechanism under large deformation conditions of the flexible membrane, a data visualization processing script was used to plot the cross-correlation coefficient (ZNCC) decay comparison curve within a 20mm peeling stroke, as shown below. Figure 2 As shown.

[0049] Figure 2 This is a dynamic comparison of the ZNCC coefficient values ​​as a function of peeling displacement between traditional pure Lagrange tracing and the incremental tracing method of this application. Figure 2 As shown in the data evolution trend, when the reference state update mechanism is not enabled (dashed trajectory in the figure), with the large tearing of the peeling interface, the internal voxel mesh is severely distorted. The ZNCC coefficient quickly drops below the algorithm failure tolerance threshold (0.6) when the slip displacement reaches 4.5 mm, causing the subsequent three-dimensional displacement calculation to completely diverge and collapse. However, after adopting the update mechanism based on local strain gradient monitoring described in this embodiment (solid trajectory in the figure), whenever the strain gradient approaches the mesh matching limit, the system immediately triggers a state reset, so that the ZNCC coefficient exhibits a "sawtooth" shape throughout the entire large deformation peeling cycle, but always remains in a high confidence range above 0.85. This visualization result strongly proves that the proposed solution ensures the absolute robustness of continuous tracking of the internal strain field.

[0050] This step extracts endogenous speckle through internal tomography and performs in-situ tracking. Combined with an incremental reference state update mechanism driven by an adaptive local strain threshold, it completely overcomes the industry pain point of accurately capturing the actual deformation of the interface in the peel test of semi-transparent large deformation materials and the high degree of computational dereliction. It provides high-fidelity and continuous low-level three-dimensional strain field data support for subsequent micro-failure analysis.

[0051] Step S3: Based on the real three-dimensional strain field of the interface, extract the microfiber topology backbone direction of the microcrack tip, drive the multi-degree-of-freedom displacement stage to perform dynamic spatial pose compensation, and obtain the force-slip displacement curve.

[0052] In this embodiment, extracting the microfiber topological backbone direction at the microcrack tip includes: performing edge detection and three-dimensional morphological refinement on the microfiberized region at the microcrack tip corresponding to the extreme strain gradient region in the real three-dimensional strain field of the interface, extracting a topological skeleton reflecting the spatial connectivity of the microfiber group; calculating the weighted average direction vector of each branch structure in the topological skeleton, and determining the weighted average direction vector as the microfiber topological backbone direction. At the peeling front of the adhesive layer, the polymer material undergoes severe yielding and forms complex microfiber phenomena. First, the three-dimensional Canny edge detection operator is used to segment the boundary of the extreme region in the three-dimensional strain field where the strain gradient is significantly higher than the neighborhood background (e.g., the gradient value is greater than three times the standard deviation of the local mean), defining the microfiberized region. Subsequently, the three-dimensional medial axis transformation method is introduced to perform three-dimensional morphological refinement operation on the voxel set of this region. This algorithm employs a topology-preserving iterative stripping technique to remove non-core voxels from the surface of microfibers layer by layer until only the core connectivity centerline of a single voxel remains, thus successfully extracting a topological skeleton that characterizes the spatial connectivity and geometric morphology of the micro-fiber cluster. After obtaining this topological skeleton, branching nodes and endpoints are identified by traversing the 26 neighborhoods of the skeleton voxels, resolving the complex mesh skeleton into multiple independent linear branch structures. To eliminate pseudo-branches caused by minor local image distortions and computational noise, a branch length threshold (e.g., 10 voxels) is set, and non-main branch burrs with lengths less than this threshold are pruned and removed. For the remaining valid independent branches, principal component analysis (PCA) is used to fit the local direction vectors of each branch structure, and then a weighted average direction vector representing the overall stripping trend is calculated.

[0053] The formula for calculating the weighted average direction vector is as follows:

[0054] (5)

[0055] In the above formula, This represents the final extracted microfiber topology backbone direction vector; This represents the total number of independent branch structures retained in the topological skeleton after three-dimensional morphological refinement and extraction and removal of non-main branch burrs. Indicates by the first The three-dimensional local direction unit vector obtained by fitting the spatial coordinate sequence of a branch structure using principal component analysis; Indicates the first The weighting coefficient of each branch structure is specifically taken as the total number of voxels or spatial Euclidean length contained in the branch skeleton. This parameter is used to characterize the dominance of thicker or longer microfibers in the stress state of the micro-fiber group.

[0056] Furthermore, the dynamic spatial pose compensation of the multi-degree-of-freedom displacement stage includes: calculating the angle comparison difference between the microfiber topology trunk direction and the preset macroscopic peeling angle reference value, and obtaining the translational deviation of the root node at the tip of the microcrack; constructing a three-dimensional homogeneous transformation matrix as a spatial compensation matrix based on the angle comparison difference and the translational deviation; performing pose prediction compensation on the spatial compensation matrix using an unscented Kalman filter algorithm, and obtaining the target extension and retraction of each servo actuator of the multi-degree-of-freedom displacement stage through inverse kinematics calculation; and driving the linkage of each servo actuator according to the target extension and retraction to adjust the spatial pose of the test sample.

[0057] Specifically, during the actual peeling process of flexible materials, due to the microscopic inhomogeneity of the internal properties of the adhesive layer, the actual propagation direction of the microcrack tip often dynamically deviates from the set macroscopic test angle (e.g., standard). or (Macroscopic stripping angle). At this point, the microfiber topology backbone direction vector calculated in real time will be... The Euler angles of the ideal space vector representing the preset macroscopic peeling angle reference value are compared and subtracted to obtain the comparison difference value including spatial yaw, pitch, and roll deviations. Simultaneously, the current three-dimensional spatial coordinates of the root node at the tip of the microcrack in the global coordinate system are extracted and subtracted from its initial calibration coordinates to obtain the translational deviation. Based on the above angle comparison difference and translational deviation, a fourth-order three-dimensional homogeneous transformation matrix, i.e., the spatial compensation matrix, is constructed, consisting of rotational and translational components. The base supporting the test sample of the covering film adopts a six-degree-of-freedom parallel displacement stage based on the Stewart configuration. After receiving the spatial compensation matrix, the vision servo controller, combined with the unscented Kalman filter algorithm, performs pose prediction compensation for the algorithm processing delay caused by the time consumption of front-end visual image acquisition and feature extraction. Subsequently, it calls the internal inverse kinematics solution module to calculate the target extension length of the six high-precision servo electric cylinders at the bottom of the displacement stage in real time. The controller issues multi-axis synchronous interpolation commands at a millisecond-level response frequency to drive the electric cylinders in linkage, dynamically correcting the physical spatial attitude of the test sample at high frequency. During this dynamic compensation process, the shear stress and normal stress fed back by the tensile sensor of the testing machine, as well as the real-time displacement of the displacement stage, are recorded simultaneously to obtain a high-fidelity force-slip displacement curve.

[0058] This step reduces the control closed-loop target of pose adjustment from the macroscopic appearance contour to the actual force vector at the microscopic crack tip, completely eliminating the test environment error caused by the microscopic local geometric distortion of the material, achieving absolute spatial consistency of the boundary conditions for the micromechanical test of the material, and effectively preventing the stress vector drift phenomenon during the peeling stress process.

[0059] Step S4: Input the force-slip displacement curve and the real three-dimensional strain field of the interface into the pre-trained cross-modal deep neural network, and output the corrected capping film peeling force.

[0060] In this embodiment, the pre-trained cross-modal deep neural network is a cross-modal mask autoencoder network. The pre-training process of the cross-modal mask autoencoder network includes: performing a preset ratio of random masking on the visual features of the input real three-dimensional strain field of the interface on the time axis and spatial axis; and synchronously inputting the mechanical features of the force-slip displacement curve and the remaining unmasked visual features into the feature encoder of the cross-modal mask autoencoder network.

[0061] Specifically, in multimodal deep learning, the real 3D strain field of the interface, as high-dimensional dense visual data, contains far more information entropy than the force-slip displacement curve, which is a low-dimensional one-dimensional time series. If these two modalities of data are directly and indiscriminately input into the network, the deep neural network will overfit the high-dimensional visual features due to its "shortcut learning" tendency, completely ignoring the constraints of macroscopic mechanical features, thus causing a serious "modal collapse" phenomenon. To overcome this defect, this embodiment adopts a patch-based random masking strategy, randomly occluding a large number of 3D strain field visual features at a very high ratio (e.g., setting the masking rate to 75%) in both the spatial grid and temporal frame dimensions.

[0062] Before performing cross-modal feature fusion, a one-dimensional temporal convolutional network (1D-CNN) is first used to perform feature upsizing and sequence segmentation on the force-slip displacement curve, generating a sequence of mechanical feature tokens with the same dimension as the visual feature patch. Then, the remaining 25% of visible visual feature patches are concatenated with this mechanical feature token sequence and fused using sequence dimension concatenation and position encoding, and fed into a Transformer-based cross-modal feature encoder to extract the joint representation vector. To ensure efficient extraction of cross-modal features, preferably, the cross-modal feature encoder contains 12 Transformer encoding layers, each configured with 16 multi-head attention mechanisms.

[0063] Furthermore, the pre-training process of the cross-modal masked autoencoder network also includes: using the decoder of the cross-modal masked autoencoder network, the mechanical features in the unmasked state are used as cross-modal cues to reconstruct the masked visual spatial features in the latent space; the network parameters are updated by backpropagation using the reconstruction error between the reconstructed visual spatial features and the original visual features as the loss function. During this decoding process, since most visual information has been discarded, the network is forced to learn to mine the deep physical mapping relationship between one-dimensional mechanical signals and three-dimensional visual strain in the latent space. The decoder uses the mechanical features as cues, and the cross-attention mechanism prompts the network to infer and fill in the missing microscopic strain distribution based on the macroscopic stress state. To ensure the physical continuity and accuracy of the reconstruction, a pure reconstruction error loss function based on mean squared error (MSE) is constructed during the pre-training stage.

[0064] The formula for calculating the pure reconstruction error loss function is as follows:

[0065] (6)

[0066] In the formula, This represents the total reconstruction error loss function of a cross-modal mask autoencoder network during the self-supervised pre-training phase. A set of indices representing visual feature blocks obscured by a random mask; This represents the total number of visual feature blocks that are occluded. This indicates the first step of the decoder's cross-reconstruction of the output in the latent space. A visual feature vector of the three-dimensional strain field at an occluded location; This represents the visual feature vector of the real, unobstructed 3D strain field corresponding to this location (i.e., Ground Truth). This represents the squared L2 norm between the reconstructed vector and the true vector, also known as the mean square error term, used to constrain the physical precision of continuous spatial deformation.

[0067] After completing the self-supervised pre-training based on reconstruction error, some low-level parameters of the cross-modal feature encoder were frozen. A multilayer perceptron (MLP) classification head and regression head were then connected to the network end. Supervised fine-tuning training was performed using a dataset with manually labeled micro-failure states. At this point, the cross-entropy loss function was used to update the network parameters to output high-precision classification results of micro-failure mechanisms, and the corrected capping film peeling force was simultaneously output through the regression head. To verify the high fidelity of the network correction, dynamic peeling mechanics data of a sample from the same test batch were extracted for visualization experiments, such as... Figure 3 As shown.

[0068] Figure 3 This is a comparison of the force-slip displacement dynamic evolution curves before and after cross-modal network correction. Figure 3 As shown in the figure, the thin solid line with severe burrs represents the "original macroscopic load signal" directly acquired by the tensile sensor of the universal testing machine. Due to the spatial drift of the micro-stress vector, local stick-slip effect, and mechanical compliance of the equipment during the peeling process of the flexible film, the original curve exhibits extremely high-frequency oscillations, making it difficult to accurately reflect the intrinsic interfacial bonding strength of the material. The thick solid line with a smooth transition represents the "corrected capping film peeling force" output after inputting the true three-dimensional strain field of the interface through a cross-modal deep neural network. Through cross-modal joint decoding of vision and mechanics, the network not only effectively filters out spurious oscillation peaks caused by the stiffness of the testing system but also accurately compensates for the energy dissipation of the underlying interface during the microcrack initiation stage. Statistical results show that the root mean square variance of the corrected curve is reduced by 76.4% compared to the uncorrected curve, achieving a leap from macroscopic empirical estimation to microscopic physical quantification in the mechanical characterization of peeling strength.

[0069] This step achieves high-precision, disturbance-resistant numerical correction of peeling force by establishing a cross-modal feature reconstruction network with strong physical causal relationships and deeply integrating it with the underlying industrial knowledge graph.

[0070] Example 2, Figure 4 A machine vision-based capping film peel force detection system is presented, including:

[0071] The data acquisition module is used to acquire visual image sequences of the peeling interface, macroscopic load signals, and internal tomographic scan images of the adhesive layer inside the cover film during the peeling process.

[0072] The strain field construction module is used to perform three-dimensional displacement calculation on the depth profile of the adhesive layer using an internal volume digital image correlation algorithm to obtain the true three-dimensional strain field of the interface.

[0073] The pose compensation module is used to extract the microfiber topology backbone direction of the microcrack tip based on the real three-dimensional strain field of the interface, drive the multi-degree-of-freedom displacement stage to perform dynamic spatial pose compensation, and obtain the force-slip displacement curve.

[0074] The peel force correction module is used to input the force-slip displacement curve and the real three-dimensional strain field of the interface into a pre-trained cross-modal deep neural network and output the corrected peel force of the cover film.

[0075] The above formulas are all dimensionless calculations. The formulas are derived from software simulations based on a large amount of collected data to obtain the most recent real-world results. The preset parameters in the formulas are set by those skilled in the art according to the actual situation.

[0076] The above embodiments can be implemented, in whole or in part, by software, hardware, firmware, or any other combination thereof. When implemented using software, the above embodiments can be implemented, in whole or in part, in the form of a computer program product.

[0077] Those skilled in the art will recognize that the modules and algorithm steps of the various examples described in conjunction with the embodiments disclosed herein can be implemented in electronic hardware, or a combination of computer software and electronic hardware. Whether these functions are implemented in hardware or software depends on the specific application and design constraints of the technical solution. Those skilled in the art can use different methods to implement the described functions for each specific application, but such implementation should not be considered beyond the scope of this application.

[0078] In addition, the functional modules in the various embodiments of this application can be integrated into one processing module, or each module can exist physically separately, or two or more modules can be integrated into one module.

[0079] The above description is merely a specific embodiment of this application, but the scope of protection of this application is not limited thereto. Any variations or substitutions that can be easily conceived by those skilled in the art within the scope of the technology disclosed in this application should be included within the scope of protection of this application. Therefore, the scope of protection of this application should be determined by the scope of the claims.

[0080] In conclusion, the above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the protection scope of the present invention.

Claims

1. A machine vision-based cover film peel force detection method, characterized by, Includes the following steps: Acquire visual image sequences of the peeling interface, macroscopic load signals, and internal tomographic scan images of the adhesive layer inside the cover film during the peeling process; The acquisition of the visual image sequence includes: using an event camera to capture the dynamic changes in light intensity within the microscopic field of view of the peeling interface, and generating an asynchronous visual event stream consisting of pixel coordinates, timestamps, and polarity changes as the visual image sequence; the acquisition of the internal tomographic scanning image includes: using a near-infrared swept frequency light source to penetrate the light-transmitting substrate of the flexible printed circuit board, performing depth profile three-dimensional interferometric imaging on the peeling process of the adhesive layer, and generating an internal tomographic scanning image. The three-dimensional displacement of the depth profile of the adhesive layer is calculated by using an internal volume digital image correlation algorithm to obtain the true three-dimensional strain field of the interface. Based on the real three-dimensional strain field of the interface, the microfiber topology backbone direction of the microcrack tip is extracted, and a multi-degree-of-freedom displacement stage is driven to perform dynamic spatial pose compensation to obtain the force-slip displacement curve. The force-slip displacement curve and the actual three-dimensional strain field of the interface are input into a pre-trained cross-modal deep neural network, which outputs the corrected capping film peeling force.

2. The machine vision-based method for detecting the peel force of a cover film according to claim 1, characterized in that, The step of using an internal volume digital image correlation algorithm to perform three-dimensional displacement calculation of the depth profile of the adhesive layer includes: The morphological features of the filler particles and micropores inside the colloidal material are extracted from the internal chromatographic scanning images and used as endogenous speckle markers. Using the endogenous speckle markers as a three-dimensional voxel tracking reference, the dynamic three-dimensional deformation of the interface adhesive layer is tracked through an internal volume digital image correlation algorithm.

3. The machine vision-based method for detecting the peel force of a cover film according to claim 2, characterized in that, The method of tracking the dynamic three-dimensional deformation of the interfacial adhesive layer using an internal volume digital image correlation algorithm includes: Monitor the local strain gradient at the peeling interface during the tracking process; When the local strain gradient exceeds a preset threshold, a reference state update mechanism is triggered, and the voxels of the internal tomographic image of the current frame are used as the reference benchmark for the next calculation cycle to calculate the true yield strain of the interfacial colloid.

4. The machine vision-based method for detecting the peel force of a cover film according to claim 1, characterized in that, The extraction of the microfiber topological backbone direction at the microcrack tip includes: Edge detection and three-dimensional morphological refinement are performed on the microfibrillated region at the tip of the microcrack corresponding to the extreme region of the strain gradient in the real three-dimensional strain field of the interface to extract the topological skeleton that reflects the spatial connectivity of the micro-fibrillated group. Calculate the weighted average direction vector of each branch structure in the topological skeleton, and determine the weighted average direction vector as the main direction of the microfiber topology.

5. The machine vision-based method for detecting the peel force of a cover film according to claim 1, characterized in that, The driving of the multi-degree-of-freedom displacement stage for dynamic spatial pose compensation includes: Calculate the angle difference between the main direction of the microfiber topology and the preset macroscopic peeling angle reference value, and obtain the translational deviation of the root node at the tip of the microcrack. Based on the angle comparison difference and the translation deviation, a three-dimensional homogeneous transformation matrix is ​​constructed as a spatial compensation matrix; The pose prediction compensation of the spatial compensation matrix is ​​performed using the unscented Kalman filter algorithm, and the target extension and retraction of each servo actuator of the multi-degree-of-freedom displacement stage is obtained by inverse kinematics calculation. The target expansion and contraction amount drives the linkage of each servo actuator to adjust the spatial attitude of the test sample.

6. The machine vision-based method for detecting the peel force of a cover film according to claim 1, characterized in that, The pre-trained cross-modal deep neural network is a cross-modal mask autoencoder network, and the pre-training process of the cross-modal mask autoencoder network includes: On the time axis and space axis, the visual features of the input interface's real three-dimensional strain field are randomly masked with a preset ratio. The mechanical characteristics of the force-slip displacement curve and the remaining unmasked visual characteristics are simultaneously input into the feature encoder of the cross-modal mask autoencoder network.

7. The machine vision-based method for detecting the peel force of a cover film according to claim 6, characterized in that, The pre-training process of the cross-modal mask autoencoder network also includes: The decoder of the cross-modal masked autoencoder network uses the mechanical features in the unmasked state as cross-modal cue words to cross-reconstruct the visual spatial features that are masked in the latent space. The network parameters are updated by backpropagation using the reconstruction error between the reconstructed visual spatial features and the original visual features as the loss function.

8. A system using the machine vision-based cover film peel force detection method as described in any one of claims 1-7, characterized in that, include: The data acquisition module is used to acquire visual image sequences of the peeling interface, macroscopic load signals, and internal tomographic scan images of the adhesive layer inside the cover film during the peeling process. The strain field construction module is used to perform three-dimensional displacement calculation on the depth profile of the adhesive layer using an internal volume digital image correlation algorithm to obtain the true three-dimensional strain field of the interface. The pose compensation module is used to extract the microfiber topology backbone direction of the microcrack tip based on the real three-dimensional strain field of the interface, drive the multi-degree-of-freedom displacement stage to perform dynamic spatial pose compensation, and obtain the force-slip displacement curve. The peel force correction module is used to input the force-slip displacement curve and the real three-dimensional strain field of the interface into a pre-trained cross-modal deep neural network and output the corrected peel force of the cover film.

Images