Abnormality detection method and system for flexible circuit board patch

Abstract

The invention relates to an anomaly detection technology, and discloses an anomaly detection method and system for a flexible circuit board patch, and the method comprises the steps: obtaining a patch image in a preset flexible circuit board, carrying out the illumination normalization processing and background suppression processing of the patch image, and obtaining a background suppression patch image, and carrying out abnormal bending identification on the background suppression patch image based on multi-scale gradient fusion and local structure consistency analysis, judging whether the patch in the flexible circuit board has abnormal bending according to a bending identification result, if yes, judging that an abnormal detection result is abnormal bending, and if not, judging that the patch in the flexible circuit board has abnormal bending. And otherwise, calculating an offset coefficient of the normalized patch image based on double-flow alignment and differential significance enhancement, judging whether the offset coefficient is greater than a preset coefficient threshold, if so, judging that the abnormal detection result is abnormal offset, and if not, judging that the abnormal detection result is abnormal. According to the invention, the communication circuit board patch abnormity detection precision can be improved.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of anomaly detection, in particular to an anomaly detection method and system for a flexible circuit board patch. BACKGROUND

[0002] As a core component of modern electronic devices, flexible circuit boards are widely used in fields such as foldable screen mobile phones, wearable devices, and aerospace electronics due to their bendability and lightweight advantages. In the production and manufacturing process of flexible circuit boards, quality detection in the patching process is a key link to ensure product performance. However, due to the material properties of flexible circuit boards and the complexity of the production environment, existing patch anomaly detection techniques face many challenges.

[0003] In existing anomaly detection methods for flexible circuit board patches, the accuracy of abnormal bending recognition of the patch is insufficient. Abnormal bending of flexible circuit boards often manifests as micro-scale structural distortion, involving local topological changes, curvature abnormalities, and other fine features. Existing bending detection methods are mostly based on single-scale gradient changes or geometric shape analysis, which cannot capture topological vortex features at multiple scales (such as microstructure distortion caused by stress concentration) and are difficult to quantify bending degree through curvature tensors, vibration instability, and other physical properties. This leads to high miss detection rates for slight bending or complex deformation areas, and the inability to distinguish between elastic deformation and plastic bending, resulting in poor detection robustness. Moreover, patch offset detection is difficult to adapt to flexible deformation. Due to the bendability of the flexible substrate, the relative positional relationship between the patch component and the pad is in a dynamically changing manifold space. Existing offset detection methods are mostly based on position comparison in a rigid coordinate system, which cannot effectively handle overall coordinate deviations caused by flexible deformation, resulting in low detection accuracy. SUMMARY

[0004] The present application provides an anomaly detection method and system for a flexible circuit board patch, which mainly aims to solve the problem of low detection accuracy in existing anomaly detection methods for flexible circuit board patches.

[0005] To achieve the above-mentioned purpose, the present application provides an anomaly detection method for a flexible circuit board patch, which comprises: Obtaining a patch image in a predetermined flexible circuit board, performing illumination normalization processing on the patch image to obtain a normalized patch image, and performing background suppression processing on the normalized patch image to obtain a background-suppressed patch image; Performing abnormal bending recognition on the background-suppressed patch image based on multi-scale gradient fusion and local structure consistency analysis to obtain a bending recognition result; Determining whether the patch in the flexible circuit board has abnormal bending according to the bending recognition result; If so, determining that the anomaly detection result is abnormal bending; If not, calculate the shift coefficient of the normalized patch image based on the double manifold alignment and differential saliency enhancement; Determine whether the shift coefficient is greater than a preset coefficient threshold value; If greater, determine that the abnormality detection result is an abnormal shift, otherwise determine that the abnormality detection result is no abnormality.

[0006] Optionally, the illumination normalization processing of the patch image to obtain the normalized patch image comprises: Collecting a multi-light source patch image in the flexible circuit board; Performing illumination invariance projection on the multi-light source patch image to obtain a single-channel illumination invariance feature map; Performing adaptive local contrast equalization processing on the single-channel illumination invariance feature map to obtain an adaptive feature map; Based on gradient domain fusion reconstruction, the adaptive feature map is used to normalize the patch image to obtain a normalized patch image.

[0007] Optionally, the background suppression processing of the normalized patch image to obtain a background suppression patch image comprises: Extracting the structural features of the normalized patch image, constructing a background structure model of the flexible substrate by analyzing the gradient distribution characteristics in the neighborhood of each pixel of the normalized patch image; Generating a background estimation map according to the structural features using the background structure model; Detecting salient regions in the background estimation map, calculating the foreground saliency value of each salient region by analyzing the gradient consistency, texture complexity and contrast intensity of each salient region; Based on the foreground value, background segmentation is realized to obtain a binary foreground mask; According to the binary foreground mask, multi-scale background suppression fusion is performed on the normalized patch image to obtain a background suppression patch image.

[0008] Optionally, the multi-scale background suppression fusion of the normalized patch image based on the binary foreground mask to obtain a background suppression patch image comprises: Based on a multi-scale space, a multi-scale image list of the normalized patch image is constructed; Performing pixel subtraction operation of images in the multi-scale image list and the background structure model in different scale spaces to obtain a background difference image list; Based on the foreground mask, foreground adaptive gain is performed on the background difference image list to obtain a gain image list; Performing multi-scale reconstruction on the gain image list to obtain a reconstructed image list; Fusing the list of reconstructed images based on the foreground saliency value to obtain a background-suppressed patch image.

[0009] Optionally, the abnormal bending identification on the background-suppressed patch image based on the multi-scale gradient fusion and local structure consistency analysis obtains a bending identification result, comprising: calculating a complex-valued gradient field of the background-suppressed patch image at different scales to obtain a multi-scale complex-valued gradient tensor; topological vortex detection on the multi-scale complex-valued gradient tensor to obtain a multi-scale vortex feature map; manifold consistency analysis on the multi-scale complex-valued gradient tensor and the multi-scale vortex feature map to obtain a curvature anomaly map; phonon vibration simulation on the curvature anomaly map based on the original gradient tensor of the background-suppressed patch image to obtain a vibration instability heat map; generating a pixel-level bending probability map according to the multi-scale vortex feature map, the curvature anomaly map, and the vibration instability heat map; phase transition criticality detection on the pixel-level bending probability map to obtain a binary bending identification result.

[0010] Optionally, the judging whether the patch in the flexible circuit board has abnormal bending according to the bending identification result comprises: identifying an abnormal area of the patch according to the bending identification result to obtain a total number of pixels of the abnormal area; identifying the number of abnormal pixel points of the abnormal area; calculating the ratio of the number of abnormal pixel points to the total number of pixels to obtain a critical point density; if the critical point density is greater than or equal to a preset density threshold, it is determined that there is abnormal bending, otherwise it is determined that there is no abnormal bending.

[0011] Optionally, the calculating the offset coefficient of the normalized patch image based on the dual-manifold alignment and differential saliency enhancement comprises: constructing a dual-flow manifold for the normalized patch image to obtain a patch manifold feature matrix and a pad manifold feature matrix; Riemannian alignment of the manifold feature matrix and the pad manifold feature matrix according to a preset Riemannian metric to obtain an alignment error matrix; constructing a quantum gradient difference map according to the alignment error matrix and the normalized patch image; saliency sparse coding on the quantum gradient difference map to obtain an offset saliency map; calculating the offset coefficient according to the offset saliency map and the alignment error matrix.

[0012] Optionally, the double-flow type construction is performed on the normalized patch image to obtain a patch manifold feature matrix and a pad manifold feature matrix, comprising: The patch element segmentation is performed on the normalized patch image to obtain a patch segmentation mask; The geometric feature, appearance feature and texture feature of the normalized patch image are extracted according to the patch segmentation mask to obtain a patch feature vector; The patch manifold feature is constructed according to the patch feature vector; The design parameters of the pad in the normalized patch image are obtained according to a preset PCB design file; The standard position feature of the pad in the normalized patch image is extracted according to the design parameters to obtain a pad manifold feature matrix.

[0013] Optionally, the offset coefficient is calculated according to the offset saliency map and the alignment error matrix, comprising: A multi-scale feature tensor of the offset saliency map is constructed, and the multi-scale feature tensor is fused with the alignment error matrix to obtain a fusion feature; The physical perception weighting is performed on the fusion feature to obtain a physical weighted feature; The physical weighted feature is compressed into a five-dimensional global feature vector through spatial average pooling; The five-dimensional global feature vector is nonlinearly mapped by using a preset fully connected neural network to obtain an offset coefficient.

[0014] To solve the above problems, the application further provides an abnormality detection system for a flexible circuit board patch, which comprises: An image processing module is configured to acquire a patch image in a preset flexible circuit board, perform illumination normalization on the patch image to obtain a normalized patch image, and perform background suppression on the normalized patch image to obtain a background-suppressed patch image; A bending recognition module is configured to select a first target user group from the candidate user group according to the matching value, and recommend the product to be recommended to the first target user group; A bending judgment module is configured to judge whether the patch in the flexible circuit board has abnormal bending according to the bending recognition result, and if yes, determine that the abnormality detection result is that there is abnormal bending, and if not, calculate an offset coefficient of the normalized patch image based on double manifold alignment and differential saliency enhancement; An offset judgment module is configured to judge whether the offset coefficient is greater than a preset coefficient threshold, and if yes, determine that the abnormality detection result is that there is abnormal offset, and if not, determine that the abnormality detection result is that there is no abnormality.

[0015] The embodiment of the present application obtains a patch image in a preset flexible circuit board, performs light normalization processing on the patch image to obtain a normalized patch image, performs background suppression processing on the normalized patch image to obtain a background-suppressed patch image, performs abnormal bending recognition on the background-suppressed patch image based on multi-scale gradient fusion and local structure consistency analysis to obtain a bending recognition result, judges whether the patch in the flexible circuit board has abnormal bending according to the bending recognition result, if yes, determines that the abnormal detection result is abnormal bending, and if not, calculates an offset coefficient of the normalized patch image based on double-manifold alignment and differential saliency enhancement, judges whether the offset coefficient is greater than a preset coefficient threshold, if yes, determines that the abnormal detection result is abnormal offset, and if not, determines that the abnormal detection result is no abnormality. Therefore, the abnormal detection method and system for the patch of the flexible circuit board provided by the present application can solve the problem of low detection accuracy of the existing abnormal detection method for the patch of the flexible circuit board. BRIEF DESCRIPTION OF DRAWINGS

[0016] Figure 1 A flowchart of an abnormal detection method for a patch of a flexible circuit board provided by an embodiment of the present application is shown. Figure 2 A function module diagram of an abnormal detection system for a patch of a flexible circuit board provided by an embodiment of the present application is shown.

[0017] The implementation, functional features and advantages of the present application will be further described with reference to the embodiments and the accompanying drawings. DETAILED DESCRIPTION

[0018] It should be understood that the specific embodiments described herein are merely intended to explain the present application and not to limit the present application.

[0019] An abnormal detection method for a patch of a flexible circuit board is provided by an embodiment of the present application. The execution subject of the abnormal detection method for the patch of the flexible circuit board includes but is not limited to at least one of electronic devices capable of being configured to execute the method provided by the embodiment of the present application, such as a server, a terminal, etc. In other words, the abnormal detection method for the patch of the flexible circuit board can be executed by software or hardware installed in a terminal device or a server device, and the software can be a blockchain platform. The server includes but is not limited to a single server, a server cluster, a cloud server, a cloud server cluster, etc. The server can be a stand-alone server, or a cloud server providing cloud services, cloud databases, cloud computing, cloud functions, cloud storage, network services, cloud communication, middleware services, domain name services, security services, content distribution networks (CDN), and basic cloud computing services such as big data and artificial intelligence platforms.

[0020] Referring to Figure 1 Fig. 1 is a flow diagram of an abnormality detection method for a flexible circuit board patch according to an embodiment of the present application. In this embodiment, the abnormality detection method for the flexible circuit board patch comprises the following steps: S1, obtaining a preset patch image in a flexible circuit board, performing illumination normalization processing on the patch image to obtain a normalized patch image, and performing background suppression processing on the normalized patch image to obtain a background-suppressed patch image.

[0021] In the embodiment, the flexible circuit board is a printed circuit board made of a flexible insulating substrate, and has the characteristics of bendability, thinness and high reliability.

[0022] In the embodiment, the illumination normalization processing on the patch image to obtain a normalized patch image comprises the following steps: Collecting a multi-light-source patch image in the flexible circuit board; Performing illumination invariance projection on the multi-light-source patch image to obtain a single-channel illumination invariance feature map; Performing adaptive local contrast equalization processing on the single-channel illumination invariance feature map to obtain an adaptive feature map; Performing normalization processing on the patch image based on gradient domain fusion reconstruction using the adaptive feature map to obtain a normalized patch image.

[0023] In detail, the collection of the multi-light-source patch image in the flexible circuit board is achieved by controlling multiple light sources of different angles and types (such as coaxial light, ring light of a specific angle and point light source) to irradiate the flexible circuit board in sequence or synchronously, and using a color camera to collect the patch RGB image under the irradiation of each type of light source.

[0024] In detail, the illumination invariance projection on the multi-light-source patch image is achieved by applying logarithmic operation to compress the dynamic range and enhance the visibility of the low-illumination area, and then using a preset projection matrix to perform linear combination operation on the multi-channel data in the logarithmic domain; this projection aims to map the original RGB information to a new feature space, in which the components in a specific direction are relatively insensitive to changes in illumination intensity and direction, and finally the components in the specific projection direction are extracted to generate a single-channel image.

[0025] In the embodiment, the background suppression processing on the normalized patch image to obtain a background-suppressed patch image comprises the following steps: Extracting the structural features of the normalized patch image, and constructing a background structure model of the flexible substrate by analyzing the gradient distribution characteristics in the neighborhood of each pixel in the normalized patch image; generating a background estimation map according to the structure features by using the background structure model; detecting salient regions in the background estimation map, calculating foreground saliency values of each salient region by analyzing gradient consistency, texture complexity and contrast intensity of each salient region; implementing background segmentation based on the foreground values to obtain a binary foreground mask; performing multi-scale background suppression fusion on the normalized patch image according to the binary foreground mask to obtain a background suppression patch image.

[0026] In detail, the background structure model of the flexible substrate is constructed by analyzing the gradient distribution characteristics of each pixel neighborhood in the normalized patch image, quantifying the direction and intensity of the substrate texture by calculating the local structure tensor (including the eigenvalues and direction of the gradient covariance matrix); the structure features are smoothed and aggregated based on multi-scale anisotropic Gaussian filtering to construct the background structure model of the flexible substrate.

[0027] In detail, the background estimation map is generated according to the structure features by using the background structure model, and the local texture features are expanded to the whole image by a spatial extrapolation algorithm to generate a background estimation map consistent with the size of the input image.

[0028] In detail, the salient regions in the background estimation map are detected by combining phase consistency edge detection and superpixel segmentation to divide the perception unit.

[0029] In detail, the foreground saliency values of each salient region are calculated by analyzing the gradient consistency, texture complexity and contrast intensity of each salient region, which are linearly weighted and fused to obtain the foreground saliency values.

[0030] In detail, the background segmentation is implemented based on the foreground values to obtain a binary foreground mask, and the foreground saliency values are mapped to the whole image pixels to generate an initial binary segmentation map using an adaptive threshold; the spatial continuity constraint is integrated by a graph cut optimization algorithm to minimize the energy function to eliminate isolated noise points, fill in the foreground holes, and accurately fit the geometric boundaries of the patch elements, finally outputting a high-precision binary foreground mask.

[0031] In detail, the binary foreground mask is a two-dimensional matrix (for example, a mask matrix of 1920x1080 corresponding to an image of 1920x1080) completely same as the size of the original image, and the value of each pixel in the matrix only contains two states of 0 (background) or 1 (foreground): 0 (black): marked as a flexible substrate background area, indicating that the area needs to be suppressed, 1 (white): marked as a patch element foreground area, indicating that the area needs to be retained or enhanced.

[0032] In the embodiment of the present application, the multi-scale background suppression fusion of the normalized patch image according to the binary foreground mask comprises: constructing a multi-scale image list of the normalized patch image based on multi-scale space; performing pixel subtraction operation of images in the multi-scale image list and the background structure model in different scale spaces to obtain a background difference image list; performing foreground adaptive gain on the background difference image list based on the foreground mask to obtain a gain image list; performing multi-scale reconstruction on the gain image list to obtain a reconstructed image list; performing saliency weighted fusion on the reconstructed image list based on the foreground saliency value to obtain a background suppression patch image.

[0033] In detail, the multi-scale image list of the normalized patch image based on multi-scale space is generated by using an interpolation algorithm to generate three different resolution image scales: original resolution, 1 / 2 down-sampling resolution and 1 / 4 down-sampling resolution. This process constructs a multi-level expression system containing macro texture to micro detail under the premise of maintaining the geometric features of the image edge, wherein the original scale retains the fine structure of the element, the medium scale balances noise control and feature retention, and the coarse scale focuses on the overall texture pattern of the substrate, forming a complete scale space representation basis.

[0034] In detail, the pixel subtraction operation of images in the multi-scale image list and the background structure model in different scale spaces to obtain a background difference image list is to perform pixel-level subtraction operation on the image of the corresponding scale and the background structure model with adaptive scaling in each independent scale space.

[0035] In detail, the foreground adaptive gain on the background difference image list based on the foreground mask to obtain a gain image list means that 2 times gain enhancement feature response is applied to the foreground area (patch element) marked by the mask, and 0.7 times gain is applied to the background area to suppress residual noise.

[0036] In detail, the multi-scale reconstruction of the gain image list to obtain a reconstructed image list is to reconstruct the gain images of each scale to the original input resolution through bicubic interpolation to restore spatial geometric consistency.

[0037] In detail, the significant weighting fusion of the reconstructed image list based on the foreground significant value is to combine the foreground significant value mapping and adopt a dynamic weight strategy to fuse three reconstructed images.

[0038] S2, based on multi-scale gradient fusion and local structure consistency analysis, the abnormal bending recognition of the background suppression patch image is obtained.

[0039] In the embodiment of the application, the abnormal bending recognition of the background suppression patch image based on multi-scale gradient fusion and local structure consistency analysis to obtain the bending recognition result comprises: calculating the complex-valued gradient field of the background suppression patch image at different scales to obtain a multi-scale complex-valued gradient tensor; topological vortex detection is performed on the multi-scale complex-valued gradient tensor to obtain a multi-scale vortex feature map; manifold consistency analysis is performed on the multi-scale complex-valued gradient tensor and the multi-scale vortex feature map to obtain a curvature anomaly map; phonon vibration simulation is performed on the curvature anomaly map based on the original gradient tensor of the background suppression patch image to obtain a vibration instability thermodynamic map; a pixel-level bending probability map is generated according to the multi-scale vortex feature map, the curvature anomaly map and the vibration instability thermodynamic map; phase transition criticality detection is performed on the pixel-level bending probability map to obtain a binary bending recognition result.

[0040] In detail, the calculation of the complex-valued gradient field of the background suppression patch image at different scales is to calculate the complex-valued gradient field at three characteristic scales (corresponding to the standard deviations of the Gaussian kernel function: 1 pixel, 2.5 pixels and 4 pixels) through an improved gradient operator. The operator encodes the traditional gradient amplitude and phase information into a complex number, in which the real part represents the intensity change rate and the imaginary part represents the direction continuity. The generated multi-scale complex-valued gradient tensor not only captures the pixel intensity change, but also quantifies the local structure geometric continuity through the phase angle.

[0041] In detail, the standard deviation of the Gaussian kernel function represents the spatial action range of multi-scale analysis.

[0042] In detail, the topological vortex detection on the multi-scale complex-valued gradient tensor is quantum topological analysis based on the complex-valued gradient tensor, and the topological vortex core point is identified by calculating the Berry phase accumulation of each pixel position. When the gradient phase around a point has a mutation of 2π integer times, the point is marked as a vortex defect center, which corresponds to the microstructure distortion caused by stress concentration in the material; through a multi-scale cross-validation strategy, the vortex distribution pattern is detected in three scale spaces synchronously, and the pseudo-vortex points caused by noise are filtered out, and finally the multi-scale vortex feature map marking the real physical distortion position is output.

[0043] In detail, the Berry phase is an important "geometric phase" in quantum mechanics, which is used to describe the additional phase change of the wave function after the quantum system undergoes slow evolution in the parameter space. Specifically, when the Hamiltonian (a quantity describing the energy of the system) of a quantum system depends on some external parameters (such as the direction of the magnetic field, the strength of the electric field, etc.), and these parameters are slowly changed in the "parameter space" along a closed path (satisfying the "adiabatic approximation", i.e. the system is always in a certain instantaneous energy eigenstate), and finally returns to the initial parameter state, the wave function of the system will not only accumulate a "dynamic phase" due to energy evolution, but also an additional phase related to the geometric properties of the path. This additional phase is the Berry phase.

[0044] In detail, the manifold consistency analysis on the multi-scale complex-valued gradient tensor and the multi-scale vortex feature map is to construct a local Riemannian manifold in the neighborhood of the vortex core point, and to describe the geometric deformation characteristics of the gradient field through the metric tensor. Specifically, the Christoffel symbol is calculated to derive the curvature tensor, and the scalar curvature is extracted to quantify the bending degree of the local surface: the curvature of the flat area tends to zero, while the abnormally bent area presents high curvature characteristics; by establishing a curvature threshold model (curvature > 0.05 is abnormal), the differential geometric quantity is converted into a visual curvature anomaly map, in which the highlighted area accurately corresponds to the material plastic deformation position.

[0045] In detail, the local Riemannian manifold is a mathematical model in differential geometry that describes curved spaces, and its core is to define the local geometric properties of the space through the metric tensor. In image processing, this model maps the pixel gradient field in the neighborhood of the vortex point to a curved two-dimensional manifold, where: each pixel point corresponds to a coordinate point on the manifold; the gradient direction corresponds to the direction of the basis vector defined by the local coordinate system; and the gradient amplitude corresponds to the scaling ratio (metric length) of the basis vector.

[0046] In detail, the phonon vibration simulation of the curvature anomaly map based on the original gradient tensor of the background-suppressed patch image is based on constructing a lattice dynamics model based on the original gradient tensor, regarding the image pixels as mass points, and deriving the force constant matrix from the gradient covariance. The phonon spectrum distribution is obtained by solving the eigen vibration equation, and special attention is paid to the spatial distribution of the low-frequency soft mode (less than 0.1 THz) - these thermodynamic instability regions represent points of weakened lattice binding energy; combined with the geometric constraints of the curvature anomaly map, the dissipation path of vibration energy in the bending region is simulated to generate a vibration instability thermodynamic map showing the vulnerable positions of the structure.

[0047] In detail, the lattice dynamics model is a theoretical framework in condensed matter physics for describing the collective vibration of crystal atoms, which is innovatively migrated to the image processing field in the present scheme for analyzing the mechanical instability characteristics of the microstructure of the flexible circuit board. The atomic mass corresponds to the pixel gray value in the image, the atomic displacement corresponds to the pixel displacement in the image, and the force constant corresponds to the gradient covariance matrix in the image. First, each pixel of the background-suppressed patch image is regarded as a lattice atom with a specific "mass" (gray value) and "position" (coordinate), the gradient covariance matrix between adjacent pixels is calculated using the original gradient tensor, and the force constant matrix representing the interaction strength between pixels is derived; by constructing a dynamics matrix containing mass-normalized force constants, the eigen equation of the Hermitian matrix is solved to obtain the eigen frequency spectrum and eigen vibration mode of the system; finally, the low-frequency soft mode distribution below the critical frequency is extracted to obtain the lattice dynamics model.

[0048] In detail, the pixel-level bending probability map is generated according to the multi-scale vortex feature map, the curvature anomaly map, and the vibration instability thermodynamic map, with vortex points as nodes and curvature gradients as edges. The node attributes include topological charge, curvature value, and vibration frequency. The implicit association between nodes is learned through a three-layer message passing network, and the sigmoid activation function is used in the decoding layer to map the aggregated features to 0-1 continuous probability values.

[0049] In detail, the phase transition criticality detection on the pixel-level bending probability map to obtain the binarized bending recognition result converts the probability field into an order parameter field by calculating the generalized free energy density containing the gradient term. By numerically solving the free energy minimum condition, the phase transition critical points of the second derivative of the free energy (free energy) are identified, which correspond to the phase boundary of the material from elastic deformation to plastic bending; finally, the critical region is subjected to morphological optimization and binarization processing, and a binary mask identifying the abnormal bending position is output.

[0050] S3, judging whether the patch in the flexible circuit board has abnormal bending according to the bending recognition result.

[0051] ​In the embodiment of the present application, the step of judging whether the patch in the flexible circuit board has abnormal bending according to the bending recognition result comprises: identifying an abnormal area of the patch according to the bending recognition result, and obtaining a total number of pixels of the abnormal area; identifying the number of abnormal pixel points of the abnormal area; calculating the ratio of the number of abnormal pixel points to the total number of pixels to obtain a critical point density; if the critical point density is greater than or equal to a preset density threshold, it is determined that there is abnormal bending, otherwise it is determined that there is no abnormal bending.

[0052] if yes, S4 is executed, and it is determined that the abnormal detection result is abnormal bending.

[0053] if no, S5 is executed, and an offset coefficient of the normalized patch image is calculated based on double manifold alignment and differential saliency enhancement.

[0054] In the embodiment of the present application, the step of calculating the offset coefficient of the normalized patch image based on double manifold alignment and differential saliency enhancement comprises: constructing a double flow type of the normalized patch image to obtain a patch manifold feature matrix and a pad manifold feature matrix; performing Riemann alignment on the manifold feature matrix and the pad manifold feature matrix according to a preset Riemann metric to obtain an alignment error matrix; constructing a quantum gradient difference map according to the alignment error matrix and the normalized patch image; performing saliency sparse coding on the quantum gradient difference map to obtain an offset saliency map; calculating the offset coefficient according to the offset saliency map and the alignment error matrix.

[0055] In detail, the step of performing Riemann alignment on the manifold feature matrix and the pad manifold feature matrix according to a preset Riemann metric is to find the best spatial transformation relationship by an optimization algorithm, eliminate the overall coordinate deviation caused by the deformation of the flexible substrate, and make the actual patch position and the theoretical pad position realize optimal coincidence in the bending space. After alignment, the spatial difference between each patch element and its corresponding pad is calculated to generate a quantitative alignment error matrix.

[0056] In detail, the step of constructing a quantum gradient difference map according to the alignment error matrix and the normalized patch image is to synchronously compare the amplitude and direction difference between the actual image gradient and the theoretical design gradient to generate a quantum gradient difference map that fuses phase information.

[0057] In detail, the significant sparse coding of the quantum gradient difference map is performed by over-complete dictionary decomposition on the quantum gradient difference map, and the image information is decomposed into three basic components of substrate texture, offset mode and random noise. Through the sparse optimization algorithm, the feature mode related to the offset is selectively enhanced, and the substrate texture and noise interference are suppressed, and finally a high-contrast offset significance map is generated.

[0058] In detail, the double-flow type shape construction of the normalized patch image obtains a patch manifold feature matrix and a pad manifold feature matrix, including: The patch element segmentation is performed on the normalized patch image to obtain a patch segmentation mask; The geometric features, appearance features and texture features of the normalized patch image are extracted according to the patch segmentation mask to obtain a patch feature vector; The patch manifold feature is constructed according to the patch feature vector; The design parameters of the pads in the normalized patch image are obtained according to a preset PCB design file; The standard position features of the pads in the normalized patch image are extracted according to the design parameters to obtain a pad manifold feature matrix.

[0059] In detail, the patch element segmentation of the normalized patch image is performed by using an instance segmentation algorithm based on deep learning, and the pixel-level boundaries of various patch elements in the image are accurately recognized by a pre-trained convolutional neural network.

[0060] In detail, the patch manifold feature is constructed according to the patch feature vector, all patch feature vectors are integrated into a structured matrix, and the structured matrix is projected into a low-dimensional embedding space by manifold learning technology (such as t-SNE or UMAP).

[0061] In detail, the design parameters of the pads in the normalized patch image are obtained according to a preset PCB design file, the circuit board layer information in the PCB design file is analyzed, and the design parameters of the pads corresponding to the current patch image are accurately extracted. Including obtaining the theoretical center coordinates, geometric dimensions (length / width / diameter), shape types (rectangle / circle / irregular shape) and rotation angles of each pad in the circuit coordinate system and other core attributes.

[0062] In the embodiment of the application, the offset coefficient is calculated according to the offset significance map and the alignment error matrix, including: A multi-scale feature tensor of the offset significance map is constructed, the multi-scale feature tensor and the alignment error matrix are fused to obtain a fusion feature; The fusion feature is physically perceived and weighted to obtain a physically weighted feature; compressing the physical weighting features into a five-dimensional global feature vector through spatial average pooling; mapping the five-dimensional global feature vector through a preset fully connected neural network to obtain an offset coefficient.

[0063] In detail, the multi-scale feature tensor for constructing the offset saliency map is to extract fine local features at the original resolution level, capture medium-scale regional features at the 1 / 2 down-sampling level, and obtain global macro features at the 1 / 4 down-sampling level. Each scale applies a Gaussian difference filter to enhance edge response, and fuses multi-scale features through maximum projection, finally outputs a three-channel feature tensor containing different granularity features.

[0064] In detail, the multi-scale feature fusion of the multi-scale feature tensor and the alignment error matrix is to first map the alignment error matrix into a heat map form, and perform channel-level splicing with the feature pyramid in the same coordinate space; then apply a channel attention mechanism to automatically learn the importance weight of each feature channel, adaptively weight the five-channel features after splicing, and highlight the features sensitive to the offset to form a unified representation of the fusion of spatial features and error intensity.

[0065] In detail, the physical perception weighting of the fused features is to load preset position sensitivity, size tolerance, and electrical criticality weight according to the element type (resistor / capacitor / IC chip); mark high-precision installation areas (such as BGA pads) through position-sensitive icons, highlight the area of small elements through size weight field, and strengthen the importance of power / signal modules through electrical criticality matrix; multiply the three groups of weight maps to generate a physical constraint matrix, and multiply each element of the fused features to ensure that the offset features of key areas obtain higher weights.

[0066] In detail, the compression of the physical weighting features into a five-dimensional global feature vector through spatial average pooling is to adopt an adaptive grid spatial average pooling strategy, divide the physical weighting features into 5 semantic regions (central core region, four-corner edge region), and calculate the arithmetic mean of each region feature vector.

[0067] In detail, the nonlinear mapping of the five-dimensional global feature vector through a preset fully connected neural network is to use the first layer 32 node ReLU function activation in the fully connected neural network to learn feature combination, the second layer 16 node Tanh function activation to capture nonlinear relationship, and the third layer Sigmoid function output to compress the result to the range of 0-1.

[0068] S6, judging whether the offset coefficient is greater than a preset coefficient threshold.

[0069] If greater, then S7 is executed, and it is determined that the abnormal detection result is that there is abnormal offset, otherwise S8 is executed, and it is determined that the abnormal detection result is that there is no abnormality.

[0070] As Figure 2 shown, it is a functional module diagram of an abnormal detection system for a flexible circuit board patch provided by an embodiment of the present application.

[0071] The abnormal detection system 100 for the flexible circuit board patch can be installed in an electronic device. According to the functions implemented, the abnormal detection system 100 for the flexible circuit board patch can include an image processing module 101, a bending recognition module 102, a bending judgment module 103, and an offset judgment module 104. The modules of the present application can also be referred to as units, which refer to a series of computer program segments that can be executed by an electronic device processor and can complete a fixed function, and are stored in the memory of the electronic device.

[0072] In the present embodiment, the functions of each module / unit are as follows: The image processing module 101 is configured to obtain a patch image in a preset flexible circuit board, perform illumination normalization processing on the patch image to obtain a normalized patch image, and perform background suppression processing on the normalized patch image to obtain a background-suppressed patch image. The bending recognition module 102 is configured to select a first target user group from the candidate user group according to the matching value, and recommend the product to be recommended to the first target user group. The bending judgment module 103 is configured to determine whether the patch in the flexible circuit board has abnormal bending according to the bending recognition result, and if so, determine that the abnormal detection result is that there is abnormal bending, and if not, calculate an offset coefficient of the normalized patch image based on double manifold alignment and differential saliency enhancement. The offset judgment module 104 is configured to determine whether the offset coefficient is greater than a preset coefficient threshold, and if so, determine that the abnormal detection result is that there is abnormal offset, otherwise determine that the abnormal detection result is that there is no abnormality.

[0073] In detail, each module in the abnormal detection system 100 for the flexible circuit board patch in the embodiment of the present application uses the same technical means as the abnormal detection method for the flexible circuit board patch described in the above Figure 1 , and can produce the same technical effects, which will not be described here.

[0074] In the embodiments of the present application, it should be understood that the disclosed system and method can be implemented in other manners. For example, the system embodiments described above are merely schematic. For example, the division of the modules is merely a logical function division. There can be another division manner for the actual implementation.

[0075] The modules illustrated as separated components can or can not be physically separated, and the components illustrated as modules can or can not be physical units, i.e., can be located in one place, or can be distributed on a plurality of network units. Some or all of the modules can be selected according to the actual needs to achieve the purposes of the embodiments.

[0076] In addition, each functional module in each embodiment of the present application can be integrated in one processing unit, or each unit can be physically present separately, or two or more units can be integrated in one unit. The integrated unit can be realized in the form of hardware, or in the form of hardware plus software functional modules.

[0077] It is obvious for those skilled in the art that the present application is not limited to the details of the above exemplary embodiments, and the present application can be implemented in other specific forms without departing from the spirit or essential characteristics of the present application.

[0078] Therefore, from any point of view, the embodiments should be regarded as exemplary and non-limiting, and the scope of the present application is defined by the appended claims rather than the above description, and it is intended to include all changes falling within the meaning and scope of the equivalent elements of the claims. Any reference signs in the claims should not be regarded as limiting the claims involved.

[0079] The embodiments of the present application can acquire and process related data based on artificial intelligence technology. Among them, artificial intelligence (Artificial Intelligence, AI) is to use digital computers or digital computer controlled machines to simulate, extend and expand human intelligence, perceive the environment, acquire knowledge and use knowledge to obtain the best results.

[0080] In addition, it is obvious that the word "comprising" does not exclude other units or steps, and the singular does not exclude the plural. The plurality of units or systems stated in the system claims can also be implemented by one unit or system through software or hardware. The words first, second, etc. are used to indicate names and do not represent any particular order.

[0081] Finally, it should be noted that the above examples are merely intended to illustrate the technical solutions of the present application and not to limit the present application. Although the present application has been described in detail with reference to the preferred embodiments, those skilled in the art should understand that the technical solutions of the present application can be modified or equivalently replaced without departing from the spirit and scope of the technical solutions of the present application.

Claims

1. An anomaly detection method for flexible circuit board patch mounting, characterized in that, The method includes: Obtain a preset patch image in a flexible circuit board, perform illumination normalization processing on the patch image to obtain a normalized patch image, and perform background suppression processing on the normalized patch image to obtain a background suppressed patch image. Based on multi-scale gradient fusion and local structure consistency analysis, abnormal bending recognition is performed on the background suppression patch image to obtain bending recognition results. Based on the bending identification results, it is determined whether there are abnormal bends in the patches of the flexible circuit board; If present, the abnormal detection result is determined to be an abnormal bend. If it does not exist, the offset coefficient of the normalized patch image is calculated based on dual manifold alignment and differential saliency enhancement; Determine whether the offset coefficient is greater than a preset coefficient threshold; If the value is greater than the value, the anomaly detection result is determined to be an abnormal offset; otherwise, the anomaly detection result is determined to be no anomaly.

2. The anomaly detection method for flexible circuit board patch as described in claim 1, characterized in that, The step of performing illumination normalization processing on the patch image to obtain a normalized patch image includes: Acquire images of the multi-light source patch in the flexible circuit board; The multi-light source patch image is subjected to illumination invariant projection to obtain a single-channel illumination invariant feature map; An adaptive local contrast equalization process is performed on the single-channel illumination invariance feature map to obtain an adaptive feature map. Based on gradient domain fusion reconstruction, the adaptive feature map is used to normalize the patch image to obtain a normalized patch image.

3. The anomaly detection method for flexible circuit board patch as described in claim 1, characterized in that, The step of performing background suppression processing on the normalized patch image to obtain a background-suppressed patch image includes: The structural features of the normalized patch image are extracted, and a background structure model of the flexible substrate is constructed by analyzing the gradient distribution characteristics in the neighborhood of each pixel of the normalized patch image. The background structure model is used to generate a background estimation map based on the structural features; The salient regions in the background estimation map are detected, and the foreground salience value of each salient region is calculated by analyzing the gradient consistency, texture complexity and contrast intensity of each salient region. Background segmentation is performed based on the foreground value to obtain a binary foreground mask; Multi-scale background suppression fusion is performed on the normalized patch image based on the binary foreground mask to obtain a background suppression patch image.

4. The anomaly detection method for flexible circuit board patch as described in claim 3, characterized in that, The step of performing multi-scale background suppression fusion on the normalized patch image based on the binary foreground mask to obtain a background suppression patch image includes: A multi-scale image list is constructed based on the normalized patch image in a multi-scale space; Perform pixel subtraction operations between the images in the multi-scale image list and the background structure model in different scale spaces to obtain a background difference image list; Based on the foreground mask, perform foreground adaptive gain on the background difference image list to obtain a gain image list; Multi-scale reconstruction is performed on the gain image list to obtain a reconstructed image list; The reconstructed image list is saliency-weighted and fused based on the foreground saliency value to obtain a background suppression patch image.

5. The anomaly detection method for flexible circuit board patch as described in claim 1, characterized in that, The abnormal bending identification of the background suppression patch image based on multi-scale gradient fusion and local structure consistency analysis yields bending identification results, including: Calculate the complex numerical gradient fields of the background suppression patch image at different scales to obtain the multi-scale complex numerical gradient tensor; Topological vortex detection is performed on the multi-scale complex numerical gradient tensor to obtain a multi-scale vortex feature map; Manifold consistency analysis was performed on the multi-scale complex numerical gradient tensor and the multi-scale vortex feature map to obtain a curvature anomaly map. Phonon vibration simulation is performed on the curvature anomaly map based on the original gradient tensor of the background suppression patch image to obtain a vibration instability thermogram. A pixel-level bending probability map is generated based on the multi-scale vortex feature map, the curvature anomaly map, and the vibration instability thermogram. Phase transition critical detection is performed on the pixel-level bending probability map to obtain a binarized bending recognition result.

6. The anomaly detection method for flexible circuit board patch as described in claim 1, characterized in that, The step of determining whether there are abnormal bends in the patches of the flexible circuit board based on the bending recognition result includes: Based on the bending recognition results, identify abnormal areas of the patch and obtain the total number of pixels in the abnormal areas; Identify the number of abnormal pixels in the abnormal region; The critical point density is obtained by calculating the ratio of the number of abnormal pixels to the total number of pixels. If the density at the critical point is greater than or equal to a preset density threshold, then an abnormal bend is determined to exist; otherwise, an abnormal bend is determined not to exist.

7. The anomaly detection method for flexible circuit board patch as described in claim 1, characterized in that, The calculation of the offset coefficient of the normalized patch image based on dual manifold alignment and differential saliency enhancement includes: The normalized patch image is subjected to dual manifold construction to obtain the patch manifold feature matrix and the pad manifold feature matrix; The manifold feature matrix and the pad manifold feature matrix are Riemann aligned according to a preset Riemann metric to obtain an alignment error matrix; A quantum gradient difference map is constructed based on the alignment error matrix and the normalized patch image; The quantum gradient difference map is saliency-sparse encoded to obtain an offset saliency map; The offset coefficients are calculated based on the offset saliency map and the alignment error matrix.

8. The anomaly detection method for flexible circuit board patch as described in claim 7, characterized in that, The step of constructing a dual manifold for the normalized patch image to obtain the patch manifold feature matrix and the pad manifold feature matrix includes: The normalized patch image is segmented into patch elements to obtain a patch segmentation mask; Based on the patch segmentation mask, the geometric features, appearance features, and texture features of the normalized patch image are extracted to obtain the patch feature vector; Construct patch manifold features based on the patch feature vectors; The design parameters of the pads in the normalized patch image are obtained according to the preset PCB design file; Based on the design parameters, the standard position features of the pads in the normalized patch image are extracted to obtain the pad manifold feature matrix.

9. The anomaly detection method for flexible circuit board patch as described in claim 7, characterized in that, The step of calculating the offset coefficient based on the offset saliency map and the alignment error matrix includes: Construct a multi-scale feature tensor of the offset saliency map, and fuse the multi-scale feature tensor with the alignment error matrix to obtain the fused features; The fused features are then subjected to physical perception weighting to obtain physically weighted features; The physically weighted features are compressed into a five-dimensional global feature vector by spatial average pooling. The offset coefficients are obtained by nonlinearly mapping the five-dimensional global feature vector using a pre-defined fully connected neural network.

10. An anomaly detection system for flexible circuit board patch mounting, characterized in that, The system includes: The image processing module is used to acquire a preset patch image in a flexible circuit board, perform illumination normalization processing on the patch image to obtain a normalized patch image, and perform background suppression processing on the normalized patch image to obtain a background suppressed patch image. The bending recognition module is used to select a first target user group from the candidate user group based on the matching value, and recommend the product to be recommended to the first target user group. The bending judgment module is used to determine whether there is abnormal bending in the patch of the flexible circuit board based on the bending recognition result. If there is, the abnormal detection result is determined to be abnormal bending. If there is no abnormal bending, the offset coefficient of the normalized patch image is calculated based on dual manifold alignment and differential saliency enhancement. The offset judgment module is used to determine whether the offset coefficient is greater than a preset coefficient threshold. If it is greater, the abnormal detection result is determined to be an abnormal offset; otherwise, the abnormal detection result is determined to be no abnormality.

Images