Method and equipment for detecting defects of AI enabling flexible circuit board

Abstract

The invention discloses an AI enabling flexible circuit board defect detection method and detection equipment, and relates to the technical field of AI detection. According to the invention, the dynamic image and the deformation data of the flexible circuit board are synchronously collected, the real-time deformation field model is fused and constructed, the receptive field and the weight of feature extraction are dynamically adjusted through the deformable attention calibration module, and defect feature offset caused by dynamic deformation is corrected. A lightweight multi-modal fusion algorithm is adopted, and visible light and infrared image data simplification and fusion detection are combined, so that comprehensive identification of surface and subsurface defects is realized. And collecting production line process parameters, establishing an association relationship between the defects and the process parameters through a graph neural network model, generating a process adjustment instruction, and feeding back the process adjustment instruction to a production line control system to form closed-loop regulation and control. According to the method and the equipment, the problems of missed detection and false detection of defects in a dynamic scene, adjustment lag caused by disjunction of the defects and the process, poor multi-equipment compatibility and the like are solved, the detection precision and efficiency are improved, and production line process optimization is facilitated.

Description

Technical Field

[0001] This invention relates to the field of AI inspection technology, specifically to an AI-enabled method and equipment for detecting defects in flexible printed circuit boards. Background Technology

[0002] Flexible printed circuit boards (FPCBs), with their thinness, flexibility, and high integration, have become core connection components for electronic devices such as smartphones and tablets. Their production quality directly determines the reliability of the end products. As electronic devices become smaller and more high-performance, the circuit density of FPCBs continues to increase, and the board thickness is further reduced. During high-speed production line transport, they are prone to dynamic bending, stretching, and torsional deformation due to factors such as the traction of the transport rollers and their own gravity.

[0003] Existing methods for detecting defects in flexible printed circuit boards (FPCBs) are mostly based on static image acquisition or simple AI algorithms, which are difficult to adapt to the needs of defect feature extraction in dynamic deformation scenarios. This leads to defect feature offset and blurring, resulting in serious missed detections and false detections. At the same time, existing detection technologies can only output defect results and cannot establish a correlation between defects and production process parameters. This makes production line process adjustments dependent on human experience, resulting in delayed and low-precision adjustments. It is impossible to suppress defect generation at the source, which seriously restricts the improvement of FPCB production efficiency and quality.

[0004] In view of the above, this application is hereby submitted. Summary of the Invention

[0005] The purpose of this invention is to provide an AI-enabled method and equipment for detecting defects in flexible printed circuit boards, so as to solve the problems mentioned in the background art.

[0006] To address the aforementioned technical problems, this invention provides an AI-enabled defect detection method for flexible printed circuit boards, comprising the following steps: S1 acquires dynamic detection images and deformation data of flexible circuit boards. S12 constructs a real-time deformation field model based on the dynamic detection images and deformation data. S13 uses a deformable attention calibration module to dynamically adjust the receptive field and weight of feature extraction according to the deformation field model to complete the adaptive calibration of defect features. S2 uses a lightweight multimodal fusion algorithm to perform multidimensional detection on the calibrated defect features and obtain the defect detection results. S3 collects production line process parameters and establishes the correlation between defect detection results and process parameters through a graph neural network model; S4 outputs process adjustment instructions based on the aforementioned correlation and feeds them back to the production line control system; this enables accurate defect detection in dynamic deformation scenarios of flexible circuit boards, avoiding missed or false detections caused by feature offsets. At the same time, lightweight design ensures real-time detection, and combined with process correlation analysis, it achieves linkage between defect prevention and process optimization.

[0007] Furthermore, in S1, dynamic detection images are acquired by a high-speed industrial camera, and acceleration and curvature data of the flexible circuit board are acquired by an attitude sensor as deformation data; the dynamic state data of the flexible circuit board is accurately obtained, providing a high-quality data source for subsequent deformation field modeling and improving the accuracy of deformation perception.

[0008] Furthermore, in S11, a real-time deformation field model is constructed by fusing pixel features and deformation data from dynamically detected images and using a temporal interpolation algorithm; this improves the real-time performance and accuracy of the deformation field model, ensuring that the model can dynamically match the deformation state of the flexible circuit board, and providing a reliable basis for feature calibration.

[0009] Furthermore, in S2, the lightweight multimodal fusion algorithm simplifies multimodal data by screening key spectral bands and infrared channels, and then uses a lightweight attention fusion network for feature fusion; while ensuring the accuracy of multimodal detection, it significantly reduces the amount of data processing and computational complexity, and adapts to the deployment requirements of embedded devices.

[0010] Furthermore, in S3, when establishing the correlation, the defect detection results are first classified and labeled, then aligned with the production line process parameters of the corresponding time period, and input into the graph neural network model for training; the quantitative correlation between defects and process parameters is clarified, the accuracy of correlation analysis is improved, and a more reliable decision-making basis is provided for subsequent process adjustments.

[0011] Furthermore, the AI ​​computing module uses an FPGA chip as the core computing unit and is equipped with a quantized and compressed neural network model; this enables high-speed inference of the neural network model, reduces equipment power consumption and size, and meets the stringent hardware requirements of real-time detection on the production line.

[0012] Furthermore, the process linkage module establishes communication with the production line control system through an industrial Ethernet interface, supporting real-time transmission and execution feedback of process adjustment commands; ensuring rapid transmission and execution of process adjustment commands, realizing closed-loop linkage between detection and production, and improving the timeliness of process optimization.

[0013] An AI-enabled flexible printed circuit board (FPCB) defect detection device includes an image acquisition module, a deformation sensing module, an AI computing module, a process linkage module, and an output module. The image acquisition module acquires images of the FPCB for inspection, the deformation sensing module acquires deformation data of the FPCB, the AI ​​computing module is equipped with a deformable attention calibration module, a lightweight multimodal fusion algorithm, and a graph neural network model, the process linkage module establishes communication between the AI ​​computing module and the production line control system, and the output module outputs defect detection results and process adjustment instructions. This device achieves integrated dynamic detection, multimodal analysis, and process linkage, with a compact structure adapted to production line needs. It can be directly embedded into existing production processes, improving the collaborative efficiency of detection and production optimization.

[0014] Compared with the prior art, the beneficial effects of the present invention are: The dynamic deformation perception and calibration scheme significantly improves the detection accuracy in dynamic scenes. Dynamic image acquisition and deformation data acquisition are performed simultaneously. By constructing a real-time deformation field model, the deformation state of each region of the plate is accurately captured. Combined with a deformable attention calibration network, the receptive field and weights of feature extraction are dynamically adjusted, effectively correcting defect feature shifts caused by dynamic deformation. This completely solves the problem of missed and false detections in high-speed production lines using traditional static detection methods. Furthermore, it requires no additional detection time, fully adapting to high-speed production cycles and significantly improving detection efficiency and dynamic adaptability.

[0015] The lightweight multimodal fusion design achieves a balance between comprehensiveness and real-time performance in defect detection. This invention integrates a miniature infrared acquisition unit into the image acquisition module, simultaneously acquiring surface and subsurface defect features. Through data simplification techniques such as channel filtering and feature reduction, combined with a lightweight fusion network and quantization compression optimization, the computational complexity is significantly reduced while maintaining detection accuracy. This addresses the pain points of traditional multimodal detection equipment, such as large size, slow processing, and difficulty in integration. It can accurately identify visible defects like surface scratches and short circuits, and effectively capture hidden defects such as internal delamination, resulting in more comprehensive detection dimensions and more flexible equipment deployment.

[0016] The defect-process linkage closed-loop mechanism achieves deep collaboration from detection to production optimization. This invention uses a process linkage module to collect key process parameters of the production line in real time, utilizes a graph neural network model to mine the nonlinear correlation between defects and process parameters, accurately locates the root cause of defects, and then feeds back process adjustment instructions, verified through simulation, to the production line control system, forming an automated closed-loop control. This solution breaks through the limitations of traditional separation of detection and process, solves the problems of lag and low accuracy in manual adjustments, suppresses defect generation at the source, significantly reduces the production line defect rate, reduces manual intervention costs, and improves the intelligence level of the production process.

[0017] The integrated design enhances the practicality and adaptability of the equipment. This invention organically integrates image acquisition, deformation sensing, AI computing, process linkage, and output modules. Each module works collaboratively and is customized for flexible circuit board inspection scenarios. For example, the AI ​​computing module uses a dedicated architecture to ensure efficient algorithm operation, the process linkage module supports industrial standard communication protocols to ensure compatibility, and the output module integrates display, alarm, and data storage functions to improve operational convenience. This design avoids compatibility issues caused by splicing multiple devices. The equipment has a compact structure and can be directly embedded into existing production processes without large-scale production line modifications, significantly reducing the application threshold and investment costs. Attached Figure Description

[0018] Figure 1 A flowchart of an AI-enabled defect detection method for flexible printed circuit boards. Detailed Implementation

[0019] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0020] Please see Figure 1 This invention provides a technical solution: an AI-enabled method and equipment for detecting defects in flexible printed circuit boards (PCBs). This embodiment is primarily applied to a high-speed production line for PCBs used in smartphones. This production line mainly produces PCBs for connecting the motherboard and display screen of smartphones. The PCBs are thin, narrow, and have a fixed length. Production is fast, and defects such as scratches, internal delamination, short circuits, and blistering of the cover film are prone to occur during the production process. Scratches are mostly caused by friction from the transfer rollers; internal delamination is caused by deviations in the lamination process parameters; short circuits are caused by incomplete etching; and blistering of the cover film is caused by uneven lamination temperature.

[0021] The detection method employed in this embodiment focuses on dynamic deformation sensing, lightweight multimodal fusion, and defect-process linkage as its core technical approaches. The supporting detection equipment integrates an image acquisition module, a deformation sensing module, an AI computing module, a process linkage module, and an output module, achieving fully automated detection from dynamic data acquisition and intelligent analysis to process feedback. The unique technical approach in this embodiment lies in its targeted innovation for the dynamic detection scenario of flexible printed circuit boards, addressing the shortcomings of existing publicly available technologies in dynamic deformation adaptation, lightweight integration, and process linkage, thereby significantly improving detection accuracy and production line adaptability.

[0022] Step S1: Based on integrated acquisition equipment, simultaneously acquire dynamic inspection images and deformation data of the flexible printed circuit board (FPCB). FPCBs are thin, flexible, and bendable. During high-speed production line transport, they undergo dynamic deformation due to the traction force of the transport rollers, their own weight, and ambient airflow, including continuous bending, slight stretching, and torsion. Existing inspection methods and equipment mostly use static acquisition methods, which can only inspect stationary or slowly moving FPCBs. Directly applying these methods to high-speed dynamic scenarios can lead to defect feature shifts and blurring, resulting in numerous missed and false detections. Therefore, this step requires the simultaneous acquisition of dynamic inspection images and deformation data using integrated acquisition equipment to provide basic data support for subsequent defect feature calibration and ensure the accuracy of inspection in dynamic scenarios.

[0023] Based on the principles of computer vision imaging and the theory of multi-source sensor data fusion, target detection in dynamic scenes requires the simultaneous acquisition of both visual and motion state information of the target. Accurate feature extraction is achieved through information complementarity. This step combines the real-time requirements of industrial visual inspection with the mechanical deformation characteristics of flexible circuit boards, designing an integrated "visual acquisition + attitude sensing" acquisition device to achieve simultaneous dual data acquisition. Specific technical methods are as follows: The integrated acquisition equipment used in this step includes an image acquisition submodule and a deformation sensing submodule, which work together through a data synchronizer. The core of the image acquisition submodule consists of a high-speed area scan camera, a ring LED light source, and a light source controller. The high-speed area scan camera has high resolution, a frame rate adapted to the production line speed, and its lens focal length is adjusted according to the size of the inspection area to ensure clear capture of the surface and edge details of the flexible circuit board. The ring LED light source adopts a diffuse reflection design, providing good light uniformity, effectively suppressing reflections on the flexible circuit board surface, and improving image contrast. The light source brightness can be adjusted in real time via the controller. The light source controller connects to the subsequent AI computing module, enabling automatic brightness adaptation.

[0024] The core of the deformation sensing submodule consists of a miniature three-axis accelerometer, a three-axis gyroscope integrated unit, a wireless transmission module, and a miniature power supply unit. The integrated sensor is small and lightweight, and its measurement range covers the deformation range during the flexible circuit board production process. Its sampling frequency matches the camera frame rate to ensure data synchronization. The wireless transmission module uses low-power technology, and its transmission distance is adapted to the width of the production line, avoiding interference with other equipment. The miniature power supply unit uses a button battery with a capacity adapted to the sensor's operating time, ensuring the detection of a single board. The sensor is fixed to the non-functional area of ​​the flexible circuit board using high-temperature resistant double-sided adhesive and moves synchronously with the board.

[0025] The data synchronizer is a key coordinating component of the acquisition equipment. It synchronizes the camera and sensor via trigger signals, preventing data timing discrepancies. During installation, the high-speed area array camera and ring light source are coaxially fixed to a frame above the production line's conveyor rollers, with the lens vertically aligned with the conveyor belt surface. The detection area covers the entire width of the flexible printed circuit board. The data synchronizer connects to the production line's photoelectric sensors. When the flexible printed circuit board moves with the conveyor belt to the entrance of the detection area, the photoelectric sensors trigger the data synchronizer, which simultaneously sends a start signal to both the high-speed area array camera and the deformation sensing submodule. The camera continuously captures dynamic images of the flexible printed circuit board at a set frame rate until the board completely leaves the detection area. The sensors simultaneously acquire the board's acceleration, angular velocity, and tilt angle data, with each frame corresponding to a set of sensor data. The acquired image data is transmitted via industrial Ethernet to the image preprocessing unit of the AI ​​computing module, while the sensor data is transmitted wirelessly to the data preprocessing unit of the AI ​​computing module. Both sets of data carry timestamps for subsequent timing alignment.

[0026] Example: On a flexible circuit board production line for smartphones, the production line operates at a fixed length per second. The flexible circuit board dimensions are fixed (length, width, and thickness), with precision circuitry printed on its surface and positioning holes along its edges. A high-speed area scan camera is set to a fixed frame rate per second, with its resolution adjusted to an appropriate value. The lens focal length is adjusted so that the detection area precisely covers the width and length of the board, ensuring that a single image can completely represent the entire board. The brightness of the ring LED light source is adjusted to a suitable value to eliminate reflections from the cover film on the board surface, achieving optimal contrast between the circuitry and the substrate. The integrated sensor of the deformation sensing submodule is attached near the positioning holes on the edge of the board; this area has no functional circuitry and will not affect subsequent processing.

[0027] When the board enters the detection area, the photoelectric sensor triggers the data synchronizer, and the camera begins continuous shooting. Each frame clearly captures the details of the lines, the state of the covering film, and the edge contours on the board surface. The integrated sensor simultaneously collects the board's acceleration data (reflecting the board's vibration and tensile state), angular velocity data (reflecting the board's torsional state), and tilt angle data (reflecting the board's bending state) during transmission. After acquisition, the image data is stably transmitted to the AI ​​computing module via industrial Ethernet, and the sensor data is transmitted to the AI ​​computing module via a wireless transmission module. The two sets of data are precisely aligned according to timestamps, preparing for subsequent deformation field modeling.

[0028] In existing publicly available technologies, the testing equipment disclosed in the flexible circuit board testing methods only uses a single camera to statically acquire images and lacks a deformation sensing component, thus failing to consider the impact of dynamic deformation on the testing. Furthermore, the multi-source data fusion technology disclosed in the multi-modal data fusion framework does not integrate attitude sensing and visual acquisition synchronously into the flexible circuit board testing equipment, nor does it employ a sensing mode that moves synchronously with the board. The unique technical approach of this step and its supporting equipment lies in designing an integrated acquisition device combining a high-speed camera and a miniature integrated sensor. A data synchronizer achieves temporal alignment between visual data and deformation data, and the deformation sensing submodule moves synchronously with the board to acquire dynamic data.

[0029] Compared to existing technologies, this step and equipment can effectively capture defect features and board movement under dynamic deformation, avoiding defect feature shifts caused by deformation, significantly reducing the missed detection rate and significantly improving detection accuracy; the synchronous acquisition mode does not require additional detection time, fully adapting to the cycle requirements of high-speed production lines, and significantly improving detection efficiency compared to static acquisition equipment; the deformation sensing submodule adopts a miniaturized design, is lightweight and small in size, and will not affect the normal transmission of flexible circuit boards, significantly enhancing equipment adaptability.

[0030] Step S11: Construct a real-time deformation field model based on the AI ​​computing module, integrating dynamic detection images and deformation data: In the dynamic detection images acquired in Step S1, the defect features of the flexible circuit board will undergo geometric shifts due to board deformation. Directly using this data for defect identification will lead to model misjudgment. Deformation data can only reflect the overall motion state of the board and cannot intuitively represent the deformation distribution in different areas of the board. Therefore, it is necessary to integrate the two data using the AI ​​computing module to construct a real-time deformation field model, accurately describing the deformation state of the board at the time of detection, clarifying the offset pattern of each pixel, and providing accurate deformation basis for subsequent defect feature calibration. The AI ​​computing module must have efficient data processing capabilities to ensure the real-time nature of model construction.

[0031] Based on image registration theory in computer vision and deformation field modeling principles in mechanics, a deformation field model can describe the deformation state of an object through the displacement vectors of pixels. This step, based on the synchronous data acquired in step S1, utilizes the parallel processing capabilities of the AI ​​computing module to construct a deformation field model through temporal interpolation and feature matching techniques, ensuring that the model can reflect the dynamic deformation of the plate in real time. Specific technical methods are as follows: The core execution component of this step is the AI ​​computing module. This module uses an industrial-grade core computing chip, memory unit, storage unit, and interface expansion unit. It is equipped with an embedded operating system and integrates algorithm modules such as data preprocessing, feature extraction, temporal alignment, and deformation field modeling. The specific technical process is divided into four parts: In the data preprocessing stage, the AI ​​computing module performs grayscale conversion and noise reduction on the dynamic detection image. Grayscale conversion uses a weighted average method to preserve the contrast between the line and the substrate; noise reduction uses a Gaussian filtering method to eliminate high-frequency noise in the image and avoid noise interference with feature point extraction; at the same time, the attitude sensor data is smoothed by using a moving average method to eliminate sensor measurement noise and preserve the true deformation trend of the board.

[0032] In the feature point extraction stage, the AI ​​computing module calls the built-in feature extraction algorithm and uses the SIFT algorithm to extract key feature points in the dynamic detection image. Stable feature points such as the edges of positioning holes and line intersections on the board are selected, and the pixel coordinates of each feature point are recorded. At the same time, based on the acceleration, angular velocity and tilt angle data collected by the attitude sensor, the overall deformation parameters of the board are calculated, including bending angle, elongation and torsion angle.

[0033] In the temporal alignment stage, the AI ​​computing module accurately aligns the preprocessed image feature point data with the deformation parameters based on the timestamps carried in the data in step S1, ensuring that each set of feature point data corresponds to a unique deformation parameter, thereby achieving deep fusion of visual data and sensor data.

[0034] In the deformation field modeling stage, the AI ​​computing module uses a temporal interpolation algorithm combined with feature matching results to construct a real-time deformation field model. Using a standard image of an undeformed flexible circuit board as a reference (the standard image is pre-stored in the AI ​​computing module's storage unit), the dynamic image acquired in step S1 is matched with the standard image for feature points, obtaining the pixel displacements of the matched feature point pairs. Combined with the aligned deformation parameters, the displacement vectors of non-feature point regions are supplemented using a temporal interpolation algorithm, forming a deformation field model covering the entire board area. This model is presented in matrix form, where each element represents the x and y displacement of the corresponding pixel. The real-time update frequency is consistent with the camera frame rate, and the modeling results are temporarily stored in memory for subsequent steps.

[0035] Example: Following the application scenario in step S1, after receiving the synchronously acquired dynamic images and deformation data, the AI ​​computing module immediately initiates the preprocessing process: A weighted average method is used to convert the color image into a grayscale image, with the lines appearing dark and the substrate light; Gaussian filtering is used to process the grayscale image to eliminate noise points caused by vibration during transmission. The attitude sensor data is processed using a moving average method to eliminate the sensor's own electronic noise, resulting in smooth acceleration, angular velocity, and tilt curves.

[0036] The AI ​​processing module uses the SIFT algorithm to extract feature points from the grayscale image. It selects the four vertices of the positioning holes on the board edge and key points at line intersections, extracting a fixed number of feature points and recording the pixel coordinates of each. Based on the attitude sensor data, it calculates that the bending angle, stretching rate, and torsion angle of the board at the detection time are fixed values. The feature point coordinates are aligned with the aforementioned deformation parameters according to the timestamp, ensuring that each feature point corresponds to the current deformation state of the board.

[0037] Using a pre-stored image of an undeformed standard flexible circuit board as a reference, the AI ​​computing module matches feature points in the dynamic image with corresponding feature points in the standard image, calculating the displacement vector of each feature point. Combined with the aligned deformation parameters, a temporal interpolation algorithm is used to supplement the displacement vectors of other pixels, constructing a real-time deformation field model of the board at that moment. The model clearly shows the pixel shift caused by bending in the middle area of ​​the board and the displacement caused by stretching in the edge area. The modeling results are temporarily stored in memory, providing a precise basis for subsequent calibration steps.

[0038] In existing publicly available technologies, the deformation field modeling methods disclosed in the graph neural network application methods only perform feature matching modeling based on image data, without incorporating attitude sensing data, and do not employ a dedicated AI computing module for real-time processing. Furthermore, commonly used deformation modeling equipment in industrial fields has low computational efficiency and cannot meet the real-time requirements of high-speed production lines. The unique technical approach of this step and its supporting AI computing module lies in integrating data fusion and deformation field modeling algorithms into the AI ​​computing module. It fuses image feature point data and attitude sensing data to construct a model, supplements displacement information in non-feature point regions through temporal interpolation algorithms, and synchronizes the model update frequency with the camera frame rate.

[0039] Compared to existing technologies, the deformation field model constructed by this step and equipment can not only accurately reflect the local deformation of each region of the plate, but also reflect the overall motion trend of the plate, significantly improving modeling accuracy. The parallel processing capability of the AI ​​computing module ensures the real-time nature of model construction, with the update frequency synchronized with the camera frame rate, avoiding calibration deviations caused by modeling lag, and significantly improving calibration accuracy compared to traditional modeling equipment. At the same time, the modeling algorithm is integrated into the AI ​​computing module, eliminating the need for additional dedicated modeling equipment, greatly improving equipment integration, and significantly reducing the amount of computation compared to traditional methods, fully adapting to the cycle time requirements of high-speed production lines.

[0040] Step S12: Adaptive calibration of defect features is achieved through the built-in calibration network in the AI ​​computing module: The deformation field model constructed in step S11 has clearly defined the offset rules of each pixel on the plate, but the defect features in the dynamic image are still in an offset state. If directly used for defect recognition, it will lead to model misjudgment. Existing calibration methods mostly use fixed parameter adjustments, which cannot adapt to the deformation differences in different regions, and the calibration equipment is separate from the detection equipment, resulting in low integration. Therefore, a deformable attention calibration network needs to be built into the AI ​​computing module to dynamically adjust the receptive field and weights of feature extraction according to the deformation field model, adaptively calibrate the offset defect features, restore the true shape of the defect, and improve the integration of the equipment.

[0041] Based on the principles of attention mechanisms in deep learning and the theory of image geometric transformation, the deformable attention mechanism achieves precise capture of target features by dynamically adjusting the receptive field. This step, based on pixel displacement information provided by the deformation field model, designs a deformable attention calibration network in the AI ​​computation module to perform targeted calibration of defect features, ensuring calibration accuracy and real-time performance. Specific technical methods are as follows: The core execution component of this step is the AI ​​computing module, which incorporates a deformable attention calibration network. This network, built on the Transformer architecture, includes a feature input layer, a deformation information fusion layer, a deformable attention layer, and a feature output layer. After pre-training, the network parameters are stored in the AI ​​computing module's memory. The specific technical process consists of four parts: the feature input layer retrieves the preprocessed dynamic image features from step S1 and the deformation field model data constructed in step S11 from the AI ​​computing module's memory, and simultaneously inputs them to the deformation information fusion layer.

[0042] The deformation information fusion layer uses a fully connected network to fuse the displacement vector of the deformation field model with image features and outputs fused features. The parameters of this layer are optimized through pre-training to ensure that the fused features can accurately reflect the correlation between defect features and deformation state.

[0043] The deformable attention layer is the core layer, dynamically adjusting the size and position of the receptive field based on the fused features. The AI ​​computation module calculates the adjustment parameters of the receptive field based on the displacement vector of each pixel in the deformation field model: for regions with large deformation, the receptive field range is increased to ensure complete coverage of the shifted defect features; for regions with small deformation, the receptive field range is reduced to improve the accuracy of feature extraction. The adjustment of the receptive field is achieved by adjusting the dimension of the attention weight matrix without changing the overall structure of the network.

[0044] In the feature weight allocation and calibration phase, the deformable attention layer assigns feature weights based on the degree of deformation in the deformation field model. Regions with greater deformation receive higher weights to ensure that significantly offset defect features are focused on during calibration; conversely, regions with less deformation receive lower weights to reduce unnecessary computation. Weight allocation is implemented using the Sigmoid function, mapping the degree of deformation to a fixed range of weight values. Subsequently, based on the adjusted receptive field and weight allocation results, attention weighting is applied to the fused features to correct offset defect features. During calibration, using pre-stored undeformed standard defect features in the AI ​​computing module as a reference, defect morphology is restored through feature alignment, and a calibrated defect feature map is output and temporarily stored in memory for subsequent defect detection.

[0045] Example: Continuing from the application scenario in step S11, the middle area of ​​the board undergoes significant deformation due to bending, resulting in a scratch defect in this area. The deformation causes the scratch's orientation in the image to deviate from its actual direction. The edge area experiences less deformation, but exhibits a blistering defect in the cover film. The AI ​​computing module calls the preprocessed dynamic image features and deformation field model data from the memory unit and inputs them into the built-in deformable attention calibration network.

[0046] The deformation information fusion layer fuses the displacement vector of the central region with the scratch features, outputting fused features. The deformable attention layer increases the receptive field based on the large deformation parameters of the central region to ensure coverage of the offset scratch features; it also reduces the receptive field based on the small deformation parameters of the edge regions to accurately capture bubbling features. During the feature weight allocation stage, the central region has a large degree of deformation and is assigned a high weight, while the edge regions have a lower weight.

[0047] During calibration, the network uses pre-stored standard scratch and blister features as references, correcting the directional shift of the scratches through attention weighting to restore their true shape; simultaneously, it precisely calibrates the position and size of the blister features. After calibration, the feature map output by the AI ​​computing module shows that both scratches and blister features have been restored to their true shape in their undeformed state. The feature map is temporarily stored in memory, providing accurate feature data for subsequent detection steps.

[0048] In existing publicly available technologies, the deformable attention module disclosed in the multimodal data fusion framework is only used for target detection in static images. It does not incorporate deformation field models for dynamic adjustment and is not integrated into a unified AI computing module. Traditional calibration equipment uses a fixed receptive field design, resulting in poor generalization ability and low integration due to its separation from the detection equipment. The unique technical approach of this step and its supporting AI computing module lies in embedding a deformable attention calibration network within the AI ​​computing module. This network fuses deformation field model data with image features, driving the network to dynamically adjust the receptive field and feature weights to achieve adaptive calibration of defect features.

[0049] Compared to existing technologies, this process and equipment can accurately calibrate defect features in different deformation areas, avoiding identification deviations caused by deformation and significantly improving defect identification accuracy. The network does not require parameter adjustments for different deformation types, significantly enhancing its generalization ability and adapting to the detection needs of different types of flexible circuit boards. The calibration network is integrated into the AI ​​computing module, eliminating the need for additional dedicated calibration equipment, greatly improving equipment integration and significantly reducing its size. At the same time, the computational load of this module is significantly reduced compared to traditional calibration methods, enabling real-time calibration and fully adapting to the cycle time requirements of high-speed production lines.

[0050] Step S2: Achieve multi-dimensional defect detection based on a lightweight multimodal acquisition and processing unit: Although the calibrated defect features restore the true shape, a single visual feature is insufficient to fully characterize the attributes of the defect. For example, surface scratches can be identified by visual features, but the visual features of sub-surface defects such as internal delamination are not obvious, making them prone to missed detection. Existing multimodal detection equipment is bulky and computationally complex, employing large processing units that cannot be integrated into conventional production line inspection equipment. Therefore, a miniature infrared acquisition unit needs to be added to the image acquisition module, and a lightweight multimodal fusion algorithm needs to be built into the AI ​​processing module to achieve comprehensive detection of surface and sub-surface defects, while ensuring real-time detection and equipment integration.

[0051] Based on multimodal data fusion theory and the lightweight design principles of deep learning models, multimodal fusion can improve detection accuracy by complementing the feature advantages of different modalities, while model lightweighting can reduce computational load through techniques such as channel pruning and quantization compression. This step acquires multi-source data through the multimodal acquisition unit of the image acquisition module, and the lightweight network built into the AI ​​computing module achieves fusion detection, meeting the integration requirements of the production line. Specific technical methods are as follows: The equipment involved in this step includes a miniature infrared acquisition unit for the image acquisition module and a lightweight fusion network unit for the AI ​​computing module. The image acquisition module adds a miniature infrared acquisition unit to the existing high-speed area scan camera. This small unit can be directly integrated into the image acquisition module's rack and works synchronously with the high-speed area scan camera to acquire infrared images of the flexible circuit board. Visible light images are used to capture the texture and morphological features of surface defects (scratches, blistering, short circuits), while infrared images are used to capture the temperature difference features of subsurface defects (internal delamination). Because the thermal conductivity of internal delamination regions differs from that of normal regions, they exhibit obvious temperature anomalies in the infrared images.

[0052] The AI ​​computing module incorporates a lightweight multimodal fusion network. This network, built upon a lightweight backbone network, includes a feature extraction branch, an attention fusion layer, and a detection output layer. After lightweight optimization and quantization compression, the network is stored in the AI ​​computing module's storage unit. The specific technical process consists of four parts: the multimodal data acquisition and synchronization stage, where a miniature infrared acquisition unit and a high-speed area array camera, under the control of a data synchronizer, simultaneously acquire visible light images (already calibrated) and infrared images of the flexible circuit board. The image data is transmitted to the AI ​​computing module via industrial Ethernet.

[0053] During the data simplification phase, the AI ​​processing module performs channel filtering on the infrared images, retaining key channels sensitive to temperature differences and eliminating invalid channels, significantly reducing the amount of infrared image data. For the calibrated visible light images, dimensionality reduction is performed, using principal component analysis to retain core features and remove redundant features. Through data simplification, the complexity of subsequent fusion operations is reduced while ensuring feature integrity.

[0054] In the feature fusion and detection phase, the AI ​​computing module utilizes a built-in lightweight fusion network. The feature extraction branch is divided into a visible light feature branch and an infrared feature branch, both employing lightweight convolutional layers to extract features and reduce network parameters. The attention fusion layer uses a lightweight attention mechanism to weight and fuse the features from the two branches, emphasizing the feature weights of defective regions. The detection output layer uses an improved detection head to simplify the network structure and improve detection speed. After network training, quantization compression techniques are used to compress the network model, further reducing model size and computational load.

[0055] In the output stage, the AI ​​computing module classifies and locates defects in the fused features through the detection output layer, outputs the defect type (scratches, internal delamination, short circuits, blistering of the cover film), defect location coordinates and defect size, forming the final defect detection result, which is temporarily stored in the storage unit and synchronously transmitted to the output module.

[0056] Example: Continuing from the application scenario in step S12, in addition to surface scratches and blistering of the cover film, the flexible circuit board for this smartphone also has an internal delamination defect in the middle area. This defect has no obvious features in the visible light image, but only appears as an abnormal area with a localized low temperature in the infrared image. Under the control of the data synchronizer, the miniature infrared acquisition unit of the image acquisition module and the high-speed area array camera synchronously acquire infrared images and calibrated visible light images of the board. The image data is transmitted to the AI ​​computing module via industrial Ethernet.

[0057] The AI ​​computing module initiates a data simplification process, filtering infrared images by channel selection, retaining channels sensitive to low-temperature anomalies and eliminating other invalid channels, thus reducing the data volume. For visible light image features, principal component analysis is used for dimensionality reduction, retaining the core features of scratches and blistering. Subsequently, the built-in lightweight fusion network is invoked: the visible light feature branch extracts the texture features of scratches and the morphological features of blistering, while the infrared feature branch extracts the temperature anomaly features of internal layering. The attention fusion layer performs weighted fusion of the three types of defect features, emphasizing the weight of internal layering features. The detection output layer accurately identifies the three types of defects: scratches, cover film blistering, and internal layering. It outputs that scratches are located in the middle area of ​​the board, blistering in the edge area, and internal layering in the middle area, while also outputting the location coordinates and size of each type of defect.

[0058] The test results are temporarily stored in the AI ​​computing module's storage unit and simultaneously transmitted to the output module, providing data support for subsequent process linkage. The entire testing process is completed in a short time, perfectly matching the transmission speed of the production line.

[0059] In existing publicly available technologies, the multimodal fusion algorithms disclosed in publicly available multimodal data fusion frameworks employ complex network architectures, resulting in high computational demands and requiring large computing devices, making them unsuitable for operation on embedded AI computing modules. Furthermore, the infrared acquisition units of traditional multimodal detection devices are bulky and cannot be integrated into conventional image acquisition modules, resulting in low integration. The unique technical approach of this process and its supporting equipment lies in integrating a miniature infrared acquisition unit into the image acquisition module and embedding a multimodal fusion algorithm based on a lightweight backbone network into the AI ​​computing module. This is combined with channel filtering, dimensionality reduction, and quantization compression techniques to achieve a lightweight design.

[0060] Compared to existing technologies, this process and equipment achieve simultaneous detection of surface and subsurface defects, significantly reducing the false negative rate. Lightweight network and data simplification technologies significantly reduce model size and computational load, enabling real-time operation on embedded AI computing modules and greatly improving detection speed. The miniature infrared acquisition unit is small in size and low in cost, and can be directly integrated into the rack of existing image acquisition modules without occupying additional production line space, significantly improving equipment integration and adapting to production line installation requirements. At the same time, quantization compression technology further reduces the power consumption of the AI ​​computing module, significantly reducing equipment operating costs.

[0061] Step S3: Establish the correlation between defects and process parameters through the collaboration of the process linkage module and the AI ​​computing module: Existing defect detection methods and equipment can only output defect results, but cannot identify the root cause of the defects. This leads to production line process adjustments relying on manual experience, resulting in low accuracy and long cycles. Furthermore, the lack of dedicated process linkage equipment prevents real-time interaction between defect data and production line process parameters. Therefore, a process linkage module needs to be configured to enable communication between the AI ​​computing module and the production line control system, collect production line process parameters, and establish a correlation between these parameters and defect detection results through the built-in correlation model of the AI ​​computing module. This allows for precise location of the process causes of defects, providing a scientific basis for subsequent process adjustments.

[0062] Based on the correlation analysis theory in data mining and the advantages of graph neural networks in relationship modeling, graph neural networks can effectively capture complex relationships between nodes. This step involves collecting process parameters through a process linkage module, constructing the correlation between defects and process parameters using a built-in correlation model in the AI ​​computing module, and achieving data interaction through the process linkage module, meeting the needs of industrial automation and AI inspection integration. Specific technical methods are as follows: This step involves equipment including a process linkage module and an AI computing module. The core of the process linkage module consists of an industrial communication unit, a protocol conversion unit, an instruction parsing unit, and an isolation protection unit, used to enable communication and data interaction between the AI ​​computing module and the production line control system. The AI ​​computing module has a built-in graph neural network-based correlation model used to construct the correlation between defects and process parameters. The specific technical process is divided into four parts: In the process parameter acquisition stage, the process linkage module establishes a connection with the production line control system through the industrial communication unit, acquiring key process parameters related to defects, including temperature, pressure, and time in the lamination process; etching solution concentration, etching time, and etching temperature in the etching process; lamination temperature, pressure, and time in the cover film lamination process; and transmission speed and transmission roller pressure in the transmission process. The acquisition frequency is consistent with the defect detection frequency to ensure that the defect result for each flexible circuit board corresponds to a unique set of process parameters. The acquired data is converted into a format recognizable by the AI ​​computing module through the protocol conversion unit and transmitted to the AI ​​computing module via the isolation protection unit.

[0063] In the data preprocessing stage, the AI ​​computing module performs outlier removal and normalization on the collected process parameters. Outlier removal uses the 3σ criterion to remove abnormal parameters caused by sensor malfunctions or human error. Normalization uses the min-max method to map process parameters of different magnitudes to the same range, avoiding the impact of parameter magnitude differences on model training. Simultaneously, the defect detection results are encoded, converting defect type, defect quantity, and defect severity into numerical features to form a defect feature vector.

[0064] In the association model construction and training phase, the AI ​​computing module incorporates an association model based on a graph neural network. The model includes an input layer, a graph convolutional layer, a fully connected layer, and an output layer. The input layer receives preprocessed process parameter vectors and defect feature vectors, treating them as graph nodes. The graph convolutional layer, as the core layer, captures the association between process parameters and defect features by calculating the adjacency matrix between nodes, extracting associated features. The fully connected layer performs nonlinear transformations on the associated features to enhance feature representation. The output layer outputs the association weights between process parameters and defect features, clarifying the degree of influence of each process parameter on the defects. The model is pre-trained using historical production data. During training, dropout technology is used to prevent overfitting and improve the model's generalization ability. The pre-trained model parameters are stored in the storage unit of the AI ​​computing module.

[0065] In the correlation result output stage, the AI ​​calculation module inputs the preprocessed process parameter vector and defect feature vector into the correlation model, obtains the correlation weight through model calculation, clarifies the main process causes of defects, and temporarily stores the correlation results in the storage unit and transmits them synchronously to the process linkage module to provide a basis for the generation of subsequent process adjustment instructions.

[0066] Example: Following the application scenario of step S2, the internal delamination, short circuits, and blistering defects of the cover film on this board are all related to the process parameters of the corresponding process. The process linkage module establishes a connection with the production line control system through the industrial communication unit to collect the process parameters during the production of this board: pressing temperature, pressure, and time; etching solution concentration, etching time, and temperature; bonding temperature, pressure, and time; and conveying speed and conveying roller pressure. The collected data is converted into a format recognizable by the AI ​​computing module by the protocol conversion unit and then transmitted to the AI ​​computing module through the isolation protection unit.

[0067] The AI ​​computing module initiates a data preprocessing process, using the 3σ criterion to remove outliers in the etching solution concentration and mapping all process parameters to a fixed range using the min-max method. Defect detection results are encoded as numerical features: internal layering, short circuits, and blistering are encoded with corresponding values. The defect count is 3, and the severity is medium, forming a defect feature vector. Subsequently, the built-in correlation model is invoked, inputting the process parameter vector and the defect feature vector. The graph convolutional layer calculates that the highest correlation weight is between pressing temperature and internal layering, etching time and short circuits, and bonding temperature and blistering. The fully connected layer reinforces these correlation features, and the output layer clearly states that low pressing temperature is the main cause of internal layering, excessive etching time is the main cause of short circuits, and uneven bonding temperature is the main cause of blistering.

[0068] The correlation results are temporarily stored in the storage unit of the AI ​​computing module and simultaneously transmitted to the process linkage module, providing accurate basis for the generation of subsequent process adjustment instructions. The entire correlation analysis process is completed automatically without manual intervention.

[0069] In existing publicly available technologies, the graph neural network correlation models disclosed in the publicly available graph neural network application methods are only used for the correlation between product quality and raw material parameters, and do not involve the correlation between flexible circuit board defects and process parameters, nor are they equipped with dedicated process linkage equipment. Commonly used correlation analysis equipment in the industrial field uses traditional statistical methods, which cannot capture complex nonlinear correlations, and is separate from detection equipment, resulting in low integration. The unique technical approach of this step and its supporting equipment lies in configuring a dedicated process linkage module to achieve communication between the AI ​​computing module and the production line control system. A defect-process correlation model based on graph neural networks is built into the AI ​​computing module, and dropout technology is used to improve the model's generalization ability.

[0070] Compared to existing technologies, this step and equipment can accurately capture the nonlinear correlation between process parameters and defects, significantly improving correlation accuracy and accurately locating the root cause of defects. The process linkage module enables real-time interaction between defect data and process parameters, eliminating the need for manual data transmission and greatly improving analysis efficiency compared to traditional statistical methods. The model has strong generalization ability and can adapt to the production processes of different types of flexible circuit boards without the need for model retraining. The correlation model is integrated into the AI ​​computing module, and the process linkage module and the AI ​​computing module work together, significantly improving equipment integration and eliminating the need for additional dedicated correlation analysis equipment, thus greatly reducing equipment investment costs.

[0071] Step S4: Based on the AI ​​computing module and the process linkage module, realize process adjustment and closed-loop feedback: The defect-process correlation established in Step S3 has clarified the process cause of the defect, but the correlation results need to be converted into specific process adjustment instructions and fed back to the production line control system to realize real-time adjustment of process parameters. In the existing technology, process adjustment is mostly done manually based on the detection results, which is lagging, has low accuracy, and lacks a closed-loop feedback mechanism. Therefore, it is necessary to generate process adjustment instructions through the AI ​​computing module, transmit them to the production line control system through the process linkage module, and receive adjustment feedback signals to realize closed-loop linkage between detection and production.

[0072] Based on industrial automation control theory and closed-loop control principles, closed-loop control can improve system stability by correcting system parameters in real time through feedback. This step, based on defect-process correlation results, generates adjustment instructions by the AI ​​calculation module, which are then fed back to the production line control system via the process linkage module. Upon receiving the adjustment feedback signal, closed-loop control is achieved. Specific technical methods are as follows: This step involves an AI computing module and a process linkage module. The AI ​​computing module has a built-in instruction generation unit and simulation verification unit, used to generate process adjustment instructions and verify their rationality. The process linkage module is responsible for instruction transmission and feedback signal reception. The specific technical process is divided into four parts: In the instruction generation stage, the instruction generation unit of the AI ​​computing module calls the pre-stored process parameter adjustment rule library based on the correlation weight and defect severity output in step S3. The rule library clearly defines the process parameter adjustment direction and adjustment range corresponding to different defect types and severity levels. For example, when the internal delamination severity is medium and the pressing temperature correlation weight is the highest, the adjustment direction is to increase the pressing temperature, and the adjustment range is determined according to the correlation weight; when the short circuit severity is medium and the etching time correlation weight is the highest, the adjustment direction is to shorten the etching time, and the adjustment range is determined according to the correlation weight. Based on the specific situation of the current defect, the corresponding rule in the rule library is matched to generate a specific process adjustment instruction.

[0073] During the instruction verification phase, the simulation verification unit of the AI ​​computing module uses simulation methods to verify the rationality of the adjustment instructions. The adjustment instructions are input into a pre-stored production line simulation model to simulate the production process of the flexible circuit board under the adjusted process parameters and predict defect occurrence. If the simulation results show a significant reduction in the defect rate, the instruction is deemed valid and sent to the process linkage module; if the defect rate does not change significantly or increases, the rule base is re-matched, and the instruction parameters are adjusted until the instruction verification is successful.

[0074] During the command transmission phase, the industrial communication unit of the process linkage module uses an industrial standard communication protocol to encapsulate verified process adjustment commands into data packets conforming to the production line control system protocol. These packets are then transmitted to the production line control system via an industrial communication cable, carrying the production batch information of the board to ensure that the commands accurately apply to the corresponding batch's production process. The isolation protection unit of the process linkage module employs opto-isolation technology to prevent electromagnetic interference from the production line from affecting command transmission, while also protecting the equipment from damage caused by voltage fluctuations.

[0075] In the closed-loop feedback phase, after receiving adjustment instructions, the production line control system adjusts the process parameters of the corresponding processes in real time: increasing the temperature of the lamination process, shortening the etching process time, and adjusting the temperature regulation device parameters of the bonding process to ensure uniform temperature distribution. After adjustment, the production line control system sends an adjustment completion signal to the process linkage module. The instruction parsing unit of the process linkage module analyzes this signal, extracts key information, and transmits it to the AI ​​calculation module. The AI ​​calculation module records the adjustment time, adjustment parameters, and adjustment effect, forming a process adjustment log, which is stored in the storage unit for subsequent effect tracking and rule base optimization.

[0076] Example: Continuing from the application scenario in step S3, the internal delamination of the board is caused by low pressing temperature, excessive etching time for short routes, and blistering by uneven bonding temperature. All defects are of medium severity. The AI ​​computing module's instruction generation unit calls the adjustment rule library and matches the corresponding adjustment rules: increase the pressing temperature, shorten the etching time, and adjust the temperature control device parameters for the bonding process to ensure uniform temperature distribution, generating specific process adjustment instructions.

[0077] The simulation verification unit of the AI ​​computing module adjusts the input instructions to the production line simulation model. The simulation shows that after adjustment, the defect rates of internal delamination, short circuits, and bubbling are significantly reduced, and the instruction verification is successful. The verified instructions are sent to the process linkage module. The industrial communication unit of the process linkage module uses the industrial standard communication protocol to encapsulate the instructions into data packets that conform to the production line control system protocol, carrying the production batch information of the board, and transmitting them to the production line control system through the industrial communication cable. The isolation protection unit effectively isolates the electromagnetic interference of the production line to ensure stable instruction transmission.

[0078] Upon receiving the instruction, the production line control system immediately adjusts the temperature controller parameters for the lamination process, increasing the lamination temperature; adjusts the timer parameters for the etching process, shortening the etching time; and adjusts multiple temperature sensors and heating devices for the bonding process to ensure uniform bonding temperature across all areas of the board. After adjustment, the production line control system sends an adjustment completion signal to the process linkage module. The instruction parsing unit analyzes this signal and transmits it to the AI ​​calculation module. The AI ​​calculation module records the adjustment parameters, time, and subsequent inspection results for this batch of boards, forming a process adjustment log, which is stored in the storage unit. Subsequent inspection results for this batch of flexible circuit boards show a significant reduction in the incidence of all three types of defects, demonstrating the significant effectiveness of the process adjustment.

[0079] In existing publicly available technologies, process adjustments in industrial automated production lines mostly employ open-loop control, lacking real-time feedback of AI detection results. The closed-loop control method disclosed in the flexible circuit board detection method does not incorporate graph neural network correlation analysis results, resulting in low adjustment accuracy and the absence of a dedicated process linkage module for stable command transmission. The unique technical approach of this step and its supporting equipment lies in integrating a command generation and simulation verification unit into the AI ​​computing module. Adjustment commands are generated and verified based on defect-process correlation weights and defect severity. Stable command transmission and closed-loop feedback are achieved through the process linkage module, which is equipped with an isolation protection unit and an industry-standard communication protocol.

[0080] Compared to existing technologies, this process and equipment achieve automated closed-loop linkage between defect detection and process adjustment, significantly reducing the lag time for process adjustment. The adjustment commands are generated based on accurate correlation analysis results and verified by simulation to ensure their rationality, resulting in significantly improved adjustment accuracy and a substantial reduction in the defect rate. The isolation protection unit of the process linkage module and the industrial standard communication protocol ensure the stability and compatibility of command transmission, making it compatible with production line control systems of different brands. The adjustment process requires no manual intervention, significantly reducing labor costs, while generating an adjustment log for easy follow-up and optimization, thereby improving the intelligence level of the production line.

[0081] Step S5: Presenting Detection Results and Equipment Status via Output Module: After detection and process adjustment are completed, the defect detection results, process adjustment instructions, and the working status of each module of the equipment need to be presented intuitively through the output module for easy real-time monitoring by operators. Existing output devices have poor display effects, limited alarm methods, and small data storage capacity, which cannot meet the actual needs of the production line. Therefore, an integrated output module is configured to realize result display, alarm prompts, and data storage functions, improving the operability and data traceability of the production line.

[0082] Based on industrial display and data storage technologies, high-definition industrial displays offer clear visuals, multi-mode alarms ensure timely operator response, and large-capacity storage meets data traceability requirements. This step utilizes an integrated output module to achieve result presentation and data management, aligning with the operational and traceability needs of industrial production. Specific technical methods are as follows: This step utilizes an integrated output module design, with core components including a high-definition industrial touchscreen, an audible and visual alarm unit, a large-capacity storage unit, and a data interface. It connects to the AI ​​computing module via the data interface to receive real-time inspection results, process adjustment information, and equipment status data. The high-definition industrial touchscreen is rack-mountable, boasts high resolution, and provides a clear display. It can display in real-time defect types, locations, quantities, severity, the operating status of each equipment module, and the execution status of process adjustment commands. It supports touch operation, allowing operators to query historical inspection data and process adjustment logs via the touchscreen.

[0083] The audible and visual alarm unit includes an alarm light and a buzzer, which are connected to the alarm signal output terminal of the AI ​​computing module. When a serious defect or equipment malfunction is detected, the AI ​​computing module sends an alarm signal to the audible and visual alarm unit, the alarm light flashes, the buzzer sounds an alarm, and the touch screen displays the cause of the alarm, ensuring that the operator can respond in a timely manner.

[0084] The high-capacity storage unit uses a mechanical hard drive, offering large capacity for storing long-term testing data, process adjustment logs, and equipment operation logs. Stored data is categorized by timestamps for easy retrieval. Data interfaces include a general data interface and a network interface, supporting the export of stored data to external storage devices or a host computer system for subsequent quality analysis and equipment maintenance.

[0085] The specific workflow is as follows: The AI ​​computing module transmits defect detection results, process adjustment instructions, and operating status data of each module of the equipment to the output module in real time; after receiving the data, the high-definition industrial touch screen displays relevant information according to a preset layout, and operators can query detailed data through the touch screen; when a serious defect or equipment failure is detected, the AI ​​computing module sends an alarm signal, the audible and visual alarm unit activates the alarm, and the touch screen displays the alarm reason simultaneously; all data is stored in real time to a large-capacity storage unit, and operators can export the data through the data interface for subsequent analysis.

[0086] Example: Continuing from the production line scenario described earlier, the high-definition industrial touchscreen of the output module is installed on the production line control panel and connected to the AI ​​computing module via a data interface. The touchscreen displays real-time defect information of the currently inspected board: scratches are located in the middle area, blistering of the cover film is located in the edge area, and internal delamination is located in the middle area, with 3 defects and moderate severity; it also displays process adjustment instructions: increase the bonding temperature, shorten the etching time, adjust the bonding temperature distribution, and the instruction execution status is "completed"; the equipment status shows that the image acquisition module, deformation sensing module, AI computing module, and process linkage module are all working normally.

[0087] When a serious short-circuit defect is subsequently detected on a board, the AI ​​computing module immediately sends an alarm signal to the output module. The red alarm light on the audible and visual alarm unit begins to flash, the buzzer emits a continuous alarm sound, and the touchscreen simultaneously displays the alarm reason: serious short-circuit defect, suggesting immediate inspection of etching process parameters. Operators can query detailed inspection data and corresponding process parameters for the board via the touchscreen, confirm the problem, and handle it promptly. All inspection data, process adjustment logs, and equipment operation logs are stored in real time to a large-capacity storage unit. At the end of the month, operators export the data to an external storage device via a general data interface for monthly quality analysis and equipment maintenance plan development.

[0088] Existing publicly available output devices mostly offer only a single display function, lacking integrated audible and visual alarms and large-capacity storage capabilities, and exhibiting poor communication stability with detection equipment. Traditional output devices also suffer from low display resolution, failing to clearly present complex defect information. The unique technical approach of the output module employed in this step lies in its integrated design, incorporating a high-definition industrial touchscreen, an audible and visual alarm unit, a large-capacity storage unit, and multiple data interfaces. This establishes stable communication with the AI ​​computing module, achieving integrated result display, alarm notification, and data storage.

[0089] Compared to existing technologies, this module offers higher display resolution, clearly presenting defect details and equipment status for easier viewing by operators; multiple alarm modes ensure timely detection of serious defects and equipment malfunctions, reducing quality risks; a large-capacity storage unit can store long-term data, meeting the traceability needs of the production line; multiple data interfaces enhance the flexibility of data export, facilitating subsequent quality analysis and equipment maintenance; and the integrated design reduces equipment footprint and investment costs while improving overall equipment coordination and stability.

[0090] Overall Collaborative Working Principle of the Equipment: In this embodiment, the AI-enabled flexible circuit board defect detection equipment integrates and installs various modules through a rack to form a complete detection system. During operation, the production line photoelectric sensor triggers the data synchronizer, and the image acquisition module and deformation sensing module synchronously acquire dynamic images and deformation data of the flexible circuit board. The data is transmitted to the AI ​​computing module. The AI ​​computing module sequentially completes data preprocessing, deformation field modeling, defect feature calibration, multimodal fusion detection, and defect-process correlation analysis, generating defect detection results and process adjustment instructions. The process linkage module transmits the process adjustment instructions to the production line control system and receives adjustment feedback signals to achieve closed-loop linkage. The output module presents the detection results, equipment status, and alarm information in real time, while storing relevant data.

[0091] Throughout the system, each module achieves collaborative control through an AI computing module. Data transmission is achieved via industrial Ethernet, wireless transmission, and industrial standard communication protocols, ensuring the real-time performance and stability of data transmission. The equipment adopts a lightweight and integrated design, adapting to the installation and cycle time requirements of high-speed production lines. Unique technical means are integrated into each step and corresponding module, significantly improving detection accuracy, efficiency, and equipment integration compared to existing publicly available technologies. This achieves full-process automation from dynamic detection to process optimization, significantly reducing production line defect rates and production costs.

Claims

1. An AI-enabled defect detection method for flexible printed circuit boards, characterized in that: Includes the following steps: S1 acquires dynamic detection images and deformation data of flexible circuit boards. S12 constructs a real-time deformation field model based on the dynamic detection images and deformation data. S13 uses a deformable attention calibration module to dynamically adjust the receptive field and weight of feature extraction according to the deformation field model to complete the adaptive calibration of defect features. S2 uses a lightweight multimodal fusion algorithm to perform multidimensional detection on the calibrated defect features and obtain the defect detection results. S3 collects production line process parameters and establishes the correlation between defect detection results and process parameters through a graph neural network model; S4 outputs process adjustment instructions based on the aforementioned correlation and feeds them back to the production line control system.

2. The AI-enabled flexible circuit board defect detection method as described in claim 1, characterized in that: In step S1, dynamic detection images are acquired using a high-speed industrial camera, and acceleration and curvature data of the flexible circuit board are acquired using an attitude sensor as deformation data.

3. The AI-enabled flexible circuit board defect detection method as described in claim 2, characterized in that: In step S11, a real-time deformation field model is constructed by fusing pixel features and deformation data from dynamically detected images and using a temporal interpolation algorithm.

4. The AI-enabled flexible circuit board defect detection method as described in claim 1, characterized in that: In S2, the lightweight multimodal fusion algorithm simplifies multimodal data by screening key spectral bands and infrared channels, and then uses a lightweight attention fusion network to perform feature fusion.

5. The AI-enabled flexible circuit board defect detection method as described in claim 1, characterized in that: In step S3, when establishing the association, the defect detection results are first classified and labeled, then aligned with the production line process parameters of the corresponding time period, and input into the graph neural network model for training.

6. The AI-enabled flexible circuit board defect detection method as described in claim 1, characterized in that: The AI ​​computing module uses an FPGA chip as the core computing unit and is equipped with a quantized and compressed neural network model.

7. The AI-enabled flexible circuit board defect detection method as described in claim 1, characterized in that: The process linkage module establishes communication with the production line control system through an industrial Ethernet interface, supporting real-time transmission and execution feedback of process adjustment commands.

8. An AI-enabled flexible circuit board defect detection device, applied to the AI-enabled flexible circuit board defect detection method according to any one of claims 1-7, characterized in that: It includes an image acquisition module, a deformation sensing module, an AI computing module, a process linkage module, and an output module. The image acquisition module is used to acquire inspection images of flexible circuit boards, the deformation sensing module is used to acquire deformation data of flexible circuit boards, the AI ​​computing module is equipped with a deformable attention calibration module, a lightweight multimodal fusion algorithm, and a graph neural network model, the process linkage module is used to establish communication between the AI ​​computing module and the production line control system, and the output module is used to output defect detection results and process adjustment instructions.

Images