Ic board detection data processing system based on edge computing

Abstract

The application discloses an IC carrier board detection data processing system based on edge computing and relates to the technical field of data processing.The application comprises a detection data analysis module and an edge-side local decision module.The detection data analysis module is used for receiving preprocessed edge feature data, defect confidence correlation data and defect judgment correlation data.The detection data analysis module comprises an edge feature unit, a defect confidence unit and a defect judgment unit.The edge-side local decision module is used for defect condition judgment and decision response.The application adapts to the characteristics of the edge computing architecture, completes the normalization conversion of three types of heterogeneous data through the edge feature unit, outputs standardized feature scores, completes candidate defect confidence evaluation and filters redundant data to lock the accurate range through the defect confidence unit, completes the final judgment of defects and the generation of root cause labels on the edge side through the defect judgment unit, and the results can directly drive the production line to execute, thereby meeting the requirements of low delay for real-time detection of the production line.

Description

Technical Field

[0001] This invention relates to the field of data processing technology, specifically to an IC carrier board detection data processing system based on edge computing. Background Technology

[0002] As advanced packaging processes evolve towards micro- and nano-scale, IC substrates, as core components carrying chip interconnection, heat dissipation, and protection, face increasingly stringent requirements for line width and hole precision. Even minute dimensional deviations and plating defects can lead to the failure of an entire batch of chips. Therefore, the full inspection requirements for IC substrates before leaving the factory are extremely stringent. Currently, online inspection of IC substrates generally adopts a multi-modal heterogeneous sensor fusion scheme, which simultaneously collects multi-dimensional data such as optical images, laser three-dimensional morphology, and electrical performance to achieve comprehensive defect coverage. However, traditional inspection systems have bottlenecks, thus requiring an IC substrate inspection data processing system based on edge computing.

[0003] The existing testing data processing system requires the complete transmission of multi-source testing data back to the remote cloud for feature processing and defect judgment. However, this method is constrained by the inherent uplink bandwidth of industrial field networks. The full data transmission of a single carrier board has a fixed time consumption. On the other hand, the cloud needs to process the massive amount of data reported by multiple workstations in parallel. The calculation queue will further lengthen the overall response time. It cannot adapt to the online testing cycle of high-speed transmission of high-density IC carrier boards and cannot meet the requirements of real-time quality inspection.

[0004] Secondly, existing detection data processing systems cannot complete the fusion calculation of multi-dimensional information at the edge because the original data feature benchmarks of different modalities are not unified and there is a lack of associative structured identifiers. The anomaly detection results of different modalities may be stored independently and in a scattered manner, which leads to a large amount of invalid data occupying transmission bandwidth and cloud computing resources, and may not be able to provide structured associated data support for subsequent defect root cause analysis. Summary of the Invention

[0005] The purpose of this invention is to provide an IC carrier board detection data processing system based on edge computing, which solves the problems mentioned in the background art.

[0006] To achieve the above objectives, the present invention provides an IC carrier board inspection data processing system based on edge computing, including an inspection data analysis module and an edge-side local decision-making module;

[0007] The detection data analysis module:

[0008] It is used to receive preprocessed edge feature data, defect confidence correlation data, and defect determination correlation data;

[0009] The detection data analysis module includes:

[0010] Edge feature unit, defect confidence unit, and defect determination unit;

[0011] The processing flow of the detection data analysis module is as follows:

[0012] Based on the confidence weight of the i-th mode, the original physical feature value of the i-th mode, the current computing power load coefficient of the edge node, the real-time contamination coefficient of the i-th sensor, the instantaneous transmission angle correction coefficient of the carrier board, and the transmission micro-vibration compensation coefficient of the i-th sensor in the edge feature data, the normalized feature score of the i-th mode is output.

[0013] Based on the historical defect correlation factor of the k-th type defect, the contribution weight of the i-th type mode to the k-th type defect, the current production line process fluctuation coefficient, the material batch adaptation coefficient of the k-th type defect, the aging attenuation coefficient of the detection head spot, the correlation factor between the distance of the k-th type candidate defect and the nearest neighbor via, and the self-calibration dynamic weight of the k-th type defect, and combined with the normalized feature score of the i-th type mode, the confidence score of the k-th candidate defect is output.

[0014] Based on the multimodal collaborative verification factor, global process threshold correction coefficient, edge node local temperature and humidity correction coefficient, adjacent carrier board spatial correlation factor in the same batch, and defect repairability weighting coefficient in the defect judgment association data, and combined with the normalized feature score of the i-th mode and the confidence of the k-th candidate defect, the final judgment value of the k-th defect is output.

[0015] Edge-side local decision-making module:

[0016] It is used to receive the normalized feature score of the i-th modality, the confidence score of the k-th candidate defect, and the final judgment value of the k-th defect, in order to determine the defect status and make a decision on how to respond.

[0017] Optionally, the detection data analysis module outputs:

[0018] The normalized feature score of the i-th modality is used to normalize the multimodal detection data into a unified and standardized feature score;

[0019] The confidence score of the kth candidate defect is used to perform preliminary screening and classification of candidate defects on the IC substrate, and the feature scores of the multimodal model are converted into confidence values ​​for the corresponding defect types.

[0020] The final determination value of the kth defect is used in conjunction with the local decision module on the edge side to achieve the final defect determination.

[0021] Optionally, the processing procedure for the edge feature unit is as follows:

[0022] A1. By dividing the IC substrate test data into three multimodal data, namely optical image mode, laser three-dimensional morphology mode and electrical performance test mode, and assigning confidence weights based on different mode types, the confidence weight of the i-th mode is obtained, so as to achieve the purpose of dynamic adjustment.

[0023] A2. By processing the detection data of multimodal types separately, the corresponding feature value data is output. The minimum and maximum values ​​of the original feature values ​​of the defect-free standard samples of the current batch of carrier plates under the i-th modality are combined to achieve the purpose of normalization.

[0024] A3. The CPU and GPU utilization rates are collected in real time through the system status monitoring module to calculate the load and output the current computing power load coefficient of the edge node.

[0025] A4. By comparing the characteristic values ​​of the standard calibration sheet with the factory standard characteristic values, the signal-to-noise ratio is compared to calculate the real-time contamination coefficient of the i-th type of sensor.

[0026] A5. By analyzing the coordinates of the Mark points on the IC carrier board, we can analyze the situation where the eigenvalues ​​are artificially high due to perspective distortion caused by the slight deflection of the carrier board during transmission, so as to calculate the instantaneous transmission deflection correction coefficient of the carrier board.

[0027] A6. By calculating the instantaneous vibration amplitude collected, the problems of image blurring caused by production line mechanical vibration and feature value false highness caused by point cloud offset are analyzed, so as to calculate the transmission micro-vibration compensation coefficient of the i-th type of sensor, and finally output the normalized feature score of the i-th type of mode.

[0028] Optionally, the processing procedure of the defect confidence unit is as follows:

[0029] B1. By introducing the normalized feature score of the i-th mode into the defect confidence unit, the purpose of cross-modal information fusion is achieved, and the common defect categories of IC carrier boards are specifically classified to realize the numbering of candidate defect types.

[0030] B2. By analyzing the defect data of the same workstation stored locally on the edge node in the last 24 hours, the number of times the same type of defect occurs in the same location is obtained, and then the historical defect correlation factor of the k-th type of defect is calculated.

[0031] B3. Based on different multimodal data, implement targeted weight assignment to obtain the contribution weight of the i-th modality to the k-th defect;

[0032] B4. By analyzing the deviation of the current process parameters of the production line, the process fluctuation coefficient of the current production line can be obtained.

[0033] B5. Obtain batch parameter data through the traceability QR code on the surface of the carrier board to match different IC carrier board materials in a targeted manner, so as to obtain the material batch adaptation coefficient of the k-th type of defect.

[0034] B6. The aging attenuation coefficient of the light spot of the detection head is analyzed by calculating the light power data collected by the light power sensor built into the detection head.

[0035] B7. By matching the pre-stored carrier board design vector diagram of the edge nodes, analyze the distance between the current candidate defect and its nearest neighbor via, and calculate the confidence level of the k-th candidate defect.

[0036] Optionally, the processing procedure of the defect determination unit is as follows:

[0037] C1. By introducing the normalized feature score of the i-th modality and the confidence score of the k-th candidate defect into the defect determination unit, the original modality features are used to cross the defect determination results to avoid single-modality misjudgment.

[0038] C2. Obtain the multimodal collaborative verification factor by statistically analyzing the normalized feature scores of the i-th mode of the three modes;

[0039] C3. By recalculating the temperature and humidity of the edge nodes, we can analyze the impact of temperature and humidity fluctuations in the edge node deployment environment on sensor accuracy and computing power stability, in order to obtain the local temperature and humidity correction coefficient of the edge nodes.

[0040] C4. By analyzing the defects of IC substrates in the same batch, the spatial correlation factor of adjacent substrates in the same batch can be obtained.

[0041] C5. By analyzing the repairability of the IC substrate, a repairability weighting coefficient for the defect is obtained. This coefficient is then combined with a global process threshold correction coefficient to finally output the final judgment value for the k-th defect.

[0042] Optionally, the edge-side local decision-making module includes a defect level analysis unit and an execution management decision-making unit.

[0043] Optionally, the defect level analysis unit specifically comprises:

[0044] When the final judgment value of the k-th defect is ≥0.9, it is classified as a severe defect.

[0045] When 0.7 ≤ the final judgment value of the kth defect < 0.9, it is classified as a minor defect.

[0046] When the final judgment value of the kth defect is <0.7, it is in the no-abnormal release level.

[0047] Optionally, the execution management decision unit specifically comprises:

[0048] When the defect is classified as severe, the edge node sends a level trigger signal to the sorting cylinder at the end of the production line via the industrial bus. The cylinder then pushes the carrier plate into the non-conforming product recycling station.

[0049] When the defect is classified as minor, the edge node sends a command to the laser marking machine next to the inspection station to mark the type and location of the defect on the designated marking area on the edge of the carrier plate, and writes it into the traceability QR code on the surface of the carrier plate.

[0050] When the board is in the no-abnormal release level, the edge node retains the no-defect release record of the carrier board and stores it locally at the edge for 30 days for subsequent traceability.

[0051] Compared with the prior art, the beneficial effects of the present invention are as follows:

[0052] I. This invention outputs the normalized feature score of the i-th modality through the edge feature unit. The edge feature unit integrates all dimensions of edge-side available parameters, such as multimodal original feature values, batch defect-free feature benchmark intervals, modal inherent confidence weights, edge node real-time computing power load coefficients, sensor real-time contamination correction coefficients, carrier plate transmission angle correction coefficients, and production line micro-vibration compensation coefficients, to complete the normalization conversion of three types of heterogeneous data: optical, laser, and electrical performance. At the same time, it offsets the systematic deviations caused by equipment status, transmission environment, and computing power fluctuations, and outputs standardized feature scores with a unified benchmark across batches and operating conditions. This provides a reliable basic input for subsequent defect judgment, significantly reduces the long-term maintenance requirements of manual calibration, and is fully compatible with the local autonomous architecture characteristics of edge computing.

[0053] Second, this invention outputs the confidence level of the k-th candidate defect through a defect confidence unit. The defect confidence unit integrates multi-dimensional local data, such as defect type identifier, modal contribution weights corresponding to different defects, historical defect correlation factors stored locally at the edge, real-time process fluctuation coefficient of the production line, material batch adaptation coefficient, detection head aging attenuation compensation coefficient, and correlation factor between defect and via distance, to complete the preliminary confidence level assessment of candidate defects at the edge. At the same time, it automatically filters redundant data without abnormal detection positions, which greatly reduces the computing power consumption of subsequent steps, adapts to the limited computing power limit of edge nodes, and the screening logic can be dynamically adapted according to the production line status and carrier board design characteristics, effectively balancing the accuracy and efficiency of the preliminary judgment, and providing an accurate range of candidate defects for the subsequent final judgment.

[0054] Third, this invention outputs the final judgment value of the k-th defect through the defect judgment unit. As the final decision output unit, the defect judgment unit integrates multi-dimensional business and environmental parameters such as multi-modal cross-cooperative verification factors, cloud asynchronous synchronous global process threshold correction coefficients, edge local environmental temperature and humidity correction coefficients, adjacent carrier board spatial correlation factors in the same batch, and defect repairability weighting coefficients. It completes the final judgment of defects and the generation of structured root cause labels at the edge. The final output result can directly drive production line sorting, labeling and other execution actions without the need for real-time cloud participation or secondary verification. The judgment logic fits the actual business needs of the production line, avoiding invalid misjudgments and missed detections. The output structured labels can be directly used for subsequent process traceability, fully meeting the low latency and autonomy requirements of real-time production line detection. Attached Figure Description

[0055] Figure 1 This is a block diagram of the system modules of the present invention;

[0056] Figure 2 This is a system flowchart of the detection data analysis module and the edge-side local decision module of the present invention;

[0057] Figure 3 This is a schematic diagram of the operation flow of the detection data analysis module of the present invention. Detailed Implementation

[0058] 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.

[0059] Please see Figures 1 to 3 This embodiment provides an IC carrier board inspection data processing system based on edge computing, including a multimodal data acquisition and processing module, an inspection data analysis module, and an edge-side local decision-making module.

[0060] Multimodal data acquisition and processing module:

[0061] This is used to acquire edge feature data, defect confidence correlation data, and defect judgment correlation data from IC carrier board inspection data, and to preprocess the acquired data so that the preprocessed data can be input into the inspection data analysis module.

[0062] The detection data analysis module includes edge feature units, defect confidence units, and defect determination units;

[0063] The processing flow of the detection data analysis module is as follows:

[0064] Firstly, based on the confidence weight of the i-th mode, the original physical feature value of the i-th mode, the current computing power load coefficient of the edge node, the real-time contamination coefficient of the i-th sensor, the instantaneous transmission angle correction coefficient of the carrier board, and the transmission micro-vibration compensation coefficient of the i-th sensor in the edge feature data, the normalized feature score of the i-th mode is output.

[0065] The normalized feature score of the i-th modality is used to normalize the multimodal detection data into a unified and standardized feature score.

[0066] Secondly: Based on the historical defect correlation factor of the k-th type of defect, the contribution weight of the i-th type of mode to the k-th type of defect, the current production line process fluctuation coefficient, the material batch adaptation coefficient of the k-th type of defect, the aging attenuation coefficient of the detection head spot, the distance correlation factor between the k-th type of candidate defect and the nearest neighbor via, and the self-calibration dynamic weight of the k-th type of defect, and combined with the normalized feature score of the i-th type of mode, the confidence score of the k-th candidate defect is output.

[0067] The confidence score of the kth candidate defect is calculated to perform preliminary screening and classification of candidate defects on the IC substrate, and the feature scores of the multimodal model are converted into confidence values ​​for the corresponding defect types.

[0068] Thirdly: Based on the multimodal collaborative verification factor, global process threshold correction coefficient, edge node local temperature and humidity correction coefficient, adjacent carrier board spatial correlation factor in the same batch, and defect repairability weighting coefficient in the defect judgment correlation data, and combined with the normalized feature score of the i-th mode and the confidence of the k-th candidate defect, the final judgment value of the k-th defect is output.

[0069] The final determination value of the kth defect is calculated and used in conjunction with the local decision-making module on the edge side to achieve the final defect determination.

[0070] Edge-side local decision-making module:

[0071] It is used to receive the normalized feature score of the i-th modality, the confidence score of the k-th candidate defect, and the final judgment value of the k-th defect, so as to determine the defect status and make a decision on how to respond.

[0072] Furthermore, the edge-side local decision-making module includes a defect level analysis unit and an execution management decision-making unit;

[0073] The defect level analysis unit is specifically as follows:

[0074] When the final judgment value of the k-th defect is ≥0.9, it is classified as a severe defect.

[0075] When 0.7 ≤ the final judgment value of the kth defect < 0.9, it is classified as a minor defect.

[0076] When the final judgment value of the kth defect is <0.7, it is in the no-abnormal release level;

[0077] The specific executive management decision-making unit is as follows:

[0078] When the defect is classified as severe, the edge node sends a level trigger signal to the sorting cylinder at the end of the production line via the industrial bus. The cylinder then pushes the carrier plate into the non-conforming product recycling station.

[0079] Edge nodes extract the structured information of the defect (defect type, coordinates, root cause label) and asynchronously transmit it back to the cloud when the network is idle. The original detection data (image, point cloud) is only cached locally at the edge for 7 days and then automatically deleted, without occupying real-time uplink bandwidth.

[0080] Furthermore, the judgment logic is as follows: defects with scores exceeding the threshold are considered irreparable and highly hazardous, and are directly judged as non-conforming products.

[0081] If three or more defects in the same range occur consecutively at the same workstation, the edge node directly sends a process warning signal to the production line PLC, triggering production line process inspection without cloud intervention.

[0082] When the defect is classified as minor, the edge node sends a command to the laser marking machine next to the inspection station to mark the type and location of the defect on the designated marking area on the edge of the carrier plate, and writes it into the traceability QR code on the surface of the carrier plate.

[0083] The edge node sends a release signal to the conveyor roller, and the carrier board flows directly into the subsequent repair process. The relevant defect data is only stored locally at the edge and used for information retrieval at the subsequent repair station. It is not actively transmitted back to the cloud.

[0084] Furthermore, the judgment logic is as follows: defects with scores in this range are repairable and low-hazard defects, and are not directly scrapped;

[0085] The results of this type of defect determination are simultaneously incorporated into the local iterative dataset for subsequent parameter iterative optimization during idle periods;

[0086] When the board is in the no-abnormal release level, the edge node retains the no-defect passage record of the carrier board and stores it locally at the edge for 30 days for subsequent traceability.

[0087] A release signal is sent to the conveyor rollers, and the carrier plate flows normally into the next production process without any additional handling.

[0088] Furthermore, the judgment logic is as follows: candidate defects with scores below the threshold are considered to be feature extraction errors or irrelevant perturbations, and are judged as having no abnormalities.

[0089] Based on the above, the three units of the system's detection data analysis module form a progressive detection closed loop adapted to the edge computing architecture. From the standardization of the underlying raw heterogeneous data, to the filtering of invalid data in the intermediate layer, and finally to the business decision output in the final layer, all core calculations are completed at the edge nodes of the production line. This fundamentally solves the high latency problem of traditional cloud-based detection architectures and adapts to the cycle time requirements of real-time quality inspection on the production line. The three units fully integrate the full-dimensional local data available on the edge side, and can complete accurate judgments without relying on full data synchronization in the cloud. At the same time, only the final high-value structured defect data is transmitted back on demand, which greatly reduces the bandwidth pressure on the industrial network and the computing load on the cloud. The output logic of the three units is completely aligned with the full-process business needs of IC substrate inspection. From feature unification to defect localization to final handling decisions, a complete self-consistent logic is formed, which directly supports the execution of automated quality inspection on the production line and subsequent root cause analysis of the process. No additional intermediate data processing links are required, realizing a fully autonomous closed loop for IC substrate inspection on the edge side.

[0090] Please refer to Figure 1 , Figure 2 as well as Figure 3 The processing flow for edge feature units is as follows:

[0091] A1. By dividing the IC substrate test data into three multimodal data, namely optical image mode, laser three-dimensional morphology mode and electrical performance test mode, and assigning confidence weights based on different mode types, the confidence weight of the i-th mode is obtained, so as to achieve the purpose of dynamic adjustment.

[0092] A2. By processing the detection data of multimodal types separately, the corresponding feature value data is output. The minimum and maximum values ​​of the original feature values ​​of the defect-free standard samples of the current batch of carrier plates under the i-th modality are combined to achieve the purpose of normalization.

[0093] A3. The CPU and GPU utilization rates are collected in real time through the system status monitoring module to calculate the load and output the current computing power load coefficient of the edge node.

[0094] A4. By comparing the characteristic values ​​of the standard calibration sheet with the factory standard characteristic values, the signal-to-noise ratio is compared to calculate the real-time contamination coefficient of the i-th type of sensor.

[0095] A5. By analyzing the coordinates of the Mark points on the IC carrier board, we can analyze the situation where the eigenvalues ​​are artificially high due to perspective distortion caused by the slight deflection of the carrier board during transmission, so as to calculate the instantaneous transmission deflection correction coefficient of the carrier board.

[0096] A6. By collecting the instantaneous vibration amplitude, we can analyze the problems of image blurring caused by production line mechanical vibration and feature value false highness caused by point cloud offset, so as to calculate the transmission micro-vibration compensation coefficient of the i-th type of sensor, and finally output the normalized feature score of the i-th type of mode.

[0097] The calculation formula for edge feature units is as follows:

[0098] ;

[0099] in:

[0100] AS i Refers to the normalized feature score of the i-th mode;

[0101] i refers to the modality number, which corresponds to the type identifier of multiple heterogeneous acquisition data, such as 1 for optical image mode, 2 for laser three-dimensional morphology mode and 3 for electrical performance test mode.

[0102] ASA i The confidence weight of the i-th mode is a fixed value preset by the system. Based on the factory accuracy calibration of the three types of sensors, this embodiment presets the confidence weight to 0.85 for the optical mode, 0.92 for the laser mode, and 0.98 for the electrical performance mode. This parameter is introduced to distinguish the detection reliability priority of different modes, increase the weight ratio of high-precision sensor features, and reduce the interference of low-precision sensor data on the final result.

[0103] ASB i This refers to the original physical feature value under the i-th mode, that is, the original physical feature value extracted from the detection position under the i-th mode. Different modes correspond to different acquisition paths. The optical mode is obtained by taking grayscale images of the carrier plate surface by the linear CCD camera at the detection station and calculating them by the edge extraction operator. The laser mode is obtained by scanning the carrier plate surface with the linear laser displacement sensor at the detection station to obtain a three-dimensional point cloud and calculating the size. The electrical performance mode is obtained by applying pressure to the carrier plate pins by the continuity test probe.

[0104] Specifically:

[0105] The optical mode uses a CCD camera to acquire 8-bit grayscale images, and the Sobel edge extraction operator is used to calculate the linewidth deviation and connected area of ​​the defect region, outputting physical size values ​​in μm units.

[0106] The laser mode uses a laser displacement sensor to scan and obtain a three-dimensional point cloud, calculates the height difference and step offset of the defect area, and outputs physical size values ​​in μm units.

[0107] Electrical performance mode: The digital multimeter applies voltage to the pins of the carrier board through the continuity probe, collects the continuity resistance value, and outputs the electrical performance value in mΩ units;

[0108] The original physical eigenvalue ASB under the i-th mode i The introduction of this is the core raw input for defect detection, which directly reflects the degree of physical feature anomaly at the corresponding location on the carrier board and provides a basic data source for subsequent normalization calculations;

[0109] ASC i and ASD i These refer to the minimum and maximum original feature values ​​of the defect-free standard samples of the current batch of carrier plates under the i-th modality. They can be obtained by pre-inspecting five defect-free standard carrier plates before each batch of production and statistically analyzing the corresponding feature values. The calculation formulas are ASC and ASC, respectively. i =min(ASC) i，1 , ASC i，2 , ..., ASC i，5 ), ASD i =max(ASC) i，1 , ASC i，2 , ..., ASC i，5 ), where ASC i，1 To ASC i，5 These are the original physical characteristic values ​​of the i-th type of 5 standard carrier plates;

[0110] ASC i and ASD i The introduction of this feature provides a benchmark range for feature normalization, eliminates the systematic bias caused by the difference in the baseline of the basic features of different batches of carrier boards, and ensures the consistency of the normalization score across batches.

[0111] ASE refers to the current computing load coefficient of the edge node. It can be calculated in real time by the system status monitoring module built into the edge feature unit by collecting CPU and GPU utilization. The CPU utilization (take the average of the last 5 times and convert it to a decimal of 0-1) and NPU inference load (take the average of the last 5 times and convert it to a decimal of 0-1) can be collected every 1 second. The final formula is ASE = min [max(CPU average load, NPU average load), 0.9], that is, when the load exceeds 90%, it is calculated as 0.9.

[0112] The introduction of the edge node current computing power load factor ASE automatically reduces the complexity of feature calculation when the edge node computing power is tight, balances the real-time performance and accuracy of detection, and avoids detection lag and data packet loss caused by computing power overload.

[0113] ASF iThe real-time contamination coefficient refers to that of the i-th type of sensor. It is calculated by comparing the characteristic values ​​of the standard calibration sample collected every 10 minutes with the factory standard characteristic values, using the following formula: ASF. i = Factory standard signal-to-noise ratio ÷ Actual standard sample feature signal-to-noise ratio. When the value is lower than 0.8, a sensor cleaning alarm is triggered.

[0114] The optical lenses and laser heads on the edge side operate in the production line environment for a long time, which will accumulate solder fumes and anti-static dust, directly reducing the signal-to-noise ratio of the collected data. Most existing systems only perform periodic calibration and do not incorporate it into the calculation in real time.

[0115] Real-time contamination coefficient ASF of this type i sensor i The introduction of this technology is used to compensate for the decrease in signal-to-noise ratio caused by dust and smoke pollution on the sensor lens and probe in real time, so as to maintain the stability of detection accuracy without frequent manual calibration.

[0116] ASG refers to the instantaneous transmission deflection correction coefficient of the carrier plate, which can be calculated by collecting the coordinates of the Mark point of the carrier plate by the vision positioning camera at the entrance of the inspection station. The calculation formula is ASG = 1 - |actual deflection angle| × 0.04, where the actual deflection angle is the angle between the reference edge of the carrier plate and the reference line of the transmission roller, and the value ranges from 0.92 to 1.0.

[0117] Even slight offsets of the carrier plate on the conveyor rollers can cause systematic deviations in the features acquired by optical and laser acquisition. Most existing systems rely on mechanical positioning to ignore this effect.

[0118] The introduction of the instantaneous transmission angle correction coefficient ASG is used to offset the problem of falsely high feature values ​​caused by perspective distortion due to the slight angle of the carrier during transmission. This can improve the detection accuracy of size features without the need for additional high-precision mechanical positioning structures.

[0119] ASH i The transmission micro-vibration compensation coefficient, referring to the i-th type of sensor, can be calculated by collecting the instantaneous vibration amplitude from the accelerometer built into the transmission roller conveyor at the detection station. The calculation formula is ASH. i = 1 - amplitude × 50, where amplitude is the collected vertical vibration amplitude value in mm, with a value range of 0.95-1.0;

[0120] Minor mechanical vibrations of the production line roller conveyor can cause random shifts in the features acquired by optical and laser acquisition. Most existing systems only perform mechanical vibration reduction hardware processing and do not incorporate real-time vibration data into feature correction.

[0121] The transmission micro-vibration compensation coefficient ASH of the i-th type of sensor iThe introduction of this technology is intended to counteract the image blurring and point cloud offset caused by mechanical vibrations in the production line, thereby reducing the performance requirements of the production line vibration damping hardware and making it suitable for the installation environment of ordinary industrial production lines.

[0122] Based on the above, this edge feature unit is a low-level data processing unit adapted to the edge computing architecture. All calculations are completed locally on the production line edge node, without the need to send the original multi-source heterogeneous data back to the cloud. This reduces the system's dependence on uplink transmission bandwidth from the source and avoids the latency caused by the original data transmission. This unit is also compatible with the dynamic change characteristics of the computing power of the edge node. It can adjust the calculation weight according to the real-time load on the edge side to ensure the stability of the edge processing. At the same time, it incorporates local dimensions such as the sensor status and carrier board transmission status that can be directly collected on the edge side into the correction logic. It can complete the offsetting of the deviation of the original feature without relying on the remote calibration of the cloud, which is in line with the local autonomous architecture characteristics of edge computing.

[0123] Normalized feature score AS of the i-th mode i It serves as the standardized foundational input for all subsequent calculations in the entire detection system. It transforms the three completely heterogeneous raw acquisition data—optical, laser, and electrical performance data—which are not uniform with the benchmark, into standardized feature scores within a unified range. This eliminates the systematic bias of the raw data caused by different modalities, different equipment states, and different transmission environments. It eliminates the need for differentiated threshold adaptation for different modalities in the subsequent defect judgment process, simplifies the calculation logic on the edge side, and avoids subsequent judgment errors caused by heterogeneous data characteristics.

[0124] Normalized feature score AS of the i-th mode i The larger the result, the higher the degree to which the original features extracted from the corresponding modality at the detection location deviate from the normal defect-free baseline, and the higher the probability that there is an identifiable anomaly of the corresponding modality at that location. The smaller the result, the closer the features of the corresponding modality are to the baseline state of the normal carrier plate, and the higher the probability that there is no identifiable anomaly of the corresponding modality at that location.

[0125] Please refer to Figure 1 , Figure 2 as well as Figure 3 The processing flow for defect confidence units is as follows:

[0126] B1. By introducing the normalized feature score of the i-th mode into the defect confidence unit, the purpose of cross-modal information fusion is achieved, and the common defect categories of IC carrier boards are specifically classified to realize the numbering of candidate defect types.

[0127] B2. By analyzing the defect data of the same workstation stored locally on the edge node in the last 24 hours, the number of times the same type of defect occurs in the same location is obtained, and then the historical defect correlation factor of the k-th type of defect is calculated.

[0128] B3. Based on different multimodal data, implement targeted weight assignment to obtain the contribution weight of the i-th modality to the k-th defect;

[0129] B4. By analyzing the deviation of the current process parameters of the production line, the process fluctuation coefficient of the current production line can be obtained.

[0130] B5. Obtain batch parameter data through the traceability QR code on the surface of the carrier board to match different IC carrier board materials in a targeted manner, so as to obtain the material batch adaptation coefficient of the k-th type of defect.

[0131] B6. The aging attenuation coefficient of the light spot of the detection head is analyzed by calculating the light power data collected by the light power sensor built into the detection head.

[0132] B7. By matching the pre-stored carrier board design vector diagram of the edge nodes, analyze the distance between the current candidate defect and its nearest neighbor via, and calculate the confidence level of the k-th candidate defect.

[0133] The calculation formula for the defect confidence unit is as follows:

[0134] ;

[0135] in:

[0136] BD k The confidence level refers to the k-th type of candidate defect;

[0137] The introduction of the normalized feature score ASi for the i-th modality enables cross-modal information fusion, which not only solves the pain point that different types of detection data cannot be calculated uniformly, but also significantly reduces the computing power consumption on the edge side, quickly completes the preliminary classification and priority sorting of defects, filters out abnormal data, and avoids invalid calculations occupying edge resources.

[0138] k refers to the type number of the candidate defect, which corresponds to the common defect categories of IC substrates. It can be preset with enumeration values ​​and includes common defect types such as open circuit, short circuit, uneven plating, hole position misalignment and foreign matter attachment. It is used to provide classification identifiers for the differentiated weight allocation and correction logic of different defect types, and to support the classification and subsequent source tracing of defects.

[0139] BDA k The historical defect correlation factor, referring to the k-th type of defect, can be obtained by statistical analysis of the most recent 24-hour defect data from the same workstation stored locally on the edge node. The calculation formula is BDA. k = 0.8 + 0.4 × (number of times the same type of defect occurs at the same location ÷ total number of inspection plates × 1000), with a value range of 0.8-1.2. The total number of inspection plates is the total number of inspection plates at this workstation in the most recent 24 hours.

[0140] Historical defect correlation factor BDA for type k defect k The introduction of this technology utilizes historical detection data stored locally on the edge side to improve the detection priority of recurring potential process defects and help quickly locate production line process fluctuations.

[0141] BDB k，i This refers to the contribution weight of the i-th mode to the k-th defect, i.e., a fixed value preset by the system. It can be calibrated based on the significance of different defects in each mode. For example, an open-circuit defect corresponds to a weight of 0.7 for the electrical performance mode, 0.2 for the optical mode, and 0.1 for the laser mode. ;

[0142] The contribution weight of the i-th mode to the k-th defect is BDB. k The introduction of this method is used to match the most sensitive detection modality for different defect types, improve the detection capability of specific defects, and reduce the interference of irrelevant modal data.

[0143] BDC refers to the current production line process fluctuation coefficient, which can be obtained in real time by edge nodes communicating with the production line PLC through the industrial bus. The value range is 1.0-1.3. It is automatically issued by the production line according to the deviation of the current process parameters (such as electroplating current and pressing temperature). It is used to automatically improve the defect detection sensitivity when the production line process is unstable, capture batch defects caused by process fluctuations in a timely manner, and avoid the outflow of batch defective products.

[0144] BDD k The material batch adaptation coefficient, which refers to the k-th type of defect, can be obtained by scanning the traceability QR code on the surface of the carrier plate by the edge node to obtain the batch parameters and match the preset threshold offset of different material batches. The value range is 0.9-1.1.

[0145] There are slight differences in the substrate, copper foil thickness, and plating material of IC substrates from different batches. The feature threshold of the same type of defect will shift. Most existing systems use a uniform threshold and do not dynamically adjust with batch. The introduction of this parameter is to adapt to the differences in feature baselines of substrates and plating materials from different batches, so that the detection parameters do not need to be recalibrated every time the batch is changed, thus shortening the changeover time of the production line.

[0146] BDE refers to the aging attenuation coefficient of the light spot in the detection head. It can be calculated by the light power sensor built into the detection head in real time by collecting the light emission power. The calculation formula is BDE = min [1 + (rated power - actual power) ÷ rated power × 1.5, 1.15]. The rated power is a fixed value calibrated at the factory. When the power is lower than 90% of the rated value, a replacement alarm is triggered.

[0147] The light emission power of laser / optical detection heads will decrease linearly with the duration of use, resulting in an overall low value of the collected features. Most existing systems only perform quarterly calibration and do not correct in real time.

[0148] The introduction of the spot aging attenuation coefficient (BDE) of the detection head is used to compensate for the attenuation of the light emission power of laser and optical detection heads after long-term use, extend the calibration cycle of the detection head, and reduce equipment maintenance costs.

[0149] BDF k The distance correlation factor between the k-th candidate defect and its nearest neighbor via is calculated by matching the defect coordinates with the pre-stored carrier plate design vector diagram of the edge nodes. The calculation formula is BDF. k =min[1+(50μm-actual spacing)÷10μm×0.04,1.16],actual spacing is in μm,spacing determination accuracy is ±5μm,weight boosting logic is activated when the spacing is less than 50μm;

[0150] The vias of IC substrates are the core structure of interlayer interconnection. If the distance between the defect and the edge of the via is less than 30μm, it will greatly increase the implicit reliability risk of hole wall cracking and interconnect short circuit in subsequent electroplating and lamination processes. Even if the size of the defect itself is much smaller than the conventional judgment threshold, the detection priority needs to be increased. However, the mainstream detection algorithm of the existing system only judges the size and shape of the defect itself, and hardly associates the spatial position of the defect with the surrounding key interconnect structure in real time.

[0151] The correlation factor between the distance between the k-th type candidate defect and the nearest neighbor via (BDF) k The introduction of this feature is intended to increase the detection priority of defects that are too close to vias, prevent latent reliability defects from flowing into subsequent processes, and improve the interconnect reliability of the carrier board.

[0152] The introduction of the min function is used to truncate the upper limit of confidence, which is the built-in numerical constraint logic of the system. If the original calculation result is greater than 1, output 1; otherwise, output the original result. This ensures that the unit output is always within the normalized range of 0-1, which matches the logic of the subsequent judgment threshold and avoids the score overflow problem caused by coefficient superposition.

[0153] Based on the above, this defect confidence unit is an intermediate layer defect screening unit on the edge side. All calculations are completed locally on the edge node without cloud involvement. This unit integrates historical detection data stored locally on the edge, real-time synchronized process status data from the production line, and pre-stored carrier board design structure data. It can complete the correlation calculation of multi-dimensional data without sending data back to the cloud, ensuring the real-time performance of defect screening, adapting to the dynamic status of the current production line and the design characteristics of the carrier board, and filtering out most abnormal detection locations in advance, significantly reducing the amount of calculation in the subsequent final judgment stage, and adapting to the limited computing power limit of the edge node.

[0154] The confidence level of the k-th candidate defect (BD) kThe calculation completed the preliminary screening and classification of candidate defects, converted the standardized multimodal feature scores into confidence values ​​corresponding to specific defect types, and completed the preliminary filtering of non-abnormal locations, retaining only candidate defects that meet the preliminary thresholds to enter the next stage of judgment. This significantly reduced the proportion of invalid calculations on the edge side and provided preliminary defect type and location markings for subsequent final judgment and root cause analysis, reducing redundant calculations in subsequent stages.

[0155] The confidence level of the k-th candidate defect (BD) k The larger the value, the higher the degree to which the candidate position matches the corresponding defect type feature, the higher the probability that it belongs to a real defect, and the higher the priority of the potential impact of the defect; the smaller the value, the lower the probability that the candidate position belongs to a real defect, and the more likely it is a feature extraction error or an irrelevant small perturbation, which will be filtered out in the preliminary screening stage and will not be included in the subsequent calculation.

[0156] Please refer to Figure 1 , Figure 2 as well as Figure 3 The processing flow of the defect determination unit is as follows:

[0157] C1. By introducing the normalized feature score of the i-th modality and the confidence score of the k-th candidate defect into the defect determination unit, the original modality features are used to cross the defect determination results to avoid single-modality misjudgment.

[0158] C2. Obtain the multimodal collaborative verification factor by statistically analyzing the normalized feature scores of the i-th mode of the three modes;

[0159] C3. By recalculating the temperature and humidity of the edge nodes, we can analyze the impact of temperature and humidity fluctuations in the edge node deployment environment on sensor accuracy and computing power stability, in order to obtain the local temperature and humidity correction coefficient of the edge nodes.

[0160] C4. By analyzing the defects of IC substrates in the same batch, the spatial correlation factor of adjacent substrates in the same batch can be obtained.

[0161] C5. By analyzing the repairability of the IC substrate, a repairability weighting coefficient for the defect is obtained. Combined with the global process threshold correction coefficient, the final judgment value of the kth defect is finally output.

[0162] The calculation formula for the defect judgment unit is as follows:

[0163] ;

[0164] in:

[0165] CR k Refers to the final judgment value of the k-th defect;

[0166] Normalized feature score AS of the i-th mode i The confidence level BD of the k-th candidate defect k The introduction of this feature allows for cross-validation of the initial defect judgment results of the defect confidence unit using the original modal features of the edge feature unit, avoiding misjudgment of a single modality. At the same time, combined with business logic such as process rules and repairability, it directly outputs the final decision score that can be executed directly.

[0167] CRA stands for Multimodal Collaborative Verification Factor, which is the normalized feature score of the i-th mode of three modes, calculated by edge nodes. i The scores were calculated based on the three modalities of AS. i If all values ​​are ≥0.6, then the value is 0.2. If there is only one type of AS... i If the value is ≥0.6, the value is -0.2; otherwise, the value is 0.

[0168] The introduction of the multimodal collaborative verification factor (CRA) reduces the probability of single-modal false positives by utilizing cross-validation of multimodal data, thereby improving the accuracy of defect identification and reducing unnecessary false positives.

[0169] CRB stands for Global Process Threshold Correction Coefficient, which is generated by the cloud platform and distributed to edge nodes weekly after compiling defect rate data for the entire production line. The value ranges from 0.95 to 1.05.

[0170] The introduction of the Global Process Threshold Correction Factor (CRB) enables the synchronous alignment of edge-side detection standards with global process standards, adapting to process iterations across the entire production line without the need for manual threshold adjustment at each edge node.

[0171] CRC refers to the local temperature and humidity correction coefficient of the edge node. It is calculated by collecting environmental data in real time from the temperature and humidity sensors built into the edge node. The calculation formula is CRC = 1 + min(|average temperature - 25| ÷ 5, 1) × 0.03 + min(|average humidity - 50| ÷ 10, 1) × 0.03, where the temperature unit is ℃ and the humidity unit is %RH, and the value range is 1-1.06.

[0172] The introduction of the local temperature and humidity correction coefficient (CRC) at edge nodes compensates for the impact of temperature and humidity fluctuations in the edge node deployment environment on sensor accuracy and computing power stability, ensuring the stability of detection accuracy without the need for additional constant temperature and humidity cabinets.

[0173] CRD refers to the spatial correlation factor of adjacent carriers in the same batch. It can be obtained by statistically analyzing the defect location data of the most recent 5 carriers cached by the edge node. If the overlap rate between the current candidate defect location and the defects in the same location of the previous 3 carriers is ≥60% (i.e., the same location means the defect coordinate difference is ≤20μm, and the overlap rate means the proportion of the same type of defect in the coordinate range of the previous 3 carriers), then the value is 1.08; otherwise, the value is 1.0.

[0174] Production defects in IC substrates (such as uneven plating and hole misalignment) are mostly batch-continuous. The probability of the same defect appearing at the same position on adjacent substrates on the same roller conveyor is many times that of random defects. Most existing systems only make independent judgments on a single substrate and do not associate them with adjacent samples.

[0175] The introduction of the spatial correlation factor (CRD) for adjacent carriers in the same batch utilizes the correlation of detection data from adjacent carriers on the edge side to improve the detection capability of batch-related continuous defects and promptly identify continuous process problems in the production line.

[0176] CRE refers to the repairability weighting coefficient of defects, which can be obtained by matching the edge nodes with the preset defect repairability library. The value is 0.92 for repairable defects and 1.08 for unrepairable defects. The defect repairability library is preset by process engineers and can be updated synchronously with the production line process iteration without adjusting the edge node parameters for each machine.

[0177] Defects on IC substrates are divided into two categories: those that can be repaired (such as minor uneven plating and surface foreign matter) and those that cannot be repaired (such as open circuits in inner layers and excessive hole offset). Existing detection algorithms only determine whether defects exist and do not adjust the output level in combination with repairability in the final decision-making stage, which can easily lead to unnecessary scrapping of good products.

[0178] The introduction of the repairability weighted coefficient (CRE) in the defect determination stage combines repairability to adjust the determination threshold, reducing the false scrapping of repairable defects, while preventing unrepairable defects from being missed and flowing into the next process, thus optimizing the yield management logic of the production line.

[0179] Based on the above, this defect determination unit is the final decision-making unit on the edge side. All core calculations are completed locally on the edge node. This unit combines multimodal cross-validation logic, edge-local adjacent sample association logic, defect business attribute logic, and edge operating environment status data, so that the final determination is fully adapted to the edge side operating environment and the actual business needs of the production line. At the same time, only the structured defect data that is determined to be of high value is sent back to the cloud on demand, without the need to send back the original detection data and intermediate calculation results, which fundamentally reduces the uplink bandwidth requirements of the system and meets the low latency requirements of real-time decision-making in the production line.

[0180] The final judgment value CR of the kth defect k It is the system's final decision output, which directly corresponds to the automated execution actions of the production line. No additional manual judgment or cloud-based secondary calculation is required. Edge nodes can directly trigger corresponding production line operations based on the result. At the same time, the output structured root cause labels can be directly used for subsequent process traceability analysis without the need for secondary data structuring processing, thus realizing the direct connection between edge detection and production line execution and process optimization.

[0181] The final judgment value CR of the kth defect k The larger the value, the higher the authenticity of the defect and the greater the potential harm, and the more important it is to prioritize strict handling logic; the smaller the value, the higher the probability that the candidate defect is a misjudgment, and it can be directly filtered without handling or consuming uplink bandwidth to send back relevant data.

[0182] It is worth noting that this embodiment presents an iterative loop, specifically as follows:

[0183] It should be noted that the iteration is triggered during idle periods of the production line (such as material change intervals and equipment standby times), and does not occupy the computing resources of normal testing. The specific process is as follows:

[0184] S1. Sample screening: Extract valid samples from the historical detection data stored locally on the edge nodes that are judged as high-confidence candidates by the defect confidence unit and finally confirmed as real defects by the defect judgment unit, and eliminate the interference of misjudged samples.

[0185] S2. Modal Hit Rate Statistics: For each modality i, the percentage of successful responses of that modality in all valid samples is calculated, i.e., the normalized feature score AS of the i-th modality output by that modality when a real defect occurs. i The proportion of times the anomaly threshold is reached out of the total number of valid samples is denoted as the actual recognition hit rate (PP) for that modality. i ;

[0186] S3, Weight Update: The updated ASA is calculated using an iterative formula. i Replace the original ASA in the edge feature unit i parameter;

[0187] Furthermore, the iterative formula is as follows:

[0188] ;

[0189] In the above formula, ASA i，t+1 The confidence weight refers to the i-th mode after the (t+1)-th iteration;

[0190] In the above formula, ASA i，t The confidence weight refers to the i-th mode after the t-th iteration;

[0191] In the above formula, PAQ refers to the iterative smoothing coefficient, which is set to 0.9 in this embodiment to balance the proportion of historical weights and new statistical results and avoid sudden changes in weights;

[0192] In the above formula, PP i The actual identification hit rate of modality i obtained from the statistical analysis is calculated from the modal response of real defect samples;

[0193] S4. Convergence Verification: Calculate the weight difference before and after the update, and determine whether the convergence condition is met. If not, repeat the above process. If it is met, terminate the current iteration.

[0194] Furthermore, the convergence condition is as follows, and the iteration terminates when any condition is met;

[0195] Condition 1, Maximum Iteration Threshold: Each calibration round shall be performed at most 20 iterations. Once this number is reached, the iteration shall be terminated regardless of whether the weight is stable, in order to avoid the iteration process occupying the computing resources of the edge node for a long time and affecting the normal detection business.

[0196] Condition 2, Weight Fluctuation Threshold: When the ASA of all modes after this iteration i Compared to the previous iteration of ASA i When the absolute value of the difference is less than the preset minimum fluctuation threshold, such as 0.002, it is determined that the weight has been adapted to the actual detection characteristics of the current scene, and the iteration is terminated.

[0197] In this iterative system, the factory-preset confidence weight ASA for the i-th mode is... i Environmental interference and equipment installation deviations may cause discrepancies between the actual recognition accuracy of each modality and the preset value. Iterative calibration can adapt to the actual environment and make up for the limitations of factory calibration. Furthermore, the sensor may experience performance drift and aging during long-term use, and the recognition reliability of each modality will change over time. Dynamic iterative weights can continuously adapt to the performance changes of the sensor without the need for manual on-site recalibration of parameters periodically.

[0198] Different batches of IC substrates have differences in substrate, manufacturing process and circuit design. The significance of different defects in each mode will change. Iterative weights can be adapted to the testing needs of different batches of products, so that the testing parameters do not need to be manually adjusted every time the batch is changed.

[0199] The confidence weights (ASA) of the i-th mode after iteration i Dynamic weights that adapt to the current deployment environment, device status, and product characteristics can more accurately reflect the actual detection reliability of each modality, making the standardized feature scores output by the edge feature units more closely match the real anomaly representation of the current scenario, improving the accuracy of feature normalization from the source, and providing more reliable input for the calculation of subsequent formulas;

[0200] Furthermore, the iteration process is completed locally on the edge node, without the need to upload the original test data to the cloud, thus not occupying the uplink bandwidth of the industrial network or leaking production data. This aligns with the local autonomy architecture of edge computing. The iteration process does not interrupt the normal testing process and is only executed when the computing power of the edge node is idle. It does not affect the testing cycle time and real-time requirements of the production line. The system can maintain the stability of testing accuracy over a long period of time and will not experience accuracy decay due to equipment aging, environmental changes, or product replacement. This significantly reduces the long-term maintenance cost of the system and reduces the frequency of manual calibration.

[0201] Although embodiments of the invention have been shown and described, it will be understood by those skilled in the art that various changes, modifications, substitutions and alterations can be made to these embodiments without departing from the principles and spirit of the invention, the scope of which is defined by the appended claims and their equivalents.

Claims

1. An IC carrier board inspection data processing system based on edge computing, characterized in that, This includes a detection data analysis module and an edge-side local decision-making module; The detection data analysis module: It is used to receive preprocessed edge feature data, defect confidence correlation data, and defect determination correlation data; The detection data analysis module includes: Edge feature unit, defect confidence unit, and defect determination unit; The processing flow of the detection data analysis module is as follows: Based on the confidence weight of the i-th mode, the original physical feature value of the i-th mode, the current computing power load coefficient of the edge node, the real-time contamination coefficient of the i-th sensor, the instantaneous transmission angle correction coefficient of the carrier board, and the transmission micro-vibration compensation coefficient of the i-th sensor in the edge feature data, the normalized feature score of the i-th mode is output. Based on the historical defect correlation factor of the k-th type of defect, the contribution weight of the i-th type of mode to the k-th type of defect, the current production line process fluctuation coefficient, the material batch adaptation coefficient of the k-th type of defect, the aging attenuation coefficient of the detection head spot, and the distance correlation factor between the k-th type of candidate defect and the nearest neighbor via, and combined with the normalized feature score of the i-th type of mode, the confidence score of the k-th candidate defect is output. Based on the multimodal collaborative verification factor, global process threshold correction coefficient, edge node local temperature and humidity correction coefficient, adjacent carrier board spatial correlation factor in the same batch, and defect repairability weighting coefficient in the defect judgment association data, and combined with the normalized feature score of the i-th mode and the confidence of the k-th candidate defect, the final judgment value of the k-th defect is output. Edge-side local decision-making module: It is used to receive the normalized feature score of the i-th modality, the confidence score of the k-th candidate defect, and the final judgment value of the k-th defect, in order to determine the defect status and make a decision on how to respond.

2. The IC carrier board detection data processing system based on edge computing according to claim 1, characterized in that: The output of the detection data analysis module is: The normalized feature score of the i-th modality is used to normalize the multimodal detection data into a unified and standardized feature score; The confidence score of the kth candidate defect is used to perform preliminary screening and classification of candidate defects on the IC substrate, and the feature scores of the multimodal model are converted into confidence values ​​for the corresponding defect types. The final determination value of the kth defect is used in conjunction with the local decision module on the edge side to achieve the final defect determination.

3. The IC carrier board detection data processing system based on edge computing according to claim 2, characterized in that: The processing procedure for the edge feature unit is as follows: A1. By dividing the IC substrate test data into three multimodal data, namely optical image mode, laser three-dimensional morphology mode and electrical performance test mode, and assigning confidence weights based on different mode types, the confidence weight of the i-th mode is obtained, so as to achieve the purpose of dynamic adjustment. A2. By processing the detection data of multimodal types separately, the corresponding feature value data is output. The minimum and maximum values ​​of the original feature values ​​of the defect-free standard samples of the current batch of carrier plates under the i-th modality are combined to achieve the purpose of normalization. A3. The CPU and GPU utilization rates are collected in real time through the system status monitoring module to calculate the load and output the current computing power load coefficient of the edge node. A4. By comparing the characteristic values ​​of the standard calibration sheet with the factory standard characteristic values, the signal-to-noise ratio is compared to calculate the real-time contamination coefficient of the i-th type of sensor. A5. By analyzing the coordinates of the Mark points on the IC carrier board, we can analyze the situation where the eigenvalues ​​are artificially high due to perspective distortion caused by the slight deflection of the carrier board during transmission, so as to calculate the instantaneous transmission deflection correction coefficient of the carrier board. A6. By calculating the instantaneous vibration amplitude collected, the problems of image blurring caused by production line mechanical vibration and feature value false highness caused by point cloud offset are analyzed, so as to calculate the transmission micro-vibration compensation coefficient of the i-th type of sensor, and finally output the normalized feature score of the i-th type of mode.

4. The IC carrier board detection data processing system based on edge computing according to claim 3, characterized in that: The processing procedure for the defect confidence unit is as follows: B1. By introducing the normalized feature score of the i-th mode into the defect confidence unit, the purpose of cross-modal information fusion is achieved, and the common defect categories of IC carrier boards are specifically classified to realize the numbering of candidate defect types. B2. By analyzing the defect data of the same workstation stored locally on the edge node in the last 24 hours, the number of times the same type of defect occurs in the same location is obtained, and then the historical defect correlation factor of the k-th type of defect is calculated. B3. Based on different multimodal data, implement targeted weight assignment to obtain the contribution weight of the i-th modality to the k-th defect; B4. By analyzing the deviation of the current process parameters of the production line, the process fluctuation coefficient of the current production line can be obtained. B5. Obtain batch parameter data through the traceability QR code on the surface of the carrier board to match different IC carrier board materials in a targeted manner, so as to obtain the material batch adaptation coefficient of the k-th type of defect. B6. The aging attenuation coefficient of the light spot of the detection head is analyzed by calculating the light power data collected by the light power sensor built into the detection head. B7. By matching the pre-stored carrier board design vector diagram of the edge nodes, analyze the distance between the current candidate defect and its nearest neighbor via, and calculate the confidence level of the k-th candidate defect.

5. The IC carrier board detection data processing system based on edge computing according to claim 4, characterized in that: The processing procedure of the defect determination unit is as follows: C1. By introducing the normalized feature score of the i-th modality and the confidence score of the k-th candidate defect into the defect determination unit, the original modality features are used to cross the defect determination results to avoid single-modality misjudgment. C2. Obtain the multimodal collaborative verification factor by statistically analyzing the normalized feature scores of the i-th mode of the three modes; C3. By recalculating the temperature and humidity of the edge nodes, we can analyze the impact of temperature and humidity fluctuations in the edge node deployment environment on sensor accuracy and computing power stability, in order to obtain the local temperature and humidity correction coefficient of the edge nodes. C4. By analyzing the defects of IC substrates in the same batch, the spatial correlation factor of adjacent substrates in the same batch can be obtained. C5. By analyzing the repairability of the IC substrate, a repairability weighting coefficient for the defect is obtained. This coefficient is then combined with a global process threshold correction coefficient to finally output the final judgment value for the k-th defect.

6. The IC carrier board detection data processing system based on edge computing according to claim 1, characterized in that: The edge-side local decision-making module includes a defect level analysis unit and an execution management decision-making unit.

7. The IC carrier board detection data processing system based on edge computing according to claim 6, characterized in that: The defect level analysis unit is specifically: When the final judgment value of the k-th defect is ≥0.9, it is classified as a severe defect. When 0.7 ≤ the final judgment value of the kth defect < 0.9, it is classified as a minor defect. When the final judgment value of the kth defect is <0.7, it is in the no-abnormal release level.

8. The IC carrier board detection data processing system based on edge computing according to claim 7, characterized in that: The execution management decision-making unit is specifically: When the defect is classified as severe, the edge node sends a level trigger signal to the sorting cylinder at the end of the production line via the industrial bus. The cylinder then pushes the carrier plate into the non-conforming product recycling station. When the defect is classified as minor, the edge node sends a command to the laser marking machine next to the inspection station to mark the type and location of the defect on the designated marking area on the edge of the carrier plate, and writes it into the traceability QR code on the surface of the carrier plate. When the board is in the no-abnormal release level, the edge node retains the no-defect release record of the carrier board and stores it locally at the edge for 30 days for subsequent traceability.

Images