PCB intelligent detection and defect production parameter tracing method

Abstract

The invention relates to the technical field of PCB (Printed Circuit Board) intelligent detection and defect production, and particularly discloses a PCB intelligent detection and defect production parameter tracing method, which comprises the following steps of: firstly, acquiring defect images, physical positions and circuit design layout data by utilizing automatic optical detection equipment, and generating defect context fingerprints; then, a production process traceability data flow is backtracked based on fingerprints, process abnormal signatures are extracted, causal association knowledge elements are constructed, then a dynamic defect knowledge graph is converged, and the system automatically matches a high-risk process mode and generates an early warning by analyzing a dynamic behavior mode of historical data and real-time parameters; predicting defect positions and types in combination with circuit design characteristics; according to the method, the limitation of traditional static threshold monitoring is broken through, the whole-process intelligence from defect positioning and process tracing to risk early warning is realized, the problem diagnosis efficiency and the quality control precision are remarkably improved, and a structured and inheritable quality knowledge system is provided for PCB manufacturing.

Description

Technical Field

[0001] This invention relates to the field of intelligent PCB board inspection and defect production technology, and more specifically, to a method for tracing intelligent PCB board inspection and defect production parameters. Background Technology

[0002] In advanced manufacturing, especially in the production of high-precision electronic products such as PCBs, ensuring product quality and improving yield are long-term challenges and core issues for enterprises. The formation mechanism of defects is extremely complex. It is not caused by a single static parameter exceeding its limit, but rather stems from the dynamic evolution of multiple production parameters within a specific time window, and is closely coupled with the product's own design layout characteristics. This implicit, non-linear causal relationship between the dynamic behavior of the process and the local environment of the product constitutes a technical black box in quality control, making it difficult to diagnose the causes of defects and leading to recurring quality problems.

[0003] Currently, the solutions commonly used in the industry mainly rely on statistical process control (SPC) and human expert experience. SPC monitors key production parameters by setting fixed upper and lower thresholds, triggering alarms when parameter values ​​exceed these limits. For defects that have already occurred, a quality analysis team composed of experienced engineers and technicians typically conducts post-incident investigations by reviewing production reports, comparing historical data, and combining personal experience to try to find possible causes. This approach constitutes the traditional quality management and root cause analysis model.

[0004] The drawbacks of traditional methods are becoming increasingly apparent. Statistical process control methods based on static thresholds are powerless against process fluctuations that, while within acceptable limits, exhibit abnormal dynamic behavior, and cannot provide early warnings of potential defects caused by process instability. Expert diagnostic models relying on manual intervention are not only inefficient and slow to respond, but their conclusions are often subjective, making it difficult to establish a unified and transferable knowledge system among different experts, and they are unable to address the complex cross-domain relationships behind massive amounts of production data. These methods generally treat process data and defect data separately, lacking a systematic approach to establish precise causal relationships between process dynamic behavior and specific defect patterns within a specific circuit context. This results in a large amount of preventative maintenance and reactive quality responses, which are costly and have limited effectiveness. Summary of the Invention

[0005] In view of this, in order to solve the problems mentioned in the background technology, a method for intelligent inspection and defect production parameter traceability of PCB boards is proposed.

[0006] The objective of this invention can be achieved through the following technical solution: This invention provides a method for intelligent inspection and defect production parameter traceability of PCB boards, including the following steps: S1, defect location and data acquisition: acquire defect location and layout combined data including defect images, physical location information and circuit design layout files.

[0007] S2. Defect Context Fingerprint Generation: Based on the joint data of defect location and layout, a defect context fingerprint is generated by combining the inherent morphological features of the defect with the environmental context features.

[0008] S3. Process traceability data stream capture: Based on the identity and time information carried in the defect context fingerprint, trace back and capture the process traceability data stream synchronized with the defect.

[0009] S4. Process Anomaly Signature Extraction: Perform dynamic behavior analysis on the process traceability data stream to extract process anomaly signatures that represent the dynamic instability of the process.

[0010] S5. Establishment of causal knowledge elements: Pair and bind process exception signatures with defect context fingerprints to establish causal knowledge elements.

[0011] S6. Dynamic Defect Knowledge Graph Aggregation: Receives causal knowledge elements, aggregates and iteratively updates the dynamic defect knowledge graph by strengthening or creating nodes and relationships.

[0012] S7. High-risk process pattern set parsing: Traverse the dynamic defect knowledge graph, filter out process anomaly signatures that are related to high-risk defect types and whose causal relationship weight values ​​exceed the preset threshold, and parse them into a high-risk process pattern set.

[0013] S8. Matching of pending warning events: Detect dynamic behaviors in the real-time production data stream that match the set of high-risk process patterns, and generate pending warning events containing information about the current production batch.

[0014] S9. Contextualized Intervention Command Generation: Receives pending warning events, queries the dynamic defect knowledge graph, and combines it with the circuit design layout characteristics of the PCB board to be processed to generate contextualized intervention commands that include risk prediction and location.

[0015] Compared with the prior art, the embodiments of the present invention have at least the following advantages or beneficial effects: (1) The present invention performs dynamic behavior analysis on multiple parameters in the production process, rather than simply monitoring static thresholds, so that the system can identify early dynamic signs that lead to defects. This means that before the defects actually form, the system can capture the unstable state in the process and issue an early warning, thus winning a valuable time window for intervention measures. This shifts the focus of quality control from post-production detection and screening to real-time prevention in the process, effectively avoiding the occurrence of batch quality accidents.

[0016] (2) This invention constructs an objective and structured causal relationship network by forcibly binding each confirmed defect to the specific process anomaly signature traced back to it and aggregating them into a dynamic knowledge graph. This makes root cause analysis no longer dependent on personal experience and subjective speculation, but can be quickly and accurately located based on a data-driven knowledge graph, greatly improving the efficiency and depth of problem diagnosis, and realizing the systematic accumulation and inheritance of quality knowledge.

[0017] (3) This invention can generate contextualized and highly accurate intervention instructions, significantly improving the efficiency and effectiveness of problem handling. When the system issues an early warning of a high-risk process mode, it not only knows what happened, but also accurately predicts what kind of defect the risk will cause and which area of ​​the circuit board is most likely to appear by querying the knowledge graph and combining it with the specific layout characteristics of the product to be processed. This context-aware prediction makes the intervention instructions generated by the system no longer a general alarm, but an action guide that includes specific location and clear handling suggestions, guiding operators or automated systems to perform surgical-like precise intervention, avoiding blind adjustments and shortening downtime. Attached Figure Description

[0018] To more clearly illustrate the technical solutions of the embodiments of the present invention, the accompanying drawings used in the description of the embodiments will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0019] Figure 1 This is a schematic diagram of the method steps of the present invention.

[0020] Figure 2 This is a schematic diagram illustrating the defect context fingerprint generation process and feature composition of the present invention.

[0021] Figure 3 This is a schematic diagram illustrating the principle of generating contextualized intervention instructions according to the present invention. Detailed Implementation

[0022] 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.

[0023] Please see Figure 1This invention provides a method for intelligent inspection and defect production parameter traceability of PCB boards, including: S1, defect location and data acquisition: acquiring defect location and layout combined data including defect images, physical location information and circuit design layout files.

[0024] It should be noted that the implementation process for obtaining defect location and layout joint data is as follows: First, a pre-processed PCB board is placed on the inspection platform of an automated optical inspection device. The device is started, and its built-in high-resolution line scan camera, with the aid of a light source, scans the entire surface of the PCB board line by line. The image processing unit inside the device compares the real-time captured images with preset standard images. Once an inconsistency is found, such as an open circuit, short circuit, or copper foil residue, the system determines it as a defect. At this point, the system captures and saves a high-resolution digital image centered on the defect. Simultaneously, the device, based on its internal motion control system and encoder readings, accurately records the position of the defect center point in the device's worktable coordinate system. Through coordinate transformation, this is converted into physical position information in the PCB coordinate system with a specified reference point as the origin. The coordinate conversion is based on existing technology and will not be elaborated here. After scanning and positioning are completed, the device or a cooperating barcode reader scans the pre-made unique identification code on the PCB board, usually a QR code or barcode, and decodes it to obtain a string of characters. The system uses this string of characters as a query keyword to access the server storing circuit design files and retrieve the circuit design layout file that perfectly matches the unique identifier. Finally, the system executes a data packaging program to create a new data unit, integrating the digital image, physical location information, and retrieved circuit design layout file generated in the previous steps to form a structured defect location and layout combined data.

[0025] Please see Figure 2 S2, Defect Context Fingerprint Generation: Based on the joint data of defect location and layout, a defect context fingerprint is generated by combining the inherent morphological features of the defect with the environmental context features.

[0026] In a specific embodiment of the present invention, the specific process of generating a defect context fingerprint includes: performing morphological analysis on the defect region in the digital image of the defect location and layout joint data, and extracting its inherent morphological features such as size, shape, and gray value.

[0027] It should be noted that the system first parses the data, sending the digital image, physical location information, and circuit design layout file to different processing modules. The first module is responsible for extracting inherent morphological features. It receives the digital image and performs morphological analysis on the identified defect areas in the image. This is a series of image processing operations aimed at quantifying the geometric and visual attributes of the defects. Specifically, the program first uses grayscale thresholding or edge detection algorithms to accurately separate the defect pixels from the background, forming a binary defect mask. Then, it calculates the total number of pixels in the mask area and multiplies it by the actual area represented by a single pixel to obtain the size of the defect. Next, by calculating the relationship between the perimeter and area of ​​the defect area outline, it calculates the compactness to quantify the complexity of its shape, whether it tends towards a circle or an irregular elongated shape. At the same time, the program calculates the average grayscale value of all original pixels belonging to the defect area as the grayscale feature of the defect. This set of quantified values—size, shape, and grayscale value—together constitutes the inherent morphological features.

[0028] It should also be noted that the specific formula for calculating tightness is as follows: ,in, Indicates firmness, These represent the area and perimeter of the defect region, respectively. Pi is approximately equal to 3.14. If the compactness is close to 1, it indicates that the defect shape is relatively regular and close to a circle. It may be a circular defect such as a hole or an unevenly sized solder joint. If the compactness is much less than 1, it indicates that the defect shape is irregular and may be a slender defect such as a scratch or a crack.

[0029] The physical location information of defects is mapped onto the circuit design layout file, and the surrounding circuit environment is analyzed to extract environmental context features such as the type of nearby components, local wiring density, and copper foil coverage.

[0030] In a specific embodiment of the present invention, the step of extracting environmental context features such as the type of adjacent components, local wiring density, and copper foil coverage includes: defining an analysis area on the circuit design layout file with the physical location information of the defect as the center.

[0031] Search for circuit design elements within the analysis area and identify the types of adjacent components.

[0032] The local cabling density is quantified by calculating the ratio of the total length of cabling within the analysis area to the area of ​​the analysis area.

[0033] The copper foil coverage rate is quantified by calculating the ratio of the total copper foil coverage area to the area of ​​the analysis region.

[0034] It should be noted that the second processing module maps the physical location information of the defect onto the circuit design layout file. This module loads the circuit design layout file and, based on the defect's board coordinates, precisely locates the defect's center point in the layout data. Then, using this point as the center, a virtual analysis area is defined, for example, a circular area with a radius of 3 mm. Within this area, the program retrieves all circuit design elements, identifies the types of adjacent components, measures the distance between the center points of adjacent pins on the components to obtain the pin pitch, calculates the local wiring density by dividing the total length of all traces within the analysis area by the area, and calculates the copper foil coverage by dividing the total area of ​​all copper foil-covered areas by the total area of ​​the analysis area. This data, extracted from the layout design and describing the environment surrounding the defect, constitutes the environmental context features.

[0035] It should also be noted that the total length of the wiring within the area is obtained by summing the lengths of all line segment geometric objects within the specified area through analyzing the circuit design layout file; the total copper foil coverage area within the area is obtained by summing the geometric areas of all copper foil fills, traces, and pads within the specified area through analyzing the circuit design layout file.

[0036] By combining inherent morphological features with environmental context features for encoding, a defect context fingerprint is formed.

[0037] It should be noted that, finally, the third module combines and encodes the inherent morphological features output from the first two modules with the environmental context features. It arranges all extracted numerical and categorical features in a predefined order to form a long feature vector. To generate a unique, fixed-length identifier, this vector can be input into a hash function to generate a unique string. This final string or vector is the defect context fingerprint representing the defect and its local physical environment.

[0038] S3. Process traceability data stream capture: Based on the identity and time information carried in the defect context fingerprint, trace back and capture the process traceability data stream synchronized with the defect.

[0039] In a specific embodiment of the present invention, the specific process of tracing back and capturing the process source data stream synchronized with the defect includes: parsing the unique identification code of the PCB board and the timestamp information of the defect formation carried in the defect context fingerprint.

[0040] Based on the unique identification code and timestamp, all production stations that the circuit board passed through in the manufacturing execution system were traced back, and the key processes related to the formation of the defect were identified.

[0041] From the equipment controllers and sensors of key processes, multiple raw production parameters that change continuously within the corresponding time window are collected to form a process traceability data stream that is strictly synchronized with the time of the defect.

[0042] It should be noted that the implementation process of capturing traceability data flow is as follows: The system first receives the defect context fingerprint generated in the previous step. Based on this fingerprint, the system performs a parsing operation, extracting the unique identification code of the PCB board and the timestamp information when the defect was recorded from its associated data. This timestamp information is the timestamp information of defect formation, as it represents the latest time point when the defect was confirmed to exist. After obtaining these two key pieces of information, the system communicates with the Manufacturing Execution System (MES). It uses the unique identification code as a query command, requesting the MES to return the complete processing record of the PCB board since production began. The MES provides a list containing all production stations and their corresponding entry and exit times. The system program then compares the timestamp information of defect formation with this list, and through reverse chronological search, locates the last processing step completed before the timestamp. This step is then identified as the key process most relevant to defect formation. After determining the key process, the system determines a precise time window based on the equipment number of that process and the entry and exit times of the PCB board at that process. Finally, the system accesses the controller or data server of the equipment corresponding to the critical process and requests the export of multiple raw production parameters continuously recorded by all relevant sensors on the equipment within a defined time window. These parameter data streams are integrated and packaged to form a process traceability data stream that is strictly aligned with the time of the defect.

[0043] It's also important to note that a Manufacturing Execution System (MES) is a factory-level information management software. Its function is to monitor, track, and manage the production process of all products in the workshop in real time. It stores detailed information such as the production stations each product passes through, the materials used, the operators, and the time. A production station refers to a specific work unit on the production line, such as a solder paste printer, a pick-and-place machine, or a reflow oven. Each station completes a specific manufacturing or assembly task. A critical process is the specific processing step among all production stations that, through time backtracking analysis, is determined to be the most likely cause of the current defect. A time window is a specific time interval defined by a start time and an end time. It is set based on the precise moments when a specific PCB board enters and leaves the critical process, and is recorded by the MES. Multiple raw production parameters refer to multiple unprocessed data streams directly collected from sensors on the equipment of the critical process, such as real-time temperature readings of multiple temperature zones in a reflow oven, conveyor belt speed, nitrogen flow rate, etc. These data collectively constitute a comprehensive description of the operational status of that process.

[0044] S4. Process Anomaly Signature Extraction: Perform dynamic behavior analysis on the process traceability data stream to extract process anomaly signatures that represent the dynamic instability of the process.

[0045] In a specific embodiment of the present invention, the specific process of extracting the process anomaly signature representing the dynamic instability of the process includes: performing dynamic behavior analysis on the multiple time series parameters contained in the process tracing data stream.

[0046] Identify and quantify dynamic behavior patterns exhibited by the collaboration between parameters, wherein the dynamic behavior patterns are abnormal fluctuations of a single parameter or unexpected linkages between multiple parameters.

[0047] The quantified dynamic behavior pattern is encoded into a feature vector to form a process anomaly signature.

[0048] It should be noted that the system receives the process traceability data stream output from the previous key process. This data stream contains multiple time-series parameters collected within a specific time window. The system then initiates dynamic behavior analysis on this data stream. This analysis does not simply check whether each parameter value exceeds the preset static upper and lower limits, but rather delves into the dynamic characteristics of the parameters changing over time and the interactions between parameters. For the analysis of a single parameter, such as the temperature curve of a certain temperature zone, the slope of the temperature curve for that zone is taken as its rate of temperature change. If the rate of temperature change is greater than the set reference rate of temperature change, it indicates that there is an abnormal fluctuation of a single parameter in that temperature zone. For the analysis of multiple parameters, the system selects parameter pairs that may be correlated based on process knowledge, such as placement speed and pressure value, and calculates their cross-correlation function within the time window. By finding the peak value in the cross-correlation function, it can determine whether there are unexpected linkages and whether there is a delayed or leading relationship in the linkage. Once one or more dynamic behavior patterns are identified, the system quantifies them. For example, for periodic oscillations, the quantification result is the frequency and amplitude of the oscillation; for the hysteresis of the temperature rise curve, the quantification result is the time delay in seconds relative to the standard curve; for unexpected linkages, the quantification result is the maximum correlation coefficient and the corresponding time lag. Finally, the system arranges and combines the values ​​of all quantified dynamic behavior patterns in a predefined order, encoding them into a fixed-dimensional numerical feature vector. This feature vector is the digital description of the dynamic instability in this production process and is defined as the process anomaly signature.

[0049] It should also be noted that a formula is needed to explicitly explain how to quantify the unexpected interactions between multiple parameters. A common method is to calculate the normalized cross-correlation coefficient, as shown in the following formula: ,in, Represents two time series parameters and With a time delay of The normalized cross-correlation coefficient at time t, the closer its absolute value is to 1, the stronger the linear correlation between the two parameters. The step size representing the time delay is an integer, calculated by different... Value This allows us to identify the time delay relationships where the correlation is strongest. and These represent two different time series parameters at the 1st... The and the first The values ​​at each time sampling point; and These are the average values ​​of the two time series parameters over the entire time window; and These are the standard deviations of the two time series parameters over the entire time window, used for normalization to eliminate the influence of different parameter units and numerical ranges; This represents the total number of sampling points within the time window. This value is set based on the length of the time window and the data acquisition frequency; for example, a 5-minute time window might acquire data once per second. That is, 300.

[0050] The process anomaly signature is a feature vector uniquely representing the dynamic instability of a specific process. Its data structure is a fixed-length array of floating-point numbers, where each element quantifies a specific dynamic behavior. Its function is to serve as input to a machine learning model or a node in a knowledge graph, providing a standardized description of the anomalous process. Dynamic behavior analysis, distinct from static threshold judgment, processes time-series data using mathematical and statistical tools. Its function is to reveal patterns, trends, and interrelationships in data changes over time. Dynamic behavior patterns are specific, quantifiable phenomena identified in dynamic behavior analysis, such as anomalous fluctuations in a single parameter or unexpected linkages between multiple parameters. Anomalous fluctuations in a single parameter refer to abnormal, rapid, or periodic changes in a single sensor signal, rather than the value itself exceeding limits. Unexpected linkages between multiple parameters refer to two or more production parameters that should be independent or correlated according to a specific pattern exhibiting abnormal synchronous or asynchronous correlations within a certain time period.

[0051] S5. Establishment of causal knowledge elements: Pair and bind process exception signatures with defect context fingerprints to establish causal knowledge elements.

[0052] In a specific embodiment of the present invention, the specific process of establishing causal association knowledge elements includes: referencing the defect context fingerprint corresponding to the process exception signature in time and space.

[0053] By forcibly pairing and binding process exception signatures with defect context fingerprints, a causal relationship is established from the dynamic behavior of a specific process to the specific defect form under a specific environment.

[0054] The pairing and binding relationship is encapsulated into an independent and indivisible data unit, forming a causal knowledge element.

[0055] It should be noted that the implementation process for establishing causal association knowledge elements is as follows: The system uses the underlying association information carried by the process anomaly signature—that is, the unique identifier of the PCB board—to reference the defect context fingerprint generated in step S2 for the same defect on the same PCB board, which corresponds to it in time and space. The mechanism ensuring time and space correspondence is that the entire data flow is transmitted with the unique PCB board and its specific defects as the core throughout. Therefore, the process anomaly signature and defect context fingerprint obtained in this step naturally belong to the same causal event chain. After loading these two data units, the process anomaly signature and the defect context fingerprint, into memory simultaneously, the system performs a forced pairing and binding operation. This operation creates a new data structure at the program logic level, inseparably linking the process anomaly signature and the defect context fingerprint as two fundamental components. This binding operation is not based on probability or statistical inference, but rather a deterministic assertion based on factual tracing, that is, explicitly establishing a causal relationship from the dynamic behavior of a specific process to the specific defect form under a specific environment. Finally, the system encapsulates this pairing and binding relationship, storing it as an independent, indivisible data unit. This newly generated data unit solidifies the connection between cause and effect at the data level, forming a causal knowledge element, which is then output to provide basic materials for building a knowledge graph.

[0056] S6. Dynamic Defect Knowledge Graph Aggregation: Receives causal knowledge elements, aggregates and iteratively updates the dynamic defect knowledge graph by strengthening or creating nodes and relationships.

[0057] In a specific embodiment of the present invention, the specific process of gathering and iteratively updating the dynamic defect knowledge graph by strengthening or creating nodes and relationships includes: if there are already node pairs and their relationship edges in the knowledge network structure that are similar to the process anomaly signature and defect context fingerprint in the causal association knowledge element, then strengthen the causal relationship weight value corresponding to the relationship edge.

[0058] If there are no node pairs and their relational edges in the knowledge network structure that are similar to the process anomaly signature and defect context fingerprint in the causal knowledge element, then new nodes are created for the process anomaly signature and defect context fingerprint, and new relational edges and initial causal relation weight values ​​are established between the new nodes.

[0059] It should be noted that the system receives causal relationship knowledge elements generated by the previous key process. Upon receiving these elements, the system uses them as input and interacts with the existing knowledge network structure. This knowledge network structure is a graph-like database composed of nodes and edges. The system first parses the causal relationship knowledge elements, extracting the process anomaly signature and defect context fingerprint. Then, the system queries the knowledge network structure to determine if there are nodes similar to the newly received signature and fingerprint. If the query results show that there is already a node in the knowledge network structure that is highly similar to the newly received process anomaly signature and also highly similar to the newly received defect context fingerprint, and there is already an edge representing a causal relationship between these two nodes, then the system will strengthen the causal relationship weight value corresponding to that node. If the query results show that there are no similar nodes, or only partially similar nodes exist but they are not related, the system will create new nodes and association paths. Specifically, the system will create a new cause node for the new process anomaly signature, a new result node for the new defect context fingerprint, and then establish a directed edge representing a causal relationship between these two new nodes, assigning an initial causal relationship weight value to this edge. By continuously receiving new causal knowledge elements and repeatedly performing the above-mentioned query, reinforcement, or creation operations, the system gradually builds and iteratively updates a dynamic defect knowledge graph that describes the complex relationships between defect morphology, circuit context, and process behavior.

[0060] It should also be noted that the specific process by which the system queries the knowledge network structure to determine whether there are nodes similar to the two newly received signatures and fingerprints includes: 1) Extracting node vectors: Extracting the stored process anomaly signature vector and defect context fingerprint vector from the nodes; 2) Calculating similarity: Calculating the similarity between the new process anomaly signature vector and the node's process anomaly signature vector, and the similarity between the new defect context fingerprint vector and the node's defect context fingerprint vector. The similarity can be calculated using cosine similarity, which is the ratio between the sum of the products of the two vectors in the corresponding dimensions and the product of the magnitudes of the two vectors; 3) Setting similarity thresholds: The process anomaly signature similarity threshold can be set to 0.8, and the defect context fingerprint similarity threshold can also be set to 0.8. Both are set based on historical data and actual needs; 4) Comprehensive judgment: If the similarity between the new process anomaly signature vector and the node's process anomaly signature vector exceeds the process anomaly signature similarity threshold, and the similarity between the new defect context fingerprint vector and the node's defect context fingerprint vector exceeds the defect context fingerprint similarity threshold, then the two nodes are considered similar; otherwise, they are considered dissimilar.

[0061] It should also be noted that when reinforcing the weights of causal relationships, a mathematical formula can be used to precisely define the update rule, as follows: ,in, Representative at the The updated causal relationship weight value after observing the same or similar causal relationships; This represents the causal relationship weight value before the update, and its initial value is set to 0 when the relationship is first created. It is a forgetting factor or historical weight coefficient, with a value between 0 and 1. The setting is based on the desired proportion of new evidence and historical evidence in the weight update. For example, setting it to 0.9 means that historical evidence accounts for 90% of the weight and new evidence accounts for 10%, so that the weight value changes more smoothly and avoids drastic fluctuations due to a single false alarm. Representing the In this algorithm, the occurrence of this event is indicated by receiving a confirmed causal relationship knowledge element each time. The value is set to 1.

[0062] The dynamic defect knowledge graph is a continuously evolving knowledge base used to represent the causal relationships of defects in manufacturing. Its data structure is a graph, where nodes represent specific process anomaly signatures or defect context fingerprints, and edges represent the causal relationships between them. The attributes of each edge store a causal relationship weight value. Its function is to aggregate isolated causal events into a global knowledge network to reveal deeper patterns. The causal relationship weight value is a numerical value attached to the edge in the knowledge network structure. Its function is to quantify the credibility or frequency of the causal relationship that a specific cause leads to a specific result; a higher value indicates a stronger correlation. An edge is a directed link in the knowledge network structure, pointing from a node representing the cause to a node representing the result, explicitly indicating the causal connection between the two.

[0063] S7. High-risk process pattern set parsing: Traverse the dynamic defect knowledge graph, filter out process anomaly signatures that are related to high-risk defect types and whose causal relationship weight values ​​exceed the preset threshold, and parse them into a high-risk process pattern set.

[0064] In a specific embodiment of the present invention, the specific process of parsing into a high-risk process pattern set includes: filtering out all relation edges in the dynamic defect knowledge graph where all target nodes are of high-frequency or high-risk defect types.

[0065] Among the filtered relationship edges, further filter out relationship edges whose causal relationship weight values ​​exceed a preset threshold.

[0066] The source nodes of the re-selected relational edges, i.e., the process anomaly signatures, are aggregated to generate a high-risk process pattern set.

[0067] It should be noted that the system first receives the dynamic defect knowledge graph aggregated from the previous key process. After receiving it, the system initiates a traversal program to systematically scan and analyze all nodes and relational edges stored within the knowledge graph. This program examines each relational edge one by one. When examining an edge, it simultaneously reads the source node (process anomaly signature) and the target node (defect context fingerprint) connected to that edge, as well as the edge's own attributes, especially the causal relationship weight value. Next, the filtering process includes two conditions. The first condition is to determine whether the target node, i.e., the defect context fingerprint, belongs to a high-frequency or high-risk defect type. This determination is made by comparing the defect morphology and location information parsed from the defect context fingerprint with a predefined directory of high-risk defects. The second condition is to check whether the causal relationship weight value of the relational edge connecting this pair of causal nodes exceeds a preset threshold. Only when both conditions are met—that is, a process anomaly signature is not only associated with a high-risk defect, but this association has been repeatedly confirmed to the point that its weight value is sufficiently high—will the process anomaly signature be filtered out. After traversing and analyzing the entire dynamic defect knowledge graph, all selected process anomaly signatures are collected. Finally, the system summarizes these collected process anomaly signatures to generate a specific target pattern list for real-time production line monitoring; this list is defined as the high-risk process pattern set.

[0068] In one specific embodiment of the present invention, the preset threshold is a numerical value used to filter out causal relationships that are not yet clear or occur infrequently. The value is set based on the analysis of the statistical distribution of historical data and the trade-off between false positive rate and false negative rate. Specifically, it can be set to 10, which means that only when a causal relationship is independently confirmed more than 10 times will its corresponding process anomaly signature be considered high risk.

[0069] It should also be noted that high-frequency or high-risk defect types are a predefined defect classification directory, whose data structure can be a list of strings or a hash table with severity level labels. Its classification is based on long-term production experience and quality statistical analysis. For example, open circuits or short circuits that lead to direct product scrapping, or defects occurring beneath critical chips, are all classified as high-risk defect types.

[0070] S8. Matching of pending warning events: Detect dynamic behaviors in the real-time production data stream that match the set of high-risk process patterns, and generate pending warning events containing information about the current production batch.

[0071] In a specific embodiment of the present invention, the step of detecting dynamic behaviors that match the high-risk process pattern set in the real-time production data stream and generating a pending early warning event containing the current production batch information includes: using the same dynamic behavior analysis method as the refining process abnormal signature, to detect and generate the current process feature vector in real time within the sliding time window of the real-time production data stream.

[0072] It should be noted that the system first receives a set of high-risk process patterns containing clearly defined target feature vectors generated by the previous key process and loads it into a real-time monitoring module. Simultaneously, this monitoring module continuously acquires real-time production data streams from the sensors of the equipment at the specified key processes on the current production line. This data stream enters a sliding window analyzer in a continuous time series format. The analyzer employs the same dynamic behavior analysis method as in step S4, performing real-time detection and calculation on the real-time production data stream within the sliding window, and extracting a feature vector characterizing the current process dynamics in real time.

[0073] Calculate the vector distance between the current process feature vector and the abnormal signatures of each process in the high-risk process pattern set.

[0074] A pattern is considered a successful match when the vector distance is less than a preset matching threshold.

[0075] It should be noted that the system then matches this real-time generated feature vector with each pattern in the high-risk process pattern set, which is each pre-stored process anomaly signature. The matching operation is performed by calculating the distance or similarity between two feature vectors. Once the system detects that the currently generated feature vector is highly similar to any pattern in the high-risk process pattern set, that is, their vector distance is less than a preset matching threshold, it is determined to be a successfully matched pattern.

[0076] It should also be noted that the formula for calculating the vector distance between the current process feature vector and the abnormal signatures of each process in the high-risk process pattern set is as follows: ,in, Represents the feature vector generated in real time With a certain pattern in the high-risk process pattern set The Euclidean distance between them This represents the current process feature vector extracted from the real-time production data stream. This represents a process anomaly signature feature vector stored in a set of high-risk process patterns. It is a vector The One portion, It is a vector The One portion, This represents the summation of all components of a vector. The formula calculates the linear distance between two vectors in multidimensional space; the smaller the distance value, the more similar the two vectors are, meaning the current dynamic behavior of the process is closer to a known high-risk pattern.

[0077] In one specific embodiment of the present invention, the preset matching threshold can be 0.85, based on the following: through historical data verification, when the similarity is ≥0.85, the accuracy of the association between process anomaly signature and defect type exceeds 90%. This threshold is widely used in industrial scenarios such as PCB defect detection after optimization with actual production data.

[0078] The process exception signature corresponding to the pattern that was successfully matched in the packaging.

[0079] Query the Manufacturing Execution System to obtain information on the production batch currently being processed in this process.

[0080] The process exception signature and the current production batch information are integrated to generate a pending warning event.

[0081] It should be noted that upon detecting a successful match, the system immediately triggers an alert generation procedure. This procedure packages structured information, including specific details of the successfully matched high-risk pattern and production batch information currently being processed in that process, obtained by querying the Manufacturing Execution System. This packaged structured information is ultimately generated as a pending alert event and sent to the subsequent processing unit.

[0082] Please see Figure 3 S9. Contextualized intervention instruction generation: Receives warning events to be processed, queries the dynamic defect knowledge graph and combines it with the circuit design layout characteristics of the PCB board to be processed to generate contextualized intervention instructions that include risk prediction and location.

[0083] In a specific embodiment of the present invention, the specific process of generating contextualized intervention instructions that include risk prediction and location includes: querying the dynamic defect knowledge graph again based on the process anomaly signature in the early warning event to obtain all defect context fingerprint information associated with it.

[0084] Obtain and analyze the circuit design layout features of the PCB board that is about to enter this process.

[0085] By combining defect context fingerprint information with circuit design layout features, we can determine the specific defect types and high-risk areas that the abnormal signature in this process may cause on the PCB board to be processed.

[0086] Risk predictions for specific defect types are generated based on the judgment results, and high-risk areas are clearly located based on the analysis of circuit design layout characteristics.

[0087] It's important to note that, firstly, the system analyzes the event and extracts the successfully matching risk pattern, i.e., the signature feature vector of that specific process anomaly. The system then uses this feature vector as a query keyword to query the dynamic defect knowledge graph again. The purpose of this query is to retrieve all defect context fingerprints associated with this risk pattern through causal edges. Since a process anomaly can lead to defects in various environments, the query result may be a set of one or more defect context fingerprints. Simultaneously, the system interacts with the manufacturing execution system to obtain the production batch number of the PCB board about to enter the problematic process and retrieves its corresponding circuit design layout file. The system analyzes this layout file, extracting key circuit design layout features. Then, the system enters the judgment phase. It compares the defect context fingerprint information obtained from the knowledge graph with the circuit design layout features of the circuit board to be processed. For example, if the knowledge graph indicates that the risk pattern is prone to short circuits near BGA chip pins, the system will locate the positions of all BGA chips in the layout of the circuit board to be processed. Through this combined analysis, the system can determine the most likely specific defect type and high-risk area on the upcoming circuit board design, based on this risk pattern. Finally, based on this accurate judgment, the system automatically generates an instruction. This instruction, in text or code form, explicitly includes the risk prediction—what defect will occur—and the location—the location where the defect is most likely to occur.

[0088] Specific defect types are clear classifications of the physical form and electrical properties of defects, such as copper foil breakage, solder ball short circuits, or pad over-etching. High-risk areas are the specific physical coordinate ranges on the circuit board to be processed, marked after combining risk prediction and its own layout feature analysis, indicating the areas most prone to predicted defects. Risk prediction is a specific description of potential quality problems based on causal relationships in a knowledge graph. Location mapping involves mapping the abstract defect location information from risk prediction to precise coordinates or areas on the specific circuit board.

[0089] The above content is merely an example and illustration of the concept of the present invention. Those skilled in the art can make various modifications or additions to the specific embodiments described, or use similar methods to replace them, as long as they do not deviate from the concept of the invention or exceed the scope defined by the present invention, and all such modifications and additions should fall within the protection scope of the present invention.

Claims

1. A method for intelligent inspection and defect production parameter traceability of PCB boards, characterized in that, Includes the following steps: S1. Defect localization and data acquisition: Acquire joint defect localization and layout data, including defect images, physical location information, and circuit design layout files; S2. Defect Context Fingerprint Generation: Based on the joint data of defect location and layout, a defect context fingerprint is generated by combining the inherent morphological features of the defect with the environmental context features. S3. Process traceability data stream capture: Based on the identity and time information carried in the defect context fingerprint, trace back and capture the process traceability data stream synchronized with the defect; S4. Process Anomaly Signature Extraction: Perform dynamic behavior analysis on the process traceability data stream to extract process anomaly signatures that characterize the dynamic instability of the process. S5. Establishment of causal knowledge elements: Pair and bind process exception signatures with defect context fingerprints to establish causal knowledge elements; S6. Dynamic Defect Knowledge Graph Aggregation: Receives causal knowledge elements, aggregates and iteratively updates the dynamic defect knowledge graph by strengthening or creating nodes and relationships; S7. High-risk process pattern set parsing: Traverse the dynamic defect knowledge graph, filter out process anomaly signatures that are related to high-risk defect types and whose causal relationship weight values ​​exceed the preset threshold, and parse them into a high-risk process pattern set. S8. Matching of pending warning events: Detect dynamic behaviors in the real-time production data stream that match the set of high-risk process patterns, and generate pending warning events that include risk level and current production batch information. S9. Contextualized Intervention Command Generation: Receives pending warning events, queries the dynamic defect knowledge graph, and combines it with the circuit design layout characteristics of the PCB board to be processed to generate contextualized intervention commands that include risk prediction and location.

2. The method for intelligent inspection and defect production parameter traceability of PCB boards according to claim 1, characterized in that: The specific process for generating the defect context fingerprint includes: Morphological analysis of defect regions is performed in digital images of defect location and layout combined data to extract their inherent morphological features such as size, shape, and gray value; The physical location information of the defect is mapped onto the circuit design layout file, and the surrounding circuit environment is analyzed to extract environmental context features such as the type of nearby components, local wiring density, and copper foil coverage. By combining inherent morphological features with environmental context features for encoding, a defect context fingerprint is formed.

3. The method for intelligent inspection and defect production parameter traceability of PCB boards according to claim 2, characterized in that: The steps for extracting environmental context features such as neighboring component types, local wiring density, and copper foil coverage include: A region of analysis is defined on the circuit design layout file, centered on the physical location information of the defect. Search for circuit design elements within the analysis area and identify the types of adjacent components; The local cabling density is quantified by calculating the ratio of the total length of cabling within the analysis area to the area of ​​the analysis area. The copper foil coverage rate is quantified by calculating the ratio of the total copper foil coverage area to the area of ​​the analysis region.

4. The method for intelligent inspection and defect production parameter traceability of PCB boards according to claim 1, characterized in that: The specific process of tracing back and capturing the source data stream synchronized with the defect includes: Parse the unique identifier of the PCB board and the timestamp information of the defect formation carried in the defect context fingerprint; Based on the unique identification code and timestamp, all production stations that the circuit board passed through in the manufacturing execution system were traced back, and the key processes related to the formation of defects were identified. From the equipment controllers and sensors of key processes, multiple raw production parameters that change continuously within the corresponding time window are collected to form a process traceability data stream that is strictly synchronized with the time of the defect.

5. The method for intelligent inspection and defect production parameter traceability of PCB boards according to claim 1, characterized in that: The specific process for extracting the process anomaly signature representing the dynamic instability of the process includes: Dynamic behavior analysis of multiple time series parameters contained in the process traceability data stream; Identify and quantify the dynamic behavior patterns exhibited by the collaboration between parameters, wherein the dynamic behavior patterns are abnormal fluctuations of a single parameter or unexpected linkages between multiple parameters. The quantified dynamic behavior pattern is encoded into a feature vector to form a process anomaly signature.

6. The method for intelligent inspection and defect production parameter traceability of PCB boards according to claim 1, characterized in that: The specific process of establishing causal association knowledge elements includes: The defect context fingerprint corresponding to the reference and process exception signature in time and space; By forcibly pairing and binding process anomaly signatures with defect context fingerprints, a causal relationship is established from the dynamic behavior of a specific process to the specific defect form under a specific environment. The pairing and binding relationship is encapsulated into an independent and indivisible data unit, forming a causal knowledge element.

7. The method for intelligent inspection and defect production parameter traceability of PCB boards according to claim 1, characterized in that: The specific process of strengthening or creating nodes and relationships, aggregating and iteratively updating the dynamic defect knowledge graph includes: If there already exist node pairs and their relation edges in the knowledge network structure that are similar to the process anomaly signature and defect context fingerprint in the causal knowledge element, then strengthen the causal relation weight value corresponding to the relation edge. If there are no node pairs and their relational edges in the knowledge network structure that are similar to the process anomaly signature and defect context fingerprint in the causal knowledge element, then new nodes are created for the process anomaly signature and defect context fingerprint, and new relational edges and initial causal relation weight values ​​are established between the new nodes.

8. The method for intelligent inspection and defect production parameter traceability of PCB boards according to claim 1, characterized in that: The specific process of parsing into a high-risk process pattern set includes: Filter out all relation edges in the dynamic defect knowledge graph where the target node is a high-frequency or high-risk defect type; Among the filtered relationship edges, further filter out relationship edges whose causal relationship weight values ​​exceed a preset threshold; The source nodes of the re-selected relational edges, i.e., the process anomaly signatures, are aggregated to generate a high-risk process pattern set.

9. The method for intelligent inspection and defect production parameter traceability of PCB boards according to claim 1, characterized in that: The step of detecting dynamic behaviors matching a set of high-risk process patterns in real-time production data streams and generating pending early warning events containing risk level and current production batch information includes: Using the same dynamic behavior analysis method as the abnormal signature of the refining process, the current process feature vector is detected and generated in real time within the sliding time window of the real-time production data stream; Calculate the vector distance between the current process feature vector and the abnormal signatures of each process in the high-risk process pattern set; When the vector distance is less than the preset matching threshold, the pattern is determined to be a successful match; The process exception signature corresponding to the pattern that was successfully matched in the packaging; Query the dynamic defect knowledge graph to obtain the risk level corresponding to the defect type associated with the abnormal signature of the process; Query the Manufacturing Execution System to obtain information on the production batch currently being processed in this process. By integrating process anomaly signatures, risk levels, and current production batch information, a pending early warning event is generated.

10. The method for intelligent inspection and defect production parameter traceability of PCB boards according to claim 1, characterized in that: The specific process of generating contextualized intervention instructions that include risk prediction and location includes: Based on the process anomaly signature in the pending warning event, query the dynamic defect knowledge graph again to obtain all associated defect context fingerprint information. Obtain and analyze the circuit design layout features of the PCB board that is about to enter this process; By combining defect context fingerprint information with circuit design layout features, we can determine the specific defect types and high-risk areas that the abnormal signature in this process may cause on the PCB board to be processed. Risk predictions for specific defect types are generated based on the judgment results, and high-risk areas are clearly located based on the analysis of circuit design layout characteristics.

Images