A semiconductor device soldering quality on-line detection system

Abstract

This invention discloses an online inspection system for welding quality in semiconductor equipment, relating to the field of semiconductor packaging inspection technology. It includes: a data fusion module that achieves time-series alignment of multi-source heterogeneous data through standard industrial communication protocols; a knowledge graph construction module that establishes a four-element association relationship between defect type, process parameters, equipment status, and material batches; a process monitoring module that calculates CPK in real time and sets dynamic thresholds; a threshold decision module that dynamically adjusts the defect judgment tolerance based on CPK deviation using a fuzzy control algorithm; a knowledge graph reasoning module that uses a graph neural network to perform similar subgraph retrieval and output root cause probability ranking; and a process feedback module that maps root causes to upstream equipment correction instructions and updates the knowledge graph. This invention achieves a real-time closed loop of detection-evaluation-diagnosis-optimization, significantly improving the accuracy of welding quality inspection and the process adaptability.

Description

Technical Field

[0001] This invention relates to the field of semiconductor packaging inspection technology, specifically to an online inspection system for the welding quality of semiconductor equipment. Background Technology

[0002] As semiconductor packaging technology develops towards high density and miniaturization, soldering quality inspection has become a key link in ensuring chip reliability. Currently, the industry generally adopts the mainstream solution of combining automated optical inspection (AOI) and X-ray inspection (AXI). The appearance and internal structure of solder joints are inspected offline by independently deployed inspection equipment, and defects are judged based on static thresholds. Some advanced production lines have introduced statistical process control (SPC) methods and used process capability index (CPK) to periodically evaluate the stability of the production line. However, existing mainstream solutions have significant drawbacks in practical applications: First, AOI and AXI data are collected independently, lacking a time-series alignment mechanism, making it difficult to correlate and analyze multi-source data; second, defect judgment uses fixed thresholds, which cannot adapt to process fluctuations, resulting in a large number of misjudgments when equipment ages or materials change; third, CPK calculation is disconnected from defect detection, and process capability assessment results are not fed back to the detection standards in real time; more importantly, existing systems lack root cause tracing capabilities, and can only provide a binary "pass / fail" result when faced with complex defects, unable to pinpoint specific process parameters (such as reflow peak temperature, solder paste viscosity, etc.), causing process optimization to rely on manual experience and resulting in lagging closed-loop control. The root cause of the above problems lies in the silos of testing data, the static nature of judgment standards, and the separation of quality assessment and process control. Existing technologies have failed to establish a closed-loop system of "data fusion - dynamic judgment - root cause tracing - process feedback", which makes it difficult to meet the urgent needs of advanced packaging for online control of welding quality. Therefore, there is an urgent need for an online detection system that can integrate multi-source heterogeneous data, dynamically adjust detection thresholds based on process capabilities, and utilize knowledge graphs to achieve automatic root cause tracing and adaptive optimization of process parameters. Summary of the Invention

[0003] To address the aforementioned technical problems, an online inspection system for the welding quality of semiconductor equipment is provided. This technical solution solves the problems of isolated inspection data, static judgment standards, and the separation of quality assessment and process control. Existing technologies have failed to establish a closed-loop system of "data fusion - dynamic judgment - root cause tracing - process feedback".

[0004] To achieve the above objectives, the technical solution adopted by the present invention is as follows: An online inspection system for the welding quality of semiconductor equipment, comprising: The data fusion module collects data from automatic optical inspection, X-ray inspection, and reflow oven through standard industrial communication protocols, achieves time alignment through a precise time synchronization mechanism, and constructs a detection data packet containing feature vectors. A knowledge graph module is constructed. After preprocessing the detection data package, the four-element association relationship and its association strength of defect type, process parameter, equipment status and material batch are extracted to construct and update the welding defect knowledge graph. The process monitoring module extracts a subset of process parameters from the detection data packet, calculates the production line process capability index CPK in real time based on the subset of process parameters, sets a first threshold and a second threshold based on historical process capability data, and compares CPK with the first threshold and the second threshold to generate a multi-level process stability rating signal. The threshold decision module dynamically adjusts the defect judgment threshold set based on the deviation between the process stability rating signal and the target process capability using a fuzzy control algorithm. When CPK is lower than the first threshold, it automatically tightens the judgment tolerance of solder joint void rate and pillow effect. When CPK is higher than the second threshold, it moderately relaxes the judgment tolerance to optimize detection efficiency. Based on the adjusted threshold set, it performs defect judgment and generates abnormal feature markers. The knowledge graph reasoning module receives abnormal feature tags as graph structure query conditions, performs similar subgraph retrieval in the welding defect knowledge graph database, and outputs a root cause probability ranking list. The process feedback module generates upstream equipment correction instructions based on the process deviation items ranked first in the root cause probability sorting list, and then feeds back the correction execution results to update the knowledge graph database.

[0005] Preferably, the data fusion module includes: The multi-protocol communication interface unit establishes a real-time data connection with the automatic optical inspection equipment, X-ray inspection equipment, and reflow oven via OPC UA or TCP / IP protocol; The timing synchronization unit performs microsecond-level timing alignment of data streams from different detection devices based on hardware trigger signals or the IEEE 1588 precise time protocol. The feature vector encapsulation unit converts the aligned automatic optical inspection data, X-ray inspection data, and reflow oven data into solder joint appearance feature vectors, internal structure perspective feature vectors, and temperature curve feature vectors, respectively, and encapsulates them into the inspection data package containing a unified timestamp identifier.

[0006] Preferably, the knowledge graph construction module includes: The data preprocessing unit is used to clean, standardize, and align the historical data of the detection data packet to generate a structured process log. The entity extraction unit extracts defect type entities, process parameter entities, equipment status entities, and material batch entities from the structured process log, and generates a unique identifier for each entity. The relation modeling unit, based on timestamp association and co-occurrence frequency analysis, establishes a four-element association relationship between defect type entities and process parameter entities, equipment status entities, and material batch entities, and generates relation edges and weights. The graph storage and update unit stores the defect type entity, process parameter entity, equipment status entity, material batch entity and their four-element association relationship as graph structure data, and updates the welding defect knowledge graph database incrementally based on real-time detection data.

[0007] Preferably, the process monitoring module performs the following steps: The reflux peak temperature, heating slope, and holding time are extracted from the detection data packet as a subset of process parameters. Obtain the upper and lower limits of the specifications of the process parameter subset, and calculate the actual fluctuation range of the process parameter subset to quantify the degree of dispersion; Based on the upper limit of the specification, the lower limit of the specification, and the actual fluctuation range, the process capability index CPK is obtained by calculating the ratio of the specification tolerance to six times the actual fluctuation range. The first threshold and the second threshold are determined based on the central trend and fluctuation range of the historical process capability data, wherein the first threshold corresponds to the lower limit threshold of the process capability and the second threshold corresponds to the upper limit threshold of the process capability. The CPK is compared with the first threshold and the second threshold. When the CPK is lower than the first threshold, a high-risk rating is generated; when it is between the first threshold and the second threshold, a medium-risk rating is generated; and when it is higher than the second threshold, a low-risk rating is generated, which serves as a multi-level process stability rating signal.

[0008] Preferably, the threshold decision module employs the following fuzzy control steps: The deviation and rate of change of the CPK from the preset process capability standard value are fuzzified to generate fuzzy input quantities; Fuzzy inference is performed based on a preset fuzzy rule base to calculate the correction coefficient of the defect judgment threshold set. The fuzzy rule base includes rules that increase the correction coefficient when the deviation is at a negative high level and the rate of change is negative. The fuzziness is resolved by the centroid method, and the specific adjustment amount of the judgment tolerance is output.

[0009] Preferably, the threshold decision module includes: The multi-defect collaborative adjustment unit receives the specific adjustment amount of the judgment tolerance and sets judgment tolerances for solder joint void rate and pillow effect respectively; when CPK is lower than the first threshold, the judgment tolerance of each defect type is tightened differentially based on a preset tightening coefficient, wherein the tightening coefficient of solder joint void rate is greater than the tightening coefficient of pillow effect; and when the CPK is higher than the second threshold, the judgment tolerance of each defect type is widened proportionally based on a preset relaxation coefficient.

[0010] Preferably, the threshold decision module further includes: The defect determination execution unit performs point-by-point defect determination on the solder joint appearance inspection data and internal structure perspective inspection data in the data packet based on the adjusted determination tolerance; and when a defect is determined to exist, it extracts the defect location, defect type and confidence level to generate anomaly feature markers.

[0011] Preferably, the knowledge graph reasoning module includes: The query condition generation unit maps the abnormal feature markers into graph-structured query conditions that include defect type attributes, process parameter constraints, and equipment status constraints. The subgraph matching unit traverses candidate subgraphs with similar structures to the query conditions in the welding defect knowledge graph database based on graph neural networks, calculates structural similarity scores, and selects the target subgraph with the highest similarity.

[0012] Preferably, the knowledge graph reasoning module further includes: The root cause assessment unit calculates the confidence score of each root cause based on the degree of correlation and influence between each root cause node and the defect node in the target subgraph. The sorting output unit generates a root cause probability sorting list by sorting the confidence scores in descending order, and outputs the root cause items at the top of the sorting list according to their confidence scores.

[0013] Preferably, the process feedback module includes: The instruction generation and issuance unit parses the high-confidence process deviation items in the root cause probability ranking list, maps them to the process parameter correction values ​​of the upstream equipment, generates correction instructions that conform to the standard communication specifications of semiconductor equipment, and issues them. The feedback update unit receives correction execution feedback data returned by upstream equipment, extracts the actual adjustment amount of process parameters and the post-adjustment test pass rate, and associates the actual adjustment amount and the post-adjustment test pass rate to the knowledge graph database to optimize the reliability assessment of the corresponding root cause node.

[0014] Compared with the prior art, the beneficial effects of the present invention are as follows: This invention proposes a multi-source data fusion architecture based on standard industrial communication protocols and precise time synchronization mechanisms. This architecture achieves microsecond-level time alignment and unified encapsulation of AOI, AXI, and reflow oven data, eliminating data silos and providing a data foundation for multi-dimensional quality analysis. It also proposes a comparison mechanism based on the process capability index CPK and dynamic thresholds to achieve real-time quantitative rating of process stability, addressing the technical problem that traditional static thresholds cannot adapt to process fluctuations. Furthermore, it proposes a dynamic adjustment strategy for defect judgment thresholds based on fuzzy control algorithms. This strategy adaptively tightens or relaxes the judgment tolerance for solder joint void rate and pillow effect based on CPK deviation and change rate, improving detection sensitivity when process capability is insufficient and optimizing detection efficiency when process capability is abundant, thus achieving quality control and capacity balance. Finally, it proposes a similar subgraph retrieval and root cause probability ranking mechanism based on graph neural networks. This mechanism maps abnormal features to graph structure query conditions, and through structural similarity calculation and confidence assessment, achieves automatic tracing from defect phenomena to process root causes, replacing manual experience analysis. Attached Figure Description

[0015] Figure 1 This is a system framework diagram of the present invention. Detailed Implementation

[0016] The following description is intended to disclose the invention and enable those skilled in the art to implement it. The preferred embodiments described below are merely examples, and other obvious variations will occur to those skilled in the art.

[0017] Reference Figure 1 As shown, an online inspection system for the welding quality of semiconductor equipment includes: The data fusion module collects data from automatic optical inspection, X-ray inspection, and reflow oven through standard industrial communication protocols, achieves time alignment through a precise time synchronization mechanism, and constructs a detection data packet containing feature vectors. The data fusion module includes: The multi-protocol communication interface unit establishes real-time data connections with the automated optical inspection equipment, X-ray inspection equipment, and reflow oven via OPC UA or TCP / IP protocols. At the communication protocol configuration level, the system achieves data access from multiple sources through standard industrial communication protocols. The OPC UA protocol configuration requires defining specific communication parameters, including serial communication indicators such as baud rate, data bits, and parity bits, as well as the node ID and namespace definition of the OPC UA information model. Simultaneously, a TCP / IP three-way handshake connection mechanism and a heartbeat detection mechanism are established to ensure communication stability. For hardware interface adaptation of the automated optical inspection equipment (AOI), X-ray inspection equipment (AXI), and reflow oven, corresponding physical interfaces need to be configured according to the equipment type, including GigE Vision or Camera Link interfaces for high-speed image transmission, RS-485 interfaces for temperature control equipment communication, and standard Ethernet ports for general data interaction, and corresponding driver adaptation methods are completed.

[0018] The timing synchronization unit performs microsecond-level timing alignment of data streams from different detection devices based on hardware trigger signals or the IEEE 1588 precise time protocol. Regarding the timing alignment mechanism, the system uses hardware trigger signals or the IEEE 1588 precise time protocol to achieve microsecond-level timing synchronization. For the hardware trigger scheme, the trigger source type needs to be configured, including encoder zero-position pulses, photoelectric gating signals, or external triggers. The electrical characteristics of the trigger signal (such as TTL level or differential signal) must be clearly defined, and a trigger delay compensation algorithm must be designed to eliminate microsecond-level signal jitter. For the IEEE 1588 protocol scheme, the master clock selection algorithm, clock synchronization period (recommended to be set to 1 second), timestamp accuracy (to achieve microsecond-level accuracy), and network delay measurement and compensation mechanism (based on a delay request-response mechanism) must be disclosed.

[0019] The feature vector encapsulation unit converts the aligned automatic optical inspection data, X-ray inspection data, and reflow oven data into solder joint appearance feature vectors, internal structure perspective feature vectors, and temperature curve feature vectors, respectively, and encapsulates them into the inspection data package containing a unified timestamp identifier. Specifically, in the feature vector construction stage, for solder joint appearance feature extraction, the system uses gray-level co-occurrence matrix (GLCM) parameters for texture analysis, combines Canny or Sobel operators for edge detection, and extracts specific quantitative indicators such as crack length, contamination area, and color deviation as feature dimensions. For internal structure perspective feature extraction, internal defect features are obtained through an X-ray image void ratio calculation algorithm (based on threshold segmentation and pixel statistics) and a pillow effect height difference measurement method (using a sub-pixel-level edge localization algorithm). For temperature curve feature extraction, a sliding window extreme value detection algorithm is used to identify the reflow peak temperature, the least squares method is used to fit and calculate the heating slope, and a time-temperature curve alignment method is implemented. Finally, the above multi-dimensional features are encapsulated into an inspection data package containing a unified timestamp identifier.

[0020] A knowledge graph module is constructed to preprocess the detection data package and extract the four-element association relationship and its association strength of defect type, process parameter, equipment status and material batch, and to construct and update the welding defect knowledge graph database. The knowledge graph construction module includes: The data preprocessing unit is used to clean, standardize, and align the historical data of the detection data package to generate a structured process log. In the data preprocessing stage, the system first performs data cleaning, employing interpolation or deletion strategies for missing values, detecting outliers based on the 3σ principle or IQR method, and applying Kalman filtering or median filtering to remove noise. Subsequently, data standardization is performed, including Min-Max normalization or Z-score standardization parameter configuration, achieving uniformity of temperature dimensions (°C to °F conversion) and standardization of time format (using ISO 8601 standard). Finally, a structured process log is generated, which must conform to a predefined JSON or XML Schema format, clearly defining field names, data types, and constraints. The entity extraction unit extracts defect type entities, process parameter entities, equipment status entities, and material batch entities from the structured process log, and generates a unique identifier for each entity. The relation modeling unit, based on timestamp association and co-occurrence frequency analysis, establishes a four-element association relationship between defect type entities and process parameter entities, equipment status entities, and material batch entities, and generates relation edges and weights. The graph storage and update unit stores the defect type entities, process parameter entities, equipment status entities, material batch entities, and their four-element associations as graph structure data. It incrementally updates the welding defect knowledge graph database based on real-time detection data. Regarding the extraction of four-element associations, the system extracts four key entities from the preprocessed data using entity recognition algorithms: for defect type entities, identification is based on AOI defect code mapping rules (e.g., mapping "CRACK" to cracks and "VOID" to voids); for process parameter entities, key parameters such as peak temperature, chain speed, and nitrogen flow rate are extracted from reflow oven logs; for equipment status entities, status data such as pick-and-place machine vacuum and printer blade pressure are collected in real-time; and for material batch entities, batch information is extracted using SN code parsing rules (including supplier code, date code, and serial number). Simultaneously, unique identifiers are generated for each type of entity, using a UUID generation algorithm or a hash-based entity encoding method to ensure entity uniqueness.

[0021] In addition, in the association strength calculation stage, the system adopts the co-occurrence frequency analysis method, sets the sliding time window size (e.g., ±30 minutes), and applies the Jaccard similarity or cosine similarity calculation formula to measure the degree of association between entities; for the quantification of relation edge weights, the system discloses the weight normalization method (mapped to the 0-1 interval) and the time decay factor (to give more weight to recent data), thereby constructing a quaternary association relationship and its association strength that includes defect type, process parameters, equipment status and material batch.

[0022] The process monitoring module extracts a subset of process parameters from the detection data packet, calculates the production line process capability index CPK in real time based on the dispersion of the process parameter subset, sets a first threshold and a second threshold based on the statistical distribution characteristics of historical process capability data, and compares the CPK with the first threshold and the second threshold to generate a multi-level process stability rating signal. The process monitoring module performs the following steps: The reflux peak temperature, heating slope, and holding time are extracted from the detection data packet as a subset of process parameters. Obtain the upper and lower limits of the specifications of the process parameter subset, and calculate the actual fluctuation range of the process parameter subset to quantify the degree of dispersion; Based on the upper limit of the specification, the lower limit of the specification, and the actual fluctuation range, the process capability index CPK is obtained by calculating the ratio of the specification tolerance to six times the actual fluctuation range. The first threshold and the second threshold are determined based on the central trend and fluctuation range of the historical process capability data, wherein the first threshold corresponds to the lower limit threshold of the process capability and the second threshold corresponds to the upper limit threshold of the process capability. The CPK is compared with the first threshold and the second threshold. When the CPK is lower than the first threshold, a high-risk rating is generated; when it is between the first threshold and the second threshold, a medium-risk rating is generated; and when it is higher than the second threshold, a low-risk rating is generated, which serves as a multi-level process stability rating signal.

[0023] The specific process of the process monitoring module is as follows: Regarding the CPK calculation algorithm, the system first extracts a subset of process parameters from the detection data packet, specifying the reflux peak temperature, heating slope, and holding time. The sampling frequency is set to 10 times per second, and a moving average window is used for data smoothing. For the setting of the upper and lower limits of specifications (USL / LSL), the parameters are determined according to the IPC-A-610 standard, customer specifications, or process window experiments. If based on historical best values, the Top 10% mean screening algorithm is used. The actual fluctuation range is calculated using the standard deviation σ to quantify the degree of dispersion, clearly distinguishing the application scenarios of sample standard deviation and population standard deviation, and calculating the Cp and Cpk exponents (where Cpk considers the process center offset). Regarding the threshold setting algorithm, the system collects historical process capability data based on a 30-day sliding window, calculates the central tendency (arithmetic mean or median) and fluctuation range (standard deviation or interquartile range), and then determines the first and second thresholds. Specific algorithms can use the first threshold = μ - 3σ and the second threshold = μ + 3σ, or be set based on percentile P5 / P95. A dynamic update mechanism is also established, setting an hourly or daily update cycle, and triggering threshold recalculation when CPK deviates from the threshold more than N times. Regarding rating signal generation, the system generates multi-level process stability rating signals based on the comparison between CPK and the thresholds, clearly defining three standards: a high-risk rating (corresponding to digital signal 0) is generated when CPK < 1.0; a medium-risk rating (corresponding to digital signal 1) is generated when 1.0 ≤ CPK < 1.33; and a low-risk rating (corresponding to digital signal 2) is generated when CPK ≥ 1.33. The signal output supports specific circuit implementations for digital signals (0 / 1 / 2) or analog signals (4-20mA / 0-10V).

[0024] The threshold decision module dynamically adjusts the defect judgment threshold set based on the deviation between the process stability rating signal and the target process capability using a fuzzy control algorithm. When the CPK is lower than the first threshold, the judgment tolerance of the solder joint void rate and pillow effect is automatically tightened. When the CPK is higher than the second threshold, the judgment tolerance is appropriately relaxed to optimize the detection efficiency. Based on the adjusted threshold set, the module performs defect judgment on the multi-dimensional feature data after parsing the data packet to generate abnormal feature markers. The threshold decision module employs the following fuzzy control steps: The deviation and rate of change of the CPK from the preset process capability standard value are fuzzified to generate fuzzy input quantities; Fuzzy inference is performed based on a preset fuzzy rule base to calculate the correction coefficient of the defect judgment threshold set. The fuzzy rule base includes rules that increase the correction coefficient when the deviation is at a negative high level and the rate of change is negative. The fuzziness is resolved by the centroid method, and the specific adjustment amount of the judgment tolerance is output.

[0025] In terms of fuzzy control algorithm implementation, the input variables (CPK deviation e and its rate of change de / dt) are first fuzzified using membership functions (triangular, trapezoidal, or Gaussian), and the universe of discourse is divided into seven or five levels: NB / NM / NS / ZO / PS / PM / PB. Then, inference is performed based on a preset fuzzy rule base, which contains a complete rule table (e.g., "IF e is NB AND de / dt is NS THEN correction coefficient is PB"), with a recommended number of 21-49 rules. Finally, the fuzziness is resolved using the centroid method (weighted average method), and the judgment tolerance adjustment amount ΔT is output, with its physical unit specified as pixels, millimeters, or percentages.

[0026] The threshold decision module includes: The multi-defect collaborative adjustment unit receives the specific adjustment amount of the judgment tolerance and sets judgment tolerances for solder joint void rate and pillow effect respectively; when CPK is lower than the first threshold, the judgment tolerance of each defect type is tightened differentially based on a preset tightening coefficient, wherein the tightening coefficient of solder joint void rate is greater than the tightening coefficient of pillow effect; and when the CPK is higher than the second threshold, the judgment tolerance of each defect type is widened proportionally based on a preset relaxation coefficient.

[0027] Judgment tolerances are set separately for solder joint void rate and pillow effect. The default threshold for solder joint void rate is set to 25% (tightened to 15%), and the default threshold for pillow effect height difference is set to 50μm (tightened to 30μm). When CPK is lower than the first threshold, the tolerances for each defect are tightened differentially based on a preset tightening coefficient α (e.g., 0.6, determined by historical debugging data or expert experience), and the void rate tightening coefficient is greater than that for the pillow effect (based on the weight analysis of the impact of defects on reliability). When CPK is higher than the second threshold, the judgment tolerances are relaxed proportionally based on a preset relaxation coefficient β (e.g., 1.2) to optimize detection efficiency.

[0028] The threshold decision module also includes: The defect determination execution unit performs point-by-point defect determination on the solder joint appearance inspection data and internal structure perspective inspection data in the data packet based on the adjusted determination tolerance; and when a defect is determined to exist, it extracts the defect location, defect type and confidence level to generate anomaly feature markers.

[0029] The knowledge graph reasoning module receives the abnormal feature tags as graph structure query conditions, performs similar subgraph retrieval in the welding defect knowledge graph database, and outputs a root cause probability ranking list. Defect determination is performed based on the adjusted threshold set. First, the region of interest (ROI) is extracted by traversing the detection data packets, and a double threshold hysteresis comparison logic is used to avoid jitter. When a defect is determined to exist, an abnormal feature label containing defect ID, location coordinates (X,Y), defect type code, confidence probability and timestamp is generated, and this label is used as the input condition for subsequent knowledge graph reasoning.

[0030] The knowledge graph reasoning module includes: The query condition generation unit maps the abnormal feature markers into graph-structured query conditions that include defect type attributes, process parameter constraints, and equipment status constraints. The subgraph matching unit traverses candidate subgraphs with similar structures to the query conditions in the welding defect knowledge graph database based on graph neural networks, calculates structural similarity scores, and selects the target subgraph with the highest similarity.

[0031] The knowledge graph reasoning module also includes: The root cause assessment unit calculates the confidence score of each root cause based on the degree of correlation and influence between each root cause node and the defect node in the target subgraph. The sorting output unit generates a root cause probability sorting list by sorting the confidence scores in descending order, and outputs the root cause items at the top of the sorting list according to their confidence scores.

[0032] The knowledge graph reasoning module process is as follows: In the query condition generation stage, the system maps the received abnormal feature tags into graph database query statements. Specifically, it can use Cypher query templates or Gremlin traversal statements, while setting attribute constraints, including encoding methods for process parameter constraints (e.g., peak temperature > 245℃) and equipment status constraints (e.g., vacuum degree < -50kPa). This constructs a graph structure query condition containing defect type attributes, process parameter constraints, and equipment status constraints. In the similarity subgraph retrieval stage, the system traverses candidate subgraphs in the welding defect knowledge graph database based on a graph neural network (GNN). The GNN architecture can be GCN, GraphSAGE, or GAT, with 2-3 layers, using ReLU or LeakyReLU activation functions, and node embedding dimensions set to 64 or 128. Target subgraphs are selected by calculating structural similarity scores. Similarity calculation methods include cosine similarity, Euclidean distance, or graph edit distance (GED), and the Top-K algorithm is used to select the K candidate subgraphs with the highest similarity (e.g., K=5). In the root cause assessment and ranking stage, the system calculates confidence scores based on the degree of association and influence intensity between each root cause node and defect node in the target subgraph. The degree of association is quantified by a path length (hop count) normalization method (e.g., 1 / hop count), and the influence intensity is calculated by an edge weight aggregation algorithm (weighted summation or max pooling). The confidence score can be calculated using Bayesian inference formulas or the RootCauseRank algorithm based on PageRank. Finally, a root cause probability ranking list is generated by a descending sorting algorithm (quick sort or heap sort), and the top-ranked root cause items are selected as output based on a confidence threshold (e.g., >0.7) or Top-N selection (e.g., N=3).

[0033] The process feedback module generates upstream equipment correction instructions based on the process deviation items ranked first in the root cause probability sorting list, and feeds back the correction execution results to update the knowledge graph database.

[0034] The process feedback module includes: The instruction generation and issuance unit parses the high-confidence process deviation items in the root cause probability ranking list, maps them to the process parameter correction values ​​of the upstream equipment, generates correction instructions that conform to the standard communication specifications of semiconductor equipment, and issues them. The feedback update unit receives correction execution feedback data returned by upstream equipment, extracts the actual adjustment amount of process parameters and the post-adjustment test pass rate, and associates the actual adjustment amount and the post-adjustment test pass rate to the knowledge graph database to optimize the reliability assessment of the corresponding root cause node.

[0035] The foregoing has shown and described the basic principles, main features, and advantages of the present invention. Those skilled in the art should understand that the present invention is not limited to the above embodiments. The embodiments and descriptions in the specification are merely principles of the invention. Various changes and modifications can be made to the invention without departing from its spirit and scope, and all such changes and modifications fall within the scope of the claimed invention. The scope of protection claimed by the appended claims and their equivalents is defined.

Claims

1. An online inspection system for the welding quality of semiconductor equipment, characterized in that, include: The data fusion module collects data from automatic optical inspection, X-ray inspection, and reflow oven through standard industrial communication protocols, achieves time alignment through a precise time synchronization mechanism, and constructs a detection data packet containing feature vectors. A knowledge graph module is constructed. After preprocessing the detection data package, the four-element association relationship and its association strength of defect type, process parameter, equipment status and material batch are extracted to construct and update the welding defect knowledge graph. The process monitoring module extracts a subset of process parameters from the detection data packet, calculates the production line process capability index CPK in real time based on the subset of process parameters, sets a first threshold and a second threshold based on historical process capability data, and compares CPK with the first threshold and the second threshold to generate a multi-level process stability rating signal. The threshold decision module dynamically adjusts the defect judgment threshold set based on the deviation between the process stability rating signal and the target process capability using a fuzzy control algorithm. When CPK is lower than the first threshold, it automatically tightens the judgment tolerance of solder joint void rate and pillow effect. When CPK is higher than the second threshold, it moderately relaxes the judgment tolerance to optimize detection efficiency. Based on the adjusted threshold set, it performs defect judgment and generates abnormal feature markers. The knowledge graph reasoning module receives abnormal feature tags as graph structure query conditions, performs similar subgraph retrieval in the welding defect knowledge graph database, and outputs a root cause probability ranking list. The process feedback module generates upstream equipment correction instructions based on the process deviation items ranked first in the root cause probability sorting list, and then feeds back the correction execution results to update the knowledge graph database.

2. The online inspection system for semiconductor equipment welding quality according to claim 1, characterized in that, The data fusion module includes: The multi-protocol communication interface unit establishes a real-time data connection with the automatic optical inspection equipment, X-ray inspection equipment, and reflow oven via OPC UA or TCP / IP protocol; The timing synchronization unit performs microsecond-level timing alignment of data streams from different detection devices based on hardware trigger signals or the IEEE 1588 precise time protocol. The feature vector encapsulation unit converts the aligned automatic optical inspection data, X-ray inspection data, and reflow oven data into solder joint appearance feature vectors, internal structure perspective feature vectors, and temperature curve feature vectors, respectively, and encapsulates them into the inspection data package containing a unified timestamp identifier.

3. The online inspection system for semiconductor equipment welding quality according to claim 1, characterized in that, The knowledge graph construction module includes: The data preprocessing unit is used to clean, standardize, and align the historical data of the detection data packet to generate a structured process log. The entity extraction unit extracts defect type entities, process parameter entities, equipment status entities, and material batch entities from the structured process log, and generates a unique identifier for each entity. The relation modeling unit, based on timestamp association and co-occurrence frequency analysis, establishes a four-element association relationship between defect type entities and process parameter entities, equipment status entities, and material batch entities, and generates relation edges and weights. The graph storage and update unit stores the defect type entity, process parameter entity, equipment status entity, material batch entity and their four-element association relationship as graph structure data, and updates the welding defect knowledge graph database incrementally based on real-time detection data.

4. The online inspection system for semiconductor equipment welding quality according to claim 1, characterized in that, The process monitoring module performs the following steps: The reflux peak temperature, heating slope, and holding time are extracted from the detection data packet as a subset of process parameters. Obtain the upper and lower limits of the specifications of the process parameter subset, and calculate the actual fluctuation range of the process parameter subset to quantify the degree of dispersion; Based on the upper limit of the specification, the lower limit of the specification, and the actual fluctuation range, the process capability index CPK is obtained by calculating the ratio of the specification tolerance to six times the actual fluctuation range. The first threshold and the second threshold are determined based on the central trend and fluctuation range of the historical process capability data, wherein the first threshold corresponds to the lower limit threshold of the process capability and the second threshold corresponds to the upper limit threshold of the process capability. The CPK is compared with the first threshold and the second threshold. When the CPK is lower than the first threshold, a high-risk rating is generated; when it is between the first threshold and the second threshold, a medium-risk rating is generated; and when it is higher than the second threshold, a low-risk rating is generated, which serves as a multi-level process stability rating signal.

5. The online inspection system for semiconductor equipment welding quality according to claim 1, characterized in that, The threshold decision module employs the following fuzzy control steps: The deviation and rate of change of the CPK from the preset process capability standard value are fuzzified to generate fuzzy input quantities; Fuzzy inference is performed based on a preset fuzzy rule base to calculate the correction coefficient of the defect judgment threshold set. The fuzzy rule base includes rules that increase the correction coefficient when the deviation is at a negative high level and the rate of change is negative. The fuzziness is resolved by the centroid method, and the specific adjustment amount of the judgment tolerance is output.

6. The online inspection system for semiconductor equipment welding quality according to claim 5, characterized in that, The threshold decision module includes: The multi-defect collaborative adjustment unit receives the specific adjustment amount of the judgment tolerance and sets judgment tolerances for solder joint void rate and pillow effect respectively; when CPK is lower than the first threshold, the judgment tolerance of each defect type is tightened differentially based on a preset tightening coefficient, wherein the tightening coefficient of solder joint void rate is greater than the tightening coefficient of pillow effect; and when the CPK is higher than the second threshold, the judgment tolerance of each defect type is widened proportionally based on a preset relaxation coefficient.

7. The online inspection system for semiconductor equipment welding quality according to claim 6, characterized in that, The threshold decision module also includes: The defect determination execution unit performs point-by-point defect determination on the solder joint appearance inspection data and internal structure perspective inspection data in the data packet based on the adjusted determination tolerance; and when a defect is determined to exist, it extracts the defect location, defect type and confidence level to generate anomaly feature markers.

8. The online inspection system for semiconductor equipment welding quality according to claim 1, characterized in that, The knowledge graph reasoning module includes: The query condition generation unit maps the abnormal feature markers into graph-structured query conditions that include defect type attributes, process parameter constraints, and equipment status constraints. The subgraph matching unit traverses candidate subgraphs with similar structures to the query conditions in the welding defect knowledge graph database based on graph neural networks, calculates structural similarity scores, and selects the target subgraph with the highest similarity.

9. The online inspection system for semiconductor equipment welding quality according to claim 8, characterized in that, The knowledge graph reasoning module also includes: The root cause assessment unit calculates the confidence score of each root cause based on the degree of correlation and influence between each root cause node and the defect node in the target subgraph. The sorting output unit generates a root cause probability sorting list by sorting the confidence scores in descending order, and outputs the root cause items at the top of the sorting list according to their confidence scores.

10. The online inspection system for semiconductor equipment welding quality according to claim 1, characterized in that, The process feedback module includes: The instruction generation and issuance unit parses the high-confidence process deviation items in the root cause probability ranking list, maps them to the process parameter correction values ​​of the upstream equipment, generates correction instructions that conform to the standard communication specifications of semiconductor equipment, and issues them. The feedback update unit receives correction execution feedback data returned by upstream equipment, extracts the actual adjustment amount of process parameters and the test pass rate after adjustment, and stores the actual adjustment amount and the test pass rate after adjustment in the knowledge graph database to optimize the reliability assessment of the corresponding root cause node.

Images