An IC carrier board production line process parameter data processing method and system

Abstract

This invention belongs to the field of IC substrate production line process parameter data processing technology, and relates to a method and system for processing IC substrate production line process parameter data. The method includes: unifying and robustly standardizing multi-source process data; verifying synchronization, authentication, and consistency to generate confidence levels and aggregating process events according to substrate identifiers; constructing a substrate-level heterogeneous spatiotemporal correlation graph to form candidate paths for anomaly propagation; uniformly outputting the basic process model to achieve anomaly identification; locating causal chains under candidate path constraints and providing interventionable root causes; driving collaborative parameter tuning based on root causes and completing the mapping of hard boundaries, ramp rates, and process window constraints, triggering freeze or minimum action and auditing. The technical solution of this application achieves integrated anomaly identification, root cause location, and safety parameter tuning through substrate-level cross-process correlation modeling and dynamic discrimination and tracing collaborative closed loop, which can improve the consistency of location tracing and the stability of discrimination, and meet the needs of stable monitoring and linkage optimization under strongly coupled process conditions.

Description

Technical Field

[0001] This invention belongs to the field of IC substrate production line process parameter data processing technology, and specifically relates to a method and system for processing IC substrate production line process parameter data. Background Technology

[0002] In existing technologies, process parameter data processing for IC substrate production lines typically relies on the centralized aggregation of data from production equipment, quality inspection records, and environmental monitoring data. This is achieved through statistical analysis and rule-based judgment of multi-process data to meet the needs of process stability monitoring, anomaly warning, and quality traceability. However, existing process parameter data processing methods have some significant shortcomings in terms of cross-process data consistency and closed-loop optimization collaboration.

[0003] In practical applications, production lines cover multiple high-precision processes such as laser drilling, electroplating, lamination, plasma treatment, and surface treatment. The parameters are multi-dimensional and the processes are tightly coupled. Data sources differ in format, protocol, unit, and sampling rate. Even after centralized aggregation, semantic alignment and correlation modeling accuracy are often limited. The overall stability of the traceability chain is weak in cross-equipment and cross-process scenarios. At the same time, anomaly judgment often uses relatively fixed thresholds and single-point alarm logic, which has limited adaptability to equipment state evolution, environmental fluctuations, and differences in substrates. This results in weak stability of the judgment criteria, which in turn leads to a low overall response efficiency of cross-process linkage optimization.

[0004] Therefore, existing technologies often suffer from limitations in the consistency control of process anomaly location and traceability results under the condition of strong coupling of multiple processes on IC substrates, and weak adaptive stability of anomaly identification to changes in operating conditions. These are the shortcomings of existing technologies.

[0005] In view of this, it is very necessary to provide a method and system for processing process parameter data of IC substrate production line in order to solve the above-mentioned defects in the prior art. Summary of the Invention

[0006] The purpose of this invention is to address the shortcomings of existing technologies, such as limited control over the consistency of process anomaly location and traceability results under strong coupling of multiple processes in IC substrates, and weak adaptive stability of anomaly discrimination to changes in operating conditions. This invention provides a method and system for processing process parameter data in IC substrate production lines to solve the aforementioned technical problems.

[0007] To achieve the above objectives, the present invention provides the following technical solution: In a first aspect, this application provides a method for processing process parameter data of an IC substrate production line, including: Acquire process parameters, equipment status, quality inspection and environmental data for the target process nodes in the IC substrate production line and complete unit unification and robust standardization; Based on the results of time synchronization verification, message authentication digest, sequence consistency verification and identity token verification, a confidence field is generated and encapsulated into a process event message. The process event message is aggregated according to the unique identifier of the IC carrier board. A heterogeneous spatiotemporal correlation graph is constructed using the unique identifier of the IC carrier board as the root, and directed edges are generated based on time series relationships, influence weights, and confidence gating to construct a set of candidate paths for anomaly propagation. Based on the unified output of the process fundamental model, dynamic quantile thresholds, anomaly type distribution, anomaly scores, quality risk vectors, key parameter trajectory predictions, and out-of-distribution discrimination results are obtained and anomaly identification is completed. Under the constraint of the set of candidate paths for abnormal propagation, structural causal chain localization is performed and counterfactual effects are calculated. The output includes the set of interventionable root causes, lag window, evidence package and linkage sequence. Based on the set of interventionable root causes, generate collaborative parameter adjustment instructions and execute executable mappings of equipment hard boundary constraints, ramp rate constraints, and process window constraints. Based on the confidence field and out-of-distribution discrimination results, trigger freeze linkage or minimum action mode and record audit information.

[0008] By adopting the above technical solution, the multi-source heterogeneous process data of IC substrate production line is uniformly and standardized, and cross-process correlation modeling is performed for substrate objects. This enables adaptive discrimination of process anomalies, interpretable cause tracing, and safe and executable collaborative parameter tuning closed loop. It can stably output consistent anomaly location and traceability conclusions under strong coupling of multiple processes and maintain the adaptive stability of the discrimination criteria. This meets the requirements of consistent and controllable process anomaly location and traceability results, stable and reliable anomaly discrimination with changes in operating conditions, and controllable linkage optimization response.

[0009] This process involves acquiring process parameters, equipment status, quality inspection, and environmental data for target process nodes and standardizing and unifying units to ensure comparability in dimensions and scales across different equipment and processes, forming a unified data input foundation and enhancing the usability and consistency of cross-process analysis. Based on time synchronization verification, message authentication digests, sequence consistency verification, and identity token verification results, a confidence field is generated and encapsulated as a process event message. This message is then aggregated according to the unique identifier of the IC carrier board, enabling quantifiable quality characterization in terms of time consistency, integrity, and source credibility, which runs through subsequent analysis links, thereby improving the continuity of the traceability chain and the consistency of conclusions. Finally, a heterogeneous spatiotemporal correlation graph is constructed using the unique identifier of the IC carrier board as the root, and directed edges are generated based on time series relationships, influence weights, and confidence gating. This constructs a set of candidate paths for anomaly propagation, allowing cross-process influence relationships to be constrained at the structural level and converging the analytical boundary of anomaly propagation, improving the stability and interpretability of the localization process.

[0010] Furthermore, based on the unified output of the process fundamental model, dynamic quantile thresholds, anomaly type distributions, anomaly scores, quality risk vectors, key parameter trajectory predictions, and out-of-distribution discrimination results are obtained, and anomaly identification is completed. This ensures that threshold determination and risk assessment remain adaptively consistent with operating condition fluctuations while also considering the differentiation of anomaly patterns and the forward-looking characterization of future risks, enhancing the stability and reliability of anomaly discrimination criteria. Under the constraint of the anomaly propagation candidate path set, structural causal chain location is performed and counterfactual effects are calculated, outputting the set of intrusive root causes, lag windows, evidence packages, and linkage sequences, thus transforming the causal attribution results from correlations. The process transforms the target into an operable causal chain and forms auditable evidence to support the conclusions, ensuring consistency and verifiability. Based on the set of operable root causes, it generates collaborative parameter adjustment instructions and executes executable mappings of equipment hard boundary constraints, ramp rate constraints, and process window constraints. Simultaneously, based on the confidence field and out-of-distribution discrimination results, it triggers freeze linkage or minimum action mode and records audit information, ensuring that linkage parameter adjustment is implemented within the safety boundary and maintains a robust action strategy when data credibility or operating conditions deviate. This achieves controllable response and improved operational reliability for cross-process linkage optimization.

[0011] Preferably, the steps of generating a confidence field based on the results of time synchronization verification, message authentication digest, sequence consistency verification, and identity token verification, and encapsulating it into a process event message, include: The system verifies the consistency of time synchronization and provides the time synchronization consistency verification result; it verifies the message authentication digest of process event messages and provides the digest verification result; it verifies the sequential consistency of process event messages and provides the sequential consistency verification result; and it verifies the identity token and provides the token verification result. A computable confidence field is generated based on the time synchronization consistency verification result, digest verification result, sequence consistency verification result, and token verification result, and then written into the process event message field contract.

[0012] By adopting the above technical solution, a combination of time synchronization consistency verification, message authentication digest verification, sequence consistency verification and identity token verification is used. The results are then fused to generate a computable confidence field, which is written into the message field contract. This enables the quantitative expression of the trustworthy status of process event messages and consistent transmission throughout the entire chain. This can improve the stability of cross-process data aggregation and traceability conclusions and enhance the reliability of anomaly detection input.

[0013] Preferably, the step of generating a computable confidence field further includes: Replay detection is introduced for sequential consistency verification, and the confidence level of the token verification result is adjusted when sequence number replay is detected; A key rotation version consistency check is introduced for message authentication digest verification, and the credibility level of the digest verification result is adjusted when the key rotation version is inconsistent; A whitelist version consistency check is introduced for identity token verification, and the trust level of the token verification result is adjusted when the whitelist versions are inconsistent; Changes in the credibility level are written into the original evidence fragment index of the evidence package.

[0014] By adopting the above technical solution, the credibility level of each verification result is adjusted in conjunction with replay detection, key rotation version consistency verification and whitelist version consistency verification, and the change in credibility level is written into the original evidence fragment index of the evidence package. This enables traceable labeling of credibility fluctuations in abnormal data and solidification of evidence closed loop, which can improve the interpretability of audit review and reduce the risk of misjudgment caused by key or whitelist updates.

[0015] Preferably, the steps of constructing a heterogeneous spatiotemporal correlation graph and constructing a set of candidate paths for anomaly propagation include: The heterogeneous node types are defined as processes, equipment, events, parameters, quality, and environment. Generate time-series related edges based on time-series relationships and record lag features; generate influence-related edges based on influence weights and record the weight sources; generate confidence gates based on confidence fields and write the confidence gates into the directed edge generation rules. Based on the directed edge generation rules, a set of candidate paths for anomaly propagation is constructed and used to converge the cross-process causal search space and the linkage adjustment boundary.

[0016] By adopting the above technical solution, a heterogeneous set of nodes including processes, equipment, events, parameters, quality, and environment is used, and hysteresis features, weight sources, and confidence gating are introduced into the directed edge generation rules. This enables a unified expression and propagation boundary constraint of cross-process influence relationships at the structural level, which can improve the convergence efficiency of causal localization and enhance the determinism of linkage adjustment range control.

[0017] Preferably, the step of constructing the set of candidate paths for anomaly propagation further includes: Similarity candidate edges are generated based on similarity features, and a similarity score is formed; Based on the lag features, lag candidate edges are generated and lag scores are formed; Based on confidence gating, similarity candidate edges and lagging candidate edges are gating and filtering, and a gating score is generated; The path weights of candidate propagation paths are determined based on similarity scores, lag scores, and gating scores. An abnormal propagation candidate path set is formed according to the path weights, and the path weights are written into the evidence package.

[0018] By adopting the above technical solution, the path weight is determined by combining similarity score, lag score and gating score and written into the evidence package. This realizes the sorting and organization of the candidate propagation path set and the solidification of evidence, which can improve the interpretability of abnormal propagation inference and enhance the stability of key propagation link identification.

[0019] Preferably, the steps for obtaining dynamic quantile thresholds, anomaly type distributions, anomaly scores, quality risk vectors, key parameter trajectory predictions, and out-of-distribution discrimination results based on the unified output of the process fundamental model, and completing anomaly identification, include: The event sequence representation is obtained by encoding the event sequence of process event messages; Relation-aware encoding is performed on heterogeneous spatiotemporal correlation maps to obtain map representations; The event sequence representation and the graph representation are integrated to form a unified representation. Based on the unified representation, the dynamic quantile threshold, anomaly type distribution, anomaly score, quality risk vector, key parameter trajectory prediction and out-of-distribution discrimination results are output. The dynamic quantile threshold, anomaly score and out-of-distribution discrimination results are used together for anomaly identification and severity level classification.

[0020] By adopting the above technical solution, the event sequence representation and the graph representation are integrated into a unified representation, and the dynamic quantile threshold, anomaly score and out-of-distribution discrimination result are jointly output by the unified representation for hierarchical identification. This achieves consistent representation-driven threshold determination and risk assessment, which can improve the adaptive stability of the anomaly severity level classification to the fluctuation of working conditions and reduce the discrimination inconsistency caused by threshold drift.

[0021] Preferably, the steps for locating structural causal chains and calculating counterfactual effects under the constraints of the candidate path set for abnormal propagation, and outputting the set of intractable root causes, lag windows, evidence packages, and linkage sequences, include: Determine the set of variables on the subgraph induced by the set of candidate propagation paths for anomalies; A priori mask matrix is ​​generated based on the set of candidate paths for anomaly propagation, and the priori mask matrix is ​​used to restrict the candidate edges of the variable set. Under the constraint of the candidate edges, the directed dependencies between variables are learned and the acyclic constraint is introduced. Based on counterfactual intervention, the counterfactual effect of the root variable on the target quality variable is calculated, and an intrusive root cause set is formed according to the counterfactual effect. The intrusive root cause set is associated with the abnormal propagation candidate path set and written into the evidence package, and the linkage order and lag window are output.

[0022] By adopting the above technical solution, the variable set is determined by using the candidate path induced subgraph and the edge candidate is restricted by the prior mask matrix. After learning the directed dependency by combining the acyclic constraint, the counterfactual effect is calculated to form an intervenable root cause set. This realizes the intervenable causal output that couples the propagation prior with the structural causal inference, which can improve the verifiability of root cause location and enhance the reliability of the linkage sequence and the output of the lag window.

[0023] Preferably, the steps of generating collaborative parameter adjustment instructions based on the set of interventionable root causes and executing executable mappings of equipment hard boundary constraints, ramp rate constraints, and process window constraints, and triggering freeze linkage or minimum action mode based on the confidence field and out-of-distribution discrimination results and recording audit information include: Based on the set of operable root causes, the target process node and the target adjustment parameter set are determined, and collaborative parameter adjustment instructions are generated and written into candidate action vectors, execution time windows and idempotent keys. Perform equipment hard boundary constraint mapping, ramp rate constraint mapping, and process window constraint mapping on the candidate action vector to obtain the executable action vector and write the collaborative parameter adjustment instruction. Generate constraint trigger reason code and mapping residual and write audit information during each constraint mapping. Based on the confidence field and the out-of-distribution discrimination results, determine the freeze linkage mode or the minimum action mode, write the collaborative parameter adjustment instruction, and write the exit score as audit information; Issue collaborative parameter adjustment instructions and obtain execution receipts. Perform consistency verification on the execution receipts based on idempotent keys and write the verification results into the audit information.

[0024] By adopting the above technical solution, hard boundary mapping, ramp rate mapping and process window mapping of candidate action vectors are used, combined with exit score to trigger freeze linkage or minimum action mode and record receipt consistency audit, so as to realize the implementation and auditable closed loop of collaborative parameter tuning instructions under execution constraints. This can improve the safety determinism of linkage actions and enhance execution consistency and process traceability.

[0025] Secondly, this application also provides a process parameter data processing system for an IC substrate production line, comprising: The standardized data acquisition unit is used to obtain process parameters, equipment status, quality inspection and environmental data of the target process nodes in the IC substrate production line and to complete unit unification and robust standardization. The event trust unit is used to generate a confidence field based on the results of time synchronization verification, message authentication digest, sequence consistency verification and identity token verification, and encapsulate it into a process event message, and aggregate the process event message according to the unique identifier of the IC carrier board. The graph reasoning unit is used to construct a heterogeneous spatiotemporal correlation graph with the unique identifier of the IC carrier board as the root, and generate directed edges based on time series relationships, influence weights and confidence gating, construct a set of candidate paths for anomaly propagation, and obtain dynamic quantile thresholds, anomaly type distribution, anomaly scores, quality risk vectors, key parameter trajectory prediction and out-of-distribution discrimination results based on the unified output of the process basic model, and complete anomaly identification. The causal decision unit is used to locate the structural causal chain and calculate the counterfactual effect under the constraint of the set of candidate paths for abnormal propagation, and outputs the set of interventionable root causes, lag window, evidence package and linkage sequence. The linkage audit unit is used to generate collaborative parameter adjustment instructions based on the set of interventionable root causes and execute executable mappings of equipment hard boundary constraints, ramp rate constraints and process window constraints. It also triggers freeze linkage or minimum action mode based on confidence field and out-of-distribution discrimination results and records audit information.

[0026] As a preferred option, the causal decision unit includes: Subgraph-induced subunits are used to determine the set of variables on a subgraph induced by the set of anomaly propagation candidate paths; The mask constraint subunit is used to generate a priori mask matrix based on the set of anomaly propagation candidate paths and to restrict the edge candidates of the variable set using the priori mask matrix. Under the edge candidate constraint, the directed dependency between variables is learned and an acyclic constraint is introduced. The counterfactual assessment subunit is used to calculate the counterfactual effect of root variables on target quality variables based on counterfactual intervention and to form an interventionable root cause set according to the counterfactual effect; The evidence arrangement subunit is used to associate the set of interventionable root causes with the set of abnormal propagation candidate paths, write them into the evidence package, and output the linkage order and lag window.

[0027] As can be seen from the above technical solutions, the present invention has the following advantages: This application provides a method and system for processing process parameters of an IC substrate production line. By unifying and standardizing the multi-source heterogeneous process data of the IC substrate production line and modeling cross-process associations oriented towards the substrate object, it achieves adaptive discrimination of process anomalies, interpretable cause tracing, and safe and executable collaborative parameter tuning closed loop. It can stably output consistent anomaly location and traceability conclusions under strong coupling of multiple processes and maintain the adaptive stability of the discrimination criteria. It meets the requirements of consistent and controllable process anomaly location and traceability results, stable and reliable anomaly discrimination with changes in operating conditions, and controllable linkage optimization response. Attached Figure Description

[0028] To more clearly illustrate the technical solution of the present invention, the accompanying drawings used in the description will be briefly introduced below. Obviously, the accompanying 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.

[0029] Figure 1 This is a flowchart of a method for processing process parameter data in an IC substrate production line provided by the present invention; Figure 2 This is a schematic diagram of a process parameter data processing system for an IC substrate production line provided by the present invention.

[0030] The system comprises: 1. Standardized data collection unit; 2. Credible event unit; 3. Graph reasoning unit; 4. Causal decision-making unit; and 5. Linked auditing unit. Detailed Implementation

[0031] Various embodiments of this disclosure are described more fully below with reference to the accompanying drawings. This disclosure may have various embodiments, and adjustments and changes may be made therein. However, it should be understood that there is no intention to limit the various embodiments of this disclosure to the specific embodiments disclosed herein, but rather this disclosure should be understood to cover all adjustments, equivalents, and / or alternatives falling within the spirit and scope of the various embodiments of this disclosure.

[0032] In the following, the terms “comprising” or “may include”, which may be used in various embodiments of this disclosure, indicate the presence of the disclosed functions, operations, or elements, and do not limit the addition of one or more functions, operations, or elements. Furthermore, as used in various embodiments of this disclosure, the terms “comprising,” “having,” and their cognates are intended only to indicate a particular feature, number, step, operation, element, component, or combination of the foregoing, and should not be construed as primarily excluding the presence of one or more other features, numbers, steps, operations, elements, components, or combinations of the foregoing, or the possibility of adding one or more combinations of the foregoing.

[0033] It should be noted in advance that, in order to facilitate a clear and accurate description of the technical solutions in the embodiments of this application, the following is a brief explanation of some terms and related technologies involved in the embodiments of this application: 1. Structural causal chain: A type of causal modeling framework that describes possible causal dependencies between variables in the form of a directed graph and satisfies the acyclic constraint. It usually introduces causal assumptions and structural constraints on the basis of statistical correlation to distinguish between correlation and causal influence and to support analytical tasks such as root cause inference.

[0034] 2. Counterfactual effect: Under given causal model conditions, the causal effect is measured by replacing one variable with another value through intervention and comparing the differences in the results of the target variable. It is used to assess the potential influence of a variable as an interventionable factor on the target outcome.

[0035] 3. Out-of-distribution discrimination: This is a type of detection method used to determine whether input data deviates from the model training or historical normal data distribution. It is usually based on indicators such as feature space distance, confidence or density estimation to identify abnormal working conditions or unseen patterns to support robust discrimination and safe decision-making.

[0036] To address the limitations in cross-process traceability consistency control and weak adaptive stability of anomaly identification criteria to operating condition fluctuations caused by the strong coupling of multiple processes and the parallel existence of heterogeneous data from multiple sources during the process parameter data processing of IC substrate production lines, which makes it difficult to meet the actual needs of production sites for rapid anomaly identification, reliable root cause location, and cross-process linkage optimization, this application discloses a method and system for processing process parameter data of IC substrate production lines. By introducing a process event aggregation and cross-process correlation modeling mechanism oriented towards substrate objects, and combining dynamic threshold discrimination, causal chain location, and collaborative parameter adjustment process under safety constraints, it can uniformly process and correlate the collected process parameters, equipment status, quality inspection, and environmental data. This enables adaptive identification of abnormal states and interpretable root cause output, thereby effectively improving the consistency of anomaly location and traceability conclusions under multi-process conditions, significantly improving the stability and reliability of anomaly identification under different operating conditions, further enhancing the linkage optimization response efficiency and operational controllability of the production process, reducing reliance on manual experience, and ensuring the process stability and product quality consistency of the IC substrate production line.

[0037] The technical solutions of the embodiments of this application will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of this application, and not all embodiments. Based on the embodiments of this application, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of this application.

[0038] like Figure 1 As shown in the figure, this embodiment provides a method for processing process parameter data of an IC substrate production line, including: Step S1: Obtain process parameters, equipment status, quality inspection and environmental data for the target process node in the IC substrate production line and complete unit unification and robust standardization; Step S2: Generate a confidence field based on the results of time synchronization verification, message authentication digest, sequence consistency verification, and identity token verification, and encapsulate it into a process event message. Aggregate the process event message according to the unique identifier of the IC carrier board. Step S3: Construct a heterogeneous spatiotemporal correlation graph with the unique identifier of the IC carrier board as the root, and generate directed edges based on time series relationships, influence weights and confidence gating to construct a set of candidate paths for anomaly propagation; Step S4: Based on the unified output of the process basic model, obtain the dynamic quantile threshold, anomaly type distribution, anomaly score, quality risk vector, key parameter trajectory prediction and out-of-distribution discrimination results, and complete the anomaly identification. Step S5: Locate the structural causal chain and calculate the counterfactual effect under the constraints of the candidate path set for abnormal propagation, and output the set of interventionable root causes, lag window, evidence package and linkage sequence; Step S6: Generate collaborative parameter adjustment instructions based on the set of interventionable root causes and execute executable mappings of equipment hard boundary constraints, ramp rate constraints, and process window constraints. Trigger freeze linkage or minimum action mode based on confidence field and out-of-distribution discrimination results and record audit information.

[0039] This embodiment employs a unit-unified and robust standardization approach for process parameters, equipment status, quality inspection, and environmental data at target process nodes in the IC substrate production line. This ensures that data across equipment, processes, and sampling cycles can still form a comparable and consistent input basis. By introducing time synchronization verification, message authentication digests, sequence consistency verification, and identity token verification, and generating quantifiable confidence fields accordingly, process events are uniformly represented in terms of temporal consistency, integrity, and source credibility, thereby improving the continuity and consistency of the entire traceability chain. By constructing a heterogeneous spatiotemporal correlation graph with the unique identifier of the IC substrate as the root and introducing time-series relationships, influence weights, and confidence gating in edge generation, cross-process influence relationships are structurally constrained and the analytical boundary of anomaly propagation is converged, ensuring both interpretability and analytical efficiency. By adopting a unified output mechanism for the process basic model and introducing dynamic quantile thresholds and out-of-distribution discrimination, the anomaly discrimination criteria can remain adaptively stable with fluctuations in operating conditions, while simultaneously outputting anomaly type representations, Anomaly scoring, quality risk vectors, and key parameter trajectory prediction enhance the ability to characterize risk evolution. By locating structural causal chains and calculating counterfactual effects under the constraints of anomaly propagation candidate path sets, the causal chain expression is improved from correlation-based to an actionable causal chain expression, forming a verifiable output including lag windows, evidence packages, and linkage sequences, thus improving the stability and auditability of root cause localization. By generating collaborative parameter adjustment instructions based on the actionable root cause set and executing executable mappings of equipment hard boundary constraints, ramp rate constraints, and process window constraints, and by combining confidence fields and out-of-distribution discrimination results to trigger freeze linkage or minimum action modes and record audit information, the linkage parameter adjustment is implemented within the safety boundary and maintains a robust action strategy when data credibility is insufficient or operating conditions deviate. Overall, without relying on human experience to lead the judgment, adaptive identification, explainable causal attribution, and safety collaborative optimization of IC substrate production line process anomalies are achieved, reducing the workload of manual review and repeated parameter tuning, and improving process stability, quality consistency, and production operation controllability.

[0040] Hereinafter, steps S1 to S6 will be specifically described according to embodiments of this application.

[0041] In step S1, the core task is to establish a data foundation at the target process node that can be consistently aligned across equipment and processes, enabling process parameter data processing to move from "heterogeneous acquisition" to a computable state of "unified semantics and unified dimensions," and providing stable input for subsequent reliable encapsulation, spatiotemporal correlation modeling, and model inference. The input of this step is the original multi-source data stream of each target process node, and the output is the structured data record after unit unification and robust standardization, along with a missing mask vector and basic statistics for subsequent consistency verification.

[0042] In this embodiment of the application, process parameters, equipment status, quality inspection and environmental data of the target process node in the IC substrate production line can be obtained and unit unification and robust standardization can be achieved.

[0043] Specifically, target process nodes can be selected from core processes such as laser drilling, electroplating, lamination, plasma treatment, and surface treatment, and their priority is determined by a three-dimensional score of "quality contribution - cross-process propagation - manipulability". For example, candidate nodes... The contribution score can be calculated using:

[0044] Among them, contribution score Used for metric nodes Key parameter deviation magnitude For the magnitude of quality deviation The average sensitivity, For the number of historical samples, To prevent positive constants with a denominator of zero.

[0045] Further, interventionability score The cross-process propagation score can be determined by the number of writable registers, adjustable range, response time, number of interlocks, and safety boundaries of the device. The ranking score can be determined by the spatiotemporal correlation strength and downstream quality sensitivity, and can be achieved by:

[0046] Among them, the comprehensive ranking score Used to normalize the priority of nodes in the candidate node set and select the top priority nodes accordingly. Each target process node has a fixed field contract, unit range, and register whitelist.

[0047] After the target process node is determined, multi-source data acquisition can be performed via SECS-GEM (SEMI Equipment Communications Standard / Generic Equipment Model) to collect process data from the equipment. Low-latency communication is carried out through industrial Ethernet field links. The acquired data can include four categories: process parameters, equipment status, quality inspection data, and environmental data, which are then encapsulated into a unified set of fields. For example, event messages can include a unique identifier for the carrier board. Batch Identification Equipment identification Process sequence code Channel signage The data includes a timestamp, event type, structured data body, and several validation fields.

[0048] In this embodiment, the units are uniformly and preferentially fixed in the field contract according to the engineering dimensions of the process specifications, so that the same field can be directly compared when reported by different devices. For example, the temperature field, pressure field, flow field, and processing time field are uniformly fixed in dimensions, and the unit identifier is fixed in the field contract to avoid caliber drift.

[0049] Furthermore, regarding robust standardization, the median during the stable period can be used for continuous parameters. Absolute deviation from median Construct a robust scale and combine it with a pruning function to suppress extreme values. For example, for the original features... Possible methods:

[0050] Among them, standardization features Used to express original features Relative to the median during the stable period The offset magnitude, The median absolute deviation during the stable period is given, and 1.4826 is the homogenization coefficient. It is a positive constant. To truncate the threshold, This is a truncation function.

[0051] Furthermore, in some embodiments of this application, in order to enable the model to explicitly distinguish between missing and zero values, a missing mask vector with the same dimension as the input features can be constructed. Missing mask vector The value can be 0 or 1, where a value of 1 indicates that the dimension is valid, and a value of 0 indicates that the dimension is missing.

[0052] For example, the plasma processing step can collect process parameters such as radio frequency power density, working gas pressure, gas flow rate, mixing ratio and processing time; the electroplating step can collect current density, voltage, electroplating time, bath temperature and chemical concentration; and the lamination step can collect temperature curve, vacuum degree, pressurization curve and holding time. After outlier removal, missing value marking and short window feature extraction are completed at the edge side, it enters the event-based encapsulation.

[0053] Thus far, step S1 has established a unified field contract and register whitelist through the three-dimensional scoring and screening of the target process nodes, and constructed a computable data foundation through robust standardization and missing masking, so that multi-source data can be stably expressed under a unified dimension and robust scale, providing a reliable input foundation for subsequent trusted encapsulation, spatiotemporal correlation modeling and model inference.

[0054] In step S2, the core task is to transform the structured data records formed in step S1 into verifiable, traceable, and quantifiable process event messages, and to form an event aggregation sequence that runs through the entire process using the unique identifier of the IC carrier board. This allows subsequent directed edge generation of heterogeneous spatiotemporal correlation graphs, out-of-distribution discrimination of the process basic model, and safe exit logic for collaborative parameter adjustment to all share the same confidence field and original evidence fragments. The input to this step is the structured data records of the target process node and equipment-side verification materials, and the output is a set of process event messages with a calculable confidence field and its event sequence aggregated by the unique identifier of the IC carrier board.

[0055] In this embodiment, a confidence field can be generated based on the results of time synchronization verification, message authentication digest, sequence consistency verification, and identity token verification, and then encapsulated as a process event message. The process event message is then aggregated according to the unique identifier of the IC carrier board.

[0056] Specifically, process event messages can adopt a fixed-field contract, which includes a unique IC carrier identifier, batch identifier, equipment identifier, process sequence code, timestamp, event type, structured data body, and evidence fields for verification. To ensure that the reported times from different devices are comparable under the same reference, time synchronization consistency can be verified, and the verification result can be provided. The deviation between the device's local time and the centrally aligned time can be written as a calculable quantity. This can be written as:

[0057] Among them, time deviation Used to express the device's local time Relative to center alignment time The offset magnitude and time synchronization consistency verification result can be determined by... The result is obtained by comparing it with the preset synchronization error threshold and writing the comparison result into the verification result field of the event message.

[0058] Furthermore, in some embodiments of this application, to prevent messages from being tampered with during transmission and aggregation, HMAC (Hash-based Message Authentication Code) can be used to perform message authentication digest verification on process event messages and provide the digest verification result. This digest can be calculated using a key bound to an identity token to deterministically serialize the message body. Possible methods include:

[0059] Among them, message authentication digest Used to characterize the message body Integrity protection, key Bound to the identity token and located by key index, As a deterministic serialization function, the digest verification result is obtained by comparing the recalculated digest on the receiving side with the digest carried in the message, and the comparison result is written into the event message field contract.

[0060] Meanwhile, to avoid message out-of-order delivery leading to distortion of the cross-process causal chain, it is necessary to verify the sequential consistency of process event messages and provide the sequential consistency verification result. Sequential consistency verification can be based on constraints formed by monotonically increasing sequence numbers and idempotent keys. The sequential consistency verification result is obtained by whether the sequence number satisfies monotonicity and whether replay detection is triggered. Furthermore, to ensure the credibility of the identity source, the identity token can also be verified and the token verification result provided. Token verification can include signature verification, validity period verification, and whitelist version consistency verification, and the token verification result is written into the event message field contract.

[0061] After the above four types of verification results are obtained, a computable confidence field is generated based on the time synchronization consistency verification result, digest verification result, sequence consistency verification result, and token verification result. This computable confidence field is then written into the process event message field contract. For example, to ensure the confidence field can be used for subsequent gating, weighting, and elimination scoring, the four types of verification results can be mapped to a unified trust level and then combined into a confidence field. This can be written as:

[0062] Among them, the confidence field Used to express the overall credibility of this process event message. The result of time synchronization consistency verification using the Sigmoid function. Summary verification results Sequence consistency check results And token verification result It can take 0 or 1, or a graded score, with weights. , , and This is used to express the contribution ratio of different verification items to the overall confidence level. For example, when the time synchronization consistency verification result is 0, it means that the time deviation exceeds the preset synchronization error threshold, and the confidence field will decrease significantly, thereby automatically suppressing the contribution of this event to the propagation path in subsequent confidence gating.

[0063] Furthermore, to enhance robustness against attacks and configuration drift during the generation of the confidence field, replay detection can be introduced into the sequence consistency check. When sequence number replay is detected, the confidence level of the token verification result can be adjusted, and this confidence level change, along with the replay evidence, can be written into the evidence package. Simultaneously, a key rotation version consistency check can be introduced into the message authentication digest check. When the key rotation version is inconsistent, the confidence level of the digest check result can be adjusted, and this confidence level change, along with the rotation version evidence, can be written into the evidence package. Similarly, a whitelist version consistency check can be introduced into the identity token check. When the whitelist version is inconsistent, the confidence level of the token verification result can be adjusted, and this confidence level change, along with the whitelist version evidence, can be written into the evidence package. Based on this, the aforementioned confidence level changes can be written into the original evidence fragment index of the evidence package, enabling subsequent causal localization and audit queries to directly locate the original evidence fragment where the confidence level change occurred.

[0064] In this embodiment of the application, after encapsulating a single process event message, it can be identified by the unique identifier of the IC carrier board. Aggregation is performed to form an event sequence with the unique identifier of the IC carrier as the primary key. Within the sequence, a stable order is established based on timestamps and sequence numbers, serving as the common input for subsequent event sequence encoding and the construction of heterogeneous spatiotemporal correlation maps. For example, the unique identifier of the same IC carrier... Process event messages generated during processes such as lamination, electroplating, and drilling can be aggregated into the same event sequence. The event sequence can retain the confidence field, summary verification result, and sequence verification result of each message for subsequent gating and evidence package reference.

[0065] Thus, step S2 forms a computable confidence field through time synchronization consistency verification, message authentication digest verification, sequence consistency verification, and identity token verification. The confidence field is then written into the process event message field contract and aggregated according to the IC carrier's unique identifier to form a stable event sequence. This constitutes a traceable and quantifiable event-based data system, providing a unified and reliable input foundation for subsequent graph modeling, anomaly identification, and security linkage.

[0066] In step S3, the core task is to represent processes, equipment, events, parameters, quality, and environment in the same heterogeneous spatiotemporal correlation graph, using the unique identifier of the IC carrier board as the root. Directed edges are generated using time-series relationships, influence weights, and confidence gating, enabling the set of anomaly propagation candidate paths to converge across the cross-process causal search space and simultaneously define the linkage adjustment boundary. The input to this step is the event sequence aggregated by the unique identifier of the IC carrier board and various correlation features; the output is the evidence fragment containing the heterogeneous spatiotemporal correlation graph, the set of anomaly propagation candidate paths, and their path weights.

[0067] In this embodiment, a heterogeneous spatiotemporal correlation graph can be constructed using the unique identifier of the IC carrier board as the root, and directed edges can be generated based on time series relationships, influence weights, and confidence gating, thereby constructing a set of candidate paths for anomaly propagation.

[0068] Specifically, heterogeneous node types can be categorized by process, equipment, event, parameter, quality, and environment, and each IC carrier board can be uniquely identified. Different associated node instances are written into the same graph instance, ensuring consistency in object granularity across the graph instances. Based on this, time-series related edges can be generated and lag features recorded. These lag features can express the time delay window of upstream parameter changes on downstream quality changes. Simultaneously, impact-related edges are generated based on impact weights, and the sources of these weights can include historical sensitivity estimates, process window boundaries, and quality inspection statistics. Finally, confidence gating is generated based on the confidence field and written into the directed edge generation rules, automatically suppressing the contribution of low-confidence field event messages to edge generation and edge weights.

[0069] Furthermore, to construct a set of candidate paths for anomaly propagation, similarity evidence, lag evidence, and gating evidence need to be unified into candidate edge scores and path scores. To this end, in some embodiments of this application, similarity candidate edges can be generated based on similarity features, forming a similarity score. These similarity features can be composed of the correlation of parameter trajectories, consistency of segmented slopes, and similarity of event types. Lag candidate edges can be generated based on lag features, forming a lag score. The lag score can be determined jointly by the lag window and the downstream response strength. Confidence gating is used to screen similarity candidate edges and lag candidate edges, forming a gating score, thus weakening or eliminating candidate edges with low confidence levels. Finally, the path weights of the candidate propagation paths are determined based on the similarity score, lag score, and gating score. An anomaly propagation candidate path set is formed according to the path weights, and the path weights are written into the evidence package. For example, the similarity score, lag score, and gating score can be fused in a weighted manner. This can be written as:

[0070] Among them, candidate edge weights Similarity score is used to quantify the degree to which the directed edge supports anomaly propagation. Lag score is used to express the similarity of parameter trajectories or event patterns. Gating score is used to express the consistency between timing delay and response strength. Used to express the gate retention strength generated by the confidence field, weighting coefficient , and This is used to balance the contribution proportions of the three types of evidence. The gating score can be obtained by combining the confidence field with a preset consistency threshold, so that the edge weights automatically decrease when the confidence field is insufficient. Possible methods include:

[0071] Among them, the confidence field From process event messages, preset consistency threshold Indicator function used to limit the minimum level of confidence required for participation in propagation reasoning. Used to express gating filters in a computable form.

[0072] Furthermore, a set of candidate paths for anomaly propagation can be constructed based on directed edge generation rules and used to converge the cross-process causal search space and linkage adjustment boundary. This ensures that subsequent structural causal chain learning is performed only in the subgraph induced by the candidate path set, thereby reducing spurious causality caused by irrelevant edges. For example, path weights can be obtained by multiplying or weighted summing the weights of each edge within the path, and the path weights and the weight source indexes of the edges within the path can be written into the evidence package to facilitate subsequent identification of "which edge is supported by which type of evidence".

[0073] Thus far, step S3 constructs a heterogeneous spatiotemporal correlation graph with the unique identifier of the IC carrier board as the root, and generates directed edges based on time series relationships, influence weights, and confidence gating. Then, path weights are formed using similarity scores, lag scores, and gating scores, and a set of candidate paths for anomaly propagation is constructed, thereby forming a propagation evidence system of "object granular graph modeling - gating propagation - interpretable path weights", which provides a reliable structural input basis for subsequent anomaly identification, causal localization, and linkage boundary limitation.

[0074] In step S4, the core task is to input the event sequence and heterogeneous spatiotemporal correlation map into the process basic model to form a unified representation. This unified representation then outputs a dynamic quantile threshold, anomaly type distribution, anomaly score, quality risk vector, key parameter trajectory prediction, and out-of-distribution discrimination results all at once. This allows anomaly identification and severity level classification to simultaneously utilize threshold information, score information, and out-of-distribution information, avoiding misjudgments caused by single threshold drift. The input to this step is the event sequence and map instance, and the output is the unified multi-task output result and anomaly identification conclusion.

[0075] In the embodiments of this application, dynamic quantile thresholds, anomaly type distributions, anomaly scores, quality risk vectors, key parameter trajectory predictions, and out-of-distribution discrimination results can be obtained based on the unified output of the process basic model, and anomaly identification can be completed.

[0076] Specifically, the event sequence of process event messages can be encoded to obtain an event sequence representation. The event sequence encoding can employ a multi-layer Transformer encoder, which incorporates event type embedding, time position encoding, and the standardized features obtained in step S1 as input. Furthermore, to improve robustness to missing data and noise, the missing data mask vector can be used as part of the attention mask, preventing the missing dimension from participating in attention normalization.

[0077] Based on this, relation-aware encoding can be performed on heterogeneous spatiotemporal correlation graphs to obtain graph representations. Graph encoding can use a relation-aware graph Transformer, which takes node type embedding, edge type embedding and candidate edge weights as input, and injects the confidence field into the edge attention coefficient through gating score, so that the contribution of low confidence edges to graph representation is suppressed.

[0078] Furthermore, the event sequence representation and the graph representation are fused to form a unified representation. To ensure that the unified representation is interpretable and controllable, a fusion gating coefficient can be used to weight the two types of representations. This can be written as:

[0079] Among them, unified representation A unified representation for carrying out process states, event sequence characterization Used to represent the dynamics of events ordered by time, graphical representation Used to express cross-process structural relationships, incorporating gating coefficients. Used to balance the contribution ratio of the two types of information, and can be determined by and The result is obtained after splicing and output through a fully connected layer.

[0080] Based on this, a unified representation can be used to output dynamic quantile thresholds, anomaly type distributions, anomaly scores, quality risk vectors, key parameter trajectory predictions, and out-of-distribution discrimination results. Specifically, the dynamic quantile threshold can be output from the unified representation via a quantile regression head, the anomaly type distribution can be output from a Softmax classification head, the anomaly score can be output from a regression head, the quality risk vector can be output from a risk embedding head, and the key parameter trajectory prediction can be output from a sequence prediction head. For example, to ensure stable multi-task training and avoid manual weighting, an uncertainty-weighted multi-task loss can be used. Possible methods include:

[0081] Among them, multi-task weight Used for automatic adjustment of the The contribution of each task's loss to the total loss, and the uncertainty parameters. For learnable parameters, total loss across multiple tasks Used for training the basic process model, task loss These can be respectively associated with quantile regression loss, classification cross-entropy loss, rating regression loss, trajectory prediction loss, and out-of-distribution discrimination auxiliary loss.

[0082] Furthermore, the out-of-distribution discrimination result can be obtained by forming a distance score from the unified representation and the training distribution statistics, and then comparing the distance score with a preset out-of-distribution threshold to obtain the out-of-distribution discrimination result. This can be written as:

[0083] Among them, out-of-distribution scores Used for quantifying unified representation Relative training distribution center The degree of deviation, covariance matrix Used to express the scale and correlation structure of the training distribution, the out-of-distribution discrimination result is... The result is obtained by comparing it with the preset distribution threshold and used as one of the trigger inputs for subsequent freeze linkage or minimum action mode.

[0084] Based on this, dynamic quantile thresholds, anomaly scores, and out-of-distribution discrimination results can be used together for anomaly identification and severity classification. For example, dynamic quantile thresholds can be used as the "exceeding threshold" criterion, anomaly scores as the "intensity" criterion, and out-of-distribution discrimination results as the "confidence region" criterion. When the dynamic quantile threshold is triggered and the anomaly score exceeds a preset severity threshold while the out-of-distribution discrimination result is negative, it is determined to be a high-severity anomaly. When the out-of-distribution discrimination result is positive, a conservative strategy is prioritized, and the anomaly type distribution is used as an auxiliary indicator.

[0085] Thus, step S4 forms a unified representation using event sequence encoding and relation-aware graph encoding. This unified representation outputs dynamic quantile thresholds, anomaly type distributions, anomaly scores, quality risk vectors, key parameter trajectory predictions, and out-of-distribution discrimination results all at once. The dynamic quantile thresholds, anomaly scores, and out-of-distribution discrimination results are then combined for anomaly identification and severity level classification, thereby forming an anomaly identification system of "unified representation - multi-task unified output - three-criteria joint identification," providing a reliable input basis for subsequent causal localization and security linkage.

[0086] In step S5, the core task is to transform the set of candidate propagation paths for anomaly propagation into prior constraints for learning structural causal chains. This involves determining the set of variables in the induced subgraph of the candidate path set and learning the directed dependencies between variables. Simultaneously, acyclic constraints are introduced to ensure the legitimacy of the causal structure. Then, counterfactual intervention is performed on the root cause variables to calculate the counterfactual effect, thereby outputting the set of operable root causes, lag windows, evidence packages, and linkage sequences, enabling subsequent linkage actions to generate interpretable causal evidence. The inputs to this step are the set of candidate propagation paths for anomaly propagation, graph instances, and anomaly identification results; the outputs are the set of operable root causes, lag windows, evidence packages, and linkage sequences.

[0087] In the embodiments of this application, structural causal chain localization and counterfactual effects can be performed under the constraint of the set of candidate paths for abnormal propagation, and the set of intrusive root causes, lag window, evidence package and linkage sequence can be output.

[0088] Specifically, a set of variables can be determined on the subgraph induced by the set of anomaly propagation candidate paths. This set can include key process parameters, key equipment status variables, key environmental variables, and target quality variables. A time lag window is set for each variable to match the lag features recorded in step S3. Simultaneously, a priori mask matrix is ​​generated based on the anomaly propagation candidate path set and used to restrict the candidate edges of the variable set, ensuring that only edges appearing in the candidate path set are allowed to enter the learning process. This can be written as:

[0089] Among them, the prior mask matrix Used to express the allowed relations for edge candidates, the set of candidate paths. From the set of candidate propagation paths for anomalies, Representing variables to variable Directed candidate edges, indicator function Used to write "whether it appears in the candidate path set" as a computable mask.

[0090] Building upon this, directed dependencies between variables can be learned under edge candidate constraints, and acyclic constraints can be introduced. For example, the causal structure can be represented as a weighted adjacency matrix. ,in Representing variables to variable The causal strength is determined, and the strength of disallowed edges is fixed to 0 using a priori masking matrices. To ensure the structure is acyclic, an acyclic constraint loss can be introduced. Possible approaches include:

[0091] Among them, the loss without a cycle constraint Used to punish directed cycles in causal structures, matrix trace Used to extract the sum of the diagonal elements and dimension of a matrix. Hadamard product is the size of the variable set. The squared treatment of the edge strength is used to ensure non-negativity. The acyclic constraint loss is optimized together with the data fitting loss during training, thereby learning the directed dependencies that satisfy the acyclic structure.

[0092] Furthermore, the counterfactual effect of root cause variables on the target quality variable can be calculated based on counterfactual intervention, and an intrusive root cause set can be formed according to the counterfactual effect. For example, intervention values ​​are applied to candidate root cause variables, and the counterfactual results of the target quality variable are obtained through structural equation modeling, then compared with the results without intervention to form the counterfactual effect. This can be written as:

[0093] Among them, the counterfactual effect Used to measure the root dependent variable Apply intervention value Post-target quality variables Expected changes The expected value of the target quality variable without intervention. To determine the expected outcome after intervention, the set of interventionable root causes is obtained by ranking the absolute values ​​of counterfactual effects, and the path weights of the set of interventionable root causes and the set of candidate paths for abnormal propagation are correlated and written into the evidence package.

[0094] Furthermore, the set of interventionable root causes and the set of candidate propagation paths for abnormalities are associated and written into the evidence package, and the linkage order and lag window are output. The linkage order can be determined according to the topological order of the causal structure, so that the upstream interventionable root causes are adjusted first and the lag window is considered, making the adjustment action observable within the response window of the target quality variable; the lag window can be obtained by statistical analysis of the lag characteristics recorded in step S3, and the lag window is written into the evidence package to constrain the execution time window of subsequent collaborative parameter adjustment instructions.

[0095] Thus far, step S5 forms a priori mask matrix and limits the candidate edges under the constraint of the candidate path set for anomaly propagation. Then, under the constraint of the candidate edges, it learns the acyclic causal structure and calculates the counterfactual effect, thereby outputting the set of interventionable root causes, lag window, evidence package and linkage order, forming a causal decision system of "candidate path prior - acyclic causal structure - counterfactual interventionable root causes", providing interpretable basis and time boundary conditions for subsequent adjustment of collaborative parameters.

[0096] In step S6, the core task is to map the set of interventionable root causes into coordinated parameter adjustment instructions, and to sequentially project candidate action vectors onto equipment hard boundary constraints, ramp rate constraints, and process window constraints to obtain executable action vectors. Simultaneously, based on the confidence field and out-of-distribution discrimination results, a freeze linkage or minimum action mode is triggered, and the constraint trigger reason code, mapping residual, exit score, and execution receipt consistency verification results are written into the audit information, ensuring that the linkage actions are both executable and traceable. The inputs to this step are the set of interventionable root causes, the hysteresis window, and the model output; the outputs are executable action vectors and audit information.

[0097] In this embodiment, collaborative parameter adjustment instructions can be generated based on the set of interventionable root causes, and executable mapping of equipment hard boundary constraints, ramp rate constraints, and process window constraints can be executed. Furthermore, based on the confidence field and out-of-distribution discrimination results, freeze linkage or minimum action mode can be triggered and audit information can be recorded.

[0098] Specifically, the target process node and target adjustment parameter set can be determined based on the set of interventionable root causes. The target adjustment parameter set can be obtained by mapping the root cause variable to the writable register of the equipment, and the mapping relationship is written into the evidence package to ensure the traceability of the action. Then, the collaborative parameter adjustment instruction is generated and written into the candidate action vector, execution time window and idempotent key, so that the instruction with the same idempotent key will not cause repeated execution when it is repeatedly issued. Further, the equipment hard boundary constraint mapping, ramp rate constraint mapping and process window constraint mapping are performed on the candidate action vector to obtain the executable action vector and write it into the collaborative parameter adjustment instruction. During each constraint mapping, the constraint trigger cause code and mapping residual are generated and written into the audit information.

[0099] For example, to write the constraint mapping as a computable procedure, the candidate action vector can be denoted as the candidate action vector. The hard boundary projection of the equipment is denoted as the hard boundary projection. Let the slope rate projection be denoted as the slope rate projection. The process window projection is denoted as the process window projection. Then the executable action vector can be expressed as a stepwise projection. It can be written as:

[0100] Among them, executable action vectors The three projections are used to express the final action issued after all constraints are satisfied. Hard boundary projection trims the action components to the upper and lower limits allowed by the equipment. Ramp rate projection limits the rate of change of the action to a preset ramp rate threshold. Process window projection restricts the action to the process window constraint set. The order of these three projections ensures that hard constraints are satisfied first, followed by dynamic constraints, and then process constraints. Based on this, the mapping residual can be defined as the difference between the candidate action vector and the executable action vector. This residual quantifies the magnitude of the constraint correction on the action, and is written into the audit information for subsequent cause analysis.

[0101] Furthermore, the freeze linkage mode or minimum action mode can be determined based on the confidence field and the out-of-distribution discrimination result, and a coordination parameter adjustment instruction can be written, along with the exit score as audit information. For example, to write the freeze linkage mode as a computable trigger condition, the freeze flag can be written as an indicator function. This can be achieved using:

[0102] Among them, the freeze sign Confidence field used to indicate whether to enter freeze linkage mode. Used to express the credibility of an event, with a preset consistency threshold. Used to define the minimum level of confidence, out-of-distribution score Used to express the degree to which a uniform representation deviates from the training distribution, with a preset out-of-distribution threshold. Used to limit out-of-distribution risks; when the freeze flag is 1, the collaborative parameter adjustment instruction enters the freeze linkage mode and only records audit information without executing any actions. The minimum action mode can be triggered when the freeze flag is 0 but the exit score is high, causing the executable action vector to be scaled to the minimum feasible range.

[0103] Furthermore, the exit score is used to uniformly quantify the combined risks of missing data, communication, interlocking, out-of-distribution, and low confidence. This can be written as:

[0104] Among them, the elimination score Used to express the degree of risk in coordinated execution, and the proportion of missing data. Based on missing mask statistics, the communication failure rate Used to express the degree of communication anomaly and the proportion of interlock risk. Used to express the degree of interlocking or protection status triggering, out-of-distribution scoring. Used to express out-of-distribution risk, confidence score field. Used to express the degree of credibility, weighting coefficient to This is used to balance the contributions of various risk factors. For example, when the proportion of interlocked risk is high, even if the confidence field is high, it will lead to an increase in the exit score, thereby triggering the minimum action mode or the direct freeze linkage mode.

[0105] Finally, a collaborative parameter adjustment command can be issued and an execution receipt can be obtained. The execution receipt is then validated for consistency based on an idempotent key, and the validation result is written into the audit information. The execution receipt may include the execution time, the actual effective action, the device-side confirmation code, and the failure reason code. The failure reason code and the constraint trigger reason code are jointly written into the audit information to form a closed-loop evidence chain. For example, the execution time window can be determined by the lag window output in step S5, allowing the action's effectiveness and quality response to be verified within the same observation window, and the verification result is archived as part of the audit information.

[0106] At this point, step S6 generates collaborative parameter adjustment instructions based on the set of interventionable root causes, and obtains executable action vectors through the stepwise projection of equipment hard boundary constraints, ramp rate constraints, and process window constraints. Then, based on the confidence field and the out-of-distribution discrimination results, it triggers the freeze linkage or minimum action mode and records audit information, thereby forming a safety linkage system of "causal interventionable root causes - constraint executable actions - risk exit protection". This system ensures that the linkage is both engineering executable and traceable throughout the process, and provides a reliable audit data foundation for subsequent continuous optimization.

[0107] In summary, this method achieves reliable convergence of cross-process heterogeneous spatiotemporal correlation graphs and anomaly propagation candidate paths by robustly standardizing and maintaining consistent standards for multi-source process data of IC substrate production lines. It combines time synchronization, message authentication, sequence consistency, and identity token verification to generate computable confidence fields and integrates them with event aggregation, graph modeling, and linkage control. Based on a unified representation of the process foundation model, it outputs dynamic quantile thresholds, anomaly scores, quality risk vectors, key parameter trajectory predictions, and out-of-distribution discrimination results. Furthermore, under candidate path constraints, it completes structural causal chain localization and counterfactual effect assessment. It can provide an interpretable evidence package to provide an actionable root cause set, lag window, and linkage sequence. Under hard boundary, ramp rate, and process window constraints, it generates executable collaborative adjustment instructions. Combined with frozen linkage and minimum action mode, it reduces misadjustments and risk spillover, minimizes manual investigation and downtime costs, improves anomaly identification accuracy, root cause determination reliability, and linkage adjustment security, thereby enhancing the quality stability and yield improvement sustainability of IC substrate production lines.

[0108] It should be noted that, although the embodiments in this application are based on... Figure 1 Steps S1 to S6 are described sequentially, but this does not mean that steps S1 to S6 must be performed in a strict order. The reason this embodiment follows this order is... Figure 1The order in which steps S1 to S6 are described is provided to facilitate understanding of the technical solutions of the embodiments of this application by those skilled in the art. In other words, in the embodiments of this application, the order of steps S1 to S6 can be appropriately adjusted according to actual needs.

[0109] In some embodiments of this application, a method for processing process parameters of an IC substrate production line is applied to an IC substrate production line that includes processes such as laser drilling, electroplating, lamination, plasma treatment, and surface treatment. The production substrates include two batches: ABF and BT. Quality inspection covers indicators such as aperture, plating thickness, line width and spacing, and surface roughness. This method operates collaboratively on both the edge and center sides, completing a full implementation process from data acquisition, anomaly identification, root cause localization, to collaborative adjustment through process event messages and event aggregation sequences with the IC substrate's unique identifier as the primary key.

[0110] The complete implementation process may include the following steps: Step 1: Determine the target process node and solidify the field contract and data collection criteria.

[0111] In this embodiment, candidate nodes are first screened across the entire production line. The candidate set includes five core process nodes: plasma treatment, electroplating for hole filling, laser drilling, lamination, and surface treatment. For any candidate node... Calculate the quality contribution score and combine it with the interventionability score and cross-process propagation score to form a comprehensive ranking score. This can be written as:

[0112]

[0113] Among them, contribution score Used for quantizing candidate nodes Key parameter deviation magnitude For the magnitude of quality deviation Average sensitivity, number of historical samples Take the last two months 100 event samples, positive constants Pick Used to avoid a denominator of zero; comprehensive sorting of scores Used for normalizing and ranking candidate node priorities, and for interventional scoring. The propagation score is derived from the number of writable registers, the adjustable range, and the number of interlocks. It is obtained from the spatiotemporal correlation strength and downstream quality sensitivity. This embodiment selects the previous... Each target process node is assigned a specific field, and for each type of node, a field contract, unit range, slot definition, and register whitelist are fixed, so that subsequent collected process parameters, equipment status, quality inspection, and environmental data can be aligned under the same field system.

[0114] Step 2: Collect data from multiple sources and complete unit unification and robust standardization to form a missing mask.

[0115] In this embodiment, a unique identifier is set for each IC carrier board. Four types of data are collected at the target process nodes: process parameters, equipment status, quality inspection data, and environmental data. For example, data collected for the plasma treatment process include RF power density of 120W, working gas pressure of 2.6kPa, argon flow rate of 35sccm, oxygen flow rate of 5sccm, and processing time of 90s; data collected for the electroplating and hole-filling process includes current density. Bath temperature Electroplating time 780s; additive concentration calculated as a dimensionless ratio of 0.85 according to the formula; vacuum degree -92kPa and heating rate during the lamination process. Holding pressure: 1.25 MPa; Holding time: 600 s; Environmental data acquisition temperature. With a relative humidity of 42%.

[0116] Perform robust normalization on continuous numeric fields and prune extreme values. This can be done by:

[0117] Among them, standardization features Used to express original features Median during relative stability period The offset magnitude, The truncation threshold is the median absolute deviation during the steady-state period. A value of 5.0 is used to suppress extreme outliers; a positive constant. Consistent with step one, construct a missing mask vector for the missing fields. Missing mask vector The component has the same dimension as the input feature and is either 0 or 1, where 1 indicates that the dimension is valid and 0 indicates that the dimension is missing. In this embodiment, if the additive concentration sensor of a certain electroplating equipment does not report within 20 seconds, the missing component of the corresponding dimension is set to 0, and the missing semantics are preserved in the subsequent model input instead of being replaced by a zero value.

[0118] Step 3: Generate process event messages and complete time synchronization verification, message authentication digest verification, sequence consistency verification, and identity token verification.

[0119] In this embodiment, each data acquisition record is encapsulated as a process event message, and the message fields include at least... The data includes batch identifier, equipment identifier, process sequence code, timestamp, event type, structured data body, and validation fields. It verifies time synchronization consistency and provides the result. Time deviation can be expressed as: Among them, time deviation Used to express the device's local time Relative center alignment time The offset magnitude, in this embodiment, uses PTP alignment and sets a preset synchronization error threshold of 2ms, when The time synchronization consistency check result is deemed passed.

[0120] Perform message authentication digest verification on process event messages and provide the digest verification result. The digest calculation can be performed using:

[0121] Among them, message authentication digest Used to characterize the message body Integrity protection, key Bound to the identity token and located by key index, This is a deterministic serialization function. It verifies the sequential consistency of process event messages and provides the result. This embodiment employs a dual constraint of monotonically increasing sequence numbers and idempotent keys, and enables replay detection to identify duplicate sequence numbers. It also verifies the identity token and provides the verification result. This embodiment enables signature verification, validity period verification, and whitelist version consistency verification, and writes each verification result into the message field contract for subsequent evidence association.

[0122] Step four involves generating a computable confidence field based on the four types of verification results and writing it into the event field contract. Simultaneously, the event sequence is aggregated according to the unique identifier of the IC carrier board. In this embodiment, the four types of verification results are mapped to hierarchical scores and combined to obtain the confidence field. It can be written as:

[0123] Among them, the confidence field Used to express the overall credibility of this process event message. The score for time synchronization consistency verification using the Sigmoid function. Abstract verification result scoring Sequence consistency check result score Token verification result scoring Take a value between 0 and 1; in this embodiment, take... , , and If serial number replay is detected, the credibility level of the token verification result is downgraded and the change in credibility level is written to the original evidence fragment index of the evidence package; if key rotation version inconsistency is detected, the credibility level of the digest verification result is downgraded and written to the evidence package index; if whitelist version inconsistency is detected, the credibility level of the token verification result is downgraded and written to the evidence package index.

[0124] Subsequently, the same Process event messages generated at each stage are aggregated into event sequences based on timestamps and sequence numbers, forming a common input for subsequent event sequence encoding and graph construction. In this embodiment, the event sequence length for a specific IC carrier board within 8 minutes is 240 messages, and each message in the sequence carries a confidence field. To support subsequent gating and elimination decisions.

[0125] Step 5: Construct a heterogeneous spatiotemporal correlation graph using the unique identifier of the IC carrier board as the root and generate a set of candidate paths for anomaly propagation. In this embodiment, processes, equipment, events, parameters, quality, and environment are used as the set of heterogeneous node types, and are integrated within the same... A graph construction example at the granular level. Time-series related edges are generated based on time-series relationships, and lag features are recorded. Influence-related edges are generated based on influence weights, and the source of those weights is recorded. Confidence gating is generated based on the confidence field and written into the directed edge generation rules. Edge weights are calculated for candidate edges and path weights are formed at the candidate path level. This can be written as:

[0126]

[0127] Among them, candidate edge weights Similarity score is used to quantify the degree to which the directed edge supports anomaly propagation. Lag score is used to express the similarity of parameter trajectories or event patterns. Gating score is used to express the consistency between timing delay and response strength. Used to express the confidence level field The resulting gating retention strength; preset consistency threshold Take 0.65; in this embodiment, take , and Candidate edges with weights below 0.30 are eliminated, thus forming a set of candidate paths for anomaly propagation. Furthermore, the path weights and weight source indices are written into the evidence package to provide structural priors for subsequent convergence of the causal search space and constraint of linkage boundaries.

[0128] Step six involves constructing and training a basic process model to generate a unified representation and output dynamic quantile thresholds, anomaly type distributions, anomaly scores, quality risk vectors, key parameter trajectory predictions, and out-of-distribution discrimination results. In this embodiment, the basic process model employs a dual-encoder cascade structure, comprising an event sequence Transformer encoder and a relation-aware heterogeneous graph Transformer encoder, and outputs a unified representation at the fusion layer. The event sequence Transformer encoder contains 6 coding layers, each with a hidden dimension of 256, 8 attention heads, and a feedforward layer with a dimension of 1024. The input consists of field embeddings and temporal location encodings of the event sequence. The field embeddings are composed of normalized features, discrete event types, and device identifiers. The relation-aware heterogeneous graph Transformer encoder contains 4 coding layers, each with a hidden dimension of 256, 8 attention heads, and a relation type embedding dimension of 64. The input consists of node type embeddings and edge weights, with gating scores injected into the attention layer to suppress low-confidence edges. The fusion layer uses gating fusion to form a unified representation. This can be written as:

[0129] Among them, unified representation A unified representation for carrying out process states, event sequence characterization Expressing temporal dynamics, graphical representation Expressing cross-process structural relationships and integrating gating coefficients The output is formed by splicing two representations and passing them through a fully connected layer, with the value limited to between 0 and 1.

[0130] The multi-task output head includes a quantile regression head, an anomaly type classification head, an anomaly score regression head, a quality risk vector output head, a trajectory prediction head, and an out-of-distribution discriminant head. This embodiment selects quantile levels. Output dynamic quantile thresholds; classify anomalies into 6 categories and output anomaly type distributions; the quality risk vector dimension is 16; the key parameter trajectory prediction covers 12 key parameters and the prediction window is the next 60 seconds.

[0131] The training data is taken from the production data of the most recent 60 days, and is categorized as follows: Samples were constructed, with the training and validation sets divided into batches of 8:2. Data augmentation included masking and noise augmentation: masking augmentation randomly masked the input feature dimensions with a probability of 0.10 and simultaneously updated the missing mask vector; noise augmentation injected Gaussian noise with a mean of 0 and a standard deviation of 0.02 into continuous features while maintaining consistent unit units. Training employed the AdamW optimizer with a learning rate of [missing information]. The weight decay is 0.01, batch size is 64, and training epochs are 30. To avoid manual weight adjustment, an uncertain weighted multi-task loss is used. Possible methods include:

[0132]

[0133] Among them, multi-task weight Used for automatic adjustment of the The contribution of each task's loss, and the uncertainty parameter. For learnable parameters, total loss across multiple tasks Used for training the basic process model, task loss These correspond to quantile regression loss, classification cross-entropy loss, score regression loss, risk vector consistency loss, trajectory prediction loss, and out-of-distribution discrimination auxiliary loss, respectively.

[0134] The out-of-distribution discrimination result is obtained by comparing a distance score generated from a unified representation and training distribution statistics with a preset out-of-distribution threshold. This can be written as:

[0135] Among them, out-of-distribution scores Used for quantifying unified representation Relative training distribution center The degree of deviation, covariance matrix Used to express the training distribution scale and correlation structure; in this embodiment, a preset out-of-distribution threshold of 35.0 is set, and when the out-of-distribution score exceeds this threshold, the out-of-distribution discrimination result is determined to be yes.

[0136] In an online inference, an IC substrate exhibited a deviation in the mean plating thickness during the electroplating process. The process base model output a dynamic quantile threshold of 3.1 at the 0.95 quantile, an anomaly score of 0.82, and a probability of the anomaly type falling under the "plating thickness anomaly" category of 0.74. The quality risk vector significantly increased in the 3rd and 9th dimensions. Trajectory prediction indicated an upward trend in current density within the next 60 seconds. The out-of-distribution score was 18.6, and the out-of-distribution discrimination result was negative, thus triggering the subsequent causal localization process.

[0137] Step seven involves learning the structural causal chain and calculating the counterfactual effect under the constraints of the candidate path set for abnormal propagation, outputting the set of interventionable root causes, lag window, evidence package, and linkage sequence. In this embodiment, within the candidate path set... A set of variables is determined on the induced subgraph, including electroplating current density, bath temperature, additive ratio, plasma power density, lamination vacuum, ambient humidity, and the target quality variable, coating thickness, etc., and a hysteresis window is set for each variable. A priori mask matrix is ​​generated based on the candidate path set, and edge candidates are restricted. This can be written as:

[0138] Among them, the prior mask matrix This is used to express the allowed edge candidate relationship. Then, under the edge candidate constraint, directed dependencies between variables are learned, and acyclic constraints are introduced. This embodiment uses a weighted adjacency matrix. The loss is acyclic and unconstrained. It can be achieved by:

[0139] Among them, the loss without a cycle constraint Used to penalize directed cycles, variable dimension Size of the variable set.

[0140] After completing the structure learning, counterfactual interventions are performed on the candidate root dependent variables, and the counterfactual effect on the target quality variable is calculated. This can be written as:

[0141] Among them, the counterfactual effect Used to measure the root dependent variable Apply intervention value Post-target quality variables The desired change. This embodiment applies an intervention value to the electroplating current density variable. The counterfactual effect of coating thickness deviation was found to be -0.38; an intervention value was applied to the bath temperature variable. The counterfactual effect was -0.21; an intervention value was applied to the additive ratio. The counterfactual effect was obtained as -0.17, thus forming an intrusive root cause set. The output hysteresis window was 90s, and the linkage sequence was "current density → bath temperature → additive ratio". At the same time, the candidate path weights, prior mask matrix indexes and counterfactual effect calculation fragments were written into the evidence package.

[0142] Step eight involves generating collaborative parameter adjustment instructions and executing the executable mapping of equipment hard boundary constraints, ramp rate constraints, and process window constraints. Simultaneously, based on the confidence field and out-of-distribution discrimination results, a freeze linkage or minimum action mode is triggered, and audit information is recorded. In this embodiment, the target process node is determined to be the electroplating process based on the set of interventionable root causes, and the target adjustment parameter set is determined to be current density, bath temperature, and additive ratio. Candidate action vectors are then generated. And write the execution time window and idempotency key. Candidate action vectors Pick , respectively, indicate that the current density is reduced. Adjusting the temperature of the bath solution The additive ratio is reduced by 0.05. The candidate action vectors are then subjected to hard boundary projection, ramp rate projection, and process window projection in sequence to obtain the executable action vectors. This can be written as:

[0143] Among them, executable action vectors Used to express the final action after all constraints are satisfied, hard boundary projection is used to trim the action to the upper and lower limits allowed by the device. The allowable current density range in this embodiment is... The ramp rate projection is used to limit the rate of change of motion within a preset ramp rate threshold. In this embodiment, the current density ramp rate threshold is: The process window projection is used to restrict actions within the process window constraint set. In this embodiment, the additive ratio process window is [0.70, 0.90]. The mapping residual is defined as the difference between the candidate action vector and the executable action vector, and a constraint trigger reason code and the mapping residual are generated and written into the audit information during each projection.

[0144] Then, based on the confidence field and out-of-distribution score, a decision is made on whether to freeze the linkage or the minimum action mode. Possible methods include:

[0145] Among them, the freeze sign Confidence field used to indicate whether to enter freeze linkage mode. To determine the credibility of an event, a consensus threshold is preset. Consistent with step five, out-of-distribution scoring To quantify out-of-distribution risk, a pre-defined out-of-distribution threshold is used. Consistent with step six. In this embodiment, the confidence field of the electroplating key event... Out-of-distribution score The freeze flag is set to 0, thus allowing the execution of linked actions. Simultaneously, an elimination score is calculated and used to switch to the minimum action mode when necessary. This can be written as:

[0146] Among them, the elimination score Used to quantify the risk of coordinated execution, missing proportion The communication failure rate was obtained from statistics on missing masks. The interlock risk ratio is obtained from link retransmission and timeout statistics. The results are obtained from statistics on interlock status and protection status; this embodiment takes... , , , as well as The elimination score threshold is set to 3.0. This embodiment calculates a missing rate of 0.02, a communication failure rate of 0.00, an interlock risk rate of 0.10, and an elimination score of 1.02, thereby avoiding triggering the minimum action mode and executing the linkage with an executable action vector.

[0147] A collaborative parameter adjustment command is issued and an execution receipt is obtained. The execution receipt includes the execution time, the actual effective action, the device confirmation code, and the failure reason code. Consistency verification is performed on the execution receipt based on an idempotent key, and the verification result is written into the audit information. Within 15 minutes of execution, quality inspection shows that the coating thickness deviation decreased from +0.45 to +0.12, the anomaly score decreased from 0.82 to 0.34, the quality risk vector decreased, and the out-of-distribution score remained within the threshold, forming a complete closed loop.

[0148] Through the complete implementation process described above, this method can govern multi-source data from IC substrate production lines using a unified field contract and robust scale. It uses a computable confidence field to integrate event aggregation, graph modeling, and linkage control. Under the constraints of propagation candidate paths, it achieves interpretable causal localization and counterfactual effect assessment. This results in the output of an actionable root cause set and linkage sequence, and the generation of executable adjustment instructions under hard boundary, ramp rate, and process window constraints. Combined with out-of-distribution discrimination and exit scoring, it achieves safe degradation of frozen linkage and minimum action mode, thereby reducing misjudgment and misadjustment and manual investigation costs, improving the accuracy of anomaly identification, the reliability of root cause localization, and the safety of collaborative adjustment, and promoting stable yield improvement and continuous controllable optimization of the production process.

[0149] It should be understood that the step numbers identified by "Step 1, Step 2" and other similar forms in the above embodiments are only used to distinguish different steps and do not limit the steps to be executed in the order of these numbers. The specific execution order of each step can be adjusted according to its functional requirements and the inherent logic in the actual application scenario. The above step numbers should not be interpreted as a limitation on the implementation process of the embodiments of this application.

[0150] like Figure 2 As shown, the following is an embodiment of an IC substrate production line process parameter data processing system provided by this disclosure. This IC substrate production line process parameter data processing system and the IC substrate production line process parameter data processing methods of the above embodiments belong to the same inventive concept. For details not described in detail in the embodiments of the IC substrate production line process parameter data processing system, please refer to the embodiments of the above IC substrate production line process parameter data processing methods.

[0151] Based on the same concept, another embodiment of this application provides an IC substrate production line process parameter data processing system, including: Data acquisition standardization unit 1 is used to acquire process parameters, equipment status, quality inspection and environmental data of target process nodes in the IC substrate production line and complete unit unification and robust standardization; Event Trust Unit 2 is used to generate a confidence field based on the results of time synchronization verification, message authentication digest, sequence consistency verification and identity token verification, and encapsulate it into a process event message, and aggregate the process event message according to the unique identifier of the IC carrier board. The graph reasoning unit 3 is used to construct a heterogeneous spatiotemporal correlation graph with the unique identifier of the IC carrier board as the root and generate directed edges based on time series relationships, influence weights and confidence gating. It constructs a set of candidate paths for anomaly propagation and obtains dynamic quantile thresholds, anomaly type distribution, anomaly scores, quality risk vectors, key parameter trajectory prediction and out-of-distribution discrimination results based on the unified output of the process basic model, and completes anomaly identification. Causal decision unit 4 is used to locate the structural causal chain and calculate the counterfactual effect under the constraint of the set of candidate paths for abnormal propagation, and outputs the set of interventionable root causes, lag window, evidence package and linkage order. The linkage audit unit 5 is used to generate collaborative parameter adjustment instructions based on the set of interventionable root causes and execute executable mappings of equipment hard boundary constraints, ramp rate constraints and process window constraints. It also triggers freeze linkage or minimum action mode based on confidence field and out-of-distribution discrimination results and records audit information.

[0152] In some embodiments of this application, the causal decision unit 4 includes: Subgraph-induced subunits are used to determine the set of variables on a subgraph induced by the set of anomaly propagation candidate paths; The mask constraint subunit is used to generate a priori mask matrix based on the set of anomaly propagation candidate paths and to restrict the edge candidates of the variable set using the priori mask matrix. Under the edge candidate constraint, the directed dependency between variables is learned and an acyclic constraint is introduced. The counterfactual assessment subunit is used to calculate the counterfactual effect of root variables on target quality variables based on counterfactual intervention and to form an interventionable root cause set according to the counterfactual effect; The evidence arrangement subunit is used to associate the set of interventionable root causes with the set of abnormal propagation candidate paths, write them into the evidence package, and output the linkage order and lag window.

[0153] The above-disclosed embodiments are merely preferred embodiments of the present invention, but the present invention is not limited thereto. Any non-creative variations that can be conceived by those skilled in the art, as well as any improvements and modifications made without departing from the principles of the present invention, should fall within the protection scope of the present invention.

Claims

1. A method for processing process parameter data in an IC substrate production line, characterized in that, include: Acquire process parameters, equipment status, quality inspection and environmental data for the target process nodes in the IC substrate production line and complete unit unification and robust standardization; Based on the results of time synchronization verification, message authentication digest, sequence consistency verification and identity token verification, a confidence field is generated and encapsulated into a process event message. The process event message is aggregated according to the unique identifier of the IC carrier board. A heterogeneous spatiotemporal correlation graph is constructed using the unique identifier of the IC carrier board as the root, and directed edges are generated based on time series relationships, influence weights, and confidence gating to construct a set of candidate paths for anomaly propagation. Based on the unified output of the process fundamental model, dynamic quantile thresholds, anomaly type distribution, anomaly scores, quality risk vectors, key parameter trajectory predictions, and out-of-distribution discrimination results are obtained and anomaly identification is completed. Under the constraint of the set of candidate paths for abnormal propagation, structural causal chain localization is performed and counterfactual effects are calculated. The output includes the set of interventionable root causes, lag window, evidence package and linkage sequence. Based on the set of interventionable root causes, generate collaborative parameter adjustment instructions and execute executable mappings of equipment hard boundary constraints, ramp rate constraints, and process window constraints. Based on the confidence field and out-of-distribution discrimination results, trigger freeze linkage or minimum action mode and record audit information.

2. The IC substrate production line process parameter data processing method as described in claim 1, characterized in that, The steps involved in generating a confidence field and encapsulating it into a process event message based on the results of time synchronization verification, message authentication digest, sequence consistency verification, and identity token verification include: The system verifies the consistency of time synchronization and provides the time synchronization consistency verification result; it verifies the message authentication digest of process event messages and provides the digest verification result; it verifies the sequential consistency of process event messages and provides the sequential consistency verification result; and it verifies the identity token and provides the token verification result. A computable confidence field is generated based on the time synchronization consistency verification result, digest verification result, sequence consistency verification result, and token verification result, and then written into the process event message field contract.

3. The IC substrate production line process parameter data processing method as described in claim 2, characterized in that, The steps for generating a computable confidence field further include: Replay detection is introduced for sequential consistency verification, and the confidence level of the token verification result is adjusted when sequence number replay is detected; A key rotation version consistency check is introduced for message authentication digest verification, and the credibility level of the digest verification result is adjusted when the key rotation version is inconsistent; A whitelist version consistency check is introduced for identity token verification, and the trust level of the token verification result is adjusted when the whitelist versions are inconsistent; Changes in the credibility level are written into the original evidence fragment index of the evidence package.

4. The IC substrate production line process parameter data processing method as described in claim 1, characterized in that, The steps for constructing a heterogeneous spatiotemporal correlation graph and building a set of candidate paths for anomaly propagation include: The heterogeneous node types are defined as processes, equipment, events, parameters, quality, and environment. Generate time-series related edges based on time-series relationships and record lag features; generate influence-related edges based on influence weights and record the weight sources; generate confidence gates based on confidence fields and write the confidence gates into the directed edge generation rules. Based on the directed edge generation rules, a set of candidate paths for anomaly propagation is constructed and used to converge the cross-process causal search space and the linkage adjustment boundary.

5. The IC substrate production line process parameter data processing method as described in claim 4, characterized in that, The steps of constructing a set of candidate paths for anomaly propagation further include: Similarity candidate edges are generated based on similarity features, and a similarity score is formed; Based on the lag features, lag candidate edges are generated and lag scores are formed; Based on confidence gating, similarity candidate edges and lagging candidate edges are gating and filtering, and a gating score is generated; The path weights of candidate propagation paths are determined based on similarity scores, lag scores, and gating scores. An abnormal propagation candidate path set is formed according to the path weights, and the path weights are written into the evidence package.

6. The IC substrate production line process parameter data processing method as described in claim 1, characterized in that, The steps for obtaining dynamic quantile thresholds, anomaly type distributions, anomaly scores, quality risk vectors, key parameter trajectory predictions, and out-of-distribution discrimination results based on the unified output of the process fundamental model, and completing anomaly identification, include: The event sequence representation is obtained by encoding the event sequence of process event messages; Relation-aware encoding is performed on heterogeneous spatiotemporal correlation maps to obtain map representations; The event sequence representation and the graph representation are integrated to form a unified representation. Based on the unified representation, the dynamic quantile threshold, anomaly type distribution, anomaly score, quality risk vector, key parameter trajectory prediction and out-of-distribution discrimination results are output. The dynamic quantile threshold, anomaly score and out-of-distribution discrimination results are used together for anomaly identification and severity level classification.

7. The IC substrate production line process parameter data processing method as described in claim 5, characterized in that, Under the constraint of the set of candidate paths for abnormal propagation, the structural causal chain is located and the counterfactual effect is calculated. The steps that output the set of intractable root causes, lag windows, evidence packages, and linkage sequences are included: Determine the set of variables on the subgraph induced by the set of candidate propagation paths for anomalies; A priori mask matrix is ​​generated based on the set of candidate paths for anomaly propagation, and the priori mask matrix is ​​used to restrict the candidate edges of the variable set. Under the constraint of the candidate edges, the directed dependencies between variables are learned and the acyclic constraint is introduced. Based on counterfactual intervention, the counterfactual effect of the root variable on the target quality variable is calculated, and an intrusive root cause set is formed according to the counterfactual effect. The intrusive root cause set is associated with the abnormal propagation candidate path set and written into the evidence package, and the linkage order and lag window are output.

8. The IC substrate production line process parameter data processing method as described in claim 7, characterized in that, The steps of generating collaborative parameter adjustment instructions based on the set of interventionable root causes and executing executable mappings of equipment hard boundary constraints, ramp rate constraints, and process window constraints, and triggering freeze linkage or minimum action mode based on confidence field and out-of-distribution discrimination results and recording audit information include: Based on the set of operable root causes, the target process node and the target adjustment parameter set are determined, and collaborative parameter adjustment instructions are generated and written into candidate action vectors, execution time windows and idempotent keys. Perform equipment hard boundary constraint mapping, ramp rate constraint mapping, and process window constraint mapping on the candidate action vector to obtain the executable action vector and write the collaborative parameter adjustment instruction. Generate constraint trigger reason code and mapping residual and write audit information during each constraint mapping. Based on the confidence field and the out-of-distribution discrimination results, determine the freeze linkage mode or the minimum action mode, write the collaborative parameter adjustment instruction, and write the exit score as audit information; Issue collaborative parameter adjustment instructions and obtain execution receipts. Perform consistency verification on the execution receipts based on idempotent keys and write the verification results into the audit information.

9. A process parameter data processing system for an IC substrate production line, characterized in that, include: The standardized data acquisition unit is used to obtain process parameters, equipment status, quality inspection and environmental data of the target process nodes in the IC substrate production line and to complete unit unification and robust standardization. The event trust unit is used to generate a confidence field based on the results of time synchronization verification, message authentication digest, sequence consistency verification and identity token verification, and encapsulate it into a process event message, and aggregate the process event message according to the unique identifier of the IC carrier board. The graph reasoning unit is used to construct a heterogeneous spatiotemporal correlation graph with the unique identifier of the IC carrier board as the root, and generate directed edges based on time series relationships, influence weights and confidence gating, construct a set of candidate paths for anomaly propagation, and obtain dynamic quantile thresholds, anomaly type distribution, anomaly scores, quality risk vectors, key parameter trajectory prediction and out-of-distribution discrimination results based on the unified output of the process basic model, and complete anomaly identification. The causal decision unit is used to locate the structural causal chain and calculate the counterfactual effect under the constraint of the set of candidate paths for abnormal propagation, and outputs the set of interventionable root causes, lag window, evidence package and linkage sequence. The linkage audit unit is used to generate collaborative parameter adjustment instructions based on the set of interventionable root causes and execute executable mappings of equipment hard boundary constraints, ramp rate constraints and process window constraints. It also triggers freeze linkage or minimum action mode based on confidence field and out-of-distribution discrimination results and records audit information.

10. The IC substrate production line process parameter data processing system as described in claim 9, characterized in that, Causal decision-making units include: Subgraph-induced subunits are used to determine the set of variables on a subgraph induced by the set of anomaly propagation candidate paths; The mask constraint subunit is used to generate a priori mask matrix based on the set of anomaly propagation candidate paths and to restrict the edge candidates of the variable set using the priori mask matrix. Under the edge candidate constraint, the directed dependency between variables is learned and an acyclic constraint is introduced. The counterfactual assessment subunit is used to calculate the counterfactual effect of root variables on target quality variables based on counterfactual intervention and to form an interventionable root cause set according to the counterfactual effect; The evidence arrangement subunit is used to associate the set of interventionable root causes with the set of abnormal propagation candidate paths, write them into the evidence package, and output the linkage order and lag window.

Images