Adaptive semiconductor data asset decision-making arrangement method, computer device and readable storage medium

Abstract

The application provides a self-adaptive semiconductor data asset decision-making arrangement method, a computer device and a readable storage medium. It relates to the technical field of data asset management and intelligent decision-making. It includes footprint collection and intention extraction, causal constraint and semantic pedigree construction, budgeted risk community modeling, online decision engine with context constraint, strategy contract generation and zero-trust execution, counterfactual twin re-labeling and closed-loop retraining. The application realizes online optimal arrangement and auditable closed-loop governance under the constraints of privacy, compliance, carbon and cost, etc., for data retention, desensitization, migration and access quota.

Description

Technical Field

[0001] This invention relates to the fields of data asset management and intelligent decision-making technology, and in particular to online policy orchestration and governance for causal explainability and multiple budget constraints. Background Technology

[0002] Current enterprise data asset governance systems mostly focus on visualizing or statically statistically analyzing various business data, operation logs, and device events. They cannot automatically extract user intent or operational intent, resource dependency chains, and sensitive information domains from real usage footprints. They also cannot dynamically balance multiple budget boundaries such as privacy budgets, compliance rules, energy consumption constraints, and operating costs. Furthermore, they lack the ability to implement automated decision-making and one-click auditing based on unified causal / semantic constraints in business systems, device terminals, or data planes.

[0003] This lack of capability is particularly pronounced in industries with extremely high requirements for real-time performance, compliance, and data lineage, such as new energy vehicle sales and semiconductor manufacturing. For example, new energy vehicle sales platforms struggle to infer users' true intentions based on their activity patterns and to make actionable decisions in conjunction with inventory, pricing, and compliance rules. Meanwhile, semiconductor production lines struggle to automatically infer process dependencies and adjust parameters within compliance constraints based on engineer operation data, equipment logs, and process data.

[0004] Therefore, there is an urgent need for an end-to-end governance approach that uses digital footprints as an entry point, enables online decision-making under the constraints of causal relationships and semantic lineage, and securely implements it across different systems through strategy contracts, in order to solve the aforementioned key technical problems in specific industries. Summary of the Invention

[0005] The purpose of this invention is to overcome the deficiencies of the prior art and provide an adaptive semiconductor data asset decision orchestration method, computer device, and readable storage medium.

[0006] This invention is a closed-loop method of footprint → intent → causality / genealogy → budget → online decision → strategy contract → audit → counterfactual → retraining. It learns a causal DAG on the initial intent-dependency graph and solidifies semantic constraints with RDF / SHACL. It uses multi-budget joint loss and multiplier pool to handle privacy / compliance / carbon / cost coupling. It adopts a multi-armed gambling machine or MPC for online solution and outputs interval-based results. The decision is compiled into a strategy contract and executed by the Sidecar with zero trust and on-chain trace. The strategy is recalibrated and continuously retrained through counterfactual twins to achieve long-term adaptive steady state.

[0007] For semiconductor manufacturing and design companies, this solution addresses the technical challenges of unified online optimization and orchestration of process data retention strategies, sensitive parameter desensitization strength, cross-domain model calls, and equipment resource quotas under multiple constraints such as highly coupled process parameters, unstable equipment status, difficulty in tracing quality indicators, and privacy / compliance / cost / energy consumption, while ensuring auditable governance across the entire supply chain.

[0008] The present invention adopts the following technical solution: An adaptive semiconductor data asset decision orchestration method includes: S100. Collect digital footprints generated by business systems, IoT devices, logs, and interactive terminals in the target semiconductor management scenario. Perform unified time alignment and session segmentation on the digital footprints. Based on sequence pattern mining and weakly supervised natural language processing models, extract query / processing intents, dependency chains, sensitive domains, and service level agreement targets from the footprints to generate an initial intent-dependency graph.

[0009] S200, Causal Constraints and Semantic Genealogy Construction: Based on the initial intent-dependency graph, action nodes, indicator nodes, and compliance constraints are first set, and the PC algorithm is used to perform conditional independence testing. Combining the independence test results with domain prior learning of causal directed acyclic graphs, DRF / SHACL is used to define data lineage constraints, quality thresholds, and compliance rules to achieve verifiable constraints on what data - what action - affects which indicator.

[0010] S300, Modeling Budgeted Risk Community: Configuring multi-dimensional budgets for each data link and target action (privacy) (Budget, compliance penalties, carbon allowances and cost caps, etc.), and set system-level capacity vectors based on historical capacity configurations, SLA targets and compliance requirements. Unitize the budget to obtain Simultaneously, construct a value proxy function. Used to approximate the business benefits under a given context and action.

[0011] S400, Context-Constrained Online Decision Engine: Employs a constrained context multi-armed gambler or model predictive control (MPC), taking context = footprint state + causal parent set + budget margin as input, and maintaining online learning model parameters. (Including value proxy function parameters) Risk prediction / uncertainty estimation parameters and multiplier pool parameters This approach maximizes value representation (marginal business revenue / utilization efficiency) while satisfying multiple budget and SLA constraints, and outputs conformal risk prediction ranges based on an online-updated risk model. And the corresponding threshold determination results.

[0012] Furthermore, it establishes interval-based predictions for risk indicators such as SLA default risk, privacy breach risk, and budget overrun risk, and outputs risk point estimates. With uncertainty With confidence coefficient Constructing risk confidence intervals Based on this, a joint loss function and a Lagrange multiplier pool are constructed. It supports exchangeability and coupling constraints between budgets.

[0013] S500, Policy Contract Generation and Zero Trust Execution: Decisions are compiled into policy contracts, which include action parameters, constraints, audit points and rollback hooks; the sidecar executor verifies and distributes the policy contracts locally on the data plane, and records evidence (hash, timestamp and execution result) in an immutable audit ledger. S600, Counterfactual Twin Recalibration and Closed-Loop Retraining: When concept drift / performance degradation / increased risk of violation are detected, a counterfactual twin trajectory is generated based on the causal graph to evaluate the difference in benefits between unexecuted / alternative actions; accordingly, the value agent, confidence interval and Lagrange multiplier are updated, triggering online retraining and parameter rollback to form a continuous adaptive closed loop.

[0014] Furthermore, step S100 includes: collecting business events, sensor status, and front-end interaction footprints based on a distributed message publish / subscribe collection framework (such as the open-source distributed streaming messaging platform Kafka), Message Queue Telemetry Transport Protocol (MQTT), Full-Duplex Network Communication Protocol (WebSocket), and log collection pipeline; performing millisecond-level time alignment through Network Time Protocol (NTP) and segmenting sessions according to a 25-minute no-interaction rule.

[0015] By leveraging Bidirectional Long Short-Term Memory (BiLSTM) networks and PrefixSpan sequence pattern mining, query / processing intents, behavioral dependency chains, sensitive data domains, and service level agreement targets are extracted from event sequences. Events and intents are constructed into confidence edges in the form {src, dst, Δt, prob, session}, where src represents the starting node (source node) of the dependency relationship, which can be an event node or an intent node; dst represents the ending node (target node) of the dependency relationship, used to characterize the subsequent events / intents that src may influence or point to; and Δt represents the time interval between the occurrence of src and the occurrence of dst, Δt = t_dst - t_src, with units that can be set to seconds, minutes, or days depending on the data sampling granularity. The `prob` parameter is used to characterize the lag effect and sequence of management actions or behavior chains. `prob` represents the confidence or probability value of the dependency relationship between `src` and `dst` being valid, preferably ranging from 0 to 1. The larger the value, the more reliable the dependency relationship. `session` represents the session identifier or session window number to which the confidence edge belongs. It is used to isolate data from different time periods, different plots, or different management cycles to avoid erroneous connections across sessions, thereby ensuring that the intent-dependency initial graph construction process is clear, traceable, and facilitates subsequent causal structure inference and policy learning. Low-confidence edges are pruned, high-frequency edges are smoothed using EWMA, and hash chains and topological verification are used to ensure sequence integrity and directional consistency. Finally, a machine-readable intent-dependency initial graph is generated for the causal construction stage.

[0016] Furthermore, step S200 includes: wherein, edge direction optimization may employ the GES algorithm and score it using the BIC criterion to determine the causal edge direction; data fields, lineage links, quality thresholds, and cross-domain compliance rules may be further described based on DRF / SHACL to generate a queryable semantic genealogy; the causal chain of action → indicator may be written into the graph database in RDF or JSON-LD format, and confidence, path length, and conditional independence test evidence may be recorded; and gray-scale intervention may be performed through the sidecar method to verify the effectiveness of the causal edge in real time, thereby supporting subsequent budgetary modeling and online decision-making.

[0017] Further step S300 includes: unifying the dimensions of each budgeted consumption and defining a budget surplus vector; iteratively updating the Lagrange multipliers based on the budget surplus and consumption rate, increasing the constraint weights corresponding to over-limit budgets and decreasing the constraint weights corresponding to budgets with sufficient surpluses; and conducting budget feasibility assessments of candidate actions based on the relationship between data links, actions, and indicators / constraints, correcting actions that exceed any budget boundary or trigger coupled constraints through pruning, rollback, or re-scoring. After action execution, updating the budget surplus and Lagrange multiplier pool based on actual privacy consumption, compliance / violation results, carbon emission consumption, and cost expenditures, thereby forming a budgeted risk community model that can be dynamically adjusted with changes in links and actions, providing a unified constraint interface for subsequent budgeted modeling and online decision-making.

[0018] Further step S400 includes: based on the footprint state obtained in step S100, the causal parent set obtained in step S200, and the budget vector obtained in step S300, combining and encoding the three into a context vector for the current time step. Defined by retention period Storage tier Access / Bandwidth Quotas Desensitization intensity Cross-domain migration actions Composed action vectors Define the value proxy function To characterize marginal business benefits and usage efficiency, and to measure budget / risk consumption. Constructing a single-step expected objective function Combined with constraint functions The Lagrangian function is formed, and the optimal action is solved within the set of possible actions. The solution can be obtained through online learning or rolling optimization using constrained context multi-armed gamblers or model predictive control (MPC), and interval-based predictions of risk indicators are given based on conformity inference. The upper confidence bound and the risk threshold The comparison triggers a conservative degradation strategy, and finally, the context representation and Lagrange multiplier parameters are updated online based on the observed actual consumption and risk feedback. To adapt to environmental drift and SLA changes and continuously optimize decision output.

[0019] In a second aspect, the present invention also provides a computer device, including a memory and a processor, wherein the memory stores computer program instructions, and when the processor executes the computer program instructions, it performs the steps of the above-described adaptive semiconductor data asset decision orchestration method.

[0020] A computer-readable storage medium having stored thereon program instructions that, when executed by a processor, perform the steps of the adaptive semiconductor data asset decision orchestration method as described above.

[0021] The beneficial effects of this invention are: (1) This invention transforms behavior records scattered across business systems, IoT devices, logs and interactive terminals into parsable intent-dependency graphs through footprint collection and intent extraction, providing a unified basis for subsequent decision-making.

[0022] (2) The present invention introduces causal constraints and semantic spectrum, which not only realizes the causal interpretability of data-action-indicator, but also ensures the verifiability of quality and compliance through RDF / SHACL.

[0023] (3) The present invention designs a budgeted risk community model that can simultaneously manage privacy budgets, compliance penalties, carbon allowances and cost caps, realize the convertibility and coupling between multi-dimensional budgets, and improve decision robustness.

[0024] (4) This invention proposes a context-constrained online decision engine that uses multi-armed gambling machines or model predictive control methods to dynamically output the optimal strategy under the premise of meeting budget and SLA, and includes a conformity confidence interval to improve the reliability of the results.

[0025] (5) This invention translates decisions into executable, auditable and rollbackable contracts through a strategy contract mechanism, and uses the Sidecar zero-trust executor to issue and verify them, ensuring the security and traceability of strategy execution.

[0026] (6) The present invention combines counterfactual twins and closed-loop retraining, which can dynamically recalibrate and retrain when drift or risk is detected, so that the system can continuously adapt and operate efficiently in the face of environmental changes. Attached Figure Description

[0027] Figure 1 This is a flowchart of the steps of the present invention.

[0028] Figure 2 A schematic diagram illustrating the evolution of context-constrained online decision loops and budget margins. Detailed Implementation

[0029] To make the objectives, technical solutions, and advantages of this invention clearer, the technical solutions of this invention are described clearly and completely below. Obviously, the described embodiments are only some embodiments of this invention, not all embodiments. Based on the embodiments of this invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of this invention.

[0030] like Figures 1-2 As shown, the adaptive semiconductor data asset decision orchestration method of the present invention includes: S100, Footprint Collection and Intent Extraction: Collect various footprint data generated by wafer production business systems, IoT devices, logs, and interactive terminals in semiconductor manufacturing scenarios. Perform timestamp calibration (unified time alignment) and session-based segmentation on the footprint data to form a footprint sequence organized by session and time order. Based on a weakly supervised natural language processing model, perform multi-label classification and entity recognition on query statements, interface descriptions, log text, and configuration fields in the session. Combined with a preset SLA dictionary and rule templates, identify SLA (Service Level Agreement) targets and their constraints, such as availability, latency, throughput, and data freshness. Combined with privacy / compliance vocabularies and entity recognition results, mark and process sensitive data to ensure compliance with privacy protection requirements and ensure the legality and compliance of the data through compliance checks. (The privacy / compliance terminology is obtained from a pre-built rule base, which is derived from applicable laws and regulations, industry standards, internal data classification and grading specifications, and historical audit samples. The entity recognition results are obtained by an entity recognition model recognizing text, log field names, and metadata descriptions. The entity recognition model can be trained using a rule-based recognizer or a machine learning-based named entity recognition model.)

[0031] Data fields protected by laws, regulations, or organizational data classification and grading standards are marked as sensitive fields, and access control, de-identification processing, or cross-domain transfer is restricted to the sensitive fields according to the preset compliance rule base.

[0032] Based on this, taking the time-ordered footprint sequence within the session as input, each request or processing operation is mapped to the corresponding DPU identifier. A sequence pattern mining algorithm is used to mine high-frequency call sequences and preceding and following items patterns that meet the support threshold. Adjacent DPUs in the preceding and following items patterns are regarded as having call or data dependency relationships, thereby recovering the dependency chain.

[0033] The connection confidence is calculated based on the frequency of occurrence, conditional probability, and lift of the dependency chain in historical sessions. Low-confidence connections with confidence below a preset threshold or not in the Top-K are weakened or deleted to achieve sparsity. The retained connections are then checked for consistency with the existing service topology configuration, type constraints, and time causality constraints. Inconsistent connections are eliminated or downweighted. Finally, an initial intent-dependency graph that has undergone sparsification and consistency checks is generated according to the unified naming of DPU.

[0034] S200, Causal Constraints and Semantic Spectrum Construction: Based on the initial intent-dependency graph obtained in step S100, the footprint sequence (event / intent sequence and its features after session segmentation) obtained in step S100 is used as sample data. Conditional independence is tested, and structural search learning is performed within the candidate edge set defined by the initial intent-dependency graph to obtain a causal directed acyclic graph (DAG). Subsequently, the data objects, processing actions, and indicator results involved in the footprint sequence are modeled as data entity classes, action classes, and indicator classes in RDF (setting action nodes, indicator nodes, and compliance constraints), and their identification methods, attribute sets, and allowed relationships are standardized according to the domain ontology. Using RDF predicates such as hasInput, hasOutput, derivedFrom, and effectsMetric, the dependencies between upstream data → processing action → downstream data and processing action → indicator revealed by the DAG are encoded as lineage paths to form a queryable semantic spectrum of data-action-indicator (conditional independence is tested using PC algorithm).

[0035] Based on privacy and compliance guidelines and data quality standards, this paper defines lineage integrity constraints, quality threshold constraints, and compliance rules using SHACL shape constraints. Lineage integrity constraints limit the necessary input / output and derivation relationships between data entities and processing actions. Quality threshold constraints limit the value range of quality metrics such as missing rate, latency, and drift score, and their association with the corresponding data entities, processing actions, and metric entities. Compliance rules limit the storage area, access subjects, permitted action types, and cross-domain transfer conditions for sensitive data.

[0036] During system operation, new footprints continuously collected from business systems, IoT devices, logs, and interactive terminals are mapped into incremental data-action-metric triplets and attached to the existing semantic hierarchy according to the DAG. The RDF graph containing the new footprint information is submitted to the SHACL validation engine, which performs online validation on the links corresponding to the new footprints according to lineage constraints, quality threshold constraints, and compliance rules. When any quality threshold is exceeded and / or any compliance shape validation fails, a violation status is marked on the corresponding link and an alarm is triggered.

[0037] S300, Budgeted Risk Community Modeling: Based on the data-action-indicator semantic spectrum obtained in step S200, each (data) link in the semantic spectrum and its processing action nodes are defined as budget modeling units. For each link or action... Construct the budget vector Budget modeling unit In other words, It is a multidimensional characterization vector of its resource / risk budget, where each component Each dimension of the budget vector corresponds to a predefined budget dimension (such as storage usage, bandwidth / access frequency, computational overhead, privacy leakage risk, availability / latency risk, compliance violation penalties, carbon quotas and cost limits, etc.) and its budget value on that unit. Each dimension of the budget vector is used to characterize the resource or risk budget of the link or action in terms of storage usage, bandwidth / access frequency, computational overhead, privacy leakage risk, availability / latency risk, etc.

[0038] Based on historical capacity configuration, SLA (Service Level Agreement) targets, and compliance requirements, a system-level capacity (budget cap) vector is pre-defined. ,in Indicates the first The overall available capacity or tolerable risk ceiling for each budget dimension (such as storage, bandwidth / access, compute, privacy risk, availability / latency risk, etc.), and the budget unit. Budget vector Perform unitization processing, let If necessary, further Normalize to a preset dimension range to obtain a dimensionless and horizontally comparable unitized budget vector.

[0039] Control variables of the link or action (Including retention period, storage tier, access / bandwidth quota, masking strength, and cross-domain migration actions, etc.) For optimization decisions, each budget dimension is considered. Define constraint functions: in, This represents the total budget or risk ceiling for this dimension. Indicates a link or action The unit consumption coefficient in this dimension.

[0040] Construct the joint loss function: in, To characterize the objective of balancing business revenue and risk costs, This refers to the shadow price of the budget dimension, maintained by the Lagrange multiplier pool. The Lagrange multiplier pool is used to implement collaborative constraints and exchangeability relationships between different links and different budget dimensions.

[0041] For observed abnormal or missing budget components, a robust imputation strategy is first employed based on the historical distribution of similar links or actions, replacing the corresponding budget dimension with the median or preset quantile of historical data to mitigate the impact of extreme values. Subsequently, upper and lower quantile thresholds are set according to the empirical distribution or risk preference of each budget dimension. For unitized budget components that exceed the upper quantile or fall below the lower quantile, quantile clipping is performed, cutting them to within the threshold range to ensure the stability of the overall budget space.

[0042] By using online optimization or rolling gradient updates, the decision variables are... The price of shadows in the multiplier pool Iterative updates are performed to track the dynamic evolution of shadow prices over time and load, and the constraint strength of each budget dimension is dynamically adjusted to achieve budgeted risk community modeling between links and actions on a global scale.

[0043] S400, Context-Constrained Online Decision Engine: Based on the footprint state obtained in step S100, the causal parent set obtained in step S200, and the unitized budget vector obtained in step S300, the three are combined and encoded into the current time step. Context vector: in, This represents the session features, SLA objectives, sensitive areas, and quality status obtained from footprint collection and intent extraction. This represents the causal parent set feature of the node corresponding to the current link or action in a causal directed acyclic graph, where the node is denoted as . Its causal parent set is defined as and will The values / statistics / embedding representations (optionally including edge confidence and causal strength) of the parent nodes are aggregated and encoded to obtain the following: , This represents the remaining budget or risk margin across each budget dimension. Specifically, in this expression, Indicates from node To the node A directed edge represents a node. For nodes The causal relationship or dependency. (Arrow symbol) Indicates the direction of causality or dependency, meaning yes The causal parent node (i.e. State or behavior influence (Result). Symbol Indicates "belongs to", so Indicates from node To the node The directed edges belong to the graph The set of edges in the array. Here This represents a specific directed acyclic graph (DAG) used to describe and model causal or dependency relationships, where edges are directed and there are no cycles. The entire expression means from node To the node The directed edges are graphs A part of it indicates that there is a clear causal or dependency relationship between them.

[0044] The decision variables, such as retention period, storage tier, access / bandwidth quota, de-identification strength, and cross-domain migration method, are combined into an action vector: in For retention period, For storage level, For access / bandwidth quotas, For desensitization intensity, Actions related to cross-domain migration.

[0045] With the goal of balancing business revenue and risk costs, the following single-step expected objective function is constructed: in, In the context The final selected and actually issued action will be executed. The business revenue obtained, Indicated in terms of budget or risk dimension On the consumption or risk measurement, The shadow price from the Lagrange multiplier pool in step S300 is used to characterize the tightness of the budget in each dimension (i.e., the marginal cost per unit budget). ( For a moment (The generated set of candidate actions). To meet budget and compliance constraints, constraint functions can be further defined:

[0046] in This is the upper limit of the budget or risk threshold.

[0047] Construct the Lagrangian function within a constrained context multi-armed gambler or model predictive control (MPC) framework: When making decisions online, within the set of actionable actions Internal solution: At time t, the online decision engine searches within the set of possible actions A for the Lagrange cost. Minimum candidate action And this is taken as the optimal action to be finally selected and executed, where Composed of business revenue items and constraints / penalties for each budget / risk dimension (weighted by shadow price). Together they constitute the whole.

[0048] Through online learning or rolling optimization, based on observed actual consumption The shadow price (Lagrange multiplier) is updated using the projection subgradient method: in: Indicates the first Online decision-making moments in the budget / risk dimension The current value of the Lagrange multiplier (shadow price) on the dimension is used to characterize the tension of the constraint and determine the penalty weight for the corresponding consumption / risk.

[0049] The learning step size is used to control the pace of updates.

[0050] Indicating in decision-making actions and context Below, the gradient or degree of violation of budget constraints.

[0051] This represents a non-negative truncation operator; if the updated value is negative, it is truncated to zero.

[0052] Through this update mechanism, the Lagrange multipliers (shadow prices) will adaptively adjust with the dynamic changes in the budget surplus, ensuring that the constraint strength is dynamically adjusted according to actual consumption and budget changes.

[0053] For time Each candidate strategy The decision engine is based on historical feedback and context. For relevant risk indicators (Including SLA default risk, privacy breach risk, budget overrun risk, etc.) A range-based prediction is made, for example, given: ...; where... Indicates time The generated candidate policy set, Indicates the first There are several candidate strategies, and Specifically refers to the decision engine from Selected from and at time The actual strategies / actions implemented.

[0054] This expression means: in the context Take action below At that time, the first The true value of each risk indicator The interval prediction is that it falls within the lower bound. With the upper realm The constructed prediction interval Inside.

[0055] Among them, the upper confidence boundary It can be calculated through conformity inference or quantile regression models, such as: For point estimates of risk indicators, For uncertainty estimation, This is the confidence coefficient. From this, the risk index can be obtained. upper confidence level The upper confidence bound is compared with the risk threshold vector predefined in step S200. A comparison is made (the threshold can be calibrated in step S300 based on historical capacity configuration, SLA targets, and compliance requirements), when a certain dimension exists. satisfy:

[0056] When the risk exceeds the limit in the confidence sense, the strategy is determined to have exceeded the limit.

[0057] When the risk exceeds the limit, the online decision engine triggers a conservative degradation strategy, selecting the action that satisfies the requirement from the candidate action set. Conservative measures, such as shortening retention periods, lowering storage tiers, tightening access / bandwidth quotas, increasing desensitization strength, and / or prohibiting cross-domain migration, are taken to reduce overall risk exposure. Simultaneously, this decision and its feedback are used to update the context representation and Lagrange multiplier pool parameters, enabling the engine to continuously adapt and adjust as the environment drifts and SLA changes.

[0058] S500, Strategy Contract Generation and Zero-Trust Execution: The online decision output from step S400 is compiled into a strategy contract. This strategy contract includes Guard conditions for defining the scope of application and triggering conditions, audit points for recording key decisions and state changes, and rollback hooks for restoring a safe state in case of risk exceeding limits or execution anomalies. On the data plane side, Sidecar components deployed around each business node perform in-situ verification of incoming data access requests, processing requests, and cross-domain migration requests. When the Guard conditions are met, the execution of the corresponding strategy contract is triggered, and specific strategies such as retention period, storage level, access / bandwidth quota, de-identification strength, and cross-domain migration actions are issued according to the action templates defined in the contract.

[0059] During each strategy contract execution, the Sidecar component automatically collects execution evidence and constructs an execution record: It standardizes and concatenates the strategy contract version identifier, trigger request context summary, and decision result involved in the execution, and calculates the execution hash using a preset cryptographic hash algorithm. It reads and records the strategy trigger time and execution completion time from a trusted time source as timestamp information. The final result of this decision (including the allow / deny / downgrade determination, the actual retention period, storage level, access / bandwidth quota, de-identification strength, cross-domain migration action, and whether an alarm or rollback was triggered) is used as the result field. The execution evidence (including the execution hash, timestamp, and result) is appended to an immutable audit ledger or audit chain for post-event evidence collection, compliance review, and traceable auditing of the strategy contract execution process.

[0060] S600, Counterfactual Twin Recalibration and Closed-Loop Retraining: A monitoring and evaluation module is set up in the control plane. This module subscribes to the optimal strategy, confidence interval, and risk threshold output by the online decision engine in step S400 based on the context state, as well as the execution results, actual SLA performance, violation alarms, and rollback records written to the immutable audit ledger in step S500. Based on the decision log and audit evidence, it continuously monitors the context distribution, business performance indicators, and compliance risk indicators using a sliding time window. For concept drift, the monitoring and evaluation module compares the context and feedback distribution within the current window with the historical calibration window, and calculates the drift degree using statistical tests or distribution difference measures. For performance degradation, key performance indicators such as SLA default rate, response latency, and error rate are statistically analyzed and compared with preset target ranges or historical baselines to obtain the performance deviation. For increased compliance risks, the number of compliance alerts, strategy rollbacks, and the frequency of risk exceeding limits are statistically analyzed to determine the risk intensity. When any indicator satisfies the following conditions for multiple consecutive windows:

[0061] and These are the concept drift metrics. Performance deviation and risk intensity The preset judgment threshold is used to trigger alarms / rollbacks and other actions when the corresponding indicator exceeds the threshold for multiple consecutive windows.

[0062] The system determines that there is currently a significant increase in concepts drift, performance degradation, and / or compliance risks, thereby triggering a counterfactual twin recalibration and closed-loop retraining process. Once the triggering conditions are met, the system uses the causal directed acyclic graph (DAG) constructed in step S200 to generate corresponding counterfactual twin trajectories for the observed real-world footprints in the audit ledger. By replacing key actions or parameters on the DAG, the system simulates the counterfactual benefits and risks of alternative action sets under the same context, calculating the benefit and risk differences between the alternative actions and the current actions. Based on this, the value proxy function and risk confidence interval in step S300 are recalibrated, the shadow prices in the Lagrange multiplier pool are adjusted, and incremental training of the online model parameters in step S400 is triggered. This creates a continuously adaptive closed loop of prediction-execution-audit-counterfactual-retraining to address dynamic changes in environmental distribution, business load, and compliance requirements.

[0063] Furthermore, step S100 includes: collecting business events, sensor status, and front-end interaction footprints based on a distributed message publish / subscribe collection framework (such as the open-source distributed streaming messaging platform Kafka), Message Queue Telemetry Transport Protocol (MQTT), full-duplex network communication protocol (WebSocket), and log collection pipeline. Millisecond-level time alignment is performed using the Network Time Protocol (NTP), and sessions are segmented according to a 25-minute no-interaction rule.

[0064] We use Bidirectional Long Short-Term Memory (BiLSTM) networks and PrefixSpan to extract query / processing intents, behavioral dependency chains, sensitive data domains, and service level agreement targets from event sequences.

[0065] Confidence edges are constructed for events and intentions in the form of {src, dst, Δt, prob, session}, where src represents the starting node (source node) of the dependency relationship, which can be an event node or an intention node. dst represents the ending node (target node) of the dependency relationship, used to characterize the subsequent events / intentions that src may influence or point to. Δt represents the time interval between the occurrence of src and the occurrence of dst, Δt = t_dst - t_src, and the unit can be set to seconds, minutes, or days depending on the data sampling granularity, used to characterize the lag effect and sequence of management actions or behavioral chains. prob represents the confidence or probability value of the dependency relationship between src and dst being valid, preferably ranging from 0 to 1, with a larger value indicating a more reliable dependency relationship. The session represents the session identifier or session window number to which the confidence edge belongs. It is used to isolate data from different time periods, different plots of land, or different management cycles, avoiding erroneous connections across sessions. This ensures that the intent-dependency initial graph construction process is clear, traceable, and facilitates subsequent causal structure inference and policy learning. Low-confidence edges are pruned, high-frequency edges are smoothed using EWMA, and hash chains and topological verification are used to ensure sequence integrity and directional consistency. Finally, a machine-readable intent-dependency initial graph is generated for the causal construction phase.

[0066] Furthermore, step S200 includes: wherein edge direction optimization may employ the GES algorithm and score it using the BIC criterion to determine the causal edge direction; data fields, lineage links, quality thresholds, and cross-domain compliance rules may be further described based on DRF / SHACL to generate a queryable semantic genealogy; the causal chain of action → indicator may be written into the graph database in RDF or JSON-LD format, and confidence, path length, and conditional independence test evidence may be recorded; and gray-scale intervention may be performed through the sidecar method to verify the effectiveness of the causal edge in real time, thereby supporting subsequent budgetary modeling and online decision-making.

[0067] Further step S300 includes: unifying the dimensions of each budgeted consumption and defining a budget surplus vector; iteratively updating the Lagrange multipliers based on the budget surplus and consumption rate, increasing the constraint weights corresponding to over-limit budgets and decreasing the constraint weights corresponding to budgets with sufficient surpluses; and conducting budget feasibility assessments of candidate actions based on the relationship between data links, actions, and indicators / constraints. Actions exceeding any budget boundary or triggering coupled constraints are corrected through pruning, rollback, or re-scoring. After action execution, the budget surplus and Lagrange multiplier pool are updated based on actual privacy consumption, compliance / violation results, carbon emission consumption, and cost expenditures, thereby forming a budgeted risk community model that can be dynamically adjusted with changes in links and actions, providing a unified constraint interface for subsequent budgeted modeling and online decision-making.

[0068] In a second aspect, the present invention also provides a computer device, including a memory and a processor, wherein the memory has computer program instructions, and when the processor executes the computer program instructions, it performs the steps of the above-described adaptive semiconductor data asset decision orchestration method.

[0069] Thirdly, the present invention also provides a computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, implements the steps of the above-described adaptive semiconductor data asset decision orchestration method.

[0070] Example like Figure 1 As shown, in this embodiment, the adaptive semiconductor data asset decision orchestration method is deployed on the control plane and data plane of the wafer fab production system, applied to a typical etching-cleaning-inspection production process. The method executes steps S100 to S600 sequentially, with each step linked by explicit intermediate variables and results. Ultimately, a set of specific and explicit data retention, desensitization, migration, and access quota policies is output and implemented in the production system in the form of policy contracts. To ensure the effectiveness and operability of the policies, they must be quantitatively linked to key business indicators such as production efficiency and sales volume. Specifically, the strategies of this invention must ensure a significant reduction in key performance indicators during production or sales, such as production latency, resource waste, system load, production costs, and inventory management and market demand response speed in the sales process, without increasing compliance / privacy risks.

[0071] For example, when an alternative strategy can significantly reduce production latency or budget overrun risk without increasing compliance or privacy risks, the strategy should directly translate into improved production efficiency. Specifically:

[0072] Improve production efficiency: By reducing downtime or optimizing the resource allocation of the production line, the production quantity per unit time (such as the number of units produced per hour) can be increased, thereby improving output and production capacity.

[0073] Cost reduction: By optimizing resource allocation (such as energy consumption and material usage), waste in the production process can be reduced, thereby lowering production costs and increasing profit margins.

[0074] Similarly, adjustments to the strategy should have a direct impact on the sales process, including but not limited to: Increase sales volume: For example, by optimizing inventory management or supply chain response time, reducing inventory backlog, and improving the speed of response to market demand, sales volume and customer satisfaction can be increased.

[0075] Enhance market responsiveness: Through data analysis and optimized pricing and marketing strategies, ensure timely adjustments can be made to address market demand fluctuations and improve sales conversion rates.

[0076] Therefore, if an alternative strategy can significantly improve production efficiency (e.g., increase production volume and reduce production costs) or increase sales volume (e.g., increase customer conversion rates and reduce inventory backlog) without increasing compliance / privacy risks, the contract should be updated; otherwise, the previous version of the contract should be maintained or rolled back to ensure quantifiable effects on production and sales.

[0077] To demonstrate the specificity of the embodiment, this embodiment selects production data of a specific batch, Lot_ID=LotA123 (24 wafers), during the period from 08:00 to 12:00 (UTC) on January 20, 2026, as example input, and provides the specific strategy parameters and execution effects of this batch at different process stages. For ease of reproduction and verification, the example data used in this embodiment is given in structured record form, including at least: sensor timing for etching and cleaning, detection results (yield and defect distribution for each wafer), MES process logs, maintenance / quality text, unified footprint (FP), session segmentation results, session intent, sensitive area annotation, SLA target, candidate strategy set, risk interval prediction, and risk threshold, etc.; some representative data are illustrated in the following steps.

[0078] Step S100: Footprint Collection and Intent Extraction: The purpose of step S100 is to unify and organize the multi-source data footprints of etching machines, cleaning machines, inspection equipment, manufacturing execution systems (MES), and maintenance systems into session-organized sequence data, and extract production-related query / processing intents, dependency chains, sensitive domains, and SLA objectives from them to generate a machine-readable intent-dependency initial graph, providing a unified input for subsequent causal modeling and budget modeling.

[0079] Regarding input data, raw sampling records are collected from production equipment and IoT devices, such as the temperature, chamber pressure, and power of the etching equipment Etcher01, the flow rate and chemical concentration of the cleaning equipment Cleaner01, and the yield index and defect distribution of the testing equipment Tester01. For example, the sampling record of Etcher01 at 08:00:00 on 2026-01-20 is: temp=65.199℃, pressure=21.805 mTorr, power=1457.6 W; the sampling record at 08:00:01 is: temp=64.945℃, pressure=22.131 mTorr, power=1471.1 W; an abnormal period of temperature rise and chamber pressure fluctuation occurs from 09:30 to 10:00, which is used to support the subsequent "Temp_high→Press_unstable" mode. **On the cleaning side,** Cleaner01's sampling records at 08:20:00 on 2026-01-20 were: flow=11.813 L / min, conc=3.632%; the sampling records at 08:20:05 were: flow=12.328 L / min, conc=3.521%; a disturbance period of decreasing flow and increasing concentration occurred between 10:05 and 10:40. On the inspection side, Tester01 output batch inspection results at 11:50:00 on January 20, 2026. For example, wafer W01 had a defect_rate of 0.0155, yield of 98.18%, and top 1 defect pattern of pattern_shift; wafer W02 had a defect_rate of 0.0188, yield of 97.85%, and top 1 defect pattern of particle; and wafer W03 had a defect_rate of 0.0226, yield of 97.38%, and top 1 defect pattern of unknown. Simultaneously, operation logs and quality results were collected from the MES system, maintenance system, and quality system, including step records (Step_ID, Lot_ID, Recipe_ID), engineer maintenance records, alarm descriptions, yield analysis reports, and quality judgment conclusions.For example, MES records show that LotA123 executed the etching step EtchStep17 (Recipe_ID=Recipe_X9, operator_id=op_1295) from 08:05 to 08:55, the cleaning step CleanStep_03 (Recipe_ID=Recipe_C2, operator_id=op_2888) from 09:00 to 10:30, and the detection step TestStep_01 (Recipe_ID=Recipe_T1, operator_id=op_5657) from 11:20 to 11:50. To achieve a unified representation of multi-source data, the local time of each system is first converted to a unified UTC timestamp through the plant's time synchronization service. The standard footprint record is denoted as:

[0080] Where src_system represents the system from which the event originates, entity_id represents the Lot_ID or device ID, event_type represents the event type, payload represents the event payload, and t represents the unified timestamp. For example, the Etcher01 sensor sampling is encoded as: ⟨iot, Etcher01, sensor_sample, {temp_c:65.199, pressure_mTorr:21.805,power_w:1457.6}, 2026-01-20 08:00:00>; the MES process step start event is encoded as: ⟨MES, EtchStep_17, step_start, {process_segment:ETCH, recipe_id:Recipe_X9, operator_id:op_1295}, 2026-01-20 08:05:00>; and the maintenance text event is encoded as: ⟨maintenance, Etcher01,text_report, {text:"Chamber pressure oscillation observed..."}, 2026-01-20 08:35:00>. To facilitate conversational processing, sensor events can be sampled and aggregated at fixed steps (e.g., one event every 10 seconds for Etcher01 and one event every 30 seconds for Cleaner01) to maintain the readability and computability of the FP sequence.

[0081] Then, using `<lot_id, process segment>` as the session key, all footprints (FPs) are sorted by time. When the time interval between adjacent records exceeds a preset threshold (e.g., 30 minutes), they are automatically split into new sessions, resulting in a session set. Each session S_i corresponds to the complete footprint of a batch of wafers in the etching-cleaning-inspection chain. In this embodiment, the session segmentation results are shown below: S1 corresponds to the ETCH segment (t_start=2026-01-20 08:05:00, t_end=2026-01-20 08:55:00), S2 corresponds to the CLEAN segment (t_start=09:00:00, t_end=10:30:00), and S3 corresponds to the TEST segment (t_start=11:20:00, t_end=11:50:00). Simultaneously, natural language text is extracted from text fields such as maintenance work orders, MES notes, and quality analysis reports to form a text set.

[0082] For example, text1 is a maintenance work order: "Chamber pressure oscillation observed. Recommend checking RF power calibration. Possible impact on yield."; text2 is a quality analysis: "Defect signature trending to particle contamination; correlate withetch pressure instability. Limit access to Recipe_Param and operator_idfields." In terms of production intent recognition, based on weakly supervised rules and a lightweight classification model, a multi-label production intent set is output for each session S_i: For example, the output for S1 is {fault detection; yield optimization}, the output for S2 is {cycle stabilization; capacity improvement}, and the output for S3 is {defect root cause analysis; yield optimization}.

[0083] Regarding sensitive field identification, entity recognition and pattern matching are performed on the fields in the payload. Sensitivity levels are marked for fields such as operator_id, wafer_sn, and recipe parameter Recipe_Param, resulting in: For example, mark Recipe_Param as highly sensitive, and mark operator_id and wafer_sn as moderately sensitive, and use them as input for subsequent de-identification and access control policies.

[0084] Regarding the construction of candidate edges, a sequence pattern mining algorithm is used to mine event sequences, which can yield typical event patterns, such as: This is then abstracted into DPU nodes and their candidate dependency edges, and confidence and lift are calculated. Combined with SLA targets extracted from the process specifications, such as etching yield not less than 97% and cleaning cycle not exceeding 2 hours, SLA_targets are formed.

[0085] For example, the yield target for the etching stage is set as yield_pct≥97.0 (by batch), the cycle target for the cleaning stage is cycle_time_min≤120 (by batch), and the defect rate target for the inspection stage is defect_rate≤0.030 (by batch).

[0086] Define a set of DPU nodes V (e.g., DPU_Temp_Etcher01, DPU_Press_Etcher01, DPU_Yield_Tester01, etc.), and denote the set of candidate dependency edges after confidence and consistency filtering as E0, thus obtaining the initial intent-dependency graph: Intent – ​​Rely on the initial graph G0 as an intermediate input for the subsequent step S200.

[0087] Step S200: Causal Constraints and Semantic Genealogy Construction: Step S200 includes: based on the initial intent-dependency graph G0, learning the causal relationships between process parameters, operational actions, and quality indicators, and using RDF and SHACL to solidify them into a queryable semantic hierarchy and constraint rules, providing a causal and compliance basis for budget modeling and online decision-making.

[0088] First, select key variables from G0 to form a variable set V, for example: A causal structure learning method based on conditional independence testing and score search (such as PC algorithm, GES / BIC score) is used to learn the causal structure of the variable set V, resulting in a causal directed acyclic graph: in, Let X_i be the set of causal edges, and each directed edge X_i→X_j carries causal strength and confidence.

[0089] Subsequently, data entities, action entities, and indicator entities are mapped to RDF class instances. Using predicates such as `:hasInput`, `:hasOutput`, `:derivedFrom`, and `:affectsMetric`, causal relationships are written into the RDF graph, forming a semantic hierarchy `Graph_RDF`. Based on requirements such as quality thresholds, process parameter ranges, and sensitive data access control, a SHACL constraint set `Shapes` is constructed. Thus, step S200 yields the causal graph `G_DAG`, the semantic hierarchy `Graph_RDF`, and the quality and compliance constraint set `Shapes`.

[0090] In this embodiment, the representative causal edges obtained from structure learning include: Temp_Etcher01→Press_Etcher01 (confidence 0.81, causal strength 0.32), Press_Etcher01→Yield (confidence 0.77, causal strength -0.28), and Clean_Time→Defect_Rate (confidence 0.73, causal strength 0.21). Meanwhile, access control and migration constraints are solidified into constraint rules, for example: (1) Recipe_Param is only allowed to be accessed under the intent {yield optimization, defect tracing}, and is denied by default; (2) operator_id must undergo strong desensitization using hash or mask; (3) Recipe_Param prohibits original migration across plant areas, and only allows migration of aggregated features.

[0091] Step S300: Budgeted Risk Community Modeling Step S300 includes: configuring multi-dimensional budgets for privacy, compliance, carbon emissions, and cost around the etching-cleaning-inspection data link and its related actions; constructing a joint objective function and a Lagrange multiplier pool to provide budget constraints and a shadow price basis for subsequent strategy optimization. This applies to each important data link or action. (For example, Etcher01 log retention, recipe parameter access, cross-plant quality data migration, etc.) Define a budget vector:

[0092] in, Indicates privacy budget, Indicates compliance risk budget, Indicates carbon emission quotas. This indicates the cost budget.

[0093] Define the global budget upper limit vector: Privacy Budget. It represents the maximum privacy resources allocated to data operations (such as data storage, processing, or transmission) to ensure that data processing complies with privacy protection requirements and avoids the disclosure of sensitive information.

[0094] Regulatory Risk Budget. This represents a budget for risks related to law and compliance, ensuring that the company complies with relevant regulations during operations and avoids the risk of compliance violations.

[0095] Carbon quota. This represents the amount of carbon emissions a company is allowed to emit during its production process, aiming to reduce its environmental impact and align with sustainable development goals.

[0096] Cost Budget. It defines the maximum allowable expenditures during operation, including various production costs such as resource consumption, material procurement, and energy use.

[0097] By summing the allocated budgets for each link / action, we obtain the current budget balance: : Indicates the current budget balance, i.e. the remaining available budget.

[0098] : Represents the global budget ceiling, which includes the sum of budgets from multiple dimensions (such as privacy budget, compliance budget, carbon emission budget, and cost budget).

[0099] : Represents each data link or action The allocated budget may include expenditures across different budget dimensions, such as privacy, compliance, carbon emissions, and costs.

[0100] here The remaining budget is calculated by summing all allocated budgets. This formula is typically used in budget management and resource allocation processes to ensure that resource consumption does not exceed the global budget limit across all budget dimensions.

[0101] With the support of causal graph G_DAG and semantic hierarchy Graph_RDF, using retention period τ, storage level ℓ, access quota q, de-identification strength d, cross-domain migration m, etc. as decision variables a, we estimate the business revenue R(x,a) under context x and action a and the consumption c_k(x,a) in each budget dimension, and construct a joint objective function: in, This represents the joint objective function, which, in a given context... and decision-making actions The objective function aims to maximize revenue while considering costs incurred within budget constraints. It seeks to balance business gains with the penalties of budget expenditures to optimize the overall system performance.

[0102] The expected business reward is defined in the given context. and decision-making actions The expected return. Specifically, It is a business revenue function that measures revenue in a specific context. Under ) to take a certain decision action ( The profits obtained. This is the expected value calculation, which represents the average return calculated under different possible scenarios. The purpose of this part is to maximize the actual benefits or profits brought about by the decision.

[0103] λ_k is the shadow price maintained by the Lagrange multiplier pool, used to reflect the current tightness of each budget dimension.

[0104] Budget consumption is represented in a given context. and decision-making actions Below, budget dimension Resource consumption. For example. This could represent expenditures on privacy budgets, compliance risks, carbon emissions, or production costs. It measures the resource consumption or budget usage associated with a particular decision-making action. This section measures resource consumption across each budget dimension, combined with… Weighted penalties are applied to consumption to ensure that decisions do not exceed budget limits.

[0105] It is a penalty for budget expenditure. It applies to all expenditures across the budget dimension ( ) and the corresponding shadow price ( Multiply and sum the costs of each budget dimension to calculate its impact on the objective function. If a budget dimension has a high cost and tight constraints (i.e., ...), the impact is calculated. If the budget dimension is large, then the penalty on the objective function is also large, forcing the system to avoid excessive resource consumption in that dimension.

[0106] In this embodiment, the statistical period is the shift window (4 hours), and the budget upper limit is set as follows: B^cost = 200 yuan, B^carbon = 1.2 kgCO2e, B^reg = 1.0 (normalized dimension), B^priv = 2.0 (ε upper limit). Before reaching the cleaning stage, the cost has been 80 yuan, carbon 0.4 kgCO2e, compliance 0.2, and privacy ε = 0.6. Therefore, the budget surplus is r = (ε surplus 1.4, compliance surplus 0.8, carbon surplus 0.8, cost surplus 120). The multiplier pool {λ_k} can be updated online as the surplus changes to characterize the "marginal cost per unit budget".

[0107] Step S400: Online decision engine with context constraints: Step S400 includes: under the given context (footprint state, causal parent set, budget margin), using a constrained context multi-armed gambler or model predictive control algorithm, calculating the specific data management strategy applicable to the current process segment, including retention period, storage level, access quota, desensitization strength and cross-domain migration method, and performing an interval assessment of the risk of the strategy.

[0108] At a certain decision time t, construct the context vector: in, Session characteristics, current batch yield trend, and defect patterns from step S100 Causal parent set values ​​from G_DAG The budget surplus r is obtained from step S300.

[0109] Correspondingly define the action vector: in, Indicates the log retention period. Indicates storage hierarchy, This indicates access and bandwidth quotas. Indicates the degree of desensitization. This indicates a cross-plant migration strategy.

[0110] In the given and candidate actions In this case, the single-step objective function is: Budget constraints must also be met: The decision engine solves for the optimal action online from the set of possible actions. ^*, and calculate the confidence interval L_k(a| for each candidate action a) of the risk index. ), U_k(a| When there exists a dimension k that satisfies U_k(a| When the value of the candidate strategy exceeds B_k, it is considered that the risk of the candidate strategy exceeds the limit in this dimension, and a conservative downgrade strategy needs to be triggered.

[0111] In this embodiment, t=2026-01-20 11:00:00 is selected, and a set of candidate strategies is given (only three representative items are listed): A1: τ=30 days, ℓ=Cold, q=800 times / hour, d=medium, m=none; A2: τ=60 days, ℓ=Warm, q=1200 times / hour, d=strong, m=only migrate aggregated features; A3: τ=90 days, ℓ=Warm, q=2000 times / hour, d=strong desensitization and Recipe_Param only aggregates / interval output, m=only migrates aggregated features and prohibits the original Recipe_Param from crossing plant areas. An example of risk interval prediction for A1 is as follows: SLA default risk , U=0.2823; Privacy leakage risk , U=0.3329; Risk of budget overrun , U=0.0913. The risk thresholds are set as follows: θ_sla=0.25, θ_privacy=0.20, θ_budget=0.18. Therefore, A1 is removed because U_sla>θ_sla and U_privacy>θ_privacy. When the detection segment shows an increase in Defect_Rate and the causal parent set indicates a high abnormal contribution from Press_Etcher01, the optimal strategy output by the online decision is A3, i.e.: τ_t=90 days, ℓ_t=Warm, q_t=2000 times / hour (limited to intent {yield optimization / defect tracing}), d_t=strong desensitization (operator_id / wafer_sn hash or mask, Recipe_Param only intervalized / aggregated values), m_t=only migrate the aggregated defect distribution and statistical characteristics (prohibiting the original Recipe_Param from crossing factory boundaries), and the risk upper confidence bound is used to ensure that each dimension does not exceed the budget limit.

[0112] Step S500: Strategy Contract Generation and Zero-Trust Execution Step S500 includes: taking the optimal strategy output from step S400. The code is compiled into an executable strategy contract, which is then issued and enforced in a zero-trust manner within the Sidecar execution component, and immutable audit evidence is recorded. This step is primarily based on the action vector obtained in step S400. ^* and context Generate a strategy contract that includes triggering conditions, action parameters, and auditing rules. No new mathematical formulas will be introduced. The aforementioned decision variables and context representations will only be referenced in the implementation.

[0113] In this embodiment, the strategy contract includes at least: trigger condition {LotA123, process segment = detection, intent = defect root cause}, action parameters {τ = 90 days, ℓ = Warm, q = 2000 times / hour, d = strong desensitization, m = aggregate migration only and prohibit the original Recipe_Param from crossing plant areas}, audit fields {visitor ID, query fingerprint, desensitized version number, migration destination domain}, violation handling {over-quota rejection, alarm and write to audit ledger}, and records the hit constraint rules and budget margin changes to ensure traceability of execution.

[0114] Step S600: Counterfactual twin recalibration and closed-loop retraining: Step S600 includes generating counterfactual twin trajectories using a causal graph when changes in context distribution, system performance degradation, or increased risk of violation are detected; evaluating the difference in returns and risks of unexecuted or alternative strategies; recalibrating value agents and budget multipliers; and incrementally retraining or rolling back the online decision model to achieve long-term adaptive optimization of the strategy.

[0115] In this step, an alternative action a′ is constructed on the causal graph G_DAG for context x, and the counterfactual benefit and risk are calculated: By comparing the profit and risk differences between the current strategy and the counterfactual strategy, the λ_k in the joint objective function is adjusted, and the relevant strategy contracts are rolled back or updated.

[0116] In this embodiment, when a drift in the distribution of key Etcher01 variables relative to the historical baseline is detected (e.g., temperature / cavity pressure drift) or a deviation from the SLA target is detected (e.g., yield_pct below 97% or cycle_time exceeding 120 minutes), an alternative action a′ is constructed (e.g., reducing q or adjusting m to "available only in this plant area"). The counterfactual benefits are compared. Budget consumption of alternative strategies This triggers a Lagrange multiplier adjustment and strategy update. If the alternative strategy significantly improves productivity, reduces costs, or mitigates over-budget risk without increasing compliance / privacy risks, the contract is updated. Otherwise, the contract is maintained or rolled back to the previous version.

[0117] Improve production efficiency: If the alternative strategy can increase the production quantity per unit time (e.g., the number of units produced per hour) by more than 10% by optimizing the production process (e.g., by reducing downtime, improving the resource utilization of the production line), then the contract is renewed to improve capacity and production efficiency.

[0118] Reduce production costs: If alternative strategies can optimize resource allocation (such as reducing energy consumption or material waste) and reduce production costs (e.g., a 5% or greater reduction in production costs per unit), then renew the contract to ensure improved profit margins.

[0119] Reduce the risk of budget overruns: If alternative strategies optimize budget allocation or reduce the risk of resource consumption (e.g., reducing the risk of budget overruns by more than 15%), then update the contract to ensure more precise and efficient budget control.

[0120] If the alternative strategy can significantly improve these quantifiable business metrics without increasing compliance / privacy risks, then the strategy will be updated. If it fails to significantly improve these metrics, then the previous version of the contract will be maintained or rolled back.

[0121] As can be seen from the above embodiments, in the semiconductor manufacturing scenario, the method of the present invention, from step S100 (footprint collection and intent extraction), step S200 (causal constraints and semantic spectrum construction), step S300 (budgeted risk community modeling), step S400 (context-constrained online decision-making), step S500 (strategy contract generation and zero-trust execution), to step S600 (counterfactual twin recalibration and closed-loop retraining), provides clear objectives, input data, intermediate variable design, and intermediate result forms in each step. It also standardizes the definitions of key variables and objective functions through formulas, ultimately outputting specific and executable data management and production decision-making strategies such as retention period, hierarchical storage scheme, access quota, desensitization strength, and cross-domain migration rules. Furthermore, the embodiments demonstrate a closed loop of specific input—specific decision—specific output / effect: providing examples of actual input data for batches / time windows, budget upper limits and margin values, and output strategy parameters (τ, ...). The specific values ​​of (q, d, m), and the candidate elimination and optimal output process achieved by comparing the risk confidence boundary with the threshold, are distinguished from the abstract restatement of the invention content.

[0122] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention, and not to limit them; although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some of the technical features; and these modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the spirit and scope of the technical solutions of the embodiments of the present invention.

Claims

1. An adaptive semiconductor data asset decision orchestration method, characterized in that, include: S100. Footprint Collection and Intent Extraction: Collect data from semiconductor management scenarios and perform unified time alignment and session segmentation on multi-source digital footprints; S200. Causal Constraints and Semantic Lineage Construction: Based on the initial intent-dependency graph, this paper combines independence testing and domain prior learning to create a causal directed acyclic graph; DRF / SHACL is used to define data lineage, quality thresholds, and compliance rules. S300. Budgeted Risk Community Modeling: Configure multi-dimensional budgets for each data link and target action, construct joint loss functions and Lagrange multiplier pools, and support exchange and coupling constraints between budgets; S400. Context-Constrained Online Decision Engine: Employs a constrained context multi-armed gambler or model predictive control, taking context = footprint state + causal parent set + budget margin as input, maximizing value proxy while satisfying multiple budgets and SLA, and outputting risk thresholds with conforming confidence intervals; S500. Strategy Contract Generation and Zero Trust Execution: Compiles decisions into strategy contracts, including action parameters, constraints, audit points, and rollback hooks; S600. Counterfactual Twin Recalibration and Closed-Loop Retraining: When concept drift / performance degradation / increased risk of violation are detected, counterfactual twin trajectories are generated based on causal graphs to evaluate the difference in benefits between unexecuted / alternative actions.

2. The method according to claim 1, characterized in that, Step S100 includes extracting query / processing intents, dependency chains, sensitive domains, and SLA targets from the event footprint based on sequence pattern mining and weak supervision, and generating an initial intent-dependency graph. Specifically, this includes: collecting business events, sensor status, and front-end interaction footprints based on a distributed message publishing / subscription collection framework; performing millisecond-level time alignment using NTP and segmenting sessions according to a 25-minute no-interaction rule; extracting query / processing intents, behavioral dependency chains, sensitive data domains, and SLA targets from the event sequence using bidirectional LSTM and sequence pattern mining; constructing confidence edges between events and intents in the form of {src,dst,Δt,prob,session}; pruning low-confidence edges; smoothing high-frequency edges using EWMA; and using hash chains and topology verification to ensure sequence integrity and directional consistency. Finally, a machine-readable initial intent-dependency graph is generated for the causal construction stage.

3. The method according to claim 1, characterized in that, Step S200 includes: setting action nodes, indicator nodes, and compliance constraints based on the initial intent-dependency graph and combining domain priors; using the PC algorithm to perform conditional independence testing and optimizing edge directions with GES / BIC to learn a causal directed acyclic graph; describing data fields, lineage links, quality thresholds, and cross-domain compliance rules based on DRF / SHACL and generating a queryable semantic genealogy; writing the action → indicator causal chain into the graph database in RDF / JSON-LD format and recording confidence, path length, and independence test evidence; and verifying the effectiveness of causal edges in real time through Sidecar grayscale intervention to support subsequent budgeted modeling and online decision-making.

4. The method according to claim 1, characterized in that, Step S300 includes unifying the dimensions of each budget consumption and defining a budget surplus vector; iteratively updating the Lagrange multipliers based on the budget surplus and consumption rate, increasing the constraint weights corresponding to over-limit budgets and decreasing the constraint weights corresponding to budgets with sufficient surpluses; and conducting budget feasibility assessments of candidate actions based on the relationship between data links, actions, and indicators / constraints. Actions that exceed any budget boundary or trigger coupled constraints are corrected by pruning, rollback, or re-scoring. After the action is executed, the budget surplus and Lagrange multiplier pool are updated by writing back based on actual privacy consumption, compliance and non-compliance results, carbon emission consumption, and cost expenditures, thereby forming a budgeted risk community model that can be dynamically adjusted with changes in links and actions, providing a unified constraint interface for subsequent budgeted modeling and online decision-making.

5. The method according to claim 1, characterized in that, Step S300 includes: a multidimensional budget including a privacy budget, penalties for compliance violations, carbon allowances, and cost caps.

6. The method according to claim 1, characterized in that, Step S400 includes: defining a value proxy function to characterize marginal business revenue / usage efficiency, performing rolling optimization using a constrained context multi-armed gambler or MPC, generating decision actions regarding retention period, storage tier, access / bandwidth quota, desensitization strength, and cross-domain migration based on the optimization results, providing intervalized predictions based on conformance inference and setting risk thresholds and conservative degradation strategies, and updating the context representation and multiplier parameters online to adapt to environmental drift and SLA changes.

7. The method according to claim 1, characterized in that, S500 includes: on-site verification and distribution on the data plane via the Sidecar executor, and recording evidence in an immutable audit ledger, wherein the evidence includes hashes, timestamps and execution results.

8. The method according to claim 1, characterized in that, S600 includes updating the value agent, confidence interval, and Lagrange multipliers accordingly, triggering online retraining and parameter rollback, forming a continuous adaptive closed loop.

9. A computer device comprising a memory and a processor, wherein the memory stores a computer program, characterized in that, When the processor executes program instructions, it performs the adaptive semiconductor data asset decision orchestration method as described in any one of claims 1-8.

10. A computer-readable storage medium having stored thereon stored program instructions, characterized in that, When the program instructions are executed by the processor, the adaptive semiconductor data asset decision orchestration method as described in any one of claims 1-8 is performed.

Images