A data element driving-based fine chemical process safety data labeling method

Abstract

This application discloses a data-driven method for annotating safety data in fine chemical processes. The method, executed by a processor, includes: constructing a data element specification system for safety data in fine chemical processes; building an annotation model based on a pre-trained language model and a data element knowledge graph, using the data element specification system as a prior knowledge base and input constraints to generate a structured schema and annotation constraint rules for the annotation task; driving a multi-model collaborative annotation engine based on the annotation constraint rules to perform constrained annotation on the raw process safety data of the DCS / SIS system, attaching a confidence score to each annotation result and performing automated quality gating to replace manual verification, outputting the annotation results as data element instances; and jointly iteratively optimizing the data element specification system and the annotation engine based on the annotation execution data and model performance indicators. This application achieves fully automated and objective control of annotation quality, accurate differentiation of safety semantics at measurement points, and cross-project reuse of annotation products.

Description

Technical Field

[0001] This application relates to the fields of artificial intelligence and industrial data processing technology, specifically to a data element-driven method for labeling safety data in fine chemical processes. Background Technology

[0002] Existing data annotation methods and their problems: (1) General data annotation method: The current data annotation process uses a "manually defined label → manually labeled → reviewed" approach. The label system is defined temporarily for each annotation project, lacking domain knowledge constraints, and has the following problems: Label definitions vary from project to project / person and cannot be reused across projects; lack of semantic constraints leads to high rates of ambiguity and conflict in labeling; and the lack of standardized descriptions of labeling results makes it difficult to interoperate with downstream AI models and information systems.

[0003] (2) Current status of industrial data labeling: Data labeling in the industrial sector is still in its early stages, and current practices include: Abnormal operating conditions are manually marked directly on the DCS screen without structured output; alarm management consultation involves manual classification of alarm root causes with inconsistent classification standards; SIS assessment involves manual organization of interlocking events without standardized data description formats; HAZOP analysis uses Word / Excel records, lacking structured expressions that can be processed by machines; vibration, corrosion, and other monitoring data in equipment health management are scattered across various subsystems, lacking unified semantic descriptions; gas detection alarm data is not aligned with environmental safety standards, making automatic compliance determination impossible.

[0004] (3) Rule-based industrial data annotation method: Some industrial data annotation already employs rule-based automated methods, which extract information from raw data through predefined regular expressions, keyword matching, and threshold comparisons. These methods have the following limitations: High rule maintenance costs: The field of fine chemical process safety involves 165 core data elements and 10 data types. The constraint rules for each data element are different, and the cost of manually maintaining thousands of rules is extremely high and prone to errors. Inability to handle fuzzy / boundary situations: When the original data is not expressed in a standardized way (such as non-standard descriptions like "temperature is too high" in HAZOP reports), rule-based methods cannot correctly map it to the data element classification. Poor generalization ability: Rules are bound to specific devices and operating conditions, and the same set of rules cannot be migrated to devices with different process conditions. When the device expands production or the process changes, all rules need to be rewritten. Lack of semantic understanding: Keyword matching cannot distinguish the contextual semantics (such as "high pressure" may correspond to normal fluctuations, alarm status, or interlock triggering under different operating conditions), resulting in mapping errors.

[0005] (4) Current status of machine learning methods in general data annotation: In the field of general natural language processing, annotation methods based on pre-trained language models (such as BERT and GPT) have made significant progress. However, these methods cannot be directly applied to the field of industrial process safety because: Industrial process safety data is highly domain-specific: general NER models cannot identify the essential semantic differences in safety between "bromination reactor liquid phase temperature" and "jacket temperature," requiring domain data element specifications as prior knowledge injection; industrial data has multimodal characteristics: various data sources such as DCS time-series data, alarm log text, P&ID drawings, and HAZOP reports require differentiated model architectures, while general annotation models only provide a unified text processing framework; industrial scenarios have extremely high requirements for annotation reliability: incorrect annotation may lead to misjudgment of safety interlock logic or incorrect compliance audit conclusions, requiring quality assurance mechanisms such as confidence gating and multi-model integration, while general methods only provide the output of a single model; there are complex physical and temporal constraints among industrial data elements (such as alarm cancellation time ≥ generation time, interlock trigger variables must be consistent with SIF definitions), and general annotation models do not have constraint reasoning capabilities. Summary of the Invention

[0006] The main objective of this application is to provide a data element-driven method for labeling safety data in fine chemical processes, aiming to solve the following technical problems: Existing methods for labeling safety data in fine chemical processes lack unified semantic definitions for labels, resulting in inconsistent labeling results for the same data across different projects and personnel; labeling results cannot be reused or exchanged across projects and enterprises; labeling quality relies on human experience, and review lacks objective criteria; industrial data labeling is homogenized with general labeling methods, failing to utilize domain knowledge; process safety data cannot be automatically processed by machines; labeling specifications are coupled with labeling execution, preventing independent reuse of specifications; data with different safety semantics due to different measurement locations or equipment parts within the same secondary category cannot be distinguished and encoded; labeling constraint strategies for different types of data elements are not differentiated; rule-based labeling methods have poor generalization ability, high rule maintenance costs, and cannot handle fuzzy semantics; manual labeling processes are inefficient, inconsistent, and difficult to scale.

[0007] To achieve the above objectives, the first aspect of this application provides a data element-driven method for labeling safety data in fine chemical processes, executed by a processor, comprising: constructing a data element specification system for fine chemical processes; constructing a labeling model based on a pre-trained language model and a data element knowledge graph, using the data element specification system for fine chemical processes as a prior knowledge base and constraint input for machine learning, generating a structured schema for the labeling task, and generating labeling constraint rules based on the structured schema for the labeling task; driving a multi-model collaborative labeling engine based on the labeling constraint rules to perform constrained labeling on the raw process safety data of the DCS / SIS system of the fine chemical production unit, attaching a confidence score to each labeling result from the model output, and performing automated quality gating based on the confidence score to replace traditional manual verification; outputting the labeling results in the form of data element instances; and jointly iteratively optimizing the data element specification system and the multi-model collaborative labeling engine based on the data and model performance indicators of the labeling execution process.

[0008] Optionally, the step of constructing an annotation model based on a pre-trained language model and a data element knowledge graph, using the fine chemical process safety data element specification system as a prior knowledge base and constraint input for machine learning, and generating a structured schema for the annotation task, includes: using a pre-trained language model finely tuned with corpus from the industrial process safety domain to encode the data element definition text and the original process safety data fragments to be annotated into semantic vectors respectively; automatically establishing a mapping relationship between the annotation task and data elements through cosine similarity calculation; constructing a data element relationship knowledge graph with data elements as nodes and semantic relationships between data elements as edges; using a graph attention network for relationship reasoning to automatically discover implicit association constraints between data elements; and automatically generating a structured schema for the annotation task based on the data type and value domain attributes of the data elements.

[0009] Optionally, a multi-model collaborative annotation engine driven by annotation constraint rules performs constrained annotation on the raw process safety data of the DCS / SIS system of a fine chemical production unit. Each annotation result is accompanied by a confidence score from the model output. Automated quality gating is performed based on the confidence score, including: scheduling the corresponding annotation model according to the data type of the data element to pre-annotate the raw process safety data from the DCS / SIS system of the fine chemical production unit; calculating the model confidence score for each annotation result and performing automated quality gating based on the confidence score: For high-confidence samples with a confidence score not lower than a first threshold, their annotation results are automatically approved; for medium-confidence samples with a confidence score between the first and second thresholds, multiple heterogeneous models are used to independently predict and determine the final annotation result through voting, and the consistency rate between models is calculated as an annotation quality indicator; for low-confidence samples with a confidence score lower than the second threshold, a hybrid strategy of uncertainty sampling and diversity sampling is used to select samples from the active learning pool for annotation supplementation, and the model weights are incrementally updated.

[0010] Optionally, the method further includes: the constraint verification engine automatically performs multi-dimensional constraint verification before the annotation results are written. The multi-dimensional constraint verification includes: value range constraint verification, type constraint verification, time sequence constraint verification, association constraint verification, structural constraint verification, and measurement point location code consistency verification. The measurement point location code consistency verification automatically matches the corresponding alarm threshold and interlocking logic according to the measurement point location segment in the data element identifier. When the measurement point location segment indicates the liquid phase position, the corresponding alarm threshold and interlocking logic are matched, and the alarm threshold of the liquid phase position is rejected from being misused for the jacket position data.

[0011] Optionally, the data element specification system and the multi-model collaborative annotation engine are jointly iteratively optimized, including: automatic calibration of the value range based on the actual distribution of the labeled values, automatic supplementation of the enumeration based on the distribution predicted by the classification model, automatic discovery of association rules based on frequent pattern mining, automatic expansion of the measurement point location segment based on cluster analysis, online updating of the model based on efficient parameter fine-tuning, and automatic monitoring of annotation quality based on distribution drift detection.

[0012] Optionally, the pre-trained language model is BERT-base-chinese or RoBERTa-wwm-ext-large, and the semantic vector is 768-dimensional or 1024-dimensional; the mapping threshold of the cosine similarity is not less than 0.85. When the highest similarity exceeds the mapping threshold, a mapping is automatically established; when the highest similarity is lower than the mapping threshold, it is marked as a mapping to be confirmed and enters the active learning process; the pre-trained language model is used for domain-adaptive fine-tuning using corpus from the field of fine chemical process safety. The fine-tuning corpus includes IEC 61511 standard text, ISA-18.2 standard text, HAZOP analysis report corpus, DCS operation log corpus, and functional safety assessment report corpus.

[0013] Optionally, the edge relationship types of the data element relationship knowledge graph include reference relationships, temporal dependency relationships, triggering relationships, combination relationships, and association constraint relationships; the graph attention network takes the initial features and edge relationship type encoding of the data elements as input, and obtains an enhanced representation of each data element after K layers of graph attention propagation, where K≥2; the association constraint strength between data elements is calculated based on the enhanced representation, and association constraint rules are automatically established when the association constraint strength exceeds the constraint threshold; the implicit association constraints include cross-data element type association constraints, including cross-type constraints that the batch status should be marked as abnormal interruption after the process variable triggers the safety instrument function, and cross-type constraints that associate the safety instrument function trigger event when the combustible gas concentration exceeds the standard.

[0014] Optionally, the data types in the annotation constraint rules include integers, real numbers, booleans, enumerations, timestamps, durations, and structures; the multi-model collaborative annotation engine driven by the annotation constraint rules includes: using a fine-tuned BERT classification model for enumeration types, wherein the fine-tuned BERT classification model adds a fully connected classification head to the BERT pre-trained model, uses the enumeration value domain code of the data element as the classification label, and trains using the cross-entropy loss function; using a temporal anomaly detection model for real number types, wherein the temporal anomaly detection model adopts a two-stage concatenated architecture of Isolation Forest and AutoEncoder, wherein the AutoEncoder uses an LSTM codec to learn the temporal reconstruction features of sensor readings under normal operating conditions and outputs a multi-dimensional reconstruction error vector, and Isolation Forest performs anomaly scoring on the multi-dimensional reconstruction error vector, with the values ​​of normal data points directly output as annotation values; using a regular parser in conjunction with a temporal logic check finite automaton for timestamp and duration types, wherein the regular parser matches and extracts the time expression in the text and automatically converts it to ISO. The 8601 standard format uses a finite automaton with temporal logic verification to check whether the extracted results satisfy the temporal constraints between events. For structural types, a multi-task sequence labeling model is used, which adopts an architecture of a shared underlying encoder plus a task-specific CRF decoder head. Each sub-attribute corresponds to an independent CRF decoder head, and BIO label sequences are used for training. For Boolean types, a rule engine and a binary classification model are used in a concatenated architecture. First, the rule engine checks whether there is a clear truth condition. If it is satisfied, the Boolean value is directly output. Otherwise, the model with a binary classification head added to BERT makes a judgment based on the contextual semantics.

[0015] Optionally, the first confidence threshold and the second confidence threshold are adaptively adjusted based on the sliding window statistics of the most recent N batches of annotation quality data, where N≥5; when the violation rate of the automatically passed annotation results in subsequent constraint verification is lower than the first violation threshold, the first confidence threshold is gradually increased; when the backlog of low confidence queue samples exceeds the preset capacity limit, the second confidence threshold is appropriately decreased; the multiple heterogeneous models include the BERT-base-chinese fine-tuning model, the RoBERTa-wwm-ext fine-tuning model, and the RoBERTa-wwm-ext-large fine-tuning model; for discrete annotation tasks, majority voting or weighted voting is used to determine the final annotation result, and for continuous annotation tasks, weighted averaging is used; the consistency rate between models is defined as the ratio of the number of models corresponding to the most frequently occurring value in the prediction results to the total number of models; when the consistency rate is lower than the consistency rate threshold, the processor automatically triggers the model fine-tuning process, using the prediction difference information of the sample and its multiple models as training signal increments to update the model weights.

[0016] Optionally, the range constraint verification includes checking whether the labeled value of the interval range satisfies the open or closed interval constraint, checking whether the labeled value of the enumerated range is in the predefined enumeration set, and checking whether the labeled value of the discrete set range is one of the set elements; the type constraint verification includes checking the numerical precision and dimension of real number types, rejecting decimal input for integer types, and verifying ISO for timestamp types. 8601 format compliance and time zone consistency; Duration type validation: Duration format and non-negativity are checked; Boolean type only allows true or false; Time sequence constraint validation: Time sequence constraint validation rules are automatically generated based on the time sequence dependency relationship edges of the knowledge graph, including: alarm cancellation time is not earlier than alarm generation time, interlock recovery time is not earlier than trigger time, and safety instrument function trigger time is within the corresponding work section time window; Association constraint validation: Cross-data element association consistency validation rules are automatically generated based on the reference relationship and constraint relationship edges of the knowledge graph, including: the associated tag number in the alarm event must exist in the equipment master table, the safety instrument function trigger variable must be consistent with the trigger variable in the safety instrument function definition, and the recipe temperature setpoint should be lower than the alarm high limit of the corresponding measuring point location; Structural constraint validation: This includes checking whether all required sub-attributes of the structural type data element have been filled in, whether the value of each sub-attribute meets its independent data type and value range constraints, and rejecting writing when required sub-attributes are missing and triggering model re-inference.

[0017] This application discloses a data element-driven method for annotating safety data in fine chemical processes, executed by a processor. The method includes: constructing a data element specification system for fine chemical processes; building an annotation model based on a pre-trained language model and a data element knowledge graph, using the data element specification system as a prior knowledge base and constraint input for machine learning to generate a structured schema for the annotation task; generating annotation constraint rules based on the structured schema; driving a multi-model collaborative annotation engine based on the annotation constraint rules to perform constrained annotation on the raw process safety data of the DCS / SIS system of a fine chemical production unit, attaching a confidence score to each annotation result from the model output, and performing automated quality gating based on the confidence score to replace traditional manual verification; outputting the annotation results as data element instances; and jointly iteratively optimizing the data element specification system and the multi-model collaborative annotation engine based on the data and model performance indicators during the annotation execution process. Through the aforementioned technical means, this application achieves the following technical effects: First, by constructing a multi-level data element specification system covering the process safety business domain, all labeled items share the same data element specification system, the semantic definition of labeled tags is unified, eliminating labeling ambiguities between projects and personnel, and the labeling results are output as standardized data element instances with unified encoding and schema, enabling cross-project and cross-system exchange and reuse. Second, by using the data element specification system as a prior knowledge base and constraint input for machine learning, expert knowledge in the field of fine chemical process safety is encoded into machine-executable labeling constraints, and automatic reasoning is achieved through knowledge graph injection into graph neural networks, enabling effective utilization of domain knowledge. Third, by implementing automated quality gating based on confidence scores, combining value range constraints, type constraints, and time sequence constraints to provide objective legality criteria, and employing a multi-model integrated voting mechanism, fully automated quality assurance is achieved without relying on human experience to judge the correctness of labels. Fourth, by using measurement point location segment coding, data elements with different safety semantics under the same secondary category due to different measurement locations are distinguished, ensuring that gas phase temperature, liquid phase temperature, jacket temperature, etc., are not confused due to the same category, and each independently carries interlocking logic and alarm threshold constraints. Fifth, by using a multi-model collaborative annotation engine, differentiated annotation strategies and dedicated annotation model architectures are adopted for different data types, significantly reducing type-related annotation errors and improving annotation accuracy. Sixth, by using a semantic matching method based on a pre-trained language model, non-standard descriptions and fuzzy expressions can be handled, no longer relying on precise keyword matching, and can be transferred across devices and operating conditions, improving annotation generalization. Attached Figure Description

[0018] Figure 1 A flowchart illustrating a data element-driven method for labeling safety data in fine chemical processes, provided in an embodiment of this application; Figure 2A schematic diagram of a multi-level classification structure of a data element specification system provided in this application embodiment; Figure 3 A schematic diagram of a measurement point location segment coding rule provided in an embodiment of this application; Figure 4 A schematic diagram of a data type system provided in an embodiment of this application; Figure 5 This application provides a schematic diagram illustrating the mapping relationship between annotation tasks and data elements. Figure 6 This application provides a schematic diagram of an automatic generation process for annotation constraint rules. Figure 7 A schematic diagram of a multi-model collaborative annotation engine architecture provided in this application embodiment; Figure 8 This application provides a schematic diagram of a standardized output process for annotation results. Figure 9 This is a schematic diagram of a joint iterative optimization process provided in an embodiment of this application.

[0019] The realization of the purpose, functional features and advantages of this application will be further explained in conjunction with the embodiments and with reference to the accompanying drawings. Detailed Implementation

[0020] It should be understood that the specific embodiments described herein are merely illustrative of this application and are not intended to limit this application.

[0021] The most relevant existing technologies are rule-based industrial data annotation methods and pre-trained language model annotation methods in the field of general natural language processing.

[0022] Rule-based industrial data annotation methods extract information from raw data through predefined regular expressions, keyword matching, and threshold comparison. The typical workflow is: manually writing extraction rules → applying the rules to the raw data → outputting the extraction results → manual review. This method suffers from high rule maintenance costs, inability to handle fuzzy semantics, and poor generalization ability; it also fails to correctly map data elements to classifications when the raw data is not properly represented.

[0023] General pre-trained language model annotation methods use models such as BERT and GPT to perform named entity recognition or classification of text. The common processing flow is: input text → model encoding → output annotation results. This method cannot be directly applied to the field of industrial process safety because: general models cannot recognize the essential semantic differences between "liquid phase temperature" and "jacket temperature" in industrial data; industrial data has multimodal characteristics, requiring differentiated model architectures; industrial scenarios have extremely high requirements for annotation reliability, necessitating quality assurance mechanisms such as confidence gating and multi-model integration; and complex physical and temporal constraints exist between industrial data elements, which general models lack constraint reasoning capabilities.

[0024] Furthermore, existing annotation methods couple annotation specifications (label system definitions) with annotation execution (labeling process), making it impossible to reuse specifications independently; data with different safety semantics due to different measurement locations within the same secondary category cannot be encoded differently; and annotation constraint strategies for different types of data elements are not systematically differentiated. These issues result in industrial process safety data annotation failing to meet practical application requirements in terms of consistency, accuracy, reusability, and automation.

[0025] refer to Figure 1 The first embodiment of this application provides a data-element-driven method for labeling safety data in fine chemical processes, addressing the technical problems mentioned in the background art, such as the lack of unified semantic definitions for labeling tags, reliance on manual experience for labeling quality, inability to distinguish the encoding of safety semantics at measurement point locations, and coupling between labeling specifications and execution. This method can be executed by a processor, which can be located on a terminal or server. The execution process of this method is as follows: Step S101: Construct a data element specification system for safety in fine chemical processes.

[0026] A multi-level data element classification system covering the process safety business domain is constructed by the processor. (Reference) Figure 2The primary classification includes 9 categories: A. Process Variables, B. Alarm Management, C. Safety Instrumented Systems, D. Equipment and Facilities, E. Risk Analysis, F. Maintenance Management, G. Environmental Safety, H. Process Formulation, and Z. Extension Reserves. The secondary classification includes more than 34 subcategories, including 6 in category A (A1 Reaction Process Variables, A2 Separation Process Variables, A3 Heat Transfer Process Variables, A4 Storage and Transportation Process Variables, A5 Utility Variables, A6 Component Analysis Variables), 5 in category B (B1 Alarm Events, B2 Alarm Configuration, B3 Alarm Performance, B4 Alarm Justification, B5 Alarm Root Causes), 5 in category C (C1 SIF Definitions, C2 SIL Data, C3 Interlock Logic, C4 Validation Testing, C5 Operational Assessment), 5 in category D (D1 Dynamic Equipment, D2 Static Equipment, D3 Control Instruments, D4 Safety Instrumented Systems, D5 Electrical Equipment), and 4 in category E (E1 HAZOP, E2 LOPA, E3...). SIL rating, E4 accident data), 4 categories of F (F1 work order management, F2 failure mode, F3 predictive maintenance, F4 spare parts management), 3 categories of G (G1 gas detection, G2 emission monitoring, G3 occupational health), 2 categories of H (H1 formula parameters, H2 batch data), and category Z is an extended reserved category that supports enterprise-defined data elements.

[0027] Each data element contains an attribute set including identifier, name, definition, data type, and value range. (See reference) Figure 4 The data types include 10 types: INT integer, REAL real number, STR string, BOOL boolean, ENUM enumeration, TS timestamp, DUR duration, ARR array, STC structure, and SEQ sequence.

[0028] Identifiers are generated according to the coding rule of standard owner code-classification code-measurement point location segment-property code-sequence number, with the format ZCT-AABBN-SSS, where ZCT is the standard owner code, AA is the combination of primary classification code and secondary classification sequence number, BB is the subdivision of measurement point location, N is the data element property code (T represents time-varying / measured value, C represents configuration / static value, E represents event, R represents analysis result), and SSS is the sequence number from 001 to 999.

[0029] refer to Figure 3The BB segment measurement point location code is used to distinguish different data elements with different safety semantics due to different measurement locations or equipment parts within the same secondary category. The judgment criteria are: if two measured values ​​trigger different interlocking logics, alarm thresholds, or produce different safety consequences, they should be defined as different data elements and use different BB values; if they belong to redundant sensors at the same measurement location, they belong to the same data element and use the same BB value, distinguished by the measurement point tag number. Commonly used BB segment location codes include: 00 = Default segment / No distinction, 01 = Gas phase / Top, 02 = Liquid phase / Bottom, 03 = Jacket / Outer shell, 04 = Tube side / Internal, 05 = Inlet, 06 = Outlet, 07 = Middle / Middle section, 08-99 = Reserved.

[0030] In one embodiment of this application, the fine chemical process safety data element specification system includes: a multi-level classification based on process safety business domains, divided into multiple primary categories and multiple secondary categories; wherein, the primary categories include process variables, alarm management, safety instruments, equipment and facilities, risk analysis, maintenance management, environmental safety, process formulation, and extended reservations; each data element contains an attribute set including an identifier, name, definition, data type, and value range; wherein, the data types include integers, real numbers, strings, booleans, enumerations, timestamps, durations, arrays, structures, and sequences; the identifier is generated according to the encoding rule of specification owner code - classification code - measurement point location segment - property code - sequence number; wherein, the measurement point location segment is used to distinguish data elements under the same secondary category that trigger different interlocking logics, alarm thresholds, or produce different safety consequences due to different measurement locations, so that measurement points with different safety semantics become independent data elements carrying independent safety constraints at the encoding level.

[0031] Furthermore, the safety semantic differentiation mechanism of the BB segment is as follows: In fine chemical production, the gas phase temperature and liquid phase temperature of the same reactor trigger different interlocking logics (such as high gas phase temperature triggering emergency cooling, and high liquid phase temperature triggering feed cut-off), and have different alarm thresholds (such as gas phase temperature alarm value set to 120°C, and liquid phase temperature alarm value set to 100°C), resulting in different safety consequences (such as gas phase overheating may lead to gas phase decomposition and explosion, and liquid phase overheating may lead to uncontrolled reaction and material overflow). Therefore, at the coding level, they are distinguished as independent data elements by different BB values, rather than just as different instances of the same data element.

[0032] For example, the coding examples are as follows: ZCT-A100T-001 is the reaction process variable - temperature - time-varying value (number 1, default position); ZCT-A101T-001 is the gas phase temperature of the reactor (BB=01, gas phase position); ZCT-A102T-001 is the liquid phase temperature of the reactor (BB=02, liquid phase position); ZCT-A103T-001 is the jacket temperature of the reactor (BB=03, jacket position).

[0033] Further, two encoding modes are adopted for value domains: an enumerated value domain adopts the format of "code:name", and an interval value domain adopts mathematical interval notation, including open interval (a,b), closed interval [a,b], half-open interval, discrete set {v1,v2,...} and infinite interval. The conversion from the value domain to the labeling constraint rule is as follows: for the open interval (a,b), the labeled value x must satisfy a<x<b, and boundary values are automatically rejected; for the closed interval [a,b], the labeled value x must satisfy a≤x≤b, and boundary values are allowed; for the discrete set {v1,v2,...}, the labeled value must be one of the elements in the set.

[0034] Further, the data element specification system serves as both the prior knowledge base and the constraint condition input for the subsequent machine learning labeling process, which specifically includes: the definition text of the data element serves as the semantic anchor of the pre-trained language model, for constructing the semantic matching benchmark of the labeling task; the reference relationship between data elements serves as the edge relationship of the knowledge graph, for the relationship reasoning of the graph neural network; the value domain encoding and type constraint of the data element serve as the input parameters of the constraint verification engine, for the automatic legality verification of the labeling result; the classification encoding of the data element (including the BB segment position encoding) serves as the category label system during model training, to ensure that the labeling output is aligned with the encoding specification.

[0035] Step S102: constructing a labeling model based on the pre-trained language model and the data element knowledge graph, taking the fine chemical process safety data element specification system as the prior knowledge base and constraint condition input for machine learning, generating a structured Schema of the labeling task, and generating labeling constraint rules based on the structured Schema of the labeling task.

[0036] The processor performs semantic encoding and knowledge graph construction on the data element specification system in step S101, and automatically generates a labeling specification for the labeling task.

[0037] In an embodiment of the present application, a pre-trained language model fine-tuned with corpora in the industrial process safety field is used to encode the data element definition text and the raw process safety data segments to be labeled into semantic vectors respectively, and the mapping relationship from the labeling task to the data elements is automatically established through cosine similarity calculation; a data element relationship knowledge graph is constructed with data elements as nodes and semantic relationships between data elements as edges, and a graph attention network is used for relationship reasoning to automatically discover implicit association constraints between data elements; the structured Schema of the labeling task is automatically generated based on the data type and value domain attribute of the data elements.

[0038] Further, the semantic matching mapping based on the pre-trained language model specifically includes: The processor uses a pre-trained language model (such as BERT-base-chinese or RoBERTa-wwm-ext-large) fine-tuned with corpus from the industrial process safety domain to encode the data element definition text and the original data fragments into high-dimensional semantic vector representations, and automatically establishes the mapping relationship between the annotation task and the data elements through semantic similarity calculation.

[0039] Data element definition encoding: For each data element deᵢ, the definition text Definition(deᵢ) is input into the pre-trained language model, and a 768-dimensional semantic vector representation e(deᵢ)∈ℝ is obtained through the [CLS] token. 768 .

[0040] Raw data fragment encoding: Input the raw data fragment dⱼ to be labeled (such as DCS log text, HAZOP record paragraph, alarm event description) into the same pre-trained language model to obtain the semantic vector representation v(dⱼ)∈ℝ 768 .

[0041] Semantic similarity calculation: Calculate the cosine similarity between each data segment v(dⱼ) and all data elements e(deᵢ): sim(dⱼ, deᵢ) = cos(v(dⱼ), e(deᵢ)) = (v(dⱼ)·e(deᵢ)) / (‖v(dⱼ)‖·‖e(deᵢ)‖) Mapping decision: For each data segment dⱼ, select the data element de* with the highest semantic similarity as the mapping target: de* = argmax_{deᵢ∈DE} sim(dⱼ, deᵢ) When the highest similarity exceeds the preset threshold τ_match (e.g., 0.85), a mapping is automatically established; when it is below the threshold, it is marked as "mapping to be confirmed" and enters the subsequent active learning process.

[0042] In an optional embodiment of this application, the pre-trained language model is BERT-base-chinese or RoBERTa-wwm-ext-large, and the semantic vector is 768-dimensional or 1024-dimensional; the mapping threshold of cosine similarity is not less than 0.85. When the highest similarity exceeds the mapping threshold, the mapping is automatically established. When the highest similarity is lower than the mapping threshold, it is marked as a mapping to be confirmed and enters the active learning process; the pre-trained language model is used for domain-adaptive fine-tuning using corpus of fine chemical process safety. The fine-tuning corpus includes IEC 61511 standard text, ISA-18.2 standard text, HAZOP analysis report corpus, DCS operation log corpus, and functional safety assessment report corpus.

[0043] Furthermore, knowledge graph-assisted mapping and relational reasoning specifically include: The processor constructs a knowledge graph of relationships between data elements and uses a graph neural network for relationship reasoning to automatically discover implicit association constraints. Using 165 core data elements as nodes and semantic relationships between data elements as edges, a data element relationship graph G=(V,E) is constructed. (Reference) Figure 5 Edge relationship types include: Reference relationships: For example, the "associated tag number" of an alarm event references the "device tag number" of the device's basic parameters; Temporal dependencies: such as the temporal chain of alarm timestamp → confirmation time → cancellation time; Triggering relationships: such as the causal relationship between a process variable exceeding the high alarm level and a SIF triggering event; Compositional relationships: such as the inclusion relationship between STC composite data elements and their sub-attributes; Related constraints: such as the safety margin relationship between the formula temperature setpoint and the alarm high limit.

[0044] A graph attention network is used to learn the encoding of the knowledge graph. The input consists of initial features of nodes (data elements) (including data type encoding, value range encoding, and classification hierarchy encoding) and edges (relation type encoding). After K layers (K≥2) of graph attention propagation, an enhanced representation h(deᵢ)∈ℝᵈ is obtained for each data element. The strength of the association constraints between data elements is calculated based on the enhanced representation. constraint_score(deᵢ, deⱼ) = σ(W·[h(deᵢ); h(deⱼ)]) Where σ is the sigmoid activation function, W is the learnable weight matrix, and [·; ·] denotes vector concatenation. Data elements whose constraint_score exceeds the threshold τ_constraint (e.g., 0.7) are automatically associated with constraint rules.

[0045] In an optional embodiment of this application, the edge relationship types of the data element relationship knowledge graph include reference relationships, temporal dependency relationships, triggering relationships, composition relationships, and association constraint relationships; the graph attention network takes the initial features and edge relationship type encoding of the data elements as input, and obtains an enhanced representation of each data element after K layers of graph attention propagation, where K≥2; the association constraint strength between data elements is calculated based on the enhanced representation, and association constraint rules are automatically established when the association constraint strength exceeds the constraint threshold; the implicit association constraints include cross-data element type association constraints, including cross-type constraints that the batch status should be marked as abnormal interruption after the process variable triggers the safety instrument function, and cross-type constraints that associate the safety instrument function trigger event when the combustible gas concentration exceeds the standard.

[0046] Furthermore, implicit constraints are automatically discovered: GNN's relational reasoning can discover implicit constraints that are difficult to exhaustively identify manually. For example, by analyzing the relational path "reaction temperature (BB=01 liquid phase) → alarm event → SIF trigger → batch event", the cross-type association constraint "when the liquid phase temperature triggers SIF, the batch status should be marked as abnormal interruption" is automatically discovered.

[0047] Furthermore, the schema is automatically generated based on data element attributes: the processor automatically generates a structured schema for the annotation task according to the attribute definitions in the data element specification system, including: traversing the data element classification system and extracting the {identifier, name, definition, data type, value range} of each data element; automatically matching the annotation model type according to the data type: ENUM → classification model, REAL → numerical regression / extraction model, TS → time series analysis model, STC → multi-task sequence annotation model; automatically generating annotation constraint rule files (JSON schema format) according to the value range; and automatically generating association verification rules according to the relation edges in the knowledge graph.

[0048] refer to Figure 6 The annotation constraint rules are automatically generated, including: Value range constraint rules: Labeled values ​​must be within the value range defined by the data element; out-of-bounds values ​​are automatically blocked. For interval value ranges, the system checks in real time whether the labeled values ​​meet the open / closed interval constraints; for enumerated value ranges, only predefined enumerated code values ​​are allowed. Type constraint rules: Label values ​​must conform to the data type defined in the data element; automatic alarms will be triggered if the types do not match. REAL type checks numerical precision, INT type rejects decimal input, TS type checks ISO 8601 format compliance, and DUR type checks Duration format and non-negativity.

[0049] Association constraint rules: Based on the reference and constraint relationships in the knowledge graph, consistency verification rules for association annotations are automatically generated. For example: the associated tag number in an alarm event must exist in the device master table; the trigger variable tag number in an interlocking action event must match the SIF trigger variable.

[0050] Temporal Constraint Rules: For E-type (event) data elements, time sequence verification rules are automatically generated based on temporal dependency edges in the knowledge graph. For example: alarm cancellation time ≥ alarm generation time; interlock recovery time ≥ trigger time.

[0051] Composite structure constraint rules: For STC type data elements, structural integrity verification rules are generated based on sub-attribute definitions. Each sub-attribute is independently subject to value range constraints and type constraints, and all required sub-attributes must be filled in completely.

[0052] Step S103: Based on the annotation constraint rules, the multi-model collaborative annotation engine performs constrained annotation on the raw process safety data of the DCS / SIS system of the fine chemical production unit. Each annotation result is accompanied by the confidence score of the model output. Automated quality gating is performed based on the confidence score to replace traditional manual verification.

[0053] refer to Figure 7 The processor-driven multi-model collaborative annotation engine performs full-process machine learning annotation on raw process safety data under the constraints of annotation specifications, including multi-model collaborative pre-annotation, confidence-based automated quality gating, multi-model integrated voting, and active learning closed loop.

[0054] In one embodiment of this application, the corresponding annotation model is scheduled according to the data type of the data element to pre-annotate the raw process safety data from the DCS / SIS system of the fine chemical production unit; a model confidence score is calculated for each annotation result, and automated quality gating is performed based on the confidence score: for high-confidence samples with a confidence score not lower than a first threshold, the annotation result is automatically approved; for medium-confidence samples with a confidence score between the first threshold and a second threshold, multiple heterogeneous models are used to independently predict and determine the final annotation result through voting, and the consistency rate between models is calculated as the annotation quality index; for low-confidence samples with a confidence score lower than the second threshold, a hybrid strategy of uncertainty sampling and diversity sampling is used to select samples from the active learning pool for annotation supplementation, and the model weights are incrementally updated.

[0055] In an optional embodiment of this application, the multi-model collaborative annotation engine includes: For enumeration types, a fine-tuned BERT classification model is used, which adds a fully connected classification head to the BERT pre-trained model, uses the enumeration value range code of the data element as the classification label, and is trained using the cross-entropy loss function; For real number types, a temporal anomaly detection model is used, which adopts a two-stage concatenated architecture of Isolation Forest and AutoEncoder, wherein AutoEncoder uses an LSTM codec to learn the temporal reconstruction features of sensor readings under normal operating conditions and outputs a multi-dimensional reconstruction error vector, and Isolation Forest performs anomaly scoring on the multi-dimensional reconstruction error vector, with the values ​​of normal data points directly output as labeled values; For timestamp and duration types, a regular parser is used in conjunction with a temporal logic check finite automaton, where the regular parser matches and extracts the time expression in the text and automatically converts it to ISO. The 8601 standard format uses a finite automaton for temporal logic verification to check whether the extracted results satisfy the temporal constraints between events. For structural types, a multi-task sequence labeling model is used, which adopts an architecture of a shared underlying encoder plus a task-specific CRF decoder head. Each sub-attribute corresponds to an independent CRF decoder head, and BIO label sequences are used for training. For Boolean types, a rule engine and a binary classification model are used in a concatenated architecture. First, the rule engine checks whether there is a clear truth condition. If it is met, the Boolean value is directly output. Otherwise, the model with a binary classification head added to BERT makes a judgment based on the contextual semantics.

[0056] Furthermore, the multi-model collaborative pre-labeling engine automatically schedules the corresponding labeling model to pre-label the original data based on the data element's data type. Each data type is configured with a dedicated labeling model, forming a multi-model collaborative labeling architecture. ENUM Type Labeling – A Classification Model Based on Fine-tuned BERT: Based on the BERT-base-chinese pre-trained model, a fully connected classification head (Hidden_size → num_classes) is added to form a text classification model. The input is the original data context text (such as the cause description paragraph in a HAZOP report, or the event description field in an alarm log), and the output is the classification label (corresponding to the code value in the enumerated value range) and the confidence score P(y=k|x)∈[0,1]. The training method uses labeled industrial process safety corpora as training data, with the enumerated value range code of the data elements as labels, and fine-tunes the model parameters using the cross-entropy loss function.

[0057] REAL Type Labeling – A Numerical Extraction Model Based on Temporal Anomaly Detection: This model employs a two-stage cascaded architecture combining an Isolation Forest anomaly detector and an AutoEncoder reconstruction model. The AutoEncoder uses an LSTM codec, taking a temporal data window as input and outputting a reconstruction error vector. The Isolation Forest scores the reconstruction error vector for anomalies. The input is a DCS temporal data stream (a sequence of sensor readings organized by sampling period), and the output is the extracted numerical labeling results (including timestamps and values), along with anomaly scores s(x)∈[0,1]. The AutoEncoder learns the temporal patterns of each sensor reading under normal operating conditions. When the reconstruction error exceeds a threshold, the Isolation Forest marks the data point as an anomaly; the values ​​of normal data points are directly output as label values.

[0058] TS / DUR Type Labeling – Based on Regular Expression Parsing + Temporal Logic Verification Algorithm: A deterministic parser based on an extended regular expression engine, combined with a finite automaton for temporal logic verification. Input is a timestamp field or time interval description from log text; output is a standardized timestamp or Duration value conforming to ISO 8601 format. The regular expression parser matches and extracts time expressions from the text (supporting various common formats), automatically converting them to the ISO 8601 standard format; the temporal logic verification automaton checks whether the extracted results satisfy the temporal constraints between events (e.g., elimination time ≥ generation time). Confidence calculation is based on the determinism of the regular expression matching (unique match → 1.0, multiple candidate matches → assigned according to matching degree).

[0059] STC Type Labeling – Sequence Labeling Based on Multi-Task NER Model: This model uses a BiLSTM-CRF or BERT-CRF multi-task named entity recognition model to perform sequence labeling on the sub-attributes of composite data elements. The model adopts a shared low-level encoder + task-specific top-level architecture: a shared BiLSTM / BERT encoder extracts contextual features, and each sub-attribute corresponds to an independent CRF decoder head. The input is unstructured text (such as HAZOP report paragraphs, alarm event descriptions, and interlocking action records), and the output is a BIO tag sequence (B-attribute name, I-attribute name, O) for each sub-attribute, extracting the corresponding text fragment as the sub-attribute value. Training is performed using structured STC data element instances as training data, with the position of the sub-attribute value in the original file labeled as the BIO tag sequence.

[0060] BOOL type labeling – based on a dual judgment of rule engine and classification model: rule engine (hard constraint judgment) + binary classification model (soft semantic judgment) in series architecture. First, the rule engine checks whether there is an explicit truth condition (such as "triggered / not triggered" explicitly recorded in the SIS log). If it is satisfied, a boolean value (confidence 1.0) is directly output; otherwise, the binary classification model (adding a binary classification head on the basis of BERT) makes a judgment based on the context semantics.

[0061] INT type annotation – based on a numerical parsing + range validation model: Numerical extractor (regular expression + named entity recognition) + range validation module. First, numerical values ​​are extracted from the text using regular expressions and the NER model. Then, it is validated whether the values ​​fall within the discrete integer set defined by the data element.

[0062] STR type labeling – based on a regular expression matching + semantic verification model: a regular expression matcher + a semantic verification classifier. Regular expressions (such as HG / T 20505 tag number rules, batch number format PBBA-YYYYMMDD-NNN) are used for matching and extraction, and the semantic verification classifier verifies the consistency between the extracted results and the context.

[0063] Furthermore, automated quality gating based on confidence scores: The processor attaches a confidence score conf∈[0,1] to each labeled result from the model output, and performs automated quality gating based on the confidence score, replacing the traditional manual verification process. High-confidence channel (conf≥θ1, where θ1 is a preset first threshold, e.g., θ1=0.95): Annotation results automatically pass through the quality gate and are directly written to the annotation result database. The processor records annotation metadata, including model identifier, confidence score, and passing time.

[0064] Medium confidence channel (θ2≤conf<θ1, θ2 is a preset second threshold, such as θ2=0.80): Triggers a secondary verification process. The processor scheduler integrates a voting module, using multiple heterogeneous models to independently predict the same annotation task, and uses the voting results as the final annotation. If the voting consistency rate is ≥ the threshold τ_vote (e.g., 0.8), it automatically passes; otherwise, it is marked as "needs fine-tuning".

[0065] Low-confidence channel (conf<θ2): The labeling results are not written yet, but enter the active learning queue and are marked as "training samples to be supplemented". At the same time, the processor outputs a "best guess" label value (with a low-confidence label) based on the current best model, for later reference but not as the final result.

[0066] Furthermore, a multi-model ensemble voting mechanism is implemented: for the same annotation task, the processor uses multiple heterogeneous models to make independent predictions, and the final annotation result is determined through ensemble voting, replacing the traditional cross-review process. For each annotation task, at least three heterogeneous models are configured for independent prediction: model variant A is a BERT-base-chinese fine-tuned model (12 layers, 768-dimensional hidden layers), model variant B is a RoBERTa-wwm-ext fine-tuned model (12 layers, 768-dimensional hidden layers, different pre-training strategies), and model variant C is a RoBERTa-wwm-ext-large fine-tuned model (24 layers, 1024-dimensional hidden layers, larger model capacity). For numerical annotation tasks (REAL / INT), the model variants are replaced with AutoEncoder variants with different hyperparameters (different hidden layer dimensions, different training window sizes).

[0067] Majority voting: For discrete labeling tasks such as ENUM / BOOL / INT, the value that appears most frequently among the predictions of multiple models is taken as the final label. y* = mode(yA, yB, y_C) Where yA, yB, and y_C are the prediction results of models A, B, and C, respectively.

[0068] Weighted voting: When the number of models > 3 or the performance differences between models are significant, weighted voting is used. y* = argmax_k Σᵢ wᵢ · I(yᵢ=k) Where wᵢ is the weight of model i (obtained by normalizing its F1 score on the validation set), and I is the indicator function.

[0069] Weighted average: For continuous annotation tasks such as REAL / TS / DUR, a weighted average is used. v* = Σᵢ wᵢ · vᵢ / Σᵢ wᵢ Model consistency rate calculation: As a quality indicator, the model consistency rate is calculated as follows: agreement = maxk Σᵢ I(yᵢ=k) / nummodels Consistency rate replaces the traditional Cohen's Kappa coefficient as an objective measure of annotation quality. When the consistency rate falls below a threshold τ_low (e.g., 0.6), the processor automatically triggers a model fine-tuning process: using the difference in predictions between the sample and its multiple models as training signals, the model weights are incrementally updated, rather than relying on manual arbitration.

[0070] In an optional embodiment of this application, the first confidence threshold and the second confidence threshold are adaptively adjusted based on the sliding window statistics of the most recent N batches of annotation quality data, where N≥5; when the violation rate of the automatically passed annotation results in subsequent constraint verification is lower than the first violation threshold, the first confidence threshold is gradually increased; when the backlog of low confidence queue samples exceeds the preset capacity limit, the second confidence threshold is appropriately decreased; multiple heterogeneous models include the BERT-base-chinese fine-tuning model, the RoBERTa-wwm-ext fine-tuning model, and the RoBERTa-wwm-ext-large fine-tuning model; for discrete annotation tasks, majority voting or weighted voting is used to determine the final annotation result, and for continuous annotation tasks, weighted averaging is used; the consistency rate between models is defined as the ratio of the number of models corresponding to the most frequently occurring value in the prediction results to the total number of models; when the consistency rate is lower than the consistency rate threshold, the processor automatically triggers the model fine-tuning process, using the prediction difference information of the sample and its multiple models as the training signal increment to update the model weights.

[0071] Furthermore, the active learning loop works as follows: the processor maintains an active learning pool and executes active learning strategies on low-confidence samples to minimize the need for manual annotation. All annotated samples with a confidence level < θ² automatically enter the active learning pool and are stored according to data type and annotation task.

[0072] Sample selection strategy: The processor uses a hybrid strategy of uncertainty sampling and diversity sampling to select the most informative samples from the active learning pool. Uncertainty sampling: Select the sample with the highest model prediction entropy. x* = argmax_x H(p(y|x)) = argmaxx (-Σk p(y=k|x)·log p(y=k|x)) Diversity sampling: The k-center-greedy algorithm is used to ensure the diversity of the selected samples in the feature space and avoid selecting redundant and similar samples.

[0073] Hybrid strategy: In each iteration, select the top α proportion of samples with the highest uncertainty (e.g., α=60%) and the bottom (1-α) proportion of samples with the highest diversity (e.g., 40%).

[0074] Only a small number of samples selected by active learning (less than 5% of the total sample size) are labeled by domain experts as supplementary training data. The labeling behavior of domain experts is an offline operation and does not belong to the steps of the labeling method process. The processor uses the newly labeled samples to perform incremental training on the labeling model, updates the model weights, and then re-predicts the low-confidence samples to improve the subsequent labeling accuracy. After each active learning iteration, the processor evaluates the performance metrics (F1 / Accuracy) of the model on the validation set, and stops the current round of active learning iteration when the performance improvement amplitude is lower than the incremental threshold (e.g., ΔF1<0.01).

[0075] In step S104, the constraint check engine automatically performs multi-dimensional constraint check before the labeling result is written.

[0076] In an embodiment of the present application, the constraint check engine automatically performs multi-dimensional constraint check before the labeling result is written. The multi-dimensional constraint check includes: value range constraint check, type constraint check, sequence constraint check, association constraint check, structure constraint check, and measurement point position code consistency check; wherein the measurement point position code consistency check automatically matches the corresponding alarm threshold and interlock logic according to the measurement point position segment in the data element identifier, when the measurement point position segment indicates a liquid phase position, the alarm threshold and interlock logic corresponding to the liquid phase are matched, and the misuse of the alarm threshold of the liquid phase position for the data of the jacket position is rejected.

[0077] Further, the value range constraint check: the value range checker program automatically checks whether the labeled value is within the value range defined by the data element. For the interval value range, check a<x<b or a≤x≤b; for the enumeration value range, check whether the labeled value is in the predefined enumeration set; for the discrete set value range, check whether the labeled value is one of the set elements.

[0078] Type constraint check: the type check engine automatically checks whether the data type of the labeled value is consistent with the definition of the data element. The REAL type checks the numerical precision and dimension, the INT type rejects decimals, the TS type checks the ISO 8601 format, the DUR type checks the Duration format and non-negativity, and the BOOL type only allows True / False.

[0079] Sequence constraint check: the sequence logic check program automatically checks the chronological relationship between events. The check rules are automatically generated from the sequence dependency edges of the knowledge graph. For example: alarm clearance time ≥ alarm generation time; interlock recovery time ≥ trigger time; maintenance work order completion time ≥ actual start time.

[0080] Association constraint verification: The knowledge graph inference engine automatically verifies the consistency of associations across data elements. Verification rules are automatically generated from the reference and constraint relationships of the knowledge graph. For example: the associated tag number in an alarm event must exist in the device master table; the SIF trigger variable must be consistent with the trigger variable in the SIF definition.

[0081] Structural constraint validation: The schema validator automatically checks the structural integrity of STC type data elements. It verifies whether all required sub-attributes are filled in and whether the value of each sub-attribute satisfies its independent data type and value range constraints.

[0082] BB segment location code consistency check: The location code verifier automatically checks the consistency between the BB segment code in the annotation result and the alarm threshold and interlocking logic. For example, when BB=01 (liquid phase) in the data element identifier, it automatically matches the alarm threshold and interlocking logic corresponding to the liquid phase temperature. It is strictly forbidden to misuse the alarm threshold of the liquid phase temperature for data with BB=03 (jacket).

[0083] In an optional embodiment of this application, the range constraint verification includes checking whether the labeled value of the interval range satisfies the open or closed interval constraint, checking whether the labeled value of the enumerated range is in the predefined enumeration set, and checking whether the labeled value of the discrete set range is one of the set elements; the type constraint verification includes checking the numerical precision and dimension of real number types, rejecting decimal input for integer types, and verifying ISO for timestamp types. 8601 format compliance and time zone consistency; Duration type validation: Duration format and non-negativity are checked; Boolean type only allows true or false; Time sequence constraint validation: Time sequence constraint validation rules are automatically generated based on the time sequence dependency relationship edges of the knowledge graph, including: alarm cancellation time is not earlier than alarm generation time, interlock recovery time is not earlier than trigger time, and safety instrument function trigger time is within the corresponding work section time window; Association constraint validation: Cross-data element association consistency validation rules are automatically generated based on the reference relationship and constraint relationship edges of the knowledge graph, including: the associated tag number in the alarm event must exist in the equipment master table, the safety instrument function trigger variable must be consistent with the trigger variable in the safety instrument function definition, and the recipe temperature setpoint should be lower than the alarm high limit of the corresponding measuring point location; Structural constraint validation: This includes checking whether all required sub-attributes of the structural type data element have been filled in, whether the value of each sub-attribute meets its independent data type and value range constraints, and rejecting writing when required sub-attributes are missing and triggering model re-inference.

[0084] Step S105: Output the annotation results as data element instances.

[0085] refer to Figure 8The annotation results are output in the form of data element instances. Each annotation record contains: a data element identifier (inherited from step S101, including BB segment position encoding information); a reference to the original data (timestamp / bit number / source pointing to the original data); an annotation value (a legal value that meets the value range constraints and whose format conforms to the data type requirements); and annotation metadata, including the annotation model identifier (the name and version of the model that generated the annotation result), the model confidence score (the original confidence value output by the model), the voting result (the predicted values ​​and consistency rates of each model in the ensemble voting), the constraint verification status (pass / reject / conditional pass), the annotation timestamp (TS format), and the quality gating channel identifier (high confidence / medium confidence ensemble voting / active learning).

[0086] The output format is JSON / Parquet. The schema is directly generated by the data element specification system and does not require additional definition. The schema is automatically generated according to the following rules: each data element identifier corresponds to a JSON field name; data types are mapped to the JSON schema's type (REAL→number, INT→integer, STR→string, BOOL→boolean, ENUM→string+enum, TS→string+format:datetime, DUR→string+pattern, ARR→array, STC→object); and value ranges are mapped to the JSON schema's constraints (range value range→minimum / maximum, enumerated value range→enum array).

[0087] Furthermore, integrity verification: Verify annotation coverage based on the data element specification classification system. Check if there is labeled data under each primary / secondary category; calculate the annotation completion rate for each type of data element; for STC type data elements, check the completeness of sub-attribute annotations (whether all required sub-attributes are filled in); issue alerts for categories with annotation completion rates below a set threshold.

[0088] Further, consistency verification: Based on the association constraint rules and timing constraint rules generated in step S102, batch consistency verification is performed on the labeled result set, and a conflict list is output. This includes cross-data element association consistency (whether the associated tag number in the alarm event exists in the device master table), timing logic consistency (whether the timestamps in the same event chain meet the chronological order), and value domain logic consistency (whether the relationship between the calculated value and the input value meets the physical constraints, such as the commissioning rate = 1 - bypass duration / statistical period duration).

[0089] Step S106: Based on the data and model performance metrics of the annotation execution process, jointly iteratively optimize the data element specification system and the multi-model collaborative annotation engine.

[0090] refer to Figure 9 In one embodiment of this application, the data element specification system and the multi-model collaborative annotation engine are jointly iteratively optimized, including: automatic calibration of the value range based on the actual distribution of the labeled values, automatic supplementation of the enumeration based on the distribution predicted by the classification model, automatic discovery of association rules based on frequent pattern mining, automatic expansion of the measurement point location segment based on cluster analysis, online updating of the model based on efficient parameter fine-tuning, and automatic monitoring of annotation quality based on distribution drift detection.

[0091] Furthermore, value range calibration: Based on the actual distribution of the labeled results, the processor automatically calibrates the value range definition of the data elements. If the labeled values ​​frequently touch the value range boundary (>5% of samples are within the 5% boundary range), the value range is automatically expanded; if the value range is much larger than the actual distribution range (value range span / actual span > 10), the value range is automatically narrowed. After the value range is adjusted, the processor automatically updates the corresponding parameters in the constraint verification engine.

[0092] Furthermore, the enumeration is supplemented as follows: For ENUM type data elements, the processor automatically detects uncovered categories based on the predicted distribution of the classification model. If the predicted probability distribution of the classification model for some samples shows a uniform distribution (information entropy close to the theoretical maximum value), it indicates that there may be undefined categories in the value range. Frequently occurring undefined categories are automatically included as candidate enumeration values, and when the frequency of a candidate value exceeds a threshold, it is automatically added to the value range.

[0093] Furthermore, association rule mining: based on the statistical associations of the annotation results, the processor uses Apriori or FP-Growth algorithms to discover implicit constraint relationships between data elements and supplement them into the annotation specification.

[0094] Furthermore, BB segment expansion: If, during annotation practice, new measurement locations with different security semantics are found under the same secondary category, the processor identifies data clusters not covered by existing BB values ​​based on anomaly detection models (such as DBSCAN clustering), prompts for the addition of new BB values, and records their security semantic descriptions.

[0095] Furthermore, the model undergoes online learning and incremental training: the processor continuously collects the confidence distribution and constraint verification results during the annotation process as real-time monitoring metrics for model performance. When the average confidence of the model for a certain annotation task continues to decline (the mean of the sliding window is below the threshold θ_decay, such as 0.85), the incremental training process is automatically triggered. The incremental training data comes from samples already annotated by domain experts in the active learning pool. An efficient parameter fine-tuning mechanism is used for incremental learning. While freezing the main weights of the original model, only a small number of newly added trainable parameter adapters are trained, avoiding catastrophic forgetting while significantly reducing training computational overhead.

[0096] In a specific embodiment, the parameter-efficient fine-tuning adopts the low-rank adaptation method: for the original model weight matrix W∈R^(d×k), low-rank decomposition matrices B∈R^(d×r) and A∈R^(r×k) are introduced, where r<<min(d,k) (typical value r=4~32). During forward propagation, W'=W+BA is calculated. During the training process, W is frozen and only A and B are trained, and the number of trainable parameters is only 0.1%~1% of that of the original model; when new data arrives, the LoRA adapter for the new task is trained and used in combination with the historical adapter.

[0097] Optionally, the parameter-efficient fine-tuning can also adopt alternative methods such as QLoRA (quantized LoRA, which trains the LoRA adapter on the 4-bit quantized base weight to further reduce video memory occupation), AdaLoRA (adaptive rank allocation, which dynamically allocates the rank r of different adapter pairs according to parameter importance), and Prefix Tuning (adding trainable prefix vectors before each layer of attention mechanism). After incremental training, the processor evaluates the model performance on the verification set, and updates the model version in the production environment only when the performance is improved.

[0098] Further, annotation quality monitoring based on distribution drift detection: the processor establishes a distribution drift detection mechanism for the model output confidence distribution of the annotation result, and monitors whether the confidence distribution of the annotation result has a significant shift through a statistical test method, so as to directly reflect the change of the annotation quality.

[0099] In a specific embodiment, the distribution drift detection adopts the Kolmogorov-Smirnov test: taking the annotation confidence distribution of the latest N batches (e.g., N=10) as the reference distribution Fref(x), the KS statistic D=max|Fref(x)-Fcur(x)| is calculated for the confidence distribution Fcur(x) of the current batch. When D exceeds the critical value corresponding to the preset significance level (e.g., D_crit depends on the sample size when α=0.05), it is determined that a significant drift has occurred.

[0100] Optionally, the distribution drift detection can also adopt alternative statistical methods such as Population Stability Index (PSI, a difference measure based on the frequency distribution after binning, PSI>0.25 indicates a significant shift), Maximum Mean Discrepancy (MMD, a distribution distance measure based on kernel methods), or combine multiple methods to form a voting mechanism. When a significant drift of the confidence distribution is detected, the following responses are automatically triggered: annotation alarm — marking the annotation result of the current batch as "confidence drift" and triggering manual review; model retraining — automatically triggering the incremental training process to update the model parameters with the latest annotation data. The reference distribution parameters for drift detection are refreshed synchronously with value range calibration and model version updates.

[0101] To more clearly illustrate the technical effects of the technical solution in this application, the following explanation will be based on a specific example of a "40-ton / year PBBA fine chemical production unit".

[0102] PBBA (Pentabromobenzyl Acrylate), CAS No. 59447-55-1, molecular formula C 10 H5Br5O2, with a molecular weight of 556.67, is a high-performance flame-retardant monomer. The plant is designed to have an annual capacity of 40 tons, employing an intermittent batch production process, with each batch producing approximately 500 kg of finished product. The production process consists of three stages: Stage 1 is the bromination reaction, carried out in a 500L glass-lined reactor R-101; Stage 2 is the esterification reaction, carried out in a 1000L stainless steel reactor R-201; Stage 3 is separation and purification, involving cooling and crystallization in a crystallization vessel C-101, filtration and separation in a centrifuge F-101, and drying in a vacuum dryer D-101 to obtain the finished product.

[0103] The unit is equipped with an independent SIS safety instrumented system, physically separated from the DCS, and has a total of 4 SIF loops: SIF-01 is a high-temperature interlock for bromination reaction → emergency cooling (SIL 2, trigger variable R-101 liquid phase temperature TI-101A ≥ 100°C); SIF-02 is a high-pressure interlock for reactor → emergency venting (SIL 1); SIF-03 is an interlock for HBr absorption tower → alkali spray (SIL 1); SIF-04 is a high-concentration combustible gas interlock → emergency ventilation + power cut-off (SIL 1).

[0104] A data element specification system covering 7 major categories and 37 core data elements (A / B / C / D / E / G / H) was constructed for the PBBA device. The safety semantic distinction of the BB segment position encoding is demonstrated using the bromination reactor temperature as an example. ZCT-A101T-001 (BB=01, liquid phase / bottom): Directly reflects the bromination reaction status. Overheating may cause the reaction to run out of control and overflow. Alarm high limit is 95°C. ≥100°C triggers SIF-01 emergency cooling.

[0105] ZCT-A103T-001 (BB=03, Jacket / Outer Shell): Reflects the state of the heating / cooling medium. Overheating may damage the enamel layer or cause heat transfer oil leakage. Alarm high limit is 120°C. There is no direct interlock, and the operation is triggered by intervention.

[0106] Although both are "reactor temperature" and belong to the same A1 category (reaction process variable), the alarm thresholds for liquid phase temperature and jacket temperature are different (95°C vs 120°C), their interlocking logic is different (triggered SIF-01 vs no interlock), and their safety consequences are different (reaction runaway vs equipment damage). Therefore, they are defined as independent data elements at the coding level. If only the same data element "reactor temperature" (BB=00 default segment) is used, it will be impossible to distinguish the differentiated safety constraints of the two, and the data from different locations will be mixed together during annotation.

[0107] During the annotation execution phase, taking the abnormal batch PBBA-20260615-003 as an example, this batch experienced an abnormal temperature rise during the bromination reaction stage, ultimately triggering the SIF-01 interlock. The processor extracts data from each measurement point of R-101 from the DCS historical database according to the BB segment code. The AutoEncoder model (LSTM 256→64→256, training window 64 seconds) reconstructs and detects anomalies in the time-series data of each measurement point. Liquid phase temperature (BB=01): At 02:15:30, the AutoEncoder outputs a reconstructed value of 96.5°C (actual value 96.8°C), with a reconstruction error of 0.3°C (within normal range). The model-extracted labeled value is 96.8°C. The constraint verification engine automatically matches the alarm thresholds for BB=01: low alarm at 70°C, high alarm at 95°C, and high-high interlock at 100°C. 96.8°C > 95°C → the constraint verification engine automatically labels the alarm status as "H: High", with a confidence level of 0.99. At 02:18:45, the actual value is 100.2°C > 100°C → the alarm status is automatically labeled as "HH: High-High", and marked "Triggered SIF-01", with a confidence level of 0.99.

[0108] Jacket temperature (BB=03): At the same time 02:15:30, the actual value was 105.2°C. The constraint verification engine automatically matched the alarm threshold for BB=03: low alarm 50°C, high alarm 120°C. 105.2°C < 120°C → automatically marked alarm status as "NORMAL", confidence level 0.97.

[0109] Comparison results: If the BB segment is not distinguished and the liquid phase temperature and jacket temperature are treated as the same data element, then 96.8°C and 105.2°C will use the same alarm threshold, leading to misjudgment—the jacket temperature of 105.2°C will be incorrectly labeled as "HH: High High" alarm (because the high high alarm threshold for liquid phase temperature is 100°C), while the actual jacket temperature of 105.2°C is within the normal range (the high alarm threshold for jacket is 120°C).

[0110] During the constraint verification phase, the constraint verification engine automatically detects and rejects the following constraint violations: BB segment mismatch (if the model labels the jacket temperature of 105.2°C as "HH: High High" alarm, the constraint verification engine refuses to write it and outputs "ZCT-A103T-001 (BB=03 jacket) high alarm threshold is 120°C, current value 105.2°C is not exceeded"); timing contradiction (if the model outputs the SIF-01 trigger time is not within the section time window, the constraint verification engine rejects it); safety instrument association violation (if the model outputs the SIF-01 commissioning status as "bypass" but there is a trigger event, the constraint verification engine rejects it); recipe-control mismatch (if the model outputs the recipe temperature setpoint is higher than the alarm high limit, the constraint verification engine rejects it).

[0111] During the joint iterative optimization phase, the processor performed range calibration on the Br2 feeding rate (ZCT-A101F-001): 3.2% of the samples in the AutoEncoder model's labeled values ​​touched the upper limit of the range, 50.0 kg / h. Distribution drift detection analysis confirmed that the actual feeding rate could reach 55 kg / h. The processor automatically expanded the range from [0.0, 50.0] to [0.0, 60.0] kg / h and simultaneously updated the parameters in the constraint verification engine. The processor performed incremental training on the HAZOP cause classification model: the model's average confidence score of 0.87 was lower than θ_decay=0.90, triggering incremental training. The LoRA method (rank r=8) was used to incrementally train the BERT classification model, freezing the original BERT weights. A LoRA adapter was added to the Q and V projection matrices of the self-attention layer, and only the adapter parameters (approximately 0.3% of the total parameters) were trained. After incremental training, the F1 score improved from 0.82 to 0.89. The processor performs a KS test on the output confidence distribution of the HAZOP cause classification model: the KS statistic D = 0.23 is calculated, and the critical value D_crit = 0.19 (n = 120) at a significance level of α = 0.05. Since D = 0.23 > D_crit = 0.19, it is determined that the confidence distribution of the current batch has drifted significantly, and the annotation alarm and model retraining are automatically triggered.

[0112] Through the complete implementation of this embodiment, the PBBA production unit completes structured annotation based on machine learning for 37 data element instances across 7 categories. Six annotation tasks and their constraint rules are automatically generated from the data element specification system through semantic matching of pre-trained language models and relational reasoning from knowledge graphs, eliminating the need for manual definition of annotation labels and constraints. A multi-model collaborative pre-annotation engine covers 10 data types, and a three-level quality gating system based on confidence levels achieves fully automated annotation for over 95% of samples. An active learning mechanism reduces the need for manual annotation to less than 5%. A constraint verification engine automatically detects and rejects constraint violations such as BB segment mismatches, timing contradictions, and safety instrument association violations. Annotation results are output in JSON format as data element instances, with built-in standardized encoding, schema, and model metadata, allowing direct integration with intelligent inspection systems, anomaly warning engines, and compliance audit platforms.

[0113] To address the aforementioned technical problems, a second embodiment of this application provides a data element-driven fine chemical process safety data annotation device to solve the same technical problems as the method embodiment. This device may include the following modules: The data element specification construction module is used to build a data element specification system for safety in fine chemical processes; The annotation specification generation module is used to build an annotation model based on a pre-trained language model and a data element knowledge graph. It uses the data element specification system for safety of fine chemical processes as a prior knowledge base and constraint input for machine learning to generate a structured schema for the annotation task, and generates annotation constraint rules based on the structured schema of the annotation task. The multi-model collaborative annotation module is used to drive the multi-model collaborative annotation engine to perform constrained annotation on the raw process safety data of the DCS / SIS system of fine chemical production unit based on annotation constraint rules. Each annotation result is accompanied by a confidence score output by the model. Automated quality gating is performed based on the confidence score to replace traditional manual verification. The annotation results are output in the form of data element instances. The joint iterative optimization module is used to perform joint iterative optimization of the data element specification system and the multi-model collaborative annotation engine based on the data and model performance indicators of the annotation execution process.

[0114] It is understood that the above-described device embodiments correspond to the method embodiments. Therefore, the device embodiments have all the beneficial effects of the method embodiments, and will not be repeated here.

[0115] The above are merely preferred embodiments of this application and do not limit the patent scope of this application. Any equivalent structural or procedural transformations made using the content of this application's specification and drawings, or direct or indirect applications in other related technical fields, are similarly included within the patent protection scope of this application.

Claims

1. A data-element-driven method for labeling safety data in fine chemical processes, executed by a processor, characterized in that: include: Construct a data element standard system for safety in fine chemical processes; A labeling model is constructed based on a pre-trained language model and a data element knowledge graph. The data element specification system for safety of fine chemical processes is used as a prior knowledge base and constraint input for machine learning to generate a structured schema for the labeling task. Labeling constraint rules are generated based on the structured schema for the labeling task. A multi-model collaborative annotation engine driven by annotation constraint rules performs constrained annotation on the raw process safety data of the DCS / SIS system of fine chemical production unit. Each annotation result is accompanied by a confidence score of the model output. Automated quality gating is performed based on the confidence score to replace traditional manual verification. The annotation results are output as data element instances; Based on the data and model performance metrics of the annotation execution process, the data element specification system and the multi-model collaborative annotation engine are jointly iteratively optimized.

2. The method as described in claim 1, characterized in that, The annotation model, constructed based on a pre-trained language model and a data element knowledge graph, uses the fine chemical process safety data element specification system as a prior knowledge base and constraint input for machine learning to generate a structured scheme for the annotation task, including: Using a pre-trained language model fine-tuned with corpus from the industrial process safety domain, the data element definition text and the original process safety data fragments to be labeled are encoded into semantic vectors respectively. The mapping relationship between the labeling task and the data element is automatically established through cosine similarity calculation. A data element relationship knowledge graph is constructed using data elements as nodes and semantic relationships between data elements as edges. A graph attention network is used for relationship reasoning to automatically discover implicit association constraints between data elements. The structured schema for annotation tasks is automatically generated based on the data type and value range attributes of data elements.

3. The method as described in claim 1, characterized in that, A multi-model collaborative annotation engine driven by annotation constraint rules performs constrained annotation on the raw process safety data of the DCS / SIS system of a fine chemical production unit. Each annotation result is accompanied by a confidence score from the model output. Automated quality gating is then performed based on these confidence scores, including: Based on the data type of the data element, the corresponding annotation model is scheduled to pre-annotate the raw process safety data from the DCS / SIS system of the fine chemical production unit; For each labeled result, calculate the model confidence score, and perform automated quality gating based on the confidence score: Among them, for high-confidence samples with a confidence level not lower than the first threshold, their annotation results are automatically approved; For medium-confidence samples with confidence levels between the first and second thresholds, multiple heterogeneous models are used to independently predict and the final labeling results are determined by voting. The consistency rate between models is calculated as a labeling quality indicator. For low-confidence samples with confidence levels below the second threshold, a hybrid strategy of uncertainty sampling and diversity sampling is used to select samples from the active learning pool for labeling and supplementation, and the model weights are incrementally updated.

4. The method as described in claim 1, characterized in that, Also includes: The constraint validation engine automatically performs multi-dimensional constraint validation before writing the annotation results. Multi-dimensional constraint validation includes: Value range constraint verification, type constraint verification, timing constraint verification, correlation constraint verification, structural constraint verification, and measurement point location code consistency verification; The consistency verification of the measuring point location code automatically matches the corresponding alarm threshold and interlocking logic based on the measuring point location segment in the data element identifier. When the measuring point location segment indicates the liquid phase position, it matches the corresponding alarm threshold and interlocking logic, and refuses to use the alarm threshold of the liquid phase position as the data of the jacket position.

5. The method as described in claim 1, characterized in that, Joint iterative optimization of the data element specification system and multi-model collaborative annotation engine, including: Automatic calibration of the value range based on the actual distribution of labeled values, automatic supplementation of enumeration based on the distribution predicted by the classification model, automatic discovery of association rules based on frequent pattern mining, automatic expansion of the measurement point location segment based on cluster analysis, online updating of the model based on efficient parameter fine-tuning, and automatic monitoring of labeling quality based on distribution drift detection.

6. The method as described in claim 1, characterized in that, The pre-trained language model is BERT-base-chinese or RoBERTa-wwm-ext-large, and the semantic vector is 768-dimensional or 1024-dimensional. The mapping threshold for cosine similarity is not lower than 0.

85. When the highest similarity exceeds the mapping threshold, a mapping is automatically established. When the highest similarity is lower than the mapping threshold, it is marked as a mapping to be confirmed and enters the active learning process. The pre-trained language model is used for domain-adaptive fine-tuning using a corpus of fine chemical process safety. The fine-tuning corpus includes IEC 61511 standard text, ISA-18.2 standard text, HAZOP analysis report corpus, DCS operation log corpus, and functional safety assessment report corpus.

7. The method as described in claim 1, characterized in that, The edge relationship types of the data element relation knowledge graph include reference relationships, temporal dependency relationships, triggering relationships, composition relationships, and association constraint relationships; The graph attention network takes the initial features and edge relationship type encoding of the data elements as input, and after K layers of graph attention propagation, it obtains the enhanced representation of each data element, where K≥2; Based on enhanced representation, the strength of the association constraint between data elements is calculated, and association constraint rules are automatically established when the strength of the association constraint exceeds the constraint threshold. The implicit association constraints include cross-data element type association constraints, including cross-type constraints that the batch status should be marked as abnormal interruption after the process variable triggers the safety instrument function, and cross-type constraints that associate the safety instrument function triggering event when the combustible gas concentration exceeds the standard.

8. The method as described in claim 1, characterized in that, The data types in the annotation constraint rules include integers, real numbers, booleans, enumerations, timestamps, durations, and structures; The multi-model collaborative annotation engine driven by annotation constraint rules includes: For enumerated types, a fine-tuned BERT classification model is used. The fine-tuned BERT classification model adds a fully connected classification head to the BERT pre-trained model, uses the enumerated value range code of the data element as the classification label, and is trained using the cross-entropy loss function. For real number types, a time-series anomaly detection model is used. The time-series anomaly detection model adopts a two-stage cascaded architecture of Isolation Forest and AutoEncoder. The AutoEncoder uses an LSTM codec to learn the time-series reconstruction features of sensor readings under normal operating conditions and outputs a multi-dimensional reconstruction error vector. Isolation Forest scores the multi-dimensional reconstruction error vector for anomalies, and the values ​​of normal data points are directly output as labeled values. A regular expression parser is used in conjunction with a finite automaton for timing logic checks on timestamp and duration types. The regular expression parser matches and extracts the time expressions in the text and automatically converts them into the ISO 8601 standard format. The finite automaton for timing logic checks checks whether the extraction results meet the timing constraints between events. A multi-task sequence labeling model is used for the structure type. The multi-task sequence labeling model adopts an architecture of shared low-level encoder plus task-specific CRF decoder head. Each sub-attribute corresponds to an independent CRF decoder head, and BIO tag sequences are used for training. For Boolean types, a rule engine and a binary classification model are used in a chain architecture. First, the rule engine checks whether there is an explicit truth condition. If it is met, the Boolean value is output directly. Otherwise, the model with a binary classification head added to BERT makes a judgment based on the context semantics.

9. The method as described in claim 3, characterized in that, The first confidence threshold and the second confidence threshold are adaptively adjusted based on the sliding window statistics of the annotation quality data of the most recent N batches, where N≥5; When the violation rate of automatically approved annotation results in subsequent constraint verification is lower than the first violation threshold, the first confidence threshold is gradually increased; when the backlog of low confidence queue samples exceeds the preset capacity limit, the second confidence threshold is appropriately reduced. Among them, the multiple heterogeneous models include the BERT-base-chinese fine-tuning model, the RoBERTa-wwm-ext fine-tuning model, and the RoBERTa-wwm-ext-large fine-tuning model; For discrete annotation tasks, majority voting or weighted voting is used to determine the final annotation result, while for continuous annotation tasks, a weighted average is used. The consistency rate between models is defined as the ratio of the number of models corresponding to the most frequent value in the prediction results to the total number of models. When the consistency rate is lower than the consistency rate threshold, the processor automatically triggers the model fine-tuning process, using the prediction difference information of the sample and its multiple models as training signals to incrementally update the model weights.

10. The method as described in claim 4, characterized in that, Value range constraint verification includes checking whether the labeled value of the interval value range satisfies the open interval or closed interval constraint, checking whether the labeled value of the enumerated value range is in the predefined enumeration set, and checking whether the labeled value of the discrete set value range is one of the set elements. Type constraint validation includes checking the numerical precision and dimensions of real number types, rejecting decimal input for integer types, validating ISO 8601 format compliance and time zone consistency for timestamp types, validating the Duration format and non-negativity for duration types, and allowing only true or false for Boolean types. The timing constraint verification is based on the timing dependency relationship edges of the knowledge graph, which automatically generate the timing order verification rules, including: alarm cancellation time is not earlier than alarm generation time, interlock recovery time is not earlier than trigger time, and safety instrument function trigger time is within the corresponding work section time window; The association constraint verification is based on the reference relationship and constraint relationship edge of the knowledge graph to automatically generate cross-data element association consistency verification rules, including that the associated tag number in the alarm event must exist in the device master table, the trigger variable of the safety instrument function must be consistent with the trigger variable in the safety instrument function definition, and the formula temperature set value should be lower than the alarm high limit of the corresponding measuring point location. Structural constraint validation includes checking whether all required sub-attributes of the structure type data element have been filled, whether the value of each sub-attribute satisfies its independent data type and value range constraints, and rejecting writing when a required sub-attribute is missing and triggering model re-inference.

Images