An automatic processing method of a manufacturing system based on AI interaction with structured data

Abstract

This invention relates to the field of industrial internet and artificial intelligence application technology, and discloses an automatic processing method for manufacturing systems based on AI and structured data interaction. The method includes: acquiring and encapsulating business data from a target enterprise management system; recording initial state values ​​and extracting key names to generate a set of valid key spaces; converting the data into standardized JSON data and inputting it into a large language model for reasoning to obtain a decision and an initial set of key factors including a comprehensive attention score; finding the intersection of the field key names in the set of valid key spaces and the initial set of key factors to output a cleaned set of effective attribution factors; acquiring real-time state values ​​and calculating an attention-weighted state drift index based on the comprehensive attention score; determining whether the index exceeds a dynamic drift tolerance threshold; if not, executing business operations; otherwise, updating the standardized JSON data for secondary reasoning or transferring it to a manual review queue.

Description

Technical Field

[0001] This invention relates to the field of industrial internet and artificial intelligence application technology, specifically to an automated processing method for manufacturing systems based on AI and structured data interaction. Background Technology

[0002] As the manufacturing industry becomes increasingly digitalized, some companies are introducing large language models to automate the approval and processing of various business documents. However, existing automation methods have significant technical limitations when interacting with underlying industrial systems.

[0003] Because generative language models rely on probability distributions to generate content, their outputs have a degree of uncertainty. When processing structured business data, the model may generate fields that are not actually defined in the target enterprise management system database. If intermediate execution components do not impose boundary constraints on this divergent output and directly accept it, sending requests containing illegal fields to the underlying database, it can lead to the failure of structured query commands or even system crashes, making it difficult to guarantee the security of the core business system in the manufacturing industry.

[0004] In high-concurrency scenarios of industrial production, underlying business data changes frequently. Existing automated processing systems generally employ strict data locking or indiscriminate numerical comparison mechanisms when controlling transaction concurrency. As soon as the system detects any physical state change of a business document during automated processing, it triggers a transaction rollback to prevent dirty writes. This mechanism fails to differentiate the impact of different data fields on the final decision based on specific business scenarios. This results in even reasonable state changes to non-core business data causing invalid process circuit breakers, reducing the system's automated instruction throughput.

[0005] Furthermore, to confirm whether business data has changed during processing, existing systems typically use a full table re-reading query to verify all fields of the document when performing concurrent status checks. When faced with massive concurrent business requests, this untargeted full data read consumes a large amount of connection pool resources of the target database and causes unnecessary network bandwidth consumption during cross-system communication. Summary of the Invention

[0006] To address the shortcomings of existing technologies, this invention provides an automated processing method for manufacturing systems based on AI and structured data interaction. This method solves the problems of out-of-bounds queries and system crashes caused by illegal fields output by generative language models, as well as the problems of frequent invalid rollbacks and system resource consumption caused by the lack of data importance differentiation and the use of full data reading in traditional concurrency control mechanisms.

[0007] To achieve the above objectives, this invention provides an automated processing method for manufacturing systems based on AI and structured data interaction, comprising the following steps:

[0008] The business data of the target enterprise management system is obtained, encapsulated, and the initial state values ​​of each field are recorded. The key names are then extracted to generate a set of valid key spaces.

[0009] After converting business data into standardized JSON data, it is input into a large language model for inference to obtain decision results and an initial set of key factors. The initial set of key factors includes field key names and corresponding comprehensive attention scores.

[0010] Extract the execution instructions from the decision results, calculate the intersection of the field key names in the set of legal key space and the set of initial key factors, and output the cleaned set of effective attribution factors.

[0011] Obtain the real-time state value of the corresponding field in the effective attribution factor set, and calculate the attention-weighted state drift index between the initial state value and the real-time state value by combining the comprehensive attention score.

[0012] Determine whether the attention-weighted state drift index is greater than the dynamic drift tolerance threshold. If it is not greater than the dynamic drift tolerance threshold, complete the business operation according to the execution instruction. If it is greater than the dynamic drift tolerance threshold, update the standardized JSON data with the real-time state value for secondary reasoning, or forcibly transfer the business document to the manual review queue.

[0013] Furthermore, the specific steps for obtaining business data from the target enterprise management system, recording the initial state values ​​of each field, and extracting key names to generate a set of valid key spaces are as follows:

[0014] The acquired business data is cleaned and transformed according to preset type rules and static mapping dictionary, and then encapsulated into standardized JSON data.

[0015] Assign a unique transaction identifier to the currently processed business document and record the system time when the data was retrieved as the initial timestamp;

[0016] Traverse the standardized JSON data and extract the key names of all non-empty core fields to generate a set of valid key spaces;

[0017] Bind the unique transaction identifier, initial timestamp, set of valid key spaces, and initial state values ​​of each field at the initial timestamp to generate and cache dynamic execution boundary snapshots.

[0018] Furthermore, after converting the business data into standardized JSON data, it is input into a large language model for inference to obtain decision results and an initial set of key factors. The specific steps are as follows:

[0019] Parse the document type identifier and department information in the standardized JSON data, and retrieve the matching business rule prompt word template from the preset relational database;

[0020] Standardized JSON data is converted into plain text strings and embedded into predefined placeholders in the business rule prompt word template to generate comprehensive reasoning text;

[0021] The inference request, which includes the comprehensive reasoning text and the mandatory attribution constraint parameters, is submitted to the large language model. The decision result generated by the inference of the large language model is obtained. At the same time, the mandatory attribution constraint parameters serve as system-level instruction prompts, requiring the large language model to perform self-attention calculation on the marked lexical units in the comprehensive reasoning text, extract the field key names that play a decisive role in the decision and their corresponding comprehensive attention scores, and output the initial set of key factors in the form of key-value pairs.

[0022] Furthermore, the specific steps for extracting execution instructions from the decision results, calculating the intersection of the field key names in the set of valid key spaces and the initial set of key factors, and outputting the cleaned set of effective attribution factors are as follows:

[0023] Analyze the decision results and extract execution instructions in the form of action identifiers;

[0024] Remove leading and trailing spaces and special escape characters from the elements in the set of valid keys and the set of initial key factors, and normalize the case of the elements.

[0025] The set of valid key spaces after removing leading and trailing spaces, special escape characters, and case normalization is intersected with the initial set of key factors. Illegal fields and their corresponding scores that do not belong to the set of valid key spaces in the initial set of key factors are removed to generate a set of valid attribution factors.

[0026] If the generated set of valid attribution factors is empty, or if the number of elements in the set is less than the minimum association threshold determined by the number of core required fields for different document types in the business dictionary, then the execution instruction is intercepted and the current business document is transferred to the manual review queue; otherwise, the cleaned set of valid attribution factors is output.

[0027] Furthermore, the specific steps for obtaining the real-time status values ​​of the corresponding fields in the effective attribution factor set are as follows:

[0028] When the business document completes the large language model inference and outputs a set of valid attribution factors, an asynchronous check mechanism is triggered and the current time is obtained as the execution timestamp.

[0029] The effective attribution factor set is used as the query parameter, and the key names contained in the set are dynamically concatenated to generate structured query language instructions or application programming interface request parameters as a target probe.

[0030] By using targeted probes, specific numerical values ​​or text content of corresponding fields within the effective attribution factor set at the execution timestamp are requested from the target enterprise management system and recorded as real-time status values.

[0031] Furthermore, before calculating the attention-weighted state drift index between the initial state value and the real-time state value, based on the comprehensive attention score, the independent state drift of each individual field is also calculated:

[0032] For continuous numerical fields, based on the difference between the real-time state value and the initial state value, a minimum normal number to prevent division by zero is introduced as a denominator compensation term to calculate the normalized absolute deviation.

[0033] For discrete state fields, a binary transition judgment is performed by comparing the string hash codes of the real-time state value and the initial state value.

[0034] The normalized absolute deviation and the binary jump discrimination result are uniformly mapped to a preset real number range to generate independent state drift values ​​for each field.

[0035] Furthermore, the specific steps for calculating the attention-weighted state drift exponent between the initial state value and the real-time state value are as follows:

[0036] Extract the preset sensitivity weights for different types of fields from the static business dictionary;

[0037] By introducing a preset weight allocation coefficient, the sensitivity weight is fused with the comprehensive attention score for the corresponding field output by the large language model to construct a joint weight coefficient.

[0038] The global attention-weighted state drift index is calculated by multiplying and summing the joint weight coefficients of each field with the corresponding independent state drift values.

[0039] Further, it is determined whether the attention-weighted state drift index is greater than the dynamic drift tolerance threshold. If it is not greater than the dynamic drift tolerance threshold, the specific steps for completing the business operation according to the execution instruction are as follows:

[0040] Obtain the forced manual lock flag from the target enterprise's management system, and construct a conditional function containing a dynamic drift tolerance threshold and a forced manual lock flag based on the logic and operation principles of Boolean algebra.

[0041] Based on the calculation results of the condition function, the release command is allowed if and only if the attention-weighted state drift index is less than or equal to the dynamic drift tolerance threshold of the document type, and the forced manual lock flag indicates that the data is not occupied by other concurrent processes.

[0042] The extracted execution instructions are serialized into network request messages. A unique transaction identifier generated based on the distributed snowflake algorithm is injected into the message header to control idempotency. The message is then sent to the target enterprise management system to execute the closed loop of business operations.

[0043] Furthermore, if the value exceeds the dynamic drift tolerance threshold, the steps of updating standardized JSON data using real-time status values ​​for secondary inference, or forcibly transferring business documents to a manual review queue, include:

[0044] Intercept the execution instructions generated in the preceding order, trigger the anti-avalanche circuit breaker and rollback mechanism, send a transaction discard message to the target enterprise management system, and forcibly release the previously held dynamic execution boundary snapshot and database connection pool resources;

[0045] Based on interception and forced release, we analyze the specific triggers that cause interception.

[0046] If the specific cause is that the attention-weighted state drift index is greater than the dynamic drift tolerance threshold, then the out-of-bounds field that caused the drift to exceed the limit and the specific deviation value are extracted.

[0047] Update the corresponding fields in the standardized JSON data using the acquired real-time status values ​​to generate updated standardized JSON data;

[0048] The updated standardized JSON data is resubmitted to the large language model for secondary inference;

[0049] If the attention-weighted state drift index corresponding to the result of the second reasoning is still greater than the dynamic drift tolerance threshold, or if the specific cause is that the forced manual lock flag fed back by the target enterprise management system is one, then the automatic processing flow will be completely interrupted and an abnormal warning report will be generated when the circuit is interrupted.

[0050] If the circuit breaker is triggered because the attention-weighted state drift index is greater than the dynamic drift tolerance threshold, the extracted out-of-bounds field, the specific deviation value, and the natural language review opinion output by the large language model will be combined to generate an abnormal warning report.

[0051] If the circuit breaker is triggered by physical table locking in the target enterprise's management system, the basic document information and the table locking conflict log will be concatenated to generate an anomaly warning report.

[0052] Finally, the abnormal warning report is pushed to the abnormal monitoring dashboard, and the abnormal documents are forcibly transferred to the manual review queue.

[0053] Furthermore, it also includes the step of building a closed-loop control system based on complete transaction lifecycle data:

[0054] The business data, the comprehensive reasoning text submitted to the large language model, the cleaned set of effective attribution factors, and the final execution judgment result are packaged into a composite sample record.

[0055] Obtain the manually corrected judgment conclusions and attribution fields from the manual review queue as the true labels;

[0056] Using composite sample records and their corresponding real labels as incremental training corpora, the network weight matrix of the multi-layer self-attention computation block inside the large language model is updated through the backpropagation algorithm to calibrate the attention allocation preference of the large language model.

[0057] This invention provides an automated processing method for manufacturing systems based on the interaction of AI and structured data. It offers the following advantages:

[0058] 1. This invention generates a set of valid key spaces by extracting the actual key names existing in the target system, and performs an intersection operation with the initial set of key factors output by the large language model, thereby eliminating undefined fields that do not belong to the underlying database. This data cleaning mechanism effectively limits the uncertainty of the generative language model output, avoids the system directly issuing instructions containing illegal fields, eliminates the risk of system errors and downtime caused by out-of-bounds queries, and ensures the security of the underlying business system in the manufacturing industry.

[0059] 2. This invention calculates the attention-weighted state drift index between the initial and real-time state values ​​by combining the sensitivity weights of the static business dictionary with the comprehensive attention score output by the large language model, and introduces a dynamic drift tolerance threshold for instruction release determination. This mechanism changes the conventional concurrency control approach that triggers rollback upon data change, enabling the system to intercept data mutations that substantially affect the final decision while tolerating reasonable data deviations. This improves the throughput efficiency of automated processing while preventing concurrent write conflicts.

[0060] 3. During concurrent state verification, this invention uses a cleaned set of valid attribution factors as query parameters to dynamically construct request instructions and target the system to obtain real-time state values ​​of core fields. This targeted acquisition mechanism utilizes the attribution boundaries determined in the model's preceding steps, performing secondary state readings only on fields strongly correlated with the current decision. This avoids full table scans of business documents, thereby reducing database query pressure and network bandwidth consumption during concurrent system verification. Attached Figure Description

[0061] Figure 1 This is a flowchart of the automated processing method for manufacturing business according to an embodiment of the present invention;

[0062] Figure 2 This is a comparison diagram of the multi-dimensional technical effects of embodiments of the present invention;

[0063] Figure 3 This diagram illustrates the verification of the concurrent state drift control mechanism in an embodiment of the present invention. Detailed Implementation

[0064] The technical solutions in the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0065] See attached document Figure 1 This invention provides an automated processing method for a manufacturing system based on the interaction of AI and structured data, comprising the following steps:

[0066] S10, Perform business data extraction and legal key space definition. Based on the unique identifier of the business document, the data extraction and encapsulation module reads basic information, material information, inventory status, amount information, natural language application reason, production progress, quality data, and supplier credit rating fields from the target enterprise management system. The data extraction and encapsulation module records the reading time as the initial timestamp, saves the initial state value of each field at the initial timestamp, and outputs standardized JSON data.

[0067] S20, Execute intelligent decision-making reasoning. The AI ​​interaction module reads standardized JSON data and the corresponding business rule prompt word templates, and constructs a request carrier to be submitted to the large language model. The request carrier contains a mandatory attribution constraint parameter, which instructs the large language model to output the field names that play a decisive role in its decision and their corresponding comprehensive attention scores while outputting the decision result, generating an initial set of key factors containing the above key-value pair mappings;

[0068] S30, Execution Result Parsing and Factor Cleaning. The result parsing module verifies the JSON results returned by the large language model. It extracts decision states including pass, rejection, or warning, extracts review opinions in natural language form, and extracts execution instructions in action identifier form. The result parsing module calculates the intersection of the field keys in the valid key space set and the initial key factor set, removes fields that do not belong to the valid key space set and their corresponding scores, and outputs the cleaned set of valid attribution factors.

[0069] S40 executes asynchronous polling and state drift calculation. The scheduled task execution module triggers an asynchronous check mechanism at the execution timestamp. The scheduled task execution module uses the set of effective attribution factors as query parameters to request the real-time state value at the execution timestamp from the target enterprise management system. The scheduled task execution module combines the sensitivity weights preset in the business dictionary to calculate the attention-weighted state drift index between the initial state value and the real-time state value;

[0070] S50 performs heterogeneous lock adaptation and closed-loop control. The scheduled task execution module determines whether the attention-weighted state drift index is greater than the system's preset dynamic drift tolerance threshold. If it is not greater than the dynamic drift tolerance threshold, the scheduled task execution module calls the open interface of the target enterprise management system or applies for a mutex lock in the middle layer and completes the business operation according to the execution instructions. If it is greater than the dynamic drift tolerance threshold, the system will update the corresponding fields in the standardized JSON data using real-time state values ​​and resubmit the updated business document to the AI ​​interaction module for secondary reasoning; if the secondary reasoning result is still determined to be greater than the dynamic drift tolerance threshold, or if a physical table lock is encountered in the target enterprise management system, the system will forcibly transfer the abnormal document to the manual review queue. The database and process update module stores all processing records at the end of the process.

[0071] See attached document Figure 1 Regarding the business data extraction and legal key space definition in step S10, the specific sub-steps include:

[0072] In this embodiment, to ensure data consistency across heterogeneous systems at different levels, the data extraction and encapsulation module establishes a communication connection with the target enterprise management system and acquires business data. The target enterprise management system encompasses an Enterprise Resource Planning (ERP) system, Manufacturing Execution System (MRP) system, Manufacturing Operations Management System (MOO) system, Manufacturing Collaboration Platform, Supply Chain Management System (SLM) system, Customer Relationship Management System (CRM) system, Product Lifecycle Management System (PWC) system, Human Resources System, Warehouse Management System, Transportation Management System, Business Process Management System (BPM) system, Office Automation System, and Government Approval Platform. As a preferred approach, the communication connection is established through methods including calling application programming interfaces (APIs), establishing direct database connections, reading local files, reading network shared files, and subscribing to message queues. For scenarios involving concurrent acquisition of multi-source data, the system employs a network time protocol for global clock synchronization to eliminate timestamp drift between different physical computing nodes. For the parsing and connection establishment of different system underlying network protocols, those skilled in the art can perform conventional configurations based on the official open documentation of the corresponding system; the network connection establishment process is well-known in the field and will not be elaborated upon here. The acquired business data corresponds to documents in various business scenarios. The document types include purchase requisitions, sales orders, production work orders, quality anomaly reports, inventory aging analysis tables, accounts receivable and payable tables, R&D material and labor cost requisitions, personnel approval forms, project initiation documents, master production schedules (MPS), and material requirements planning (MRP).

[0073] After acquiring business data, it needs to undergo standardized dimensionality reduction processing. The data extraction and encapsulation module cleans the business data and encapsulates it into standardized JSON data. During this process, the system, based on a pre-mounted static mapping dictionary, forcibly converts heterogeneous fields from various sources into a unified internal basic data type. Core fields include basic information fields, material information fields, inventory status fields, amount information fields, natural language application reason fields, production-related fields, quality data fields, and credit risk fields. Basic information fields include document number, creation time, applicant, and department. Inventory status fields include current inventory, safety stock, and available inventory. Amount information fields include application amount, contract amount, and amount already paid. Natural language application reason fields include business background text entered by business personnel on the front-end page. Production-related fields include work order number and planned start time. Quality data fields include defect rate and number of rework attempts. Credit risk fields include supplier rating and overdue days. If any null values ​​or illegal characters are found in the extracted fields, the system assigns corresponding default safety values ​​according to preset type rules to prevent calculation errors such as denominators approaching zero or null pointer overflows during subsequent numerical calculations. The data extraction and encapsulation module converts each extracted core field into key-value pairs, generating standardized JSON data with a consistent structure. Through the above data type alignment and null value compensation mechanisms, data format differences between heterogeneous systems are completely eliminated.

[0074] Based on the encapsulated data described above, the data extraction and encapsulation module traverses the standardized JSON data to generate a valid key space set. This set represents the actual physical field boundaries of the current business document within the target enterprise management system. Specifically, the data extraction and encapsulation module extracts all key names from the standardized JSON data and combines them into a one-dimensional array or hash set, denoted as the valid key space set. This set serves as a validation benchmark in subsequent steps, preventing the AI ​​model from outputting false fields that do not exist in the target enterprise management system when generating decision results, thereby effectively mitigating the risk of system downtime caused by out-of-bounds queries in the underlying database.

[0075] To ensure the rigor of subsequent concurrent state verification, the data extraction and encapsulation module generates a dynamic execution boundary snapshot of the business document and persists it in cache. In this embodiment, the data extraction and encapsulation module assigns a unique transaction identifier to the currently processed business document and obtains the current system time as the initial timestamp. The relevant system parameter extraction and set definition logic is shown in the following formula:

[0076] ;

[0077] ;

[0078] In the formula, Indicates the initial timestamp;

[0079] This indicates the current system time after alignment with the Network Time Protocol.

[0080] Represents the set of valid key spaces;

[0081] This represents the total number of key names, and its value is a positive integer greater than or equal to 1. This value is determined by the number of non-empty core fields actually contained in the current business document.

[0082] Represents the first in the set of valid key spaces A specific key name. The range of values ​​is Positive integers between [a certain range].

[0083] The data extraction and encapsulation module records the specific numerical or textual content corresponding to each key name at the initial timestamp, denoted as the initial state value. This initial state value reflects the physical truth value of the business document at the moment of extraction, constituting the static physical benchmark for subsequent asynchronous state drift comparison. The data extraction and encapsulation module binds the unique transaction identifier, the initial timestamp, the set of valid key spaces, and all initial state values ​​into an independent data object, using this data object as a dynamic execution boundary snapshot. To balance high-concurrency read performance with distributed disaster recovery requirements, the dynamic execution boundary snapshot is stored in the system's in-memory database or distributed cache component, providing underlying benchmark data for subsequent concurrent conflict detection.

[0084] The intelligent decision-making reasoning in step S20 specifically includes the following sub-steps:

[0085] S201, In this embodiment, based on the structured data extracted in the aforementioned steps, the system needs to dynamically match the corresponding business judgment logic. The AI ​​interaction module starts the dynamic prompt word engine to load the matching context rules. The system maintains a relational database at its core, which stores business rule prompt word templates covering multiple business scenarios. The AI ​​interaction module parses the document type identifier and department information in the standardized JSON data, uses this information as query conditions, and retrieves and extracts the corresponding business rule prompt word template from the relational database. As a preferred approach, the business rule prompt word template includes preset expert role settings, business compliance verification boundaries, and handling strategies for similar historical situations. Through this dynamic loading mechanism, the system transforms the underlying fixed code logic into configurable natural language rules to adapt to the approval requirements of different levels and positions.

[0086] S202, after obtaining the template, in order to construct a text sequence that conforms to the model input specifications, the AI ​​interaction module performs context construction and data serialization operations. The AI ​​interaction module converts the structured standardized JSON data into plain text strings and embeds these plain text strings into predefined placeholders within the business rule prompt word template. During this assembly process, the system separates numerical fields and natural language application reason fields into paragraphs and uses system-preset separators for boundary labeling to prevent feature confusion regarding the physical meaning of fields during model parsing. After the above preprocessing logic, the system generates a comprehensive reasoning text with a complete logical structure. This comprehensive reasoning text includes the current physical state of the document and expert experience rules for determining whether the physical state is compliant, and its overall character length is strictly controlled within the maximum number of context terms supported by the model.

[0087] S203, after completing the text sequence combination, this system relies on a pre-trained large language model to perform deep semantic feature extraction and classification regression calculation. The AI ​​interaction module submits an inference request containing mandatory attribution constraint parameters to the large language model. In this embodiment, the selected large language model adopts a generative neural network model based on a multi-layer transformer decoder architecture. The internal hierarchical structure of this neural network model, from bottom to top, includes an input lexicalization layer, a multi-layer self-attention calculation block, a feedforward neural network layer, and a final linear classification and normalization exponent output layer.

[0088] During the model's input phase, the AI ​​interaction module transforms the comprehensive reasoning text into a high-dimensional sequence of lexical embedding vectors through an input lexicalization layer. Addressing the inherent output divergence inherent in generative language models, this solution sets strict structured output parameters in the application programming interface's request carrier. The AI ​​interaction module instructs the large language model to return the processing result in a predefined JSON data structure. More importantly, a mandatory attribution constraint parameter is embedded in the request carrier. This mandatory attribution constraint parameter, acting as a system-level instruction, requires the large language model to backtrack its internal reasoning path before outputting the final decision state. The large language model needs to identify the fields that play a decisive role in its decision within its analytical context and extract the keys of these fields and their corresponding comprehensive attention scores into an independent key-value pair mapping dictionary, which the system defines as an initial set of key factors carrying weight coefficients.

[0089] S204, the system receives and records the initial set of key factors calculated by the large language model based on the self-attention mechanism. Those skilled in the art will know that the reasoning mechanism within the large language model is generally built upon a multi-layer self-attention transformer architecture. When processing comprehensive reasoning text, the attention layer of the large language model calculates a weight score for each input token. In this specific attribution and source tracing scenario, the internal feature extraction logic of the large language model responding to the mandatory attribution constraint parameters relies on the quantitative evaluation of the attention weights. Its core self-attention score aggregation logic can be explained by the following formula with an anti-overflow term:

[0090] ;

[0091] In the formula, This represents the comprehensive attention score generated by the model for the labeled lexical units of a specific business field. This score physically reflects the degree of influence of the current input field on the final output decision state.

[0092] This represents the total number of attention heads in a multi-head attention mechanism. Its value is a positive integer greater than one, and the specific value is determined by the number of pre-trained network layers of the selected neural network model. This indicates the current attention head being computed. This represents the output weight matrix of the current attention head; Represents the query matrix; The transpose of the key matrix; The three matrices, representing the value matrices, are all obtained by linear mapping from the high-dimensional vector sequence of the comprehensive reasoning text. The dimension scalar representing the key vector; This represents the square root of the scalar in that dimension, used to prevent the gradient vanishing of the normalized exponential function in high-dimensional space during scaling dot product attention calculations; This represents the smallest positive real constant preset by the system. This represents the normalization exponential function, used to map the scaled dot product result to a probability distribution that sums to one.

[0093] After generating decision criteria, the large language model, based on the aforementioned calculation process, ranks the marked words corresponding to business fields in the comprehensive reasoning text according to their comprehensive attention scores. To avoid biased judgments caused by relying solely on a single extreme value, the system employs a threshold filtering mechanism based on multi-dimensional weighted logic. The large language model not only evaluates the absolute value of the comprehensive attention score of a single word, but also combines the word frequency of the word in the context with the field attribute weights assigned in the static business dictionary for aggregate scoring. The large language model selects marked words whose aggregate scores exceed a preset importance threshold, decodes them in reverse to their corresponding field keys, and extracts them along with the aggregate score value, assembling them into a key-value pair mapping dictionary.

[0094] As a preferred approach, the preset importance threshold is determined by the large language model during the fine-tuning training phase. In the model building and training steps, training samples are derived from archived real business documents from the company's history, while training labels consist of a set of core decision fields annotated by human business experts. The fine-tuning training process uses the cross-entropy loss function to calculate the difference between the predicted attribution field probability distribution and the true label distribution, and continuously updates the network weight matrix of the attention layer through backpropagation until the loss value on the validation set converges to an acceptable range, thus solidifying the corresponding preset importance threshold. The extracted dictionary containing the weight distribution serves as the initial set of key factors, which is output along with the final decision result. This mechanism constrains the model's complex nonlinear reasoning process at the underlying network weight level, giving it logical interpretability at the engineering level and providing precise parameter basis for subsequent targeted anti-concurrency verification based on underlying physical data.

[0095] The result analysis and factor cleaning in step S30 specifically includes the following sub-steps:

[0096] S301, In this embodiment, the system needs to transform unstructured model cognition into deterministic instructions executable by the underlying industrial system. The result parsing module receives the processing results returned by the large language model and performs format verification on the processing results based on a preset data structure. The system's underlying layer is configured with a standardized data structure validator. The result parsing module determines whether the returned data conforms to the system's required format by comparing the key-value pair structure, data type, and nesting level. To address potential data loss or type mismatch issues during the verification process, the system is configured with automatic retry and exception throwing mechanisms to prevent illegal data from penetrating downstream components. After successful verification, the result parsing module extracts multi-dimensional business features from the returned data.

[0097] S302, based on the above verification results, the result parsing module analyzes the decision results, extracting decision states including pass, rejection, or warning, extracting review opinions in natural language form, and extracting execution instructions in action identifier form. As a preferred approach, the decision state is represented by a strict preset enumeration value, used to map the macro-level compliance conclusion of business documents. The review opinion serves as explanatory textual evidence for subsequent manual traceability. The execution instructions are represented by action codes that can be directly parsed by the underlying application programming interface. The action codes correspond to the pass operation, rejection operation, sending warning notifications to the relevant responsible person's email or enterprise collaborative office software (such as DingTalk, WeChat Work) operation, and reporting anomalies to the next higher-level approver or risk control center operation in the business system. At the same time, the result parsing module extracts the initial key factor set generated by the large language model. This set consists of a series of key-value pairs containing field names and their corresponding comprehensive attention scores.

[0098] S303, based on the technical characteristics of generative artificial intelligence models that rely on probability distributions to generate text, its output initial key factor set may contain illegal fields undefined in the target enterprise management system's database architecture. To eliminate such data uncertainty, the result parsing module calculates the intersection of the legal key space set and the key names in the initial key factor set, removing fields that do not belong to the legal key space set and their corresponding score values. Before performing specific algebraic mapping, the system removes leading and trailing spaces and special escape characters from the set elements using regular expressions, and converts all English letters to the same uppercase and lowercase format to prevent matching failures due to character encoding or case differences. After completing string normalization preprocessing, the system performs strict field filtering based on the intersection operation principle in set theory. In this embodiment, the cleaning logic of the factor intersection filter can be explained by the following set operation formula:

[0099] ;

[0100] In the formula, This represents the set of effective attribution factors after cleaning.

[0101] This represents all the key names contained in the initial key factor set output by the large language model in the previous processing stage;

[0102] This represents the set of valid key spaces that the system retrieves from the underlying layer and persists in the cache.

[0103] This represents the set intersection operator.

[0104] The technical purpose of this formula is to force the divergent output boundaries of the large language model to be pruned within the legal parameter range of the underlying industrial database through rigorous set intersection operations. This deterministic algebraic mapping ensures that the subsequently generated dynamic query probes fully conform to the table structure definition of the target enterprise management system, thereby avoiding the target enterprise management system from throwing Structured Query Language execution exceptions or causing database service crashes due to query out-of-bounds errors. After calculation, the result parsing module outputs a cleaned set of effective attribution factors (internally retaining only legal key names and their associated attention scores).

[0105] S304. To ensure high availability and absolute data security of the underlying execution system in the manufacturing industry, and to avoid system deadlock caused by relying solely on a single logical judgment, the system performs multi-dimensional joint judgment on the cleaned set state. The result parsing module evaluates the cardinality state and business complexity weight of the effective attribution factor set. If the effective attribution factor set is empty, or if the number of elements contained in the set is less than the minimum association threshold determined by the number of core required fields for different document types in the business dictionary, it indicates that the judgment basis generated by the large language model has deviated from the physical data environment of the actual business document. At this time, the system triggers the fault tolerance degradation mechanism, intercepts the issuance of the current decision instruction (i.e., intercepts the above-extracted execution instruction), and suspends the business document along with the review opinion generated by the large language model, transferring it to the manual review queue. The minimum association threshold is determined by the number of core required fields for different document types in the business dictionary, and its value range is a positive integer greater than or equal to 1. If the effective attribution factor set is not empty and meets the above threshold requirements, the result parsing module outputs the set as a valid parameter, providing accurate targeted input data for the next step of asynchronous state probe generation.

[0106] The asynchronous polling and state drift calculation in step S40 specifically includes the following sub-steps:

[0107] S401, In this embodiment, since deep neural network inference of large language models is usually accompanied by uncontrollable time delays, the underlying data of the main business system may change during this period. To avoid blocking the main business thread of the target enterprise management system, the scheduled task execution module triggers an asynchronous check mechanism at the execution timestamp. As a preferred approach, the scheduled task execution module is configured with an event-driven non-blocking thread pool. When the inference logic of the business document at the artificial intelligence model end is completed and a valid attribution factor set is output, the system calls the underlying clock interface to obtain the current high-precision network synchronization time and defines it as the execution timestamp. Based on the above timestamp, the scheduled task execution module wakes up the dormant polling thread to prepare for concurrent verification of possible physical state changes of the document during model inference.

[0108] S402, after obtaining the execution timestamp, to reduce the input / output query pressure on the underlying production database, the system abandons the traditional concurrency detection method of full table field scanning. The scheduled task execution module uses the set of valid attribution factors as query parameters to request the real-time status value under the execution timestamp from the target enterprise management system. Specifically, the system only extracts the key names contained in the set of valid attribution factors and uses these key names to dynamically concatenate into Structured Query Language (SCL) instructions or Application Programming Interface (API) request parameters. This targeted probe mechanism fully utilizes the attribution boundaries already output by the artificial intelligence model, performing secondary reading only on core fields strongly related to the current decision. This operation ensures the rigor of concurrency verification while maximizing the conservation of network bandwidth and database connection pool resources. The system records the specific numerical or text content returned by the above targeted query and persists it as the real-time status value of the current business document.

[0109] S403, for the acquired real-time status value and the initial status value cached prior, the system needs to quantify the degree of physical deviation between the two within the time window. Since industrial business data includes continuous numerical variables and discrete textual variables, the system adopts heterogeneous data branching calculation logic. For continuous numerical fields such as inventory quantity and amount, the system calculates their normalized absolute deviation; to prevent division overflow when the original status value is zero, a very small positive number is introduced as a denominator compensation term. For discrete status fields such as approval status and supplier level, the system performs binary transition discrimination based on string hash comparison. At this stage, the system uniformly maps the deviations of the above two heterogeneous fields to a preset real number interval, generating independent state drift values ​​for each field. In this embodiment, the calculation logic of independent state drift values ​​can be explained by the following piecewise function:

[0110] ;

[0111] In the formula, Indicates the first in the set of effective attribution factors The independent state drift of each field;

[0112] This indicates the real-time status value obtained by querying under the execution timestamp;

[0113] This indicates the initial state value recorded for this field at the initial timestamp;

[0114] This represents the system's preset minimum normal value for preventing division by zero, which is typically set to 10. -6 It is used to maintain the validity of mathematical operations when the initial state value approaches zero;

[0115] This represents the function that takes the maximum value.

[0116] This represents the Kronecker function, which outputs 1 when the string hash code of the real-time state value is exactly the same as the string hash code of the initial state value, and outputs 0 otherwise.

[0117] The technical purpose of this piecewise function is to uniformly transform underlying business data with different physical dimensions and data types into dimensionless pure numerical drift metrics, providing a standardized input benchmark for subsequent aggregation calculations.

[0118] S404, after completing the deviation quantification of a single field, the system needs to combine the global business perspective with the local cognition of the artificial intelligence model to perform multi-dimensional sensitivity fusion. The scheduled task execution module calculates the attention-weighted state drift index between the initial state value and the real-time state value based on the sensitivity weights preset in the business dictionary. As a preferred approach, the static business dictionary configures benchmark sensitivity weights for different types of fields. For example, the sensitivity weight of the purchase amount field is usually manually set to be greater than that of the ordinary remarks field. The system fuses this static sensitivity weight with the dynamic comprehensive attention score output by the large language model to construct a joint weight coefficient. Subsequently, the system multiplies and sums the joint weight coefficients of each field with the corresponding independent state drift amount to obtain the global attention-weighted state drift index. In this embodiment, the aggregation logic of the attention-weighted state drift index can be explained by the following mathematical model:

[0119] ;

[0120] In the formula, This represents the attention-weighted state drift index ultimately calculated by the system.

[0121] This represents the total number of fields included in the set of valid attribution factors, and its value ranges from positive integers greater than or equal to 1.

[0122] This indicates that the value assigned by the underlying static business dictionary is... The sensitivity weight of each field, whose value ranges from 0 to 1;

[0123] This represents the comprehensive attention score generated by the large language model for this field during the preceding inference step;

[0124] This represents the system's preset weight allocation coefficient, which ranges from 0 to 1. This weight allocation coefficient is used to adjust the dominant ratio of human business experience and artificial intelligence reasoning in concurrent conflict detection. Its specific value is dynamically configured by the system administrator based on the fault tolerance of the specific business scenario.

[0125] The technical purpose of this formula is to break away from the one-sided concurrency control mechanism that relies solely on simple numerical comparisons. By introducing a dual weighting of business importance and model decision dependence, the system can achieve adaptive concurrency tolerance adjustment. For example, when a core indicator that plays a decisive role in model decision-making and is highly sensitive to business changes slightly, the formula will exponentially amplify its overall drift, forcibly triggering subsequent interception mechanisms; while when a non-core discrete state field changes, it outputs a smaller drift exponent, allowing the business process to continue. This multi-dimensional weighted logic effectively avoids the performance loss caused by the mindless rollback of global transactions due to updates of irrelevant fields in traditional optimistic locking mechanisms, significantly improving instruction throughput efficiency in high-concurrency scenarios in manufacturing.

[0126] The action execution and transaction control in step S50 specifically includes the following sub-steps:

[0127] S501, in this embodiment, after quantifying the underlying physical state deviation, the system needs to determine whether to send the model-generated decision instruction to the target enterprise management system based on multi-dimensional conditions. The action execution and transaction control module obtains the attention-weighted state drift index output by the preceding steps and, combined with the real-time table-level lock status fed back by the target enterprise management system, performs the final interception calculation. Considering the potential concurrent write conflicts in complex industrial production network environments, the system abandons the arbitrary judgment method of comparing only a single threshold and introduces a dynamic drift tolerance threshold and a forced manual lock flag for joint evaluation. Before making a specific instruction release judgment, the system constructs a multi-condition combined gating logic based on the logic and operation principles of Boolean algebra. In this embodiment, the final execution judgment logic can be explained by the following condition function:

[0128] ;

[0129] In the formula, This indicates the final execution judgment result. A value of 1 represents allowing the instruction to proceed, while a value of 0 represents blocking the instruction.

[0130] This represents the attention-weighted state drift index calculated by the system in the previous stage;

[0131] This represents the dynamic drift tolerance threshold dynamically configured by the system for different business documents. The value of this threshold is a real number ranging from 0 to 10. Its specific value is determined by the underlying static business dictionary based on the financial impact level or production safety level of the document. For example, the threshold for financial documents is usually set to a small value close to 0.

[0132] This indicates a forced manual lock flag fed back by the target enterprise management system. The flag is set to 1 when the corresponding data row in the target database is being occupied by other high-priority manual processes, and 0 otherwise.

[0133] The technical purpose of this formula is to establish a dual protection barrier based on data stability and the underlying transaction occupancy state. It ensures that automatic machine intervention in business processes is only permitted under these dual security conditions: core data drift is within a preset safe range, and the underlying physical data records are not locked by concurrent processes. This effectively prevents the risks of dirty writes and deadlocks in concurrent environments.

[0134] S502, based on the calculation result of the above conditional function, the action execution and transaction control module executes the corresponding transaction branch logic. If the final execution judgment result is one, it indicates that within the time window of the deep network inference of the artificial intelligence model, the core data of the business document has not undergone a physical mutation sufficient to overturn the original decision, and the current data is not locked. At this time, the action execution and transaction control module extracts the execution instruction containing the action identifier. Regardless of whether the previous decision status is passed or rejected, it is serialized into a network request message conforming to the application programming interface specification of the target enterprise management system, and the corresponding status transition instruction is issued. To prevent repeated execution of instructions caused by network jitter in the industrial field, the system injects a unique transaction identifier generated based on the distributed snowflake algorithm into the header of the network request message to achieve strict idempotency control of the interface call. For the underlying filtering of the interface idempotency verification, those skilled in the art can perform conventional design based on the unique key constraint of the database or the Bloom filter of the distributed cache. The specific implementation method is a well-known technology in the field and will not be described in detail here. After the instruction is successfully issued, the system receives the submission success confirmation response returned by the underlying database, completing the automatic processing closed loop of the current business document.

[0135] S503: If the final execution result is zero, it indicates that the underlying physical data has significantly shifted, or that business records have been taken over by other concurrent transactions. In response to this high-risk situation, the system immediately triggers the avalanche protection circuit breaker and rollback mechanism. The action execution and transaction control module intercepts the execution instructions generated by the artificial intelligence model and sends a transaction abandonment message to the target enterprise management system, forcibly releasing the previously held dynamic execution boundary snapshot and database connection pool resources.

[0136] Based on the above interception actions, the system further analyzes the specific reasons why the final execution judgment result is zero. If the blocking reason is that the forced manual lock flag reported by the target enterprise's management system is one (i.e.... This indicates that the document is being used by a high-priority manual process. The system skips the secondary reasoning step and directly and completely shuts down the automatic processing flow. If the reason for the blockage is that the attention-weighted state drift index is greater than the dynamic drift tolerance threshold (i.e., ... The system extracts the out-of-bounds fields that cause the drift to exceed the limit and their specific deviation values, updates the corresponding fields in the standardized JSON data using real-time status values, and resubmits the updated business documents to the AI ​​interaction module for secondary reasoning.

[0137] If the final execution decision result generated by the secondary inference is still zero, the system will also completely shut down the automatic processing flow. When any of the above-mentioned circuit breaker conditions occur, the system generates an anomaly warning report. If the circuit breaker is caused by the attention-weighted state drift index exceeding the dynamic drift tolerance threshold, the report is composed of the out-of-bounds field, the specific deviation value, and the natural language review opinion output by the large language model; if the circuit breaker is caused by underlying physical table locking, the report is composed of the document's basic information and the table locking conflict log. This anomaly warning report is automatically pushed to the anomaly monitoring dashboard of the manufacturing operations management system, forcibly transferring the abnormal document to the manual review queue.

[0138] S504, to construct a closed-loop control system with self-iterative capabilities and prevent model performance from conceptual drift due to changes in industrial scenarios, persistently archives complete transaction lifecycle data in the action execution and transaction control modules. The system packages business data, comprehensive reasoning text generated by the large language model, cleaned effective attribution factor sets, and final execution judgment results into composite sample records and writes them to an independent distributed file system. In the subsequent offline optimization phase, these composite sample records are used as incremental training corpora for fine-tuning the large language model. Specifically, the comprehensive reasoning text of business documents is lexicalized into a high-dimensional vector sequence as input features; while the judgment conclusions and attribution fields finalized by business experts in the manual review queue serve as the true labels. The fine-tuning process employs a gradient descent-based backpropagation algorithm, continuously updating the network weight matrix of the multi-layer self-attention calculation blocks by calculating the cross-entropy loss function between the model's predicted probability distribution and the true labels, until the loss value on the validation set converges to a preset range. Through this data feedback mechanism based on real-world operating conditions, the system can continuously calibrate the attention allocation preferences of the large language model in specific industrial scenarios, further reducing the decision conflict rate in complex industrial situations.

[0139] To more clearly illustrate the technical solution of the present invention and its practical application effects, a specific application scenario is listed below: high-concurrency approval of purchase requisition forms.

[0140] Specific application examples:

[0141] Suppose a large electronics manufacturing company is running a target enterprise management system (SAP ERP). A salesperson submits an "Urgent Material Purchase Request Form" on the system front end. The fully automated processing flow of the system of this invention is as follows:

[0142] S10 (Extraction and Encapsulation): The data extraction and encapsulation module detects the creation of a purchase requisition (unique transaction identifier: PO-20260310-001). The system starts at the initial timestamp. (e.g., 10:00:00) Extract business data from the ERP system. After cleaning, generate standardized JSON data. The initial status values ​​recorded include:

[0143] Current inventory: 50 pieces

[0144] Safety stock: 100 units

[0145] Number of applications: 80

[0146] Reason for application: "Due to a last-minute additional order from major customer Huawei, the production line urgently needs to stock up on materials; therefore, we are requesting expedited procurement." The system extracts all existing key names and generates a set of valid key spaces. (like: ).

[0147] S20 (Intelligent Reasoning): The AI ​​interaction module embeds the aforementioned JSON data into the "Procurement Review" business rule prompt template and submits it to the large language model. The large language model not only outputs the decision conclusion but also, under the influence of forced attribution constraint parameters, outputs the initial set of key factors and its comprehensive attention score in reverse:

[0148] (Current inventory): 0.6 (plays a decisive role)

[0149] (Natural Language Reasoning): 0.3 (Secondary determining factor)

[0150] (Supplier sentiment): 0.1 (Illusory fictional field generated by the large language model)

[0151] S30 (Parsing and Cleaning): The result parsing module verifies the validity of the JSON format returned by the large language model. It extracts the decision status as "Pass" and executes the command as "Approve".

[0152] Entering factor intersection filtering, the set of legal key spaces is... Intersection calculation with the initial set of key factors ( ).because Keys not in the valid key space are completely removed. The cleaned set of valid attribution factors retains only those that are not in the valid key space. (0.6) and (0.3).

[0153] S40 (Asynchronous Polling and State Drift Calculation): The time has now reached 10:00:05 (execution timestamp). During this 5-second asynchronous inference period, another production order has just picked up 10 units of the material from the warehouse.

[0154] The scheduled task execution module sends targeted probes to the ERP system only based on valid attribution factors to obtain... The real-time status value changed to 40 items.

[0155] The system calculates the independent state drift. : .

[0156] Combined with the sensitivity weights configured in the business dictionary (assuming inventory weights) After allocation, the calculated value is 0.8. The system then calculates the final attention-weighted state drift index. It is 0.16.

[0157] S50 (Heterogeneous Lock Adaptation and Closed-Loop Control): Action Execution and Transaction Control Module Judgment: The system presets a dynamic drift tolerance threshold for this type of routine purchase order. The value is 0.3. This is because the calculated drift index 0.16 ≤ 0.3, and the underlying database indicates that the table was not manually locked. ).

[0158] The system determined that although inventory had shifted, the shift was not severe enough to overturn the logical baseline that "current inventory is still far below safety stock." Therefore, the system automatically called the underlying ERP interface to update the document status to "approved" and persisted the entire data chain as a composite sample record, completing a closed loop without human intervention.

[0159] Experimental verification and effect comparison:

[0160] To verify the effectiveness and reliability of the method of the present invention in real industrial scenarios, the applicant conducted offline simulation experiments based on the real MES / ERP historical operating condition dataset of a leading home appliance manufacturer.

[0161] Experimental setup:

[0162] Test dataset: 10,000 historical business documents (including procurement, quality, work order anomalies, etc.) were randomly selected, and 15% of the underlying data was randomly injected into the simulation environment to tamper with it during high concurrency (simulating the risk of dirty writes).

[0163] Control group A (traditional rule engine): uses hard-coded static threshold logic.

[0164] Control group B (naked large language model access): directly uses the large model to make judgments and attempt to issue commands, without configuring the legal key space constraint and state drift detection mechanism of this invention (i.e. lacking S30 and S40 / S50 mechanisms).

[0165] Experimental group (the scheme of this invention): Fully enabled forced attribution factor cleaning, attention-weighted state drift calculation and closed-loop control.

[0166] Comparison of indicators and experimental results

[0167]

[0168] Experimental conclusion:

[0169] Experiments have shown that this invention, by creatively introducing valid key space verification and a set of effective attribution factors, restricts the uncontrollable output of AI to 100% within the boundaries of physical database security, completely eliminating the risk of system downtime caused by out-of-bounds queries. At the same time, by calculating the attention-weighted state drift index, it breaks through the deadlock bottleneck of the traditional concurrency control's "black and white" approach, intercepting malicious dirty writes while tolerating reasonable data drift, thereby increasing the unmanned closed-loop rate of automated manufacturing processes to over 91.5%.

[0170] The above-mentioned technical effects can be fully demonstrated by combining the experimental data in the attached figures:

[0171] Combined with appendix Figure 2 Quantitative analysis:

[0172] Figure 2 The three core indicators of the three different processing solutions (rule engine, bare large model access, and the solution of this invention) in complex industrial scenarios are compared horizontally in the form of a bar chart. The vertical axis of the chart represents the percentage (%) of each indicator.

[0173] Accuracy of understanding complex reasons (corresponding to the light gray bars in the figure): Traditional rule engines based on hard coding lack natural language processing capabilities, resulting in an accuracy of only 12.5% ​​for understanding complex reasons. A large number of documents containing unstructured business backgrounds cannot be effectively parsed. However, after introducing the large language model, the accuracy of the bare large model access solution and the solution of this invention soared to 94.2% and 94.5% respectively, proving that this invention perfectly preserves the excellent ability of the large language model in deep semantic feature extraction.

[0174] Out-of-bounds query downtime rate (corresponding to the dark black bars in the graph): This is a core indicator for measuring the underlying security of the system. The graph shows that traditional rule engines, due to their strict static table structure limitations, have a 0% downtime rate; however, bare-metal large-model access solutions, due to AI illusions and uncontrollable divergent output, directly generate structured query commands, leading to an out-of-bounds query downtime rate as high as 8.7%, which is unacceptable in core manufacturing systems. The solution of this invention, by executing a factor cleaning logic of "intersection of legal key space sets," forcibly prunes out-of-bounds fields, successfully reducing the out-of-bounds query downtime rate back to 0%, achieving a perfect fusion of AI intelligence and industrial-grade absolute security.

[0175] The fully automated closed-loop rate (corresponding to the gray bars in the figure): Limited by low comprehension capabilities and mechanical logic that leads to deadlocks under concurrency, the fully automated closed-loop rate of traditional rule engines is only 15%; while the bare large model access solution has improved this to 76%, the lack of effective state verification in concurrent environments often leads to dirty write errors, forcing manual rollbacks. The solution of this invention balances high-precision inference and concurrency fault tolerance, significantly improving this indicator to 91.5%, greatly reducing the cost of manual review.

[0176] Combined with appendix Figure 3 Logical explanation:

[0177] like Figure 3 As shown in the figure, the graph visually illustrates the anti-collision circuit breaker process of this invention when processing 30 high-concurrency business documents in the form of a line scatter plot. The horizontal axis of the graph is the "concurrent business document processing sequence (sorted by timestamp)", and the vertical axis is the "attention-weighted state drift index" calculated by the system.

[0178] System dynamic tolerance threshold (corresponding to the horizontal dashed line in the figure): A horizontal dashed line with a vertical axis of 0.3 is set in the figure, representing the "interception threshold baseline" and the "system dynamic tolerance threshold". This threshold is dynamically configured by the system based on the business dictionary, forming the watershed between command release and interception.

[0179] Reasonable data drift release (corresponding to the hollow circle data points in the diagram): The "Real-time Document Drift Index (Hollow Circles)" connected by the black solid line in the diagram represents the amount of physical state change that occurs in each document within the asynchronous inference time window. It can be seen that most documents (such as serial numbers 1, 2, and 4 to 22, etc.) experienced minor state drifts in the underlying database (index between 0 and 0.25), but none exceeded the safety baseline of 0.3. This invention's system allows these operations to continue to be released and complete the closed loop, completely breaking the deadlock bottleneck of traditional optimistic locking's "global rollback if there is even the slightest data change," thus ensuring high business throughput.

[0180] Interception of Malicious Dirty Writes and Abnormal Operations (corresponding to solid triangle data points in the diagram): The "abnormal operations that were circuit-broken / intercepted (solid triangles)" highlighted in the diagram represent business documents with abnormally high drift indices. For example, the drift index of document number 3 reached approximately 0.33, while the drift index of documents with serial numbers 23, 24, and 25 experienced a sharp change (approaching 0.96 at its highest). This indicates that during the AI's processing, the underlying core business data (such as inventory being heavily seized by other work orders) underwent a drastic change sufficient to overturn the original decision-making benchmark. Once the hollow circle crosses the dotted line and becomes a solid triangle, the system action execution and transaction control module will immediately trigger a circuit breaker to intercept the automated instruction, thereby effectively preventing security incidents caused by concurrent write conflicts.

Claims

1. An automated processing method for manufacturing systems based on AI and structured data interaction, characterized in that, Includes the following steps: The business data of the target enterprise management system is obtained, encapsulated, and the initial state values ​​of each field are recorded. The key names are then extracted to generate a set of valid key spaces. After the business data is converted into standardized JSON data, it is input into a large language model for inference to obtain decision results and an initial set of key factors. The initial set of key factors includes field key names and corresponding comprehensive attention scores. Extract the execution instructions from the decision results, calculate the intersection of the field key names in the set of legal key spaces and the set of initial key factors, and output the cleaned set of effective attribution factors. Obtain the real-time state value of the corresponding field in the set of effective attribution factors, and calculate the attention-weighted state drift index between the initial state value and the real-time state value by combining the comprehensive attention score. If the attention-weighted state drift index is greater than the dynamic drift tolerance threshold, the business operation is completed according to the execution instruction. If it is greater than the dynamic drift tolerance threshold, the standardized JSON data is updated using the real-time state value for secondary reasoning, or the business document is forcibly transferred to the manual review queue.

2. The automated processing method for a manufacturing system based on AI and structured data interaction as described in claim 1, characterized in that, The specific steps for obtaining business data from the target enterprise management system, recording the initial state values ​​of each field, and extracting key names to generate a set of valid key spaces are as follows: The acquired business data is cleaned and transformed according to preset type rules and a static mapping dictionary, and then encapsulated into the standardized JSON data. Assign a unique transaction identifier to the currently processed business document and record the system time when the data was retrieved as the initial timestamp; Traverse the standardized JSON data and extract the key names of all non-empty core fields to generate the set of valid key spaces; The unique transaction identifier, the initial timestamp, the set of valid key spaces, and the initial state values ​​of each field under the initial timestamp are bound together to generate and cache dynamic execution boundary snapshots.

3. The automated processing method for a manufacturing system based on AI and structured data interaction according to claim 2, characterized in that, The specific steps for converting the business data into standardized JSON data and inputting it into a large language model for inference to obtain decision results and an initial set of key factors are as follows: Parse the document type identifier and department information in the standardized JSON data, and retrieve the matching business rule prompt word template from the preset relational database; The standardized JSON data is converted into a plain text string and embedded into a predefined placeholder in the business rule prompt word template to generate comprehensive reasoning text; The inference request containing the comprehensive inference text and the mandatory attribution constraint parameters is submitted to the large language model to obtain the decision result generated by the inference of the large language model. At the same time, the mandatory attribution constraint parameters serve as system-level instruction prompts, requiring the large language model to perform self-attention calculation on the marked lexical units in the comprehensive inference text, extract the field key names that play a decisive role in the decision and the corresponding comprehensive attention scores, and output the initial key factor set in the form of key-value pairs.

4. The automated processing method for a manufacturing system based on AI and structured data interaction as described in claim 3, characterized in that, The specific steps for extracting the execution instructions from the decision results, calculating the intersection of the field key names in the set of legal key spaces and the set of initial key factors, and outputting the cleaned set of effective attribution factors are as follows: Analyze the decision results and extract the execution instructions in the form of action identifiers; The elements in the legal key space set and the initial key factor set are subjected to the removal of leading and trailing spaces and special escape characters, as well as case normalization. Perform an intersection operation between the set of legal key spaces (after removing leading and trailing spaces, special escape characters, and case normalization) and the initial key factor set, and remove illegal fields and their corresponding scores from the initial key factor set that do not belong to the set of legal key spaces to generate a set of valid attribution factors. If the generated set of valid attribution factors is empty, or if the number of elements in the set is less than the minimum association threshold determined by the number of core required fields for different document types in the business dictionary, then the execution instruction is intercepted and the current business document is transferred to the manual review queue; otherwise, the cleaned set of valid attribution factors is output.

5. The automated processing method for a manufacturing system based on AI and structured data interaction according to claim 4, characterized in that, The specific steps for obtaining the real-time status value of the corresponding field in the set of effective attribution factors are as follows: When the business document completes the large language model inference and outputs the set of effective attribution factors, an asynchronous check mechanism is triggered and the current time is obtained as the execution timestamp. The set of effective attribution factors is used as query parameters, and the key names contained in the set are dynamically concatenated to generate structured query language instructions or application programming interface request parameters as targeted probes. The targeted probe requests the specific numerical or textual content of the corresponding field in the effective attribution factor set at the execution timestamp from the target enterprise management system and records it as the real-time status value.

6. The automated processing method for a manufacturing system based on AI and structured data interaction according to claim 1, characterized in that, Before calculating the attention-weighted state drift index between the initial state value and the real-time state value by combining the comprehensive attention score, the method further includes calculating the independent state drift of a single field: For continuous numerical fields, based on the difference between the real-time state value and the initial state value, a minimum normal number to prevent the division of zero is introduced as a denominator compensation term to calculate the normalized absolute deviation. For discrete state fields, a binary transition judgment is performed by comparing the string hash codes of the real-time state value and the initial state value. The normalized absolute deviation and the binary jump discrimination result are uniformly mapped to a preset real number interval to generate the independent state drift corresponding to each field.

7. The automated processing method for a manufacturing system based on AI and structured data interaction as described in claim 6, characterized in that, The specific steps for calculating the attention-weighted state drift index between the initial state value and the real-time state value are as follows: Extract the preset sensitivity weights for different types of fields from the static business dictionary; A preset weight allocation coefficient is introduced, and the sensitivity weight is fused with the comprehensive attention score for the corresponding field output by the large language model to construct a joint weight coefficient. The global attention-weighted state drift index is obtained by multiplying and summing the joint weight coefficients of each field with the corresponding independent state drift values.

8. The automated processing method for a manufacturing system based on AI and structured data interaction according to claim 1, characterized in that, The specific steps for determining whether the attention-weighted state drift index is greater than the dynamic drift tolerance threshold, and if it is not greater than the dynamic drift tolerance threshold, to complete the business operation according to the execution instruction are as follows: Obtain the forced manual lock flag fed back by the target enterprise management system, and construct a conditional function containing the dynamic drift tolerance threshold and the forced manual lock flag based on the logic and operation principle of Boolean algebra; Based on the calculation result of the condition function, the release command is allowed if and only if the attention-weighted state drift index is less than or equal to the dynamic drift tolerance threshold of the document type, and the forced manual lock flag indicates that the data is not occupied by other concurrent processes. The extracted execution instructions are serialized into a network request message. A unique transaction identifier generated based on the distributed snowflake algorithm is injected into the message header to control idempotency. The message is then sent to the target enterprise management system to execute the business operation closed loop.

9. The automated processing method for a manufacturing system based on AI and structured data interaction according to claim 2, characterized in that, The steps of updating the standardized JSON data using the real-time status value for secondary inference or forcibly transferring the business document to the manual review queue if the value exceeds the dynamic drift tolerance threshold include: Intercept the execution instructions generated in the preceding order, trigger the anti-avalanche circuit breaker and rollback mechanism, send a transaction discard message to the target enterprise management system, and forcibly release the previously held dynamic execution boundary snapshot and database connection pool resources; Based on the aforementioned interception and forced release, analyze the specific triggers that caused the interception; If the specific cause is that the attention-weighted state drift index is greater than the dynamic drift tolerance threshold, then the out-of-bounds field that caused the drift to exceed the limit and the specific deviation value are extracted. The corresponding fields in the standardized JSON data are updated using the obtained real-time status values ​​to generate updated standardized JSON data; The updated standardized JSON data is resubmitted to the large language model for secondary inference. If the attention-weighted state drift index corresponding to the secondary reasoning result is still greater than the dynamic drift tolerance threshold, or if the specific cause is a forced manual lock flag fed back by the target enterprise management system, then the automatic processing flow will be completely interrupted, and an abnormal warning report will be generated when the circuit breaker occurs. If the circuit breaker is triggered because the attention-weighted state drift index is greater than the dynamic drift tolerance threshold, the extracted out-of-bounds field, the specific deviation value, and the natural language review opinion output by the large language model will be concatenated to generate the abnormal warning report. If the circuit breaker is triggered by physical table locking in the target enterprise management system, the basic document information and the table locking conflict log will be concatenated to generate the abnormal warning report. Finally, the abnormal warning report is pushed to the abnormal monitoring dashboard, and the abnormal documents are forcibly transferred to the manual review queue.

10. The automated processing method for a manufacturing system based on AI and structured data interaction according to claim 9, characterized in that, It also includes the step of building a closed-loop control system based on complete transaction lifecycle data: The business data, the comprehensive reasoning text submitted to the large language model, the cleaned set of effective attribution factors, and the final execution judgment result are packaged into a composite sample record. Obtain the manually corrected judgment conclusions and attribution fields from the manual review queue as the true labels; Using the composite sample records and their corresponding real labels as incremental training corpora, the network weight matrix of the multi-layer self-attention computation block within the large language model is updated through the backpropagation algorithm to calibrate the attention allocation preference of the large language model.

Images