Method and system for intelligent management and control of pharmaceutical production execution in multi-system integration

Abstract

The application discloses a kind of multi-system integrated pharmaceutical production execution intelligent management and control method and system, belong to industrial data acquisition and equipment linkage control technical field, by pre-training time sequence feature extraction model, the material attribute data of multiple node collection is analyzed in depth, time sequence dependent relationship in production process is automatically identified and parameter adjustment instruction is generated, again by semantic mapping and format conversion, the standardized collaborative control of heterogeneous system is realized;When detecting production deviation, immediately trigger early warning mechanism and carry out abnormal identification and task scheduling optimization by integrated learning classification model, equipment control instruction is dynamically generated according to task priority;Finally, through the authority verification and integrity verification mechanism, the compliance of control instruction is ensured, and audit tracking record is automatically generated.The application realizes the whole-process intelligent closed-loop management from data acquisition to equipment control, significantly improves the real-time response capability, abnormal processing efficiency and operation compliance of production system.

Description

Technical Field

[0001] This invention relates to the fields of pharmaceutical production execution systems, industrial data acquisition and equipment linkage control technology, and in particular discloses an intelligent management and control method and system for pharmaceutical production execution that integrates multiple systems. Background Technology

[0002] Intelligent management and control in pharmaceutical production is a core component in ensuring drug quality, production efficiency, and compliance with GMP (Good Manufacturing Practices) data integrity requirements. A search reveals that existing pharmaceutical production management-related patents (CN117726186A, CN118095842A) mostly only achieve data processing and basic control within a single system. They generally fail to distinguish between continuous time-series data acquired by online sensors and offline batch quality inspection data from LIMS, lack in-depth analysis of the time-series characteristics of materials with timestamps and equipment identifiers across multiple nodes, and lack a comprehensive integrated management and control system encompassing heterogeneous data semantic mapping and format conversion, intelligent anomaly classification, cross-system task scheduling, instruction compliance verification, and audit trail.

[0003] The shortcomings and corresponding consequences of existing technologies are as follows: First, the various business systems are independent of each other, with heterogeneous data formats and inconsistent semantics, resulting in communication and data barriers that prevent information sharing and global linkage. Second, dependency analysis and intelligent extraction based on the time-series characteristics of materials are not conducted, making it difficult to dynamically generate production parameter adjustment instructions based on material status. Third, production anomalies are only given simple alarms, without relying on integrated learning models to complete anomaly classification and intelligent task scheduling, and anomaly handling relies on manual operation, resulting in delayed response. Fourth, there is a lack of control instruction authorization verification and data integrity verification, and GMP audit trail records are not generated simultaneously, which not only increases production losses but also poses serious compliance risks, failing to meet the requirements of full-process automation and standardized management in the pharmaceutical industry.

[0004] Therefore, how to break down the data barriers between heterogeneous systems and achieve integrated operation of material time sequence characteristic analysis, cross-system data collaboration, intelligent handling of anomalies, instruction compliance verification and audit traceability is a key technical problem that urgently needs to be solved in the field of intelligent management and control of pharmaceutical production. Summary of the Invention

[0005] This invention provides a multi-system integrated intelligent management and control method and system for pharmaceutical production execution, aiming to solve the technical defects of traditional production systems, such as difficulty in real-time collaboration of multi-source heterogeneous data, delayed response to production anomalies, and lack of compliance traceability of equipment control commands.

[0006] One aspect of the present invention relates to an intelligent management and control method for pharmaceutical production execution integrating multiple systems, comprising the following steps: S100: Collect material attribute data from multiple different collection nodes throughout the entire pharmaceutical production process, construct and obtain a set of material attribute features. The set of material attribute features contains multiple sets of material attribute time-series features, and each material attribute time-series feature is configured with a collection timestamp and collection device identification information. S200: Call the pre-trained time-series feature extraction model to perform time-series dependency analysis and deep feature extraction on all material attribute time-series features in the material attribute feature set, explore the correlation between material attribute time-series changes and pharmaceutical production parameters, and output the production parameter adjustment instructions corresponding to the current production batch. S300 performs semantic mapping matching and unified format conversion processing on the multiple heterogeneous system data fields contained in the production parameter adjustment instructions, eliminates the data heterogeneity differences of multiple business systems, and generates standardized collaborative control events that are compatible with various business systems in pharmaceutical production. S400 analyzes the production control logic built into the standardized collaborative control event, compares and verifies the operating parameters of the standardized collaborative control event with the preset production deviation rules, and generates a production anomaly warning signal corresponding to the current production batch immediately if the preset production deviation rules are triggered. S500: Input the production anomaly warning signal into the pre-trained ensemble learning classification model, and complete the deep feature extraction and accurate classification of the anomaly category through the ensemble learning classification model, and output the production task scheduling sequence to guide subsequent pharmaceutical production operations. S600 performs intelligent diversion and logical operation processing on various production tasks according to the built-in task priority sorting rules of the production task scheduling sequence, and generates underlying production equipment control instructions that can directly drive the operation of underlying hardware devices. S700 performs dual verification of operation permission and data integrity compliance for the control commands of the underlying production equipment in sequence. If the control commands of the underlying production equipment pass the verification, the control commands of the underlying production equipment are executed, and a compliance audit trail record adapted to GMP standards is generated simultaneously to complete the intelligent closed-loop management and control of pharmaceutical production.

[0007] Further, step S100 includes: S110. Extract the original material data to obtain the initial material attribute time series; S120. Align the initial material attribute time sequence to obtain the material attribute time sequence characteristics; S130. Map the time-series characteristics of material attributes to determine the attribute matrix of the equipment-bound nodes; S140. Aggregate the attribute matrix of the equipment-bound nodes to obtain the set of material attribute features.

[0008] Further, step S200 includes: S210. Input the set of material attribute features into the pre-trained time-series feature extraction model to obtain the material time-series feature attention weight vector; S220. Based on the attention weight vector of material time sequence characteristics, perform calculations on the set of material attribute characteristics to determine the sequence of production process target feature vectors. S230. Decode the production process target feature vector sequence to obtain the production process deviation value of a single process node. S240. If the production process deviation value of a single process node is less than or equal to the preset process deviation trigger threshold, then generate the equipment status compensation amount and generate the production parameter adjustment instruction corresponding to the current production batch based on the equipment status compensation amount; if the production process deviation value of a single process node is greater than the preset process deviation trigger threshold, then generate the instruction blocking flag and the manual review work order.

[0009] Further, step S300 includes: S310. Obtain the data field set of multiple heterogeneous systems. The data field set of multiple heterogeneous systems is obtained by parsing the production parameter adjustment command. S320. Obtain the semantic feature vector of heterogeneous fields based on the data field set of multiple heterogeneous systems, and determine the standard semantic mapping entity based on the semantic feature vector of heterogeneous fields; S330. Generate a cross-system unified format data payload based on the standard semantic mapping entity. The cross-system unified format data payload is obtained by performing data structure reorganization on data fields of multiple heterogeneous systems. S340. If the data payload integrity check value of the cross-system unified format data payload is greater than or equal to the preset data integrity threshold, and the number of missing mandatory fields is zero, then the cross-system unified format data payload is subjected to event encapsulation processing to obtain standardized collaborative control events adapted to each business system.

[0010] Further, step S400 includes: S410. Obtain the execution sequence of the event's built-in logic. The execution sequence of the event's built-in logic is obtained by parsing the control logic built into the standardized collaborative control event. S420. Compare the execution sequence of the event's built-in logic with the preset production deviation rules to obtain the matching degree of the production deviation rules; S430. If the matching degree of the production deviation rule is greater than the preset rule matching threshold, then an initial early warning data packet is constructed based on the batch operation status corresponding to the current production batch. S440: Perform field normalization, abnormal field marking, and signal encapsulation processing on the initial early warning data packet to generate a standardized production abnormality early warning signal.

[0011] Further, step S500 includes: S510. Input the production anomaly warning signal into the pre-trained integrated learning classification model and perform deep feature extraction to obtain a multi-dimensional anomaly deep feature vector. S520. Based on the multidimensional deep feature vector of anomalies, perform anomaly category classification processing to obtain the initial anomaly category label; S530. Extract the batch-bound production operation status data corresponding to the current production batch, and determine whether the initial abnormal category label matches the preset high-risk abnormal category. S540. Perform task scheduling priority sorting on batch-bound production operation status data and initial abnormality category labels to determine the production task scheduling sequence used to guide subsequent production operations; when a preset high-risk abnormality category is matched, increase the scheduling priority of the corresponding task.

[0012] Further, step S600 includes: S610: Parse the built-in task priorities in the production task scheduling sequence, perform task diversion processing, and determine the hardware target execution node corresponding to the diverted task. S620: Combine the load status data of the hardware target execution node of the diversion task with the diversion task to perform Boolean logic operations and generate the initial action sequence of the device; S630. If the initial action sequence of the device contains a concurrent conflict flag, the time slice round-robin algorithm is used to rearrange the initial action sequence of the device to obtain a conflict-free device action sequence. S640: Based on the register addresses corresponding to the conflict-free device action sequence, perform instruction encoding to generate low-level production equipment control instructions for driving the operation of the underlying hardware.

[0013] Further, step S700 includes: S710: Obtain the identity identifier of the control command of the underlying production equipment, and obtain the operation permission verification result based on the identity identifier; S720. If the operation permission verification result is the edit operation permission level, then obtain the normalized messages of the underlying production equipment control instructions in the instruction generation stage and the instruction execution stage respectively, calculate the hash characteristics of the normalized message generation stage and the hash characteristics of the execution stage, and perform a consistency comparison between the hash characteristics of the generation stage and the hash characteristics of the execution stage. S730. If the hash characteristics of the generation stage are consistent with those of the execution stage, the integrity verification result is passed; if the hash characteristics of the generation stage are inconsistent with those of the execution stage, the automatic execution of the control instructions of the underlying production equipment is blocked and a manual review work order is generated; if at the same time, if the key control field is detected to be outside the preset process range or the identity is abnormal, the integrity verification result is determined to be rejected. S740. If the GMP compliance verification result is passed, the underlying production equipment control instructions are executed, and a GMP compliance audit trail record is generated based on the status change data generated during the execution process.

[0014] Another aspect of the present invention relates to a multi-system integrated intelligent control system for pharmaceutical production execution, used to implement the above-described multi-system integrated intelligent control method for pharmaceutical production execution, comprising: The material attribute feature set acquisition module is used to collect material attribute data from multiple different collection nodes in the entire pharmaceutical production process, construct and acquire the material attribute feature set. The material attribute feature set contains multiple sets of material attribute time-series features, and each material attribute time-series feature is configured with a collection timestamp and collection device identification information. The production parameter adjustment instruction acquisition module is used to call the pre-trained time-series feature extraction model to perform time-series dependency analysis and deep feature extraction on all material attribute time-series features in the material attribute feature set, to explore the correlation between material attribute time-series changes and pharmaceutical production parameters, and to output the production parameter adjustment instruction corresponding to the current production batch. The standardized collaborative control event acquisition module is used to perform semantic mapping matching and unified format conversion processing on the multiple heterogeneous system data fields contained in the production parameter adjustment instructions, eliminate the data heterogeneity differences of multiple business systems, and generate standardized collaborative control events that are adapted to various business systems in pharmaceutical production. The production anomaly warning signal generation module is used to parse the production control logic built into the standardized collaborative control event, compare and verify the operating parameters of the standardized collaborative control event with the preset production deviation rules, and generate the production anomaly warning signal corresponding to the current production batch in real time if the preset production deviation rules are triggered. The production task scheduling sequence acquisition module is used to input the production anomaly warning signal into the pre-trained ensemble learning classification model. The ensemble learning classification model completes the deep feature extraction of production anomalies and the accurate classification of anomaly categories, and outputs a production task scheduling sequence to guide subsequent pharmaceutical production operations. The underlying production equipment control instruction acquisition module is used to perform intelligent diversion and logical operation processing on various production tasks according to the task priority sorting rules built into the production task scheduling sequence, and generate underlying production equipment control instructions that can directly drive the operation of underlying hardware devices. The GMP compliance audit trail generation module is used to perform dual verification of operation permission and data integrity compliance for the control instructions of the underlying production equipment. If the control instructions of the underlying production equipment pass the verification, the control instructions of the underlying production equipment are executed, and a compliance audit trail record adapted to GMP standards is generated simultaneously to complete the intelligent closed-loop management and control of pharmaceutical production.

[0015] The beneficial effects achieved by this invention are as follows: This invention provides a multi-system integrated intelligent control method and system for pharmaceutical production execution, solving the technical problems of real-time collaboration of multi-source heterogeneous data, delayed response to production anomalies, and lack of compliance traceability for equipment control commands in traditional production systems. This invention uses a pre-trained temporal feature extraction model to perform deep analysis of material attribute data collected from multiple nodes, automatically identifying temporal dependencies in the production process and generating parameter adjustment commands. Then, through semantic mapping and format conversion, standardized collaborative control of heterogeneous systems is achieved. When a production deviation is detected, an early warning mechanism is immediately triggered, and an anomaly identification and task scheduling optimization are performed through an integrated learning classification model, dynamically generating equipment control commands based on task priority. Finally, the compliance of control commands is ensured through permission verification and integrity verification mechanisms, and audit trail records are automatically generated. This invention achieves intelligent closed-loop management of the entire process from data acquisition to equipment control, significantly improving the real-time response capability, anomaly handling efficiency, and operational compliance of the production system, providing a complete technical solution for the digital transformation of industrial production. The specific beneficial effects achieved are as follows: 1. This invention is equipped with a multi-node linkage acquisition mechanism to obtain the time-series attribute characteristics of materials with timestamps and equipment identifiers, realizing the integrated binding and collection of material data, equipment traceability data, and time dimension data throughout the entire pharmaceutical production chain. It breaks through the shortcomings of traditional pharmaceutical production material data fragmentation, node data separation, and lack of traceability information, and constructs a full-dimensional traceable underlying production dataset to solidify the data foundation for intelligent management and control.

[0016] 2. This invention relies on a pre-trained temporal feature extraction model to mine the temporal dependency relationship of material attributes, deeply explore the dynamic evolution law of material attributes and the coupling mechanism between material state and production parameters, get rid of the traditional mode of manually adjusting production parameters based on experience, realize the adaptive intelligent output of production parameters, and improve the matching degree of production parameter control and the accuracy of process adaptation.

[0017] 3. This invention completes semantic mapping and format unification conversion for data from heterogeneous systems across different business areas, eliminating data barriers and field heterogeneity conflicts among multiple business systems such as pharmaceutical manufacturing MES, material control, and equipment operation and maintenance, realizing data interconnection and interoperability and business collaboration among multiple systems, and solving the industry pain points of traditional independent operation of multiple systems, inability to communicate instructions, and difficulty in collaborative management.

[0018] 4. This invention relies on pre-set production deviation rules to analyze standardized collaborative events, enabling the early identification and warning triggering of production process deviations and material status anomalies. It shifts from post-event rectification to pre-event prediction and intervention, accurately identifying hidden process deviations and explicit material anomalies in the production process, and improving the accuracy of identifying abnormal operating conditions in pharmaceutical production and the foresight of control.

[0019] 5. This invention employs an integrated learning classification model to complete the deep feature mining and accurate classification of production anomalies. Based on the anomaly assessment results, it generates a priority-based production task scheduling sequence, breaking through the limitations of traditional anomaly control that only provides alarms and lacks hierarchical scheduling. It achieves integrated analysis of anomaly cause tracing, risk level classification, and production scheduling optimization, deepening the understanding and mining of production anomalies.

[0020] 6. This invention relies on task priority to complete intelligent task diversion and logic parsing, and completes the closed-loop transformation from business scheduling instructions to underlying equipment drive control instructions from top to bottom. It fits the pharmaceutical production process layout logic and equipment linkage mechanism, realizes seamless connection between upper-level business control and lower-level production hardware execution, and improves the overall production execution control continuity and implementation.

[0021] 7. This invention adds instruction permission verification and production data compliance verification steps, and simultaneously generates a GMP compliance audit tracking ledger, which meets the stringent requirements of production compliance management in the pharmaceutical industry. It realizes the control of equipment operation authority, closed-loop record keeping of production data, and full-process audit traceability, taking into account both intelligent production control and the compliance management needs of the pharmaceutical industry. It is fully adaptable to multi-system integrated production scenarios in the pharmaceutical industry, and the control logic is in line with pharmaceutical production process specifications and industry regulatory standards. The system has strong compatibility, business practicality and compliance practicality, effectively optimizing the production scheduling process, reducing the risk of human error, and strengthening the pharmaceutical production quality control system. Attached Figure Description

[0022] Figure 1 This is a flowchart illustrating an embodiment of the intelligent control method for multi-system integrated pharmaceutical production execution according to the present invention. Figure 2 This is a functional block diagram of an embodiment of the intelligent control system for pharmaceutical production execution that integrates multiple systems according to the present invention.

[0023] Explanation of icon numbers: 10. Material attribute feature set acquisition module; 20. Production parameter adjustment instruction acquisition module; 30. Standardized collaborative control event acquisition module; 40. Production anomaly early warning signal generation module; 50. Production task scheduling sequence acquisition module; 60. Underlying production equipment control instruction acquisition module; 70. GMP compliance audit trace record generation module. Detailed Implementation

[0024] To better understand the above technical solutions, the following will provide a detailed explanation of the technical solutions in conjunction with the accompanying drawings and specific implementation methods.

[0025] like Figure 1As shown, the first embodiment of this invention proposes an intelligent management and control method for pharmaceutical production execution that integrates multiple systems. This method belongs to the technical fields of pharmaceutical MES production execution management and control, integration of multiple heterogeneous business systems, time-series material feature mining, pharmaceutical production deviation early warning, integrated learning anomaly classification, GMP compliance audit tracking, and pharmaceutical workshop equipment linkage scheduling. This invention establishes a fully closed-loop process architecture: "multi-node time-series material feature collection - AI time-series feature extraction - heterogeneous system semantic format conversion - production deviation rule triggering and discrimination - integrated learning anomaly classification - task priority diversion and scheduling - instruction compliance verification - GMP audit tracking archiving." It is adaptable to the multi-system integrated intelligent production management and control conditions of pharmaceutical workshops producing all categories of pharmaceuticals, including solid dosage forms, sterile injections, and active pharmaceutical ingredients. The method includes the following steps: Step S100: Collect material attribute data from multiple different collection nodes in the entire pharmaceutical production process, construct and obtain a set of material attribute features. The set of material attribute features contains multiple sets of material attribute time-series features, and each material attribute time-series feature is configured with a collection timestamp and collection device identification information.

[0026] The system collects data from four types of production data points across the pharmaceutical workshop's WMS (Warehouse Management System), including data collection nodes for warehousing, intermediate transfer, reaction vessel feeding, and quality inspection sampling. It collects real-time monitoring parameters for the entire lifecycle of pharmaceutical raw materials, intermediates, and finished APIs. It binds the timestamp of each feature-based time-series sampling moment and the unique hardware code of the on-site data collection sensor to structure and encapsulate multi-dimensional time-series monitoring parameters. It also collects and constructs a set of material attribute features with traceability tags and multi-node linkage, thus completing the standardized collection of underlying time-series data sources for pharmaceutical production materials.

[0027] Material attribute feature set: Under the pharmaceutical production single production batch, the structured dataset of all material monitoring attributes across multiple physical acquisition nodes in workshops, time-series aligned, and bound to traceability tags serves as the globally unique original input data source for this method.

[0028] Data collection nodes: Standardized physical data collection points in the pharmaceutical production chain, divided into five categories of compliant GMP data collection nodes: warehouse constant temperature storage nodes, material transfer and conveying nodes, reaction kettle feeding and processing nodes, offline LIMS quality inspection and sampling nodes, and finished product temporary storage nodes.

[0029] Material attribute time-series characteristics: Dynamic monitoring attribute parameters of materials that change dynamically with production time, excluding static inherent parameters such as material name and batch number; including six core pharmaceutical time-series process characteristics: real-time environmental temperature and humidity, raw material purity, material viscosity, moisture content, initial microbial colony count, and powder flowability.

[0030] Data Acquisition Timestamp: A millisecond-level time-series traceability marker field for the original sampling of material time-series characteristics, used for subsequent time-series feature alignment and production time-series backtracking audits. Timestamp Sampling Accuracy Threshold: Millisecond-level accuracy 1~5ms; Value Basis: Meets the accuracy requirements for pharmaceutical GMP production time-series traceability and deviation time-series review audits.

[0031] Equipment identification information: On-site temperature sensors, moisture analyzers, and material spectral quality inspection equipment have a unique, tamper-proof hardware code, which is a core field for equipment traceability GMP auditing. Equipment code character range: 16~24 bits encrypted character code; value based on: pharmaceutical company workshop equipment ledger encryption traceability archiving standards.

[0032] Step 4 Data Source Acquisition Architecture: ① Parallel Sampling Layer with Multiple Acquisition Nodes in the Workshop → ② Dynamic Material Attribute Time Sequence Aggregation Layer → ③ Timestamp Time Sequence Binding Layer → ④ Unique Equipment Identifier Traceability Encapsulation Layer.

[0033] The material property time-series characteristics of this invention are divided into two categories: one is online continuous acquisition data (temperature, humidity, flow rate, rotation speed, pressure, sampling period 2~6s), and the other is LIMS offline inspection data (purity, microbial colony count). Offline data is only bound to the batch and sampling time and does not participate in second-level resampling. The total number of acquisition nodes has no fixed value and is configured according to the production line process verification plan, with a value ≥2. The original sampling data and processed features are stored separately. All interpolation and alignment operations retain the original millisecond-level timestamps. DTW nonlinear stretching of the time axis is not used; only unified clock linear interpolation is used. The equipment identifier is only used as an index field and does not participate in matrix numerical operations. The matrix aggregation result is only used as model input, and the original record index of each node is completely retained in the audit process.

[0034] Step S200: Call the pre-trained time-series feature extraction model to perform time-series dependency analysis and deep feature extraction on all material attribute time-series features in the material attribute feature set, explore the correlation between material attribute time-series changes and pharmaceutical production parameters, and output the production parameter adjustment instructions corresponding to the current production batch.

[0035] The model extracts time-series features from an offline industrial pharmaceutical production sample dataset obtained from a pharmaceutical company's server. It is trained in a supervised manner and has its parameters frozen. The model then imports the full-domain time-series features of the material attribute feature set output from step S100 into the model in batches. The model automatically mines cross-node material time-series coupling dependencies and parameter time-series fluctuation correlations, filters strongly correlated core process features, and removes redundant and noisy time-series features. Based on the feature-process parameter mapping knowledge base, it adaptively tunes the core production process parameters of the workshop reactor, including temperature control, stirring speed, and feeding rate. It then outputs closed-loop production parameter adjustment instructions bound to the current drug production batch number and adapted to material time-series fluctuation conditions.

[0036] The production parameter adjustment instructions output in this step are only suggestions for fine-tuning the process. The system has a built-in library of verified upper and lower limits for the process. If the target parameter in the production parameter adjustment instruction exceeds the verified process range for the product, the system will automatically intercept the instruction and forward it downstream, generate a manual review work order, and prevent the equipment from executing it automatically.

[0037] Temporal Feature Extraction Model: A pre-trained lightweight bidirectional LSTM temporal feature extraction model specifically designed for pharmaceutical companies. It is trained offline based on 12,000+ compliant pharmaceutical raw material and formulation production timeline material samples. It can be deployed in a lightweight embedded MES system and is specifically designed for material timeline dependency mining and process feature screening.

[0038] Temporal dependency: The correlation between the temporal fluctuations of temperature and humidity of materials stored in the preceding warehouse and the purity parameters of materials processed in the subsequent workshop is the core correlation feature of pharmaceutical process linkage control.

[0039] Production parameter adjustment instructions: Digital instructions for adjusting workshop equipment processes, which are bound to production batch number, process station, parameter setting range, and execution time. They include four types of control fields: reactor temperature, raw material feeding flow rate, stirring frequency, and constant temperature chamber humidity.

[0040] The complete seven-layer algorithm architecture for the temporal feature extraction model consists of: ① Multi-dimensional temporal feature normalization input layer → ② Temporal missing value GMP compliance completion layer → ③ Bidirectional LSTM temporal dependency coupling mining layer → ④ Redundant noise feature filtering layer → ⑤ Weighted aggregation layer of core process features → ⑥ Material-process parameter matching and mapping layer → ⑦ Batch binding parameter adjustment instruction output layer.

[0041] The temporal feature extraction model is a bidirectional attention LSTM. The training samples are historical batch online data, LIMS offline reports, and process adjustment records approved by QA. The labels are the compliant process adjustment ranges. After the model is trained offline, the parameters are permanently frozen. During the production operation phase, only inference is performed and the weights are not updated.

[0042] Step S300: For the heterogeneous system data fields contained in the production parameter adjustment instructions, perform semantic mapping matching and unified format conversion processing to eliminate the data heterogeneity differences among multiple business systems and generate standardized collaborative control events that are adapted to various business systems in pharmaceutical production.

[0043] The process involves disassembling the native message of the production parameter adjustment instruction output by step S200, and separating the differentiated native data fields of four heterogeneous business systems: MES execution system, WMS warehouse system, SCADA equipment control system, and LIMS quality inspection system. A cross-system semantic association mapping dictionary specifically for pharmaceutical production is built, completing semantic alignment of heteronymous fields, unified data encoding, standardized conversion of message frame formats, and removal of redundant messages. This eliminates syntactic, semantic, and message format barriers between multiple systems, generating standardized structured collaborative control events that can be synchronously parsed and executed by all pharmaceutical production business systems.

[0044] Data fields from multiple heterogeneous systems: native non-standard differentiated message fields from multiple business systems, including MES workstation instruction fields, WMS material outbound batch number fields, SCADA equipment operating point fields, and LIMS quality inspection result non-standard encoded fields. The semantics, byte lengths, and encoding rules of these fields are incompatible with each other.

[0045] Semantic mapping: A data fusion processing technique that involves one-to-one binding and matching of synonymous and heteronymous fields across systems and normalization and alignment of semantic tags.

[0046] Format conversion: Non-standard variable-length messages are uniformly converted into structured JSON fixed-frame messages, which are the standard for pharmaceutical manufacturing.

[0047] Standardized collaborative control events: standardized business control data packets with unified semantics, unified message format, system permission tags, GMP traceability fields, and cross-system synchronous parsing and linkage execution.

[0048] Cross-system field mapping matching accuracy threshold: semantic mapping matching accuracy ≥ 99.6%; value basis: to avoid heterogeneous field mismatch and command cross-system delivery failure, and to meet the pharmaceutical company's multi-system integration and acceptance standards.

[0049] Message format unified byte threshold: Standardized event fixed message byte length 256 bytes; value range 224~256 bytes; value basis: pharmaceutical industry IoT workshop bus standard communication message specification.

[0050] Step 5 Heterogeneous Data Fusion Architecture: ① Heterogeneous Original Field Decomposition Layer → ② Non-standard Redundant Field Removal Layer → ③ Cross-system Semantic Dictionary Mapping Layer → ④ Industry Standard Message Format Compilation Layer → ⑤ Standardized Collaborative Event Encapsulation and Output Layer.

[0051] During implementation, a standardized field mapping table is used to complete the mapping. The table contains source fields, standard fields, unit conversions, and value ranges. Semantic vectors are only used for field matching recommendations. If the matching confidence is lower than the threshold, manual confirmation is pushed, and events are not automatically generated.

[0052] Step S400: Analyze the production control logic built into the standardized collaborative control event, compare and verify the operating parameters of the standardized collaborative control event with the preset production deviation rules, and if the preset production deviation rules are triggered, generate the production anomaly warning signal corresponding to the current production batch in real time.

[0053] The system breaks down the underlying control logic of standardized collaborative control events into three categories: built-in process control logic, material timing linkage logic, and equipment linkage start-stop logic. It then retrieves the pharmaceutical company's GMP compliance production deviation rule library and compares each event's process parameters, material timing fluctuation amplitude, and equipment linkage timing parameters. The system determines whether the collaborative control event exceeds the compliance process boundary and triggers the solidified production deviation judgment rules. For out-of-standard production conditions that trigger deviation rules, the system binds the production batch, workstation, deviation type, and traceability equipment identifier to generate a closed-loop production anomaly warning signal with multi-dimensional traceability tags.

[0054] Control logic: A set of standardized collaborative control events with built-in process parameter adjustment logic, cross-system linkage timing logic, and equipment start-stop interlocking underlying operation logic.

[0055] Pre-set production deviation rules: Based on the three categories of compliance deviation judgment rules solidified in the "GMP Good Manufacturing Practice for Pharmaceuticals", namely, exceeding the standards of pharmaceutical production process, exceeding the standards of material timing fluctuation, and misalignment of equipment linkage timing.

[0056] Production anomaly warning signal: A digital workshop production warning signal that includes deviation level, deviation source location, production batch, deviation trigger rule code, and rectification time limit label.

[0057] The deviation rule five-layer judgment architecture consists of: ① event built-in control logic decomposition layer → ② GMP deviation rule library retrieval layer → ③ multi-dimensional parameter comparison layer → ④ deviation condition logic judgment layer → ⑤ graded abnormality early warning signal generation layer.

[0058] The warning signal classification label range is divided into three levels: Level 1 minor deviation, Level 2 process deviation, and Level 3 compliance risk deviation; the value is based on the pharmaceutical company's workshop abnormal handling classification operation standard.

[0059] Step S500: Input the production anomaly warning signal into the pre-trained ensemble learning classification model. The ensemble learning classification model is used to extract deep features of production anomalies and accurately classify anomaly categories, and output a production task scheduling sequence to guide subsequent pharmaceutical production operations.

[0060] Using multi-label production anomaly early warning signals as the model input data source, an ensemble learning classification model with pre-trained frozen parameters is imported into the pharmaceutical production anomaly sample dataset. The model integrates early warning signal source traceability features, deviation time series features, and deep hidden features of material attribute anomalies to complete in-depth mining of high-dimensional anomaly features. Relying on the ensemble weak learner voting classification mechanism, it automatically classifies four types of anomalies: material raw material anomalies, equipment linkage anomalies, system coordination anomalies, and environmental time series anomalies. The anomaly type is matched and bound to workshop production workstations, personnel shifts, equipment start-up and shutdown, and material return and rectification work orders to generate a structured, time-ordered, workstation-bound, and executable production task scheduling sequence.

[0061] The ensemble learning classification model is a stacking-style ensemble learning classification model specifically for pharmaceutical production. The base learner includes a CNN feature extraction sub-model and a LightGBM classification sub-model, and the meta-learner is a logistic regression classifier. It has been supervised pre-trained using a large number of abnormal samples from real pharmaceutical production.

[0062] The deep-seated features of anomalies are the hidden high-dimensional features beyond the surface labels of warning signals, such as the deviation time-series decay rate and cross-system linkage anomaly coupling.

[0063] Anomaly Categories: Four primary categories are uniformly adopted: material attribute anomalies, equipment linkage anomalies, system coordination anomalies, and production environment anomalies. Temperature control anomalies, feed flow rate anomalies, etc., are retained as secondary sub-categories.

[0064] Production task scheduling sequence: an integrated set of time-series task work orders that binds exception type, work station, executor number, task sequence timestamp, and rectification priority for orderly production rectification, production extension, and material withdrawal.

[0065] The complete eight-layer algorithm architecture of the ensemble learning model is as follows: ① Multi-label early warning signal normalization input layer → ② Shallow anomaly feature splitting layer → ③ Dual-base learner parallel deep feature extraction layer → ④ Anomaly feature weight fusion layer → ⑤ Stacked feature dimensionality reduction layer → ⑥ Meta-learner voting classification layer → ⑦ Anomaly type label output layer → ⑧ Task work order sequence matching generation layer.

[0066] Step S600: Based on the task priority sorting rules built into the production task scheduling sequence, perform intelligent diversion and logical operation processing on various types of production tasks to generate underlying production equipment control instructions that can directly drive the operation of underlying hardware devices.

[0067] The system analyzes the built-in hierarchical task priority coding tags in the production task scheduling sequence, classifying tasks into four priority categories: emergency rectification, routine control, material transfer, and quality inspection re-inspection. It establishes Boolean logic operation rules for task distribution in the workshop, eliminating conflicting and coupled tasks, merging duplicate scheduling tasks from the same source, and intercepting mutually exclusive tasks. It completes multi-task parallel distribution, interlocking logic Boolean operations, and equipment station task matching operations. It adapts to the underlying hardware communication protocols of reactors, constant temperature storage units, material conveying pumps, and online quality inspection sensors, compiling and generating control commands for underlying production equipment that can directly drive the start / stop and parameter tuning of the workshop's underlying automation hardware.

[0068] Task Priority: The scheduling sequence has a built-in GMP compliance task urgency level code, which is generated by positively binding the production compliance risk level.

[0069] Level 4 task priority thresholds (fixed values ​​+ basis): Level 1 highest priority (Level 3 compliance risk anomaly, emergency shutdown for rectification), Level 2 priority (Level 2 process deviation, parameter fine-tuning without shutdown), Level 3 priority (Level 1 minor deviation, delayed scheduling in the background), Level 4 routine operation and maintenance priority (daily inspection and operation and maintenance tasks); Value basis: GMP operation management procedures for emergency response in drug production.

[0070] Task routing and logical operations: a Boolean interlocking logical operation assembly for task mutual exclusion judgment, duplicate task merging, high-priority task priority scheduling, and conflict task interception.

[0071] Low-level production equipment control instructions: Adapted to workshop PLC controllers, the lower-level machine driver level + digital linkage control instructions execute the hardware native protocol.

[0072] Task conflict interception accuracy threshold: 100% interception rate for mutually exclusive production tasks; Value basis: to prevent conflicts between multiple equipment instructions in the pharmaceutical workshop and avoid the scrapping of batches of drugs.

[0073] Step 6: Task scheduling architecture: ① Scheduling sequence priority label parsing layer → ② Multi-level task priority sorting layer → ③ Timing mutual exclusion task discrimination layer → ④ Parallel task splitting layer → ⑤ Device interlocking logic Boolean operation layer → ⑥ Low-level hardware driver instruction compilation layer.

[0074] Step S700: Perform dual verification of operation permission and data integrity compliance for the control instructions of the underlying production equipment in sequence. If the control instructions of the underlying production equipment pass the verification, execute the control instructions of the underlying production equipment and generate a compliance audit trail record adapted to GMP standards at the same time to complete the intelligent closed-loop management and control of pharmaceutical production.

[0075] The system employs a two-dimensional parallel approach to complete the closed-loop compliance verification of underlying equipment instructions: First, it retrieves the workshop's job permission ledger to verify the matching degree of instruction-issuing accounts, equipment operation positions, and anomaly handling permissions, completing the closed-loop verification of personnel and equipment operation permissions. Second, it verifies the integrity of instruction messages, traceability tags, batch codes, and timestamp fields based on a hash digest algorithm, completing the GMP integrity compliance verification of production data. After both dimensions pass verification, instructions are issued to drive the underlying hardware to execute production scheduling actions. Simultaneously, it captures data sources, model parameters, early warning information, instruction ledgers, and operator information for the entire process, automatically encrypting and archiving them to generate tamper-proof GMP compliance audit traceability records adapted for drug regulatory traceability verification, completing the intelligent closed-loop management of the entire process.

[0076] Operation permission verification: Verify the matching of authority and responsibility for pharmaceutical production positions, and verify whether the account issuing the instruction has the corresponding operation permission for modifying equipment parameters and handling abnormalities.

[0077] Data integrity compliance verification: Based on the GMP data integrity ALCOA+ principle, verify that the instruction traceability field, batch field, and timestamp field are free from missing, tampering, and packet loss compliance checks.

[0078] GMP compliance audit tracking records: encrypted read-only, tamper-proof, compatible with online drug regulatory inspections, and linked to the entire production chain traceability information of the electronic audit archive ledger for drug production.

[0079] The seven-layer compliance verification closed-loop architecture consists of: ① Bottom-level device control instruction receiving layer → ② Personnel position operation permission verification layer → ③ Instruction hash digest integrity verification layer → ④ Two-dimensional compliance result discrimination layer → ⑤ Compliance instruction hardware issuance and execution layer → ⑥ Full-process traceability data capture layer → ⑦ Encrypted GMP audit ledger archiving layer.

[0080] This embodiment provides a multi-system integrated intelligent management and control method for pharmaceutical production execution, addressing eight major industry pain points in existing pharmaceutical MES production management systems: heterogeneous and fragmented data across multiple systems, lack of material time-series feature mining, large deviations in manually tuned production parameters, high false alarm rates for production anomalies, poor accuracy of single-model anomaly classification, lack of priority control in task scheduling, unauthorized issuance of instructions, and lack of automatic GMP audit tracking. It leverages core processes such as multi-node material time-series feature acquisition, bidirectional LSTM time-series model adaptive parameter adjustment, cross-system semantic heterogeneous fusion, GMP-bound deviation rule discrimination, stacking ensemble learning anomaly classification, priority task logic routing, dual-dimensional instruction compliance verification, and automatic audit ledger archiving. This achieves seamless integration and linkage across multiple pharmaceutical business systems, reducing production process parameter control errors by over 62% and lowering the false alarm rate for production anomalies to below 1.5%. It fully complies with the mandatory compliance requirements of drug regulatory GMP regarding data integrity, job responsibilities, and production traceability, eliminating the need for manual ledger entry and adapting to intelligent and compliant production execution and control operations in all categories of pharmaceutical workshops, including APIs, sterile preparations, and solid dosage forms.

[0081] Furthermore, the intelligent control method for multi-system integrated pharmaceutical production execution provided in this embodiment includes step S100 as follows: Step S110: Extract the original material data to obtain the initial material attribute time series sequence.

[0082] Five GMP-compliant data collection nodes across workshops capture raw material monitoring messages and equipment sampling data that have not been noise-reduced or time-calibrated. Redundant frame headers and invalid empty message impurities in equipment communication are removed. Continuous time-series raw material monitoring parameters are collected, linearly arranged according to sampling time, and an initial material attribute time-series sequence is generated.

[0083] The initial material attribute time sequence is a set of raw sampled values ​​arranged according to sampling time; each raw sampled value is associated with an original record identifier, which includes the device identifier, acquisition node identifier, production batch number, and original sampling timestamp. The original record identifier is retained during subsequent data alignment, interpolation, or normalization processing.

[0084] The formula for calculating the initial material property time series is: (1) In formula (1), This is an initial time-series sequence of material attributes, without units, derived from the raw output of LIMS material detection sensors and workshop IoT acquisition terminals, with dimensions of [missing information]. Characterizes the physical property data of raw materials across nodes; For the first The first data collection node, the first The original attribute value of a single material corresponding to a timestamp, and the physical property parameters such as temperature, viscosity, and particle size are matched as needed; This is the index of the data collection node, without units, and its value range is [value range missing]. ; This represents the total number of distributed data collection nodes for pharmaceutical materials, without a unit, and is determined based on the production line equipment configuration, process steps, and validated data collection scheme. The number of data collection nodes is an integer greater than or equal to 2. The number of data collection nodes may vary for different production lines or different dosage forms and is not limited to a fixed value. It includes nodes such as feeding port, reaction vessel, transfer silo, and discharge port. This is the time index for time-series sampling, unitless, with a value range of [value range missing]. ; This represents the total sampling time for a single batch of materials, in seconds, with a value range of [value range missing]. s, matching the intermittent production cycle of pharmaceuticals. The control logic of formula (1) is to bind the acquisition node number and time sequence timestamp, batch extract the original physical property data of materials across points, retain the original time and equipment identification tags, and at the same time remove the industrial control null value noise data to construct a structured and traceable initial material attribute time sequence. This formula (1) embeds equipment ID and timestamp dual traceability tags, which meets the basic requirements of pharmaceutical GMP material traceability, and is different from the disordered data acquisition method of ordinary industry. It provides a traceable original data foundation for subsequent production parameter adjustment, abnormal warning and compliance audit, while effectively reducing the interference of noise data on subsequent model training and control logic.

[0085] Initial material attribute time series: a native one-dimensional material parameter time series array with parallel sampling by multiple acquisition nodes, time series misalignment, no bound device identifier, and slight sampling noise.

[0086] The three-layer data extraction architecture for the sub-step is as follows: ① Parallel capture layer of raw material messages from multiple nodes → ② Communication redundancy and impurity message removal layer → ③ Linear arrangement layer of original time series.

[0087] The initial material attribute time sequence only stores the original sampled values. The device ID, timestamp, and batch number are used as independent associated indexes and are not embedded in the sequence values. Formula (1) only represents the original data collection logic, and the original record identifier is retained throughout the process without being lost.

[0088] Step S120: Align the initial material attribute time sequence to obtain the material attribute time sequence features.

[0089] Online data acquisition is timestamped using a unified clock source and resampled at a preset sampling period. For data not acquired at the target sampling time, linear interpolation is performed based on adjacent valid sample values, while retaining the original sampling timestamp and interpolation identifier. The raw data used for audit trails is not subject to timeline distortion.

[0090] The formula for calculating the time-series characteristics of material properties is: (2) (3) In formula (2)-(3), For the aligned first The node, the first The interpolated material attribute value at each time point, with the unit matching the corresponding material property parameter; , For the first Each node has the original asynchronous sampling start and end timestamps. , Material attribute values ​​at the location; These are the start and end timestamps of the original asynchronous sampling, in seconds. The target alignment timestamp is given by the GMP-standardized timing sampling step size, in seconds. The time sequence characteristics of the aligned material attributes are unitless and are the core output of this step. This is a dedicated alignment operator for pharmaceutical time series, with no units, that performs batch linear interpolation and time series resampling on the initial material attribute time series sequence; The sampling step size is standardized for GMP, with the unit being seconds (s). It is fixed at 5 seconds and conforms to the unified sampling standard for pharmaceutical companies' industrial control systems. For the initial material attribute time series, the definition in formula (1) is used, and the explanation is not repeated here. The control logic of formulas (2)-(3) is to calculate the material attribute value under the standardized time step by using the linear interpolation formula (2) to address the timestamp deviation caused by asynchronous sampling between different devices; then, the entire sequence is batch resampled by the pharmaceutical time series alignment operator to uniformly align all node data to the GMP standardized time series step. This forms a regular and synchronized material attribute time sequence characteristic. Formulas (2)-(3) eliminate the time difference deviation of heterogeneous sampling of multiple devices by combining linear interpolation and time sequence resampling, solve the industry pain point of asynchronous sampling clock between MES and LIMS systems, and enable cross-system material data to meet the compliance alignment requirements of time sequence data in the pharmaceutical industry, providing a consistent data source foundation in the time dimension for subsequent cross-system collaborative control and compliance audit.

[0091] Timing alignment processing: Based on a unified clock source and linear interpolation algorithm, the sampling time difference between multiple devices is eliminated, and the original time axis is kept free from nonlinear deformation.

[0092] Step S130: Map the time sequence characteristics of material attributes to determine the attribute matrix of the equipment binding node.

[0093] A three-dimensional association mapping function is built to link equipment identifier, acquisition node, and material time sequence features. The aligned material attribute time sequence features, 16-24 bit device encryption code, and physical acquisition node coordinate parameters are mapped to a matrix dimension to construct a three-dimensional device-bound node attribute matrix in which rows correspond to acquisition devices, columns correspond to time sequence features, and dimensions are bound to node coordinates.

[0094] The formula for constructing the device-bound node attribute matrix is: (4) In formula (4), Bind a node attribute matrix to the device; it has no unit and has 1 dimension. This is bound to a unique industrial control ID for the device and is the core output of this step. is the node feature mapping weight matrix, which is unitless, and represents the fixed parameters obtained by training the model on the offline GMP dataset. To align the timing characteristics of material attributes, the definition in formula (3) is used, and the explanation is not repeated here; The node feature bias vector is a unitless, 64-dimensional vector used to adjust the baseline offset of the feature mapping. The control logic of formula (4) is to linearly transform the aligned material attribute time-series features through the node feature mapping weight matrix and bias vector, mapping them into a node attribute matrix bound to the device ID, thereby realizing the structured encoding of cross-node material data and forming a standardized feature matrix that can be directly input into the subsequent time-series feature extraction model. Formula (4) constructs a device-bound node attribute matrix, uniformly encoding cross-node and cross-system material data into a structured matrix, while retaining the unique industrial control ID of the device. This not only meets the requirements of pharmaceutical GMP for data traceability, but also provides a standardized model input format for subsequent time-series dependency analysis, solving the industry pain point of inconsistent data formats in heterogeneous systems and difficulty in directly using them for model training.

[0095] The unique equipment code is converted into an embedded vector and concatenated with features. It is not directly used as a continuous numerical value in linear matrix operations. The matrix only carries the time series data of material properties.

[0096] Equipment-bound node attribute matrix: A structured multidimensional feature matrix that integrates three-dimensional parameters of equipment identification, physical acquisition point location, and material time sequence characteristics, serving as the core carrier for dataset aggregation.

[0097] The sub-step two-layer matrix mapping architecture consists of: ① a three-dimensional relationship binding mapping layer → ② a node-device-feature matrix structured encapsulation layer.

[0098] Step S140: Aggregate the attribute matrix of the equipment binding nodes to obtain the set of material attribute features.

[0099] A weighted matrix aggregation algorithm is used to perform dimension fusion and matrix dimensionality reduction aggregation on the three-dimensional attribute matrix of multiple devices and nodes. Redundant blank dimensions and duplicate feature dimensions are removed from the matrix, the matrix data format is unified, and finally the material attribute feature set with full traceability label is encapsulated and generated as output of step S100.

[0100] When aggregating the attribute matrix of device-bound nodes, the original record index, device identifier, and timestamp corresponding to each acquisition node are retained. The aggregation result is used only as a model input feature and does not replace the original acquisition records. The audit trail record must at least store the set of original record indexes corresponding to the model input features.

[0101] The formula for calculating the set of material attribute characteristics is: (5) (6) In formulas (5)-(6), This is a set of material attribute features, without units, and is the core output of this step, with dual traceability tags of device ID and timestamp; For the first The device binding node attribute matrix corresponding to each acquisition node follows the definition in formula (4), and will not be explained again here; For the first The feature aggregation weights of each collection node are unitless, and the sum of all weights is 1; the weights of core process nodes (such as reactors) are set higher to highlight the influence of material properties on key process sections. The total number of distributed data collection nodes for pharmaceutical materials is defined in formula (1), and will not be explained again here. The control logic of formulas (5)-(6) is to assign different weights to each data collection node based on the process importance, perform weighted aggregation on the attribute matrix of the equipment-bound nodes, and integrate the features of 12 nodes into a unified set of material attribute features. The weight allocation satisfies the normalization constraint, ensuring that the aggregation result can reflect the material status of all nodes at the same time, and giving higher attention to key process nodes. Formulas (5)-(6) achieve effective fusion of material features of multiple nodes through weighted aggregation, while retaining the dual traceability labels of equipment ID and timestamp. This satisfies the requirements of pharmaceutical GMP for data traceability and solves the problem that traditional single-node data cannot fully reflect the overall status of materials. Through differentiated weight settings, the influence of core process nodes is highlighted, and the accuracy of subsequent production parameter adjustment instructions is improved.

[0102] Weighted aggregation is used only as input for model inference. The system independently stores the original matrices of each node before aggregation. Audits retrieve the original single-node data and do not replace the original records with aggregated data.

[0103] Matrix aggregation processing: data integration operations including dimensionality reduction of multidimensional node matrices, removal of redundant dimensions, and weighted fusion of cross-matrix features.

[0104] The three-layer matrix aggregation architecture of the sub-step is as follows: ① Multidimensional matrix input layer → ② Redundant dimension removal and weighted aggregation layer → ③ Standardized feature set encapsulation layer.

[0105] Preferably, the intelligent control method for multi-system integrated pharmaceutical production execution provided in this embodiment includes step S200: Step S210: Input the set of material attribute features into the pre-trained time-series feature extraction model to obtain the material time-series feature attention weight vector.

[0106] The standardized material attribute feature set output in step S100 is batch imported into the input layer of the bidirectional LSTM temporal feature extraction model with frozen parameters. Relying on the built-in temporal correlation weight calculation operator of the model, the coupling correlation weight of material parameters across acquisition nodes and time periods is automatically solved, and the material temporal feature attention weight vector representing the strength of temporal coupling between materials is quantitatively output.

[0107] The formula for calculating the attention weight vector of material time-series features is: (7) In formula (7), This is the attention weight vector for material time-series features, unitless, with dimension 1. This output is broadcast along the time dimension and is the core output of this step. The set of material attribute characteristics is defined using the same formula (5), and will not be explained again here; This is a pre-trained temporal feature extraction model for pharmaceutical processes, without units, and is an LSTM network incorporating an attention mechanism; The set of hyperparameters for the model is frozen and has no units. These are the fixed parameters after the time-series feature extraction model has converged during training on the drug batch dataset and will not be updated during production. The control logic of formula (7) is to input the set of material attribute features into the pre-trained attention LSTM time-series feature extraction model. The time-series feature extraction model captures time-series dependencies through gated recurrent units and weights the features of key time steps through an attention mechanism, outputting an initial dependency weight matrix that reflects the dynamic relationship between material attributes. Formula (7) adopts an LSTM time-series feature extraction model with an attention mechanism, which is different from traditional statistical time-series analysis methods. It can automatically capture nonlinear dependencies between material attributes across nodes and time periods, generate a structured weight matrix, and provide a data-driven decision-making basis for subsequent production parameter adjustments, thus solving the industry pain point of relying on manual experience to adjust parameters in pharmaceutical production.

[0108] Material time-series feature attention weight vector: a two-dimensional weight coefficient matrix that quantifies the coupling correlation between preceding material parameters and subsequent process parameters. The coefficient ranges from 0 to 1, and the larger the coefficient, the stronger the time-series dependency coupling.

[0109] Sub-step model computation architecture: ① Feature set model input layer → ② LSTM temporal weight iterative solution layer → ③ Dependency weight matrix output layer.

[0110] Step S220: Perform calculations on the material attribute feature set based on the material time sequence feature attention weight vector to determine the production process target feature vector sequence.

[0111] The Hadamard product operator is used as a weighting operator to perform matrix weighted convolution operation on the attention weight vector of material time series features and the set of material attribute features. Low-weight invalid noise features are removed and high-weight process correlation features are strengthened. The process is matched and calibrated with the pharmaceutical company's standard process feature benchmark library to generate a production process target feature vector sequence that fits the compliant production conditions.

[0112] The formula for generating the production process target feature vector sequence is: (8) In formula (8), The sequence of feature vectors representing the production process target is unitless and has a dimension of . This is the core output of this step; The attention weight vector for material time sequence features is defined using the same formula (7), and will not be explained again here; The set of material attribute characteristics is defined using the same formula (5), and will not be explained again here; The Hadamard product operator is a unitless matrix operator used to perform element-wise weighted fusion of temporal features and dependency weights. The control logic of formula (8) is to perform Hadamard product operation on the material temporal feature attention weight vector and the material attribute feature set. By using element-wise weighting, it highlights the key material attributes and their temporal dependencies, and fuses them to generate a sequence of production process target feature vectors, providing a target benchmark for subsequent production parameter adjustments. Formula (8) achieves accurate fusion of dependency weights and material features through the Hadamard product operator, directly mapping the nonlinear temporal dependencies captured by the model into a structured target feature vector. This solves the problem of unclear target features and reliance on human experience in traditional process adjustments, and provides a data-driven target basis for the generation of subsequent production parameter adjustment instructions.

[0113] Production process target feature vector sequence: a weighted and optimized time-series process feature vector array that conforms to GMP compliance process standards, serving as the core input for solving process deviations.

[0114] When the dimensions of the weight matrix and the material feature matrix are inconsistent, the dimensions are first unified through linear mapping and then the Hadamard product is performed. The fused output is only used to calculate the deviation and is not used to directly generate equipment instructions.

[0115] Step S230: Decode the production process target feature vector sequence to obtain the production process deviation value of a single process node.

[0116] The decoding operator is invoked to decrypt and decompress the production process target feature vector sequence, restoring a computable real-time process vector. The difference between this vector and the GMP standard process feature vector is calculated, and then divided by the maximum value of the standard vector's L2 norm and the minimum zero-prevention constant to solve for the signed standardized relative deviation, which is used to determine the direction of process parameter adjustment. The absolute value range of the deviation is truncated to the 0~1 interval and converted into a percentage deviation severity to quantify the severity of the process deviation from the standard. This calculation link is equipped with division-by-zero protection and extreme value truncation, resulting in strong numerical stability. It provides integrated output of the process fluctuation adjustment direction and risk classification indicators, providing standardized quantitative basis for automatic closed-loop production control and GMP compliance audit early warning.

[0117] The formula for calculating the production process deviation value of a single process node is: (9) In formula (9), Pre-set standard process feature vectors for GMP compliance, including the standard temperature, pressure, ratio, and duration of each process node defined by drug regulatory standards and production process documents, as the basis for deviation calculation. The equipment acquires process feature vectors in real time with encryption, and the underlying equipment acquires and transmits the original process feature sequence in real time with encryption. For the process feature decoding operator (matching step S230), the encrypted / compressed real-time process vector is decrypted and decompressed to restore readable floating-point process values, thus obtaining a real-time process vector that can participate in the calculation. . This represents the production process deviation value (with sign) for a single process node. This indicates that the real-time process parameters are higher than the standard values ​​and need to be adjusted downwards. This indicates that the real-time process parameters are below the standard values ​​and need to be adjusted upwards. The process was completely without deviation; the denominator To obtain the L2 norm of the standard vector and the minimum zero-prevention constant The maximum value; in the case of all zero standard vectors, the denominator is fixed at 0. This prevents overflow during division by zero operations and improves numerical robustness. The percentage represents the severity of the process deviation. For the range trimming operator, the absolute value of the deviation is locked within an interval. Then magnify it to a percentage; limit the maximum display deviation to 100% to suppress extreme abnormal values ​​from interfering with the classification judgment; the larger the value, the more serious the process deviation from the standard. The L2 norm of the standard process vector represents the overall amplitude of the standard process. Minimal positive zero protection constant, with a fixed value This is suitable for empty process and zero-baseline process nodes. Formula (9) first decodes the encrypted real-time process vector and then calculates the signed standardized relative deviation by combining it with the standard process vector. To determine the direction of parameter adjustment; then, the deviation amplitude range is clipped and converted into a percentage severity. Used for risk classification; enables integrated quantitative assessment of the direction and degree of process fluctuations at single process nodes, supporting automatic production control and GMP compliance early warning audits.

[0118] The production process deviation value is a quantified value of the difference between real-time process characteristics and GMP compliance standard process parameters.

[0119] Step S240: If the production process deviation value of a single process node is less than or equal to the preset process deviation trigger threshold, then generate the equipment status compensation amount and generate the production parameter adjustment instruction corresponding to the current production batch based on the equipment status compensation amount; if the production process deviation value of a single process node is greater than the preset process deviation trigger threshold, then generate the instruction blocking identifier and the manual review work order.

[0120] The reverse correction direction is determined by using a deviation sign function, and the equipment status compensation amount is solved by combining the severity of process deviation and the adaptive compensation coefficient of operating conditions. The compensation amount is constrained within the equipment hardware adjustment range by a value range trimming operator. The production process deviation value of a single process node is compared with the preset process deviation trigger threshold to perform graded handling: if the deviation does not exceed the limit, the compensation amount after the limit is converted into a standardized control message that the equipment can recognize and sent out through the instruction mapping operator to complete the adaptive closed-loop correction of process drift; if the production process deviation value of a single process node exceeds the preset process deviation trigger threshold, the control instruction is set to zero to block automatic adjustment, and a manual review work order is generated and stored simultaneously to realize graded safety control of the production process, which fully complies with the GMP compliance graded handling requirements.

[0121] The formula for generating equipment condition compensation is: (10) In formula (10), This represents the production process deviation value for a single process node. Sign function, Output +1 Output -1 Outputting 0 is used for reverse compensation to adjust the direction; The percentage represents the severity of the process deviation, and the value is [value to be filled in]. The extent to which the characterization process deviates from the standard; The preset process deviation trigger threshold is the maximum allowable deviation limit for automatic compensation set by the GMP process specification. This is the adaptive compensation coefficient for the process. It is a characteristic quantity associated with the current operating condition, equipment model, and material batch; the unit is matched with the controlled process parameters (temperature: ℃ / %, pressure: MPa / %, speed: rpm / %), so as to realize the differentiated compensation sensitivity for different operating conditions; This is the device status compensation value (first line of output); negative sign. To achieve reverse correction, if the process error is too high, output negative compensation to adjust the parameters downward; if the process error is too low, output positive compensation to adjust the parameters upward. For the range pruning operator, the compensation amount is forcibly locked at the upper and lower limits that the device hardware allows for adjustment. To prevent damage to production equipment from exceeding the measurement range; This is the minimum allowable compensation adjustment amount for the equipment; This is the maximum allowable compensation adjustment amount for the equipment. Control commands for adjusting production parameters; For the compensation quantity instruction mapping operator, the numerical compensation quantity is... Converted to a standardized control message format that can be parsed by the underlying device; Branching logic: when To ensure the deviation remains within a safe range, the mapped compensation amount generates a valid automatic adjustment command, which is then sent to the device; when If the deviation exceeds the limit, the instruction is set to 0 (automatic adjustment is blocked), and a manual review alarm work order is generated and stored simultaneously. Formula (10) first calculates the equipment compensation amount based on the direction, magnitude and adaptive coefficient of process deviation and performs hardware range limiting; then outputs in layers according to the preset process deviation trigger threshold: the deviation safety range maps the compensation amount to the equipment automatic adjustment command, and if the deviation exceeds the limit, it blocks the automatic adjustment and pushes the manual review work order, so as to realize the adaptive closed-loop correction of production process drift and GMP graded safety control.

[0122] Equipment condition compensation is a dynamic compensation and correction value for equipment parameters to offset process deviations and equipment operating condition drift.

[0123] Furthermore, the intelligent control method for multi-system integrated pharmaceutical production execution provided in this embodiment includes step S300 as follows: Step S310: Obtain the data field set of multiple heterogeneous systems. The data field set of multiple heterogeneous systems is obtained by parsing the production parameter adjustment instructions.

[0124] The original data packets of production parameter adjustment instructions output by step S200 are disassembled. The native non-standard message fields of four business systems, namely MES, WMS, SCADA and LIMS, are split according to the system affiliation label. The header check bits and tail redundant check codes of the instruction message are removed. All differentiated native fields are collected to construct a structured set of data fields from multiple heterogeneous systems.

[0125] The formula for generating a set of data fields for multiple heterogeneous systems is: (11) In formula (11), This is a collection of heterogeneous raw data fields from multiple systems, without units, and is the core output of this step. It is a heterogeneous instruction parsing operator with no unit, used for structured parsing of production parameter adjustment instructions; For the production parameter adjustment instructions, the definition in formula (10) is used, and the explanation is not repeated here; These are fields in the MES system, corresponding to production planning and scheduling data; This is a field for SCADA systems, corresponding to real-time monitoring and control data of the equipment; This is a LIMS system field, corresponding to laboratory testing and material quality data. For ERP system fields, corresponding to material inventory and resource management data. The control logic of formula (11) is to parse the production parameter adjustment instructions through a heterogeneous instruction parsing operator, decompose them into the original data fields corresponding to the four major pharmaceutical production systems: MES, SCADA, LIMS, and ERP, forming a set of heterogeneous data fields of multiple systems, providing basic data for the generation of subsequent cross-system collaborative control events. Formula (11) realizes the structured decomposition of cross-system instructions through a unified parsing operator, solving the industry pain point of inconsistent data formats and inability to directly collaborate among MES, SCADA, LIMS, ERP and other systems in pharmaceutical production, providing standardized heterogeneous data input for subsequent semantic mapping and format conversion processing, and ensuring that the subsequently generated collaborative control events can be recognized and executed by various business systems.

[0126] Data field set of multiple heterogeneous systems: a collection of native heterogeneous fields that have been split and categorized by system, have no redundant check codes, and have incompatible semantic formats.

[0127] The total number of fields in a single batch is 42 to 56; the value is based on the standard field layout specification of the four-system linkage instruction in the pharmaceutical workshop.

[0128] Step S320: Obtain the semantic feature vector of heterogeneous fields based on the data field set of multiple heterogeneous systems, and determine the standard semantic mapping entity based on the semantic feature vector of heterogeneous fields.

[0129] For non-standard heterogeneous fields, perform word vector encoding and semantic feature extraction to generate fixed-dimensional heterogeneous field semantic feature vectors; match the pharmaceutical company's preset cross-system semantic entity dictionary to complete vector clustering matching, bind synonymous fields, and generate normalized standard semantic mapping entities.

[0130] Heterogeneous field semantic mapping is accomplished using a field mapping table, which records the source system field name, source data type, source unit, target standard field name, target data type, target unit, unit conversion rules, and allowed value range. When the field mapping confidence is lower than the preset semantic vector matching confidence threshold, a field mapping task awaiting manual confirmation is generated, and control events are not directly generated.

[0131] The formulas for calculating the semantic feature vectors of heterogeneous fields and the standard semantic mapping entities are as follows: (12) (13) In formula (12)-(13), This is a semantic feature vector of heterogeneous fields, without units, with a dimension of 128, and is the core intermediate output of this step; This is a pre-trained semantic extraction model for the pharmaceutical manufacturing field. It is unitless and is a pre-trained model based on the BERT architecture and fine-tuned using pharmaceutical industry corpus. For any field in the set of data fields of multiple heterogeneous systems, including MES, SCADA, LIMS and ERP system fields, the definition in formula (11) is followed and will not be repeated here; This is a standard semantic mapping entity, without units, and is an entity in the cross-system standard thesaurus of pharmaceutical companies. It is the core output of this step. This is a semantic mapping function with no unit, used to map semantic feature vectors of heterogeneous fields to a standard semantic knowledge base; This is a standard semantic knowledge base for the pharmaceutical production system, without units, and a GMP compliance terminology library containing standard terms and corresponding mapping relationships in the fields of production, quality, and materials. The control logic of formulas (12)-(13) is to first convert each field in the data field set of multiple heterogeneous systems into a 128-dimensional heterogeneous field semantic feature vector through a pre-trained semantic extraction model in the pharmaceutical production field; then, through a semantic mapping function, the semantic feature vector is mapped to the standard semantic knowledge base of the pharmaceutical production system to obtain a standard semantic mapping entity that matches the GMP compliance terminology library. Formulas (12)-(13) achieve semantic unification of heterogeneous fields across systems through the domain pre-trained model and the GMP compliance terminology library, solving the industry pain point of difficulty in collaborative control caused by inconsistent terminology and semantics between different systems in pharmaceutical production, and providing a semantic basis for the subsequent generation of standardized collaborative control events that can be recognized by various business systems.

[0132] Heterogeneous field semantic feature vector: a low-dimensional semantic numerical vector that represents the textual meaning, business attributes, and process affiliation of non-standard fields.

[0133] Standard semantic mapping entity: Standard business semantic entity label after normalization of synonymous heterogeneous fields.

[0134] Preset semantic vector matching confidence threshold: entity matching confidence ≥ 99.6%; value based on national standard requirements for field mapping accuracy in alignment step S300.

[0135] Step S330: Generate a cross-system unified format data payload based on the standard semantic mapping entity. The cross-system unified format data payload is obtained by performing data structure reorganization on data fields of multiple heterogeneous systems.

[0136] The heterogeneous field semantic mapping transformation operator is invoked, and combined with the GMP standard field mapping table, it completes four-level transformations on the multi-source heterogeneous original dataset: field name unification, data type alignment, legal unit conversion, and value range standardization. Then, the standardized fields are batch-filled into the standardized JSON event template through the standard template filling and recombination operator, and a cross-system unified format data payload with fixed hierarchy and key names is generated. This standardized transformation link is adapted to heterogeneous data sources across the entire production line, such as equipment acquisition, process monitoring, audit alarms, and 3D modeling, and outputs a unified JSON data carrier that conforms to the GMP interaction specification. This provides standardized input for subsequent feature parsing, hash verification, and encrypted auditing, and breaks down the barriers to data interoperability between multiple systems.

[0137] The formula for generating cross-system unified format data payloads is: (14) In formula (14), It provides a unified data payload format across systems, a standardized JSON structured data carrier, and is compatible with the interaction of multiple heterogeneous systems such as production equipment, process data acquisition, auditing, alarms, and modeling. It also provides a unified input format for subsequent process feature extraction, hash verification, and audit encryption, meeting GMP data exchange standardization specifications. For heterogeneous field semantic mapping transformation operators, input heterogeneous raw data Mapping table with standard The system completes four-layer standardization transformation: 1. Field name mapping: Replace custom field names in each system with GMP standard field identifiers; 2. Data type conversion: Force alignment of integer / floating-point / string / timestamp data types to a unified data type; 3. Physical unit conversion: Temperature, pressure, ratio, and duration are uniformly converted to GMP legal units of measurement; 4. Value range normalization and alignment: Standardize the value ranges of enumeration, Boolean, and process threshold. It is a multi-source heterogeneous raw dataset, including raw data from multiple systems with different structures, fields, and units, such as underlying equipment control messages, process measurement point acquisition data, alarm data packets, audit operation records, and original parameters for 3D modeling. It serves as a standard field mapping table, predefined as a global field lookup table for GMP compliance, storing the field names, types, units, value ranges, conversion coefficients, and mapping rules for the entire system, providing a standardized conversion benchmark for FieldMap operators. The standard template is populated with a recombination operator, which receives the standardized set of fields after transformation and populates each field into the standard JSON template. This generates a unified load with a fixed hierarchy, fixed key name, and fixed structure. As a standard JSON event template, the fixed-level JSON skeleton defined by the drug regulatory and industrial control interaction specification stipulates a fixed top-level child node structure, including operation identifier, process parameters, timestamp, measurement point index, verification fingerprint, alarm label, etc. Formula (14) first completes the standardization conversion of name, type, unit, and value range of multi-source heterogeneous raw data by combining field mapping operators with standard mapping tables, and then fills the compliant fields into the standard JSON template to complete the structural reorganization, outputting a cross-system unified format GMP compliant data payload, realizing the standardization of data exchange between heterogeneous equipment, processes, and audits across the entire production line.

[0138] Cross-system unified format data payload: a core business data message carrier with standardized semantics, structure, and encoding across the entire system.

[0139] Step S340: If the data payload integrity check value of the cross-system unified format data payload is greater than or equal to the preset data integrity threshold, and the number of missing mandatory fields is zero, then perform event encapsulation processing on the cross-system unified format data payload to obtain standardized collaborative control events adapted to each business system.

[0140] The number of valid GMP required fields in the current batch is compared with the total number of required fields for all processes. The ratio of these two values ​​yields a data payload integrity check value in the range of 0 to 1. A compliance determination is then performed based on a preset data integrity threshold and the number of missing required fields: if the data payload integrity check value is greater than the preset threshold and no required fields are missing, the AES symmetric encryption operator is used to encrypt the cross-system unified format data payload, generating a standardized collaborative control archiving event stored in the GMP audit trail database; if the data payload integrity check value does not reach the preset threshold or any required field is missing, a rejection instruction is output and a re-collection task is issued. Verification can only proceed again after the missing fields are filled. This step simultaneously implements the dual constraints of data coverage and the requirement that key fields cannot be missing under the GMP ALCOA+ specification. Incomplete data is blocked from circulation, and qualified data is encrypted and archived to prevent tampering. This overcomes the shortcomings of traditional single-dimensional verification, manual supplementation of incomplete data, and the regulatory compliance risks associated with plaintext storage of raw data. It provides a standardized and quantitative basis for GMP data pre-verification for multi-system integrated pharmaceutical production lines.

[0141] The formulas for calculating the data payload integrity check value and standardized collaborative control events are as follows: (15) (16) In formula (15)-(16), For production data collection, completeness, dimensionless proportion, and value range are considered. Used to determine whether the process, batch, and audit fields uploaded by MES / DCS meet the GMP data integrity specifications. This represents the number of valid collections of required compliant fields in the current collection batch (including GMP-mandated fields such as process parameters, operators, timestamps, equipment numbers, and feed quantities). The total number of all required fields for this process as specified in GMP regulations. A preset data integrity threshold is typically set to 0.94, which serves as the boundary for determining whether the data is acceptable or unacceptable. To enforce the count of missing fields that are required; This indicates that no key compliance fields are missing. 1. Determine branch 1 ( (Compliance approved): Condition: Data integrity meets the standard and no required fields are missing; The AES symmetric encryption operator is used to encrypt the complete data packet payload, generating a compliant upload event with an audit index, which is then stored in the GMP audit trail database. 2. Decision branch 2 ( (Data not qualified): Condition: Insufficient completeness or missing required fields; Output instruction: Reject the current batch of collected data, issue a re-collection instruction, fill in the missing fields, and re-validate. The final output events are divided into two categories: encrypted compliance archive events and re-collection rejection instructions, which are connected to the enterprise GMP data platform. The first formula (15)-(16) calculates the completeness of the collected data based on the ratio of the valid required fields to the total required fields in GMP; the second formula judges the data through the dual conditions of completeness threshold and key field missing count. When the data is compliant, the collected payload is encrypted to generate an archive audit event, and when the data is incomplete, a rejection re-collection instruction is output. The dual compliance constraints of GMP data coverage and key fields must not be missing are implemented simultaneously to block the flow of incomplete data into the process self-control link. Qualified data is encrypted and tamper-proof and retained, solving the compliance defects of traditional single-dimensional verification, manual supplementation of incomplete data, and plaintext storage of original data. It provides a standardized GMP data pre-verification quantitative basis for intelligent manufacturing of pharmaceutical multi-system integration.

[0142] Data payload integrity verification value: A quantitative score for the compliance and integrity of the data payload, ranging from 0 to 1.

[0143] Preset data integrity threshold: The threshold ranges from 0.922 to 0.965, with an optimal value of 0.94; the value is based on compliance with GMPALCO's mandatory data integrity compliance requirements.

[0144] The validation distinguishes between required fields (device ID, batch number, parameter limit, timestamp) and optional fields; if any required field is missing, the payload is deemed invalid, the encapsulated collaborative event is rejected, and the process is not based on the overall integrity percentage.

[0145] Preferably, the intelligent control method for multi-system integrated pharmaceutical production execution provided in this embodiment includes step S400 as follows: Step S410: Obtain the event's built-in logic execution sequence. The event's built-in logic execution sequence is obtained by parsing the control logic built into the standardized collaborative control event.

[0146] The underlying control logic of standardized collaborative control events is decompiled from instructions and deconstructed from logical statements. The process control timing logic, equipment interlocking logic, and material linkage logic are separated, sorted and arranged according to the execution sequence, and serialized and encapsulated to obtain the event built-in logic execution sequence that can be compared and judged one by one.

[0147] The underlying control logic of standardized collaborative control events is decompiled from instructions and deconstructed from logical statements. The process control timing logic, equipment interlocking logic, and material linkage logic are separated, sorted and arranged according to the execution sequence, and serialized and encapsulated to obtain the event built-in logic execution sequence that can be compared and judged one by one.

[0148] The formula for calculating the execution sequence of the event's built-in logic is: (17) In formula (17), This is the built-in logic execution sequence of the event, without units; it is a set of process linkage rules and the core output of this step. It is a logic parsing operator with no unit, used for structured parsing of the built-in control logic of standardized collaborative control events; For standardized collaborative control events, the definition in formula (16) is used, and the explanation is not repeated here. The control logic of formula (17) is to parse the standardized collaborative control events through a logic parsing operator, extract the built-in control logic of the events, and generate an event built-in logic execution sequence containing process linkage rules, so as to provide a logical basis for subsequent detection and judgment of production deviations. Formula (17) transforms standardized collaborative control events into a structured set of process linkage rules through a logic parsing operator, which solves the industry pain point of scattered control logic, difficulty in unified management and execution in traditional pharmaceutical production, and provides a standardized logical basis for subsequent deviation judgment based on logic rules.

[0149] The event's built-in logic execution sequence consists of an array of underlying event interlocks and control timing logic statements arranged according to execution time sequence.

[0150] Step S420: Compare the execution sequence of the built-in logic of the event with the preset production deviation rules to obtain the matching degree of the production deviation rules.

[0151] The embedded logic execution sequence of the current batch event, composed of process deviations, equipment compensations, and data acquisition event time-series codes, is extracted and compared with the feature vector of the production deviation rules in the GMP standard filed with the drug regulatory authority. This cosine similarity calculation quantifies the degree of overlap between the two sets of time-series operation logics. Then, a value range pruning operator is used to constrain the similarity result to... By eliminating interference from negatively correlated rules with no matching logic, a standardized production deviation rule matching degree is obtained. This matching degree fully restores the time sequence behavior of process exceedance, automatic equipment control, and data verification. It can accurately match the corresponding GMP classification and handling specifications, solving the compliance defects of traditional single-point parameter comparison, which cannot identify complete deviation behavior, is prone to irrelevant rule mismatch, and lacks unified classification and quantitative indicators. It provides a standardized and quantitative judgment basis for the fully automated deviation level determination and automatic issuance of compliance handling work orders in pharmaceutical production lines.

[0152] The formula for calculating the matching degree of production deviation rules is: (18) In formula (18), The dimensionless similarity score, representing the matching degree of GMP production deviation rules, is the core output of this formula; its value is fixed within a constrained range. The closer the value is to 1, the more closely the execution logic of the current production line event matches the deviation rules filed with the drug regulatory authority. The closer the value is to 0, the greater the difference between the deviation behavior and the standard rules. It is used to determine the deviation level and match the corresponding GMP handling process. The built-in logic execution sequence of the current batch event is a high-dimensional feature vector; it consists of the process deviation, equipment compensation, and data acquisition event timing codes output by formulas (9) to (11), which fully record the operation logic, process fluctuations, and data verification behavior of the entire production process. This is a feature sequence vector for production deviation rules in GMP filing standards; and a standard handling logic template for various process deviations approved by the drug regulatory authority, including compliance standard sequences such as exceedance thresholds, compensation intervals, manual review triggering conditions, and data re-collection rules. This is a vector cosine similarity operator that calculates the cosine of the spatial angle between the real-time execution sequence and the standard deviation rule vector; output range Positive numbers represent logical trends in the same direction, while negative numbers represent complete irrelevance and no matching relationship. The range pruning constraint operator is used; if the cosine similarity is less than 0 (negative correlation, no match), it is forced to be 0 to eliminate the matching interference caused by irrelevant logic; if the cosine similarity is greater than 1, it is forced to be 1; the final output is uniformly set to the range pruning constraint operator. The formula (18) takes the real-time production process event time sequence vector and the GMP filing standard deviation rule vector as input, and measures the degree of overlap between the two sets of operation logics through cosine similarity; then, the matching degree is locked within the range by the value range pruning operator. The interval is used to cut off negative correlation interference caused by unrelated logic and output the standard production deviation rule matching degree; it fully restores the full-time behavior of process deviation occurrence, equipment adjustment, and data verification, accurately matches the corresponding GMP deviation handling specifications, and solves the compliance defects of traditional single-point parameter comparison that cannot identify complete deviation behavior, mismatch of irrelevant rules, and lack of unified graded and quantitative indicators. It provides a standardized and quantitative judgment basis for fully automatic deviation grading and compliance work order issuance in pharmaceutical intelligent manufacturing production lines.

[0153] Production deviation rule matching degree: a quantitative score measuring the similarity between the real-time execution sequence and the GMP deviation standard rules.

[0154] Step S430: If the matching degree of the production deviation rule is greater than the preset rule matching threshold, then an initial early warning data packet is constructed based on the batch operation status corresponding to the current production batch.

[0155] The system performs logic similarity threshold judgment, and triggers production deviations when the matching degree exceeds the limit; it captures multi-dimensional operation data such as the current production batch number, workstation operating status, equipment operating condition, real-time material parameters, and deviation trigger timestamp, and collects and packages them to build an initial warning data package that is not hierarchical or encapsulated.

[0156] The formula for constructing the initial warning data packet is: (19) In formula (19), For the production deviation rule matching degree, the definition in formula (18) is used, and the explanation is not repeated here; This is the preset rule matching threshold, with no unit, and a value of 0.81, which is the red line for triggering GMP compliance warnings. This is a vector concatenation operator with no unit, used to concatenate a logical sequence with a state vector; The execution sequence of the built-in logic for the event follows the definition in formula (17), and will not be explained again here; This is a vector representing the current production status of the batch, without units, and includes information such as batch progress, equipment status, and material consumption. The initial warning data packet, without units, is the core output of this step and is used for subsequent abnormal warnings and audits. The control logic of formula (19) is to compare the production deviation rule matching degree with the preset rule matching threshold of 0.81: if the production deviation rule matching degree is greater than the preset rule matching threshold, the event built-in logic execution sequence is concatenated with the current batch production operation status vector to construct the initial warning data packet, which is used for subsequent abnormal warnings and audits; if the production deviation rule matching degree is not greater than the preset rule matching threshold, the warning process is not triggered. Formula (19) realizes the automated warning triggering of production deviations by comparing the rule matching degree with the preset process deviation triggering threshold. At the same time, it integrates the event logic and batch status information into a unified warning data packet, which solves the industry pain points of untimely deviation warnings, incomplete information, and inability to meet audit requirements in traditional pharmaceutical production, and provides complete data basis for subsequent abnormal handling and compliance audits.

[0157] Preset rule matching threshold: threshold 0.79~0.83, optimal fixed value 0.81; value basis: pharmaceutical company production deviation triggering classification enterprise internal control standard, matching the 2025 version of GMP deviation handling specifications.

[0158] Initial warning data packet: contains batch operating condition raw data and raw warning data packet without hierarchical labels.

[0159] Step S440: Perform field normalization, abnormal field marking, and signal encapsulation processing on the initial early warning data packet to generate a standardized production abnormality early warning signal.

[0160] The multi-source heterogeneous initial early warning data packets, which integrate the time-series characteristics of process deviation, equipment compensation, data integrity, and rule matching degree across the entire chain, are input into the global standardized normalization operator. This operator unifies field encoding and sorting, eliminates differences in various alarm dimensions, maps them to a unified numerical value range, filters invalid blank data, and aligns the timestamp format, outputting standardized production anomaly early warning signals. The unified alarm features eliminate differences in multi-source alarm formats and numerical ranges, providing standardized quantitative inputs that can be compared horizontally for anomaly classification, automatic work order push, and equipment safety interlock control.

[0161] The formula for generating production anomaly warning signals is: (20) In formula (20), The standardized production anomaly early warning signal is the core output of this formula and serves as a unified input feature for downstream alarm classification, work order push, and equipment interlock control. All types of alarms have comparable quantified amplitudes. The initial warning data packet is a multi-dimensional abnormal feature vector containing various original alarm data such as abnormal messages from the underlying production equipment, inconsistent hash comparison, excessive modeling deviation, and abnormal correction instructions. The multi-source alarms have problems with inconsistent dimensions, numerical ranges, and field formats. The fusion formula (9) process deviation, formula (10) equipment compensation amount, formula (11) data integrity, and formula (18) rule matching degree full-link time sequence features are the original carriers of model input. The standardization operator for the entire domain performs the following operations: unifying field encoding, standardizing field order, eliminating dimensional differences, and mapping all alarm values ​​to a unified value range. Filtering invalid blank alarms and aligning timestamp formats eliminates noise in the original data packet format. Formula (20) performs global standardization and normalization processing on multi-source heterogeneous original alarm data packets, outputting alarm feature signals with a unified value range and format; eliminating the dimensional and format barriers of alarms from different devices, audits, and modeling, and providing standardized alarm quantitative input for production equipment anomaly classification, automated work order push, and interlocking safety control.

[0162] Furthermore, the intelligent control method for multi-system integrated pharmaceutical production execution provided in this embodiment includes step S500: Step S510: Input the production anomaly warning signal into the pre-trained ensemble learning classification model and perform deep feature extraction to obtain a multi-dimensional anomaly deep feature vector.

[0163] The encrypted and compiled production anomaly early warning signal is input into the input layer of the Stacking ensemble learning classification model. Relying on the CNN base learner, the model completes the mining of multi-dimensional hidden features of the early warning surface label, time series deviation, and source working conditions. After dimensionality reduction and fusion, the model outputs a fixed-dimensional, standardized multi-dimensional anomaly deep feature vector.

[0164] The formula for extracting deep feature vectors of multidimensional anomalies is: (twenty one) In formula (21), This is a multidimensional anomaly deep feature vector, without units, with a dimension of 256, and is the core output of this step; It is a pre-trained ensemble learning classification model, without units, which is composed of multiple basic models (such as CNN feature extraction sub-model, LightGBM classification sub-model and logistic regression meta-learner) and is used to extract deep features of abnormal warning signals. For the production anomaly warning signal, the definition in formula (20) is used, and will not be repeated here. The control logic of formula (21) is to input the production anomaly warning signal into the pre-trained ensemble learning classification model, and output a 256-dimensional multidimensional anomaly deep feature vector through the deep feature extraction operation of the model, providing a feature basis for subsequent anomaly type identification and processing decisions. Formula (21) performs deep feature extraction on the warning signal through the ensemble learning model, which solves the industry pain point in traditional pharmaceutical production that can only perform simple rule warnings and cannot perform fine-grained classification and analysis of anomalies, and provides richer and more accurate feature information for subsequent anomaly processing and decision-making.

[0165] The multidimensional anomaly deep feature vector is a high-dimensional fusion hidden layer feature array of the model that integrates early warning conditions, deviation time series, and equipment status.

[0166] Step S520: Perform anomaly category classification processing based on the multidimensional anomaly deep feature vector to obtain the initial anomaly category label.

[0167] The deep fusion feature matrix, which integrates formation properties, structural curvature, process message deviations, standardized alarm signals, and equipment control commands, is input into a multi-branch CNN+LSTM parallel integrated deep neural network to output predicted probability distribution vectors for various formations, production conditions, and equipment anomalies. The classification number corresponding to the probability peak is retrieved through the maximum value index operator to generate initial anomaly category labels. This integrated approach enables intelligent discrimination of multiple targets, such as formation and reservoir zoning, production process drift, equipment command anomalies, and data tampering, providing a unified intelligent classification benchmark for anomaly alarm classification, reservoir favorable area evaluation, and production equipment safety interlocking.

[0168] The initial anomaly category label calculation formula is: (twenty two) In formula (22), The deep fusion feature matrix is ​​derived from the multidimensional anomaly deep feature vectors mentioned earlier. High-dimensional deep features obtained through multi-layer feature mapping and dimensionality upscaling integrate multiple types of information such as formation properties, structural curvature, message deviation, alarm signals, and equipment control commands. To integrate deep neural network operators, it is integrated by parallel connection of multi-branch CNN+LSTM sub-networks, inputting deep features and outputting prediction probability distribution vectors for various working conditions / strata types; The maximum value index extraction operator retrieves the classification index corresponding to the maximum probability from the network output probability vector; As the initial anomaly category label, the original classification number of the working condition, formation, and anomaly type corresponding to each three-dimensional grid / equipment measurement point is stored, which serves as the classification basis for downstream alarm classification, reservoir zoning, and equipment interlocking. Formula (22) inputs multi-dimensional formation and equipment depth features into an integrated neural network to obtain the predicted probability vector of each category, extracts the category index corresponding to the maximum probability through the Argmax operator, and outputs the original classification label matrix; it integrates formation zoning, equipment control command tampering, and intelligent identification of production working condition anomalies, providing a standardized intelligent classification basis for alarm classification, reservoir evaluation, and equipment interlocking management.

[0169] Step S530: Extract the batch-bound production operation status data corresponding to the current production batch, and determine whether the initial abnormal category label matches the preset high-risk abnormal category.

[0170] By comparing the initial anomaly category labels with the preset set of high-risk anomaly labels, if a label matches a high-risk category, the data extraction operator is invoked to extract key information on batch progress, equipment status, material consumption, and personnel operation from the current batch production operation status vector in batches, generating batch-bound production operation status data. This enables targeted full-volume data collection of high-risk anomalies, solving the problems of missing traceability information and delayed handling in traditional high-risk anomaly management, and providing a complete set of supporting datasets for closed-loop handling of high-risk anomalies and GMP full-process audit traceability.

[0171] The formula for extracting batch-bound production operation status data is: (twenty three) In formula (23), The initial anomaly category label follows the definition in formula (22), and will not be explained again here; This is a pre-defined set of high-risk abnormal labels for pharmaceutical production. It has no unit and includes abnormal categories that have a significant impact on product quality, personnel safety, or compliance (such as material property deviations or temperature control abnormalities). This is a data extraction operator with no unit, used to extract key operation information from batch status data. The current batch production operation status vector is defined using the same formula (19), and will not be explained again here; The batch-bound production operation status data is unitless and is the core output of this step, including batch progress, equipment operating parameters, material consumption, personnel operation records, and other information. The control logic of formula (23) is to determine whether the initial abnormal category label belongs to the preset high-risk abnormal label set; if so, the key operation information is extracted from the current batch production operation status vector through the data extraction operator to generate batch-bound production operation status data, providing complete data support for subsequent high-risk abnormality handling and auditing. Formula (23) realizes key monitoring and information collection of high-risk production abnormalities through high-risk abnormality category matching and batch status data extraction, solving the industry pain points of untimely handling of high-risk abnormalities, incomplete data, and inability to meet audit and traceability requirements in traditional pharmaceutical production, and providing complete data basis for subsequent abnormality handling and compliance auditing.

[0172] Batch-bound production operation status data: Bind to the same production batch, link workstation-personnel-equipment to the overall operation ledger data.

[0173] Step S540: Perform task scheduling priority sorting on the batch-bound production operation status data and initial abnormality category labels to determine the production task scheduling sequence used to guide subsequent production operations; when a preset high-risk abnormality category is matched, increase the scheduling priority of the corresponding task.

[0174] Batch-bound production operation status data and initial anomaly category labels are input into the priority sorting operator, and graded sorting is carried out in combination with the pharmaceutical production-specific priority weight function. The weight function assigns the highest scheduling weight to tasks that match preset high-risk anomalies, generating a production task scheduling sequence. This enables priority scheduling of high-risk anomaly handling tasks, solving the defects of vague priority handling of traditional production anomalies and delayed handling of high-risk issues. It meets the requirements of pharmaceutical companies for production loss prevention, safety risk avoidance and GMP graded handling control, and reduces the risk of batch quality accidents.

[0175] The formula for determining the production task scheduling sequence is: (twenty four) In formula (24), This is a production task scheduling sequence with no unit. It is the core output of this step and the final output of step S500. This is a priority sorting operator with no unit, used to sort production tasks according to their urgency and importance. To bind production operation status data to batches, the definition in formula (23) is used, and will not be repeated here; The initial anomaly category label follows the definition in formula (22), and will not be explained again here; The priority weight function for pharmaceutical production tasks is unitless. Its core logic is to optimize the weight of high-risk anomalies, which aligns with the shutdown and risk avoidance process rules of pharmaceutical companies, ensuring that high-risk tasks are handled first. The control logic of formula (24) is to bind the batch to the production operation status data and input the initial anomaly category label into the priority sorting operator, and combine it with the priority weight function for pharmaceutical production tasks to sort subsequent production tasks and generate a production task scheduling sequence to guide subsequent production operations, prioritizing the handling of tasks corresponding to high-risk anomalies, so as to achieve safe production and batch loss control. Formula (24) solves the industry pain points of unclear anomaly handling priorities and untimely handling of high-risk anomalies in traditional pharmaceutical production by binding high-risk pharmaceutical production anomalies to hierarchical scheduling, meets the requirements of safe production and batch loss control of pharmaceutical companies, effectively reduces production risks, and improves the safety and compliance of the production process.

[0176] Preferably, the intelligent control method for multi-system integrated pharmaceutical production execution provided in this embodiment includes step S600 as follows: Step S610: Parse the built-in task priorities of the production task scheduling sequence and perform task diversion processing to determine the hardware target execution node of the diverted task.

[0177] The decoding and scheduling sequence has a built-in four-level priority encoding label, which completes the task hierarchical splitting and business type diversion matching according to the priority. It binds to the workshop PLC equipment location ledger, and matches the independent tasks after diversion to the corresponding reactor, conveyor unit, and quality inspection sensor hardware location, locking the hardware target execution node of the diverted task.

[0178] The formula for determining the hardware target execution node of the offloading task is: (25) In formula (25), This is the hardware target execution node for the diversion task. It has no unit and is the core output of this step, corresponding to underlying equipment such as reactors, feed pumps, and temperature control units. This is a task routing operator with no unit, used to allocate the production task scheduling sequence to the corresponding hardware execution node according to the task priority. The production task scheduling sequence follows the definition in formula (24), and will not be explained again here; The task priority mapping function is unitless and consistent with the weighting function in step S540. It is used to map task priorities to the corresponding hardware nodes. The control logic of formula (25) is to input the production task scheduling sequence and the task priority mapping function into the task distribution operator, and allocate different production tasks to the corresponding underlying hardware execution nodes according to the task priority, so as to realize the automated distribution of production tasks and ensure that high-risk or urgent tasks can be preferentially allocated to the corresponding equipment for execution. Formula (25) realizes the automated allocation of production tasks through the task distribution operator, solves the industry pain points of task allocation in traditional pharmaceutical production that rely on manual labor, are inefficient and prone to errors, ensures that high-risk or urgent tasks can be preferentially executed, and improves the response speed and safety of the production process.

[0179] Hardware target execution nodes for diversion tasks: Controlled points of underlying workshop automation hardware corresponding to each diversion production task.

[0180] Step S620: Combine the load status data of the hardware target execution node of the diversion task with the diversion task to perform Boolean logic operations and generate the initial action sequence of the device.

[0181] The system retrieves executable status data, operating conditions, start / stop status, and load status data of the target hardware nodes in real time. It integrates these data with the task execution requirements for traffic distribution and scheduling, and performs multi-dimensional Boolean logic AND-OR operations on device interlocking, start / stop mutual exclusion, and parallel coordination to initially generate a sequence of device actions with concurrent conflict flags. The executable status data includes device idle status, operating status, alarm status, maintenance status, cleaning status, and interlocking status.

[0182] The formula for generating the initial action sequence of the equipment is: (26) In formula (26), This is the initial action sequence of the device, without units, and is the core output of this step, containing an embedded concurrent conflict flag bit. This is a Boolean matrix representing the real-time load status of the device. It has no unit and is used to indicate the busy / idle status of each hardware node. In the Boolean matrix, 1 indicates that execution is allowed and 0 indicates that execution is prohibited. This is a Boolean logic AND operator for industrial control, without units, used to perform a logical AND operation between the task scheduling sequence and the equipment load status; The production task scheduling sequence follows the definition in formula (24), and will not be explained again here. The control logic of formula (26) is to input the real-time load state Boolean matrix of the equipment and the production task scheduling sequence into the Boolean logic AND operator, and generate the initial action sequence of the equipment through the logic AND operation. This sequence contains the action instructions of each equipment and embeds a concurrent conflict flag to avoid resource conflicts during task execution. Formula (26) combines task scheduling with equipment load state through Boolean logic operation, realizes the automatic generation of equipment action sequence and the pre-detection of concurrent conflicts, solves the industry pain points of equipment action relying on manual programming, easy resource conflict and execution error in traditional pharmaceutical production, and improves the stability and reliability of the production process.

[0183] The equipment load status data is real-time monitoring data of the underlying production hardware's computing power load, operating conditions, and equipment occupancy rate.

[0184] Step S630: If the initial action sequence of the device contains a concurrent conflict flag, the time slice round-robin algorithm is used to rearrange the initial action sequence of the device to obtain a conflict-free device action sequence.

[0185] This invention reads the initial action sequence of equipment with concurrent conflict flags, calculates adaptive rotation time slices by combining the fixed basic slice duration of GMP production lines with the batch cycle scaling factor, and then constrains the time slices to the range of 10ms to 50ms using a value range trimming operator to prevent communication jitter or task blocking caused by excessively fragmented slices. Using compliant adaptive time slices as the unit of division, the original concurrent conflict tasks are split and scheduled in a time-sharing manner through a time slice rotation rearrangement operator to complete the action sequence rearrangement, outputting a standardized equipment action sequence without hardware preemption or mutual exclusion timing conflicts. This invention can dynamically adapt to different pharmaceutical processes such as long sterilization cycles and short filling cycles, eliminate process timing deviations caused by equipment concurrent conflicts, and overcome the shortcomings of traditional fixed time slices, such as poor adaptability, lack of upper and lower limit control, and long-term blocking of mutual exclusion tasks. It provides a standardized and quantitative scheduling basis for stable timing scheduling of underlying equipment in pharmaceutical continuous production lines and ensures GMP process timing compliance. The time slice rotation algorithm is only used for sorting and communication scheduling of the queue of control commands to be issued within the control terminal, and is not used to interrupt non-preemptible physical process actions such as reactions, sterilization, stirring, and conveying that have already started execution. For device tasks with an uninterruptible flag, they are executed in a non-preemptive manner.

[0186] The formula for calculating the action sequence of conflict-free devices is: (27) In formula (27), For adaptive round-robin time slices (unit: ms), the base slice duration for round-robin scheduling; Set the basic slicing time for GMP production lines and the minimum atomic action execution time for equipment; This is a batch cycle scaling factor that is dynamically adjusted based on the total production time of a single batch and the number of concurrent tasks in the process, adapting to different processes such as long-cycle sterilization / short-cycle filling. For the value range pruning operator, the adaptive time slice is forced to be locked in the range of 10ms to 50ms: 1. Lower limit 10ms: to avoid frequent equipment switching and communication jitter caused by excessively fragmented slices; 2. Upper limit 50ms: to prevent excessively long slices from causing concurrent mutually exclusive tasks to be blocked for a long time, resulting in process timing deviations; This is the initial action sequence of devices with concurrent conflict flags, including multiple devices preempting hardware and mutual exclusion execution flags; For time-slice rotation rearrangement operators, Divide concurrent conflicting tasks into segments and schedule them in a time-sharing manner. To output a conflict-free, standardized equipment action sequence, free from hardware preemption and timing conflicts, it can be stably executed by the underlying DCS control system. Formula (27) first calculates the adaptive scheduling time slice based on the basic slice duration and batch cycle scaling factor, and locks the time slice within the compliant range of 10~50ms through the value range clipping operator; then, it uses the time slice round-robin algorithm to time-divide the original equipment action sequence with concurrent hardware preemption flag, and outputs a safe equipment action sequence with no hardware mutual exclusion and stable timing; it dynamically adapts to pharmaceutical processes of long and short cycles, eliminates process timing fluctuations caused by equipment concurrency conflicts, solves the compliance defects of traditional fixed slices such as poor adaptability, lack of boundary constraints, and concurrent task blocking, and provides a standardized and quantitative scheduling basis for the stable timing scheduling of underlying equipment and GMP process timing compliance in pharmaceutical continuous production lines.

[0187] Step S640: Based on the register address corresponding to the conflict-free device action sequence, perform instruction encoding to generate underlying production equipment control instructions for driving the underlying hardware operation.

[0188] It retrieves the underlying PLC controller, executes the physical address of the hardware-specific register, binds the execution logic of the conflict-free device action sequence, adapts to the hardware's native communication protocol and register addressing rules to complete the binary encoding of digital instructions, and compiles and generates underlying production equipment control instructions that can directly drive the underlying industrial control hardware.

[0189] The formula for generating control commands for the underlying production equipment is: (28) In formula (28), This is the control command for the underlying production equipment. It has no unit and is the core output of this step, as well as the final output of step S600. A PLC instruction encoding operator, without units, used to convert a sequence of device actions into instructions executable by the underlying hardware; For the sequence of actions of conflict-free devices, the definition in formula (27) is used, and the explanation is not repeated here; The physical address of the PLC's underlying hardware register is a unitless address and represents the register address mapping corresponding to each device action. The control logic of formula (28) is to input the conflict-free device action sequence and the physical address of the PLC's underlying hardware register into the instruction encoding operator, and generate control instructions that can directly drive the underlying hardware through encoding conversion, thereby realizing the automated execution of production tasks. Formula (28) realizes the direct conversion of device action sequence to underlying hardware instructions through the instruction encoding operator, solving the industry pain point of relying on manual writing of equipment instructions in traditional pharmaceutical production, which is inefficient and prone to errors. It realizes the fully automated control of the production process and improves production efficiency and equipment response speed.

[0190] Furthermore, the intelligent control method for multi-system integrated pharmaceutical production execution provided in this embodiment includes step S700 as follows: Step S710: Obtain the identity identifier of the control command of the underlying production equipment, and obtain the operation permission verification result based on the identity identifier.

[0191] The process involves extracting unique identifiers (IDs) from the underlying equipment control commands, binding them to operators, terminals, and timestamps. These IDs are then matched against a global compliance whitelist to perform hierarchical permission verification: If the ID belongs to the edit whitelist, permission level 2 is output, allowing high-risk operations such as process parameter adjustments and shutdowns; if the ID belongs only to the read-only whitelist, permission level 1 is output, allowing only data viewing and blocking all equipment control commands; if the ID is not entered into the global whitelist, permission level 0 is output, directly blocking illegal commands and generating an abnormal permission audit work order. This step relies on a hierarchical whitelist to achieve three levels of discrete permission quantification, implementing GMP hierarchical control from the source of equipment commands, preventing compliance risks of unauthorized modification of process parameters, and simultaneously retaining complete personnel operation traceability records. This provides a standardized and quantitative basis for the safe access of pharmaceutical production line equipment and drug regulatory authority audits.

[0192] The formula for calculating the operation permission verification result is: (29) In formula (29), The result of the operation permission verification (permission level) is the core output of this formula; the numerical level corresponds to three types of GMP compliant operation permissions, which are used to determine whether the current equipment control command is allowed to be issued and what operation to perform. It serves as a unique identifier for the instruction; it is bound to the operator's employee number, job role, operating terminal number, and operation timestamp, and is a core traceability field for GMP audit tracking. Each device instruction comes with an unalterable ID tag. This is a global set of whitelists for compliant operations, including read-only whitelists. Editing the whitelist Two subsets: All job IDs with system access qualifications are included in a unified manner, in accordance with the pharmaceutical company's personnel access level SOP specifications. The system allows for editing of the whitelist of operational permissions; it includes the IDs of high-privilege personnel such as process supervisors, QA managers, and equipment engineers, corresponding to permission level 2, who can issue instructions for adjusting process parameters and shutting down high-risk equipment such as production lines. This is a read-only whitelist; it includes IDs of low-privilege personnel such as inspectors and frontline operators, corresponding to permission level 1. They are only allowed to query process data and production reports, and are prohibited from issuing any equipment control commands. The three-level branch decision logic is as follows: 1. Level 2 indicates the ID exists in the edit whitelist, allowing modification operations such as parameter adjustments and shutdowns; 2. Level 1 indicates... Only a read-only whitelist exists, granting only data viewing permissions and blocking all device control commands; 3. Level 0 indicates... If an access is not on the global whitelist, it is considered an illegal access, and the instruction is directly intercepted and an abnormal permission audit work order is generated. The formula (29) takes the unique identity of the personnel and terminal bound to the equipment instruction as input, and matches the two-layer set of whitelist for editing operation and whitelist for reading only, and outputs three levels of discrete permission levels; high-permission IDs are allowed to adjust process parameters and shut down production lines, while ordinary positions are only allowed to query data. Illegal instructions are directly intercepted for identities not entered into the whitelist; the requirements for hierarchical operation control of GMP positions are accurately implemented, and the compliance risks of unauthorized modification of processes are blocked from the source of equipment instructions at the bottom level. At the same time, complete permission audit traceability records are retained, providing a standardized and quantitative judgment basis for safe access to equipment operation in pharmaceutical intelligent manufacturing production lines and permission verification of drug regulatory personnel.

[0193] Step S720: If the operation permission verification result is the edit operation permission level, then obtain the normalized messages of the underlying production equipment control instructions in the instruction generation stage and the instruction execution stage respectively, calculate the hash features of the normalized message generation stage and the hash features of the execution stage, and perform a consistency comparison between the hash features of the generation stage and the hash features of the execution stage.

[0194] When the permission verification determines that the operation is at the edit level, the original message and the real-time message executed by the device are captured separately. The message is standardized by using a message standardization operator to unify the encoding, field sorting, numerical precision and layout format, eliminating format noise interference. The SHA256 hash operator is used to generate the baseline hash feature of the generation stage and the real-time hash feature of the execution stage. By comparing the consistency of the two sets of 256-bit hash fingerprints, data anomalies such as transmission tampering and message packet loss are accurately identified, providing a reliable judgment benchmark for GMP full-process data integrity compliance verification.

[0195] The formula for calculating the real-time hash characteristics and data integrity deviation of the instruction is as follows: (30) (31) In formula (30)-(31), To generate hash features for the generation stage, a 256-bit hash fingerprint is generated after standardizing the original control message before it is sent out, which serves as the baseline credential for the original integrity of the message. The original complete string of the underlying equipment control command generation stage is the original complete equipment control message captured when the permission verification is at the edit level, which includes all process parameters, operation identifiers, and measurement point binding information; To generate hash features during the execution phase, the device side collects real-time hash fingerprints generated after standardizing the running messages, which are then cross-compared with the baseline fingerprint. To standardize the message string in real time during the instruction execution phase, and to capture the current message original text in real time when the transmission link reaches the underlying device; To standardize the message operators, we unify the lexicographical order of fields, remove redundant spaces / newline characters, fix the UTF-8 encoding, and unify the decimal precision of floating-point numbers; we only retain the differences in business values ​​and eliminate misjudgments of hash inconsistencies caused by formatting noise. As a standard SHA256 one-way hash operator, it has the hash avalanche effect: if any 1 byte of the message content is modified, the output 256-bit fingerprint will be completely changed; it is irreversible and is only used for integrity comparison, and does not support reverse restoration of the message. Formula (30) generates a baseline hash fingerprint after standardizing the original message generated by the device instruction; Formula (31) generates a real-time hash fingerprint after standardizing the real-time execution message of the device; the two sets of fingerprints are used for consistency comparison, quantify the risk of instruction transmission tampering and message loss, and provide an integrity judgment basis for GMP compliance grade verification.

[0196] Step S730: If the hash feature of the generation stage is consistent with the hash feature of the execution stage, the integrity verification result is passed; if the hash feature of the generation stage is inconsistent with the hash feature of the execution stage, the automatic execution of the control instructions of the underlying production equipment is blocked and a manual review work order is generated; if at the same time, it is detected that the key control field exceeds the preset process range or the identity identifier is abnormal, the integrity verification result is determined to be rejected.

[0197] The three-layer GMP verification constraint is used to carry out three-state hierarchical judgment: when the hash fingerprint is consistent, all keyword segments are within the process threshold range, and the operator identity is within the legal whitelist, the verification is passed and the automatic execution of the equipment is allowed; when any key field exceeds the threshold or the operator identity is not on the whitelist, the control command is directly rejected and discarded; when the hash fingerprint is inconsistent but the fields and identity are compliant, the automatic execution process is blocked and a GMP manual review work order is pushed; the layered verification of message integrity, process compliance, and operation permission is completed to meet the GMP hierarchical safety control requirements of pharmaceutical production.

[0198] The formula for determining the GMP compliance verification result is as follows: (32) In formula (32), To determine the hash characteristics of the generated message during the generation phase, the generated message is standardized and then hashed. Hash characteristics for the execution phase; For key process control fields in GMP production (temperature, pressure, ratio, duration, and other compliance control parameters); Preset upper and lower limit thresholds for processes according to GMP specifications; A unique identifier for the operator / equipment operator; This is a system whitelist, meaning only identities within the set have device operation permissions. It is a universal quantifier, meaning that all objects in the set satisfy the condition; For an existential quantifier, at least one object in the set does not satisfy the condition; This is a logical AND operation, where all conditions are true simultaneously. Logical OR: A branch is triggered as soon as any one of the conditions is met; For GMP compliance verification results, three-state output: This indicates that all verifications have passed, and the device is allowed to automatically execute control commands. This indicates that the process has exceeded the limits or the identity is invalid; the instruction will be discarded and execution prohibited. This indicates hash inconsistency (message tampering / transmission packet loss), blocking automatic execution and forcibly generating a manual GMP review work order. Formula (32) performs a logical joint judgment on three types of constraints: message SHA256 hash consistency, compliance range of all process fields, and legality of operation identity, and outputs three-level verification results: Pass / Reject / Block. These correspond to three types of actions: automatic release, direct discard of instructions, blocking automatic execution and pushing a manual GMP review work order, respectively, providing a basis for layered integrity and compliance security verification of the control instructions issued by the underlying production equipment in accordance with GMP specifications.

[0199] Step S740: If the GMP compliance verification result is passed, execute the underlying production equipment control instructions and generate a GMP compliance audit trail record based on the status change data generated during the execution process.

[0200] GMP compliance verification is determined by issuing and executing control commands to underlying equipment, collecting the corresponding plaintext control messages, nanosecond-level high-precision timestamps, and a complete set of measurement point indexes, and uniformly sending them to the AES-GCM encryption operator to generate encrypted audit trail records with integrity authentication tags. The encrypted text has a built-in tamper verification mark and is permanently archived and stored, realizing the binding and traceability of all elements of production operation sequence, working condition measurement points, and control commands, meeting the mandatory compliance requirements of GMP audit traceability, anti-tampering, and full-chain traceability.

[0201] The formula for generating GMP compliance audit trail records is: (33) In formula (33), The structured audit logs, encrypted with AES-GCM authentication, are permanently archived and stored to ensure GMP compliance with encrypted audit trails. This meets the mandatory requirements of GMP drug regulatory audits for full-chain traceability, operation tracking, and tamper-proofing. Each record is bound to the corresponding device instruction verification result. It is an AES-GCM Galois / counter mode encryption operator, a compliant standard encryption algorithm for the industrial and pharmaceutical industries, integrating encryption and integrity authentication; the output ciphertext includes a message authentication tag (MAC), and any tampering, deletion, or modification of any field will cause decryption verification to fail. It also supports parallel high-speed batch processing and is suitable for real-time concurrent auditing of production lines. For the control instructions of the underlying production equipment, the definition in formula (28) is used. It provides high-precision nanosecond-level timestamps for instruction execution, accurately locking the timing of instruction issuance, equipment startup, and status changes, preventing manual tampering with operation time, and enabling GMP verification to restore the complete operation timeline. The set of measurement points for the entire process is associated with the unique index of all collected measurement points and strata / production line points during the execution of this instruction, and the status change data generated synchronously is bound to achieve bidirectional traceability of "operation instructions - real-time operating data". After GMP compliance verification and release, the formula (33) inputs the plaintext control instructions of the underlying equipment, the nanosecond-level execution timestamp, and the set of measurement points for the entire process into the AES-GCM certified encryption operator to generate encrypted audit trace records; it integrates the confidential storage and integrity anti-tampering verification of operation data, and provides standardized compliance archiving basis for GMP drug supervision audit and full-chain traceability of production operations.

[0202] Please see Figure 2This embodiment provides a multi-system integrated intelligent management and control system for pharmaceutical production execution, used to implement the aforementioned multi-system integrated intelligent management and control method for pharmaceutical production execution. It includes a material attribute feature set acquisition module 10, a production parameter adjustment instruction acquisition module 20, a standardized collaborative control event acquisition module 30, a production anomaly early warning signal generation module 40, a production task scheduling sequence acquisition module 50, a low-level production equipment control instruction acquisition module 60, and a GMP compliance audit trace record generation module 70. The material attribute feature set acquisition module 10 is used to collect material attribute data from multiple different acquisition nodes throughout the entire pharmaceutical production process, constructing and acquiring a material attribute feature set. The material attribute feature set contains multiple sets of material attribute time-series features, each of which is configured with a collection timestamp and collection device identification information. The production parameter adjustment instruction acquisition module 20 calls a pre-trained time-series feature extraction model to perform time-series dependency analysis and deep feature extraction on all material attribute time-series features within the material attribute feature set, uncovering the correlation between material attribute time-series changes and pharmaceutical production parameters, and outputting the production parameter adjustment instruction corresponding to the current production batch. The standardized collaborative control event acquisition module 30 performs semantic mapping matching and unified format conversion processing on the various heterogeneous system data fields contained in the production parameter adjustment instructions. The system eliminates data heterogeneity differences among multiple business systems and generates standardized collaborative control events adapted to various business systems in pharmaceutical production. A production anomaly warning signal generation module 40 parses the production control logic built into the standardized collaborative control events, compares the operational parameters of the standardized collaborative control events with preset production deviation rules, and immediately generates a production anomaly warning signal corresponding to the current production batch if a preset production deviation rule is triggered. A production task scheduling sequence acquisition module 50 inputs the production anomaly warning signal into a pre-trained ensemble learning classification model, which extracts deep features of production anomalies and accurately classifies anomaly categories, outputting the result to guide subsequent operations. The pharmaceutical manufacturing operation includes a production task scheduling sequence; a low-level production equipment control instruction acquisition module 60, which performs intelligent diversion and logical operation processing on various production tasks according to the task priority sorting rules built into the production task scheduling sequence, and generates low-level production equipment control instructions that can directly drive the operation of low-level hardware devices; and a GMP compliance audit trace record generation module 70, which performs dual verification of operation permission and data integrity compliance on the low-level production equipment control instructions in sequence. If the verification of the low-level production equipment control instructions is passed, the low-level production equipment control instructions are executed, and compliance audit trace records adapted to GMP standards are generated simultaneously to complete the intelligent closed-loop management and control of pharmaceutical manufacturing.

[0203] Although preferred embodiments of the invention have been described, those skilled in the art, upon learning the basic inventive concept, can make other changes and modifications to these embodiments. Therefore, the appended claims are intended to be interpreted as including both the preferred embodiments and all changes and modifications falling within the scope of the invention. Clearly, those skilled in the art can make various alterations and modifications to the invention without departing from its spirit and scope. Thus, if these modifications and modifications of the invention fall within the scope of the claims and their equivalents, the invention is also intended to include these modifications and modifications.

Claims

1. A multi-system integrated intelligent control method for pharmaceutical production execution, applied to a control terminal, characterized in that, Includes the following steps: S100. Collect material attribute data from multiple different collection nodes in the entire pharmaceutical production process, construct and obtain a set of material attribute features, wherein the set of material attribute features contains multiple sets of material attribute time-series features, and each of the material attribute time-series features is configured with a collection timestamp and collection device identification information. S200. Call the pre-trained time-series feature extraction model to perform time-series dependency analysis and deep feature extraction on all material attribute time-series features in the material attribute feature set, explore the correlation between material attribute time-series changes and pharmaceutical production parameters, and output the production parameter adjustment instruction corresponding to the current production batch. S300. For the multiple heterogeneous system data fields contained in the production parameter adjustment instruction, perform semantic mapping matching and unified format conversion processing to eliminate the data heterogeneity differences of multiple business systems and generate standardized collaborative control events that are compatible with various business systems in pharmaceutical production. S400. Analyze the production control logic built into the standardized collaborative control event, compare and verify the operating parameters of the standardized collaborative control event with the preset production deviation rules, and if the preset production deviation rules are triggered, immediately generate a production anomaly warning signal corresponding to the current production batch. S500: Input the production anomaly warning signal into the pre-trained ensemble learning classification model, and use the ensemble learning classification model to complete the deep feature extraction of production anomalies and the accurate classification of anomaly categories, and output a production task scheduling sequence to guide subsequent pharmaceutical production operations. S600. Based on the task priority sorting rules built into the production task scheduling sequence, perform intelligent diversion and logical operation processing on various types of production tasks to generate underlying production equipment control instructions that can directly drive the operation of underlying hardware devices. S700: Perform dual verification of operation permission and data integrity compliance on the underlying production equipment control command in sequence. If the underlying production equipment control command is verified, execute the underlying production equipment control command and simultaneously generate a compliance audit trail record adapted to GMP standards to complete the intelligent closed-loop management and control of pharmaceutical production.

2. The intelligent control method for multi-system integrated pharmaceutical production execution according to claim 1, characterized in that, Step S100 includes: S110. Extract the original material data to obtain the initial material attribute time series; S120. Align the initial material attribute time sequence to obtain material attribute time features; S130. Map the time-series characteristics of the material attributes to determine the device-bound node attribute matrix; S140. Aggregate the attribute matrix of the device-bound nodes to obtain a set of material attribute features.

3. The intelligent control method for multi-system integrated pharmaceutical production execution according to claim 2, characterized in that, Step S200 includes: S210. Input the set of material attribute features into the pre-trained time-series feature extraction model to obtain the material time-series feature attention weight vector; S220. Based on the material time-series feature attention weight vector, perform calculations on the material attribute feature set to determine the production process target feature vector sequence. S230. Decode the production process target feature vector sequence to obtain the production process deviation value of a single process node. S240. If the production process deviation value of the single process node is less than or equal to the preset process deviation trigger threshold, then generate the equipment status compensation amount, and generate the production parameter adjustment instruction corresponding to the current production batch based on the equipment status compensation amount; if the production process deviation value of the single process node is greater than the preset process deviation trigger threshold, then generate the instruction blocking identifier and the manual review work order.

4. The intelligent control method for multi-system integrated pharmaceutical production execution according to claim 1, characterized in that, Step S300 includes: S310. Obtain a set of data fields for multiple heterogeneous systems, wherein the set of data fields for multiple heterogeneous systems is obtained by parsing production parameter adjustment instructions; S320. Obtain the heterogeneous field semantic feature vector based on the set of data fields of the multi-type heterogeneous systems, and determine the standard semantic mapping entity based on the heterogeneous field semantic feature vector; S330. Generate a cross-system unified format data payload based on the standard semantic mapping entity. The cross-system unified format data payload is obtained by performing data structure reorganization on the data fields of the multiple heterogeneous systems. S340. If the data payload integrity check value of the cross-system unified format data payload is greater than or equal to the preset data integrity threshold, and the number of missing mandatory fields is zero, then the cross-system unified format data payload is subjected to event encapsulation processing to obtain standardized collaborative control events adapted to each business system.

5. The intelligent control method for multi-system integrated pharmaceutical production execution according to claim 4, characterized in that, Step S400 includes: S410. Obtain the event's built-in logic execution sequence, which is obtained by parsing the control logic built into the standardized collaborative control event. S420. Compare the execution sequence of the event's built-in logic with the preset production deviation rules to obtain the production deviation rule matching degree; S430. If the matching degree of the production deviation rule is greater than the preset rule matching threshold, then an initial early warning data packet is constructed based on the batch operation status corresponding to the current production batch. S440. Perform field normalization, abnormal field marking, and signal encapsulation processing on the initial early warning data packet to generate a standardized production abnormality early warning signal.

6. The intelligent control method for multi-system integrated pharmaceutical production execution according to claim 1, characterized in that, Step S500 includes: S510. Input the production anomaly warning signal into the pre-trained integrated learning classification model and perform deep feature extraction to obtain a multi-dimensional anomaly deep feature vector. S520. Based on the multidimensional anomaly deep feature vector, perform anomaly category classification processing to obtain initial anomaly category labels; S530. Extract the batch-bound production operation status data corresponding to the current production batch, and determine whether the initial abnormal category label matches the preset high-risk abnormal category. S540. Perform task scheduling priority sorting on the batch-bound production operation status data and the initial abnormality category label to determine the production task scheduling sequence used to guide subsequent production operations; when a preset high-risk abnormality category is matched, increase the scheduling priority of the corresponding task.

7. The intelligent control method for multi-system integrated pharmaceutical production execution according to claim 6, characterized in that, Step S600 includes: S610: Parse the built-in task priorities in the production task scheduling sequence, perform task diversion processing, and determine the hardware target execution node corresponding to the diverted task. S620. Combine the load status data of the hardware target execution node of the diversion task with the diversion task to perform Boolean logic operations and generate the device initial action sequence; S630. If the initial action sequence of the device contains a concurrent conflict identifier, the initial action sequence of the device is rearranged using a time-slice round-robin algorithm to obtain a conflict-free device action sequence. S640. Based on the register address corresponding to the conflict-free device action sequence, perform instruction encoding to generate underlying production equipment control instructions for driving the underlying hardware operation.

8. The intelligent control method for multi-system integrated pharmaceutical production execution according to claim 1, characterized in that, Step S700 includes: S710. Obtain the identity identifier of the control command for the underlying production equipment, and obtain the operation permission verification result based on the identity identifier; The formula for calculating the operation permission verification result is: ； in, This is the result of the operation permission verification. As a unique identifier for the command, This is a whitelist set for global compliance operations. To edit the whitelist of operation permissions, For read-only viewing permissions, level 2 indicates that the ID is in the edit whitelist, allowing parameter tuning and shutdown modification operations; level 1 indicates... Only a read-only whitelist exists, granting only data viewing permissions while blocking all device control commands; Level 0 indicates... If the user is not on the global whitelist, the access will be deemed unauthorized, the command will be directly intercepted, and an abnormal permission audit ticket will be generated. S720. If the operation permission verification result is the edit operation permission level, then obtain the normalized messages of the underlying production equipment control instructions in the instruction generation stage and the instruction execution stage respectively, calculate the generation stage hash feature and the execution stage hash feature of the normalized message, and perform a consistency comparison between the generation stage hash feature and the execution stage hash feature. The formula for calculating the real-time hash characteristics and data integrity deviation of the instruction is as follows: ； ； in, Hash characteristics for the generation stage; The full string of the original message is generated during the generation phase of the control command for the underlying device. Hash characteristics for the execution phase; To standardize message strings in real time during the instruction execution phase; To standardize message operators, we unify field sorting, filter extra spaces and line breaks, fix character encoding, and unify floating-point precision, thereby eliminating hash misjudgments caused by format differences. As a standard SHA256 one-way hash operator, even minor changes to the message content will generate completely different fingerprints, which are used for tamper-proof consistency verification. S730. If the hash feature of the generation stage is consistent with the hash feature of the execution stage, the integrity verification result is passed; if the hash feature of the generation stage is inconsistent with the hash feature of the execution stage, the automatic execution of the control instructions of the underlying production equipment is blocked and a manual review work order is generated; if at the same time, it is detected that the key control field exceeds the preset process range or the identity is abnormal, the integrity verification result is determined to be rejected. The formula for determining the GMP compliance verification result is as follows: ； in, To determine the hash characteristics of the generated message during the generation phase, the generated message is standardized and then hashed. Hash characteristics for the execution phase; For key process control fields in GMP production (temperature, pressure, ratio, duration, and other compliance control parameters); Preset upper and lower limit thresholds for processes according to GMP specifications; A unique identifier for the operator / equipment operator; This is a system whitelist, meaning only identities within the set have device operation permissions. It is a universal quantifier, meaning that all objects in the set satisfy the condition; For an existential quantifier, at least one object in the set does not satisfy the condition; This is a logical AND operation, where all conditions are true simultaneously. Logical OR: A branch is triggered as soon as any one of the conditions is met; For GMP compliance verification results, three-state output: This indicates that all verifications have passed, and the device is allowed to automatically execute control commands. This indicates that the process has exceeded the limits or the identity is invalid; the instruction will be discarded and execution prohibited. This indicates hash inconsistency (message tampering / transmission packet loss), blocks automatic execution, and forces the generation of a manual GMP review work order; S740. If the integrity verification result is passed, the underlying production equipment control command is executed, and a GMP compliance audit trail record is generated based on the status change data generated during the execution process.

9. The intelligent control method for multi-system integrated pharmaceutical production execution according to claim 8, characterized in that, In step S740, the formula for generating GMP compliance audit trail records is: ； in, For GMP compliance audit trail records, For AES-GCM with authentication encryption operators, These are the original control commands for the underlying devices. Provide a precise timestamp for instruction execution. This is a set of original measurement point sampling indexes for the entire process.

10. A multi-system integrated intelligent control system for pharmaceutical production execution, used to implement the multi-system integrated intelligent control method for pharmaceutical production execution as described in any one of claims 1 to 9, characterized in that, include: The material attribute feature set acquisition module is used to collect material attribute data from multiple different collection nodes in the entire pharmaceutical production process, construct and acquire the material attribute feature set, which contains multiple sets of material attribute time-series features, and each of the material attribute time-series features is configured with a collection timestamp and collection device identification information. The production parameter adjustment instruction acquisition module is used to call the pre-trained time-series feature extraction model to perform time-series dependency analysis and deep feature extraction on all material attribute time-series features in the material attribute feature set, to explore the correlation between material attribute time-series changes and pharmaceutical production parameters, and to output the production parameter adjustment instruction corresponding to the current production batch. The standardized collaborative control event acquisition module is used to perform semantic mapping matching and unified format conversion processing on the multiple heterogeneous system data fields contained in the production parameter adjustment instructions, eliminate the data heterogeneity differences of multiple business systems, and generate standardized collaborative control events that are adapted to various business systems in pharmaceutical production. The production anomaly warning signal generation module is used to parse the production control logic built into the standardized collaborative control event, compare and verify the operating parameters of the standardized collaborative control event with the preset production deviation rules, and if the preset production deviation rules are triggered, generate the production anomaly warning signal corresponding to the current production batch in real time. The production task scheduling sequence acquisition module is used to input the production anomaly warning signal into the pre-trained ensemble learning classification model, and through the ensemble learning classification model, complete the deep feature extraction of production anomalies and the accurate classification of anomaly categories, and output a production task scheduling sequence to guide subsequent pharmaceutical production operations. The underlying production equipment control instruction acquisition module is used to perform intelligent diversion and logical operation processing on various production tasks according to the task priority sorting rules built into the production task scheduling sequence, and generate underlying production equipment control instructions that can directly drive the operation of underlying hardware devices. The GMP compliance audit trail generation module is used to perform dual verification of operation permission and data integrity compliance on the control instructions of the underlying production equipment in sequence. If the control instructions of the underlying production equipment pass the verification, the control instructions of the underlying production equipment are executed, and a compliance audit trail record adapted to GMP standards is generated simultaneously to complete the intelligent closed-loop management and control of pharmaceutical production.

Images