A large language model driven industrial sensor physical model automatic construction and edge deployment method and system

Abstract

The application discloses a large language model driven industrial sensor object model automatic construction and edge deployment method, comprising the following steps: S1, preprocessing an industrial sensor protocol document to obtain to-be-analyzed content; S2, calling a large language model through a multi-format protocol document analysis engine, extracting structured information, and generating a unified intermediate representation; S3, generating target Internet of Things platform object model files and edge execution format files according to the unified intermediate representation through a multi-platform object model adaptive generator; S4, performing correctness verification on the two files through an object model correctness automatic verification engine; S5, issuing the edge execution format file that passes the verification to an edge gateway, and loading the file by the edge gateway without interrupting data collection services; and S6, parsing industrial sensor original data according to a direct mapping table in the edge execution format file by the edge gateway to obtain corresponding engineering quantity data.

Description

Technical Field

[0001] This invention relates to the fields of industrial internet, industrial IoT, artificial intelligence and edge computing technology, and in particular to a method and system for automatic construction and edge deployment of industrial sensor object models driven by a large language model. Background Technology

[0002] Industrial sites typically deploy a large number of different types of sensors and smart instruments, such as temperature and humidity sensors, pressure transmitters, flow meters, gas analyzers, vibration monitors, and power meters. These devices are manufactured by different companies and may use different communication protocols such as Modbus RTU / TCP, Profibus-DP, HART, CANopen, OPC-UA, MQTT, CoAP, LoRa, and NB-IoT. Differences in data frame formats, register address mappings, function codes, byte order, data types, scaling factors, engineering units, and alarm thresholds among different protocols and manufacturers complicate the industrial IoT device integration process.

[0003] Different communication protocols may also include DeviceNet, EtherNet / IP, PROFINET, Zigbee, WirelessHART, ISA100.11a, EtherCAT, POWERLINK, and AMQP. Taking the Modbus protocol as an example, temperature sensors from different manufacturers may use different register addresses, numerical encoding methods, and range scaling factors, creating a "last mile" problem for connecting industrial field devices.

[0004] An object model is a structured description of the attributes, services, and events of IoT devices. It is typically represented in JSON or other structured formats to enable data interoperability between devices and cloud platforms. Current object model construction methods primarily rely on engineers manually reading device technical manuals and configuring fields, or on semi-automatic parsing using regular expressions, fixed templates, and rule engines. Manual methods are time-consuming and prone to errors in register addresses, byte order, scaling factors, and engineering units. Fixed templates or rule engines have poor adaptability to unstructured documents, scanned documents, multi-page tables, multilingual manuals, and vendor-defined extended fields.

[0005] In the traditional manual approach, engineers typically need to read the original manufacturer's technical manuals one by one and fill in each field according to the JSON Schema specification of the target cloud platform. In the semi-automatic approach based on rule engines, each new device type often requires the development of specific adaptation rules, and protocol changes after device firmware upgrades may also lead to the redevelopment of rules.

[0006] Large language models possess strong capabilities in natural language understanding, structured information extraction, and document parsing, making them suitable for one-time semantic understanding and field extraction of protocol documents. However, real-time data parsing at the industrial edge typically requires low latency, determinism, and low resource consumption, making large language models unsuitable for direct participation in the edge real-time data parsing process. Therefore, how to leverage large language models to complete semantic parsing of cloud-based protocol documents while simultaneously converting the parsing results into deterministically executable rule files at the edge is a problem that needs to be addressed in the automated construction and deployment of industrial sensor object models.

[0007] Deploying large language models directly at the industrial edge still has problems such as high inference latency, large hardware resource consumption, and insufficient output stability. Therefore, it is more suitable for one-time or offline protocol document understanding and object model structure generation tasks.

[0008] The existing technology still has the following shortcomings: First, protocol document parsing lacks semantic understanding capabilities, making it difficult to handle unstructured or semi-structured protocol documents; Second, object model construction still relies on manual configuration or fixed rules, making it difficult to scale up and adapt to devices from different manufacturers; Third, object models for the same device are difficult to adapt to multiple cloud platform formats simultaneously; Fourth, the generated object model results lack a verification loop based on real sensor data; Fifth, there is a lack of effective connection between cloud-based intelligent parsing and edge-based deterministic real-time execution; Sixth, object model updates often require restarting the data acquisition service, affecting continuous operation in industrial settings. Summary of the Invention

[0009] This invention aims to provide a method and system for automatic construction and edge deployment of industrial sensor object models driven by a large language model. It performs semantic parsing of industrial sensor protocol documents through a cloud-based large language model, generates a platform-independent unified intermediate representation, and further generates the object model of the target IoT platform and an edge execution format executable at the edge. This enables the edge gateway to perform deterministic parsing of raw sensor data without calling the large language model, thereby improving the efficiency of object model construction, cross-platform adaptability, real-time edge parsing, and continuous object model updates.

[0010] The technical solution of the present invention is as follows: A method for automatic construction and edge deployment of industrial sensor object models driven by a large language model includes the following steps: S1 receives the industrial sensor protocol document and performs document preprocessing on the industrial sensor protocol document to obtain the content of the document to be parsed; S2 uses a multi-format protocol document parsing engine to call a large language model to extract structured information from the content of the document to be parsed, generating a unified intermediate representation that is independent of the IoT platform. S3 generates target IoT platform object model files and edge execution format files based on a unified intermediate representation through a multi-platform object model adaptive generator. The edge execution format file includes a direct mapping table for parsing raw data from industrial sensors at the edge. S4 verifies the correctness of the target IoT platform's object model file and edge execution format file through an automatic object model correctness verification engine; S5 sends the verified edge execution format file to the edge gateway, and the edge gateway loads the edge execution format file without interrupting the data acquisition service; S6, the edge gateway performs deterministic parsing of the raw data from industrial sensors based on the direct mapping table in the edge execution format file, and obtains the engineering quantity data corresponding to the industrial sensors.

[0011] Furthermore, S2 specifically includes the following steps: S21, the document content to be parsed is divided into blocks according to at least one of document format, chapter title, table boundary and semantic boundary, and source location information including page number information is established for the block-based document content; S22, input the segmented document content into the large language model, and extract at least one of the following: basic device information, communication parameters, data point information, register information, event information, and service command information of the industrial sensor; S23. The extraction results of the large language model are constrained and output according to the preset structured output format, and the field validation and confidence level labeling of the constrained output results are performed to obtain a unified intermediate representation.

[0012] Furthermore, the unified intermediate representation includes the device basic information field and the data point field; The basic device information field includes at least one of the following: manufacturer information, device model, protocol type, firmware version, and communication parameters; The data point field includes at least one of the following: data point identifier, data point name, data type, read / write permissions, engineering units, measurement range, scaling factor, offset, register address, function code, byte order, data length, alarm threshold, source location information, and confidence level identifier.

[0013] Furthermore, S3 specifically includes the following steps: S31, convert at least one of the data point field, event field and service field in the unified intermediate representation into the target IoT platform object model file; S32, extract the fields required for edge parsing from the unified intermediate representation, and generate an edge execution format file. The fields required for edge parsing include at least one of register address, function code, data type, byte order, data length, scaling factor, offset, engineering unit, and range. S33, precompile the field mapping relationship in the edge execution format file into a direct mapping table. The direct mapping table includes index entries and mapping values. The index entries include at least one of protocol type, device address, function code, and register address. The mapping values ​​include at least one of data point identifier, data type, byte order, data length, scaling factor, offset, engineering unit, and range.

[0014] Furthermore, S4 specifically includes the following steps: S41, Static structure verification is performed on the target IoT platform object model file and edge execution format file. Static structure verification includes at least one of the following: format legality verification, required field integrity verification, register address range verification, and data type and range matching verification. S42, parse the raw data acquired by the industrial sensor according to the edge execution format file to obtain the verification parsing value, and perform dynamic verification based on the verification parsing value; S43, input the target IoT platform object model file, edge execution format file and corresponding industrial sensor protocol document fragments into the large language model for semantic review, and obtain the semantic review result; S44. When the results of static structure verification, dynamic verification, or semantic audit do not meet the preset conditions, an error report is generated, and a re-parsing process or a manual review process is triggered.

[0015] Furthermore, dynamic verification includes at least one of value range verification, statistical distribution verification, and reference data point error verification; value range verification is used to determine whether the verification resolution value is within the range defined by the unified intermediate representation; statistical distribution verification is used to determine whether the verification resolution value corresponding to the original data collected by the multi-frame industrial sensor conforms to the preset fluctuation characteristics of the corresponding industrial sensor; reference data point error verification is used to determine whether the error between the verification resolution value and the manual reference value exceeds the preset error threshold.

[0016] Furthermore, S5 specifically includes the following steps: S51 assigns version identifiers to verified edge execution format files through the object model version management and distribution module, generates integrity verification information, and distributes it to the edge gateway through an encrypted channel; S52, the edge gateway uses the currently running edge execution format file as the currently active model, uses the newly received edge execution format file as the preloaded model, and switches the preloaded model to the currently active model during the data acquisition cycle interval; S53, when a preset number of parsing error frames occur consecutively after the switch, the edge gateway rolls back to the previous stable version of the edge execution format file.

[0017] Furthermore, S6 specifically includes the following steps: S61, the edge gateway decodes the protocol frames of the raw data from the industrial sensors to obtain at least one of the protocol type, device address, function code and register address; S62, the edge gateway queries the direct mapping table based on at least one of the protocol type, device address, function code and register address to obtain the corresponding data point parsing rule; S63, the edge gateway performs byte order conversion, data type conversion, scaling calculation and offset calculation according to the data point parsing rules to obtain the engineering quantity data; The edge execution format file contains a dedicated protocol parsing template corresponding to the industrial communication protocol. The dedicated protocol parsing template is used to perform protocol frame decoding, function code recognition, register address positioning, byte order conversion, and scaling calculation. The data point mapping relationship is determined by a direct mapping table.

[0018] Furthermore, the above method also includes a protocol change and update step, which includes: S71 receives upgraded industrial sensor protocol documents or differential protocol documents; S72, calls the large language model to identify the changes in the upgraded industrial sensor protocol document or the difference protocol document compared to the original industrial sensor protocol document; S73, Generate an update patch for the object model based on the changes; S74 distributes the object model update patch to the edge gateway that has deployed the corresponding industrial sensor model, so that the edge gateway updates the edge execution format file according to the object model update patch.

[0019] Furthermore, this invention provides a large language model-driven automatic construction and edge deployment system for industrial sensor object models, including a cloud processing platform and an edge gateway; The cloud processing platform includes a multi-format protocol document parsing engine, a unified intermediate representation layer for object models, a multi-platform object model adaptive generator, an automatic object model correctness verification engine, and an object model version management and distribution module. The multi-format protocol document parsing engine is used to call a large language model to extract structured information from industrial sensor protocol documents and generate a unified intermediate representation that is independent of the IoT platform. The unified intermediate representation layer for object models is used to store unified intermediate representations; The multi-platform object model adaptive generator is used to generate target IoT platform object model files and edge execution format files based on a unified intermediate representation; The automatic verification engine for object model correctness is used to verify the correctness of the object model files and edge execution format files of the target IoT platform; The object model version management and distribution module is used to distribute the verified edge execution format file to the edge gateway. It is also used to receive the protocol document or difference protocol document after the industrial sensor upgrade, identify the changes and generate object model update patches. The edge gateway includes an edge device model hot deployment and real-time execution engine, which is used to load edge execution format files without interrupting the data acquisition service and perform deterministic parsing of raw industrial sensor data based on a direct mapping table.

[0020] Compared with the prior art, the beneficial effects of the present invention are as follows: 1. Semantic parsing of unstructured or semi-structured industrial sensor protocol documents is performed using a large language model, reducing the workload of manual field-by-field reading and configuration of object models.

[0021] 2. By using a unified intermediate representation as a platform-independent intermediate layer, the parsing results of the same protocol document can be converted into different IoT platform object model formats and edge execution formats, thereby improving cross-platform adaptability.

[0022] 3. By precompiling the edge execution format into a direct mapping table, the edge performs deterministic parsing based on table lookup, byte order conversion, and scaling calculation, avoiding the involvement of large language models in the real-time data parsing path at the edge.

[0023] 4. By forming a verification closed loop through static structure verification, dynamic data verification, and semantic auditing of the large language model, problems such as missing fields, register errors, range mismatch, or abnormal parsing results can be detected before the object model is issued.

[0024] 5. By employing a double-buffered hot-loading and anomaly rollback mechanism, the object model update process can reduce the impact on the continuity of on-site data acquisition. Attached Figure Description

[0025] Figure 1 This is a schematic diagram of the overall system architecture provided for an embodiment of the present invention.

[0026] Figure 2 This is a schematic diagram of the method steps provided in an embodiment of the present invention. Detailed Implementation

[0027] The technical solution of the present invention will be further described below with reference to the accompanying drawings. It should be understood that the following embodiments are for illustrative purposes only and are not intended to limit the scope of protection of the present invention. Where there is no conflict, the technical features in the following embodiments can be combined with each other.

[0028] like Figure 1 and Figure 2As shown, this embodiment provides a method and system for automatic construction and edge deployment of industrial sensor object models driven by a large language model. This method can be implemented through the cooperation of a cloud processing platform and an edge gateway. The cloud processing platform is used to complete the reception, preprocessing, structured information extraction, unified intermediate representation generation, multi-platform object model generation, edge execution format generation, correctness verification, and version management and distribution of industrial sensor protocol documents. The cloud processing platform may include a multi-format protocol document parsing engine, a unified intermediate representation layer for object models, a multi-platform object model adaptive generator, an automatic object model correctness verification engine, and an object model version management and distribution module. The multi-format protocol document parsing engine uses a large language model to extract structured information from industrial sensor protocol documents, generating a unified intermediate representation independent of the IoT platform. The unified intermediate representation layer stores the unified intermediate representation. The multi-platform object model adaptive generator generates the target IoT platform object model file and edge execution format file based on the unified intermediate representation. The automatic object model correctness verification engine verifies the correctness of the target IoT platform object model file and edge execution format file. The object model version management and distribution module distributes the verified edge execution format file to the edge gateway and generates object model update patches based on upgraded or differing protocol documents. The edge gateway may include an edge-end object model hot deployment and real-time execution engine, which loads the edge execution format file without interrupting data acquisition services and performs deterministic parsing of raw industrial sensor data based on a direct mapping table.

[0029] The method in this embodiment includes the following steps: S1 receives the industrial sensor protocol document and performs document preprocessing on the industrial sensor protocol document to obtain the content of the document to be parsed.

[0030] Users upload industrial sensor protocol documents through a cloud processing platform. These documents can be in PDF, Excel, CSV, Word, HTML, XML, JSON, or plain text formats, and can also be manufacturer-provided technical manuals, register lists, communication protocol descriptions, node description files, or device description files. The multi-format protocol document parsing engine performs corresponding basic parsing operations based on the document format. For example, it performs text and table extraction on electronic PDFs, OCR recognition on scanned PDFs, extracts row and column structure from Excel or CSV files, extracts node structure from XML or JSON files, and extracts heading levels, table content, and body paragraphs from Word or HTML files.

[0031] The basic parsing operations described above are only used to obtain text, table, or node content that can be processed by large language models, and do not pre-set fixed parsing rules for specific vendor protocols or specific data point mapping relationships. The multi-format protocol document parsing engine further divides the document content into blocks according to chapter titles, table boundaries, continuous register address ranges, page number ranges, figure caption positions, or semantic themes, and records the source location information for each document block. Source location information may include page numbers, chapter names, table names, row and column positions, figure caption positions, or original file identifiers, used for subsequent field tracing, semantic review, and manual verification.

[0032] S2 uses a multi-format protocol document parsing engine to call a large language model to extract structured information from the content of the document to be parsed, generating a unified intermediate representation independent of the IoT platform.

[0033] The multi-format protocol document parsing engine inputs the segmented protocol document content into the large language model, and constrains the extraction scope and output format of the large language model by using prompt word templates for the industrial sensor protocol domain. The large language model extracts at least one of the following from the protocol document: basic device information, communication parameters, data point information, register information, event information, and service command information.

[0034] The basic device information may include the manufacturer name, device model, protocol type, and firmware version; communication parameters may include baud rate, data bits, stop bits, parity check method, device address, node identifier, and topic name or port number; data point information may include data point identifier, data point name, data type, read / write permissions, engineering units, measurement range, accuracy, scaling factor, offset, and alarm threshold; register information may include function code, register address, address length, byte order, data length, and register purpose; event information may include alarm code, error code meaning, status bit definition, and event output parameters; and service command information may include the meaning of writable registers, control command format, input parameters, and output parameters.

[0035] During the extraction process, the system uses a predefined schema to structurally constrain the output of the large language model, ensuring that the extracted results are output in the field format required by the unified intermediate representation. For cases where required fields are missing, field formats are invalid, register addresses exceed the protocol's allowed range, data types and ranges do not match, engineering units are not standardized, scaling factors are missing, or byte order is unclear, the system generates a verification prompt and triggers a re-extraction, self-verification, or manual review process.

[0036] Furthermore, the system attaches a confidence level identifier to each extracted field. In this embodiment, the first preset threshold can be 0.8, and the second preset threshold can be 0.6. For fields with a confidence level lower than the first preset threshold, the system re-inputs the corresponding original document fragment, field context, and extracted results into the large language model for cross-validation. For fields whose confidence level remains lower than the second preset threshold after self-validation or cannot be automatically extracted, the system marks the field as awaiting manual review and generates a manual completion list containing missing fields, source locations, and recommended completion methods. Users can complete fields through a web interface or API interface, and the completed fields are then re-entered into the unified intermediate representation.

[0037] S3, through a multi-platform object model adaptive generator, generates the target IoT platform object model file and the edge execution format file based on a unified intermediate representation. The edge execution format file includes a direct mapping table for parsing raw industrial sensor data at the edge. This direct mapping table is pre-compiled from the field mapping relationships in the edge execution format file and is used by the edge gateway to perform low-latency, deterministic parsing of the raw industrial sensor data.

[0038] In this embodiment, the unified intermediate representation is an intermediate data structure independent of the IoT platform. It is used to receive the structured extraction results of the large language model and serves as an intermediate hub for subsequent multi-platform object model generation and edge execution format generation. The unified intermediate representation may include a version number, basic device information, an attribute list, a service list, an event list, and an array of extended fields.

[0039] Basic device information may include manufacturer name, device model, protocol type, firmware version, and communication parameters. Each attribute in the attribute list may include attribute identifier, attribute name, description, data type, access permissions, engineering units, measurement range, scaling factor, offset, protocol address information, alarm threshold, source location information, and confidence level. Protocol address information may include device address, function code, register address, byte order, and data length. The service list may include service identifier, service name, input parameters, output parameters, and command writing rules. The event list may include event identifier, event type, trigger condition, output data, and alarm level. An extended field array is used to store manufacturer-defined fields or specific protocol fields.

[0040] The multi-platform object model adaptive generator reads the unified intermediate representation and performs format conversion according to the field specifications of the target IoT platform. When the target platform is Alibaba Cloud Link IoT platform, it can map the attributes, services, and events in the unified intermediate representation to the properties, services, and events fields in TSL format, respectively; when the target platform is Huawei Cloud IoT platform, it can generate a device profile format file; when the target platform is Tencent Cloud IoT platform, it can generate a data template format file; and when the target platform is a custom gateway platform, it can generate the corresponding format through a programmable adapter.

[0041] Simultaneously, the multi-platform object model adaptive generator also generates edge execution format files. These edge execution format files differ from cloud-based object model files; they remove some metadata required for cloud display, management, or interaction, retaining only the fields needed for real-time parsing at the edge, such as protocol type, device address, function code, register address, data type, byte order, data length, scaling factor, offset, engineering units, and range. These fields are pre-compiled into a direct mapping table.

[0042] The direct mapping table can include index entries and mapping values. Index entries can include at least one of the following: protocol type, device address, function code, register address, node identifier, and topic name. Mapping values ​​can include at least one of the following: data point identifier, data type, byte order, data length, scaling factor, offset, engineering units, measurement range, read / write permissions, and alarm threshold. During real-time resolution, the edge gateway queries the direct mapping table based on the index entries to obtain the corresponding data point resolution rules.

[0043] The edge execution format file can also be configured with dedicated protocol parsing templates corresponding to industrial communication protocols. These dedicated protocol parsing templates are used at the edge to perform protocol frame decoding, function code recognition, register address location, CRC check, byte order conversion, data type conversion, and scaling calculations. These templates are used for real-time data execution at the edge and are instantiated into specific device parsing rules by populating a direct mapping table; they are not used to parse protocol documents such as PDFs, Word documents, and Excel files, nor do they participate in the semantic extraction process of large language models.

[0044] S4 uses an automatic object model correctness verification engine to verify the correctness of the target IoT platform's object model file and edge execution format file.

[0045] The automatic verification engine for object model correctness performs at least one of the following on the generated target IoT platform object model file and edge execution format file: static structure verification, dynamic data verification, and semantic auditing.

[0046] Static structure verification includes format validity verification, required field integrity verification, field value range validity verification, register address range verification, data type and range matching verification, and cross-field consistency verification. Cross-field consistency verification can be used to determine whether the actual range after the scaling factor is applied is consistent with the range defined in the unified intermediate representation, or to determine whether the identifier, unit, type, and access permissions of the same data point are consistent in the cloud object model file and the edge execution format file.

[0047] Dynamic data verification involves parsing the raw industrial sensor data provided by the user according to the parsing rules in the edge execution format file to obtain verification parsing values. These verification parsing values ​​are then subjected to at least one of the following verification methods: value range verification, statistical distribution verification, and reference data point error verification. Value range verification determines whether the verification parsing value falls within the range defined by a unified intermediate representation. Statistical distribution verification determines whether the verification parsing values ​​corresponding to multiple frames of raw industrial sensor data conform to the preset fluctuation characteristics of the corresponding industrial sensor. Reference data point error verification determines whether the error between the verification parsing value and the manual reference value exceeds a preset error threshold.

[0048] Semantic auditing involves inputting the target IoT platform object model file, edge execution format file, and corresponding industrial sensor protocol document fragments into a large language model. The large language model then checks in audit mode whether the meaning of fields, data types, engineering units, measurement ranges, alarm thresholds, register addresses, and read / write permissions are consistent with the semantics of the protocol document.

[0049] When the results of static structure verification, dynamic data verification, or semantic auditing fail to meet preset conditions, the automatic object model correctness verification engine generates an error report. The error report may include error fields, error type, source location information, error cause, and recommended handling methods. The system can trigger a re-parsing process for the large language model based on the error report, or submit the relevant fields for manual review. When all the aforementioned verifications pass, the system marks the corresponding target IoT platform object model file and edge execution format file as verified, allowing them to enter the version management and distribution process.

[0050] In this embodiment, when all verifications pass, the system automatically issues the object model and allows it to be deployed to the edge. When the static structure verification passes but the dynamic data verification has doubts, the system generates a manual review request. When there are field errors, the system generates an error report and triggers the large language model to be re-parsed. When multiple re-parses still fail, the system generates a detailed list of manual interventions and requests engineers to intervene.

[0051] S5 sends the verified edge execution format file to the edge gateway, and the edge gateway loads the edge execution format file without interrupting the data acquisition service.

[0052] The object model version management and distribution module assigns version identifiers to verified edge execution format files and records the version generation time, device model, applicable protocol, change summary, and scope of application. Version identifiers can include major version number, minor version number, and build number. The system maintains a complete version chain from initial generation to all updates and supports querying historical object models by device model, protocol type, edge gateway number, or version number.

[0053] The object model version management and distribution module can distribute edge execution format files to target edge gateways via an encrypted channel. Before distribution, the system can generate integrity verification information for the edge execution format file and authenticate the target edge gateway. After receiving the edge execution format file, the edge gateway performs integrity verification and source authentication. Once the verification is successful, it enters the preloading process.

[0054] In this embodiment, the object model version management and distribution module transmits the verified edge execution format file to the target edge gateway through an encrypted channel. During transmission, DTLS or TLS protocols are used for encryption, and SHA-256 hash verification is used to ensure file integrity. Before distribution, the system performs identity authentication and object model signature verification on the edge gateway to ensure the object model's source is trustworthy. Distribution strategies can include single-device targeted distribution and batch distribution. Batch distribution can be performed by grouping devices or dividing regions into batches. After distribution is completed, the system waits for the edge gateway to return a loading confirmation. If no confirmation is received within a timeout period, a limited number of retransmissions are performed.

[0055] The edge device model hot deployment and real-time execution engine employ a dual-buffering approach to maintain the currently active model and the pre-loaded model. The older version of the edge execution format file continues to serve as the currently active model for real-time data acquisition, while the newly received edge execution format file is loaded into the pre-load buffer as the pre-loaded model. After the pre-loaded model completes format verification, version verification, and necessary trial parsing verification, the edge gateway switches the pre-loaded model to the currently active model during data acquisition cycle intervals. The switching can be implemented atomically to reduce the impact of model updates on the continuity of on-site data acquisition. After switching, if the edge gateway experiences a preset number of parsing anomaly frames consecutively, it automatically rolls back to the previous stable version of the edge execution format file. Parsing anomalies can include register address misses, CRC check failures, data length mismatches, numerical out-of-range errors, data type conversion failures, or direct mapping table lookup failures. The edge gateway performs rapid local processing on these anomalies and can upload the anomaly information to the cloud processing platform for subsequent review. In this embodiment, the preset number can be 5 frames; when the edge gateway experiences 5 consecutive parsing anomalies after switching to the new device model, the edge gateway automatically rolls back to the previous stable version.

[0056] S6, the edge gateway performs deterministic parsing of the raw data from industrial sensors based on the direct mapping table in the edge execution format file, and obtains the engineering quantity data corresponding to the industrial sensors.

[0057] The edge gateway decodes the protocol frames of the raw data from the industrial sensors to obtain at least one of the following: protocol type, device address, function code, register address, node identifier, or topic name. Based on this information, it queries the direct mapping table to obtain the corresponding data point parsing rules. The edge gateway then performs byte order conversion, data type conversion, scaling calculation, and offset calculation according to the data point parsing rules, outputting the engineering quantity data.

[0058] For example, for Modbus RTU protocol data, the edge gateway can first perform frame header identification, device address identification, function code identification, data length identification, and CRC check, and then query the direct mapping table based on the function code and register address. For OPC-UA protocol data, the edge gateway can query the corresponding data point resolution rules based on the node identifier. For MQTT protocol data, the edge gateway can query the corresponding data point resolution rules based on the Topic name and message field path. This parsing process does not call the large language model; the large language model is only used for semantic parsing and semantic auditing of protocol documents in the cloud or offline.

[0059] In the specific application scenario of this embodiment, a production workshop of an automotive parts manufacturing company deploys various heterogeneous sensor devices, including Modbus RTU devices, OPC-UA devices, MQTT devices, and custom protocol devices. The Modbus RTU devices include temperature and humidity sensors, pressure transmitters, and power meters; the OPC-UA devices include robot controllers and CNC machining centers; the MQTT devices include vibration monitors; and the custom protocol devices include welding equipment using a manufacturer-specific TCP binary protocol. Users upload the corresponding protocol documents for the above devices through the web interface of the multi-format protocol document parsing engine. These documents include PDF technical manuals, Excel register lists, XML node description files, Word protocol descriptions, and PDF documents containing binary frame format descriptions. Taking the TS-300 Modbus RTU temperature and humidity sensor as an example, the user uploads the sensor's PDF technical manual, which includes communication parameter descriptions, register mapping tables, data format descriptions, and alarm threshold descriptions. The multi-format protocol document parsing engine extracts text and recognizes tables from the PDF technical manual, dividing it into blocks according to chapter titles and register mapping table boundaries. The large language model extracts the sensor's basic device information, communication parameters, and attribute information from the segmented document content.

[0060] Communication parameters may include a baud rate of 9600, 8 data bits, no parity, and 1 stop bit. Temperature attributes may include an identifier (temperature), a Chinese name (temperature), a data type (int16), a register address (30001), a function code (04), a byte order (big), a data length of 1 register word, a scaling factor (0.1), an offset (0), a unit of °C, a range of -40 to 125, and high / low alarm thresholds. Humidity attributes may include an identifier (humidity), a Chinese name (relative humidity), a data type (uint16), a register address (30002), a function code (04), a byte order (big), a data length of 1 register word, a scaling factor (0.1), an offset (0), a unit of %RH, and a range of 0 to 100.

[0061] Based on the above extraction results, the system generates a unified intermediate representation and further generates the target IoT platform object model file and edge execution format file. Taking the generation of Huawei Cloud IoT device profile as an example, a service named Environment can be defined, and temperature and humidity attributes can be set under this service; taking the generation of edge execution format file as an example, the direct mapping table records the device identifier, protocol type, communication parameters, and the two register mapping rules for temperature and humidity.

[0062] During dynamic data verification, engineers collected several raw Modbus RTU response frames from the temperature and humidity sensor during normal operation. The object model correctness automatic verification engine parsed each response frame according to the parsing rules in the edge execution format file, and determined whether the parsed temperature and relative humidity values ​​fell within their corresponding ranges, whether they conformed to the preset fluctuation characteristics of the indoor environment, and whether the error between them and the human reference values ​​exceeded the preset error threshold. After successful verification, the object model version management and distribution module distributed the edge execution format file to the target edge gateway.

[0063] After receiving the edge execution format file, the edge gateway loads it into a pre-loading buffer and performs model switching during data acquisition cycle intervals. After switching, the edge gateway parses the response frames of the TS-300 temperature and humidity sensor during real-time acquisition. For the temperature register, the edge gateway queries the direct mapping table based on the function code and register address to obtain the corresponding data type, byte order, and scaling factor, then converts the original register value into a temperature engineering value. For the humidity register, the edge gateway outputs the relative humidity engineering value in the same way. The entire real-time parsing process does not call the large language model.

[0064] The cloud processing platform is deployed on the enterprise's private cloud or intranet, and the large language model service is also deployed on the enterprise's private cloud or intranet. Sensor protocol documents are uploaded as non-production data to the large language model service for parsing. The generated unified intermediate representation and edge execution format files are distributed to the field edge gateway via the enterprise intranet. After loading the edge execution format files, the field edge gateway parses the real-time production data from the sensors locally. The real-time production data is not uploaded to external network services.

[0065] In the IoT transformation scenario of petrochemical plants, HART field instruments, Profibus-DP valve positioners, frequency converters, and custom RS-485 gas detectors may be present on-site. The cloud processing platform is deployed on the enterprise's private cloud or intranet, and the large language model service is also deployed privately. The system inputs the technical manuals, EDD files, register tables, or communication protocol descriptions of the above-mentioned devices into the cloud processing platform. The large language model performs semantic parsing on the protocol documents and generates corresponding object models. The output edge execution format file is distributed to the local edge gateway of the plant through an encrypted intranet channel. Real-time data such as pressure, temperature, flow rate, liquid level, and gas concentration during the production process are parsed and processed only locally on the edge gateway according to the edge execution format file, avoiding the transmission of real-time production data to the external network.

[0066] The multi-format protocol document parsing engine parses HART protocol documents. The HART protocol transmits digital communication by superimposing frequency shift keying signals onto a 4-20mA analog current signal. Devices can organize data according to a hierarchical command structure of general commands and manufacturer-specific commands.

[0067] For general HART commands, the multi-format protocol document parsing engine can recognize the response formats of commands 0, 1, 2, and 3, and extract device identifiers, primary variables, secondary variables, analog current values, and device status information from the response formats. For vendor-specific commands, the multi-format protocol document parsing engine can extract proprietary command numbers, command parameters, response fields, and status bit definitions from EDD files or PDF manuals.

[0068] Taking a HART temperature transmitter as an example, the generated physical model includes primary variables, secondary variables, device status, and device configuration parameters. The primary variable is the process temperature, the secondary variable is the sensor cold junction temperature, the device status includes sensor open circuit, sensor short circuit, and data saturation status, and the device configuration parameters include upper and lower limit alarm values, damping time, and sensor type.

[0069] When industrial sensor firmware is upgraded or protocol documents are changed, users can upload the upgraded industrial sensor protocol document or a difference protocol document. The multi-format protocol document parsing engine uses a large language model to identify the changes in the upgraded protocol document compared to the original protocol document. Changes may include the addition or deletion of registers, changes in register addresses, changes in function codes, changes in scaling factors, changes in data types, changes in byte order, changes in engineering units, or changes in alarm thresholds.

[0070] The system generates an update patch for the object model based on the identified changes. The update patch can include the changed fields, the content before the change, the content after the change, the applicable device model, the applicable firmware version, and a rollback flag. The automatic object model correctness verification engine verifies the update patch. Once verification is successful, the object model version management and distribution module distributes the update patch or the updated edge execution format file to the edge gateways that have deployed the corresponding industrial sensors.

[0071] After receiving an update patch for the object model, the edge gateway updates the preloaded model according to the patch content and performs hot-loading switching during data acquisition cycle intervals. If a preset number of parsing error frames occur consecutively after the switch, the edge gateway rolls back to the previous stable version; if parsing is normal after the switch, the updated edge execution format file is marked as the current stable version.

[0072] In smart pipe network scenarios for water systems, pressure sensors from different manufacturers can have different register addresses, data types, ranges, and scaling factors. For example, manufacturer A's pressure register address is 40001, data type is INT16, range is 0 to 1.6 MPa, and scaling factor is 0.0001; manufacturer B's pressure register address is 40101, data type is FLOAT32, range is 0 to 2.5 MPa, and scaling factor is 1; manufacturer C's pressure register address is 30001, data type is UINT16, range is 0 to 1.0 MPa, and scaling factor is 0.01. A multi-format protocol document parsing engine generates an independent edge execution format file for each manufacturer's model. The edge gateway maintains the correspondence between device models and edge execution format files and automatically loads the corresponding file during device registration.

[0073] The above are merely preferred embodiments of the present invention and are not intended to limit the present invention. All equivalent substitutions, simple modifications, or improvements made within the spirit and principles of the present invention should be included within the scope of protection of the present invention.

Claims

1. A method for automatic construction and edge deployment of industrial sensor object models driven by a large language model, characterized in that: Includes the following steps: S1 receives the industrial sensor protocol document and performs document preprocessing on the industrial sensor protocol document to obtain the content of the document to be parsed; S2 uses a multi-format protocol document parsing engine to call a large language model to extract structured information from the content of the document to be parsed, generating a unified intermediate representation that is independent of the IoT platform. S3 generates target IoT platform object model files and edge execution format files based on a unified intermediate representation through a multi-platform object model adaptive generator. The edge execution format file includes a direct mapping table for parsing raw data from industrial sensors at the edge. S4 verifies the correctness of the target IoT platform's object model file and edge execution format file through an automatic object model correctness verification engine; S5 sends the verified edge execution format file to the edge gateway, and the edge gateway loads the edge execution format file without interrupting the data acquisition service; S6, the edge gateway performs deterministic parsing of the raw data from industrial sensors based on the direct mapping table in the edge execution format file, and obtains the engineering quantity data corresponding to the industrial sensors.

2. The method for automatic construction and edge deployment of industrial sensor object models driven by a large language model according to claim 1, characterized in that: S2 specifically includes the following steps: S21, the document content to be parsed is divided into blocks according to at least one of document format, chapter title, table boundary and semantic boundary, and source location information including page number information is established for the block-based document content; S22, input the segmented document content into the large language model, and extract at least one of the following: basic device information, communication parameters, data point information, register information, event information, and service command information of the industrial sensor; S23. The extraction results of the large language model are constrained and output according to the preset structured output format, and the field validation and confidence level labeling of the constrained output results are performed to obtain a unified intermediate representation.

3. The method for automatic construction and edge deployment of industrial sensor object models driven by a large language model according to claim 2, characterized in that: The unified intermediate representation includes the device basic information field and the data point field; The basic device information field includes at least one of the following: manufacturer information, device model, protocol type, firmware version, and communication parameters; The data point field includes at least one of the following: data point identifier, data point name, data type, read / write permissions, engineering units, measurement range, scaling factor, offset, register address, function code, byte order, data length, alarm threshold, source location information, and confidence level identifier.

4. The method for automatic construction and edge deployment of industrial sensor object models driven by a large language model according to claim 1, characterized in that: S3 specifically includes the following steps: S31, convert at least one of the data point field, event field and service field in the unified intermediate representation into the target IoT platform object model file; S32, extract the fields required for edge parsing from the unified intermediate representation, and generate an edge execution format file. The fields required for edge parsing include at least one of register address, function code, data type, byte order, data length, scaling factor, offset, engineering unit, and range. S33, precompile the field mapping relationship in the edge execution format file into a direct mapping table. The direct mapping table includes index entries and mapping values. The index entries include at least one of protocol type, device address, function code, and register address. The mapping values ​​include at least one of data point identifier, data type, byte order, data length, scaling factor, offset, engineering unit, and range.

5. The method for automatic construction and edge deployment of industrial sensor object models driven by a large language model according to claim 1, characterized in that: S4 specifically includes the following steps: S41, Static structure verification is performed on the target IoT platform object model file and edge execution format file. Static structure verification includes at least one of the following: format legality verification, required field integrity verification, register address range verification, and data type and range matching verification. S42, parse the raw data acquired by the industrial sensor according to the edge execution format file to obtain the verification parsing value, and perform dynamic verification based on the verification parsing value; S43, input the target IoT platform object model file, edge execution format file and corresponding industrial sensor protocol document fragments into the large language model for semantic review, and obtain the semantic review result; S44. When the results of static structure verification, dynamic verification, or semantic audit do not meet the preset conditions, an error report is generated, and a re-parsing process or a manual review process is triggered.

6. The method for automatic construction and edge deployment of industrial sensor object models driven by a large language model according to claim 5, characterized in that: Dynamic verification includes at least one of value range verification, statistical distribution verification, and reference data point error verification; value range verification is used to determine whether the verification resolution value is within the range defined by a unified intermediate representation; statistical distribution verification is used to determine whether the verification resolution value corresponding to the original data collected by the industrial sensor in multiple frames conforms to the preset fluctuation characteristics of the corresponding industrial sensor. Reference data point error verification is used to determine whether the error between the verified analytical value and the manual reference value exceeds a preset error threshold.

7. The method for automatic construction and edge deployment of industrial sensor object models driven by a large language model according to claim 1, characterized in that: S5 specifically includes the following steps: S51 assigns version identifiers to verified edge execution format files through the object model version management and distribution module, generates integrity verification information, and distributes it to the edge gateway through an encrypted channel; S52, the edge gateway uses the currently running edge execution format file as the currently active model, uses the newly received edge execution format file as the preloaded model, and switches the preloaded model to the currently active model during the data acquisition cycle interval; S53, when a preset number of parsing error frames occur consecutively after the switch, the edge gateway rolls back to the previous stable version of the edge execution format file.

8. The method for automatic construction and edge deployment of industrial sensor object models driven by a large language model according to claim 1, characterized in that: S6 specifically includes the following steps: S61, the edge gateway decodes the protocol frames of the raw data from the industrial sensors to obtain at least one of the protocol type, device address, function code and register address; S62, the edge gateway queries the direct mapping table based on at least one of the protocol type, device address, function code and register address to obtain the corresponding data point parsing rule; S63, the edge gateway performs byte order conversion, data type conversion, scaling calculation and offset calculation according to the data point parsing rules to obtain the engineering quantity data; The edge execution format file contains a dedicated protocol parsing template corresponding to the industrial communication protocol. The dedicated protocol parsing template is used to perform protocol frame decoding, function code recognition, register address positioning, byte order conversion, and scaling calculation. The data point mapping relationship is determined by a direct mapping table.

9. The method for automatic construction and edge deployment of industrial sensor object models driven by a large language model according to claim 1, characterized in that: The method further includes a protocol change and update step, which includes: S71 receives upgraded industrial sensor protocol documents or differential protocol documents; S72, calls the large language model to identify the changes in the upgraded industrial sensor protocol document or the difference protocol document compared to the original industrial sensor protocol document; S73, Generate an update patch for the object model based on the changes; S74 distributes the object model update patch to the edge gateway that has deployed the corresponding industrial sensor model, so that the edge gateway updates the edge execution format file according to the object model update patch.

10. A system for automatic construction and edge deployment of industrial sensor object models driven by a large language model, characterized in that: This includes cloud processing platforms and edge gateways; The cloud processing platform includes a multi-format protocol document parsing engine, a unified intermediate representation layer for object models, a multi-platform object model adaptive generator, an automatic object model correctness verification engine, and an object model version management and distribution module. The multi-format protocol document parsing engine is used to call a large language model to extract structured information from industrial sensor protocol documents and generate a unified intermediate representation that is independent of the IoT platform. The unified intermediate representation layer for object models is used to store unified intermediate representations; The multi-platform object model adaptive generator is used to generate target IoT platform object model files and edge execution format files based on a unified intermediate representation; The automatic verification engine for object model correctness is used to verify the correctness of the object model files and edge execution format files of the target IoT platform; The object model version management and distribution module is used to distribute the verified edge execution format file to the edge gateway. It is also used to receive the protocol document or difference protocol document after the industrial sensor upgrade, identify the changes and generate object model update patches. The edge gateway includes an edge device model hot deployment and real-time execution engine, which is used to load edge execution format files without interrupting the data acquisition service and perform deterministic parsing of raw industrial sensor data based on a direct mapping table.

Images