An industrial abnormality processing orchestration and execution method and system

Abstract

The present application belongs to but is not limited to the technical field of shale gas production, and particularly relates to an industrial exception processing scheduling and execution method and system, comprising: S1, based on task semantic description, parsing the production intention into an atomic application chain; S2, utilizing a workflow executor to perform dynamic scheduling according to a data dependency topology; S3, realizing a closed-loop automation of detection, recommendation, review, review and re-optimization through an intelligent agent. The present application provides a closed-loop system of a visual interface and an artificial feedback mechanism, and operation and maintenance personnel can replay, label and correct the detected exceptions on the interface. User operation records will form high-quality data samples and be fed back to a training module as a new training set, thereby constructing a complete closed-loop mechanism of 'data->model->feedback->data'. The design overcomes the disadvantages of the traditional system that the model lacks the ability of continuous adjustment at the operation level after deployment, so that the model life cycle management has self-adaptability and continuous evolution ability.

Description

Technical Field

[0001] This invention belongs to, but is not limited to, the field of shale gas production technology, and particularly relates to an industrial anomaly handling scheduling and execution method and system. Background Technology

[0002] Existing technology example 1: WO2020250247A2—Industrial environment anomaly detection method and system This published paper proposes to improve detection performance in industrial systems by combining sensor data and quantitative / qualitative metadata, and introduces masking technology to handle missing data issues, as well as a model recommendation mechanism for transfer learning between systems of the same scale.

[0003] While this method is innovative in anomaly detection, it primarily focuses on how to detect anomalies more accurately, without delving into how to automatically orchestrate multiple modules into a closed-loop workflow at the detection result level, or proposing mechanisms for semantic task description and intelligent agent-driven workflow planning. Furthermore, it lacks a systematic closed-loop design for continuous iteration and optimization based on user feedback after the detection model is deployed.

[0004] Example of prior art 2: US12393169—Methods and apparatus for monitoring industrial equipment The patent discloses a sensor signal data analysis and abnormal behavior identification technology for industrial equipment, which compares log events with timestamps of abnormal behavior to assist in root cause analysis; it also mentions a method that combines rule mechanisms with clustering or unsupervised clustering in the absence of expert knowledge.

[0005] However, this solution still primarily relies on the temporal overlap of events and anomalies for judgment, resulting in a relatively static and limited processing logic. It lacks a universal semantic task decomposition and workflow scheduling mechanism, and it doesn't offer the ability to automatically orchestrate multiple stages such as detection, recommendation, review, debriefing, and optimization into a unified workflow. Furthermore, the solution's module integration and cross-module dependency coordination capabilities are weak, making it difficult to handle scenarios with frequent interactions between modules in complex industrial systems.

[0006] Similar technical example 3: US20230281023A1—Ontology-based workflow automation and execution This application discloses a workflow creation, execution, and platform mechanism driven by semantic ontology, which aims to automate the connection and management of business processes in the form of ontology semantics.

[0007] While it proposes an ontology-driven approach to process automation, its application scenarios primarily lean towards business processes or general computing processes, rather than addressing the needs of industrial anomaly detection and process optimization. In the highly coupled, time-series, and dependent scenario of industrial anomaly handling, it lacks in-depth design for mechanisms such as "automatic selection of heterogeneous modules, data dependency topology scheduling, and detection-optimization closed loop." Furthermore, the solution does not disclose how to convert natural language task intent into task semantic chains, nor how to utilize user feedback for model optimization within a closed-loop mechanism.

[0008] These three similar technologies each involve aspects such as anomaly detection, ontology workflow automation, and industrial equipment monitoring, but none fully cover the complete chain of this invention: "task semantic description → intelligent agent planning → workflow executor scheduling based on data dependency topology → detection-recommendation-review-re-optimization closed loop," nor the closed-loop design of visual interaction and user feedback for model iteration. This invention has significant innovative advantages in automatic orchestration, closed-loop optimization, and cross-module collaboration. Summary of the Invention

[0009] To address the problems existing in the prior art, this invention provides an industrial anomaly handling orchestration and execution method and system.

[0010] This invention discloses an industrial anomaly handling orchestration and execution method to solve the problems of existing industrial anomaly handling processes, such as reliance on manual experience, severe system fragmentation, non-reusable processes, and low execution efficiency.

[0011] The method of this invention includes the following steps: First, natural language task input is received from operators or system terminals. An intelligent agent, combining a pre-built industrial semantic knowledge structure, performs semantic parsing and intent recognition on the natural language task, converting production goals, anomaly types, constraints, and expected output results into a structured task semantic description. This task semantic description abstracts data acquisition, feature processing, model detection, alarm generation, and process optimization into a unified task semantic unit, and describes the input source, processing rules, output goals, and upstream and downstream dependencies in a standardized field format, thereby achieving the conversion from unstructured production intent to a computable semantic representation.

[0012] Secondly, the intelligent agent, based on the input-output constraint model (Schema) of each atomic application in the pre-built capability library, performs matching analysis on the semantic description of the task, automatically completes microservice selection and process orchestration, and generates atomic application chain scripts described by a structured workflow language. These atomic applications include data access modules, signal cleaning modules, feature engineering modules, anomaly detection model modules, alarm strategy modules, and process parameter optimization modules. The scripts explicitly define the parameter configurations, data channels, and calling order of each atomic application, enabling the automated construction of anomaly handling logic.

[0013] Furthermore, the underlying workflow execution engine parses the atomic application chain scripts, converting them into a Directed Acyclic Graph (DAG) structure, where each node corresponds to an atomic application unit, and edges represent data flow dependencies. During operation, the execution engine monitors the status of each node in real time and implements dynamic scheduling based on the pre-existing data dependencies between nodes: concurrent scheduling is performed for nodes without data dependencies; synchronous blocking scheduling is performed for nodes with mandatory flow relationships; and a rollback or retry mechanism is triggered when an abnormal node fails. Through this DAG scheduling mechanism, a hybrid serial and parallel execution is achieved, improving the utilization of industrial computing resources, reducing end-to-end execution latency, and avoiding redundant computations and resource conflicts.

[0014] Upon completion, the system automatically collects detection results, alarm information, and manual review feedback, feeding these back as closed-loop data to the intelligent agent module. Based on the execution results and feedback, the intelligent agent updates the microservice orchestration strategy and adaptively optimizes the anomaly detection model parameters, achieving a continuous closed-loop evolution of detection, recommendation, review, debriefing, and further optimization, thereby continuously improving the accuracy of anomaly handling and the reliability of decision-making.

[0015] This invention also provides an industrial anomaly handling orchestration method based on workflow language. By automating the chaining of data acquisition, feature processing, model inference, alarm output, and process optimization steps, an executable workflow definition script file is generated. This script file can be loaded and parsed into a DAG topology by a workflow execution engine, progressively calling various microservice modules to complete the fully automated execution of the industrial anomaly handling process.

[0016] Furthermore, the workflow definition file is a script file that can be loaded by the workflow executor and call each microservice module step by step to complete the entire process of industrial exception handling.

[0017] Another objective of this invention is to provide an intelligent agent-driven planning method for industrial anomalies. After receiving a structured task description, the intelligent agent retrieves available microservice components from a capability library, automatically selects the required components, plans their execution order and data dependencies, and outputs a complete workflow definition file.

[0018] The capability library includes data extraction applications, filtering applications, statistical computing applications, and model inference applications. The intelligent agent automatically selects and combines components based on the task objectives.

[0019] Another object of the present invention is to provide an industrial anomaly handling orchestration and execution system comprising: The task semantic parsing module is used to parse task intents into atomic application chains.

[0020] The workflow execution module is used to dynamically invoke microservices according to the data dependency topology.

[0021] The intelligent agent module is used to achieve closed-loop automation and adaptive task planning.

[0022] The semantic interaction module is used to receive task input in natural language and convert it into a structured task description.

[0023] The visualization interface module is used to visualize the anomaly detection process, alarm generation process, and process optimization process, and supports manual feedback of results back to the training dataset.

[0024] Another object of the present invention is to provide a computer device including a memory and a processor, the memory storing a computer program, which, when executed by the processor, causes the processor to perform the steps of the industrial exception handling orchestration and execution method.

[0025] Another object of the present invention is to provide a computer-readable storage medium storing a computer program, which, when executed by a processor, causes the processor to perform the steps of the industrial exception handling orchestration and execution method.

[0026] Another objective of this invention is to provide an information data processing terminal for implementing the industrial anomaly handling orchestration and execution system.

[0027] Based on the above technical solutions and the technical problems solved, the advantages and positive effects of the technical solution to be protected by this invention are as follows: This invention realizes a complete industrial anomaly handling system, from natural language intent parsing to executable workflow script generation, and then to DAG-level dynamic scheduling execution and closed-loop optimization. It significantly improves the automation level of anomaly handling, process reusability and system scalability, and has good engineering application value and industrial promotion prospects.

[0028] This invention provides a closed-loop system with a visual interface and a human feedback mechanism, allowing maintenance personnel to replay, annotate, and correct detected anomalies on the interface. User operation records form high-quality data samples, which are fed back to the training module as new training sets, thus constructing a complete closed-loop mechanism of "data → model → feedback → data". This design overcomes the shortcomings of traditional systems that lack the ability to continuously adjust the model at the operational level after deployment, enabling adaptive and continuous evolution capabilities in model lifecycle management.

[0029] Because user-annotated results are automatically incorporated into the training system, the model can be retrained or fine-tuned periodically to adapt to the constantly changing production characteristics and data distribution in the system's operating environment. The closed-loop feedback mechanism enables the model to improve not only offline but also online, responding to new situations and continuously enhancing predictive performance and stability. Compared to traditional one-time training models, the system possesses long-term adaptive optimization capabilities.

[0030] Regarding the accuracy of anomaly detection, this invention integrates user feedback data to continuously correct model misjudgment trends, reducing false positive and false negative rates. Especially in boundary situations such as fluid accumulation and pressure differential anomalies, after users correct misclassifications through the interface, the system can absorb these cases to optimize the algorithm, significantly enhancing the model's ability to judge edge samples. Therefore, this invention not only ensures safe industrial operation but also improves the reliability and usability of detection results.

[0031] (1) The expected benefits and commercial value of the technical solution of this invention after transformation are as follows: This invention, by achieving automated closed-loop orchestration and execution of industrial anomaly handling, can significantly reduce the frequency of manual intervention by maintenance personnel, improve anomaly response speed by more than 30%, and reduce unplanned downtime. It is estimated to save medium to large-sized shale gas or oil and gas production enterprises hundreds of millions of yuan in maintenance costs annually. Simultaneously, through continuous model optimization driven by user feedback, the false alarm rate can be reduced to below 10%, improving production safety and process stability, resulting in significant economic and social benefits. This solution is easily deployed on existing industrial IoT platforms and has excellent industrialization prospects and commercial promotion value.

[0032] (2) The technical solution of this invention fills a technical gap in the industry both domestically and internationally: While existing technologies include research on single aspects such as anomaly detection and workflow automation, a complete closed-loop automation solution that organically combines natural language task intent → semantic decomposition → agent intelligent planning → dynamic scheduling of data dependency topology → detection-recommendation-review-re-optimization has yet to be seen. This is particularly true in the field of industrial anomaly handling, where a complete implementation of microservice automatic selection based on a capability library and workflow language-driven digital support platform is lacking. This invention proposes for the first time an agent-driven closed-loop workflow orchestration mechanism for industrial scenarios, filling a systemic technological gap in this field.

[0033] (3) The technical solution of the present invention solves a technical problem that people have long wanted to solve but have never been able to solve successfully: Industrial anomaly handling has long suffered from problems such as fragmented modules, excessive manual intervention, and difficulty in continuous optimization after model deployment. The industry has consistently sought to achieve a fully automated closed-loop process from anomaly detection to process optimization, but has been hampered by technical challenges such as data dependency coordination between heterogeneous modules, automatic conversion of task semantics into executable workflows, and effective feedback data return, preventing the development of an efficient and reliable system-level solution. This invention successfully addresses this long-standing industry pain point by combining task semantic description, agent planning, and dynamic topology scheduling.

[0034] (4) The technical solution of the present invention overcomes technical bias: Existing technologies generally suffer from biases such as "real-time industrial anomaly handling must rely on fixed rules and human experience," "agent-based intelligent planning is unsuitable for high-reliability industrial scenarios," and "workflow orchestration struggles to achieve dynamic data-dependent scheduling." This invention overcomes these biases, demonstrating that agents, supported by a capability library, can safely and reliably plan industrial tasks and achieve efficient dynamic execution through data-dependent topology scheduling. Furthermore, by incorporating user feedback, a continuous model evolution loop is formed, thus successfully applying advanced semantic and agentic technologies in the industrial field. Attached Figure Description

[0035] Figure 1 This is a flowchart of the industrial anomaly handling orchestration and execution method provided in the embodiments of the present invention.

[0036] Figure 2 This is a structural diagram of the industrial anomaly handling orchestration and execution system provided in an embodiment of the present invention. Detailed Implementation

[0037] To make the objectives, technical solutions, and advantages of this invention clearer, the invention will be further described in detail below with reference to embodiments. It should be understood that the specific embodiments described herein are merely illustrative and not intended to limit the invention.

[0038] In existing industrial anomaly detection and handling practices, there is a common problem of disconnect between engineering knowledge and computer implementation capabilities. Petroleum engineers have a deep understanding of processes such as oil production, gathering and transportation, injection and production, and separation, and can accurately determine the process location and handling logic of anomalies. However, due to unfamiliarity with underlying data structures, algorithm models, and programming languages, they find it difficult to translate this process experience into executable anomaly detection and handling procedures. This results in anomaly handling being highly dependent on software engineers, with rigid procedures, high adjustment costs, and difficulty in adapting to dynamic changes in field conditions.

[0039] To address the technical problems of the existing technologies, this invention proposes an industrial anomaly handling orchestration method and its execution system oriented towards process flow. This method does not require petroleum engineers to directly access underlying data or computer code; instead, it provides a process orchestration language that allows anomaly detection and handling logic to be defined at the process flow level. The system encapsulates anomaly detection, feature processing, model judgment, alarm generation, and process optimization capabilities into standardized, reusable anomaly handling operators, each corresponding to a clear process semantic and processing function. Petroleum engineers only need to select and connect the corresponding operators based on a familiar process flow diagram to complete the definition of the anomaly detection and handling process.

[0040] Simultaneously, this invention provides a workflow execution system to parse and execute the process orchestration results, automatically mapping the process definition at the process level to low-level executable exception handling tasks. This approach transforms industrial exception handling from "code-driven" to "process-driven," enabling engineers to directly utilize their own process experience to build and adjust exception handling processes, significantly improving the flexibility, maintainability, and engineering applicability of exception detection and handling.

[0041] like Figure 1 As shown, the industrial anomaly handling orchestration and execution method provided in this embodiment of the invention includes: S1, based on task semantic description, parses the production intent into an atomic application chain.

[0042] The system expresses the intent for anomaly detection and handling in natural language or simple industrial flowcharts. Using advanced semantic parsing technology, it transforms these intents into a chain of atomic-level applications. These atomic applications are executable operators specifically designed for industrial anomaly handling, covering various stages such as data acquisition, feature processing, model detection, alarms, and process optimization. In this way, petroleum engineers can translate their industrial experience into executable anomaly handling task descriptions without writing code.

[0043] S2 utilizes the workflow executor to dynamically schedule tasks according to the data dependency topology.

[0044] S3 achieves closed-loop automation of detection, recommendation, review, debriefing, and re-optimization through intelligent agents.

[0045] The task semantic description provided in this embodiment of the invention includes representing data acquisition, feature processing, model detection, alarm generation, and process optimization as a unified task semantic unit.

[0046] The workflow executor provided in this embodiment of the invention constructs a topology structure based on the data dependencies between tasks, and calls each task unit in the order of the topology structure to avoid redundant calculations and achieve parallel scheduling.

[0047] The capability library includes data extraction applications, filtering applications, statistical computing applications, and model inference applications. The intelligent agent automatically selects and combines components based on the task objectives.

[0048] In the industrial anomaly handling orchestration and execution method provided in this embodiment of the invention, the signal data processing process is the core link connecting industrial field perception and precise anomaly handling. It runs through the entire process of atomic application chain construction, workflow dynamic scheduling and intelligent agent closed-loop optimization. Relying on task semantic parsing, workflow scheduling and intelligent agent technology, it realizes the full-link automated processing of signal data from acquisition to application, from detection to optimization, and provides data support for precise identification and efficient handling of industrial anomalies.

[0049] The signal data processing process begins with a task semantic description and unfolds logically according to the steps of "acquisition-preprocessing-feature extraction-detection and analysis-result application," deeply integrating with the core technical steps in the invention embodiments. Firstly, in the S1 step atomic application chain construction phase, signal data acquisition serves as the first atomic application unit. Through the unified definition of task semantic units, the signal acquisition requirements of various sensors in the industrial field (such as pressure, temperature, and vibration sensors) are transformed into executable operators. Without manual coding, petroleum engineers can clearly define the range, frequency, and accuracy requirements of signal acquisition using natural language or simple flowcharts. The system automatically calls the corresponding data acquisition microservice to complete the real-time acquisition and initial caching of raw signals from the industrial field.

[0050] The acquired raw signal needs to undergo a preprocessing stage. This stage integrates microservice components such as filtering and statistical calculation applications from the capability library. The intelligent agent automatically selects the appropriate processing component based on the task objective. After receiving the structured task description, the intelligent agent searches for available components in the capability library. Addressing issues such as noise, outliers, and missing data in the raw signal, it automatically calls the filtering application to denoise the signal and uses the statistical calculation application to remove outliers and complete missing data. This transforms the chaotic raw signal into well-organized, usable standardized data, laying the foundation for subsequent feature processing.

[0051] The preprocessed standardized signal enters the feature processing stage. As a key link in the atomic application chain, the system uses semantic parsing technology to transform the user's feature extraction intent into corresponding atomic applications. These applications call feature processing microservices to extract key features related to industrial anomalies from the signal, such as signal peak value, fluctuation frequency, and trend change rate. This transforms one-dimensional signal data into a high-dimensional feature vector, enhancing the identification of anomaly signals. Simultaneously, the workflow executor dynamically schedules feature processing tasks with preceding data acquisition and preprocessing tasks based on data dependency topology, ensuring the orderly processing of data, avoiding redundant calculations, and improving processing efficiency.

[0052] After feature processing, the signal data enters the model detection stage. The intelligent agent calls upon the model inference application in its capability library, inputs the feature vector into the anomaly detection model, and completes the identification and judgment of abnormal signals, outputting the anomaly level, anomaly type, and preliminary handling suggestions. The detection results generate alarm information through the alarm atomic application and are simultaneously transmitted to the intelligent agent, entering a closed loop of review, debriefing, and further optimization. The intelligent agent automatically reviews the detection results, debriefs the processing effect based on historical processing data, and if there are detection deviations, automatically adjusts the feature processing parameters, model parameters, and component combination methods to optimize the signal data processing flow and achieve continuous improvement in processing accuracy.

[0053] The entire signal data processing process is driven by scripts in the workflow definition file. The workflow executor dynamically schedules each link according to the data dependency topology. The intelligent agent is responsible for component selection, process planning and optimization throughout the process without human intervention. This not only solves the problems of cumbersome traditional industrial signal processing processes and high professional requirements for operators, but also ensures the accuracy and efficiency of data processing through closed-loop optimization, providing reliable data support for the rapid and accurate handling of industrial anomalies.

[0054] In the industrial anomaly handling orchestration and execution system proposed in this invention, a collaborative mechanism runs through the entire process of task understanding, process construction, execution control, and continuous optimization. Through task semantic parsing, workflow scheduling, and intelligent agent collaborative operation, the system realizes the transformation of industrial anomaly handling from "experience-driven" to "intelligent autonomy." The system first uses task semantic description as the entry point for human-machine collaboration, uniformly modeling the anomaly detection and handling intentions expressed by petroleum engineers in natural language or flowcharts. The semantic parsing module performs structured analysis on the task description, identifying the implicit data sources, processing logic, judgment conditions, and optimization goals, and decomposes it into standardized task semantic units. Each task semantic unit corresponds to an atomic application operator with clear inputs, outputs, and functional boundaries, thereby achieving automatic mapping of industrial experience to executable logic.

[0055] Building upon this foundation, the workflow executor plays a crucial scheduling role. The system constructs a directed acyclic topology based on the data dependencies between semantic units of each task, clearly defining the sequential constraints and parallel relationships between different processing stages. During execution, the workflow executor dynamically schedules atomic application chains according to the topology sorting results. While ensuring data consistency and execution correctness, it automatically triggers parallelizable task units, achieving efficient utilization of computing resources. Simultaneously, a dependency-aware mechanism avoids redundant data processing and invalid calls, ensuring high real-time performance and reliability in the exception handling process, adapting to the complex and ever-changing operating environment of industrial sites.

[0056] Furthermore, this invention introduces an intelligent agent as the decision-making and evolutionary hub of the collaborative mechanism. Upon receiving a structured task semantic description, the intelligent agent automatically retrieves matching microservice components from its capability library, including application capabilities such as data extraction, filtering preprocessing, statistical analysis, model inference, and optimization decision-making. Based on the task objectives, it combines and orchestrates these components to generate a complete workflow definition file. This definition file is described using a unified workflow language and can be directly loaded and executed by the workflow executor, achieving decoupling between task planning and task execution.

[0057] During the execution phase, the intelligent agent continuously participates in the anomaly handling closed loop. Through comprehensive analysis of detection results, alarm feedback, and manual review information, it evaluates and corrects existing model parameters, processing strategies, and process structures. When anomaly patterns evolve or on-site conditions change, the intelligent agent can automatically adjust component selection and execution order, regenerate or optimize workflow definition files, and achieve continuous iteration of anomaly detection, handling, and process optimization. Through the above collaborative mechanism, this invention constructs a multi-role, multi-level collaborative industrial anomaly handling system, enabling the system to have adaptive, self-learning, and self-optimizing capabilities, significantly improving the automation level and engineering application value of industrial anomaly handling.

[0058] In this embodiment of the invention, workflow orchestration includes: The industrial anomaly handling process is uniformly described and orchestrated using TWL (Workflow Language) provided by the digital support platform. Multiple processing steps, such as "data acquisition → feature processing → model detection → alarm → process optimization," are automatically linked in a machine-executable manner, thereby achieving end-to-end automated collaboration across systems and algorithm modules. The TWL script declaratively describes the inputs, outputs, execution conditions, and dependencies of each processing node, enabling the standardized representation, reuse, and scheduling of complex industrial processes.

[0059] Building upon this foundation, an agent-driven workflow auto-planning mechanism is introduced. Upon receiving a structured task description, the agent first performs semantic understanding and reasoning on the task objectives and constraints, and automatically queries a capability library. This library includes various microservice components registered in the business application layer of the digital support platform, such as data extraction apps, filtering apps, statistical computing apps, model inference apps, and result analysis apps. Based on the matching relationship between task objectives and capabilities, the agent automatically selects the required functional components and plans their execution order and data dependencies, forming an execution path that satisfies logical consistency and data accessibility. Finally, the agent transforms the planning results into standardized workflow definition files, such as TWL script files, thereby achieving automatic transformation and deployment from "intent description" to "executable process."

[0060] In this embodiment of the invention, workflow dynamic execution and data materialization include: The Workflow Executor loads and executes the workflow generated in the previous step, calling each microservice in sequence to complete the data processing.

[0061] The agentic execution includes: supporting input of natural language task descriptions (such as "detecting the risk of fluid accumulation on platform X in the next 24 hours"), the system automatically triggers the workflow to complete data query, model selection, inference, and result output.

[0062] In industrial anomaly handling scenarios, traditional technical approaches often suffer from fragmented processes and excessive human intervention. For example, data acquisition and feature processing, model inference and alarm generation, and process optimization and post-mortem improvement are typically performed independently by different tools and personnel, lacking a unified semantic description and orchestration mechanism. This decentralized operation not only increases response time but also makes the effectiveness of anomaly detection highly dependent on expert experience, making it difficult to meet the real-time and closed-loop improvement requirements of complex production environments. The industrial anomaly handling orchestration and execution method proposed in this invention systematically addresses these pain points in industrial applications.

[0063] Its core lies in decomposing abstract production intentions into a chain of executable atomic applications through task semantic description. This semantic description, based on a unified ontology and task modeling framework, allows steps such as detection, recommendation, review, and optimization to be expressed as logically clear task units. The advantage of this abstract modeling is that it shields the differences in underlying implementations; different algorithm modules, data sources, or process units can be uniformly orchestrated, thereby improving the system's scalability and cross-scenario applicability.

[0064] In terms of execution mechanism, a Workflow Executor is introduced as the core task-driven scheduler. It does not execute in a fixed linear order, but rather dynamically generates the topology based on data dependencies and triggers each stage as needed. This topology-based execution approach ensures that upstream data processing results are delivered to downstream modules in a timely manner, while avoiding redundant computation and significantly improving overall processing efficiency. Compared to statically configured pipeline systems, this method can achieve adaptive scheduling adjustments in the face of data fluctuations or changes in task objectives.

[0065] The Agent plays a role in intelligent orchestration and planning throughout the industrial anomaly handling process. Its working principle is not simply component retrieval and assembly, but rather semantic-driven generation based on the industrial ontology graph as a contextual constraint foundation. Specifically, the Agent first uses the underlying industrial ontology graph to semantically align the input natural language or structured task description, mapping the involved equipment entities, process parameters, anomaly types, and their relationships to standard concept nodes and relational edges in the ontology. This clarifies the equipment hierarchy and process topology of the task, forming a task context space with semantic boundaries.

[0066] After semantic alignment, the Agent invokes the standard interface declarations of each microservice component in the capability library. These interface declarations include input data type definitions, output data structure definitions, and functional semantic tags. Under the entity relationships and data semantic constraints defined by the industrial ontology graph, the Agent performs matching, filtering, and combination planning for candidate microservice components. During the matching process, the Agent verifies whether the data type and semantic tags of the upstream component's output fields meet the input constraints of the downstream component, ensuring data flow consistency at both the type and semantic layers.

[0067] Based on the aforementioned semantic constraints and interface matching mechanism, the Agent combines and orchestrates multiple atomic functional modules to generate a structured workflow script that conforms to the syntax specifications of the underlying execution engine, achieving automatic transformation from task intent to executable process definition. This mechanism uses the industrial ontology graph as the core constraint condition for the generation logic, constituting a key technical feature that distinguishes it from traditional rule-matching orchestration.

[0068] The capability library encompasses microservice components such as data extraction, filtering, statistical computation, and model inference. The Agent autonomously selects modules and plans their order and dependencies based on target constraints and data characteristics. The resulting workflow definition file can be directly loaded by the Workflow Executor, ensuring full automation from task description to execution.

[0069] During dynamic execution and data materialization, the Workflow Executor progressively calls various microservices to complete data collection, cleaning, feature engineering, model inference, and alarm triggering. Unlike alarms triggered by static rules, this process directly couples model detection with process optimization, ensuring that optimization suggestions are fed back to the production control layer in real time. Through this dynamic execution approach, anomaly handling shifts from passive response to proactive prediction and prevention.

[0070] Finally, the design of Agentic execution enables the system to have higher interactivity and adaptability. Users only need to input the task objective in natural language, such as predicting the abnormal risk of a production platform in the next 24 hours, and the system can automatically complete the workflow construction and execution, and present the results in a visual form. This process breaks down the barriers between data, models, and decision-making, and builds a truly automated closed-loop processing mechanism, significantly improving the intelligence level of industrial anomaly management and its industrial application value.

[0071] like Figure 2 As shown, the industrial anomaly handling orchestration and execution system provided in this embodiment of the invention includes: The task semantic parsing module is used to parse task intents into atomic application chains.

[0072] The workflow execution module is used to dynamically invoke microservices according to the data dependency topology.

[0073] The intelligent agent module is used to achieve closed-loop automation and adaptive task planning.

[0074] The semantic interaction module is used to receive task input in natural language and convert it into a structured task description.

[0075] The visualization interface module is used to visualize the anomaly detection process, alarm generation process, and process optimization process, and supports manual feedback of results back to the training dataset.

[0076] The industrial anomaly handling orchestration and execution system provided in this invention takes "semantic-driven orchestration, dependency-aware scheduling, and closed-loop intelligent optimization" as its core ideas. It performs unified modeling and automated execution of the anomaly detection, analysis, handling, and review process in industrial scenarios, realizing the transformation of anomaly handling from passive response to proactive collaboration and continuous optimization.

[0077] During system runtime, the task semantic description module first performs semantic modeling of the anomaly handling task. This module receives anomaly handling intentions from operations personnel, business systems, or upstream detection systems, parses these intentions into structured semantic representations, and then maps them into a set of executable atomic application chains. Each atomic application corresponds to a standardized functional unit, such as anomaly detection, root cause analysis, impact assessment, handling suggestion generation, and manual review interface, and uses semantic tags to describe its input and output data types, dependencies, and execution constraints. This transforms the handling process, which originally relied on human experience, into a machine-understandable, programmable, and executable process model.

[0078] After obtaining the atomic application chain, the dynamic scheduling module constructs a task execution topology based on the data dependencies between atomic applications, and the Workflow Executor dynamically schedules this topology. During scheduling, a fixed execution order is not used; instead, task execution is triggered based on the actual data readiness state and dependency fulfillment, thus supporting complex process control such as parallel execution, conditional jumps, and exception rollbacks. When an atomic application completes execution and generates output data, the system automatically passes it as input to subsequent dependent tasks and dynamically adjusts the execution path based on the runtime state to adapt to factors such as data fluctuations, resource changes, and anomalies in the industrial environment.

[0079] During and after task execution, the system utilizes a closed-loop automation module for continuous monitoring and optimization. This module, centered on an intelligent agent, integrates anomaly detection results, recommended handling plans, manual review and feedback, and the final handling effect, forming a closed-loop processing flow of "detection—recommendation—review—re-optimization." Through the accumulation and analysis of historical anomaly cases, handling effects, and feedback information, the system continuously optimizes semantic mapping rules, scheduling strategies, and recommendation model parameters, making subsequent anomaly handling more accurate, efficient, and stable.

[0080] At the system level, this invention also implements the above-mentioned method in engineering through computer equipment, computer-readable storage media, and information data processing terminals. The processor in the computer equipment executes program instructions in the memory to achieve the coordinated operation of various modules; the computer-readable storage media is used to carry the program and support system deployment and migration; the information data processing terminal provides users with a human-computer interaction interface for anomaly management, process configuration, and result display, thereby constructing a complete automated and intelligent operation system for industrial anomaly handling.

[0081] As a preferred improvement of the present invention, an intelligent planning method for industrial anomaly handling is provided in an embodiment of the present invention, comprising the following steps: The system receives production intentions for industrial anomalies through a pre-set intelligent agent module, which then extracts task objectives and constraints based on natural language processing technology.

[0082] The production intent is semantically aligned using a pre-built industrial ontology graph to determine the equipment entities, process attributes, and relationship topology between entities involved in the task.

[0083] A structured task semantic description is generated based on the semantic alignment result. The task semantic description includes data objects, processing targets, and control logic.

[0084] The standard interface declarations of each microservice component in the capability library are invoked. These standard interface declarations include input constraints and output constraints.

[0085] Under the semantic constraints of the industrial ontology graph, the microservice components are automatically filtered and combined to ensure that the output data type of the upstream components is consistent with the input constraints of the downstream components.

[0086] Generate structured atomic application chain scripts that conform to the preset workflow syntax.

[0087] After the industrial anomaly handling process is completed, the detection results and manual review results are received, and the results are used as labeled samples to supplement the training set, triggering adaptive fine-tuning or retraining of the parameters of the underlying anomaly detection model, and updating the intelligent agent's screening rules for microservice components.

[0088] To verify the practical technical effectiveness of this invention, a four-month simulated deployment and test was conducted in the actual production environment of a shale gas production enterprise, involving three gas extraction platforms and processing approximately 100,000 sensor data points daily. The test results are as follows: 1. Evidence for improvements in anomaly detection accuracy and false alarm rate: Before deployment, the traditional rule-based offline model approach had an average false positive rate of 28.7% and a false negative rate of 16.2%. After deploying the system of this invention, through closed-loop optimization based on user feedback (collecting over 2,800 valid labeled samples for monthly model fine-tuning), the false positive rate decreased to 9.8% and the false negative rate decreased to 4.9%. Particularly in the scenario of liquid accumulation risk prediction, the ability to identify edge samples was significantly improved, with an accuracy increase of 25%.

[0089] 2. Evidence regarding abnormal response time and processing efficiency: In the traditional workflow, the average time from anomaly detection to generating process optimization suggestions is 45 minutes (including multi-system switching and manual review). After adopting the Agent automatic planning and Workflow Executor dynamic scheduling of this invention, the end-to-end processing time is reduced to less than 5 minutes, improving response efficiency by 89%. The system supports natural language task input, and the average processing time for a single task, from intent description to result output, is less than 3 minutes.

[0090] 3. Evidence regarding maintenance workload and economic benefits: During the pilot program, the average daily anomaly handling time for operations and maintenance personnel decreased from 6.2 hours to 2.1 hours, and the frequency of manual intervention decreased by 66%. Unplanned downtime incidents decreased by 4, and based on an estimated loss of approximately 500,000 yuan per downtime, a total economic loss of approximately 2 million yuan was avoided over 6 months.

[0091] 4. Evidence of the effectiveness of closed-loop optimization iteration: User feedback data collected through a visual interface was automatically fed back into the training set. The model underwent four iterations of optimization during the pilot phase, with an average accuracy improvement of 3.5% on the validation set after each iteration. This demonstrates the effectiveness of the closed-loop mechanism of "data → model → feedback → data" and shows that the system possesses long-term adaptive evolution capabilities.

[0092] The test data mentioned above was recorded by the company's internal monitoring system and confirmed by a third-party audit, which fully demonstrates that the present invention has achieved significant technical effects in terms of anomaly detection accuracy, processing efficiency, reduced operation and maintenance costs, and continuous model optimization, and has achieved the expected industrial application value.

[0093] Example 1 At a natural gas compression station, maintenance personnel input the task intent as "anomaly detection and process optimization." The system first decomposes the task into atomic application chains such as data acquisition, feature processing, model detection, alarm generation, and process optimization through task semantic parsing, and generates a unified task semantic unit.

[0094] The workflow executor constructs a topology based on the data dependencies between tasks. For example, the output of feature processing is directly used as the input for model detection, and alarm generation depends on the model detection results. The executor dynamically schedules each step according to the topological order, avoiding repeated calculations on the same data and supporting parallel execution of multiple processes.

[0095] Example 2 In the monitoring of production units in refining and chemical enterprises, the system utilizes the workflow language of the digital support platform to describe anomaly detection tasks. The statements sequentially define a data acquisition module, a feature processing module, a model inference module, an alarm module, and a process optimization module, generating a workflow definition file.

[0096] This file is in standard script format and can be directly loaded by the workflow executor. During execution, the executor reads the script statements line by line and calls the corresponding microservices, enabling the entire process from data acquisition to process optimization to be automated and connected without manual configuration.

[0097] Example 3 In the scenario of predicting the risk of fluid accumulation in oil and gas fields, the intelligent agent receives a structured task description of "predicting abnormal fluid accumulation in the next 24 hours". The agent first queries the capability library and retrieves a set of components including data extraction applications, filtering applications, statistical computing applications, and model inference applications.

[0098] Based on the task objectives, the agent automatically selects the necessary components and determines the execution order: first, data extraction is performed; then, signal filtering and statistical calculations are conducted; finally, model inference is invoked to output abnormal results. After determining the component dependencies, the agent generates a directly executable workflow definition file, which is then loaded by the executor.

[0099] Example 4 At a liquefied natural gas storage and transportation station, the system deployed a task semantic parsing module, a workflow execution module, and an intelligent agent module. Maintenance personnel input natural language commands such as "detect temperature anomalies in the storage tank over the next 12 hours" into the visual interface, and the semantic parsing module converts this into a structured task description.

[0100] The semantic interaction module delegates tasks to the intelligent agent for planning, generates workflows, and then schedules and executes them through the workflow execution module. The visualization interface module displays temperature data curves, alarm statuses, and process optimization suggestions in real time, while also allowing operations and maintenance personnel to annotate and provide feedback on the results. Feedback information is fed back into the dataset for model retraining.

[0101] Specific implementation plan for intelligent planning method for industrial anomaly handling.

[0102] This solution targets the petrochemical production scenario. Based on the intelligent planning method described in the claims, it clarifies the specific implementation details, technology selection, and parameter configuration of each step to ensure that the method is feasible and reusable. It enables automated and intelligent planning for industrial anomaly handling and is suitable for the production anomaly handling needs of small and medium-sized petrochemical enterprises.

[0103] Intelligent Agent Module Deployment: The intelligent agent module is developed using Python, integrating the BERT natural language processing model, and deployed on an industrial edge server. It supports input of production intentions in the form of natural language and simple flowcharts. Operators input intentions such as "detect abnormal wellhead pressure, trigger alarm and optimize production process" through the terminal. The module automatically extracts the task objectives (pressure anomaly detection, alarm, process optimization) and constraints (response time ≤ 10 minutes, no impact on normal production).

[0104] Industrial ontology graph construction: Based on the Neo4j graph database, an industrial ontology graph is constructed, comprising an equipment entity layer, a process parameter layer, and an anomaly event layer. The equipment entity layer includes 120 entities such as wellheads and sensors; the process parameter layer includes 30 parameters such as pressure and flow rate; and the anomaly event layer covers 25 types of events such as pressure anomalies. Each layer is linked through a semantic relationship of "equipment-parameter-anomaly," supporting real-time updates and semantic queries.

[0105] Structured task semantic description generation: Based on semantic alignment results, production intent is transformed into unified task semantic units, covering 5 types of units including data acquisition and feature processing. Each unit specifies input fields (such as pressure sensor ID), output fields (such as anomaly level), and execution rules (such as sampling frequency of 1 time / minute), and stores them in JSON format to ensure that the machine can parse them.

[0106] Capability Library and Microservice Invocation: A microservice capability library is built, containing 15 atomic application units (data acquisition, filtering, etc.). Each unit corresponds to a single function and provides standard interface declarations through RESTful interfaces, clearly defining input and output constraints (e.g., the data acquisition component's input is the sensor ID, and the output is a real-time value). The intelligent agent module verifies component compatibility through interface calls.

[0107] Microservice composition and script generation: Under the semantic constraints of the ontology graph, compatible components are automatically selected and combined in the order of "data acquisition → feature processing → model detection → alarm generation → process optimization" to ensure that upstream output matches downstream input. Using BPMN 2.0 workflow syntax, structured atomic application chain scripts are generated, defining the identifiers of each node, data sources, and data flow relationships. The scripts can be directly loaded and executed by the workflow executor.

[0108] Solution Testing and Optimization: Fifty historical anomaly cases were selected for testing to ensure semantic extraction accuracy ≥96%, component matching rate ≥98%, and script generation time ≤3 seconds, meeting the real-time requirements of industrial scenarios. Based on operational feedback, the ontology graph semantic relationships and microservice interface constraints will be iteratively optimized to improve planning accuracy.

[0109] It should be noted that embodiments of the present invention can be implemented in hardware, software, or a combination of both. The hardware portion can be implemented using dedicated logic; the software portion can be stored in memory and executed by a suitable instruction execution system, such as a microprocessor or dedicated-design hardware. Those skilled in the art will understand that the above-described devices and methods can be implemented using computer-executable instructions and / or included in processor control code, for example, such code provided on a carrier medium such as a disk, CD, or DVD-ROM, a programmable memory such as read-only memory (firmware), or a data carrier such as an optical or electronic signal carrier. The devices and modules of the present invention can be implemented by hardware circuitry such as very large-scale integrated circuits or gate arrays, semiconductors such as logic chips, transistors, or programmable hardware devices such as field-programmable gate arrays, programmable logic devices, etc., or by software executed by various types of processors, or by a combination of the above-described hardware circuitry and software, such as firmware.

[0110] The above description is merely a specific embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any modifications, equivalent substitutions, and improvements made by those skilled in the art within the scope of the technology disclosed in the present invention, and within the spirit and principles of the present invention, should be covered within the scope of protection of the present invention.

Claims

1. An intelligent planning method for handling industrial anomalies, characterized in that, Includes the following steps: The system receives production intentions for industrial anomalies through a pre-set intelligent agent module, which then extracts task objectives and constraints based on natural language processing technology. The production intention is semantically aligned using a pre-built industrial ontology graph to determine the equipment entities, process attributes, and relationship topology between entities involved in the task. A structured task semantic description is generated based on the semantic alignment result. The task semantic description includes data objects, processing targets, and control logic. Call the standard interface declarations of each microservice component in the capability library, wherein the standard interface declarations include input constraints and output constraints; Under the semantic constraints of the industrial ontology graph, the microservice components are automatically filtered and combined to ensure that the output data type of the upstream components is consistent with the input constraints of the downstream components. Generate structured atomic application chain scripts that conform to the preset workflow syntax.

2. The method according to claim 1, characterized in that, The industrial ontology map includes an equipment entity layer, a process parameter layer, and an anomaly event layer, with each layer linked by predefined semantic relationships.

3. The method according to claim 1, characterized in that, The structured task semantic description represents data acquisition, feature processing, model detection, alarm generation, and process optimization as a unified task semantic unit. Each task semantic unit includes an input field, an output field, and an execution rule field.

4. The method according to claim 1, characterized in that, The microservice components are atomic application units, each corresponding to a single processing function, and are connected through explicit data input and output fields.

5. The method according to claim 1, characterized in that, The structured atomic application chain script defines task node identifiers, input data sources, output data targets, and data flow relationships between nodes in text form.

6. A workflow-driven execution method for industrial anomaly handling, characterized in that, Includes the following steps: Receive an atomic application chain script generated by the method of any one of claims 1 to 5; Parse the atomic application chain script to extract each atomic application unit and its data input / output relationships; Construct a directed acyclic graph data dependency topology based on the data input-output relationship; Perform state monitoring on the nodes in the directed acyclic graph; Under the condition of satisfying data dependency constraints, nodes without prior dependencies are scheduled in parallel. For nodes with pre-dependent nodes, sequential scheduling is performed after the dependent nodes have completed execution; Complete the industrial anomaly detection and handling process.

7. The method according to claim 6, characterized in that, In the directed acyclic graph, nodes represent atomic application units, edges represent data flow dependencies, and the execution condition of any node is that the execution status of all its predecessor nodes is complete.

8. The method according to claim 6, characterized in that, The status monitoring includes marking the pending, executing, completed, and failed states of a node.

9. The method according to claim 6, characterized in that, When a node fails to execute, retry scheduling or stopping the execution of subsequent nodes is performed based on the data dependencies of the directed acyclic graph.

10. The method according to claim 6, characterized in that, After the industrial anomaly handling process is completed, the detection results and manual review results are received, and the results are used as labeled samples to supplement the training set, triggering adaptive fine-tuning or retraining of the parameters of the underlying anomaly detection model, and updating the intelligent agent's screening rules for microservice components.

Images