Equipment fault diagnosis dynamic optimization method and system based on smart power plant

Abstract

The invention discloses an equipment fault diagnosis dynamic optimization method and system based on an intelligent power plant, and relates to the technical field of intelligent operation and maintenance of the intelligent power plant. Standardized acquisition and feature construction of multi-source operation data are carried out through unified coding, time synchronization and a semantic tag system; a mechanism model and AI model dual-channel fusion diagnosis engine is constructed, and parallel reasoning and structured diagnosis of multi-source data are output through residual calculation and consistency analysis; a dynamic optimization model containing three elements of energy efficiency, time delay and risk is constructed based on structured diagnosis, and operation parameters are reconfigured in real time through constraint solution and a control quantity mapping mechanism. According to the method, multi-source data fusion diagnosis and dynamic optimization linkage is achieved, the real-time analysis, self-adaptive optimization and continuous learning capabilities are achieved, and the intelligent level of equipment operation is improved.

Description

Technical Field

[0001] This invention relates to the field of intelligent operation and maintenance technology for smart power plants, specifically to a dynamic optimization method and system for equipment fault diagnosis in smart power plants. Background Technology

[0002] With the transformation of the energy structure and the continuous improvement of the digitalization level of the power system, smart power plants have become an important direction for modern power generation enterprises to achieve intensive management and efficient operation and maintenance. In recent years, technologies such as the Internet of Things, big data, cloud computing, and artificial intelligence have been gradually applied to the field of generator unit operation monitoring and fault diagnosis, forming an intelligent system centered on multi-source data acquisition, online monitoring and analysis, and remote operation and maintenance. However, most existing systems focus on static feature analysis and single-model judgment, making it difficult to achieve real-time multi-source data fusion and dynamic optimization decision-making. Against this backdrop, how to achieve equipment status diagnosis and dynamic optimization of operating parameters based on the smart power plant architecture has become an important research direction in the field of intelligent operation and maintenance.

[0003] Existing smart power plant equipment operation and maintenance methods generally rely on single monitoring systems or static algorithm models, lacking the ability to uniformly model multi-source heterogeneous data. Differences in data structure, timestamps, and semantics between different systems lead to inconsistencies in feature extraction and state assessment. Furthermore, while traditional mechanistic models have physical constraints, they struggle to cope with multidimensional nonlinear changes under complex operating conditions; while pure AI models, although possessing self-learning capabilities, are easily affected by data noise and sample bias, failing to form interpretable fault mechanism reasoning. At the operational optimization level, existing research typically uses offline computation methods to analyze energy efficiency and control parameters, lacking a dynamic feedback mechanism based on real-time diagnostic results. This prevents automatic adjustment of operating conditions or optimization of control quantities during operation, resulting in problems such as high energy consumption and lag in regulation. Simultaneously, mainstream systems generally lack continuous learning mechanisms based on log data, leading to long-term stagnation of model parameters and a decline in diagnostic accuracy and optimization performance over time. Summary of the Invention

[0004] In view of the above-mentioned problems, the present invention is proposed.

[0005] Therefore, the technical problem solved by this invention is: existing methods for equipment diagnosis and operation optimization in smart power plants suffer from inaccurate feature extraction due to inconsistent multi-source data structures, low model fusion degree making real-time diagnosis difficult, and lagging optimization of operating parameters lacking a dynamic optimization mechanism. The invention also addresses the problem of how to achieve adaptive dynamic optimization of equipment fault diagnosis and operating conditions under a multi-model fusion architecture.

[0006] To address the aforementioned technical problems, this invention provides the following technical solution: a dynamic optimization method for equipment fault diagnosis in smart power plants, comprising constructing a three-tiered structure of a field acquisition layer, an edge computing layer, and a data center layer; standardizing the acquisition and feature construction of multi-source operational data through a unified coding, time synchronization, and semantic tagging system; after feature construction, constructing a dual-channel fusion diagnostic engine of a mechanism model and an AI model; utilizing residual calculation and consistency analysis to output parallel inference and structured diagnosis of multi-source data; constructing a dynamic optimization model based on structured diagnosis, incorporating three elements: energy efficiency, latency, and risk; reconfiguring operational parameters in real time through constraint solving and control quantity mapping mechanisms; and constructing a sample management, incremental training, and parameter recalibration mechanism using log data output from the dynamic optimization model and structured diagnosis.

[0007] As a preferred embodiment of the dynamic optimization method for equipment fault diagnosis based on smart power plants described in this invention, the three-tier structure of the field acquisition layer, edge computing layer, and data center layer includes: the field acquisition layer interacting with DCS, BOP, and NCS systems via OPCUA, Modbus TCP, or IEC 61850 protocols; the edge computing layer performing data cleaning, timing alignment, feature extraction, and wavelet compression processing on the raw signals; and the data center layer adopting a hierarchical storage structure of real-time database, historical database, and relational database, with centralized management of feature data in a unified JSON Schema format.

[0008] As a preferred embodiment of the dynamic optimization method for equipment fault diagnosis based on smart power plants described in this invention, the dual-channel fusion diagnostic engine for constructing a mechanism model and an AI model includes a mechanism modeling layer, an AI modeling layer, a fusion inference layer, and a model management layer; the mechanism modeling layer establishes a mathematical model based on the equipment thermodynamic equation, energy balance equation, and structural characteristics; the AI ​​modeling layer trains feature data samples based on multiple algorithms; and the fusion inference layer calculates residuals.

[0009] As a preferred embodiment of the dynamic optimization method for equipment fault diagnosis based on smart power plants described in this invention, the dynamic optimization model includes a multi-objective optimization model, represented as follows: ; in, Indicates energy efficiency loss. Indicates maintenance delay. Indicates risk factors, , , The weight coefficients are dynamically adjusted based on the system's operating status. The optimization process employs a combination of the Lagrange multiplier method and Newton's iteration to solve for constraints, and the optimization results are converted into setpoints for combustion, fan, and pump valve control quantities based on a mapping table.

[0010] As a preferred embodiment of the dynamic optimization method for equipment fault diagnosis based on smart power plants described in this invention, the dynamic optimization model includes: after receiving the equipment operating status information output by the diagnostic engine, the dynamic optimization model obtains the dynamic fluctuation values ​​of load change, temperature deviation, and energy efficiency indicators through a real-time data feedback module; based on the dynamic fluctuation values, it constructs a multi-parameter optimization iterative equation including operating energy consumption, equipment load, and operating condition stability, and performs rolling updates within a continuous time window; each round of optimization iteration corrects the change step size of the control variables according to real-time data, and performs synchronous verification with the mechanism model through a feedback channel.

[0011] As a preferred embodiment of the dynamic optimization method for equipment fault diagnosis based on smart power plants described in this invention, the real-time reconfiguration of operating parameters through constraint solving and control quantity mapping mechanism includes: during the execution of the optimization control quantity, smoothing of optimization instructions is performed through a two-layer buffer control structure consisting of a logic buffer layer and a time buffer layer; the logic buffer layer automatically compares the differences between the set values ​​before and after optimization, and when the change in control quantity exceeds a threshold, the issuance of instructions is paused and manual confirmation is requested; the time buffer layer adjusts the control parameters step by step using a linear recursive update method in the range of 5 to 15 seconds.

[0012] As a preferred embodiment of the dynamic optimization method for equipment fault diagnosis based on smart power plants described in this invention, the mechanism for constructing sample management, incremental training and parameter recalibration includes sample data extraction, label binding, sample cleaning and sample balancing steps.

[0013] Sample labels are generated from confirmation records of the inspection system or maintenance work order system. Training samples use incremental learning to update model parameters, and the parameter update is expressed as follows: ; in, express Time-model parameters, express Time-model parameters, This represents the gradient update amount during training on new samples. This represents the migration coefficient.

[0014] Another objective of this invention is to provide a dynamic optimization system for equipment fault diagnosis in smart power plants. This system can construct a dynamic optimization model that includes three elements: energy efficiency, time delay, and risk, based on structured diagnosis. It can also reconfigure operating parameters in real time through constraint solving and control quantity mapping mechanisms. This solves the problems of low model fusion, difficulty in achieving real-time diagnosis, and lag in operating parameter optimization, as well as the lack of dynamic optimization mechanisms in current smart power plant equipment diagnosis and operation optimization methods.

[0015] As a preferred embodiment of the dynamic optimization system for equipment fault diagnosis based on a smart power plant according to the present invention, it includes: a multi-source data acquisition and feature construction module, a mechanism and AI fusion diagnostic modeling module, a dynamic optimization decision and operating condition reconfiguration module, and an online self-learning and model adaptive update module; the multi-source data acquisition and feature construction module is used to unify the structure of the field acquisition layer, edge computing layer, and data center layer, and to perform standardized acquisition, time synchronization, and feature vector generation of multi-source operating data from DCS, BOP, NCS, and sensors; the mechanism and AI fusion diagnostic modeling module is used to establish a dual-channel diagnostic engine of mechanism model and AI model, and to realize equipment operating status identification, anomaly detection, and structured diagnostic output through residual calculation and consistency analysis; the dynamic optimization decision and operating condition reconfiguration module is used to construct a dynamic optimization model, to solve the constraints by comprehensively considering energy efficiency, time delay, and risk factors, and to map the optimal control parameters to the control system setpoint to achieve real-time operating condition adjustment; the online self-learning and model adaptive update module is used to perform sample management, incremental training, and parameter recalibration.

[0016] Another objective of this invention is to provide a dynamic optimization device for equipment fault diagnosis based on a smart power plant, comprising a memory and a processor, wherein the memory stores a computer program, and the processor executes the computer program to implement a dynamic optimization method for equipment fault diagnosis based on a smart power plant.

[0017] Another object of the present invention is to provide a dynamic optimization storage medium for equipment fault diagnosis based on a smart power plant, wherein a computer program is stored thereon, and when the computer program is executed by a processor, the steps of the dynamic optimization method for equipment fault diagnosis based on a smart power plant are implemented.

[0018] The beneficial effects of this invention are: The dynamic optimization method for equipment fault diagnosis based on smart power plants provided by this invention constructs a three-tiered structure of field acquisition layer, edge computing layer, and data center layer to achieve standardized acquisition and unified feature construction of multi-source heterogeneous data in power plants, laying a high-quality foundation for model input. Then, through a diagnostic engine that integrates mechanistic and AI models, combined with residual calculation and consistency analysis, it achieves collaborative reasoning based on physical constraints and data-driven approaches, enabling accurate identification of abnormal states under complex operating conditions. On this basis, a dynamic optimization model incorporating three elements—energy efficiency, latency, and risk—is established. Based on the unit's operating boundaries and logical constraints, control quantities are solved and parameters are mapped in real time, forming a diagnosis-control closed loop. Finally, using diagnostic and optimization log data, a sample management, incremental training, and parameter recalibration mechanism is constructed, enabling the model to possess self-learning and adaptive capabilities.

[0019] By organically combining the above steps, a fully integrated design of data acquisition, diagnostic analysis, optimization decision-making, and self-evolutionary updates is achieved. A dynamic optimization system for fault diagnosis of smart power plant equipment is constructed from three levels: system architecture, algorithm collaboration, and intelligent optimization, which significantly improves the system's real-time performance, accuracy, and self-evolutionary capabilities. Attached Figure Description

[0020] To more clearly illustrate the technical solutions of the embodiments of the present invention, the accompanying drawings used in the description of the embodiments will be briefly introduced below. Obviously, the accompanying drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0021] Figure 1 This is an overall flowchart of a dynamic optimization method for equipment fault diagnosis based on a smart power plant, provided in Embodiment 1 of the present invention. Detailed Implementation

[0022] To make the above-mentioned objects, features, and advantages of the present invention more apparent and understandable, specific embodiments of the present invention will be described in detail below with reference to the accompanying drawings. Obviously, the described embodiments are only a part of the embodiments of the present invention, and not all of them. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort should fall within the protection scope of the present invention.

[0023] Example 1, referring to Figure 1 As an embodiment of the present invention, a dynamic optimization method for equipment fault diagnosis based on a smart power plant is provided, comprising the following steps: S1: Construct a three-tiered structure consisting of a field acquisition layer, an edge computing layer, and a data center layer. Standardize the acquisition and feature construction of multi-source operational data through a unified coding, time synchronization, and semantic tagging system.

[0024] Furthermore, the field data acquisition layer primarily uses the DCS system, BOP control system, NCS protection system, vibration monitoring device, infrared temperature measurement device, and condition-based maintenance device of the generator set and auxiliary equipment as data sources. Signal points from each system are connected to the industrial IoT gateway using standardized communication protocols such as OPC UA, Modbus TCP, and IEC 61850.

[0025] The types of signals collected include, but are not limited to: pressure, temperature, flow rate, current, voltage, power, speed, vibration, displacement, valve opening, oil temperature, oil pressure, and electrical status signals. Each signal has a sampling period, engineering range, data type, and unit attributes, and undergoes a one-time verification before access. To improve the timeliness and reliability of data acquisition, the field-level nodes uniformly adopt a dual-port redundant design, with automatic switching between primary and backup channels via link aggregation. The sampling frequency is set according to the equipment category; for example, the sampling period for high-speed rotating equipment (steam turbines, coal mills, etc.) is set to 1s–2s, while the sampling period for low-speed fluid equipment (circulating water pumps, cooling water pumps, etc.) is set to 5s–10s.

[0026] In addition, a lightweight data buffer unit is deployed at the field acquisition layer. When the network is interrupted or the main station storage is delayed, the acquisition module automatically switches to local caching mode and transmits the cached data in batches after the link is restored to ensure data continuity and integrity. All acquisition units support NTP network time synchronization and GPS / BeiDou dual-mode time synchronization to ensure that the time error across the entire plant is controlled within ±50ms.

[0027] It should be noted that the edge computing layer is located between the field data aggregation switch and the data center, and is mainly responsible for structuring, filtering, compressing, and performing preliminary feature extraction on the raw data. Each edge node runs an independent data preprocessing service process, which includes the following four core modules: The data cleaning module performs operations such as format standardization, null value imputation, mutation value removal, and duplicate sample elimination on the collected raw data. Mutation detection employs a combination of the three-σ criterion and a moving window mid-range smoothing algorithm to distinguish between real-world fluctuations and sampling anomalies.

[0028] Timing alignment module: For asynchronous data streams from different systems, it performs unified alignment through timestamp resampling and linear interpolation to ensure the consistency of data frames on the timeline.

[0029] Feature extraction module: Automatically calls the corresponding signal feature template based on the equipment type. For example, for rotating machinery, it extracts the root mean square vibration value, kurtosis factor, envelope spectrum energy, and frequency domain peak features; for thermal equipment, it extracts the temperature rise rate, pressure difference coefficient, and flow fluctuation factor; and for electrical equipment, it extracts current harmonic components, phase difference, and power factor change rate, etc.

[0030] Data compression module: Piecewise differential and wavelet compression algorithms are used for high-frequency signals to ensure that the bandwidth occupancy rate is less than 60% during real-time transmission, while maintaining the integrity of key frequency components.

[0031] After completing the above processing, the edge nodes send the structured data to the central data bus in a unified JSON Schema format. To ensure security, each data packet is accompanied by an encrypted signature and sequence number to prevent duplication and tampering.

[0032] It should also be noted that the data center layer receives standardized data streams from various edge nodes and uses a three-tiered storage architecture of real-time database, historical database, and relational database from the smart power plant platform for hierarchical storage. The real-time database (such as PI, TDengine, etc.) is used to store high-frequency operating data within 24 hours, supporting millisecond-level read / write and high-speed query; the historical database adopts a partitioned storage strategy, managed by two-level indexes of time and device, for subsequent model training calls; the relational database stores various types of non-time-series data, including device static attribute tables, point information tables, calibration parameter tables, and data quality identification tables.

[0033] All databases are connected to the data platform interface via a data access bus, forming a data access system with a single entry point. The system employs master-slave replication and hot standby mechanisms to ensure that data readability is not affected by the failure of any node.

[0034] To accommodate the computational needs of different models, the data center layer also defines rules for generating feature datasets. These rules include data window length, sliding step size, sampling alignment, and feature normalization method. An automated task scheduling system periodically executes feature set update tasks, storing the feature results in the model input table to provide an input data source for subsequent diagnostic model calls.

[0035] Furthermore, to achieve cross-system and multi-disciplinary data fusion, the system establishes a unified data semantic tagging system. Each acquired signal is automatically assigned five tags upon entry into the database: Equipment ID, Signal Type, Location, Unit, and Data Level. The Equipment ID follows a plant-level unique naming convention, consisting of a four-segment code: "Unit Number—System Number—Equipment Number—Measuring Point Number." Signal types are categorized according to dimensions such as thermal control, electrical, mechanical, and environmental. Data levels are divided into basic, analytical, and model levels based on sampling accuracy and application. This semantic tagging system is maintained in the platform's data dictionary and is available for various subsystems to access via API, enabling consistent semantic mapping and automatic recognition of data from different sources. This provides a standardized input interface for subsequent mechanism modeling and AI algorithm integration.

[0036] It should be noted that during data acquisition and transmission, the system strictly adheres to the security zoning and access control requirements of smart power plants. All acquired data is transmitted via a dedicated industrial control network, and edge nodes only open ports authorized by the firewall whitelist. The transmission protocol uses a TLS encrypted channel, and both communicating parties use certificates for two-way authentication. The data center implements hierarchical authorization through Access Control Lists (ACLs) and Role-Based Access Control (RBAC) mechanisms, allowing different levels of operations, analysis, and development personnel to access corresponding data tables or views based on their role permissions. The log audit module records each data access behavior, including access time, user identity, access type, and data range. All audit logs are periodically archived and written to the security log database.

[0037] It should also be noted that the feature data is provided via both a RESTful API and a Kafka message queue. For batch training tasks, the API interface supports querying feature data in JSON or Parquet format by time range and device ID.

[0038] For real-time diagnostic tasks, Kafka queues push incremental feature streams at the Topic level, ensuring that the model diagnostic module can obtain the latest input with low latency. The response time of the output interface must be less than 2 seconds, and data consistency is achieved by comparing check hashes with sequence numbers.

[0039] S2: After feature construction, a dual-channel fusion diagnostic engine combining mechanism model and AI model is built, which uses residual calculation and consistency analysis to output parallel inference and structured diagnosis of multi-source data.

[0040] Furthermore, the mechanism modeling layer is based on the structural characteristics, thermodynamic relationships, and energy balance equations of the equipment, and establishes corresponding mathematical models according to the plant-level equipment standard parameter tables and operating procedures.

[0041] Taking a steam turbine system as an example, its model includes steam flow balance, enthalpy difference equation, efficiency correction curve, and back pressure constraint equation. For pump equipment, the core is the flow-head characteristic curve, power input-output model, and temperature rise model. For motor equipment, the model mainly focuses on the mapping relationship between current, power factor, temperature rise, and speed. The model parameters are derived from equipment factory data, historical operating calibration values, and the feature dataset generation rules formed in stage S1, including data window length, sliding step size, sampling alignment method, and feature normalization method. By directly calling these rules, the mechanistic model can achieve dynamic mapping between parameters and real-time data.

[0042] The mechanistic models are developed within the platform's visual modeling environment. The platform includes built-in symbolic computation libraries, matrix solver libraries, signal processing libraries, and a collection of commonly used engineering functions, supporting a streamlined drag-and-drop modeling process. Each model node is connected to its corresponding data label via an input port, such as device identifier, signal type, and measurement point location, ensuring data semantic consistency within the model. All compiled mechanistic models are stored in the model library as intermediate representation files (JSON or XML structure), which can be accessed as needed by the fusion inference layer.

[0043] It should be noted that the AI ​​modeling layer relies on the smart power plant's artificial intelligence modeling platform. By learning from the data output by the feature extraction module, it establishes classification and prediction models for different equipment types. The AI ​​modeling platform supports multiple algorithm frameworks, including but not limited to support vector machines, random forests, isolated forests, multilayer perceptrons, LightGBM, and decision tree regression models.

[0044] The data input interface calls the structured data output from the edge computing layer and sends it to the central data bus data channel in a unified JSON Schema format. It utilizes Kafka real-time message streams to subscribe to the latest feature samples, enabling online model input. Training samples are extracted from historical datasets managed by a two-level index of time and device in a historical database. Label data is automatically generated from fault records and maintenance logs in the work order system. The AI ​​model training process is executed in a containerized environment, with each training task running as an independent container instance, ensuring isolation of computing resources and task traceability. Model parameter files, training logs, and evaluation results are stored in a model metadata table in a relational database, with fields including model ID, version number, input dimension, algorithm type, hyperparameter set, and performance metrics.

[0045] To maintain interface compatibility with the mechanistic model, the AI ​​model's input and output both follow a unified interface specification: input consists of device identifier + timestamp + feature vector, and output consists of anomaly score + category label + confidence level. This specification achieves automatic matching at the fusion inference layer through dynamic binding.

[0046] It should also be noted that the fusion inference layer is the core module where the mechanistic model and the AI ​​model run together, employing a fusion strategy of dual-channel parallel computing and residual consistency analysis. Within the same time window, the system separately calls the mechanistic model and the AI ​​model to calculate the theoretical output values ​​and actual predicted values ​​of the equipment. The mechanistic channel outputs physical quantity results, such as expected temperature, pressure, or vibration amplitude; the AI ​​channel outputs anomaly probabilities or state classification results. The fusion inference layer performs residual calculations on both: ; in, Represents the residual signal. This represents the real-time observations in the standardized data stream received by the data center layer. These are the predicted values ​​from the mechanistic model.

[0047] After the residual signal undergoes sliding window filtering and threshold normalization, it enters the multi-dimensional consistency analysis module. If the statistical characteristics of the residual and the anomaly confidence level output by the AI ​​model both exceed the set threshold within the same time window, the system generates a structured diagnostic event, which includes fields such as device ID, fault mode ID, trigger time, anomaly factor, and model version number, and stores it in the diagnostic event database.

[0048] This layer also implements a model switching mechanism. When the difference between the results of the AI ​​model and the mechanistic model continues to exceed a set threshold, the system triggers a model consistency check task, reloading the latest version of the parameter file from the model library. This mechanism ensures diagnostic accuracy and model stability during long-term operation.

[0049] Furthermore, the model management layer is used to uniformly manage the registration, scheduling, monitoring, and version maintenance of mechanistic models and AI models. The system automatically generates a unique model identifier during model deployment and establishes an index relationship in the database. All model files and their corresponding data source mappings are recorded in the model-data mapping table, with fields including: model type, input point table, output variable name, computation period, priority, and dependent model ID.

[0050] The task scheduling module employs a distributed scheduling framework to automate model computations. Computation priorities are assigned based on the importance of the equipment: the computation cycle for the main machine and critical auxiliary machines is set to 5 seconds, while for ordinary equipment it is 30 seconds. The scheduling module periodically triggers model execution instances according to the scheduling table and monitors the computation status, execution duration, and output integrity. When a model fails to run or outputs abnormal information, the system automatically switches to a backup node and generates an alarm record.

[0051] In addition, the model management layer provides an interface for upper-layer applications to access diagnostic results. The interface format is consistent with the defined feature data output interface design, using both RESTful API and Kafka message queues to ensure that the output of the diagnostic engine can seamlessly enter the subsequent optimization module.

[0052] It should be noted that during the integration of the mechanism and the AI ​​model, the correspondence between the established data semantic labeling system is continuously maintained. Each model input and output data carries complete label information, including device identifier, signal type, measurement point location, unit, and data level. Model calculation logs and intermediate results are recorded in the "Diagnostic Log Table" of the relational database, with fields including input time range, data source number, residual statistics, model version, and execution node. All logs are periodically summarized and compared with each data access behavior recorded by the log audit module to form a complete traceable link, ensuring the consistency and security of model input, output, and data access.

[0053] It should also be noted that the diagnostic engine is deployed in a container cluster within the smart power plant data center. Each model instance runs as an independent container, and the containers communicate with each other through a service mesh. Container images are managed uniformly, with version control and image signing performed by the DevOps platform to ensure model environment consistency. Internally, the engine uses a shared memory cache to store the input data from the most recent time window, enabling high-speed parallel computation between the mechanistic model and the AI ​​model. The three processes of data reading, computation, and result writing are completed within the same pipeline, with an average latency of no more than 2 seconds, meeting the system requirements for real-time diagnostics.

[0054] S3: Based on structured diagnostics, a dynamic optimization model is constructed that includes three elements: energy efficiency, time delay, and risk. The operating parameters are reconfigured in real time through constraint solving and control quantity mapping mechanisms.

[0055] Furthermore, the dynamic optimization module and the diagnostic engine are deployed in the same container cluster, employing a microservice architecture. Internally, the module consists of four core sub-modules: an optimization data preparation unit, an objective function construction unit, a constraint solving unit, and a control variable adjustment unit. These sub-modules interact via a shared memory channel, achieving communication latency of less than 2 seconds. The input interface directly calls the diagnostic event database and real-time data stream: the residual signal, after sliding window filtering and threshold normalization, enters the multi-dimensional consistency analysis module. If the statistical characteristics of the residual and the anomaly confidence level output by the AI ​​model both exceed a set threshold within the same time window, the system generates a structured diagnostic event, including fields such as device ID, fault mode ID, trigger time, anomaly factor, and model version number, and stores it in the diagnostic event database.

[0056] After acquiring the aforementioned diagnostic events, the dynamic optimization module automatically reads the real-time operating parameters, control loop setpoints, and historical performance curves of the corresponding devices, and establishes a time synchronization mapping with the standardized data stream received by the data center layer. Before entering the optimization engine, all data streams are processed uniformly through the "time alignment module" to ensure the consistency of the input data frames on the time axis.

[0057] It should be noted that the data preparation unit is used to construct the input vectors and state variables required for the optimization calculation. This unit periodically (5s to 30s) extracts temperature, pressure, flow rate, rotational speed, current, voltage, power, and vibration signals from the real-time database and historical database.

[0058] The data extraction process follows the rules for generating feature datasets, including data window length, sliding step size, sampling alignment, and feature normalization method. An automated task scheduling system periodically executes feature set update tasks, storing the feature results in the model input table to provide an input data source for subsequent diagnostic model calls.

[0059] The optimal input vector is defined as: ; Each of them This corresponds to a specific characteristic parameter (such as main steam temperature, feedwater flow rate, air supply volume, motor current, etc.). Furthermore, the optimized data preparation unit automatically loads the fault mode codes output by the diagnostic engine. With confidence level This forms the state extension vector: ; As a real-time input to the dynamic optimization model.

[0060] The objective function construction unit uses energy efficiency loss, maintenance delay, and risk assessment as core indicators to define a dynamic optimization objective function: ; in, Indicates energy efficiency loss. Indicates maintenance delay. Indicates risk factors, , , This represents the weighting coefficient, which is dynamically adjusted based on the system's operating status.

[0061] It should also be noted that the constraint solving unit is used to define and maintain the equipment operating boundaries, ensuring that safety limits are not exceeded during optimization adjustments. The constraint set includes: physical constraints: determining the upper and lower limits of operating variables based on the equipment characteristic curves (flow rate-head, temperature-efficiency, load-flue gas temperature, etc.) output by the mechanistic model; safety constraints: boundary values ​​from equipment operating procedures and control and protection logic, such as the upper limit of boiler main steam pressure and the lower limit of cooling water pump outlet pressure; logical constraints: ensuring logical consistency between control quantities, such as the ratio of air supply volume to primary air volume and the matching relationship between fuel and air coefficients; real-time constraints: taken from the status lock flags of abnormal equipment in the diagnostic results. If the fault mode is severe, the corresponding equipment control variables enter a frozen state and do not participate in optimization adjustments. Constraints are input to the solver in matrix form.

[0062] The control quantity adjustment unit is responsible for mapping the optimization calculation results into executable control commands. The mapping relationship is predefined in the configuration file. For example, for a boiler system, the adjustment objects include fuel quantity setting, blower damper opening, and primary air to secondary air ratio; for a steam turbine system, the adjustment objects include feedwater flow rate, extraction steam valve position, and condensate pump speed; for motors and pumps, the adjustment objects include inverter frequency, excitation current, and operating mode switching. The optimization results are sent to the corresponding DCS substation via the OPC UA interface or the control system's open API. To ensure execution safety, the control quantity adjustment module uses a two-layer buffer mechanism: a logic buffer layer, which compares the differences in setpoints before and after optimization; when the change exceeds a set threshold, manual confirmation is required before execution; and a time buffer layer, which uses a smooth, recursive update method of 5 to 15 seconds to prevent sudden changes in control quantities. The control execution status is collected in real-time via DCS feedback and re-enters the field acquisition layer data channel to achieve closed-loop tracking.

[0063] The optimization engine's operation is uniformly controlled by the scheduling module. Each optimization cycle includes the following processes: Status Update: Retrieve feature data and diagnostic events of the nearest window from the real-time database; Objective Calculation: Calculate the objective function components based on the current load and risk level; Constraint Solving: Load the safety boundary and execute optimization iterations; Control Issuance: Generate control variable adjustment instructions and write them to the DCS interface; Log Storage: Record each optimization result, calculation duration, input status, and output control variable in the optimization log table. Fields include: Task ID, Start Time, Iteration Count, F-value, Change in Control Variable, and Execution Node Number.

[0064] All log and diagnostic log tables share a primary key index, which facilitates subsequent joint analysis and model backtracking.

[0065] The dynamic optimization module maintains data synchronization with the smart power plant energy consumption analysis system and operation guidance system. Optimization results are published through a Kafka message queue designed with a feature data output interface, using the Optimization / DeviceID / Time naming convention for topics. After subscribing to this topic, the energy consumption analysis system can use the optimization results for operational indicator assessment and performance statistics. The optimization module also interfaces with the safety production system; when the risk level exceeds a set threshold, it automatically triggers safety constraint updates, prohibiting further operational adjustments.

[0066] The optimization module is deployed in a containerized manner, sharing compute nodes with the diagnostic engine. Each optimization instance runs as an independent container, with CPU and memory resources allocated according to priority within the container orchestration system. The container image pre-installs a mathematical optimization library and provides a REST API for upper-layer scheduling. Data exchange is completed via message queues, and optimization results are written to an in-memory database cache for real-time reading by the optimization visualization interface. The container's running status is monitored by the cluster monitoring system. If CPU utilization exceeds a set threshold or a response times out, a migration mechanism is triggered, rescheduling the optimization task to a standby node.

[0067] S4: Construct a sample management, incremental training, and parameter recalibration mechanism using dynamic optimization models and log data output from structured diagnostics.

[0068] Furthermore, the online self-learning and model adaptive update module is deployed in the smart power plant's big data center, establishing a bidirectional data channel with the mechanism model library, AI model library, diagnostic event database, and optimization log table. Internally, the module consists of four core subsystems: a sample data management unit, a model retraining unit, a parameter recalibration unit, and a version management and rollback unit. Its data input sources include: standardized data streams received from the data center layer, providing raw operational characteristic data; diagnostic log tables, used to obtain fault modes, model versions, and residual statistics; optimization log tables, used to load optimization target values, constraint boundaries, and changes in control variables; and maintenance work order systems and mobile inspection terminals, used to read manual confirmation results and maintenance measure records.

[0069] The aforementioned multi-source data are aggregated through a unified ETL process to form a self-learning sample library. ETL tasks are executed in batch mode, with a scheduling cycle of once at 0:00 and once at 12:00 daily to ensure that the model update frequency matches the data accumulation rate.

[0070] It should be noted that the main task of the sample data management unit is to construct a sample set that can be used for model retraining from the aforementioned multi-source data. This unit first extracts the input feature vectors for the corresponding time window from the real-time database based on the device ID and trigger time in the diagnostic event database. The window length and step size are determined by the feature dataset generation rules, which include the data window length, sliding step size, sampling alignment, and feature normalization method. Through an automated task scheduling system, the feature set update task is periodically executed, storing the feature results in the model input table to provide an input data source for subsequent diagnostic model calls. Next, the sample data management unit performs a label binding operation. If the inspection or maintenance system confirms a device malfunction, the label is set to 1; if it is confirmed as a false alarm, it is set to 0. After sample generation, this unit performs a data cleaning process, including missing value imputation (linear interpolation), outlier removal (three-σ criterion), and sample balancing (SMOTE algorithm). Finally, the samples are stored in the sample library according to a four-dimensional index structure of device ID—time—feature dimension—label, facilitating direct access by the AI ​​model layer.

[0071] It should also be noted that the model retraining unit is responsible for periodically triggering the retraining of the AI ​​model and correcting the mechanistic model parameters. AI model retraining is implemented based on a containerized environment, with each training task running as an independent container instance. The scheduling system reads sample data from the sample library for the past N days (N is configurable, defaulting to 30 days), dividing it into training, validation, and test sets in a 7:2:1 ratio. The model training process uses incremental learning: updating the model weights through a gradient transfer mechanism without discarding old parameters. The specific formula is as follows: ; in, express Time-model parameters, express Time-model parameters, This represents the gradient update amount during training on new samples. This represents the migration coefficient.

[0072] For the mechanistic model, the system automatically executes the parameter calibration process based on the newly collected operational data. The parameter calibration module uses the least squares method to refit the energy balance equation, flow rate equation, and thermal efficiency curve. After the calibration parameter file is generated, it is automatically uploaded to the model management layer, and the parameter update time and calibration sample range fields in the model metadata table are updated.

[0073] After training is complete, the model retraining unit will output a training report file containing metrics such as model version number, training start and end time, sample size, validation accuracy, and loss function value, and store it in the model training report table of the relational database.

[0074] The parameter recalibration unit's task is to dynamically correct key parameters through online computation while keeping the model version unchanged. This unit receives residual sequences from the diagnostic engine in real time. The residual trend is calculated using the exponentially weighted moving average (EWMA) method. ; in, Indicates the residual trend quantity. This represents the instantaneous residual signal from the previous time step. For smoothing coefficients, Represents the actual observed value. This represents the model's output value.

[0075] like If the parameters consistently exceed a set threshold, a parameter recalibration process is triggered. The recalibration algorithm uses an adaptive Kalman filter to update the model's state variables and outputs a corrected parameter set. Once the new parameters have been verified, they are written to the model parameter file and marked as the online calibration version. This file is then incrementally committed to the Git repository, recording the version number, commit time, number of modified parameters, and scope of the changes.

[0076] Version management and rollback units ensure the security and traceability of model updates. The system employs a Git-based distributed version control mechanism, with all model files (including AI model parameter files, mechanistic model equation files, and fusion configuration files) stored in a central repository. Each model update generates a unique version number and registers it in the database model version index table. When a new model is deployed, the system automatically initiates an A / B validation process: under the same data input conditions, the old and new versions of the model are run simultaneously, comparing the mean residual, anomaly detection rate, and computation latency. If the new version's performance metrics are all superior to the old version, the system upgrades it to the production version; otherwise, it automatically rolls back to the previous stable version. The rollback process is completed via hot-swapping in the container orchestration system, with a switching time of less than 5 seconds, ensuring uninterrupted operation of the model service.

[0077] The continuous learning mechanism automatically maintains model update tasks through a scheduling system. The task scheduler generates a scheduling table based on the model type and update cycle, which mainly includes: AI diagnostic model: daily checks on the number of new samples, and automatically triggers training tasks when the number of samples exceeds a threshold; Mechanistic model: monthly parameter recalibration tasks; Fusion model: quarterly overall performance evaluation tasks.

[0078] The running status of all tasks is recorded in the model scheduling log table, with fields including task name, start time, execution node, runtime, status code, and output report path. The scheduling system automatically generates an alarm when a task is abnormally interrupted or fails, sends it to the maintenance console via the message bus, and records detailed stack information in the log for subsequent troubleshooting.

[0079] The interaction between the online self-learning module and the data center strictly adheres to the defined security mechanisms: during data acquisition and transmission, the system strictly follows the security zoning and access control requirements of smart power plants. All acquired data is transmitted through a dedicated industrial control network, and edge nodes only open ports authorized by the firewall whitelist. The transmission protocol uses a TLS encrypted channel, and both communicating parties use certificates for two-way authentication. The data center implements hierarchical authorization through Access Control Lists (ACLs) and Role-Based Access Control (RBAC) mechanisms, allowing different levels of operations, analysis, and development personnel to access corresponding data tables or views according to their role permissions. The log auditing module records each data access behavior, including access time, user identity, access type, and data scope. All audit logs are periodically archived and written to the security log database. This security mechanism is fully implemented in the self-learning module, ensuring that access to training data and model files is traceable and controllable at both the physical and logical levels. Furthermore, the submission, retrieval, and merging operations of the model repository all require digital signature authentication, and any unauthorized modifications are rejected.

[0080] The online self-learning module, like the diagnostic and optimization modules, is deployed using a containerized architecture. Each model retraining task runs as an independent container instance, with unified management of container images. Each image contains a Python runtime environment, a deep learning framework (TensorFlow, PyTorch, or LightGBM), and an internally developed optimization library. During container operation, it communicates in real-time with the cluster monitoring system, monitoring metrics including CPU utilization, GPU memory usage, I / O bandwidth, and task execution time. If the system detects that resource utilization exceeds a threshold, it automatically triggers task migration, scheduling the running instance to an idle node. The task migration process does not affect the integrity of the model files; the model training state is saved via a checkpoint file and restored on the new node.

[0081] Example 2, one embodiment of the present invention, provides a dynamic optimization system for equipment fault diagnosis based on a smart power plant, including a multi-source data acquisition and feature construction module, a mechanism and AI fusion diagnostic modeling module, a dynamic optimization decision and operating condition reconfiguration module, and an online self-learning and model adaptive update module.

[0082] The multi-source data acquisition and feature construction module unifies the structure of the field acquisition layer, edge computing layer, and data center layer, and performs standardized acquisition, time synchronization, and feature vector generation of multi-source operational data from DCS, BOP, NCS, and sensors. The mechanism and AI fusion diagnostic modeling module establishes a dual-channel diagnostic engine with mechanism and AI models, and realizes equipment operating status identification, anomaly detection, and structured diagnostic output through residual calculation and consistency analysis. The dynamic optimization decision-making and operating condition reconfiguration module builds a dynamic optimization model, solves constraints by integrating energy efficiency, latency, and risk factors, and maps the optimal control parameters to the control system setpoint to achieve real-time operating condition adjustment. The online self-learning and model adaptive update module performs sample management, incremental training, and parameter recalibration.

[0083] This embodiment also provides a computer device, including a memory and a processor. The memory stores a computer program, and when the processor executes the computer program, it implements the dynamic optimization method for equipment fault diagnosis based on smart power plants proposed in the above embodiment.

[0084] This embodiment also provides a computer-readable storage medium storing a computer program thereon. When the computer program is executed by a processor, it implements the dynamic optimization method for equipment fault diagnosis based on a smart power plant as proposed in the above embodiment.

[0085] If the aforementioned functions are implemented as software functional units and sold or used as independent products, they can be stored in a computer-readable storage medium. Based on this understanding, the technical solution of this invention, essentially, or the part that contributes to the prior art, or a portion of the technical solution, can be embodied in the form of a software product. This computer software product is stored in a storage medium and includes several instructions to cause a computer device (which may be a personal computer, server, or network device, etc.) to execute all or part of the steps of the methods described in the various embodiments of this invention. The aforementioned storage medium includes various media capable of storing program code, such as USB flash drives, portable hard drives, read-only memory (ROM), random access memory (RAM), magnetic disks, or optical disks.

[0086] The logic and / or steps represented in the flowchart or otherwise described herein, for example, can be considered as a sequenced list of executable instructions for implementing logical functions, and can be embodied in any computer-readable medium for use by, or in conjunction with, an instruction execution system, apparatus, or device (such as a computer-based system, a processor-including system, or other system that can fetch and execute instructions from, an instruction execution system, apparatus, or device). For the purposes of this specification, "computer-readable medium" can be any means that can contain, store, communicate, propagate, or transmit programs for use by, or in conjunction with, an instruction execution system, apparatus, or device.

[0087] More specific examples of computer-readable media (a non-exhaustive list) include: electrical connections (electronic devices) having one or more wires, portable computer disk drives (magnetic devices), random access memory (RAM), read-only memory (ROM), erasable and editable read-only memory (EPROM or flash memory), fiber optic devices, and portable optical disc read-only memory (CDROM). Furthermore, computer-readable media can even be paper or other suitable media on which the program can be printed, because the program can be obtained electronically, for example, by optically scanning the paper or other medium, followed by editing, interpreting, or otherwise processing as necessary, and then stored in computer memory.

[0088] It should be understood that various parts of the present invention can be implemented in hardware, software, firmware, or a combination thereof. In the above embodiments, multiple steps or methods can be implemented in software or firmware stored in memory and executed by a suitable instruction execution system. For example, if implemented in hardware, as in another embodiment, it can be implemented using any one or a combination of the following techniques known in the art: discrete logic circuits having logic gates for implementing logical functions on data signals, application-specific integrated circuits (ASICs) having suitable combinational logic gates, programmable gate arrays (PGAs), field-programmable gate arrays (FPGAs), etc.

[0089] It should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention and not to limit it. Although the present invention has been described in detail with reference to preferred embodiments, those skilled in the art should understand that modifications or equivalent substitutions can be made to the technical solutions of the present invention without departing from the spirit and scope of the technical solutions of the present invention, and all such modifications or substitutions should be covered within the scope of the claims of the present invention.

Claims

1. A dynamic optimization method for equipment fault diagnosis based on smart power plants, characterized in that, include: A three-tiered structure of on-site data acquisition layer, edge computing layer and data center layer is constructed. Standardized acquisition and feature construction of multi-source operational data are carried out through a unified coding, time synchronization and semantic tagging system. After feature construction, a dual-channel fusion diagnostic engine combining mechanistic model and AI model is built, which uses residual calculation and consistency analysis to output parallel inference and structured diagnosis of multi-source data; Based on structured diagnostics, a dynamic optimization model is constructed that includes three elements: energy efficiency, time delay, and risk. The operating parameters are reconfigured in real time through constraint solving and control quantity mapping mechanisms. By using a dynamic optimization model and log data output from structured diagnostics, a mechanism for sample management, incremental training, and parameter recalibration is constructed.

2. The dynamic optimization method for equipment fault diagnosis based on smart power plants as described in claim 1, characterized in that: The three-tiered structure comprising the field acquisition layer, edge computing layer, and data center layer includes... The field acquisition layer interacts with the DCS, BOP, and NCS systems via OPC UA, Modbus TCP, or IEC 61850 protocols. The edge computing layer performs data cleaning, temporal alignment, feature extraction, and wavelet compression on the original signal. The data center layer adopts a hierarchical storage structure of real-time database, historical database and relational database, and uses a unified JSON Schema format for centralized management of feature data.

3. The dynamic optimization method for equipment fault diagnosis based on smart power plants as described in claim 1 or 2, characterized in that: The dual-channel fusion diagnostic engine, which integrates the construction mechanism model and the AI ​​model, includes... Mechanism modeling layer, AI modeling layer, fusion inference layer and model management layer; The mechanism modeling layer establishes a mathematical model based on the equipment's thermodynamic equations, energy balance equations, and structural characteristics; the AI ​​modeling layer trains feature data samples based on multiple algorithms; and the fusion inference layer calculates residuals.

4. The dynamic optimization method for equipment fault diagnosis based on smart power plants as described in claim 3, characterized in that: The dynamic optimization model includes, The multi-objective optimization model is expressed as follows: ； in, Indicates energy efficiency loss. Indicates maintenance delay. Indicates risk factors, , , This represents the weighting coefficient, which is dynamically adjusted based on the system's operating status. It is best to add the meaning to the expression. The optimization process performs constraint solving by combining the Lagrange multiplier method with Newton's iteration, and transforms the optimization results into setpoints for combustion, fan, and pump valve control quantities based on a mapping table.

5. The dynamic optimization method for equipment fault diagnosis based on smart power plants as described in any one of claims 1, 2, and 4, characterized in that: The dynamic optimization model includes, After receiving the equipment operating status information output by the diagnostic engine, the dynamic optimization model obtains the dynamic fluctuation values ​​of load changes, temperature deviations and energy efficiency indicators through the real-time data feedback module. A multi-parameter optimization iterative equation, which includes operating energy consumption, equipment load, and operating condition stability, is constructed based on dynamic fluctuation values, and rolling updates are performed within a continuous time window. Each iteration of the optimization process corrects the step size of the control variables based on real-time data and verifies them synchronously with the mechanistic model through a feedback channel.

6. The dynamic optimization method for equipment fault diagnosis based on smart power plants as described in claim 5, characterized in that: The real-time reconfiguration of operating parameters through constraint solving and control quantity mapping mechanism includes... During the execution of the optimization control quantity, the optimization instruction is smoothed through a two-layer buffer control structure consisting of a logic buffer layer and a time buffer layer. The logic buffer layer automatically compares the differences in set values ​​before and after optimization. When the change in control quantity exceeds the threshold, it pauses the issuance of commands and requests manual confirmation. The time buffer layer uses a linear recursive update method with intervals of 5 to 15 seconds to adjust the control parameters step by step.

7. The dynamic optimization method for equipment fault diagnosis based on smart power plants as described in any one of claims 1, 2, 4, and 6, characterized in that: The constructed sample management, incremental training, and parameter recalibration mechanism includes, Sample data extraction, label binding, sample cleaning, and sample balancing steps; Sample labels are generated from confirmation records of the inspection system or maintenance work order system. Training samples use incremental learning to update model parameters, and the parameter update is expressed as follows: ； in, express Time-model parameters, express Time-model parameters, This represents the gradient update amount during training on new samples. This represents the migration coefficient.

8. A dynamic optimization system for equipment fault diagnosis based on a smart power plant, employing the dynamic optimization method for equipment fault diagnosis based on a smart power plant as described in any one of claims 1 to 7, characterized in that: It includes a multi-source data acquisition and feature construction module, a mechanism and AI fusion diagnostic modeling module, a dynamic optimization decision-making and operating condition reconfiguration module, and an online self-learning and model adaptive update module; The multi-source data acquisition and feature construction module is used to unify the structure of the field acquisition layer, edge computing layer and data center layer, and to perform standardized acquisition, time synchronization and feature vector generation of multi-source operation data from DCS, BOP, NCS and sensors; The mechanism and AI fusion diagnostic modeling module is used to establish a dual-channel diagnostic engine of mechanism model and AI model, and realizes equipment operating status identification, anomaly detection and structured diagnostic output through residual calculation and consistency analysis; The dynamic optimization decision-making and operating condition reconfiguration module is used to construct a dynamic optimization model, solve for constraints by comprehensively considering energy efficiency, time delay and risk factors, and map the optimal control parameters to the control system setpoint to achieve real-time operating condition adjustment. The online self-learning and model adaptive update module is used to perform sample management, incremental training, and parameter recalibration.

9. A computer device comprising a memory and a processor, wherein the memory stores a computer program, characterized in that, When the processor executes the computer program, it implements the steps of the dynamic optimization method for equipment fault diagnosis based on a smart power plant as described in any one of claims 1 to 7.

10. A computer-readable storage medium having a computer program stored thereon, characterized in that, When the computer program is executed by the processor, it implements the steps of the dynamic optimization method for equipment fault diagnosis based on a smart power plant as described in any one of claims 1 to 7.

Images