Large-model-enabled equipment full-life-cycle digital twinborn decision-making system

Abstract

The invention relates to the technical field of equipment management, in particular to an equipment full-life-cycle digital twinborn decision-making system enabling a large model. Comprising a digital twin modeling unit; a large model enabling analysis unit, wherein the large model enabling analysis unit adopts a hydroelectric equipment multi-modal causal constraint analysis model; a whole-process closed-loop management and control unit; and an intelligent decision output unit. According to the method, the multi-modal causal constraint analysis model adaptive to the working condition of the hydroelectric equipment is constructed, and a causal chain verification backtracking mechanism is introduced, so that the problem that the reasoning result lacks logic verification is effectively solved, the logic consistency of a fault reasoning conclusion is guaranteed, and the reliability of decision output is improved; through a scene adaptation mode of'pre-training + fine tuning 'of a large model, multi-modal feature fusion processing and deep linkage of a workflow engine and a digital twinborn body, full-life-cycle management requirements of equipment are fully covered, and the refinement and intelligence level of hydroelectric equipment management is further improved.

Description

Technical Field

[0001] This invention relates to the field of equipment management technology, and more specifically, to a large-scale model-enabled digital twin decision-making system for the entire lifecycle of equipment. Background Technology

[0002] The full lifecycle management of hydropower equipment spans the three core stages of technological upgrading and procurement, operation and maintenance, and decommissioning, and is crucial for ensuring the safe and stable operation of hydropower enterprises and improving operational efficiency. For hydropower enterprises with large installed capacity and long operating years, the systematic and standardized nature of equipment management is particularly important. Currently, some hydropower enterprises in the industry have gradually promoted online equipment management, achieving the digitization of basic functions such as ledger recording and maintenance registration. However, overall, it remains in a decentralized management stage and has not yet built an integrated digital and intelligent operation management support system covering the entire lifecycle of equipment. Multi-source information such as real-time sensor data, design drawings, maintenance records, and historical faults related to the equipment has not been effectively integrated, and there is a lack of deep linkage between management processes and the actual status of equipment, making it difficult to achieve refined control throughout all stages.

[0003] In the existing technology, relevant patents have explored the integration of digital twins and equipment management. For example, Chinese patent CN202410996770.X discloses a high-end equipment full life cycle management platform and method based on digital twins. It covers the entire life cycle management of equipment, including equipment design simulation, manufacturing quality control, operation status monitoring, fault prediction, maintenance plan formulation, and decommissioning and recycling assessment, by constructing a digital twin model of the equipment. Another example is Chinese patent CN202410726578.9, which discloses a digital twin implementation method and device. It collects the operation data of the target equipment itself and its twin, optimizes the digital twin model through an endogenous intelligent module based on a large model, updates the twin, and provides feedback on the simulation results.

[0004] Despite the design advantages of the aforementioned technical solutions, they also suffer from the following technical shortcomings: First, insufficient integration of multi-source data and adaptation to operating conditions: CN202410996770.X does not involve deep fusion processing of heterogeneous data from multiple sources such as real-time sensor data, design drawings, and maintenance texts; CN202410726578.9's data collection scope is limited to the operating data of the equipment itself and its twin; and neither of these methods optimizes the model for specific operating conditions such as differences in high and low loads of hydropower equipment, resulting in insufficient applicability of the model to hydropower scenarios and incomplete data support. Second, the inference results lack a logical verification mechanism: CN202410996770.X does not establish a logical verification mechanism. The causal relationship verification process related to faults, as outlined in CN202410726578.9, focuses solely on the iterative optimization of the digital twin model, failing to verify the logical consistency of the reasoning conclusions. This may lead to contradictory reasoning results and affect the reliability of decisions. Thirdly, the system lacks end-to-end linkage and decision compliance verification: CN202410996770.X's management process and digital twin status linkage remain at a basic level. CN202410726578.9 does not cover the entire lifecycle and all stages of equipment upgrades, procurement, and decommissioning, and neither incorporates power industry technical standards for simulation verification of decision-making schemes, making it difficult to guarantee the compliance of maintenance, decommissioning, and upgrade decisions. Therefore, we propose a large-scale model-enabled digital twin decision-making system for the entire equipment lifecycle. Summary of the Invention

[0005] The purpose of this invention is to provide a large-scale model-enabled digital twin decision-making system for the entire lifecycle of equipment, in order to solve the problems mentioned in the background art, such as insufficient integration of multi-source data and adaptation to working conditions, lack of logical verification mechanism for inference results, and lack of full-process linkage and compliance verification of decision-making.

[0006] To address the aforementioned technical problems, the present invention aims to provide a large-scale model-enabled digital twin decision-making system for the entire lifecycle of equipment, comprising: The digital twin modeling unit adopts multi-source heterogeneous data fusion and physical-digital bidirectional linkage technology to construct a dynamic digital twin of hydropower equipment in the entire process of "technical transformation and procurement - operation and maintenance - decommissioning and disposal", so as to realize real-time synchronization of parameters between the digital twin and the physical equipment; The large-scale model-enabled analysis unit employs a multimodal causal constraint analysis model for hydropower equipment. This model combines historical operational data of the equipment to perform intelligent reasoning, extracting specific features related to faults and operational states from the equipment's data. In the pre-training phase, a general multimodal model is used as the initial foundation, and the model is optimized by associating and improving the objective parameters of hydropower equipment operation to form a pre-trained version of the multimodal causal constraint analysis model. In the fine-tuning phase, historical fault and maintenance data of core equipment in the hydropower plant are used to iteratively optimize the parameters of the multimodal causal constraint analysis model. The inference output of the multimodal causal constraint analysis model is detected through the fault causal chain mechanism of the hydropower equipment, and if contradictions exist, the model is backtracked to the feature layer for re-inference. The full-process closed-loop management and control unit adopts workflow engine and digital twin physical state binding technology to build a full life cycle process template of equipment that is adapted to the technical standards of the power industry, so as to realize the linkage and closed loop between process and equipment status. The intelligent decision output unit integrates the dynamic state of the digital twin with the reasoning conclusions of the multimodal causal constraint analysis model of the hydropower equipment, and combines the technical standards of the hydropower equipment to derive decision schemes for equipment maintenance, decommissioning, and technical upgrading.

[0007] As a further improvement to this technical solution, the digital twin modeling unit includes a multi-source heterogeneous data fusion module. The multi-source heterogeneous data fusion module is used to access real-time sensor data of hydropower equipment, equipment design drawing data, operation and maintenance record text data and historical fault data. After processing the accessed data by removing outliers, standardizing data format and filling in missing values, a weighted fusion algorithm is used to allocate weights according to the credibility of different types of data, and the preprocessed multi-source data is fused into a digital twin in a unified format to construct the basic data.

[0008] As a further improvement to this technical solution, the digital twin modeling unit also includes a physical-digital bidirectional linkage module. The physical-digital bidirectional linkage module acquires the actual operating parameters of the physical body of the hydropower equipment through edge sensors deployed on the physical body of the hydropower equipment at a preset acquisition frequency. Based on the acquired actual operating parameters of the physical body of the hydropower equipment, it updates the simulation parameters of the dynamic digital twin in real time and compares the actual operating parameters of the physical body of the hydropower equipment with the simulation parameters of the dynamic digital twin. When the deviation between the actual operating parameters of the physical body of the hydropower equipment and the simulation parameters of the dynamic digital twin exceeds a preset threshold, the parameter synchronization calibration process is triggered.

[0009] As a further improvement to this technical solution, the large model-enabled analysis unit includes a multimodal data adaptation module, a multimodal feature extraction module, a model pre-training and fine-tuning optimization module, and a causal chain verification and backtracking module; wherein: The multimodal data adaptation module is used to receive data specific to hydropower equipment, perform format standardization, noise filtering and modal consistency processing, and output standardized multimodal data that meets the input requirements of the multimodal causal constraint analysis model of hydropower equipment. The multimodal feature extraction module is used to perform targeted feature extraction, normalization, and dynamic weight fusion on standardized multimodal data to form a unified fusion feature vector related to faults and operating states. The model pre-training and fine-tuning optimization module uses a general multimodal model as the initial basis. It first performs pre-training constraint optimization by combining objective parameters of hydropower equipment operation, and then performs fine-tuning and iterative optimization using historical fault and maintenance data of core equipment in hydropower plants, outputting a scenario-adapted multimodal causal constraint analysis model for hydropower equipment. The causal chain verification and backtracking module quantifies the verification model's inference output through the causal chain mechanism of hydropower equipment failure. When a contradiction exists, it triggers the feature layer backtracking and re-inference process.

[0010] As a further improvement to this technical solution, the process by which the multimodal data adaptation module outputs standardized multimodal data that meets the input requirements of the multimodal causal constraint analysis model for hydropower equipment includes the following steps: S20.1 Receive vibration time sequence data of the physical body of the hydroelectric equipment, oil chromatography detection text data, equipment appearance inspection image data, real-time unit speed data and guide vane opening adjustment data. S20.2. An adaptive median filtering algorithm is used to process the vibration time series data, and the window radius is dynamically adjusted according to the vibration frequency; a Gaussian filtering algorithm is used to process the equipment appearance inspection image data; and invalid characters are removed from the oil chromatography detection text data through regularization. S20.3 The time-series data includes vibration time-series data, real-time unit speed data, and guide vane opening adjustment data after adaptive median filtering. Missing values ​​are filled in using linear interpolation and the sampling frequency is standardized. The equipment appearance inspection image data after Gaussian filtering is standardized in resolution. The final output is a standardized multimodal data set containing processed vibration time-series data, processed oil chromatography detection text data, processed equipment appearance inspection image data, standardized real-time unit speed data, and standardized guide vane opening adjustment data, providing suitable input for subsequent feature extraction.

[0011] As a further improvement to this technical solution, the process of generating a unified fused feature vector by the multimodal feature extraction module includes the following steps: S21.1 The processed vibration time series data is used to extract time-frequency domain features through wavelet transform. The processed oil chromatography detection text data was used to extract feature vectors using the BERT word embedding algorithm. Spatial feature vectors were extracted from the processed equipment appearance inspection image data using a ResNet50 network. Features were extracted from real-time speed data of standardized units and standardized guide vane opening adjustment data through a sliding window statistical method. ; S21.2, All unimodal features are normalized using Min-Max. Mapping to a unified interval eliminates dimensional differences; S21.3, Based on the operating condition adaptation factor of hydropower equipment Assign modal weights And satisfy Through multimodal feature fusion Integrate all normalized features to output a unified fused feature vector. ;in When in the low load range, the focus is on vibration and image features, while when in the high load range, the focus is on text and rotational speed opening features.

[0012] As a further improvement to this technical solution, the process of the model pre-training and fine-tuning optimization module outputting the final model includes the following steps: S22.1 Constructing a set of objective constraint parameters ,in The rated speed of the equipment, Design the water head for the equipment, Rated power of the equipment The permissible vibration threshold for the equipment; Normalization is performed to obtain the normalized set of constraint parameters, denoted as . ;based on Construct the correlation mapping matrix between constraint parameters and model weights. The pre-trained model weight parameter set is generated through parameter association logic. This pre-trained model weight parameter set is denoted as... The parameter correction amount is dynamically determined based on the original weight parameter set of the general multimodal model, which is denoted as . ; S22.2 Constructing a fine-tuning dataset ,in For the first The fused feature vector of each sample, For the first Fault labels for each sample; The dataset is divided into two parts according to a reasonable ratio: a dataset for model training and a dataset for model performance validation. The dataset for model training is denoted as... The dataset used for model performance validation is denoted as . ; S22.3, using the cross-entropy loss function To optimize the objective, For model weight parameters The prediction error loss function is calculated; the model weight parameters are updated using the stochastic gradient descent algorithm. The model weight parameters after rounds of iteration are denoted as During the iteration process, the parameter update step size is dynamically adjusted. The parameter update step size of the round of iteration is denoted as ; Calculation based on validation set loss function value ;when The iteration stops when the model stabilizes, and the output is a multimodal causal constraint analysis model of hydropower equipment with final optimization parameters. The final optimization parameters are recorded as follows: This is the final set of weight parameters for the model, determined after pre-training and fine-tuning.

[0013] As a further improvement to this technical solution, the process of the causal chain verification and backtracking module to complete the inference output verification and backtracking includes the following steps: S23.1 Typical causal relationship of built-in water and electricity equipment ,in This is a set of typical causal relationships for hydroelectric equipment failures. For the number of causal relationship categories, For the first The set of antecedent features for causal relationships For the first A set of consequence features for causal relationships, with each type of causal relationship configured with an inherent technical correlation strength coefficient. ; S23.2, Using the cosine similarity calculation function The degree of matching between the quantified fusion feature vector and the antecedent feature set is specifically achieved through... Calculate the antecedent matching degree, based on the antecedent matching degree and the inherent technical correlation strength coefficient. A quantitative indicator of the overall matching between computational model inference results and causal chain rules ;like If the preset reasonable threshold is not reached, the inference output is determined to be contradictory. S23.3, Feature Dimension Set with Insufficient Matching Degree between Location and Causal Chain Rules The dimension of a single conflict feature is ; S23.4. Recall the multimodal feature extraction module and use an enhancement algorithm to re-extract features from the original standardized data corresponding to the conflict dimensions, obtaining the enhanced secondary fusion feature vector. The enhancement algorithm includes wavelet packet transform overlay of time-series data, attention mechanism overlay of image data, and keyword weight enhancement overlay of text data; Input with final optimization parameters The multimodal causal constraint analysis model of hydropower equipment is used for secondary inference; S23.5. Repeat steps S23.2 to S23.4 until the causal relationship is determined. If the requirements are met or a reasonable number of backtracking attempts are reached, a conflict warning message will be output.

[0014] As a further improvement to this technical solution, the full-process closed-loop control unit includes a process template construction module, a status binding communication module, a linkage rule execution module, and a closed-loop data synchronization module; wherein: The process template construction module is based on the visualization modeling technology of the workflow engine. It transforms the technical standards of the power industry into standardized process node parameters that the engine can recognize, and constructs a digital process template for the entire stage of equipment "technical upgrade procurement - operation and maintenance - decommissioning". The digital process template supports parameterized adjustment through technical interfaces without the need for manual redefinition of the process framework. The state binding communication module adopts the OPCUA industrial communication protocol to establish a two-way data channel between the workflow engine and the dynamic digital twin of the digital twin modeling unit. Through the real-time data subscription-push mechanism, the workflow engine can obtain the physical state parameters of the equipment in the digital twin in milliseconds and transmit the process execution data back to the digital twin in real time. The linkage rule execution module has a built-in logical judgment algorithm based on the physical status parameters of the device. When the device status parameters fed back by the digital twin meet the preset technical threshold, the algorithm automatically triggers the process node jump instruction of the workflow engine without manual intervention in process switching. The closed-loop data synchronization module synchronizes the process execution data to the dynamic digital twin in real time through the time-series database writing interface, and updates the simulation parameters of the dynamic digital twin. At the same time, based on the equipment status change rate fed back by the dynamic digital twin, the execution time limit parameters of the process nodes are dynamically optimized through the priority adjustment algorithm, forming a technical-level linkage closed loop of "equipment physical status - process technical parameters - digital twin simulation parameters".

[0015] As a further improvement to this technical solution, the intelligent decision output unit includes a multi-source decision data fusion module, a technical standard parameterization adaptation module, a multi-objective decision derivation module, and a decision scheme verification module; wherein: The multi-source decision data fusion module adopts a feature-level fusion algorithm to associate and integrate the dynamic state data of the digital twin output by the digital twin modeling unit with the reasoning conclusions of the multimodal causal constraint analysis model of hydropower equipment in the large model empowerment analysis unit. Abnormal data is removed through a data consistency verification algorithm to generate a decision input dataset in a unified format. The technical standard parameterization and adaptation module is used to convert the technical standards of hydropower equipment into calculable decision constraint parameters, forming a standard parameter library, and supporting dynamic updates of standard parameters through technical interfaces. The multi-objective decision derivation module has a built-in multi-objective optimization algorithm. Combined with the input dataset of the multi-source decision data fusion module and the constraint parameters of the technical standard parameterization adaptation module, it automatically derives candidate decision schemes for equipment maintenance, decommissioning and technical upgrading, and outputs the technical indicators corresponding to the candidate decision schemes. The decision scheme verification module is used to input candidate decision schemes into a dynamic digital twin for simulation and to verify whether the state changes of the digital twin after the implementation of the scheme meet the technical standard requirements. If the simulation results meet the preset compliance conditions, the final decision scheme is output. If not, the results are fed back to the multi-objective decision derivation module to readjust and optimize the parameters until a compliant scheme is generated.

[0016] Compared with the prior art, the beneficial effects of the present invention are as follows: 1. This invention integrates multi-source heterogeneous data such as real-time data from hydropower equipment sensors, design drawings, and operation and maintenance texts to construct a dynamically updated digital twin and achieve real-time synchronization of physical and digital parameters. Furthermore, it optimizes models for specific operating conditions such as high and low load differences in hydropower equipment, effectively solving the problems of insufficient integration of multi-source data and insufficient model adaptability, thereby improving the comprehensiveness of data support and the applicability of the model to hydropower scenarios. 2. This invention constructs a multimodal causal constraint analysis model adapted to the operating conditions of hydropower equipment, introduces a causal chain verification and backtracking mechanism, effectively solves the problem of lack of logical verification of reasoning results, ensures the logical consistency of fault reasoning conclusions, and improves the reliability of decision output; 3. This invention achieves closed-loop linkage between the entire lifecycle process of equipment upgrade procurement, operation and maintenance, and decommissioning and the status of the digital twin, and combines the decision-making scheme with the digital twin simulation verification of the power industry technical standards. This effectively solves the problems of insufficient linkage throughout the entire process and lack of verification of decision compliance, and ensures the compliance of maintenance, decommissioning and upgrade decisions. 4. This invention comprehensively covers the full life cycle management needs of equipment through a scenario adaptation mode of "pre-training + fine-tuning" of a large model, multimodal feature fusion processing, and deep linkage between the workflow engine and the digital twin. It further improves the refinement and intelligence level of hydropower equipment management and realizes the upgrade from decentralized management to integrated control. Attached Figure Description

[0017] Figure 1 This is a schematic diagram of the system framework of the present invention; The meanings of the labels in the diagram are as follows: 1. Digital twin modeling unit; 10. Multi-source heterogeneous data fusion module; 11. Physical-digital bidirectional linkage module; 2. Large model-enabled analysis unit; 20. Multimodal data adaptation module; 21. Multimodal feature extraction module; 22. Model pre-training and fine-tuning optimization module; 23. Causal chain verification and backtracking module; 3. Full-process closed-loop control unit; 30. Process template construction module; 31. Status binding communication module; 32. Linkage rule execution module; 33. Closed-loop data synchronization module; 4. Intelligent decision output unit; 40. Multi-source decision data fusion module; 41. Technical standard parameterization adaptation module; 42. Multi-objective decision derivation module; 43. Decision scheme verification module. Detailed Implementation

[0018] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those of ordinary skill in the art without creative effort are within the scope of protection of the present invention.

[0019] like Figure 1 As shown, this embodiment provides a large-scale model-enabled digital twin decision-making system for the entire lifecycle of equipment, including: Digital twin modeling unit 1 adopts multi-source heterogeneous data fusion and physical-digital two-way linkage technology to construct a dynamic digital twin of hydropower equipment in the whole stage of "technical transformation and procurement - operation and maintenance - decommissioning and disposal", so as to realize the real-time synchronization of parameters between the digital twin and the physical equipment. In this embodiment, the digital twin modeling unit 1 includes a multi-source heterogeneous data fusion module 10. The multi-source heterogeneous data fusion module 10 is used to access real-time sensor data of hydropower equipment, equipment design drawing data, operation and maintenance record text data and historical fault data. After processing the accessed data by removing outliers, standardizing data format and filling in missing values, a weighted fusion algorithm is used to allocate weights according to the credibility of different types of data, and the preprocessed multi-source data is fused into a digital twin in a unified format to construct the basic data.

[0020] Specifically, real-time sensor data covers mechanical quantities (speed, vibration frequency, guide vane opening, etc.), hydraulic parameters (head, flow rate, volute pressure, etc.), and electrical quantities (active power, voltage, current, etc.) of hydropower equipment; equipment design drawing data includes 3D structural drawings of equipment, component assembly drawings, etc.; operation and maintenance record text data consists of unstructured data such as manually entered inspection logs and maintenance reports; historical fault data includes related data such as past fault types, fault occurrence time, fault handling solutions and results.

[0021] Specifically, the preprocessing of the access data is implemented as follows: outlier removal uses box plot or Z-score methods. By setting a normal data distribution range, outlier data exceeding the reasonable range is identified and marked to avoid interference from extreme values ​​on the fusion results; data format standardization adopts unified industry standard units (such as power unit kW, voltage unit kV), converting design drawings of different formats into the FBX universal format, converting maintenance record text data into UTF-8 encoded structured text, and adding timestamps and unique device identifiers to all data to ensure that the time synchronization accuracy meets the power system data acquisition specifications; missing value completion adopts an adaptation method for different types of data. Time-series sensor data is supplemented using linear interpolation, discrete design drawing data is marked with missing items by associating with similar equipment drawing information, and text-based maintenance data retains the original missing state and marks the missing identifier.

[0022] Furthermore, after preprocessing, the multi-source heterogeneous data fusion module 10 employs a weighted fusion algorithm, with weight allocation determined using the entropy weight method: First, data reliability evaluation indicators (including data timeliness, completeness, and accuracy) are constructed. After standardizing each indicator, the information entropy of each indicator is calculated, and then the objective weight of each indicator is derived from the entropy value—the weight of data timeliness is obtained by standardizing the difference between its timestamp and the current time to calculate the entropy value; the weight of data completeness is obtained by standardizing the entropy value of the proportion of missing values; and the weight of data accuracy is obtained by standardizing the entropy value of the error rate (such as the deviation rate between sensor data and calibration values). Finally, the comprehensive reliability weight of various types of data is calculated—real-time sensor data has the highest comprehensive weight due to its strong timeliness and small error; equipment design drawing data, as static basic data, has the next highest weight; and maintenance record text data and historical fault data are dynamically adjusted according to data completeness (e.g., historical fault data with 90% completeness has a higher weight than maintenance record data with 60% completeness). Meanwhile, during weighted fusion, various types of data are first mapped to the same data dimension, and then data with the same time node and the same device encoding are weighted and summed according to weight to generate a unified format (JSON format) digital twin construction base data, which is stored in a time series database (such as InfluxDB) for use when building dynamic digital twins.

[0023] In this embodiment, the digital twin modeling unit 1 further includes a physical-digital bidirectional linkage module 11. The physical-digital bidirectional linkage module 11 acquires the actual operating parameters of the physical body of the hydropower equipment through edge sensors deployed on the physical body of the hydropower equipment at a preset acquisition frequency. Based on the acquired actual operating parameters of the physical body of the hydropower equipment, it updates the simulation parameters of the dynamic digital twin in real time and compares the actual operating parameters of the physical body of the hydropower equipment with the simulation parameters of the dynamic digital twin. When the deviation between the actual operating parameters of the physical body of the hydropower equipment and the simulation parameters of the dynamic digital twin exceeds a preset threshold, the parameter synchronization calibration process is triggered.

[0024] Specifically, the deployment locations of the edge sensors are determined based on the core components of the equipment: the vibration sensor for the turbine is deployed at the main shaft bearing housing, the temperature sensor for the generator is embedded in the stator winding slot, and the oil level sensor for the main transformer is installed on the side wall of the oil tank. The acquisition frequency is set in stages according to the degree of influence of the parameters on equipment operation: critical parameters (such as generator stator temperature and main transformer voltage) are acquired at a frequency of 1 time / second, while general parameters (such as equipment ambient temperature and cooling fan speed) are acquired at a frequency of 1 time / 5 minutes, avoiding data transmission pressure caused by high-frequency acquisition. Simultaneously, the acquired data is transmitted to the edge computing gateway via the Modbus-TCP protocol. The transmission process uses the AES-256 encryption algorithm to ensure data security. The edge computing gateway maintains time synchronization with the power plant's NTP server, ensuring that the time error of the acquired data does not exceed 5 milliseconds.

[0025] Specifically, the physical-digital bidirectional linkage module 11 updates the simulation parameters of the dynamic digital twin in real time through the RESTful API interface based on the actual operating parameters of the collected physical body of the hydropower equipment. For example, when the sensor detects that the speed of generator No. 1 increases from 1500 r / min to 1505 r / min, the physical-digital bidirectional linkage module 11 immediately transmits the parameter to the digital twin modeling platform and synchronously updates the speed simulation parameters of generator No. 1 in the digital twin, with an update delay of no more than 100 milliseconds.

[0026] Meanwhile, the physical-digital bidirectional linkage module 11 compares the actual operating parameters of the physical body of the hydroelectric equipment with the simulated parameters of the dynamic digital twin in real time. The preset threshold is determined by statistically analyzing the historical normal operating data of the equipment: for sensor parameters (such as vibration frequency), the standard deviation σ of the normal operating data of the past 3 months is calculated, and the preset threshold is set to ±3σ (that is, when the parameter fluctuation exceeds this range, it is judged as an abnormal deviation); for equipment status parameters (such as guide vane opening), the preset threshold is ±5% of the design rated value (such as when the rated opening is 80%, the threshold range is 76% to 84%).

[0027] Furthermore, when the deviation between the actual operating parameters of the physical hydroelectric equipment and the simulated parameters of the dynamic digital twin exceeds a preset threshold, the physical-digital bidirectional linkage module 11 automatically triggers the parameter synchronization calibration process: First, the edge computing gateway detects the working status of the edge sensors (such as power supply voltage and communication connection status). If the sensor has a power outage or communication interruption problem, the sensor fault is repaired first and then the data is collected again. Subsequently, if the sensor is working properly, the Kalman filter algorithm is used to filter noise from the collected actual operating parameters to eliminate instantaneous data fluctuations caused by electromagnetic interference. Then, based on the filtered actual operating parameters, the corresponding simulation parameters of the dynamic digital twin are corrected through the parameter adjustment interface of the digital twin modeling platform (such as adjusting the simulated rotation speed from 1500 r / min to 1505 r / min). Finally, the adjusted simulated parameters are compared with the actual parameters of the physical equipment again. If the deviation still exceeds the preset threshold, the calibration process of the second and third steps above is repeated until the deviation is within a reasonable range. If the deviation still fails to meet the standard after three consecutive calibrations, the physical-digital bidirectional linkage module 11 sends a calibration warning message to the power plant operation and maintenance system, prompting manual intervention to check the status of the digital twin model or sensor to ensure the consistency of parameters between the dynamic digital twin and the physical equipment.

[0028] The large-scale model-enabled analysis unit 2 adopts a multimodal causal constraint analysis model for hydropower equipment. This model combines historical operating data of the equipment to conduct intelligent reasoning, extracting specific features related to faults and operating states from the equipment's specific data. In the pre-training stage, a general multimodal model is used as the initial foundation, and the model is optimized by associating and optimizing the objective parameters of hydropower equipment operation to form a pre-trained version of the multimodal causal constraint analysis model. In the fine-tuning stage, the parameters of the multimodal causal constraint analysis model are iteratively optimized using historical fault and maintenance data of the core equipment of the hydropower plant. The inference output of the multimodal causal constraint analysis model is detected through the fault causal chain mechanism of the hydropower equipment. When contradictions exist, the model is backtracked to the feature layer for re-inference. The large model-enabled analysis unit 2 includes a multimodal data adaptation module 20, a multimodal feature extraction module 21, a model pre-training and fine-tuning optimization module 22, and a causal chain verification and backtracking module 23; among which: In this embodiment, the multimodal data adaptation module 20 receives data specific to hydropower equipment, performs format standardization, noise filtering, and modal consistency processing, and outputs standardized multimodal data that meets the input requirements of the multimodal causal constraint analysis model for hydropower equipment. The process by which the multimodal data adaptation module 20 outputs standardized multimodal data that meets the input requirements of the multimodal causal constraint analysis model for hydropower equipment includes the following steps: S20.1 Receive vibration time sequence data of the physical body of the hydroelectric equipment, oil chromatography detection text data, equipment appearance inspection image data, real-time unit speed data and guide vane opening adjustment data. S20.2. An adaptive median filtering algorithm is used to process the vibration time series data, and the window radius is dynamically adjusted according to the vibration frequency; a Gaussian filtering algorithm is used to process the equipment appearance inspection image data; and invalid characters are removed from the oil chromatography detection text data through regularization. S20.3 The time-series data includes vibration time-series data, real-time unit speed data, and guide vane opening adjustment data after adaptive median filtering. Missing values ​​are filled in using linear interpolation and the sampling frequency is standardized. The equipment appearance inspection image data after Gaussian filtering is standardized in resolution. The final output is a standardized multimodal data set containing processed vibration time-series data, processed oil chromatography detection text data, processed equipment appearance inspection image data, standardized real-time unit speed data, and standardized guide vane opening adjustment data, providing suitable input for subsequent feature extraction.

[0029] Specifically, the multimodal data adaptation module 20 receives data specific to hydropower equipment, performs format standardization, noise filtering, and modal consistency processing, and outputs a standardized multimodal data set that meets the input requirements of the multimodal causal constraint analysis model for hydropower equipment, wherein: The received data includes vibration time sequence data of hydroelectric equipment (such as X / Y / Z three-dimensional vibration acceleration data at the main shaft bearing housing), oil chromatography test text data, equipment appearance inspection image data (such as appearance of stator winding end and guide vane sealing surface image), real-time unit speed data (such as measured speed at generator output end) and guide vane opening adjustment data (such as opening percentage data fed back by guide vane actuator). In the noise filtering process, an adaptive median filtering algorithm is used for the vibration time series data, with a window radius of... Based on vibration frequency Dynamic adjustment ( hour , hour , hour The equipment visual inspection image data were processed using a 3×3 core and standard deviation method. Gaussian filtering, the formula is: ,in For the filter kernel in The weight value of the position. Invalid characters are removed from the text data of oil chromatography detection using regular expressions; During modal consistency processing, time-series data (vibration time series, unit speed, guide vane opening) are padded with missing values ​​through linear interpolation and a unified sampling frequency of 10Hz is applied. Image data is standardized to a resolution of 640×480 pixels, and the final output is a standardized multimodal data set. ,in To process the vibration time series data, To process the post-oil chromatography text data, For the processed image data, To standardize the speed and opening data.

[0030] Traditional multimodal data processing solutions lack adaptability to hydropower equipment-specific data (vibration time series, oil chromatography text, maintenance images, and rotational speed). Problems include data format heterogeneity leading to fusion obstacles, the generalization of noise filtering methods resulting in poor noise suppression in hydropower scenarios, and a lack of modal consistency in multimodal data. For example, when using fixed-parameter filtering to process broadband vibration data of turbines, over-smoothing of fault features or residual noise can easily occur, severely affecting the accuracy of subsequent analysis. The multimodal data adaptation module 20 specifically receives four types of hydropower equipment-specific data and, through adaptive median filtering (dynamically adjusting the window radius based on vibration frequency), Gaussian filtering (customizing parameters for maintenance images), linear interpolation, and resolution unification, completes format standardization, noise filtering, and modal consistency processing, outputting a standardized multimodal data set that meets the input requirements of the hydropower equipment multimodal causal constraint analysis model. It solves the technical pain points of traditional multimodal preprocessing in hydropower scenarios, such as "difficulty in integrating heterogeneous data, inaccurate noise filtering, and poor modal alignment", and realizes customized preprocessing of hydropower-specific data, laying a highly adaptable and high-quality data foundation for subsequent model analysis.

[0031] In this embodiment, the multimodal feature extraction module 21 is used to perform targeted feature extraction, normalization, and dynamic weight fusion on standardized multimodal data to form a unified fused feature vector related to faults and operating states. The process of generating the unified fused feature vector by the multimodal feature extraction module 21 includes the following steps: S21.1 The processed vibration time series data is used to extract time-frequency domain features through wavelet transform. The processed oil chromatography detection text data was used to extract feature vectors using the BERT word embedding algorithm. Spatial feature vectors were extracted from the processed equipment appearance inspection image data using a ResNet50 network. Features were extracted from real-time speed data of standardized units and standardized guide vane opening adjustment data through a sliding window statistical method. ; S21.2, All unimodal features are normalized using Min-Max. Mapping to a unified interval eliminates dimensional differences; S21.3, Based on the operating condition adaptation factor of hydropower equipment Assign modal weights And satisfy Through multimodal feature fusion Integrate all normalized features to output a unified fused feature vector. ;in When in the low load range, the focus is on vibration and image features, while when in the high load range, the focus is on text and rotational speed opening features.

[0032] Specifically, single-modal feature extraction is performed separately for each data modality, as follows: right A three-level wavelet decomposition was performed using the db4 wavelet basis function to extract time-frequency domain features. (dimension is) (i.e., a feature vector with 1 row and 128 columns), the calculation formula is as follows: ; In the formula, This represents the approximation coefficients after 3-level wavelet decomposition (reflecting the low-frequency trend of the data). , , These represent the detail coefficients (reflecting high-frequency fluctuations in the data) after wavelet decomposition at levels 3, 2, and 1, respectively. This represents a function for calculating the mean of a vector. The function that calculates the variance of a vector. The function representing the calculation of vector energy (the calculation logic is as follows) , For vectors The One element, (where the vector length is...) This represents a function for calculating the standard deviation of a vector. right The BERT-base model with pre-trained weights "bert-base-chinese" is used to extract word embedding features, and the feature vector corresponding to the [CLS] identifier token output by the model is taken as... (dimension is) The calculation formula is as follows: ; In the formula, This represents the feature extraction function of the BERT model. This indicates that Chinese pre-trained weights are used. This represents the sentence-initial classification token output by the BERT model, and its corresponding feature vector can represent the semantic information of the entire text. right Spatial features are extracted using a ResNet50 model with pre-trained weights of "ImageNet" (with the last fully connected layer removed), and the output is... (dimension is) The calculation formula is as follows: ; In the formula, This represents the feature extraction function of the ResNet50 model. This indicates that weights pre-trained on the ImageNet dataset are used; right Statistical features are extracted using a sliding window (window size T = 10 data points, window step size = 5 data points), forming... (dimension is) The calculation formula is as follows: ; In the formula, These represent the first to the second elements after the sliding window is divided. Subsequences ( (The total number of subsequences after the sliding window segmentation). This represents a function that calculates the maximum value of a vector.

[0033] Specifically, the Min-Max normalization method is used in the feature normalization stage to... Mapping to the [0,1] interval respectively, the calculation formula is as follows: ; In the formula, This represents the original single-modal features to be normalized. This represents the minimum value of the feature on the model training dataset. This indicates the maximum value of the feature on the model training dataset. Represents the normalized single-modal features (correspondingly denoted as ). ).

[0034] Specifically, in the dynamic weight fusion stage: First, based on the real-time load rate of the hydropower equipment Determine the operating condition adaptation factor ( The calculation logic is as follows ,when (Low load condition) ,when (High load conditions) ); Then assign the weights to each normalized feature according to the following formula: ; In the formula, express The weight, express The weight, express The weight, express The weights, and satisfying ; Finally, a unified fused feature vector is generated by weighted summation. (dimension is) ,Depend on (Constructed by dimensional features), the calculation formula is as follows: .

[0035] Traditional multimodal feature fusion schemes suffer from two major technical drawbacks: First, single-modal feature extraction lacks specificity for the hydropower industry. Using general time-series, text, and image feature extraction methods fails to accurately capture the unique fault characteristics of hydropower equipment, such as insufficient ability to extract low-frequency fault features from turbine vibration time-series data. Second, the fusion weights are fixed, failing to consider the complex operating conditions of hydropower equipment with significant differences in high and low loads, leading to a disconnect between feature fusion and actual scenario requirements. The multimodal feature extraction module 21 uses db4 wavelet transform to extract time-domain and frequency-domain features for vibration time-series data, a BERT model to extract specialized semantic features for oil chromatography text data, a ResNet50 model to extract spatial features for maintenance image data, and a sliding window to extract statistical features for speed and opening data. Simultaneously, an operating condition adaptation factor is introduced. The multimodal feature extraction module 21 dynamically adjusts the feature weights of each mode based on the real-time load rate of the hydropower equipment (emphasizing vibration and image feature weights under low load conditions, and emphasizing text and rotational speed opening feature weights under high load conditions). Through a professional feature extraction and dynamic weight fusion mechanism, the multimodal feature extraction module 21 solves the problem of "non-exclusive features and unsuitable weights" in traditional multimodal fusion, enabling feature fusion to accurately match the operating scenario of the hydropower equipment.

[0036] In this embodiment, the model pre-training and fine-tuning optimization module 22 uses a general multimodal model as the initial basis. It first performs pre-training constraint optimization by combining objective parameters of hydropower equipment operation, and then performs fine-tuning and iterative optimization using historical fault and maintenance data of core equipment in the hydropower plant, outputting a scenario-adapted multimodal causal constraint analysis model for hydropower equipment. The process of the model pre-training and fine-tuning optimization module 22 outputting the multimodal causal constraint analysis model for hydropower equipment includes the following steps: S22.1 Constructing a set of objective constraint parameters ,in The rated speed of the equipment, Design the water head for the equipment, Rated power of the equipment The permissible vibration threshold for the equipment; Normalization is performed to obtain the normalized set of constraint parameters, denoted as . ;based on Construct the correlation mapping matrix between constraint parameters and model weights. The pre-trained model weight parameter set is generated through parameter association logic. This pre-trained model weight parameter set is denoted as... The parameter correction amount is dynamically determined based on the original weight parameter set of the general multimodal model, which is denoted as . ; S22.2 Constructing a fine-tuning dataset ,in For the first The fused feature vector of each sample, For the first Fault labels for each sample; The dataset is divided into two parts according to a reasonable ratio: a dataset for model training and a dataset for model performance validation. The dataset for model training is denoted as... The dataset used for model performance validation is denoted as . ; S22.3, using the cross-entropy loss function To optimize the objective, For model weight parameters The prediction error loss function is calculated; the model weight parameters are updated using the stochastic gradient descent algorithm. The model weight parameters after rounds of iteration are denoted as During the iteration process, the parameter update step size is dynamically adjusted. The parameter update step size of the round of iteration is denoted as ; Calculation based on validation set loss function value ;when The iteration stops when the system reaches a stable state, and the output is a multimodal causal constraint analysis model of hydropower equipment with final optimization parameters. This is the final set of weight parameters for the model, determined after pre-training and fine-tuning.

[0037] Specifically, in the pre-training phase: first, define the set of objective constraint parameters. This set contains the core operational constraint parameters of hydropower equipment, as detailed below: right The set of normalized constraint parameters is obtained by applying Min-Max normalization. The calculation formula is as follows: ; In the formula, This represents the set of minimum constraint parameters for similar hydropower equipment (with the same model and specifications as the equipment being modeled). ), This represents the set of maximum constraint parameters for similar hydropower equipment (e.g., ); And construct the association mapping matrix (dimension is) , (This represents the total dimension of the weight parameters in a general multimodal model). Initialize using a random normal distribution (mean 0, standard deviation 0.01); The pre-trained model weight parameter set is generated using the following formula. : ; In the formula, This represents the original set of weight parameters of a general multimodal model (i.e., the initial weights that have not been pre-trained). This represents the parameter correction coefficient (with a value of 0.001, used to control the degree of influence of the constraint parameters on the weights). express The transpose matrix (dimension 1) ).

[0038] Specifically, in the fine-tuning phase: first, a fine-tuning dataset is constructed. This dataset contains fused feature vectors and corresponding fault labels, as detailed below: ; In the formula, This represents the total number of samples in the fine-tuned dataset (example: =10000), Indicates the first The unified fusion feature vector of each sample (compared to the output of module 21) (Consistent format) Indicates the first Fault labels for each sample ( =1 indicates that the device is faulty. =0 indicates that the equipment is operating normally); and in a ratio of 8:2, Divided into training set With the validation set The division formula is as follows: ; Specifically, using the cross-entropy loss function To optimize the objective, the calculation formula is as follows: ; In the formula, This represents the set of model weight parameters for the current iteration. express Number of samples ( =0.8 ), This represents the model's probability of predicting faults for feature vector F (the calculation logic is as follows). , This represents the sigmoid activation function. (This represents the logits value output by the model). Represents the natural logarithm function; The model weights are updated using the stochastic gradient descent (SGD) algorithm, and the update formula is as follows: ; In the formula, Indicates the first The set of model weight parameters for each iteration. Indicates the first The set of model weight parameters for each iteration. Indicates the first Learning rate (initial value) for each iteration =0.001, decaying to 0.9 times the current value every 10 iterations). Represents the loss function exist The gradient at that point (calculated using the backpropagation algorithm); Calculate the validation set loss (i.e., the model in) The average cross-entropy loss (on the above) is calculated using the following formula: ; In the formula, express Number of samples ( ); When 5 consecutive iterations satisfy When the model loss stabilizes, the iteration stops, and the final optimized parameters are output. ;in Indicates the first The validation set loss of the round, Indicates the first The validation set loss of the round; This represents the set of model weight parameters at the point where iteration stops.

[0039] Understandably, the learning rate decays to 0.9 times its current value every 10 iterations. This is based on the engineering practice requirements for training intelligent diagnostic models in the "DL / T2561-2022 Technical Guidelines for Condition-Based Maintenance Assessment of Vertical Hydropower Generators," and is a common parameter setting method for model training in the field of condition-based maintenance of hydropower equipment. Five consecutive iterations satisfy the requirements. The stopping condition refers to the accuracy requirement of "equipment condition evaluation accuracy rate ≥ 95%" in the "DL / T2561-2022 Technical Guidelines for Condition-Based Maintenance Assessment of Vertical Hydro Generators". This threshold can ensure that the model performance meets industry standards, facilitate the judgment of parameter rationality and adjustment according to the scenario.

[0040] Traditional approaches to applying large-scale industrial models to hydropower equipment diagnosis suffer from two major shortcomings: First, they directly reuse general multimodal models without incorporating the physical operational constraints of the hydropower equipment, potentially leading to model inferences that contradict equipment mechanisms. For example, they might ignore the constraint relationship between the turbine's design head and the guide vane opening, resulting in outputs that do not conform to the equipment's operational logic. Second, the fine-tuning stage lacks labeled data of real hydropower plant faults, relying on general equipment fault data for training, resulting in low accuracy in identifying hydropower-specific faults. The model pre-training and fine-tuning optimization module 22, based on a general multimodal model, first performs pre-training constraint optimization by combining objective parameters such as the hydropower equipment's rated speed, design head, rated power, and allowable vibration threshold, enabling the model to understand the hydropower equipment's operational mechanisms from a fundamental level. Then, it uses historical fault and maintenance data from core hydropower plant equipment for fine-tuning and iterative optimization, achieving a dual optimization approach of "physical mechanism constraints + real fault data-driven optimization." This technical approach solves the problems of "mechanism disconnect and data distortion" in general large models in hydropower scenarios. The trained multimodal causal constraint analysis model for hydropower equipment not only conforms to the physical operation logic of the equipment, but also has the ability to accurately identify hydropower-specific faults.

[0041] In this embodiment, the causal chain verification and backtracking module 23 quantifies the inference output of the verification model through the causal chain mechanism of hydropower equipment failure. When a contradiction exists, the feature layer backtracking and re-inference process is triggered. The process of the causal chain verification and backtracking module 23 to complete the inference output verification and backtracking includes the following steps: S23.1 Typical causal relationship of built-in water and electricity equipment ,in This is a set of typical causal relationships for hydroelectric equipment failures. For the number of causal relationship categories, For the first The set of antecedent features for causal relationships For the first A set of consequence features for causal relationships, with each type of causal relationship configured with an inherent technical correlation strength coefficient. ; Typical causal relationships for hydroelectric equipment This embodiment provides the following example: ='Vibration amplitude of the main shaft bearing of the hydro-generator > 0.15mm' ='Abnormal wear of main shaft bearing', this causal relationship comes from the typical analysis of a hydropower station operation and maintenance case in "Dynamic Problems and Fault Diagnosis Principles and Methods of Hydropower Generator Sets"; ='Turbine guide vane opening deviation > 5%' ='Turbine efficiency reduction fault' is a fault scenario from the appendix of the industry standard "DL / T2561-2022 Technical Guidelines for Condition-Based Maintenance Assessment of Vertical Hydro Generators"; at the same time" =25 groups, covering faults in core components such as main shaft, turbine, and generator, which are based on typical fault diagnosis scenarios defined in "DL / T2561-2022 Technical Guidelines for Condition-Based Maintenance Assessment of Vertical Hydro-generators".

[0042] S23.2, Using the cosine similarity calculation function The degree of matching between the quantified fusion feature vector and the antecedent feature set is specifically achieved through... Calculate the antecedent matching degree, based on the antecedent matching degree and the inherent technical correlation strength coefficient. A quantitative indicator of the overall matching between computational model inference results and causal chain rules ;like If the preset reasonable threshold is not reached, the inference output is determined to be contradictory. S23.3, Feature Dimension Set with Insufficient Matching Degree between Location and Causal Chain Rules The dimension of a single conflict feature is ; S23.4. Recall the multimodal feature extraction module 21, and use the enhancement algorithm to re-extract features from the original standardized data corresponding to the conflict dimension, to obtain the enhanced secondary fusion feature vector. The enhancement algorithms include wavelet packet transform overlay of time-series data, attention mechanism overlay of image data, and keyword weight enhancement overlay of text data; Input with final optimization parameters The multimodal causal constraint analysis model of hydropower equipment is used for secondary inference; S23.5. Repeat steps S23.2 to S23.4 until the causal relationship is determined. If the requirements are met or a reasonable number of backtracking attempts are reached, a conflict warning message will be output.

[0043] Specifically, in the inference output verification stage: cosine similarity calculation is used. and Antecedent matching degree ,because For a set of vibration-related features, it is necessary to start from... Extract the corresponding vibration feature subvector (Right now The first 128 dimensions of features, and (The dimensions are consistent, with the first 128 dimensions representing vibration time-series features), and the calculation formula is as follows: ; In the formula, This represents the vector dot product operation. express The transpose of , The L2 norm of a vector (calculated logically as follows) , For vectors The (elements) The value range is [-1, 1], and the closer the value is to 1, the higher the matching degree. The overall matching quantitative index is calculated using the following formula. : ; In the formula, The value range is [-1, 1], and the preset reasonable threshold is 0.5. The judgment model's inference output contradicts the causal chain rules.

[0044] Specifically, in the backtracking and re-inference phase: First, locate the conflict feature set. ,calculate and Absolute value of the difference between features of each dimension ,in for The One element, for The (element 1), and the top 20 features with the largest absolute differences are used to construct the feature set. ; Subsequently, the multimodal feature extraction module 21 was called again to perform... The corresponding original standardized data are re-extracted using an enhancement algorithm, as follows: Four layers of wavelet packet transform (with the wavelet basis function still db4) are superimposed on the vibration time series data to generate enhanced vibration features. ; An enhanced image feature is generated by overlaying channel attention mechanisms (channel weight allocation via the SENet module) onto equipment appearance inspection image data. ; Enhanced text features are generated by overlaying keyword weights onto oil chromatography detection text data (multiplying the word embedding weights of fault-related keywords such as "hydrogen," "acetylene," and "temperature anomaly" by 1.2). ; The feature normalization and dynamic weight fusion process is re-executed to obtain the secondary fused feature vector. ;Will Input with final optimization parameters The multimodal causal constraint analysis model of hydropower equipment is used for secondary inference; Finally, repeat the above verification process until... (Experience threshold in the field of hydroelectric equipment fault diagnosis) or the number of backtracking iterations reaches 3 (a common number of iterations used in engineering practice to balance reasoning efficiency and accuracy); if after 3 backtracking iterations... If the threshold is still not reached, a conflict warning message will be sent to the hydropower plant's operation and maintenance system, prompting manual intervention for verification.

[0045] Traditional artificial intelligence technology often exhibits "black box" reasoning characteristics in the fault diagnosis of hydropower equipment: the model only outputs fault conclusions and cannot explain the causal logic of the fault; moreover, when the reasoning is biased, it is difficult to trace back to pinpoint the problem as a data or feature-level issue, resulting in insufficient interpretability and traceability of the fault diagnosis. The causal chain verification and backtracking module 23 incorporates a set of typical causal relationships for hydropower equipment faults, through... The quantitative indicators verify the causal rationality of the reasoning conclusions; when contradictions are found during verification, conflict characteristics can be located and a secondary reasoning process can be triggered. The causal chain verification and backtracking module 23 solves the technical defects of traditional models such as "reasoning without causality and difficulty in backtracking misjudgments", enabling hydropower equipment fault diagnosis to have causal interpretability and reasoning backtrackability, meeting the high requirements of the power system for fault diagnosis to be "interpretable, reliable and traceable".

[0046] The full-process closed-loop control unit 3 adopts workflow engine and digital twin physical state binding technology to build a full life cycle process template of equipment that is compatible with power industry technical standards, so as to realize the linkage and closed loop between process and equipment status. In this embodiment, the full-process closed-loop control unit 3 includes a process template construction module 30, a status binding communication module 31, a linkage rule execution module 32, and a closed-loop data synchronization module 33; wherein: The process template construction module 30, based on the visual modeling technology of the workflow engine, transforms the technical standards of the power industry into standardized process node parameters that the engine can recognize, and constructs a digital process template for the entire stage of equipment "technical upgrade procurement - operation and maintenance - decommissioning". The digital process template supports parameterized adjustment through technical interfaces, without the need for manual redefinition of the process framework. Specifically, the process template construction module 30 is based on the visual modeling technology of the workflow engine. It uses the industrial-grade open-source workflow engine Flowable 6.7.0 (compatible with the BPMN 2.0 standard). Through its built-in FlowableModeler visual modeling tool, users can select basic components such as "start node", "approval node" and "execution node" from the component library by dragging and dropping, and configure the logical order of nodes according to the entire life cycle of the hydropower equipment. Meanwhile, for power industry technical standards such as DL / T5161.1-2021 and the "Example of Equipment Decommissioning Management Standard", a three-step method of "clause decomposition - parameter mapping - standardized definition" is adopted to transform them into process node parameters. For example, "the vibration value of the turbine main shaft needs to be tested every 6 months" is decomposed into parameters that the engine can recognize, such as "testing cycle = 180 days" and "testing parameter = main shaft vibration value". And each node is configured with unified parameter fields such as "node ID", "executing subject" and "associated equipment parameters" (stored in JSON format). In addition, the digital process template supports parameterized adjustments via the RESTAPI interface. When making adjustments, the modified parameters are passed in via a POST request, and the engine automatically updates and generates a new version. It supports batch adjustment of parameters for the same type of nodes and retains historical versions, eliminating the need for manual redefinition of the process framework.

[0047] The state binding communication module 31 adopts the OPCUA industrial communication protocol to establish a two-way data channel between the workflow engine and the dynamic digital twin of the digital twin modeling unit 1. Through the real-time data subscription-push mechanism, the workflow engine can obtain the physical state parameters of the equipment in the digital twin in milliseconds and transmit the process execution data back to the digital twin in real time. Specifically, the state-binding communication module 31 adopts the OPCUA industrial communication protocol. The OPCUA server is deployed on the industrial server where the dynamic digital twin is located (sharing hardware with the digital twin modeling unit 1). The OPCUA server session timeout is set to 30 seconds, and the data cache threshold is set to 10MB. The OPCUA client is integrated into the workflow engine system. The client session timeout is consistent with the server, and the data cache size is set to 5MB. Node identifiers are defined according to the hierarchy of "equipment type - parameter category - parameter name" (such as "hydro turbine - vibration parameters - main shaft vibration amplitude", "generator - temperature parameters - stator winding"). The system employs OPCUA standard encryption strategies and certificate authentication to ensure communication security for the "group temperature" data. When establishing a two-way data channel, the workflow engine pushes process execution data (including "current node ID," "execution status," etc., triggered when node status changes) to the server through the OPCUA client. Simultaneously, it obtains device parameters through a "data change-triggered push + timed retrieval" mechanism—the server actively pushes data when parameter changes exceed ±5%, and the client retrieves data every 100ms to ensure millisecond-level timeliness. All interactive data uses OPCUA standard data types (numerical, status, and timestamp types) to ensure parsing consistency.

[0048] The linkage rule execution module 32 has a built-in logical judgment algorithm based on the physical status parameters of the device. When the device status parameters fed back by the digital twin meet the preset technical threshold, the algorithm automatically triggers the process node jump instruction of the workflow engine without manual intervention in process switching. Specifically, the linkage rule execution module 32 has a built-in logical judgment algorithm based on the physical state parameters of the device, which is implemented using the Drools 7.59.0 rule engine. The rule file (including the "condition part When" and the "action part Then", the condition association digital twin parameters and preset thresholds, and the action definition process jump instructions) is written in the DRL language. The rule file is stored in the workflow engine configuration directory and is automatically loaded into the rule library when the module starts and supports hot updates (automatically reload the modified rules within 10 seconds). Meanwhile, the linkage rule execution module 32 obtains the latest device parameters from the status binding communication module 31 every 100ms, encapsulates them into a device parameter object and passes it into the rule base. The Drools engine traverses the rule matching conditions and sends a jump instruction to the workflow engine when the conditions are met. The preset technical threshold is determined according to the power industry standards and equipment manuals. The threshold is stored in the "threshold_config" table of the MySQL 8.0 database, which contains fields such as "threshold ID", "parameter name" and "standard source". The threshold can be modified through a visual interface and is synchronized to the rule engine in real time after modification.

[0049] The closed-loop data synchronization module 33 synchronizes the process execution data to the dynamic digital twin in real time through the time-series database writing interface, and updates the simulation parameters of the dynamic digital twin. At the same time, based on the equipment status change rate fed back by the dynamic digital twin, the execution time limit parameters of the process nodes are dynamically optimized through the priority adjustment algorithm, forming a technical-level linkage closed loop of "equipment physical status - process technical parameters - digital twin simulation parameters".

[0050] Specifically, the closed-loop data synchronization module 33 realizes data synchronization and parameter optimization through the time-series database writing interface, selects InfluxDB2.1 as the time-series database and configures the corresponding HTTPAPI interface; the module obtains process data such as "node execution time" and "acceptance result" from the workflow engine every 50ms, writes it into the database in the format of "timestamp-device ID-process parameter-parameter value", and calls the interface of the digital twin modeling unit 1 to synchronize the simulation parameters corresponding to the process data (such as the vibration value of the twin spindle after maintenance) to the dynamic digital twin.

[0051] Meanwhile, based on the equipment status change rate fed back by the dynamic digital twin, a priority adjustment algorithm is used to optimize the execution time limit of process nodes—when the change rate exceeds a preset value (e.g., 0.003 mm / s). 2 When the time limit of the corresponding node is shortened by 30% through the workflow engine API (e.g., 72h to 50.4h), a closed loop of linkage is formed, which includes "device status - process parameter adjustment - twin parameter update - feedback of device status", to ensure that the process and device status are matched in real time.

[0052] Intelligent Decision Output Unit 4 integrates the dynamic state of the digital twin with the reasoning conclusions of the multimodal causal constraint analysis model of hydropower equipment, and combines the technical standards of hydropower equipment to derive decision-making schemes for equipment maintenance, decommissioning and technical upgrading.

[0053] In this embodiment, the intelligent decision output unit 4 includes a multi-source decision data fusion module 40, a technical standard parameterization adaptation module 41, a multi-objective decision derivation module 42, and a decision scheme verification module 43; wherein: The multi-source decision data fusion module 40 uses a feature-level fusion algorithm to associate and integrate the dynamic state data of the digital twin output by the digital twin modeling unit 1 with the reasoning conclusions of the multimodal causal constraint analysis model of hydropower equipment in the large model empowerment analysis unit 2. Abnormal data is removed through a data consistency verification algorithm to generate a decision input dataset in a unified format. Specifically, the multi-source decision data fusion module 40 adopts a feature-level fusion algorithm of "feature splicing + attention weighting" to associate and integrate the dynamic state data of the digital twin of the digital twin modeling unit 1 with the reasoning conclusions of the hydropower equipment multimodal causal constraint analysis model of the large model empowerment analysis unit 2, and generates a decision input dataset in a unified format after data consistency verification.

[0054] Meanwhile, in terms of input data acquisition, dynamic status data is pulled in real time from digital twin modeling unit 1 via the REST API interface (update frequency 50ms). The data fields include the unique identifier of the equipment, real-time operating parameters (main shaft vibration amplitude, stator winding temperature, guide vane opening, unit speed), cumulative running time, and historical fault records (including fault type, occurrence time, and processing result). The data format is JSON. The inference conclusion is obtained through the model inference result output interface (HTTP POST request) of unit 2 (update frequency 100ms). The fields include the unique identifier of the equipment (consistent with the twin data), fault type, fault risk level, fault impact range, and fault association parameters (recording the core equipment parameters that caused the fault). The data format is also JSON.

[0055] Furthermore, the feature-level fusion implementation steps are as follows: Feature extraction and standardization: The numerical real-time operating parameters in the twin dynamic state data are processed using Min-Max standardization, based on the normal operating parameter range specified in the technical standards for hydropower equipment. The processed data generates a 4-dimensional feature vector. ; in, Standardized value of spindle vibration amplitude Standardized values ​​for stator winding temperature Standardized value of guide vane opening Standardized values ​​for unit speed; conversion of cumulative operating time into standardized coefficients. (Based on a 30-year equipment design life), historical fault records are quantified into fault frequency characteristics in the range of 0-1 based on the number of faults in the past 3 years. This forms a 2D auxiliary feature vector: ; In the multimodal causal constraint analysis model for hydropower equipment, the risk level in the inference conclusion is directly mapped to a 0-1 interval value (Level 1 = 0.2, Level 2 = 0.4, ..., Level 5 = 1.0). The scope of the fault impact is quantified according to the number of affected components (single component only = 0.2, multiple components = 0.6, whole machine = 1.0). The fault correlation parameters are based on the correlation strength with the fault (model output). After normalization, features are generated, forming a 3D inference feature vector: ; in, Standardized values ​​representing the level of fault risk; A standardized value representing the extent of the fault's impact; This represents the standardized value of the correlation strength between the fault correlation parameters and the fault. Feature splicing: Will and By concatenating them in sequence, a 9-dimensional fused feature vector is obtained. ; Attention weighting: A single-hidden-layer fully connected network (9 input dimensions, 18 hidden layer dimensions, 9 output dimensions, and ReLU activation function) was constructed. This network was trained using historical decision-making data from hydropower equipment (containing 1000 valid samples, each sample including concatenated features and manually labeled feature importance weights) to learn the attention weight coefficients for each feature. (Weights sum to 1), weighted fusion features are calculated via matrix multiplication. ;in Represents the set of weight coefficients The transpose of .

[0056] Furthermore, data consistency verification employs a dual mechanism of "threshold judgment + logical association": In the threshold judgment stage, abnormal data that exceed the [0,1] interval after standardization or whose original parameters exceed the allowable range of the technical standard are removed. In the logical association stage, the fault type in the reasoning conclusion is verified to match the anomaly of the twin parameters (e.g., if the reasoning conclusion is "bearing wear," it must correspond to an abnormal spindle vibration amplitude; otherwise, it is considered logically contradictory data). Abnormal data is processed through "label removal + interpolation completion" (if anomalies are obtained three times consecutively, the data set is removed, and the moving average of the first five valid data sets is used for completion). The final generated decision input dataset is stored in JSON format, containing the device's unique identifier and fused feature vector. The original valid data snapshot and data verification result fields are stored in the time series database for subsequent modules to call.

[0057] The technical standard parameterization adaptation module 41 is used to convert the technical standards of hydropower equipment into calculable decision constraint parameters, form a standard parameter library, and support dynamic updates of standard parameters through technical interfaces. Specifically, the technical standard parameterization adaptation module 41, for hydropower equipment technical standards such as "DL / T5161.1-2021" and "Example of Equipment Decommissioning Management Standard", adopts the "clause decomposition-parameter mapping-quantification definition" method to transform the qualitative requirements and quantitative indicators in the standards into calculable decision constraint parameters (such as maintenance interval ≥180 days, decommissioning period ≤30 years, vibration amplitude after technical upgrade ≤0.15mm / s), forming a standard parameter library and storing it in a MySQL database; it supports dynamic updating of parameters through the RESTAPI technical interface, and the parameters are synchronized to the parameter library in real time after format verification during the update to ensure that the constraint parameters are consistent with the latest technical standards.

[0058] The multi-objective decision derivation module 42 has a built-in multi-objective optimization algorithm. Combined with the input dataset of the multi-source decision data fusion module 40 and the constraint parameters of the technical standard parameterization adaptation module 41, it automatically derives candidate decision schemes for equipment maintenance, decommissioning and technical upgrading, and outputs the technical indicators corresponding to the candidate decision schemes. Specifically, the multi-objective decision derivation module 42 incorporates the NSGA-II multi-objective optimization algorithm, with a population size of 100, a crossover probability of 0.8, a mutation probability of 0.05, and 200 iterations. It aims to optimize for "lowest maintenance cost, shortest downtime, and highest equipment reliability." Combining the input dataset from the multi-source decision data fusion module 40 with the constraint parameters from the technical standard parameterization adaptation module 41, it automatically derives 3-5 candidate decision schemes (including maintenance, decommissioning, and technical upgrade schemes) and outputs the corresponding technical indicators for each scheme (such as cost, duration, and reliability improvement rate for maintenance schemes, residual value recovery rate for decommissioning schemes, and return on investment for technical upgrade schemes).

[0059] The decision scheme verification module 43 is used to input candidate decision schemes into the dynamic digital twin for simulation and to verify whether the state changes of the digital twin after the implementation of the scheme meet the technical standard requirements. If the simulation results meet the preset compliance conditions, the final decision scheme is output. If not, the results are fed back to the multi-objective decision derivation module 42 to readjust and optimize the parameters until a compliant scheme is generated.

[0060] Specifically, the decision scheme verification module 43 inputs the candidate decision scheme into the dynamic digital twin for virtual operation by calling the simulation operation interface of the digital twin modeling unit 1; using the constraint parameters of the technical standard parameterization adaptation module 41 as compliance conditions, it verifies whether the state change of the twin after the implementation of the scheme meets the requirements; if the simulation result is compliant, the final decision scheme is output; if it is non-compliant, the non-compliant parameters (such as the vibration amplitude exceeding the standard after technical modification) are fed back to the multi-objective decision derivation module 42, and the parameters are adjusted and optimized before re-deriving until a compliant scheme is generated.

[0061] Those skilled in the art will understand that the process of implementing all or part of the steps of the above embodiments can be carried out by hardware or by a program instructing the relevant hardware.

[0062] The foregoing has shown and described the basic principles, main features, and advantages of the present invention. Those skilled in the art should understand that the present invention is not limited to the above embodiments. The embodiments and descriptions in the specification are merely preferred examples and are not intended to limit the invention. Various changes and modifications can be made to the invention without departing from its spirit and scope, and all such changes and modifications fall within the scope of the present invention as claimed. The scope of protection of the present invention is defined by the appended claims and their equivalents.

Claims

1. A large-scale model-enabled digital twin decision-making system for the entire lifecycle of equipment, characterized in that: include: Digital twin modeling unit (1), the digital twin modeling unit (1) adopts multi-source heterogeneous data fusion and physical-digital bidirectional linkage technology to construct a dynamic digital twin of hydropower equipment in the whole stage of "technical transformation and procurement - operation and maintenance - decommissioning and disposal", so as to realize the real-time synchronization of parameters between the digital twin and the physical equipment; The large model empowerment analysis unit (2) adopts the hydropower equipment multimodal causal constraint analysis model. The hydropower equipment multimodal causal constraint analysis model combines the historical operation data of the equipment to carry out intelligent reasoning. It extracts the fault and operation status related exclusive features from the hydropower equipment exclusive data. In the pre-training stage, the general multimodal model is used as the initial basis. The objective parameters of the hydropower equipment operation are combined with the correlation optimization to form the pre-trained version of the hydropower equipment multimodal causal constraint analysis model. In the fine-tuning stage, the parameters of the hydropower equipment multimodal causal constraint analysis model are iteratively optimized using the historical fault and maintenance data of the core equipment of the hydropower plant. The reasoning output of the hydropower equipment multimodal causal constraint analysis model is detected through the hydropower equipment fault causal chain mechanism. When there is a contradiction, it is backtracked to the feature layer for re-inference. The full-process closed-loop control unit (3) adopts the workflow engine and digital twin physical state binding technology to build a full life cycle process template of equipment that is compatible with the technical standards of the power industry, so as to realize the linkage and closed loop between process and equipment status; The intelligent decision output unit (4) integrates the dynamic state of the digital twin with the reasoning conclusions of the multimodal causal constraint analysis model of hydropower equipment, and combines the technical standards of hydropower equipment to deduce decision schemes for equipment maintenance, decommissioning and technical transformation.

2. The large-model-enabled digital twin decision-making system for the entire lifecycle of equipment as described in claim 1, characterized in that, The digital twin modeling unit (1) includes a multi-source heterogeneous data fusion module (10). The multi-source heterogeneous data fusion module (10) is used to access real-time sensor data of hydropower equipment, equipment design drawing data, operation and maintenance record text data and historical fault data. After the accessed data is processed by removing outliers, standardizing data format and filling missing values, a weighted fusion algorithm is used to allocate weights according to the credibility of different types of data, and the preprocessed multi-source data is fused into a digital twin in a unified format to construct the basic data.

3. The large-model-enabled digital twin decision-making system for the entire lifecycle of equipment as described in claim 2, characterized in that, The digital twin modeling unit (1) also includes a physical-digital bidirectional linkage module (11). The physical-digital bidirectional linkage module (11) acquires the actual operating parameters of the physical body of the hydropower equipment through edge sensors deployed on the physical body of the hydropower equipment at a preset acquisition frequency. Based on the acquired actual operating parameters of the physical body of the hydropower equipment, it updates the simulation parameters of the dynamic digital twin in real time and compares the actual operating parameters of the physical body of the hydropower equipment with the simulation parameters of the dynamic digital twin. When the deviation between the actual operating parameters of the physical body of the hydropower equipment and the simulation parameters of the dynamic digital twin exceeds a preset threshold, the parameter synchronization calibration process is triggered.

4. The large-model-enabled digital twin decision-making system for the entire lifecycle of equipment as described in claim 3, characterized in that, The large model empowerment analysis unit (2) includes a multimodal data adaptation module (20), a multimodal feature extraction module (21), a model pre-training and fine-tuning optimization module (22), and a causal chain verification and backtracking module (23); wherein: The multimodal data adaptation module (20) is used to receive dedicated data from hydropower equipment, perform format standardization, noise filtering and modal consistency processing, and output standardized multimodal data that meets the input requirements of the multimodal causal constraint analysis model of hydropower equipment. The multimodal feature extraction module (21) is used to perform targeted feature extraction, normalization and dynamic weight fusion on standardized multimodal data to form a unified fusion feature vector related to faults and operating states; The model pre-training and fine-tuning optimization module (22) takes the general multimodal model as the initial basis, first combines the objective parameters of hydropower equipment operation for pre-training constraint optimization, and then uses the historical fault and maintenance data of the core equipment of the hydropower plant for fine-tuning and iterative optimization, and outputs a scenario-adapted multimodal causal constraint analysis model of hydropower equipment. The causal chain verification and backtracking module (23) quantifies the verification model inference output through the causal chain mechanism of hydropower equipment failure. When there is a contradiction, it triggers the feature layer backtracking and re-inference process.

5. The large-model-enabled digital twin decision-making system for the entire lifecycle of equipment as described in claim 4, characterized in that, The process by which the multimodal data adaptation module (20) outputs standardized multimodal data that meets the input requirements of the multimodal causal constraint analysis model for hydropower equipment includes the following steps: S20.1 Receive vibration time sequence data of the physical body of the hydroelectric equipment, oil chromatography detection text data, equipment appearance inspection image data, real-time unit speed data and guide vane opening adjustment data. S20.

2. An adaptive median filtering algorithm is used to process the vibration time series data, and the window radius is dynamically adjusted according to the vibration frequency; a Gaussian filtering algorithm is used to process the equipment appearance inspection image data; and invalid characters are removed from the oil chromatography detection text data through regularization. S20.3 The time-series data includes vibration time-series data, real-time unit speed data, and guide vane opening adjustment data after adaptive median filtering. Missing values ​​are filled in using linear interpolation and the sampling frequency is standardized. The equipment appearance inspection image data after Gaussian filtering is standardized in resolution. The final output is a standardized multimodal data set containing processed vibration time-series data, processed oil chromatography detection text data, processed equipment appearance inspection image data, standardized real-time unit speed data, and standardized guide vane opening adjustment data, providing suitable input for subsequent feature extraction.

6. The large-model-enabled digital twin decision-making system for the entire lifecycle of equipment as described in claim 5, characterized in that, The process by which the multimodal feature extraction module (21) generates a unified fused feature vector includes the following steps: S21.1 The processed vibration time series data is used to extract time-frequency domain features through wavelet transform. The processed oil chromatography detection text data was used to extract feature vectors using the BERT word embedding algorithm. Spatial feature vectors were extracted from the processed equipment appearance inspection image data using a ResNet50 network. Features were extracted from real-time speed data of standardized units and standardized guide vane opening adjustment data through a sliding window statistical method. ; S21.2, All unimodal features are normalized using Min-Max. Mapping to a unified interval eliminates dimensional differences; S21.3, Based on the operating condition adaptation factor of hydropower equipment Assign modal weights And satisfy Through multimodal feature fusion Integrate all normalized features to output a unified fused feature vector. ;in When in the low load range, the focus is on vibration and image features, while when in the high load range, the focus is on text and rotational speed opening features.

7. The large-model-enabled digital twin decision-making system for the entire lifecycle of equipment as described in claim 6, characterized in that, The process of the model pre-training and fine-tuning optimization module (22) outputting the multimodal causal constraint analysis model of hydropower equipment includes the following steps: S22.1 Constructing a set of objective constraint parameters ,in The rated speed of the equipment, Design the water head for the equipment, Rated power of the equipment The permissible vibration threshold for the equipment; Normalization is performed to obtain the normalized set of constraint parameters, denoted as . ;based on Construct the correlation mapping matrix between constraint parameters and model weights. The pre-trained model weight parameter set is generated through parameter association logic. This pre-trained model weight parameter set is denoted as... The parameter correction amount is dynamically determined based on the original weight parameter set of the general multimodal model, which is denoted as . ; S22.2 Constructing a fine-tuning dataset ,in For the first The fused feature vector of each sample, For the first Fault labels for each sample; The dataset is divided into two parts according to a reasonable ratio: a dataset for model training and a dataset for model performance validation. The dataset for model training is denoted as... The dataset used for model performance validation is denoted as . ; S22.3, using the cross-entropy loss function To optimize the objective, For model weight parameters The prediction error loss function is calculated; the model weight parameters are updated using the stochastic gradient descent algorithm. The model weight parameters after rounds of iteration are denoted as During the iteration process, the parameter update step size is dynamically adjusted. The parameter update step size of the round of iteration is denoted as ; Calculation based on validation set loss function value ;when The iteration stops when the system reaches a stable state, and the output is a multimodal causal constraint analysis model of hydropower equipment with final optimization parameters. This is the final set of weight parameters for the model, determined after pre-training and fine-tuning.

8. The large-model-enabled digital twin decision-making system for the entire lifecycle of equipment as described in claim 7, characterized in that, The process of the causal chain verification and backtracking module (23) to complete the inference output verification and backtracking includes the following steps: S23.1 Typical causal relationship of built-in water and electricity equipment ,in This is a set of typical causal relationships for hydroelectric equipment failures. For the number of causal relationship categories, For the first The set of antecedent features for causal relationships For the first A set of consequence features for causal relationships, with each type of causal relationship configured with an inherent technical correlation strength coefficient. ; S23.2, Using the cosine similarity calculation function The degree of matching between the quantified fusion feature vector and the antecedent feature set is specifically achieved through... Calculate the antecedent matching degree, based on the antecedent matching degree and the inherent technical correlation strength coefficient. A quantitative indicator of the overall matching between computational model inference results and causal chain rules ;like If the preset reasonable threshold is not reached, the inference output is determined to be contradictory. S23.3, Feature Dimension Set with Insufficient Matching Degree between Location and Causal Chain Rules The dimension of a single conflict feature is ; S23.

4. Recall the multimodal feature extraction module (21) and use the enhancement algorithm to re-extract features from the original standardized data corresponding to the conflict dimension to obtain the enhanced secondary fusion feature vector. The enhancement algorithm includes wavelet packet transform overlay of time-series data, attention mechanism overlay of image data, and keyword weight enhancement overlay of text data; Input with final optimization parameters The multimodal causal constraint analysis model of hydropower equipment is used for secondary inference; S23.

5. Repeat steps S23.2 to S23.4 until the causal relationship is determined. If the requirements are met or a reasonable number of backtracking attempts are reached, a conflict warning message will be output.

9. The large-model-enabled digital twin decision-making system for the entire lifecycle of equipment as described in claim 8, characterized in that, The full-process closed-loop control unit (3) includes a process template construction module (30), a status binding communication module (31), a linkage rule execution module (32), and a closed-loop data synchronization module (33); wherein: The process template construction module (30) is based on the visualization modeling technology of the workflow engine. It transforms the power industry technical standards into standardized process node parameters that the engine can recognize, and constructs a digital process template for the entire stage of equipment "technical transformation and procurement - operation and maintenance - decommissioning". The digital process template supports parameterized adjustment through technical interfaces without the need for manual redefinition of the process framework. The state binding communication module (31) adopts the OPCUA industrial communication protocol to establish a two-way data channel between the workflow engine and the dynamic digital twin of the digital twin modeling unit (1). Through the real-time data subscription-push mechanism, the workflow engine can obtain the physical state parameters of the equipment in the digital twin in milliseconds and transmit the process execution data back to the digital twin in real time. The linkage rule execution module (32) has a built-in logical judgment algorithm based on the physical state parameters of the device. When the device state parameters fed back by the digital twin meet the preset technical threshold, the algorithm automatically triggers the process node jump instruction of the workflow engine without manual intervention in process switching. The closed-loop data synchronization module (33) synchronizes the process execution data to the dynamic digital twin in real time through the time-series database writing interface, and updates the simulation parameters of the dynamic digital twin. At the same time, based on the equipment status change rate fed back by the dynamic digital twin, the execution time limit parameters of the process node are dynamically optimized through the priority adjustment algorithm, forming a technical-level linkage closed loop of "equipment physical status - process technical parameters - digital twin simulation parameters".

10. The large-model-enabled digital twin decision-making system for the entire lifecycle of equipment according to claim 9, characterized in that, The intelligent decision output unit (4) includes a multi-source decision data fusion module (40), a technical standard parameterization adaptation module (41), a multi-objective decision derivation module (42), and a decision scheme verification module (43); wherein: The multi-source decision data fusion module (40) adopts a feature-level fusion algorithm to integrate the dynamic state data of the digital twin output by the digital twin modeling unit (1) with the reasoning conclusions of the hydropower equipment multimodal causal constraint analysis model in the large model empowerment analysis unit (2), and removes abnormal data through the data consistency verification algorithm to generate a decision input dataset in a unified format. The technical standard parameterization adaptation module (41) is used to convert the technical standards of hydropower equipment into calculable decision constraint parameters, form a standard parameter library, and support dynamic updates of standard parameters through technical interfaces. The multi-objective decision derivation module (42) has a built-in multi-objective optimization algorithm. Combined with the input dataset of the multi-source decision data fusion module (40) and the constraint parameters of the technical standard parameterization adaptation module (41), it automatically derives candidate decision schemes for equipment maintenance, decommissioning and technical transformation, and outputs the technical indicators corresponding to the candidate decision schemes. The decision scheme verification module (43) is used to input the candidate decision scheme into the dynamic digital twin for simulation operation, and verify whether the state change of the digital twin after the implementation of the scheme meets the technical standard requirements. If the simulation result meets the preset compliance conditions, the final decision scheme is output. If it does not meet the conditions, it is fed back to the multi-objective decision derivation module (42) to readjust and optimize the parameters until a compliant scheme is generated.

Images