690V low-voltage equipment intelligent operation and maintenance management system for LNG (Liquefied Natural Gas) station

Abstract

The invention discloses a 690V low-voltage equipment intelligent operation and maintenance management system for an LNG station, and relates to the technical field of intelligent operation and maintenance, the system comprises a visual management platform, and the visual management platform is in communication connection with the following modules: a digital chassis module, which is used for collecting multi-source heterogeneous data of 690V low-voltage equipment in the LNG station, and transmitting the multi-source heterogeneous data to the LNG station; and constructing an equipment digital twinborn model. By collecting multi-source heterogeneous data and constructing the equipment digital twin model, real-time mapping of the physical state and the digital model of the equipment is achieved, operation and maintenance personnel can visually master the real-time operation state of the equipment through the visual management platform, frequent on-site inspection is not needed, manpower and time cost is greatly saved, and meanwhile, the operation and maintenance efficiency is improved. The fault diagnosis function based on the digital twinborn model can accurately position a fault point, avoids blind troubleshooting caused by incomplete information in traditional operation and maintenance, remarkably improves the accuracy of operation and maintenance, and reduces the number; the downtime and the maintenance cost of equipment are reduced.

Description

Technical Field

[0001] This invention relates to the field of intelligent operation and maintenance technology, specifically to an intelligent operation and maintenance management system for 690V low-voltage equipment in LNG terminals. Background Technology

[0002] With the adjustment of the global energy structure and the continuous growth of natural gas demand, the rapid development of the LNG industry has placed higher demands on energy supply, storage and transportation. As an important link in energy conversion, LNG terminals undertake tasks such as storage, gasification and distribution of liquefied natural gas. In LNG terminals, 690V low-voltage equipment typically includes important electrical facilities such as transformers, distribution cabinets and control systems. In order to improve the operational efficiency and safety of LNG terminals, the management and maintenance of 690V low-voltage equipment has become particularly important.

[0003] In existing technologies, operation and maintenance management systems trigger alarms when thresholds are exceeded, but they lack correlation analysis. Operation and maintenance personnel are overwhelmed by massive amounts of isolated alarm information, making it difficult to distinguish noise from real danger signals. This can easily lead to warning fatigue and overlook truly critical hidden dangers. Therefore, how to construct a digital twin health baseline for equipment, provide early warnings based on the health status of the equipment, and realize the propagation reasoning of early warnings to locate specific components, shorten diagnosis time, and ensure the efficiency of operation and maintenance management of 690V low-voltage equipment is the problem that this invention aims to solve. To this end, an intelligent operation and maintenance management system for 690V low-voltage equipment in LNG terminals is proposed. Summary of the Invention

[0004] To solve the above technical problems, the present invention is implemented through the following technical solution: an intelligent operation and maintenance management system for 690V low-voltage equipment in LNG terminals, including a visual management platform, wherein the visual management platform is communicatively connected to the following modules: The digital chassis module is used to collect multi-source heterogeneous data from 690V low-pressure equipment in LNG terminals and build digital twin models of the equipment to achieve real-time mapping between the physical state of the equipment and the digital model. The health baseline early warning module is used to generate an adaptively adjusted health baseline by utilizing historical operating data of the equipment and the equipment's digital twin model, so as to trigger early correlation warnings; The operation and maintenance knowledge graph construction module is used to build an operation and maintenance knowledge graph containing equipment-component-fault phenomenon-root cause-remedial measures by combining structured and unstructured knowledge, including equipment manuals, fault cases, maintenance procedures and expert experience. The fault propagation reasoning and localization module is used to analyze early correlation warning information, extract the implied fault phenomena as input, traverse and reason in the operation and maintenance knowledge graph, find the fault path that best matches the current multi-parameter anomaly pattern through graph algorithms, deduce the root cause, and locate the specific physical component. The resource scheduling module, based on the fault location results, automatically generates and recommends the optimal maintenance strategy from a pre-established maintenance strategy library that includes preventive maintenance, emergency repairs, and spare parts replacement, while intelligently scheduling maintenance resources.

[0005] Preferably, the digital chassis module includes a multi-source heterogeneous data acquisition unit and a model building unit; The multi-source heterogeneous data acquisition unit is used to access multi-source heterogeneous data from 690V low-voltage equipment in the LNG terminal through IoT gateways and edge computing nodes. This data includes electrical data, status data, environmental data from 690V equipment, and process data from SCADA and DCS systems. The unit also performs preprocessing operations on the multi-source heterogeneous data, including cleaning, alignment, standardization, and labeling, to form a standardized basic dataset. The model building unit constructs a high-precision digital twin model of the equipment based on the physical mechanism of the equipment and the collected historical data, and maps the physical entity's state and behavior in real time.

[0006] Preferably, the multi-source heterogeneous data acquisition unit specifically includes: By collecting multi-source heterogeneous data through IoT gateways and edge computing nodes deployed on 690V low-voltage equipment in LNG terminals, including electrical and condition monitoring data from the 690V low-voltage equipment itself, environmental sensor data, and process data from the upper-level SCADA / DCS system, the preliminary aggregation and protocol parsing of multi-source heterogeneous data is completed. Preprocessing of the multi-source heterogeneous data received includes steps such as cleaning, alignment, standardization, and labeling. The standardized, multi-source heterogeneous data that has undergone comprehensive governance is organized and packaged according to the system-defined format to output a standardized basic dataset.

[0007] Preferably, the model building unit specifically includes: Based on the physical mechanism of the equipment, an initial digital twin model framework for the equipment is constructed. The principles of multiple disciplines, including electrical, thermodynamic and mechanical dynamics, are integrated to define the mathematical relationships of key parameters, including voltage, temperature and vibration. Historical operating data is integrated in sync, and the boundary conditions and dynamic characteristics of the model are corrected through parameter identification to ensure bidirectional calibration driven by mechanism and data. By utilizing collected multi-source heterogeneous data, a physical information neural network (PINN) is used to train the device digital twin model to perform nonlinear mapping of device behavior. Deploy the trained device digital twin model to edge computing nodes, access device sensor data in real time, map the physical entity's state and behavior, and continuously fine-tune the model parameters of the device digital twin model using new data streams to achieve state synchronization and behavior prediction closed loop between the device digital twin model and the physical device.

[0008] Preferably, the health baseline early warning module includes a health baseline self-learning unit and a multi-parameter correlation early warning unit; The health baseline self-learning unit is used to combine the equipment digital twin model and extract the equipment's historical operating data, and autonomously learn the normal operating parameter range of the equipment under different loads, seasons and working conditions through unsupervised learning machine learning algorithms to form an adaptively adjustable health baseline. The multi-parameter correlation early warning unit is used to extract real-time equipment operating parameters and compare them with the generated health baseline. When multiple parameters show an abnormal pattern of co-deviating from the health baseline, the unit uses a clustering algorithm to filter duplicate alarms and triggers an early correlation early warning to focus on key hidden dangers.

[0009] Preferably, the health baseline self-learning unit specifically includes: By combining the equipment digital twin model with historical operating data extracted from the historical database, and aligning it with time stamps and operating condition labels including season and load type, a structured dataset is constructed through noise reduction, normalization, and outlier removal to ensure the quality and consistency of the input data. The DBSCAN density-based clustering algorithm is used to group the data in the structured dataset, automatically identify data clusters under different working conditions, extract key operating parameters, including temperature change rate and vibration spectrum energy, through PCA feature dimensionality reduction, and calculate the statistical boundary (mean, standard deviation) of each working condition cluster to determine the initial health baseline range. Based on the deviation analysis between real-time running data and historical baselines, the cluster centers and boundary parameters are dynamically updated through a sliding window mechanism, and a feedback adjustment mechanism is introduced. When new data continues to deviate from the healthy baseline, the unsupervised model is retrained to form an adaptively adjustable healthy baseline.

[0010] Preferably, the multi-parameter correlation early warning unit specifically includes: Multi-source heterogeneous data of 690V low-voltage equipment is collected in real time through edge computing nodes. After data cleaning and normalization, the Mahalanobis distance between each parameter in the multi-source heterogeneous data and the healthy baseline is calculated. Timestamps and operating condition labels are marked synchronously to generate parameter deviation vectors. Based on the DBSCAN density clustering algorithm, density clustering is performed on the parameter deviation vector to identify abnormal clusters with multiple parameters deviating together. False alarms caused by pseudo-anomalies where a single parameter exceeds the limit but the overall pattern is normal are filtered out, while core abnormal patterns and their associated parameter combinations are retained. The clustering results are statistically validated (chi-square test). When the confidence level of the abnormal cluster is ≥95%, an early warning is triggered, and the parameters and operating context of the key hidden dangers are output.

[0011] Preferably, the operation and maintenance knowledge graph construction module specifically includes: The system performs word segmentation and entity recognition on structured and unstructured knowledge, including equipment manuals, failure cases, maintenance procedures and expert experience, to extract core elements including equipment, components, failure phenomena, root causes and handling measures. It also establishes the relationship between elements through semantic parsing, transforms unstructured knowledge into structured data and stores it in a multi-source knowledge base. Based on the hierarchical relationship of equipment-component-fault phenomenon-root cause-response measures defined by structured data, a knowledge graph framework is constructed through ontology modeling, clarifying the attributes and association rules of each entity, and forming an extensible semantic network model; By integrating multi-source knowledge base data, eliminating redundant information through entity alignment, verifying relationship consistency through logical reasoning, correcting erroneous associations by combining expert feedback, and continuously optimizing the completeness of the graph and the accuracy of reasoning, a standardized operation and maintenance knowledge graph is formed.

[0012] Preferably, the fault propagation reasoning and localization module specifically includes: Receive early correlation warnings from the health baseline warning module, parse the warning information, extract multi-parameter abnormal patterns, use natural language processing technology to extract fault phenomenon descriptions from unstructured alarm logs, and map them to standardized fault phenomenon nodes in the operation and maintenance knowledge graph to form an input set to be reasoned. Starting with the extracted fault phenomena, a depth-first traversal is performed in the operation and maintenance knowledge graph based on the relationship between equipment-component-fault phenomenon-root cause-handling measures. The graph matching algorithm with subgraph isomorphism is applied to analyze the similarity between multi-parameter abnormal patterns and historical fault paths, and the fault propagation chain with the highest degree of consistency with the current fault phenomenon is selected. Tracing back along the matched fault path, the root cause node is determined through causal reasoning. Combining the equipment topology and component hierarchy, the abstract root cause is mapped to a specific physical component, and the location result is output, which includes a description of the root cause, the name of the associated component, and a confidence score.

[0013] Preferably, the resource scheduling module specifically includes: Receive the fault location results output by the fault propagation reasoning and location module, including the root cause description, related component names and confidence scores. From the pre-established operation and maintenance strategy library, preliminarily screen the suitable strategies according to the fault type to form a candidate strategy set, which includes operation and maintenance strategies for preventive maintenance, emergency repair and spare parts replacement. Based on the current maintenance resource status (manpower, spare parts inventory, and equipment availability) and the operation and maintenance cost model (labor costs, spare parts costs, and downtime losses), resource utilization and cost assessments are performed on the candidate strategy set, strategies with insufficient resources or excessive costs are eliminated, and an optimized list of feasible strategies is generated. Based on the urgency of the fault and resource efficiency (shortest completion time), the list of feasible strategies is prioritized and sorted, the optimal maintenance strategy is output and the corresponding resources are scheduled (specifying the maintenance team, spare parts and time window), and the resource status is updated synchronously.

[0014] This invention provides an intelligent operation and maintenance management system for 690V low-voltage equipment in LNG terminals. It offers the following advantages: (I) The intelligent operation and maintenance management system for 690V low-voltage equipment in LNG terminals collects multi-source heterogeneous data and constructs a digital twin model of the equipment to achieve real-time mapping between the physical state of the equipment and the digital model. Operation and maintenance personnel can intuitively grasp the real-time operating status of the equipment through the visual management platform, eliminating the need for frequent on-site inspections and greatly saving manpower and time costs. At the same time, the fault diagnosis function based on the digital twin model can accurately locate the fault point, avoiding blind troubleshooting caused by incomplete information in traditional operation and maintenance, significantly improving the accuracy of operation and maintenance, and reducing equipment downtime and maintenance costs.

[0015] (ii) The 690V low-voltage equipment intelligent operation and maintenance management system for LNG terminals uses historical equipment operation data and digital twin models to generate an adaptively adjusted health baseline. By triggering early correlation warnings, it can promptly detect abnormal patterns in the early stages of equipment failure, provide early warnings of potential failures, buy time for operation and maintenance personnel to handle the situation, take measures in advance, prevent the failure from worsening, and reduce the impact of the failure on the normal operation of the LNG terminal.

[0016] (III) The intelligent operation and maintenance management system for 690V low-voltage equipment in LNG terminals, combined with the operation and maintenance knowledge graph, can quickly deduce the root cause of the fault and locate the specific physical component. Based on the fault location results, it automatically generates and recommends the optimal maintenance strategy from the operation and maintenance strategy library. At the same time, it can intelligently schedule maintenance resources to ensure the scientificity and rationality of the operation and maintenance strategy, optimize resource utilization efficiency, avoid resource waste, and reduce operation and maintenance costs. Attached Figure Description

[0017] Figure 1This is a schematic diagram of the workflow of an intelligent operation and maintenance management system for 690V low-voltage equipment in an LNG terminal according to the present invention. Figure 2 This is a data flow diagram of an intelligent operation and maintenance management system for 690V low-voltage equipment in an LNG terminal, according to the present invention. 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 skilled in the art without creative effort are within the scope of protection of the present invention.

[0019] Example 1, please refer to Figure 1 , Figure 2 This invention provides a technical solution: an intelligent operation and maintenance management system for 690V low-voltage equipment in LNG terminals, including a visual management platform, which has the following communication connections: The digital chassis module is used to collect multi-source heterogeneous data of 690V low-pressure equipment in LNG terminals and build digital twin models of the equipment to realize real-time mapping between the physical state of the equipment and the digital model. The digital chassis module includes a multi-source heterogeneous data acquisition unit and a model building unit. The multi-source heterogeneous data acquisition unit is used to access multi-source heterogeneous data from the 690V low-voltage equipment in the LNG terminal via IoT gateways and edge computing nodes. This data includes electrical data (voltage, current, power, harmonics), status data (temperature, vibration, partial discharge), environmental data (temperature, humidity, salt spray concentration) from the 690V equipment, and process data from SCADA and DCS systems. The unit performs preprocessing operations on the multi-source heterogeneous data, including cleaning, alignment, standardization, and labeling, to form a standardized basic dataset. This dataset is then collected by IoT gateways and edge computing nodes deployed on the 690V low-voltage equipment in the LNG terminal. According to reports, the system encompasses electrical and condition monitoring data from 690V low-voltage equipment, environmental sensor data, and process data from the upper-level SCADA / DCS system. It completes the initial aggregation and protocol parsing of multi-source heterogeneous data, and preprocesses the accessed multi-source heterogeneous data, including cleaning, alignment, standardization, and labeling. Specifically, it cleans outliers and missing values, performs timestamp alignment, and standardizes data formats and dimensions. It then performs structured labeling based on equipment and parameter types. Finally, the comprehensively processed standardized multi-source heterogeneous data is organized and packaged according to the system-defined format, outputting a standardized basic dataset. The specific tasks of the multi-source heterogeneous data acquisition unit are as follows: Through IoT gateways and edge computing nodes deployed on the 690V low-voltage equipment at the LNG terminal, it comprehensively collects electrical data (voltage, current, power, harmonics), status data (temperature, vibration, partial discharge), environmental data (temperature, humidity, salt spray concentration), and process data from SCADA and DCS systems from the 690V equipment. Specifically, the IoT gateways are responsible for real-time access to the electrical parameters and status signals of the equipment itself, while the edge computing nodes integrate time-series data from environmental sensors and process control systems. During data acquisition, it supports the parsing and conversion of various industrial protocols, completing the initial aggregation of multi-source data. The unit then performs systematic preprocessing of the raw multi-source data, using cleaning algorithms to remove outliers and missing values ​​to ensure data reliability. Subsequently, it performs timestamp alignment to unify data from different devices or... The system's sampling frequency and time series benchmark eliminate time deviations. In the data standardization phase, differences in units and format heterogeneity are addressed by converting data into a unified format through mapping rules. Structured tagging is performed based on equipment type and parameter category, assigning searchable semantic identifiers to the data, forming a logically clear and easily analyzable intermediate dataset. The pre-processed multi-source heterogeneous data needs to be organized and encapsulated according to the system's defined specifications. Through data model design, structured tags, time series, and metadata about equipment relationships are embedded into data packets to ensure data traceability and contextual integrity. During encapsulation, a hierarchical storage structure is used to distinguish between real-time and historical data, and multi-dimensional retrieval by equipment, time, and parameter type is supported. This results in a standardized basic dataset stored in a unified format, providing data support for the intelligent operation and maintenance of LNG terminal 690V low-pressure equipment. The model building unit constructs a high-precision digital twin model of the equipment based on its physical mechanism and collected historical data. This model maps the physical entity's state and behavior in real time. It builds an initial digital twin model framework based on the equipment's physical mechanism, integrating principles from multiple disciplines including electrical, thermodynamic, and mechanical dynamics. It defines the mathematical relationships between key parameters such as voltage, temperature, and vibration, and synchronously integrates historical operating data. Through parameter identification, it corrects the model's boundary conditions and dynamic characteristics, ensuring bidirectional calibration between mechanism and data-driven approaches. Utilizing collected multi-source heterogeneous data, it employs a Physical Information Neural Network (PINN) to train the digital twin model's nonlinear mapping capability of the equipment's behavior. Through cross-validation, it optimizes hyperparameters, improving the generalization and robustness of the digital twin model under complex operating conditions. The trained digital twin model is deployed to edge computing nodes, receiving real-time access to equipment sensor data to map the physical entity's state and behavior. New data streams are used to continuously fine-tune the model parameters, achieving state synchronization and behavior prediction closed-loop between the digital twin model and the physical equipment. The specific work of the model building unit is as follows: Based on the physical mechanism of the equipment, it integrates the principles of multiple disciplines such as electrical, thermodynamic and mechanical dynamics to establish an initial digital twin model framework for the equipment. By analyzing the working principle of the equipment, it defines the mathematical relationships of key parameters including voltage, temperature and vibration, forming a mechanism-driven static model structure. Historical operating data is introduced simultaneously, and parameter identification methods are used to correct the boundary conditions and dynamic characteristics of the model, realizing bidirectional calibration between the mechanism model and actual data. The mechanism model provides physical constraints for data analysis to avoid overfitting; historical data provides feedback on the actual behavior of the equipment, corrects the deviation of theoretical assumptions, and ensures that the model has interpretability and basic accuracy in the initial stage. By utilizing collected multi-source heterogeneous data, the nonlinear mapping capability of the device digital twin model is trained using Physical Information Neural Network (PINN). PINN embeds physical equations (Maxwell's equations, heat conduction equations) into the neural network loss function to constrain the training process to conform to physical laws. At the same time, it learns dynamic characteristics under complex working conditions through data-driven learning. During training, the dataset is divided by cross-validation to optimize hyperparameters including the number of network layers and activation functions, balancing model accuracy and computational efficiency, and improving the model's generalization ability in extreme working conditions or data-sparse scenarios. The trained device digital twin model is deployed to edge computing nodes, accessing the device sensor data stream in real time to complete the mapping from the physical entity state to the digital space. Through a real-time data-driven mechanism, the model parameters are continuously fine-tuned to achieve synchronization between the device digital twin model and the physical device state. At the same time, the new data stream is used to update the model's behavior prediction capabilities, forming a closed loop of perception-mapping-prediction-correction: the edge nodes perform real-time assessment of the device's health status and predict potential failure trends; the prediction results are fed back to the device digital twin model, triggering adaptive parameter adjustments to ensure that the model maintains high accuracy during long-term operation under device aging and environmental changes. The health baseline early warning module is used to generate an adaptively adjusted health baseline by utilizing historical operating data of the equipment and the equipment digital twin model, so as to trigger early correlation warnings. The health baseline early warning module includes a health baseline self-learning unit and a multi-parameter correlation warning unit. The health baseline self-learning unit combines the equipment's digital twin model with historical operating data. Through unsupervised machine learning algorithms, it autonomously learns the normal operating parameter ranges of the equipment under different loads, seasons, and operating conditions, forming an adaptively adjustable health baseline. It integrates the equipment's digital twin model with historical operating data extracted from the historical database, aligning it with timestamps and operating condition labels including season and load type. Through denoising, normalization, and outlier removal, a structured dataset is constructed to ensure the quality and consistency of the input data. A DBSCAN-based density-based clustering algorithm is used to group the data in the structured dataset, automatically identifying data clusters under different operating conditions. Key operating parameters, including temperature change rate and vibration spectrum energy, are extracted using PCA feature dimensionality reduction. The statistical boundaries (mean and standard deviation) of each operating condition cluster are calculated to determine the initial health baseline range. Based on the deviation analysis between real-time operating data and historical baselines, the cluster centers and boundary parameters are dynamically updated through a sliding window mechanism. A feedback adjustment mechanism is introduced: when new data continuously deviates from the health baseline, the unsupervised model is retrained to form an adaptively adjustable health baseline. The specific tasks of the health baseline self-learning unit are as follows: integrate the digital twin model of the equipment with historical operating data to build a high-quality structured dataset; use timestamp alignment technology to accurately match the time series data of voltage, temperature and vibration covered in the historical operating data with the operating condition labels (season, load type) to ensure the consistency of multi-source heterogeneous data in the time dimension; use wavelet denoising algorithm to eliminate sensor noise; and combine Z-Score normalization method to unify the data units to avoid analysis bias caused by differences in magnitude. Meanwhile, outliers are removed based on the 3σ principle to prevent extreme operating conditions or data acquisition errors from interfering with baseline modeling, ultimately forming a structured dataset containing operating condition labels, feature parameters, and timestamps. Based on the preprocessed structured dataset, the DBSCAN density clustering algorithm is used to automatically divide the data into clusters under different operating conditions without pre-specifying the number of clusters. The data distribution pattern is dynamically identified through the neighborhood radius and minimum sample number, effectively distinguishing between normal operating conditions and abnormal patterns (abnormal operating conditions and noisy data). The neighborhood radius is determined by the inflection point using a k-distance graph (k=5). The minimum number of samples (MinPts) is set according to the data density: MinPts=20 for normal operating conditions and MinPts=5 for abnormal operating conditions, thus distinguishing three types of data clusters: Cluster 1 represents normal operating conditions (high density, core points >90%), Cluster 2 represents abnormal operating conditions (low density, boundary points 30%), and Cluster 3 represents noise data (outliers, <5%). For each cluster, PCA principal component analysis is used to extract key operating parameters, including temperature change rate and vibration spectrum energy, reducing the dimensionality of high-dimensional data to 2-3 principal components while retaining more than 95% of the information. The statistical boundaries (mean, standard deviation) of each cluster are further calculated, and the probability distribution range of the healthy baseline is determined by combining kernel density estimation, thus determining the initial healthy baseline range. The baseline is dynamically updated through a sliding window mechanism. After real-time running data flows into the system, its deviation from the current healthy baseline (Mahavira distance) is calculated. If the deviation continues to exceed the threshold of 3 times the standard deviation, a feedback adjustment mechanism is triggered to expand the sliding window range and recalculate the cluster centers and boundary parameters to absorb short-term operating condition fluctuations. If the new data still deviates from the healthy baseline, the unsupervised model retraining process is started to adjust the DBSCAN parameters or the PCA principal component direction to ensure that the healthy baseline adapts to equipment aging or environmental changes. This enables the healthy baseline to have self-learning capabilities and maintain an accurate portrayal of the actual state of the equipment during long-term operation, avoiding false alarms or missed alarms caused by static healthy baselines and improving the reliability and timeliness of fault prediction. The multi-parameter correlation early warning unit is used to extract real-time equipment operating parameters and compare them with the generated health baseline. When multiple parameters show an abnormal pattern of co-deviating from the health baseline, the unit filters duplicate alarms through a clustering algorithm, thereby triggering an early correlation early warning to focus on key hidden dangers. The unit collects multi-source heterogeneous data of 690V low-voltage equipment in real time through edge computing nodes. After data cleaning and normalization, the unit calculates the Mahalanobis distance between each parameter in the multi-source heterogeneous data and the health baseline, and simultaneously marks the timestamp and operating condition label to generate a parameter deviation vector. The unit performs density clustering on the parameter deviation vector based on the DBSCAN density clustering algorithm to identify abnormal clusters of co-deviating multiple parameters. It filters duplicate alarms caused by pseudo-anomalies where a single parameter exceeds the limit but the overall pattern is normal, and retains the core abnormal pattern and its associated parameter combination. The clustering results are statistically verified (chi-square test). When the confidence of the abnormal cluster is ≥95%, an early warning is triggered, and the parameters and operating condition context of the key hidden dangers are output. The specific work of the multi-parameter correlation early warning unit is as follows: The edge computing node collects multi-source heterogeneous data in real time through various types of sensors deployed on 690V low-voltage equipment, and records the timestamp and operating condition label (load type, running time) simultaneously. After the data flows into the edge computing node, it is cleaned and normalized. Sliding window filtering is used to eliminate transient noise. The Z-Score method is used to unify the dimensions to avoid analysis bias caused by differences in parameter magnitude. Outliers are removed by the 3σ principle to prevent data acquisition errors or interference from extreme operating conditions. Then, the preprocessed multi-source heterogeneous data is mapped to the healthy baseline coordinate system, the Mahalanobis distance between each parameter and the healthy baseline is calculated, and a parameter deviation vector containing timestamp and operating condition label is generated. Using the parameter deviation vector as input, the DBSCAN density clustering algorithm automatically divides the data into clusters. Data distribution is identified through neighborhood radius and minimum sample size. During clustering, pseudo-anomalies with single parameter exceeding limits but normal overall patterns are filtered out, while core anomaly clusters and their associated parameter combinations are retained. Density reachability analysis ensures the continuity and stability of anomaly clusters, avoiding fragmented alarms caused by local noise. A chi-square test is performed on the DBSCAN clustering results to verify the statistical significance of the anomaly clusters compared to the healthy baseline. The chi-square statistic of the data distribution within the cluster and the baseline distribution is calculated, and the p-value of the chi-square test is determined based on the degrees of freedom. When the confidence level is ≥95% (p≤0.05), the anomaly pattern is considered valid. When an alarm is triggered, key hidden danger parameters and operating context (load type, runtime) are output, providing operational decision-making basis for maintenance personnel. The expressions for the Mahalanobis distances of each parameter to the healthy baseline are as follows: ; ; ; In the formula: Here, is the Mahalanobis distance, representing the distance of a sample data point relative to the multidimensional distribution of the healthy baseline; it is a scalar. x is a real-time parameter vector, which is... A column vector, representing a vector consisting of R monitored parameters (such as voltage, temperature, vibration, etc.) at a certain timestamp; The healthy baseline mean vector is a The column vectors, composed of the average values ​​of each parameter in the health baseline dataset, represent the center points of health status; S is the health baseline covariance matrix, which is a... The matrix describes the variance and covariance relationships among the parameters in the healthy baseline dataset. The diagonal elements are the variances of the parameters, and the off-diagonal elements are the covariances between two parameters. Let S be the inverse of the healthy baseline covariance matrix S, and T denote the transpose of the matrix; as x approaches... When the variation pattern of x is consistent with the parameters of the health baseline, the smaller the Mahalanobis distance, the healthier the device condition. When the variation pattern between parameters deviates from the inherent correlation in the healthy baseline, the larger the Mahalanobis distance, the higher the probability of an abnormal condition. The expression for the parameter deviation vector is as follows: ; In the formula: v is the parameter deviation vector. The calculated Mahalanobis distance is denoted by n; n is the number of data points within the time window used to construct a deviation vector; t is the time period identifier, which identifies the operating segment corresponding to the deviation vector; L is the operating condition label, which is a categorical variable representing the operating condition of the equipment within the time window; In a healthy state, the Mahalanobis distance value in the parameter deviation vector remains at a low level. When a fault begins to emerge, a series of Mahalanobis distance values ​​that gradually increase or remain at a high level will appear continuously in the parameter deviation vector, forming a high-bias cluster in clustering. The expression for the p-value in the chi-square test is as follows: ; ; In the formula: p is the p-value of the chi-square test, which represents the probability of observing the current sample data or more extreme data under the premise that the null hypothesis (H0) is true; The chi-square statistic measures the overall difference between observed frequencies and expected frequencies. To observe the frequency, the number of data points falling into each pre-divided parameter value interval in the abnormal clusters identified by DBSCAN; df represents the expected frequency, which is the number of data points in the healthy baseline dataset that fall within the same parameter value range; p represents the probability, and X represents a random variable that follows a chi-square distribution with df degrees of freedom. A small p value (≤0.05) means that, under the premise that the null hypothesis (the anomalous cluster is normal) is true, the observation of the current difference is a low-probability event, there is sufficient evidence to reject the null hypothesis, and the anomalous cluster is determined to have a significant statistical difference from the healthy baseline, which is a valid anomalous pattern, thus triggering an early warning. The smaller the p value, the more significant the difference between the anomalous pattern and the healthy state, and the higher the reliability of the early warning. The operation and maintenance knowledge graph construction module is used to build an operation and maintenance knowledge graph containing equipment-component-fault phenomenon-root cause-remedial measures by combining structured and unstructured knowledge, including equipment manuals, fault cases, maintenance procedures and expert experience. The fault propagation reasoning and localization module is used to analyze early correlation warning information, extract the implied fault phenomena as input, traverse and reason in the operation and maintenance knowledge graph, find the fault path that best matches the current multi-parameter anomaly pattern through graph algorithms, deduce the root cause, and locate the specific physical component. The resource scheduling module, based on the fault location results, automatically generates and recommends the optimal maintenance strategy from a pre-established maintenance strategy library that includes preventive maintenance, emergency repairs, and spare parts replacement, while intelligently scheduling maintenance resources.

[0020] Example 2, as Figure 1 , Figure 2 As shown, based on Embodiment 1, the present invention provides a technical solution: the operation and maintenance knowledge graph construction module specifically includes: performing word segmentation and entity recognition on structured and unstructured knowledge including equipment manuals, fault cases, maintenance procedures and expert experience; extracting core elements including equipment, components, fault phenomena, root causes and handling measures; establishing relationships between elements through semantic parsing; transforming unstructured knowledge into structured data and storing it in a multi-source knowledge base; defining the hierarchical relationship of equipment-component-fault phenomenon-root cause-handling measures based on structured data; constructing a knowledge graph framework through ontology modeling; clarifying the attributes and association rules of each entity; forming an extensible semantic network model; integrating multi-source knowledge base data; eliminating redundant information through entity alignment; verifying the consistency of relationships through logical reasoning; correcting erroneous associations by combining expert feedback; continuously optimizing the graph integrity and reasoning accuracy; and forming a standardized operation and maintenance knowledge graph. The specific tasks of the operation and maintenance knowledge graph construction module are as follows: performing word segmentation and entity recognition on multi-source knowledge, covering structured and unstructured data such as equipment manuals, fault cases, maintenance procedures and expert experience; extracting the core elements of equipment, components, fault phenomena, root causes and handling measures through natural language processing technology; establishing the relationship between elements using semantic parsing technology; clarifying the correlation between fault phenomena and components, or the causal relationship between root causes and handling measures through syntactic analysis and contextual understanding; and finally transforming unstructured knowledge into structured data and storing it in a standardized format in a multi-source knowledge base. Based on structured data, a hierarchical relationship is defined between equipment, components, fault phenomena, root causes, and remedial measures. An ontology modeling framework is constructed, clarifying the attributes and association rules of each entity to form a scalable semantic network model that enables logical connections between knowledge points. These association rules cover equipment-component relationships, component-fault phenomenon relationships, fault phenomenon-root cause relationships, root cause-remedial measures relationships, and cross-level association rules. Multi-source knowledge base data is integrated, and entity alignment technology eliminates redundant information. Logical reasoning is used to verify relationship consistency (checking the causal rationality between fault phenomena and root causes). Expert feedback is incorporated to correct erroneous associations and adjust unreasonable remedial measure recommendations, continuously optimizing the completeness and inference accuracy of the graph to form a standardized operation and maintenance knowledge graph. Equipment entity attributes include basic attributes, status attributes, and associated attributes. Basic attributes include equipment ID, equipment name, equipment type, system, manufacturer, commissioning date, and operating environment (temperature / humidity range). Status attributes include operating status (normal / fault / downtime), health score (calculated based on historical data), and maintenance cycle (last maintenance time, next maintenance schedule). Associated attributes include a list of associated components, associated fault case IDs, and associated maintenance procedure IDs. Component entity attributes include basic attributes, status attributes, and associated attributes; basic attributes include component ID, component name, component type, equipment ID, component location (specific location within the equipment), and criticality level (high / medium / low); status attributes include component status (normal / worn / failed), remaining lifespan (predicted based on usage time or monitoring data), and replacement records (replacement time, reason for replacement); associated attributes include a list of associated fault phenomena, a list of associated root causes, and associated remedial measures ID; The entity attributes of a fault phenomenon include basic attributes, monitoring attributes, and associated attributes. Basic attributes include phenomenon ID, phenomenon description (e.g., "excessive vibration" or "abnormal temperature"), phenomenon type (mechanical / electrical / thermal), and severity (minor / serious / critical). Monitoring attributes include monitoring parameters (e.g., vibration frequency, temperature value), threshold range (normal / early warning / alarm), occurrence time, and duration. Associated attributes include associated component ID, associated root cause ID, associated handling measure ID, and associated historical case ID. The root cause entity attributes include basic attributes, analytical attributes, and related attributes; basic attributes include cause ID, cause description (e.g., "insufficient bearing lubrication", "winding insulation aging"), and cause type (design defect / improper operation / environmental factors); analytical attributes include probability of occurrence (high / medium / low), scope of impact (local / global), and detection method (e.g., oil analysis, infrared thermal imaging); related attributes include associated fault phenomenon ID, associated remedial measure ID, and associated preventive measures (e.g., regular lubrication, environmental control); The entity attributes of the handling measures include basic attributes, execution attributes, and related attributes; the basic attributes include measure ID, measure type (repair / replacement / adjustment), operation steps (such as "shutdown inspection" or "replace bearing"), and required tools / spare parts; the execution attributes include execution time, execution personnel, execution result (success / failure), and subsequent verification (such as testing after restart); the related attributes include associated fault phenomenon ID, associated root cause ID, and associated preventive measures; The association rules cover equipment-component relationships, component-fault phenomenon relationships, fault phenomenon-root cause relationships, root cause-response relationship, and cross-level association rules. The device-component relationship is as follows: Rule 1: The device ID to which a component belongs must point to an existing device ID, forming a composition relationship where a device contains a component; Rule 2: The device health score must be based on the status of the associated components (when more than 50% of the components are in an abnormal state, the device health score drops to poor). The component-fault phenomenon relationship is as follows: Rule 3: The component ID associated with the fault phenomenon must point to an existing component ID, forming a causal relationship between the component fault and the phenomenon; Rule 4: The same component can be associated with multiple fault phenomena; The relationship between fault phenomenon and root cause is as follows: Rule 5: The root cause associated fault phenomenon ID must point to an existing phenomenon ID, forming a reasoning relationship where the phenomenon is caused by the cause; Rule 6: One fault phenomenon may correspond to multiple root causes, and the main cause needs to be determined through logical reasoning or expert verification. The root cause-response relationship is as follows: Rule 7: The root cause ID associated with the response measure must point to an existing cause ID, forming a solution relationship between cause and response measure; Rule 8: The response measures must have a clear priority (emergency measures take precedence over preventive measures) and be associated with preventive measures to avoid recurrence; The cross-level association rules are as follows: Rule 9: When equipment fails, it is necessary to trace the entire chain of component-failure phenomenon-root cause-response measures; Rule 10: After the response measures are implemented, the component status and equipment health score need to be updated to form a closed-loop management. The fault propagation reasoning and localization module specifically includes: receiving early correlation warnings from the health baseline warning module, parsing the warning information, extracting multi-parameter anomaly patterns, using natural language processing technology to extract fault phenomenon descriptions from unstructured alarm logs, and mapping them to standardized fault phenomenon nodes in the operation and maintenance knowledge graph to form an input set to be reasoned. Starting from the extracted fault phenomena, a depth-first traversal is performed in the operation and maintenance knowledge graph based on the relationship between equipment-component-fault phenomenon-root cause-handling measures. A graph matching algorithm with subgraph isomorphism is applied to analyze the similarity between the multi-parameter anomaly patterns and historical fault paths, and the fault propagation chain with the highest degree of consistency with the current fault phenomenon is selected. The fault propagation chain is traced backward along the matched fault path, and the root cause node is determined through causal relationship reasoning. Combining the equipment topology and component hierarchy, the abstract root cause is mapped to a specific physical component, and the localization result containing the root cause description, associated component name, and confidence score is output. The specific tasks of the fault propagation reasoning and localization module are as follows: It receives early correlation warnings from the health baseline warning module, extracts multi-parameter anomaly patterns using structured parsing technology, and simultaneously uses natural language processing technology to perform word segmentation, entity recognition, and semantic analysis on unstructured alarm logs to extract fault phenomenon descriptions. These descriptions are then mapped to predefined standardized fault phenomenon nodes in the operation and maintenance knowledge graph. Specifically, it uses word vector cosine distance similarity calculation to accurately match unstructured text with graph nodes, forming a set of inputs to be reasoned, containing multi-parameter anomaly patterns and standardized fault phenomena. After obtaining the standardized fault phenomenon input set, it uses the extracted standardized fault phenomena as a starting point and performs a depth-first traversal based on the hierarchical relationship of equipment-component-fault phenomenon-root cause-response measures. It compares the topological structure of the current multi-parameter anomaly pattern with that of historical fault paths using a subgraph isomorphism algorithm, calculating the similarity between the two in terms of node attributes and edge relationships. During the screening process, paths with high consistency in parameter combinations and complete causal chains are prioritized for matching, while low-confidence branches are eliminated. The fault propagation chain with the highest degree of consistency with the current fault phenomenon is output, clarifying the complete path from the initial fault phenomenon to the root cause. After the path matching is completed, reverse tracing and deep reasoning are performed along the selected fault propagation chain with the highest degree of consistency. The root cause node is determined through causal reasoning. Combining the equipment topology and physical connection relationship, the abstract root cause is mapped to specific physical components. The location result includes a description of the root cause, the name of the associated component, and a confidence score (calculated based on historical fault data statistics and real-time parameter deviation). If there are multiple candidate paths, the optimal solution is output through weighted voting to ensure the interpretability and engineering applicability of the location result. The expression for the similarity between multi-parameter anomaly patterns and historical failure paths is as follows: ; ; ; In the formula: This represents the comprehensive similarity between the current multi-parameter anomaly pattern and historical failure paths (value range: 0~1, higher values ​​indicate a better match). For node attribute similarity, This is the current set of abnormal parameters. This refers to the set of parameters involved in the historical path. For edge relationship similarity, This represents the percentage deviation of the current j-th parameter. This represents the percentage deviation of the corresponding parameter in the historical path, where N is the total number of parameters. These are the weighting coefficients, default. , This reflects that node matching takes precedence over edge matching, when the current parameter set is completely consistent with the historical path and the degree of deviation is the same. If the parameters do not overlap or deviate in the opposite direction, ; The expression for the fault propagation chain score that best matches the current fault phenomenon is as follows: ; ; In the formula: For the overall consistency of the fault propagation chain, Score the causal chain integrity. For the complete causal law quantity, This represents the total number of candidate paths. Weighting coefficients (default) , (Balancing parameter matching and causal chain quality), paths with high parameter similarity and complete causal chains score the highest, while paths with low parameter matching but complete causal chains score the lowest. The expression for the confidence score is as follows: ; ; ; In the formula: Final confidence level for identifying the root cause (value range: 0~1, the higher the value, the more reliable). Rate historical frequencies. This represents the number of times the historical fault path occurred. This represents the total number of historical failures. Scoring for real-time parameter deviation. The historical maximum deviation threshold for the current j-th parameter. For parameter weights, Weighting historical data (default) =0.6, reflecting that historical experience takes precedence over real-time data; paths with high historical frequency and significant deviations from real-time parameters have the highest confidence, while paths with low historical frequency but slight deviations from real-time parameters have lower confidence. The expression for outputting the optimal solution using weighted voting is as follows: ; In the formula: The optimal fault propagation chain after weighted voting. Let f be the match score of the f-th path in the q-th match, let q be the confidence score of the f-th path in the q-th match, and K be the number of repeated matches (used to eliminate randomness, default K=5). Paths with both high match score and high confidence score will be selected first. The resource scheduling module specifically includes: receiving the fault location results output by the fault propagation reasoning and location module, including the root cause description, related component names and confidence scores; initially screening suitable strategies from a pre-established maintenance strategy library based on fault type to form a candidate strategy set, including maintenance strategies for preventive maintenance, emergency repair and spare parts replacement; combining the current maintenance resource status (manpower, spare parts inventory and equipment availability) and maintenance cost model (labor costs, spare parts costs and downtime losses) to evaluate the resource occupancy and cost of the candidate strategy set, eliminating strategies with insufficient resources or excessive costs, generating an optimized list of feasible strategies; prioritizing and ranking the list of feasible strategies based on fault urgency and resource efficiency (shortest completion time), outputting the optimal maintenance strategy and scheduling the corresponding resources (specifying the maintenance team, spare parts and time window), and synchronously updating the resource status; The specific tasks of the resource scheduling module are as follows: It receives the fault location results output by the fault propagation reasoning and location module, which includes a root cause description, related component names, and confidence scores. Then, it performs preliminary screening based on the fault type from a pre-built maintenance strategy library containing preventative maintenance, emergency repair, and spare parts replacement strategies, matching suitable maintenance strategies to form a candidate strategy set. This aims to comprehensively cover the response needs under different fault scenarios, ensuring that reasonable maintenance plans can be formulated for various fault conditions. After generating the candidate strategy set, it evaluates it in conjunction with the current maintenance resource status. The maintenance resource status includes manpower, spare parts inventory, and equipment availability. Simultaneously, using an O&M cost model encompassing labor costs, spare parts costs, and downtime losses, the candidate strategy set is evaluated for resource utilization and economic efficiency. By analyzing the performance of each strategy in terms of resource utilization and cost, strategies that cannot be implemented due to insufficient resources or whose costs exceed budget standards are eliminated, generating an optimized list of feasible strategies. After obtaining the list of feasible strategies, the strategies in the list are prioritized and ranked according to fault urgency and resource efficiency (measured by the shortest completion time). Fault urgency reflects the severity of the fault's impact on system operation, while resource efficiency considers the speed and effectiveness of strategy implementation. By comprehensively considering fault urgency and resource efficiency, the optimal maintenance strategy is determined. Once the optimal maintenance strategy is determined, the corresponding resources are immediately scheduled, including assigning a suitable maintenance team, allocating necessary spare parts, and determining a reasonable time window. Simultaneously, the resource status is updated to ensure that resource scheduling is based on the latest resource information, guaranteeing the continuity and efficiency of the entire O&M process. The expression for resource occupancy assessment is as follows: ; In the formula: For strategy Resource availability (1 = available, 0 = unavailable). Based on the current available manpower, For strategy Required working hours This represents the current spare parts inventory quantity. For strategy Required number of spare parts The current available time window for the device. For strategy Estimated execution time, assuming sufficient resources ( ), If the number of available strategies is 0, then the strategy is available; otherwise, the number is 0. The more abundant the resources, the more available strategies there are. The expression for cost assessment is as follows: ; ; In the formula: For strategy Total cost This is the hourly rate (hours multiplied by the unit price). This is the cost of spare parts (quantity multiplied by unit price). This is the downtime loss (downtime multiplied by the loss per hour). For strategy Downtime caused (in hours) The downtime loss is expressed as yuan per hour, and the weight is denoted as . , , The higher the labor costs, spare parts costs, or downtime losses, the higher the total cost. The expression for the list of feasible strategies is as follows: ; In the formula: List of feasible strategies The budget threshold (unit: yuan) is the highest budget. The higher the budget, the more feasible strategies there are, and strategies with insufficient resources or cost overruns are eliminated. The expression for priority scoring is as follows: ; ; ; In the formula: For strategy Priority score (value range: 1~3, the higher the value, the higher the priority); The severity level is categorized by the degree of urgency of the fault: WG represents critical faults, HG represents severe faults, and DG represents general faults. Resource efficiency (measured by the shortest completion time), Smaller size means higher efficiency. Assigning weights based on urgency, =0.6; Policies with critical failures and short execution times have high priority.

[0021] It should be noted that, in this document, relational terms such as "first" and "second" are used only to distinguish one entity or operation from another, and do not necessarily require or imply any such actual relationship or order between these entities or operations. Furthermore, the terms "comprising," "including," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or apparatus that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such a process, method, article, or apparatus. Without further limitations, an element defined by the phrase "comprising one..." does not exclude the presence of other identical elements in the process, method, article, or apparatus that includes said element.

[0022] Although embodiments of the invention have been shown and described, it will be understood by those skilled in the art that various changes, modifications, substitutions and alterations can be made to these embodiments without departing from the principles and spirit of the invention, the scope of which is defined by the appended claims and their equivalents.

Claims

1. An intelligent operation and maintenance management system for 690V low-voltage equipment in LNG terminals, comprising a visual management platform, characterized in that, The visualization management platform has the following communication connection modules: The digital chassis module is used to collect multi-source heterogeneous data from 690V low-voltage equipment in LNG terminals and build digital twin models of the equipment. The health baseline early warning module is used to generate an adaptively adjusted health baseline by utilizing historical operating data of the equipment and the equipment's digital twin model, so as to trigger early correlation warnings; The operation and maintenance knowledge graph construction module is used to build an operation and maintenance knowledge graph that includes equipment manuals, fault cases, maintenance procedures and expert experience, and contains equipment-component-fault phenomenon-root cause-remedial measures. The fault propagation reasoning and localization module is used to analyze early correlation warning information, extract the implied fault phenomena as input, traverse and reason in the operation and maintenance knowledge graph, find the root cause of the fault path, and locate the specific physical component. The resource scheduling module generates and recommends the optimal maintenance strategy from a pre-established operation and maintenance strategy library based on the fault location results, and intelligently schedules maintenance resources.

2. The intelligent operation and maintenance management system for 690V low-voltage equipment in LNG terminals according to claim 1, characterized in that: The digital chassis module includes a multi-source heterogeneous data acquisition unit and a model building unit; The multi-source heterogeneous data acquisition unit is used to access multi-source heterogeneous data from 690V low-voltage equipment in the LNG terminal through IoT gateways and edge computing nodes. This data includes electrical data, status data, environmental data from the 690V equipment, and process data from SCADA and DCS systems. The unit also performs preprocessing operations on the multi-source heterogeneous data to form a standardized basic dataset. The model building unit constructs a high-precision digital twin model of the equipment based on the physical mechanism of the equipment and the collected historical data, and maps the physical entity's state and behavior in real time.

3. The intelligent operation and maintenance management system for 690V low-voltage equipment in LNG terminals according to claim 2, characterized in that: The multi-source heterogeneous data acquisition unit specifically includes: By collecting multi-source heterogeneous data through IoT gateways and edge computing nodes deployed on 690V low-voltage equipment in LNG terminals, the data includes electrical and condition monitoring data from the 690V low-voltage equipment itself, environmental sensor data, and process data from the upper-level SCADA / DCS system. Preprocessing of the multi-source heterogeneous data received includes steps such as cleaning, alignment, standardization, and labeling. The standardized, multi-source heterogeneous data that has undergone comprehensive governance is organized and packaged according to the system-defined format to output a standardized basic dataset.

4. The intelligent operation and maintenance management system for 690V low-voltage equipment in LNG terminals according to claim 2, characterized in that: The model building unit specifically includes: Based on the physical mechanism of the equipment, an initial digital twin model framework for the equipment is constructed, integrating the multidisciplinary principles of electrical, thermodynamic and mechanical dynamics, defining the mathematical relationships of key parameters including voltage, temperature and vibration, and simultaneously integrating historical operating data. The boundary conditions and dynamic characteristics of the model are corrected through parameter identification. By utilizing collected multi-source heterogeneous data, a physical information neural network is used to train the digital twin model of the device to perform nonlinear mapping of device behavior. Deploy the trained device digital twin model to edge computing nodes, access device sensor data in real time, map the physical entity state and behavior, and continuously fine-tune the model parameters of the device digital twin model using the new data stream.

5. The intelligent operation and maintenance management system for 690V low-voltage equipment in LNG terminals according to claim 2, characterized in that: The health baseline early warning module includes a health baseline self-learning unit and a multi-parameter correlation early warning unit; The health baseline self-learning unit is used to combine the equipment digital twin model and extract the equipment's historical operating data, and autonomously learn the normal operating parameter range of the equipment under different loads, seasons and working conditions through unsupervised learning machine learning algorithms to form a health baseline. The multi-parameter correlation early warning unit is used to extract real-time equipment operating parameters and compare them with the generated health baseline. When multiple parameters show an abnormal pattern of co-deviating from the health baseline, a clustering algorithm is used to filter duplicate alarms, thereby triggering an early correlation early warning.

6. The intelligent operation and maintenance management system for 690V low-voltage equipment in LNG terminals according to claim 5, characterized in that: The health baseline self-learning unit specifically includes: By combining the digital twin model of the equipment with historical operating data extracted from the historical database, and aligning it with time stamps and operating condition labels including season and load type, a structured dataset is constructed through noise reduction, normalization, and outlier removal. The DBSCAN density-based clustering algorithm is used to group the data in the structured dataset, automatically identify data clusters under different working conditions, extract key operating parameters, including temperature change rate and vibration spectrum energy, through PCA feature dimensionality reduction, and calculate the statistical boundary of each working condition cluster to determine the initial health baseline range. Based on the deviation analysis between real-time running data and historical baselines, the cluster centers and boundary parameters are dynamically updated through a sliding window mechanism, and a feedback adjustment mechanism is introduced. When new data continues to deviate from the healthy baseline, the unsupervised model is retrained to form an adaptively adjustable healthy baseline.

7. The intelligent operation and maintenance management system for 690V low-voltage equipment in LNG terminals according to claim 5, characterized in that: The multi-parameter correlation early warning unit specifically includes: Multi-source heterogeneous data of 690V low-voltage equipment is collected in real time through edge computing nodes. After data cleaning and normalization, the Mahalanobis distance between each parameter in the multi-source heterogeneous data and the healthy baseline is calculated. Timestamps and operating condition labels are marked synchronously to generate parameter deviation vectors. Based on the DBSCAN density clustering algorithm, density clustering is performed on the parameter deviation vector to identify abnormal clusters with multiple parameters deviating together. False alarms caused by pseudo-anomalies where a single parameter exceeds the limit but the overall pattern is normal are filtered out, while core abnormal patterns and their associated parameter combinations are retained. The clustering results are statistically verified. When the confidence level of the abnormal cluster is ≥95%, an early warning is triggered, and the parameters and operating context of the key hidden dangers are output.

8. The intelligent operation and maintenance management system for 690V low-voltage equipment in LNG terminals according to claim 1, characterized in that: The operation and maintenance knowledge graph construction module specifically includes: The system performs word segmentation and entity recognition on structured and unstructured knowledge, including equipment manuals, failure cases, maintenance procedures and expert experience, to extract core elements including equipment, components, failure phenomena, root causes and handling measures. It also establishes the relationship between elements through semantic parsing, transforms unstructured knowledge into structured data and stores it in a multi-source knowledge base. Based on the hierarchical relationship of equipment-component-fault phenomenon-root cause-response measures defined by structured data, a knowledge graph framework is constructed through ontology modeling, clarifying the attributes and association rules of each entity, and forming an extensible semantic network model; By integrating multi-source knowledge base data, eliminating redundant information through entity alignment, verifying relationship consistency through logical reasoning, correcting erroneous associations by combining expert feedback, and continuously optimizing the completeness of the graph and the accuracy of reasoning, a standardized operation and maintenance knowledge graph is formed.

9. The intelligent operation and maintenance management system for 690V low-voltage equipment in LNG terminals according to claim 8, characterized in that: The fault propagation reasoning and localization module specifically includes: Receive early correlation warnings from the health baseline warning module, parse the warning information, extract multi-parameter abnormal patterns, use natural language processing technology to extract fault phenomenon descriptions from unstructured alarm logs, and map them to standardized fault phenomenon nodes in the operation and maintenance knowledge graph to form an input set to be reasoned. Starting with the extracted fault phenomena, a depth-first traversal is performed in the operation and maintenance knowledge graph based on the relationship between equipment, components, fault phenomena, root causes, and handling measures. A graph matching algorithm with subgraph isomorphism is applied to analyze the similarity between multi-parameter abnormal patterns and historical fault paths, and the fault propagation chain with the highest degree of consistency with the current fault phenomenon is selected. Tracing back along the matched fault path, the root cause node is determined through causal reasoning. Combining the equipment topology and component hierarchy, the abstract root cause is mapped to a specific physical component, and the location result is output, which includes a description of the root cause, the name of the associated component, and a confidence score.

10. The intelligent operation and maintenance management system for 690V low-voltage equipment in LNG terminals according to claim 1, characterized in that: The resource scheduling module specifically includes: Receive the fault location results output by the fault propagation reasoning and location module, including the root cause description, related component names and confidence scores. From the pre-established operation and maintenance strategy library, preliminarily screen the suitable strategies according to the fault type to form a candidate strategy set, which includes operation and maintenance strategies for preventive maintenance, emergency repair and spare parts replacement. Based on the current maintenance resource status and operation and maintenance cost model, resource consumption and cost assessments are performed on the candidate strategy set, strategies with insufficient resources or excessive costs are eliminated, and an optimized list of feasible strategies is generated. Based on the urgency of the fault and resource efficiency, the list of feasible strategies is prioritized and sorted, the optimal maintenance strategy is output and the corresponding resources are scheduled, and the resource status is updated synchronously.

Images