Intelligent management and control early warning platform and method for fuel whole process

Abstract

The invention discloses an intelligent management and control early warning platform and method for a fuel whole process, and relates to the technical field of intelligent management, and the platform comprises a heterogeneous data collection module which is used for collecting multi-source data of the fuel whole process in real time; the space-time coupling analysis module is used for constructing a space-time model of fuel circulation and generating a multi-dimensional tensor; the causal inference module is used for identifying an abnormal causal relationship; the comprehensive risk assessment module is used for extracting risk factors and calculating a comprehensive risk index; and the grading early warning management and control module is used for automatically dividing early warning grades according to the risk indexes and generating a multi-dimensional early warning matrix so as to realize intelligent early warning management and control. According to the invention, the technical problem that abnormal events cannot be timely identified and accurately pre-warned due to multi-source data dispersion and link association deficiency in existing fuel whole-process management and control is solved, and the technical effects of accurate monitoring and real-time pre-warning of the fuel whole process are achieved through an intelligent management and control system based on space-time coupling and causal inference.

Description

Technical Field

[0001] This invention relates to the field of intelligent management technology, specifically to an intelligent control and early warning platform and method for the entire fuel process. Background Technology

[0002] With the large-scale and intelligent development of fuel supply chain management, the entire process of fuel from procurement, transportation, delivery to the plant, storage to use has formed a complex and dynamic network spanning regions and systems. Traditional fuel management methods mainly rely on manual recording and segmented monitoring, which suffers from problems such as scattered data sources and information lag, making it difficult to meet the comprehensive control needs of modern thermal power plants for safety, efficiency, and quality. In particular, in a multi-source data environment, the temporal correlation and spatial coupling relationships between different links are difficult to accurately model, resulting in the inability to detect and trace abnormal events in a timely manner. Summary of the Invention

[0003] This application provides a smart management and early warning platform and method for the entire fuel process, which solves the technical problem that the existing fuel process management and control is characterized by scattered multi-source data and missing links, resulting in abnormal events that cannot be identified and accurately warned in a timely manner.

[0004] The first aspect of this application provides an intelligent management and early warning platform for the entire fuel process. The platform includes: a heterogeneous data acquisition module for real-time acquisition of multi-source heterogeneous data at each stage of the fuel process, with timestamps and geographic location tags; a spatiotemporal coupling analysis module for constructing a spatiotemporal coupling model of fuel flow based on a graph neural network, using the multi-source heterogeneous data as input, representing the state of each stage with event nodes, and representing time delay and geographic coupling relationships with edge weights, generating a multi-dimensional spatiotemporal tensor; a causal inference module for feature embedding and entity extraction of the multi-dimensional spatiotemporal tensor, constructing a multi-entity semantic relationship network, inferring potential causal chains of abnormal events, and identifying abnormal causal relationship sets; a comprehensive risk assessment module for using the abnormal causal relationship sets as input, combining the spatiotemporal coupling model to extract multi-dimensional risk factors, and dynamically weighting them to generate a comprehensive risk index; and a hierarchical early warning and control module for automatically classifying early warning levels based on the dynamic distribution range of the comprehensive risk index, and generating a multi-dimensional early warning matrix based on anomaly type, frequency, and correlation, for hierarchical early warning and control.

[0005] The second aspect of this application provides a smart management and early warning method for the entire fuel process. The method includes: collecting multi-source heterogeneous data in real time at each stage of the fuel process, and attaching timestamps and geographic location tags; constructing a spatiotemporal coupling model of fuel flow based on a graph neural network, using the multi-source heterogeneous data as input, with event nodes representing the state of each stage and edge weights representing time delay and geographic coupling relationships, generating a multi-dimensional spatiotemporal tensor; performing feature embedding and entity extraction on the multi-dimensional spatiotemporal tensor to construct a multi-entity semantic relationship network, and inferring potential causal chains of abnormal events to identify abnormal causal relationship sets; using the abnormal causal relationship sets as input, extracting multi-dimensional risk factors in conjunction with the spatiotemporal coupling model, and dynamically weighting them to generate a comprehensive risk index; automatically classifying early warning levels based on the dynamic distribution range of the comprehensive risk index, and generating a multi-dimensional early warning matrix in conjunction with anomaly type, frequency, and correlation, for hierarchical early warning management.

[0006] One or more technical solutions provided in this application have at least the following technical effects or advantages: This application provides a smart management and early warning platform and method for the entire fuel process, relating to the field of intelligent management technology. It uses graph neural networks to model the spatiotemporal correlation of multi-source heterogeneous data, combines causal reasoning to identify anomaly sources and propagation paths, dynamically assesses the comprehensive risk index, and achieves intelligent risk classification and closed-loop management through a hierarchical early warning matrix. This improves the predictability and decision-making intelligence of the system, solving the technical problem in existing fuel process management where multi-source data is scattered and the correlation between links is missing, resulting in the inability to identify and accurately warn of abnormal events in a timely manner. It achieves the technical effect of accurate monitoring and real-time early warning of the entire fuel process through a smart management and control system based on spatiotemporal coupling and causal inference. Attached Figure Description

[0007] To more clearly illustrate the technical solutions in the embodiments of the present invention, the accompanying drawings used in the description of the embodiments will be briefly introduced below. Obviously, the accompanying drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0008] Figure 1 A schematic diagram of a smart management and early warning platform for the entire fuel process provided in this application embodiment; Figure 2 This is a schematic diagram of a smart management and early warning method for the entire fuel process, provided as an embodiment of this application.

[0009] Figure labeling: Heterogeneous data acquisition module 10, spatiotemporal coupling analysis module 20, causal inference module 30, comprehensive risk assessment module 40, hierarchical early warning and control module 50. Detailed Implementation

[0010] This application provides a smart management and early warning platform and method for the entire fuel process, which solves the technical problem that the existing fuel process management and control is characterized by scattered multi-source data and missing links, resulting in abnormal events that cannot be identified and accurately warned in a timely manner.

[0011] The technical solutions of the embodiments of this application will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only a part of the embodiments of this application, and not all of them. All other embodiments obtained by those skilled in the art based on the embodiments of this application without creative effort are within the scope of protection of this application.

[0012] It should be noted that the terms "first," "second," etc., in the specification and accompanying drawings of this application are used to distinguish similar objects and are not necessarily used to describe a specific order or sequence. It should be understood that such data can be interchanged where appropriate so that the embodiments of this application described herein can be implemented in orders other than those illustrated or described herein. Furthermore, the terms "comprising" and "having," and any variations thereof, are intended to cover non-exclusive inclusion; for example, a process, method, system, product, or server that includes a series of steps or units is not necessarily limited to those steps or units explicitly listed, but may include other steps or modules not explicitly listed or inherent to such processes, methods, products, or devices.

[0013] Example 1, as Figure 1 As shown, this application provides an intelligent management and early warning platform for the entire fuel process, the platform comprising: The heterogeneous data acquisition module 10 is used to collect multi-source heterogeneous data in real time at each stage of the fuel process, and attach timestamps and geographic location tags.

[0014] Specifically, the heterogeneous data acquisition module 10 of this application is responsible for collecting multi-source heterogeneous data in real time at each key stage of the fuel process, and attaching timestamps and geographic location tags to these data to realize the real-time collection, sorting and unified management of data from different sources and of different types.

[0015] In actual operation, data acquisition terminals need to be deployed at key stages of fuel circulation, such as loading points, transport vehicles, weighbridges at the plant entrance, sampling and preparation workshops, laboratories, and coal yards. Each acquisition terminal can connect to different types of equipment, such as weighbridges, samplers, belt scales, analyzers, cameras, and access control systems. Because these devices come from different manufacturers and use different communication protocols, the module is designed with multiple data acquisition interfaces to automatically identify the device type and complete data access. In this way, whether the data is weighing values, sampling time, laboratory indicators, or video images, it can all be automatically collected and aggregated by the system.

[0016] The collected data first undergoes preliminary processing at edge computing nodes. These edge nodes can be understood as on-site intelligent transfer stations, responsible for cleaning, denoising, format conversion, and anomaly filtering of the data before transmitting the processed results to the central system. To enhance data traceability, each data point is automatically tagged with a timestamp and a geographic location label. The timestamp accurately records the time of data collection, ensuring all data aligns on the same timeline during subsequent analysis; the geographic location label indicates the location where the data was generated, allowing the system to understand the spatial location of specific processes and establish a correspondence between time and space information.

[0017] To ensure the real-time performance and reliability of data, the heterogeneous data acquisition module 10 is internally designed with data buffering and breakpoint resume mechanisms. When the on-site network is unstable or temporarily interrupted, the system automatically caches the data locally. Once the network is restored, it will automatically re-transmit the data, ensuring that the data is not lost or corrupted. At the same time, a hierarchical transmission mechanism is adopted, prioritizing the uploading of critical data to ensure the real-time performance of core information. All data is verified during transmission to prevent omissions or tampering.

[0018] At the data aggregation level, data from different stages can be uniformly coded, categorized, and stored in a spatiotemporal data center. This data center can quickly retrieve information based on time, location, equipment, or business process. For example, the system can quickly find the entire data trajectory of a batch of fuel from its arrival at the plant to its testing, providing a reliable basis for subsequent spatiotemporal analysis, causal inference, and risk assessment.

[0019] The spatiotemporal coupling analysis module 20 is used to construct a spatiotemporal coupling model of fuel flow based on graph neural network, and takes the multi-source heterogeneous data as input, with event nodes representing the state of each link, edge weights representing the time delay and geographical coupling relationship, and generating a multidimensional spatiotemporal tensor.

[0020] Furthermore, when constructing a spatiotemporal coupling model of fuel flow based on a graph neural network, the spatiotemporal coupling analysis module 20 is also used to perform the following steps: P21: Initialize the graph structure of the spatiotemporal coupling model by taking key events in the entire fuel process as nodes and the flow logic between events as edges; P22: Collect historical multi-source heterogeneous data to train the graph neural network until the model loss function converges, and obtain the spatiotemporal coupling model that represents the spatiotemporal dynamics of fuel flow.

[0021] It should be understood that the main task of the spatiotemporal coupling analysis module 20 of this application is to establish a spatiotemporal coupling model that reflects temporal changes and spatial relationships throughout the entire fuel process, in order to comprehensively characterize the flow characteristics and correlation patterns of fuel at different stages. Through this modeling approach, the system can not only see the state of each stage itself, but also understand the temporal sequence, spatial geographical distribution, and mutual influence between stages, thereby achieving dynamic perception and intelligent analysis of the entire fuel chain.

[0022] The spatiotemporal coupling analysis module 20 uses graph neural networks as its core technology. Simply put, a graph neural network is an intelligent algorithm capable of processing the relationships between nodes. Unlike traditional models that only consider individual data points, it learns patterns from the connections between points, making it particularly suitable for business scenarios with complex flow relationships. The entire fuel flow process perfectly fits this characteristic: from loading, transportation, arrival at the plant, sampling, testing to warehousing, each stage influences and transmits information layer by layer. Therefore, a spatiotemporal map of fuel flow can be established using graph neural networks.

[0023] When constructing a spatiotemporally coupled model of fuel flow, the first step is to initialize the model's graph structure. This process uses key events in the entire fuel flow process as nodes and the flow logic between events as edges. Specifically, key events include important events in each stage of fuel procurement, transportation, storage, and use, such as fuel arriving at the storage point and fuel being used. The flow logic between events reflects the sequence and dependencies between these events; for example, fuel must arrive at the storage point before it can be used. In this way, the initialized graph structure can intuitively represent the key stages in the fuel flow process and their interrelationships, laying the foundation for subsequent spatiotemporal dynamic analysis.

[0024] After initializing the graph structure, historical multi-source data is needed to train the model. This historical multi-source data includes various types of information such as weighbridge weighing records, transportation trajectories, test results, equipment status, and environmental parameters. The system inputs this data into the graph neural network using a batch training approach, gradually inputting data in units of a sliding time window, and continuously iterating and optimizing the network parameters so that the model can automatically learn the coupling relationships and variation patterns of fuel at different stages. During training, the system calculates a loss function, which is an indicator that measures the difference between the model's predictions and the actual data. When the loss function continuously decreases and tends to stabilize, it indicates that the model has learned stable patterns and has reached a convergence state.

[0025] Ultimately, the trained spatiotemporal coupling model generates a multidimensional spatiotemporal tensor containing multi-level information about fuel status in time, space, and stages. This tensor acts like a three-dimensional map of fuel flow, which can be used to analyze whether an anomaly in a certain stage is related to delays in preceding stages or geographical conditions, and can also support subsequent causal inference and risk assessment.

[0026] Furthermore, when the spatiotemporal coupling analysis module 20 takes the multi-source heterogeneous data as input, uses event nodes to represent the state of each stage, and edge weights to represent the time delay and geographical coupling relationship, and generates a multidimensional spatiotemporal tensor, it is also used to perform the following steps: P23: The real-time collected multi-source heterogeneous data is converted into node feature vectors and edge feature vectors required by the spatiotemporal coupling model; P24: The node feature vectors and edge feature vectors are input into the spatiotemporal coupling model for forward inference, and the hidden states of all nodes output by the model in each time slice are aggregated and structurally assembled into the multidimensional spatiotemporal tensor.

[0027] Optionally, when generating multidimensional spatiotemporal tensors, the spatiotemporal coupling analysis module 20 is responsible for fusing and calculating the real-time collected multi-source heterogeneous data with the constructed spatiotemporal coupling model to achieve a structured mapping from data input to model understanding. This enables the system to automatically capture the temporal correlation and spatial dependency between different business links through intelligent computing, and output a computable, analyzable, and traceable dynamic data structure in a unified tensor form.

[0028] First, the real-time collected multi-source heterogeneous data is transformed into input features that can be directly calculated by the spatiotemporal coupling model. The system receives various data streams in real time from edge nodes and the data platform, including metering data, transportation trajectories, equipment operating status, test results, video recognition events, and environmental parameters. After preprocessing, this raw data is divided into two types of inputs: node feature vectors and edge feature vectors. The node feature vectors describe the state information of each event node, such as fuel batch number, equipment condition, sampling results, moisture content, arrival time, and operator number. When generating node feature vectors, the system uniformly encodes different types of data, normalizes quantitative data, transforms qualitative data into a computable numerical form, and combines the node's timestamp and geographic coordinates to form a structured node feature vector.

[0029] Edge feature vectors are used to characterize the connections between nodes and their spatiotemporal features. Based on the logical sequence of events, the system extracts the time delay between two nodes, such as transportation time and testing cycles, as well as geographical coupling parameters, such as distance, terrain correlation, and factory area identification. These are then combined with the flow direction and business type to form edge feature vectors with directionality and dynamic weights. This vector form allows the model to not only understand the connection between two links but also how, over what time, and over what distance they influence each other.

[0030] Subsequently, the node feature vectors and edge feature vectors are input into the spatiotemporal coupling model for forward inference. At this point, the graph neural network propagates information layer by layer according to the defined graph structure. Each node receives information input from its neighboring nodes during computation, combines it with edge weights for weighted aggregation, and thus updates its own hidden state—that is, the comprehensive feature expression of the node in a specific time and space context. For example, the system employs a time-slice layering mechanism during inference, dividing the entire fuel flow process into multiple discrete time slices in chronological order. For instance, the process of a batch of fuel from loading to testing can be divided into several time slices, such as loading, transportation, arrival at the plant, sampling, and testing. The model calculates and updates the hidden states of all nodes in each time slice and aggregates the output results in real time. The hidden states of all nodes in all time slices are structurally spliced ​​and indexed according to the time, space, and state dimensions, ultimately forming a multidimensional spatiotemporal tensor. The multidimensional spatiotemporal tensor is a high-dimensional data structure that can simultaneously contain information in both time and space dimensions, providing data support for subsequent causal inference and risk assessment.

[0031] The causal inference module 30 is used to embed features and extract entities from the multidimensional spatiotemporal tensor, construct a multi-entity semantic relationship network, infer the potential causal chain of abnormal events, and identify abnormal causal relationship sets.

[0032] Furthermore, when performing feature embedding and entity extraction on the multidimensional spatiotemporal tensor to construct a multi-entity semantic relationship network, the causal inference module 30 is also used to perform the following steps: P31: The multidimensional spatiotemporal tensor is reduced in dimensionality using a nonlinear transformation to obtain a low-dimensional dense feature embedding representation; P32: Based on the feature embedding representation, key entities are identified and extracted; P33: Semantic relationships between the key entities are defined and established according to business processes and spatial topology; P34: The key entities and semantic relationships are vectorized to construct a multi-entity semantic relationship network.

[0033] It should be understood that the main function of the causal inference module 30 of this application is to identify the hidden logical connections and causal relationships in different business processes based on the multidimensional spatiotemporal tensor output by the spatiotemporal coupling analysis module, thereby realizing the tracking of the causes of abnormal events and the location of responsibilities.

[0034] In the specific implementation process, the causal inference module 30 first performs feature embedding processing on the multidimensional spatiotemporal tensor. Since the spatiotemporal tensor contains multidimensional data such as time, space, and state, its structure is often high-dimensional and sparse. Direct computation would lead to information redundancy and excessive model complexity. Therefore, a nonlinear feature transformation method can be used to map the high-dimensional feature space to a low-dimensional dense space through a multi-layer mapping function, thereby preserving the main feature distribution patterns and eliminating noise and redundancy. This process can be understood as compressing and learning complex data, allowing the model to retain only the key features most valuable for fuel behavior evolution, resulting in a set of low-dimensional dense feature embedding vectors. Each vector corresponds to a node or event state in the spatiotemporal tensor, efficiently expressing its core attributes.

[0035] Next, based on the obtained low-dimensional dense feature embedding representation, the module begins to identify and extract key entities. Through methods such as cluster analysis, principal component analysis, and feature importance calculation, the low-dimensional feature space is filtered and aggregated to identify representative key objects throughout the fuel process. For example, entities representing different types such as suppliers, transportation batches, sampling equipment, test results, storage locations, and operators can be identified. Each entity not only contains its own attributes, such as number, time, location, and indicator value, but also carries dynamic semantic information reflected by the feature embedding vector.

[0036] Subsequently, based on business processes and spatial topology information, semantic relationships between key entities are defined and established. For example, firstly, based on fuel business logic, temporal relationships between events are established, such as transportation activities affecting arrival time, sampling activities affecting test results, and test indicators affecting blending quality. Simultaneously, spatial topology information is used to describe the geographical dependencies between entities, such as the distance between sampling points and coal yards, and the correspondence between transportation routes and storage areas. In this way, a semantic association system is formed that simultaneously incorporates business logic and spatial constraints. Each association includes directional and weight attributes; directionality describes the direction of causal propagation, and weight reflects the strength and degree of influence of the relationship.

[0037] Finally, the module vectorizes key entities and semantic relationships to construct a multi-entity semantic relationship network. Vectorization is the process of converting entities and relationships into numerical vectors, generating a unique vector representation for each entity, assigning a computable weight vector to each semantic relationship, and forming a network structure through matrix storage to construct the multi-entity semantic relationship network. This network can not only reflect direct dependencies between entities but also express potential impact paths across levels and stages. For example, delays in the transportation process may affect the receiving process through time-dependent edges, and then indirectly affect the test results through sampling time nodes. Through network propagation algorithms, the system can trace the impact path along this chain relationship and identify potential sources of anomalies.

[0038] Furthermore, when performing the inference of potential causal chains of abnormal events and identifying abnormal causal relationship sets, the causal inference module 30 is also used to perform the following steps: P35: Locate the anomalous event node indicated by the multidimensional spatiotemporal tensor in the multi-entity semantic relationship network; P36: Based on the anomalous event node, perform abductive reasoning along the directed edges in the semantic relationship network to infer the potential causal chain of the anomalous event; P37: Prune and merge multiple potential causal chains to form a systematic set of anomalous causal relationships.

[0039] In one possible embodiment of this application, the causal inference module 30 then executes the process of inferring the potential causal chain of abnormal events and identifying the set of abnormal causal relationships, thereby realizing intelligent diagnosis and traceable management of abnormalities throughout the fuel process, forming an executable logical structure that can be directly used for risk analysis and decision support.

[0040] In a multi-entity semantic relationship network, the causal inference module 30 first locates anomalous event nodes indicated by a multi-dimensional spatiotemporal tensor. During this process, the system continuously receives dynamic monitoring data streams from the spatiotemporal coupling analysis module and identifies nodes in the spatiotemporal tensor that reflect time delays, geographical offsets, numerical mutations, or state imbalances based on preset anomaly detection rules and statistical thresholds. When the behavioral indicators or flow characteristics of a node deviate from the normal range, it is automatically marked as an anomalous node, and its key attribute information is extracted, including event time, location, associated stage, and upstream and downstream dependencies. By accurately locating these nodes, the module can provide a clear starting point for subsequent causal inference.

[0041] Based on the identified anomalous event nodes, causal reasoning is performed along the directed edges in the semantic relationship network. Starting from the anomalous node, the system traverses possible upstream nodes in reverse according to the defined directional relationships in the semantic relationship network, gradually searching for potential causes that may have led to the anomaly. During the reasoning process, the system dynamically evaluates the causal weight of each edge, combining factors such as time sequence, spatial distance, and business logic to filter out logically consistent causal paths. For example, when the system detects an abnormal laboratory indicator, the reasoning engine traces back to connected nodes such as sampling operations, transportation batches, and warehousing conditions, analyzing whether these processes exhibit abnormal behaviors such as timeouts, abnormal temperature and humidity, or quality fluctuations. Simultaneously, to ensure the accuracy of the reasoning results, a multi-dimensional constraint mechanism is introduced during the reasoning process. Time constraints ensure that the causal chain conforms to the temporal logic of cause and effect; spatial constraints ensure that there are actual geographical connections or operational dependencies between nodes; and logical constraints ensure that the reasoning process does not cross business levels or unrelated links. Through the superposition of these constraints, false causal relationships can be effectively eliminated, retaining only potentially meaningful causal chains.

[0042] After obtaining multiple potential causal chains, the causal inference module 30 prunes and fuses these chains. Pruning aims to remove irrelevant or redundant causal relationships to reduce noise and improve the accuracy of the inference results. Fusion integrates chains with similar or related causal relationships to form a more systematic set of anomalous causal relationships. This process not only simplifies the representation of causal relationships but also reveals the intrinsic connections between different potential causal chains, thus providing a comprehensive and systematic view of anomalous causal relationships. The final set of anomalous causal relationships clearly demonstrates the multiple possible causes of anomalous events and their interactions, providing important causal evidence for risk assessment and early warning.

[0043] The comprehensive risk assessment module 40 is used to extract multi-dimensional risk factors by taking the abnormal causal relationship set as input and combining it with the spatiotemporal coupling model, and dynamically weighting them to generate a comprehensive risk index.

[0044] Furthermore, the comprehensive risk assessment module 40 is also used to perform the following steps: P41: Analyze the core risk factors in the set of abnormal causal relationships, locate the nodes and edges related to the core risk factors in the spatiotemporal coupling model, and extract quantified multidimensional risk factors; P42: Dynamically allocate the weight coefficients of each risk factor according to the strength and frequency of each causal path in the set of abnormal causal relationships; P43: Based on the weight coefficients, fuse the multidimensional risk factors to generate a comprehensive risk index that characterizes the global risk.

[0045] Specifically, the comprehensive risk assessment module 40 of this application is responsible for transforming complex causal analysis results into quantifiable risk indicators. This module takes the set of abnormal causal relationships identified by the causal inference module 30 as input, and extracts multidimensional risk factors in conjunction with a spatiotemporal coupling model, and then dynamically weights and generates a comprehensive risk index.

[0046] In the specific execution process, the core risk triggers within the set of abnormal causal relationships are first analyzed. Core risk triggers refer to key factors that repeatedly occur in multiple causal chains, have a wide impact, or carry high weight, such as transportation delays, sampling lags, equipment failures, laboratory deviations, and data anomalies. By calculating the frequency and propagation impact of each risk trigger in the causal chain using statistical analysis methods and ranking them by importance, the core risk triggers can be extracted. Subsequently, the nodes and edges related to these core risk triggers are located in a spatiotemporal coupling model. In this way, quantitative multidimensional risk factors can be extracted. These risk factors may include multiple dimensions such as time delays, geographical coupling relationships, equipment status, and environmental conditions, which together constitute the basic data for risk assessment.

[0047] Next, the extracted risk factors are dynamically weighted based on the strength and frequency of each causal path in the abnormal causal relationship set. To ensure a scientific and reasonable weight allocation, a hierarchical weighting mechanism is adopted. First, the causal strength (i.e., the ability to propagate abnormalities) and occurrence frequency (i.e., the historical probability of occurrence) of each causal path are calculated and normalized to an initial weight. Second, these weights are mapped to the corresponding risk factor groups, so that the weights of the risk factors can dynamically change with the system's operating status and the distribution of abnormalities. For example, when the system continuously detects transportation-related abnormalities, the weight of transportation-related risk factors will automatically increase, while the weight of laboratory-related factors with lower correlation will decrease accordingly.

[0048] Finally, after weight allocation, the multi-dimensional risk factors are fused and calculated to generate a comprehensive risk index. The fusion calculation employs a weighted aggregation model, superimposing risk factors of different categories and dimensions according to their weight coefficients. Simultaneously, a nonlinear fusion function is introduced into the calculation to handle the interactions and non-independent relationships between different risk factors, thereby avoiding risk underestimation or distortion caused by simple linear weighting. The resulting comprehensive risk index is a dynamically changing quantitative indicator over time, with its value range corresponding to the risk level interval. A higher risk index indicates a higher risk state for the system. The generation of the comprehensive risk index provides an intuitive quantitative indicator for risk management and early warning, enabling decision-makers to quickly understand the current risk situation and take appropriate measures.

[0049] The graded early warning and control module 50 is used to automatically classify early warning levels based on the dynamic distribution range of the comprehensive risk index, and generate a multi-dimensional early warning matrix by combining the anomaly type, frequency and correlation, so as to carry out graded early warning and control.

[0050] Furthermore, the hierarchical early warning and control module 50 is also used to perform the following steps: P51: Obtain the time series distribution of the historical comprehensive risk index and perform density clustering to define the threshold ranges for different warning levels; P52: Match the comprehensive risk index with the threshold ranges to classify the current warning level; P53: Integrate the current warning level with the anomaly type, frequency and correlation degree extracted from the abnormal causal relationship set, and fill them together into a preset matrix structure to generate a multi-dimensional warning matrix.

[0051] Optionally, the hierarchical early warning and control module 50 of this application is responsible for classifying early warning levels according to the comprehensive risk index and generating a multi-dimensional early warning matrix to achieve hierarchical early warning and control.

[0052] First, the system acquires time-series distribution data of the historical comprehensive risk index and performs statistical analysis using a density clustering algorithm. The system expands the risk index recorded within a certain period, such as the past three months or one year, along the time dimension and uses density peak clustering to identify the distribution characteristics and clustering trends of the risk index in different intervals. By analyzing the concentrated distribution areas and variation gradients of risk values, the system automatically defines threshold intervals for different warning levels. For example, the green interval represents normal status, the yellow interval represents mild risk, the orange interval represents moderate risk, and the red interval represents severe risk. This method differs from traditional fixed threshold divisions; instead, it uses dynamic interval divisions based on adaptive data analysis, which can automatically adjust according to changes in system operating characteristics and risk trends, making the warning level division more objective and adaptable.

[0053] Next, the calculated comprehensive risk index is matched with the defined threshold range to determine the current warning level. When the system detects that the risk index is in different risk level ranges, the module will automatically switch the corresponding risk status indicator and generate a warning signal. For example, when the comprehensive risk index falls into the orange range, the system will trigger a medium risk warning, prompting managers to pay attention to specific processes or batches; when the risk index exceeds the red range, the system will immediately enter high-risk mode.

[0054] Finally, the current warning level is integrated with the anomaly types, frequencies, and correlations extracted from the anomaly causal relationship set. This information is then populated into a pre-defined matrix structure to generate a multi-dimensional warning matrix. This multi-dimensional warning matrix is ​​a comprehensive data structure that not only includes the current risk level but also encompasses multiple dimensions such as the type of abnormal event, its frequency of occurrence, and its correlation with other events. For example, the vertical axis of the matrix represents the anomaly type, the horizontal axis represents the risk level, and the matrix cells store comprehensive indicator values, including risk weights, trigger counts, impact scope, and degree of linkage. This multi-dimensional warning information provides decision-makers with a comprehensive risk view, helping them to more effectively formulate and implement warning and control measures.

[0055] Furthermore, the hierarchical early warning and control module 50 is also used to perform the following steps: P56: Based on the mapping relationship between the numerical combination of each unit in the multidimensional early warning matrix and the current early warning level, perform hierarchical early warning control.

[0056] Specifically, after generating the multi-dimensional early warning matrix, the tiered early warning and control module 50 further implements specific tiered early warning and control strategies based on the mapping relationship between the multi-dimensional early warning matrix and the current early warning level. The system first analyzes the numerical combinations of each unit in the multi-dimensional early warning matrix. The numerical combination of each unit reflects the interaction results between multiple indicators such as risk level, anomaly type, occurrence frequency, and correlation. Based on the correspondence between these combined values ​​and the determined current early warning level, a risk response rule base is established to guide control actions at different levels.

[0057] During operation, when changes in the comprehensive risk index cause an update to the current warning level, the system automatically traverses the multi-dimensional warning matrix, retrieves all matrix units corresponding to that level, and analyzes their numerical characteristic combinations. For example, when the current level is an orange warning, the system will focus on filtering units in the matrix that simultaneously meet the requirements of medium-to-high risk weights and high correlation, in order to identify the risk types and links that have the greatest impact on the system. Subsequently, the warning response strategy engine is invoked to trigger corresponding control measures based on the strategy templates corresponding to different warning levels.

[0058] For low-level warnings, such as yellow, alert and recording measures are implemented, such as sending reminders to relevant responsible persons, marking abnormal status on the monitoring interface, and activating the risk tracking mechanism. For medium-level warnings, such as orange, local intervention measures are triggered, such as adjusting fuel transportation plans, increasing sampling frequency, or restricting the entry of abnormal batches into the warehouse. When the warning level reaches the highest level, such as red, the system will activate the emergency response procedure, including suspending relevant business processes, reporting to the management center, and coordinating with video surveillance and access control systems to implement security controls.

[0059] Furthermore, the system continuously monitors the response effectiveness during the implementation of tiered early warning and control measures. After measures are implemented, feedback data is automatically collected, and the control results are compared with the trend of risk index changes to assess the effectiveness of the measures. If the risk is not mitigated or abnormal indicators continue to rise, the system will automatically upgrade the response level to a higher level of control; conversely, when risk indicators gradually decline and stabilize within a safe range, the system will downgrade and gradually restore normal operation. This mechanism enables the system to automatically match an appropriate response level based on the real-time risk status, ensuring timely intervention in the early stages of risk and rapid control in high-risk phases, thereby minimizing the impact of abnormal events on fuel supply and production safety.

[0060] Furthermore, the hierarchical early warning and control module 50 is also used to perform the following steps: P57: For high-level early warning events, reverse path search is performed based on the spatiotemporal coupling model to track the risk source nodes and propagation paths, automatically generate event causal chains, and map the event causal chains into event responsibility chains based on the multi-entity semantic relationship network; P58: Based on the event causal chains and event responsibility chains, a closed-loop handling strategy for targeted push is matched and generated from the preset handling plan library.

[0061] It should be understood that when handling high-level early warning events, the tiered early warning and control module 50 enters the proactive intervention and closed-loop decision-making stage. This stage is used to automatically generate risk tracing, responsibility identification, and handling strategies for high-level early warning events. The core objective of this stage is to enable the system to automatically trace the source of a serious risk, clarify the causal chain of responsibility, and intelligently generate targeted emergency and handling measures based on established plans after detecting a serious risk, ensuring that risk events can be controlled quickly, accurately, and in a closed loop.

[0062] When encountering a high-level warning event, the tiered warning and control module 50 first performs a reverse path search based on a spatiotemporal coupling model. Starting from the abnormal node that triggered the warning, the search proceeds in the reverse direction along the time and space dimensions of the graph structure of the spatiotemporal coupling model, tracing back to possible upstream influencing nodes layer by layer. During the search, the system comprehensively considers the temporal order, geographical location correlation, and edge weights between nodes, representing delays or coupling strength, and uses a reverse traversal algorithm to locate the starting point and main propagation chain of the risk event. Through this process, the evolution path of the risk event can be reconstructed, generating a structured event causal chain that clearly demonstrates the complete propagation logic of the event from the initial trigger to the final abnormal result.

[0063] After generating the event causal chain, the module further maps the event causal chain to an event responsibility chain based on a multi-entity semantic relationship network. This mapping process utilizes the semantic relationships between entities defined in the multi-entity semantic relationship network to associate the nodes and edges in the causal chain with the corresponding responsible parties. For example, if a risk propagates due to a device malfunction, the maintenance department of that device will be mapped as a responsible party in the responsibility chain. Through this mapping, the module can clarify the roles and responsibilities of each responsible party in the risk propagation process, providing a basis for subsequent handling strategies.

[0064] Finally, based on the generated event causal chain and event responsibility chain, a targeted closed-loop response strategy is generated and matched from a pre-set contingency plan library. This process first analyzes the key nodes and risk types in the causal chain, extracts risk trigger characteristics such as equipment failure, process abnormalities, and operational delays, and performs feature matching with rule templates in the contingency plan library. Subsequently, the most suitable contingency plan is selected from the library based on this information. For example, if the risk source is a failure of a certain piece of equipment, the response strategy might include immediately activating backup equipment and notifying the maintenance team for emergency repairs. The generated closed-loop response strategy not only includes specific response measures but also covers monitoring and feedback mechanisms for the response process. This ensures the effective execution of response measures and allows for adjustments based on actual conditions. For example, while performing repair measures, the monitoring system tracks the repair progress and effectiveness in real time, and adjusts the response strategy promptly upon detecting new risk indicators. This mechanism can not only quickly identify the risk source and clarify the boundaries of responsibility but also intelligently generate the optimal response path based on historical contingency plans, significantly improving the system's risk response efficiency and governance transparency.

[0065] In summary, the embodiments of this application have at least the following technical effects: This application solves the problems of data dispersion and information isolation in traditional systems by uniformly collecting and spatiotemporally identifying multi-source heterogeneous data throughout the entire fuel process; it constructs a spatiotemporal coupling model through graph neural networks to accurately depict the dynamic correlation between each link; it adopts a causal inference mechanism to identify the potential causal chain of abnormal events, realizing the tracing of the root causes of anomalies and the analysis of propagation paths; it dynamically generates a comprehensive risk index based on multidimensional risk factors to improve the real-time performance and accuracy of risk assessment; and it achieves intelligent risk identification, targeted response, and fully automated management of the entire process through a hierarchical early warning and closed-loop control mechanism.

[0066] It has achieved the technical effect of precise monitoring and real-time early warning of the entire fuel process through an intelligent management and control system based on spatiotemporal coupling and causal inference.

[0067] Example 2, based on the same inventive concept as the aforementioned example of a smart management and early warning platform for the entire fuel process, such as... Figure 2 As shown, this application provides a smart management and early warning method for the entire fuel process. The system and method embodiments in this application are based on the same inventive concept. The method includes: In each stage of the fuel production process, multi-source heterogeneous data is collected in real time and appended with timestamps and geographic location tags. A spatiotemporal coupling model of fuel flow is constructed based on a graph neural network, using the multi-source heterogeneous data as input. Event nodes represent the state of each stage, and edge weights represent time delay and geographic coupling relationships, generating a multi-dimensional spatiotemporal tensor. Feature embedding and entity extraction are performed on the multi-dimensional spatiotemporal tensor to construct a multi-entity semantic relationship network and infer the potential causal chains of abnormal events, identifying abnormal causal relationship sets. Using the abnormal causal relationship sets as input, multi-dimensional risk factors are extracted in combination with the spatiotemporal coupling model, and a comprehensive risk index is dynamically weighted and generated. Based on the dynamic distribution range of the comprehensive risk index, early warning levels are automatically divided, and a multi-dimensional early warning matrix is ​​generated by combining the anomaly type, frequency, and correlation degree for hierarchical early warning control.

[0068] Furthermore, the method also includes: For high-level early warning events, reverse path search is performed based on the spatiotemporal coupling model to track the risk source nodes and propagation paths, automatically generate event causal chains, and map the event causal chains into event responsibility chains based on the multi-entity semantic relationship network; based on the event causal chains and event responsibility chains, a closed-loop response strategy is matched and generated from the preset response plan library for targeted push.

[0069] Furthermore, a spatiotemporal coupled model of fuel flow is constructed based on graph neural networks, including: The spatiotemporal coupling model is initialized with key events in the entire fuel process as nodes and the flow logic between events as edges. Historical multi-source heterogeneous data is collected to train the graph neural network until the model loss function converges, thereby obtaining a spatiotemporal coupling model that represents the spatiotemporal dynamics of fuel flow.

[0070] Furthermore, using the multi-source heterogeneous data as input, event nodes represent the state of each stage, and edge weights represent the time delay and geographical coupling relationship, a multi-dimensional spatiotemporal tensor is generated, including: The real-time collected multi-source heterogeneous data is transformed into node feature vectors and edge feature vectors required by the spatiotemporal coupling model; the node feature vectors and edge feature vectors are input into the spatiotemporal coupling model for forward inference, and the hidden states of all nodes output by the model in each time slice are aggregated and structurally assembled into the multidimensional spatiotemporal tensor.

[0071] Furthermore, feature embedding and entity extraction are performed on the multidimensional spatiotemporal tensor to construct a multi-entity semantic relationship network, including: The multidimensional spatiotemporal tensor is reduced in dimensionality using a nonlinear transformation to obtain a low-dimensional dense feature embedding representation; based on the feature embedding representation, key entities are identified and extracted; semantic relationships between the key entities are defined and established according to business processes and spatial topology; the key entities and semantic relationships are vectorized to construct a multi-entity semantic relationship network.

[0072] Furthermore, inferring the potential causal chains of anomalous events and identifying sets of anomalous causal relationships, including: In the multi-entity semantic relationship network, the abnormal event node indicated by the multi-dimensional spatiotemporal tensor is located; based on the abnormal event node, causal reasoning is performed along the directed edges in the semantic relationship network to infer the potential causal chain of the abnormal event; multiple potential causal chains are pruned and merged to form a systematic set of abnormal causal relationships.

[0073] Furthermore, using the aforementioned set of abnormal causal relationships as input, and combining it with the spatiotemporal coupling model to extract multidimensional risk factors, a comprehensive risk index is dynamically weighted and generated, including: The core risk factors in the abnormal causal relationship set are analyzed, and the nodes and edges related to the core risk factors are located in the spatiotemporal coupling model to extract quantified multidimensional risk factors. According to the strength and frequency of each causal path in the abnormal causal relationship set, the weight coefficients of each risk factor are dynamically allocated. Based on the weight coefficients, the multidimensional risk factors are fused to generate a comprehensive risk index that represents global risk.

[0074] Furthermore, based on the dynamic distribution range of the comprehensive risk index, early warning levels are automatically classified, and a multi-dimensional early warning matrix is ​​generated by combining anomaly type, frequency, and correlation, including: The time series distribution of the historical comprehensive risk index is obtained and density clustering is performed to define the threshold ranges for different warning levels. The comprehensive risk index is matched with the threshold ranges to classify the current warning level. The current warning level is then fused with the anomaly type, frequency, and correlation degree extracted from the abnormal causal relationship set and filled into a preset matrix structure to generate a multi-dimensional warning matrix.

[0075] Furthermore, based on the mapping relationship between the numerical combination of each unit in the multidimensional early warning matrix and the current early warning level, hierarchical early warning control is implemented.

[0076] It should be noted that the order of the embodiments described above is merely for descriptive purposes and does not represent the superiority or inferiority of the embodiments. Furthermore, the above description focuses on specific embodiments of this specification. Additionally, the processes depicted in the accompanying drawings do not necessarily require a specific or sequential order to achieve the desired results. In some implementations, multitasking and parallel processing are possible or may be advantageous.

[0077] The above description is only a preferred embodiment of this application and is not intended to limit this application. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of this application should be included within the protection scope of this application.

[0078] This specification and accompanying drawings are merely illustrative examples of this application and are intended to cover any and all modifications, variations, combinations, or equivalents within the scope of this application. Clearly, those skilled in the art can make various alterations and modifications to this application without departing from its scope. Therefore, if such modifications and modifications fall within the scope of this application and its equivalents, this application intends to include such modifications and modifications.

Claims

1. A smart management and early warning platform for the whole process of fuel, characterized in that, The platform comprises: a heterogeneous data collection module, configured to collect multi-source heterogeneous data in real time in each link of a fuel whole process, and to attach a time stamp and a geographical position label; a space-time coupling analysis module, configured to construct a space-time coupling model of fuel flow based on a graph neural network, to input the multi-source heterogeneous data as an input, to represent a state of each link by an event node, to represent a time delay and a geographical coupling relationship by an edge weight, and to generate a multi-dimensional space-time tensor; a causal inference module, configured to perform feature embedding and entity extraction on the multi-dimensional space-time tensor, to construct a multi-entity semantic relationship network, to infer a potential causal chain of an abnormal event, and to identify an abnormal causal relationship set; a comprehensive risk assessment module, configured to input the abnormal causal relationship set, to extract a multi-dimensional risk factor in combination with the space-time coupling model, and to dynamically weight to generate a comprehensive risk index; a hierarchical early warning control module, configured to automatically divide early warning levels based on a dynamic distribution interval of the comprehensive risk index, to generate a multi-dimensional early warning matrix in combination with an abnormal type, a frequency, and a correlation degree, and to perform hierarchical early warning control.

2. The intelligent management and early warning platform for the whole process of fuel according to claim 1, characterized in that, The hierarchical early warning control module is further configured to: for a high-level early warning event, perform reverse path search based on the space-time coupling model, track a risk source node and a propagation path, automatically generate an event causal chain, and map the event causal chain to an event responsibility chain based on the multi-entity semantic relationship network; based on the event causal chain and the event responsibility chain, match and generate a closed-loop treatment strategy for directional push from a preset treatment plan library.

3. A smart management and early warning platform for the whole process of fuel according to claim 1, characterized in that, When constructing the space-time coupling model of fuel flow based on the graph neural network, the space-time coupling analysis module is further configured to: initialize a graph structure of the space-time coupling model by taking a key event of a fuel whole process as a node and a flow logic between events as an edge; collect historical multi-source heterogeneous data to train the graph neural network until a model loss function converges, to obtain a space-time coupling model representing space-time dynamics of fuel flow.

4. A smart management and early warning platform for the whole process of fuel according to claim 2, characterized in that, When inputting the multi-source heterogeneous data as an input, representing a state of each link by an event node, representing a time delay and a geographical coupling relationship by an edge weight, and generating a multi-dimensional space-time tensor, the space-time coupling analysis module is further configured to: convert the real-time collected multi-source heterogeneous data into a node feature vector and an edge feature vector required by the space-time coupling model; input the node feature vector and the edge feature vector into the space-time coupling model for forward reasoning, converge hidden states of all nodes in each time slice output by the model, and structure and assemble the multi-dimensional space-time tensor.

5. A smart management and early warning platform for the whole process of fuel according to claim 1, characterized in that, When performing feature embedding and entity extraction on the multi-dimensional space-time tensor, and constructing a multi-entity semantic relationship network, the causal inference module is further configured to: perform dimension reduction on the multi-dimensional space-time tensor by using a nonlinear transformation to obtain a low-dimensional dense feature embedding representation; identify and extract key entities based on the feature embedding representation; define and establish semantic relationships between the key entities according to a business process and a spatial topology; vectorize the key entities and semantic relationship vectors to construct a multi-entity semantic relationship network.

6. A smart management and early warning platform for the whole process of fuel according to claim 5, characterized in that, When inferring a potential causal chain of an abnormal event and identifying an abnormal causal relationship set, the causal inference module is further configured to: In the multi-entity semantic relationship network, an abnormal event node indicated by the multi-dimensional spatio-temporal tensor is located; Based on the abnormal event node, a potential causal chain of the abnormal event is inferred by tracing the directed edges in the semantic relationship network; The multi-dimensional spatio-temporal tensor is pruned and fused to form a systematic abnormal causal relationship set.

7. A smart management and early warning platform for the whole process of fuel according to claim 1, characterized in that, The comprehensive risk assessment module is also used for: Analyzing the core risk inducement in the abnormal causal relationship set, and locating the nodes and edges related to the core risk inducement in the spatio-temporal coupling model to extract quantified multi-dimensional risk factors; According to the strength and frequency of each causal path in the abnormal causal relationship set, the weight coefficients of each risk factor are dynamically allocated; Based on the weight coefficients, the multi-dimensional risk factors are fused to generate a comprehensive risk index representing the global risk.

8. The intelligent management and early warning platform for the whole process of fuel according to claim 1, characterized in that, The hierarchical early warning control module is also used for: Obtaining the time series distribution of the historical comprehensive risk index to perform density clustering, and defining the threshold interval of different early warning levels; Matching the comprehensive risk index with the threshold interval to divide the current early warning level; Fusing the current early warning level, the abnormal type, frequency and correlation degree extracted from the abnormal causal relationship set into a preset matrix structure to generate a multi-dimensional early warning matrix.

9. A smart management and early warning platform for the whole process of fuel according to claim 8, characterized in that, The hierarchical early warning control module is also used for: According to the mapping relationship between the numerical combination of each cell in the multi-dimensional early warning matrix and the current early warning level, hierarchical early warning control is performed.

10. A method for intelligent management and early warning of the whole process of fuel, characterized in that, The method comprises: In each link of the fuel whole process, multi-source heterogeneous data are collected in real time, and time stamps and geographic location labels are attached; Based on a graph neural network, a spatio-temporal coupling model of fuel flow is constructed, and the multi-source heterogeneous data are taken as input, each link state is represented by an event node, the time delay and geographic coupling relationship are represented by edge weight, and a multi-dimensional spatio-temporal tensor is generated; The multi-dimensional spatio-temporal tensor is subjected to feature embedding and entity extraction to construct a multi-entity semantic relationship network, and a potential causal chain of an abnormal event is inferred to identify an abnormal causal relationship set; Taking the abnormal causal relationship set as input, multi-dimensional risk factors are extracted in combination with the spatio-temporal coupling model, and a comprehensive risk index is dynamically weighted generated; Based on the dynamic distribution interval of the comprehensive risk index, early warning levels are automatically divided, and a multi-dimensional early warning matrix is generated in combination with the abnormal type, frequency and correlation degree for hierarchical early warning control.

Images