Gas station equipment monitoring system and method based on Internet of Things

Abstract

The invention relates to the technical field of gas station equipment monitoring, in particular to a gas station equipment monitoring system and method based on the Internet of Things. The method specifically comprises the steps of collecting multi-dimensional operation parameters of equipment and business process node dynamics in real time, constructing a bidirectional coupling graph model of the equipment and business, and extracting high-order features of physical association and business logic interaction between the equipment by using a graph convolutional network; predicting equipment fault risks by adopting a dynamic health scoring mechanism, backtracking a risk propagation chain in combination with an attention mechanism, and mapping multi-dimensional data into interpretable fault diagnosis information through an abnormal semantic generation network; a double-layer reinforcement learning game mechanism is introduced, an inner layer coordinates equipment maintenance and a risk suppression strategy, an outer layer overall plans service efficiency and a safety target, optimization actions such as resource scheduling and oil product allocation are generated, and a perception-diagnosis-decision closed loop is formed. According to the method, full-time perception of the equipment state, active fault early warning and cross-domain collaborative optimization are realized, and the operation and maintenance efficiency and safety of the gas station are remarkably improved.

Description

Technical Field

[0001] The present invention relates to the technical field of gas station equipment monitoring, and particularly to an Internet of Things-based gas station equipment monitoring system and method. Background Art

[0002] With the expansion of the operation scale of gas stations and the complexity of service requirements, the equipment monitoring system needs to cope with challenges such as the perception of the collaborative operation status of multiple devices, the adaptation of business dynamic loads, and the prevention and control of risk propagation chains; traditional gas station equipment monitoring is based on manual inspections and local sensor alarms, making it difficult to meet the requirements of real-time data fusion analysis and cross-domain collaborative optimization.

[0003] A Chinese invention patent with the publication number CN110702852B discloses an intelligent monitoring system based on multiple oil and gas concentration sensors of the Internet of Things. The system consists of an environmental parameter acquisition platform for the gas station oil tank area based on the ZigBee network and a monitoring subsystem for the environmental oil and gas concentration sensors in the gas station oil tank area; effectively solving the problem that the existing gas station oil tank area environmental monitoring system does not accurately detect the environmental oil and gas concentration in the gas station oil tank area and give early warnings of sensor failures according to the non-linear, large lag, and complex changes in the environmental oil and gas concentration in the gas station oil tank area, thus improving the accuracy and robustness of predicting the oil and gas concentration in the gas station.

[0004] However, traditional gas station equipment monitoring still has problems such as information silos, lagging responses, and low data utilization rates. With the development of the Internet of Things, cloud computing, and big data analysis, a new system that can achieve full-scale perception of equipment status, dynamic behavior modeling, intelligent optimization scheduling, and proactive fault warning is needed to improve the intelligence and refinement level of gas station equipment management. Summary of the Invention

[0005] The object of the present invention is to propose an Internet of Things-based gas station equipment monitoring system and method for the problems existing in the background art.

[0006] The technical solution of the present invention: An Internet of Things-based gas station equipment monitoring method includes the following specific implementation steps: S1. Multidimensional operating parameters and dynamic business process nodes of key equipment in the gas station; S2. By constructing a set of devices and business process nodes and associating state vectors to describe the operating characteristics, using the attention mechanism to dynamically adjust the coupling weights, and combining with the graph convolutional network to extract high-order interaction features, an equipment-business coupling model is established; S3. By fusing the equipment status, the influence of neighbor nodes, and the coupling weights through the attention mechanism to construct a comprehensive health representation vector, using a lightweight evaluation function to quantify the equipment health score, combining with the graph time series model to predict future fault risks, and backtracking the cumulative influence weights to identify the risk chain path; S4. Construct a multi-dimensional anomaly feature vector by extracting device propagation features, use semantic generation network to map anomaly semantic categories, combine the device-business coupling model with the risk chain to deduce the anomaly causal path, and finally generate anomaly semantic recognition results; S5. By constructing an operation state vector that includes equipment health, business indicators, and safety risks, defining a multi-objective optimization function and a two-layer reinforcement game optimization strategy generation mechanism, and dynamically adjusting the strategy priority in combination with abnormal semantic interpretation, the optimal operation decision is generated; S6. Integrate the optimal operational decision and abnormal analysis results, jointly execute automated resource scheduling, combine manual intervention and continuous monitoring and feedback to dynamically adjust the strategy.

[0007] Preferably, the construction process of the device-service coupling model is: S21. Define device collection , and define a collection of business process nodes , build a set of all nodes ; Among them, e i represents the i-th key equipment node in the gas station; b j represents the jth business node of the business process; m represents the total number of key equipment in the gas station; n represents the total number of business process steps; S22. Define each device node e i and business node b j The associated state vector , : ; ; in, Represents the kth state feature of the i-th device node; represents the lth process feature of the jth service node; M represents the total number of state features in the i-th device node; L represents the total number of process features of the j-th service node; S23. Define the device-service initial coupling association matrix A: ; Among them, A uv Represents the initial connection weight between nodes u and v; S24, introduce the dynamic coupling weight tensor W(t) to perform adaptive weighted correction on the initial correlation matrix: ; ; Where W(t) represents the dynamic coupling weight tensor at time t; , A set of state vectors representing devices and services respectively; Denotes element-wise product; Denotes the device-service coupling graph after dynamic weighting; Denotes the weight learning function; S25. Apply the graph convolutional network GCN to perform joint representation learning on device and service nodes, and extract high-order interaction features between nodes: ; ; Among them, , Denote the node representation of the (l + 1)-th layer and the node representation of the l-th layer respectively; Denotes the trainable weight of the l-th layer; Denotes the ReLU activation function; D denotes the degree matrix; Denotes the dynamic weighted connection weight between node i and node j at time t; Denotes the total connection weight of node i; S26. Output the device-service coupling model G(t): ; Among them, N denotes the node set; Denotes the connection relationship between nodes after dynamic weighting; Denotes the joint embedding representation of each node.

[0008] Preferably, the risk chain path identification process is as follows: S31. For each device node , construct its comprehensive health representation vector: ; Among them, Denotes the current device state vector; Denotes the neighbor nodes coupled with the current device node e i ; Denotes the dependence strength learned by the attention mechanism; W 1 Denotes the weight matrix; Denotes the non-linear activation function ReLU; Denotes the comprehensive health representation vector of device e i ; S32. Use the comprehensive health representation vector i of device e to quantitatively score the current health state of device e i : ; Among them, Denotes the health score, that is, the health score of device e i at time t. The closer it is to 0, the more abnormal it is, ; w 2a and b respectively represent the weights and bias terms obtained from model training; here, represents the vector transpose operation; set the health risk threshold θ, when it is considered that there is a health risk and enter the prediction process; S33. Select the health representation sequences at the past time moments and input them into the graph time series prediction model GGRU: ; Further predict the future fault risk score: ; Among them, represents the health score of the prediction device e i at the future τ time steps, that is, the health score at the future time t + τ; represents the health representation of the prediction device e i at the future time t + τ; represents the graph gated recurrent unit model, taking into account both structure perception and time series memory capabilities; S34. Identify high-risk devices through the health score threshold, extract the cumulative impact weights and reverse backtrack the key upstream nodes until the termination condition, and output the risk chain path.

[0009] Preferably, the implementation process of extracting the cumulative impact weights and reverse backtracking the key upstream nodes until the termination condition and outputting the risk chain path is as follows: A1. For each device e i , according to the health score obtained from the future prediction, set the health risk threshold θ low . If it satisfies: < θ low , then determine that the device e i is a potential fault high-risk node; A2. Extract the influence weight of each edge (ei, ej) in the device-business coupling model G(t) on the node health state at time t, and construct the cumulative influence weight matrix of each node within the time window [t - NumT + 1, t]: ; Among them, represents the cumulative health impact degree of the device e j on the device e i within the past NumT time moments; A3. For each high-risk device node e i , perform the following backtracking: Starting from e i , reverse search its direct neighbor node set (i.e., the nodes that affected it in the past); Arrange in descending order according to the cumulative influence weight and select the node e whose cumulative influence weight is greater than the set threshold γ j ; Recursively backtrack the upstream influence nodes of e j and continuously construct the path. The termination conditions are: the cumulative weight drops below the threshold; the length of the backtracked path reaches the maximum depth d max ; reaching the source service node; A4. Finally, output the risk chain path Chain(e i ).

[0010] Preferably, the process of generating the abnormal semantic recognition result is as follows: S41. Extract the propagation feature vector of device e i ; ; S42. For each device e predicted to have potential high risk i , construct a set of multi-dimensional abnormal feature vectors by integrating the current state, prediction trend, business metric changes, and link propagation characteristics : ; Among them, represents the health score of device e i at time t; represents the predicted health score of device e i in the future τ time steps; represents the amplitude of health change, ; represents the current set of key business metrics; represents the propagation feature vector of device e i ; S43. Introduce the abnormal semantic generation network SGN, and the goal is to map to the specific abnormal semantic category set C through the feature vector ; Define the classification probability distribution : ; S44. According to the identified abnormal category, combine the device-business coupling graph G(t) with the risk chain Chain(e i ), and deduce the logical path of abnormal formation; S45. Output the abnormal semantic recognition result.

[0011] Preferably, the process of extracting the propagation feature vector of device e i is as follows: B1. Extract the basic attributes of the link itself: B1. Extract the basic attributes of the link itself: Chain length :From e i The shortest number of propagation steps to the terminal node; Maximum spread width : In any level, the level with the largest number of child nodes; Node aggregation changes : The change in node clustering coefficient along the chain; B2. During the propagation process, extract the risk transfer weight w(e) of each edge, i.e., the connection between devices. i ,e j ), define device e i The cumulative transmission weight of the link : , extract the average single-hop propagation strength : ; Among them, (e u ,e v ) indicates that in the risk chain Chain(e i ), device e u To device v There is a potential risk transfer relationship between them; w(e u ,e v ) indicates the description of the device e u To device v the intensity of transmission when conveying risk; B3. Define device e i The propagation eigenvector of is: .

[0012] Preferably, the abnormal semantic generation network SGN converts multimodal data into abnormal categories with business semantics. The input feature vector includes the current health score, predicted score, trend difference, business key indicators and risk chain propagation characteristics. The multi-layer perceptron MLP is used to extract various feature semantics, and the probability distribution of abnormal semantic categories is generated through the Softmax function.

[0013] Preferably, the optimal operational decision-making process is as follows: S51, construct the current gas station operation state vector S t : ; in, Indicates all devices i A vector of health scores at time t; Represents a business operation indicator vector; represents the current operating cost vector; represents the security risk chain score vector; Represents time context information; S52. Define the action set A = {a 1 , a 2 , …, a t ,..., a K}, where each action a t represents an operation strategy, and construct the optimization objective function: ; Among them, represents the expected optimization revenue of the strategy π; π represents the strategy function, which takes the current state S t as input and outputs the optimal action a t ; T represents the total number of steps in the decision-making cycle, that is, the time step length considered in this round of optimization plan; represents the discount factor, 0 < γ < 1; represents the operation cost function, which is the operation overhead corresponding to the state S t when the action a t is executed; represents the safety and service revenue function; represents the safety-cost regulation weight factor; S53. Construct a two-layer reinforcement game optimization strategy generation mechanism, a two-layer structure of inner-layer device game + outer-layer business game, and model the game tension in the strategy evolution process based on reinforcement learning to generate multi-objective optimization strategies; S54. Introduce a semantic interpretation regulation unit, abnormal semantic interpretation results, adjust the focus of the optimization strategy, and correct the action set generated by reinforcement learning through soft constraints; S55. Select the optimal strategy .

[0014] Preferably, the implementation process of the two-layer reinforcement game optimization strategy generation mechanism is as follows: C1. Participating subjects and state-action modeling: Inner-layer game subject, that is, device agent: Each key device e i is modeled as a reinforcement learning agent Agent i ; Outer-layer game subject, that is, business service agent: The gas station business process is modeled as a service agent Agent b ; C2. Inner-layer game mechanism, that is, device game: Health-chain risk control tension modeling: Model the problem as a repeated game reinforcement learning, and each Agent i learns the strategy and participates in the following local revenue game: ; Among them, represents the device agent e iCumulative revenue within the entire game cycle; Denotes the discount factor, where 0 < γ < 1; Denotes device e i After performing action a t The change in its health score; Denotes the impact of the current device failure on the risk chain propagation of other devices; Denotes the current action cost; α, β, θ denote dynamic weight coefficients; C3, Outer game mechanism, i.e., business game: Global efficiency - risk control coordination strategy: Business layer Agent b Observes the health status and task requirements of all devices, and generates a scheduling strategy based on the global state , whose revenue function: , outputs the policy guiding factors: α, β, θ, and injects the policy guiding factors: α, β, θ into the revenue functions of all device Agents i ; Among them, Denotes the cumulative operating revenue of the business agent within the entire game cycle; Denotes the current business service quality indicator; Denotes the current overall device risk level of the system; Denotes the operating cost caused by the current scheduling strategy; , , Denotes the business strategy preference weight; C4, At each moment t, the game process is as follows: Business agent Agent b Observes the global state S t , adjusts the weights α, β, θ; Each Agent i Under the guidance of the weights, executes device actions ; All device actions affect the system state S t+1 , which is fed back to the business agent; Uses the multi-agent reinforcement learning framework MADDPG to iteratively update the strategy; C5, Defines the composite revenue function of the overall system: ; Among them, Denotes the composite revenue indicator after the overall system strategy is executed; Denotes the weight control parameter of the business layer revenue in the total objective; N denotes the number of devices participating in the inner game; By setting and using the Pareto front screening mechanism, selects the optimal strategy combination of multiple objectives { , }; Among them, Represents the combination of device actions; Represents the scheduling strategy; C6. The final output is an optimization strategy for multiple objectives: the device layer action strategy table, the business layer operation strategy, and the strategy interpretive structure.

[0015] The technical solution of the present invention: An Internet of Things-based gas station equipment monitoring system, which is used to execute the above-mentioned Internet of Things-based gas station equipment monitoring method, including: An Internet of Things collaborative perception module, which is used to collect the operation data of various devices in the gas station in real time and monitor the dynamics of business process nodes at the same time; A device-business two-way modeling engine, which is used to model the association between the device state and the business process state as a two-way coupling graph structure, use the graph structure to show the physical and logical connections between devices, and define the interaction rules between device load and business flow; An abnormal semantic analysis module, which is used to construct a device health joint inference and prediction mechanism under dynamic business load and an abnormal parsing mechanism based on the device-business semantic structure. The abnormal semantic analysis module includes: a device health joint inference unit and an abnormal semantic recognition and interpretation unit; The device health joint inference unit dynamically infers the device health state through a joint inference algorithm, takes into account the current state of the device and future business load changes, accurately predicts the occurrence probability of potential faults, so as to achieve early warning and take measures in advance; The abnormal semantic recognition and interpretation unit deeply analyzes the predicted abnormality, and provides accurate fault diagnosis information for maintenance personnel by understanding the reasons, influence scope and related business processes behind the device abnormality; A dynamic operation optimization module, which is used to comprehensively optimize fault warning, business load and device state. Based on the objective function of minimizing fault risk, operation cost and customer waiting time, an optimal resource scheduling plan is generated through an optimization algorithm; An intelligent alarm and resource scheduling module, which is used for real-time alarm, automatically executes the optimization plan, schedules device resources and arranges necessary maintenance work.

[0016] Compared with the prior art, the above technical solution of the present invention has the following beneficial technical effects: A gas station equipment monitoring system and method based on the Internet of Things provided by the present invention collect real-time equipment operation parameters and dynamic data of business processes; construct a dynamic coupling graph structure, and combine a graph convolutional network to extract high-order interaction features of physical associations and business logics between devices, realizing in-depth collaborative analysis of equipment status and business loads; adopt a joint inference algorithm to predict equipment health risks, combine an attention mechanism to trace the risk propagation chain and generate interpretable fault diagnosis information, effectively solving the ambiguity problem of traditional numerical alarms; through a double-layer reinforcement learning game mechanism, balance the goals of equipment maintenance, risk suppression, and business efficiency, dynamically generate resource scheduling and oil product allocation strategies, significantly reducing operation costs and the risk of fault chain diffusion; realize the automatic execution and closed-loop feedback of optimization strategies, forming a perception-diagnosis-decision integrated system; Compared with traditional manual inspections and static monitoring, this solution breaks through the limitations of information silos, improves data utilization and operation and maintenance response speed, reduces equipment failure rates through predictive maintenance, and optimizes the overall operation safety and economy of gas stations. Brief Description of the Drawings

[0017] Figure 1 It is a system architecture diagram of a gas station equipment monitoring system based on the Internet of Things proposed by the present invention; Figure 2 It is a method flow chart of a gas station equipment monitoring method based on the Internet of Things proposed by the present invention. Detailed Embodiments

[0018] Example 1, as Figure 1 shown, a gas station equipment monitoring system based on the Internet of Things proposed by the present invention includes: an Internet of Things collaborative perception module, a device-business bidirectional modeling engine, an abnormal semantic analysis module, a dynamic operation optimization module, and an intelligent alarm and resource scheduling module.

[0019] The Internet of Things collaborative perception module collects real-time operation data of various devices in the gas station, including but not limited to the status information of oil pumps, fuel dispensers, storage tanks, and payment terminal devices, and at the same time monitors the dynamics of business process nodes (including but not limited to refueling, payment, car washing), and this data flows to the device-business bidirectional modeling engine; The device-business bidirectional modeling engine models the association between the device status and the business process status as a bidirectional coupling graph structure, uses the graph structure to show the physical and logical connections between devices, and defines the interaction rules of device loads and business flows, so as to provide a reliable basis for subsequent reasoning and analysis; The abnormal semantic analysis module constructs a device health joint inference and prediction mechanism under dynamic business loads and an abnormal parsing mechanism based on the device-business semantic structure. The abnormal semantic analysis module includes: a device health joint inference unit and an abnormal semantic recognition and interpretation unit; The device health joint inference unit dynamically infers the device health status through a joint inference algorithm, takes into account the current status of the device and future changes in business load, accurately predicts the occurrence probability of potential faults, and thus realizes early warning and takes measures in advance; The abnormal semantics recognition and interpretation unit deeply analyzes the predicted anomalies, and provides accurate fault diagnosis information for maintenance personnel by understanding the reasons, influence scope and related business processes behind the device anomalies, avoiding simple numerical alarms; The dynamic operation optimization module comprehensively optimizes fault warnings, business loads and device status. Based on the objective function of minimizing fault risks, operation costs and customer waiting times, it generates an optimal resource scheduling plan through an optimization algorithm, including but not limited to automatically switching to standby devices and adjusting the oil supply path; The intelligent alarm and resource scheduling module gives real-time alarms and automatically executes the optimization plan, schedules device resources and arranges necessary maintenance work.

[0020] Embodiment 2, as Figure 2 shown, a method for monitoring gas station equipment based on the Internet of Things proposed by the present invention is applied to a system for monitoring gas station equipment based on the Internet of Things proposed in Embodiment 1, and its specific implementation steps are as follows: S1. The Internet of Things collaborative perception module collects multi-dimensional operation parameters of key equipment in the gas station (including but not limited to oil pumps, fuel dispensers, storage tanks, pipelines, payment terminals) in real time through the deployment of a multi-source sensor network, including but not limited to flow rate, pressure, temperature, vibration, current, voltage and payment processing status; At the same time, business process nodes (including but not limited to customer queuing, payment completion, and oil product switching) are synchronously monitored as perception elements; The perception data stream is transmitted to the device-business two-way modeling engine.

[0021] S2. Based on the collected perception data stream, the device-business two-way modeling engine establishes a two-way coupling model of the device state vector and the business process state vector, uses a graph structure to represent the physical association of devices (such as the connection relationship between the oil pump and the fuel dispenser) and the business logic flow (such as the order of refueling → payment → car wash), and at the same time defines the interaction rules between the device load and the business dynamics, which is used to dynamically infer the trend of the device operation pressure with the change of the business. The specific implementation process is as follows: S21. Obtain the perception data stream and define the device set , and define the business process node set , and construct the entire node set ; Among them, e i represents the i-th key device node in the gas station, including but not limited to oil pumps, fuel dispensers, storage tanks; b jDenote the j-th business node of the business process, including but not limited to the start of refueling, payment confirmation, and fuel type switching; m represents the total number of key devices in the gas station; n represents the total number of business process steps; S22. For each device node e i and business node b j respectively associate a state vector, denoted as and respectively, which is used to describe its current operating state or business state characteristics: ; ; Among them, represents the k-th state characteristic of the i-th device node, including but not limited to flow rate, current, and vibration amplitude; represents the l-th process characteristic of the j-th business node, including but not limited to queue length, payment response time, and refueling completion rate; M represents the total number of state characteristics in the i-th device node; L represents the total number of process characteristics in the j-th business node; S23. Define the device-business initial coupling association matrix A: ; Among them, A uv represents the preliminary connection weight between nodes u and v; It should be noted that physical association: such as the oil pump driving the oil gun to supply oil; logical association: such as payment confirmation affecting the start and stop of fuel type switching; S24. In order to reflect the real-time influence intensity between device state changes and business load changes, introduce the dynamic coupling weight tensor W(t) to perform adaptive weighted correction on the initial association matrix: ; ; Among them, W(t) represents the dynamic coupling weight tensor at time t; , respectively represent the set of state vectors of devices and businesses; represents the element-wise product; represents the device-business coupling graph after dynamic weighting; represents the weight learning function (adaptive learning based on the attention mechanism in this embodiment); S25. After obtaining the dynamic coupling graph , apply the graph convolutional network (GCN) to perform joint representation learning on device and business nodes, and extract the high-order interaction features between nodes: ; ; Among them, , respectively represent the node representation of the (l + 1)-th layer and the node representation of the l-th layer; represents the trainable weight of the l-th layer; represents the activation function (ReLU is adopted in this embodiment); D represents the degree matrix; represents the dynamic weighted connection weight between node i and node j at time t; represents the total connection weight of node i (including the sum of its weighted connection values with all other device nodes and service nodes); S26. Output the device-service coupling model G(t): ; Among them, N represents the node set (device nodes + service nodes); represents the connection relationship between nodes after dynamic weighting (reflecting the time-varying association strength between devices and services); represents the joint embedding representation of each node (the device-service interaction feature vector obtained through graph convolution learning).

[0022] S3. The device health joint inference unit dynamically infers the health status and failure probability of each device at the current and future time periods according to the constructed device-service coupling model. The inference process comprehensively considers the current perception state of the device, the business load pressure, and the historical operation mode, and identifies potential failure risks in advance through predictive calculation. The specific implementation process is as follows: S31. For each device node , considering factors such as its business location, interaction neighbors, and task participation degree in the coupling graph, construct its comprehensive health representation vector: ; Among them, represents the current device state vector (including but not limited to temperature, voltage, pressure); represents the neighbor nodes (including devices and services) coupled with the current device node e i ; represents the dependence strength learned by the attention mechanism, reflecting the influence of node j on the health status of e i ; W 1 represents the weight matrix for linear transformation; represents the non-linear activation function (ReLU) to enhance the modeling ability; represents the current device e i 's comprehensive health representation vector; Accordingly, three factors, namely cross-domain influence (device and business collaboration), structural position (the role of the device in the task link), and coupling dynamics (the coupling weight changes at any time), are embedded in the representation to achieve health status perception driven by business; S32. Use device e i 's comprehensive health representation vector , input it into a lightweight neural evaluation function, and quantitatively score the current health status of device e i : ; Here, represents the vector transpose operation; Among them, represents the health score, that is, the health score of device e i at time t. The closer it is to 0, the more abnormal it indicates, ; w 2 and b respectively represent the weight and bias term obtained by model training; Set the health risk threshold θ. When < θ, it is considered that there is a health risk and enter the prediction process; S33. Select the health representation sequence of the past time steps and input it into the graph time series prediction model (i.e., Gated Graph Recurrent Unit, GGRU, graph gated recurrent unit model): ; Further predict the future failure risk score: ; Among them, represents the predicted health score of device e i at the future τ time step, that is, the health score at the future time t + τ; represents the predicted health representation of device e i at the future time t + τ; represents the graph gated recurrent unit model, which takes into account both structural perception and time series memory capabilities; S34. Use the nodes with a significant decrease in the prediction score (i.e., < θ low ) to trace back the influence path. Specifically: S3401. For each device e i , according to the health score obtained from the future prediction, set the health risk threshold θ low . If it satisfies: < θ low , then determine that device e i is a potential high-risk failure node; S3402. Extract the influence weight of each edge (ei, ej) in the device - service coupling model G(t) on the node health status at time t , and construct the cumulative influence weight matrix of each node within the time window [t−NumT + 1, t]: ; Among them, represents the cumulative health impact degree of device e j on device e i within the past NumT time instants; S3403. For each high - risk device node e i , perform the following backtracking: Starting from e i , retrieve its set of direct neighbor nodes retroactively (i.e., the nodes that affected it in the past); Arrange them in descending order according to the cumulative influence weight , and select the nodes e j whose cumulative influence weight is greater than the set threshold γ; Recursively backtrack the upstream influence nodes of e j , continuously construct the path until: the cumulative weight drops below the threshold; the length of the backtracking path reaches the maximum depth d max ; or reach the source service nodes (including but not limited to the oil tank area, fuel dispenser management server nodes); Finally, output the risk chain path Chain(e i ).

[0023] S4. The abnormal semantic recognition and interpretation unit, based on the output results of device health joint reasoning and fault trend prediction (i.e., predicting the health score of device e i at the future τ time steps and its risk chain path Chain(e i ))), further mines and identifies the potential abnormal semantics of the device, and deduces the causal chain of the abnormality, and finally generates an operation - interpretable abnormal description to provide an intuitive basis for operation and maintenance decisions. The specific implementation process is as follows: S41. Extract the propagation feature vector of device e i : (1) Extract the basic attributes of the link itself: Chain length : The shortest propagation steps from e i to the end node; Maximum propagation width : The layer with the largest number of child nodes at any level; Change in node aggregation degree : The variation of the node clustering coefficient along the chain; (2) During the propagation process, extract the risk transfer weight w(e i ,e j ) (i.e., representing the influence degree of the abnormal signal from one device to another device), and define the link cumulative propagation weight i of device e : , and extract the average single-hop propagation intensity : ; Among them, (e u ,e v ) represents that there is a potential risk transfer relationship (i.e., one hop) between device e i and device e u to device e v in the risk chain Chain(e u ,e v ) represents the propagation intensity when describing the risk transfer from device e u to device e v ; (3) Define the propagation eigenvector of device e i as: ; S42. Construct the abnormal eigenvector (abnormal multi-dimensional representation): For each device e i predicted to have potential high risk, comprehensively consider the current state, prediction trend, business metric changes, and link propagation characteristics, and construct a set of multi-dimensional abnormal eigenvectors : ; Among them, represents the health score of device e i at time t; represents the predicted health score of device e i at the future τ time steps; represents the health change amplitude, ; represents the current set of key business metrics (including but not limited to traffic, temperature, oil pressure); represents the propagation eigenvector of device e i ; S43. Introduce the abnormal semantic generation network (Semantic Generation Network, SGN), and the goal is to map through the eigenvector to the specific abnormal semantic category set C; Define the classification probability distribution : ; It should be noted that the Abnormal Semantic Generation Network (SGN) is a multi-modal, interpretable, and self-supervised abnormal semantic classification network for intelligent monitoring of business processes. Its main function is to convert multi-modal data such as equipment health scores, risk chain features, and business metrics into abnormal categories with business semantics, such as oil and gas leakage trends, signal anomalies between devices, etc.; the input feature vectors of SGN include the current health score, predicted score, trend difference, business key metrics, and risk chain propagation features. It extracts various feature semantics through a multi-modal embedding module (i.e., through a multi-layer perceptron MLP), and generates the probability distribution of abnormal semantic categories through the Softmax function; S44. According to the identified abnormal categories, combined with the device-business coupling graph G(t) and the risk chain Chain(e i ), deduce the logical path of the formation of the anomaly; S45. Output the abnormal semantic recognition results in the form of a graph + text summary, including but not limited to: Highlight the key nodes in the risk chain; mark the changes in abnormal indicators for each node; give a concise business explanation; For example: "The fuel dispenser #5 detects a continuous decrease in oil pressure, resulting from fluctuations in the tank sensor + aging of the pipeline valve leading to abnormal flow control, and there is a predicted hidden danger of oil and gas leakage."

[0024] S5. The dynamic operation optimization module introduces a reinforcement learning + game regulation mechanism + cost-risk quantification function to jointly generate optimization actions, and dynamically converts the identified equipment health trends, risk chain propagation paths, and abnormal semantic explanations and other information into executable operation optimization strategies (including but not limited to scheduling strategies, maintenance arrangements, load reconstruction, and oil product allocation), realizing a closed-loop of perception-identification-reasoning-optimization, and improving the overall safety and economic efficiency of the gas station. The specific implementation process is as follows: S51. Construct the current gas station operation state vector S t : ; Among them, represents the vector composed of the health scores of all devices e i at time t; represents the business operation index vector (including but not limited to the remaining oil volume, the number of customers queuing); represents the current operation cost vector (including but not limited to maintenance, energy, and manpower); represents the safety risk chain score vector (in this embodiment, the average value of the Chain propagation depth is adopted); represents the time context information (including but not limited to peak / low valley, weather, scheduling cycle); S52. Define the action set A = {a 1 ,a2 , …, a t , ..., a K}, each action a t represents an operation strategy, and an optimization objective function is constructed as follows: ; where, represents the expected optimization benefit of strategy π, comprehensively considering the minimization of operation cost and the maximization of safety benefit; π represents the strategy function, which takes the current state S t as input and outputs the optimal action a t ; T represents the total number of steps in the decision-making period, that is, the time step length considered in this round of optimization plan; represents the discount factor, which controls the present value of future benefits, 0 < γ < 1; represents the operation cost function, which is the operation overhead corresponding to state S t when action a t is executed; represents the safety and service benefit function, which is the reduced safety risk or increased customer satisfaction when action a t is executed; represents the safety-cost regulation weight factor, which adjusts the trade-off degree between safety and operation efficiency in the optimization process and can be dynamically adjusted; S53. To ensure multi-objective constraints (safety + cost + efficiency), a two-layer reinforcement game optimization strategy generation mechanism is constructed, including an inner-layer device game and an outer-layer business game two-layer structure. Based on reinforcement learning, the game tension in the strategy evolution process is modeled to generate multi-objective optimization strategies; S54. A semantic interpretation regulation unit is introduced. According to the abnormal semantic interpretation results output by S4, the focus of the optimization strategy is adjusted, and the action set generated by reinforcement learning is corrected through soft constraints; Exemplarily, if the abnormality stems from "oil quality problem", deployment or cleaning is preferably recommended; if there is a "secondary chain abnormality" of a device on the risk chain, the risk propagation node is preferably closed; S55. Select and output the optimal strategy .

[0025] S6. The intelligent alarm and resource scheduling module receives the optimal strategy, the abnormal semantic recognition result, the risk chain path Chain(e i ), the health score at the future time t + τ and the collected perception data stream, coordinates the resource scheduling device, automatically executes some decisions (including but not limited to switching devices and adjusting the oil supply path), and at the same time guides manual intervention for necessary repair or maintenance operations, continuously monitors the execution effect and dynamically adjusts the strategy to form a closed-loop optimization.

[0026] Embodiment 3. A method for monitoring gas station equipment based on the Internet of Things proposed by the present invention further includes a double-layer enhanced game optimization strategy generation mechanism, and its specific implementation steps are as follows: A1. Participating entities and state-action modeling: Inner-layer game entity (equipment agent): Each key equipment e i is modeled as a reinforcement learning agent Agent i with the following characteristics: State input: equipment health status , risk propagation degree ρ i (t), business dependency B i (t); Action output: including but not limited to repair / unload / share load / tolerate operation; Objective function: maximize the recovery of health trend + minimize the local risk chain propagation; Outer-layer game entity (business service agent): The gas station business process (including but not limited to refueling, dispensing, customer queuing) is modeled as a service agent Agent b , and its control: task / load allocation strategy among multiple devices; manpower scheduling, oil product resource allocation; Objective function: maximize the overall customer satisfaction / business throughput + minimize the probability of business interruption.

[0027] A2. Inner-layer game mechanism (equipment game): Health-chain risk control tension modeling: Model the problem as a repeated game reinforcement learning, and each Agent i learns a strategy and participates in the following local revenue game: ; Among them, represents the cumulative revenue of equipment agent e i during the entire game cycle; represents the discount factor, 0 < γ < 1, and the behavior closer to the current moment is more important; represents equipment e i after executing action a t , the change in its health score; represents the impact of the current equipment failure on the risk chain propagation of other equipment, and the measurement is the chain decline term of the health of adjacent equipment; represents the current action cost, including but not limited to maintenance cost, downtime loss, resource occupancy; α, β, θ represent dynamic weight coefficients, that is, the importance of health recovery, the importance of risk propagation suppression, and the influence degree of control behavior cost respectively; Accordingly, a game strategy is introduced into the risk chain items, and the strategies among devices generate a non-independent coupled game structure. If a device delays processing, its adjacent nodes (in the risk propagation graph) may bear increased burdens, forming a local non-cooperative game: If the Agent i ignores maintenance, the Agent j (dependent device) will experience a decline in health; the system uses the graph attention mechanism to estimate ChainLoss i (t), reflecting the degree of adjacent influence.

[0028] A3. Outer game mechanism (business game): Global efficiency - risk control coordination strategy: The business layer Agent b observes the health status and task requirements of all devices and generates a scheduling strategy based on the global state , and its revenue function: ; Among them, represents the cumulative operating revenue of the business agent during the entire game cycle; represents the current business service quality indicators, including but not limited to the average customer waiting time, fueling task completion rate, service delay indicators; represents the current overall device risk level of the system, aggregating the impact values of all device risk chains; represents the operating cost caused by the current scheduling strategy, including but not limited to personnel scheduling expenses, oil product resource allocation costs, service switching losses; , , represent the business strategy preference weights, namely the QoS priority, the importance of safety risk control, and the cost sensitivity degree; Output strategy guiding factors: α, β, θ (injected into the revenue functions of all device Agents i as regulation parameters); Accordingly, the device layer strategy is embedded as a dynamic variable into the business layer game: The Agent b does not directly control the device, but affects the device strategy direction through dynamic weights (α, β, θ) to achieve soft game guidance.

[0029] A4. Double-layer game collaborative evolution mechanism: At each moment t, the game process is as follows: The business agent Agent b observes the global state S t , and adjusts the weights α, β, θ; Each Agent i performs device actions under the guidance of the weights ; All device actions affect the system state S t+1 , and feedback to the business agent; Use a multi-agent reinforcement learning framework (MADDPG is adopted in this embodiment) to iteratively update the policy; Accordingly: construct a business policy to regulate the device policy → the device response generates system feedback → the system feedback optimizes the business policy, forming a human-device-business multi-agent collaborative game feedback closed-loop.

[0030] A5. Game convergence and policy selection mechanism: Define the composite revenue function of the overall system: ; Among them, represents the composite revenue index after the overall system policy is executed; represents the weight control parameter of the business layer revenue in the total target; N represents the number of devices participating in the inner-layer game; Accordingly: by setting the Pareto frontier screening mechanism, it is judged whether there exists: a set of device action combinations , a set of scheduling policies , such that any unilateral adjustment will cause the overall revenue to decrease, which is regarded as the output of the optimal collaborative policy.

[0031] A6. The final output is an optimization strategy with multiple objectives: the device layer action policy table, the business layer operation policy, and the policy interpretive structure.

[0032] The embodiments of the present invention have been described in detail above with reference to the accompanying drawings. However, the present invention is not limited thereto. Various changes can be made without departing from the spirit of the present invention within the knowledge scope of those skilled in the art.

Claims

1. A gas station equipment monitoring method based on the Internet of Things, characterized in that: The specific implementation steps include the following: S1. Multi-dimensional operating parameters and business process node dynamics of key equipment in gas stations; S2. By constructing a set of equipment and business process nodes and associating state vectors to describe the operation characteristics, the attention mechanism is used to dynamically adjust the coupling weights, and the graph convolutional network is combined to extract high-order interaction features to establish a device-business coupling model; S3. The comprehensive health representation vector is constructed by integrating the device status, neighbor node influence and coupling weight through the attention mechanism. The lightweight evaluation function is used to quantify the device health score. The graph time series model is combined to predict future failure risks, and the cumulative impact weight is traced back to identify the risk chain path. S4. Construct a multi-dimensional anomaly feature vector by extracting device propagation features, use semantic generation network to map anomaly semantic categories, combine the device-business coupling model with the risk chain to deduce the anomaly causal path, and finally generate anomaly semantic recognition results; S5. By constructing an operation state vector that includes equipment health, business indicators, and safety risks, defining a multi-objective optimization function and a two-layer reinforcement game optimization strategy generation mechanism, and dynamically adjusting the strategy priority in combination with abnormal semantic interpretation, the optimal operation decision is generated; S6. Integrate the optimal operational decision and abnormal analysis results, jointly execute automated resource scheduling, combine manual intervention and continuous monitoring and feedback to dynamically adjust the strategy.

2. A gas station equipment monitoring method based on the Internet of Things according to claim 1, characterized in that: The construction process of the device-business coupling model is as follows: S21. Define device collection , and define a collection of business process nodes , build a set of all nodes ; Among them, e i represents the i-th key equipment node in the gas station; b j represents the jth business node of the business process; m represents the total number of key equipment in the gas station; n represents the total number of business process steps; S22. Define each device node e i and business node b j The associated state vector , : ； ； in, Represents the kth state feature of the i-th device node; represents the lth process feature of the jth service node; M represents the total number of state features in the i-th device node; L represents the total number of process features of the j-th service node; S23. Define the device-service initial coupling association matrix A: ； Among them, A uv Represents the initial connection weight between nodes u and v; S24, introduce the dynamic coupling weight tensor W(t) to perform adaptive weighted correction on the initial correlation matrix: ； ； Where W(t) represents the dynamic coupling weight tensor at time t; , A set of state vectors representing devices and services respectively; represents element-wise product; Represents a dynamically weighted device-service coupling graph; represents the weight learning function; S25. Apply graph convolutional network GCN to perform joint representation learning on devices and business nodes and extract high-order interaction features between nodes: ； ； in, , They represent the node representation of the l+1th layer and the node representation of the lth layer respectively; Represents the trainable weights of the lth layer; represents the ReLU activation function; D represents the degree matrix; represents the dynamic weighted connection weight between node i and node j at time t; represents the total connection weight of node i; S26, output device-business coupling model G(t): ; Where N represents the node set; Represents the connection relationship between nodes after dynamic weighting; Represents the joint embedding representation of each node.

3. A gas station equipment monitoring method based on the Internet of Things according to claim 2, characterized in that: The risk chain path identification process is as follows: S31. For each device node , construct its comprehensive health representation vector: ； in, Represents the current device state vector; Indicates the current device node e i Coupled neighbor nodes; represents the dependency strength learned by the attention mechanism; W1 represents the weight matrix; Represents the nonlinear activation function ReLU; Indicates device e i A comprehensive health representation vector; S32. Use equipment i Comprehensive health representation vector , for device e i Quantify the current health status: ; in, Indicates the health score, i.e., the device e i At time t, the closer the health score is to 0, the more abnormal it is. ; w2 and b represent the weight and bias terms obtained from model training respectively; Represents a vector transpose operation; Set the health risk threshold value θ, when When <θ, it is considered that there is a health risk and the prediction process begins; S33, Select the past sequence of health representations at each moment , input to the graph time series prediction model GGRU: ; Further prediction of future failure risk score: ; in, Represents the predicted device e i The health score at the future τ time step, that is, the health score at the future time t+τ; Represents the predicted device e i Health representation at future time t+τ; Representation graph gated recurrent unit model, taking into account both structure perception and time series memory capabilities; S34. Identify high-risk equipment through health score thresholds, extract cumulative impact weights, trace back to key upstream nodes until termination conditions, and output the risk chain path.

4. A gas station equipment monitoring method based on the Internet of Things according to claim 3, characterized in that: The implementation process of extracting the cumulative impact weight and tracing back to the key upstream nodes until the termination condition and outputting the risk chain path is as follows: A1. For each device e i , based on the health score predicted in the future , set the health risk threshold θ low , if it satisfies: <θ low , then determine the device e i A node with a high risk of potential failure; A2. Extract the impact weight of each edge (ei, ej) on the node health status at time t in the device-business coupling model G(t) , construct the cumulative influence weight matrix of each node in the time window [t−NumT+1,t]: ； in, Indicates device e j Device e in the history of NumT moments i the cumulative health impacts of A3. For each high-risk device node e i , perform the following backtrace: From e i Start from, and retrieve its direct neighbor node set in reverse (i.e. nodes that influenced it in the past); According to the cumulative impact weight Sort in descending order and select nodes e whose cumulative influence weight is greater than the set threshold γ j ; Recursive backtracking j The upstream influencing node continues to build the path, and the termination conditions are: the cumulative weight drops below the threshold; the length of the traceback path reaches the maximum depth d max ; Reach the source business node; A4. Final output risk chain path Chain(e i ).

5. A gas station equipment monitoring method based on the Internet of Things according to claim 4, characterized in that: The process of generating abnormal semantic recognition results is as follows: S41, extraction equipment i The propagation eigenvector of ; S42. For each device predicted to have a potential high risk i , combining the current status, forecast trend, business indicator changes and link propagation characteristics, to construct a set of multi-dimensional anomaly feature vectors : ; in, Indicates device e i Health score at time t; Represents the predicted device e i The health score at the next τ time steps; Indicates the extent of health changes, ; Represents the current set of key business indicators; Indicates device e i The propagation eigenvector of S43, introduce the abnormal semantic generation network SGN, the goal is to pass the feature vector Mapped to a specific abnormal semantic category set C; Defining Classification Probability Distribution : ; S44. According to the identified abnormal category, combine the device-business coupling graph G(t) and the risk chain Chain(e i ), deduce the logical path of abnormal formation; S45. Output abnormal semantic recognition results.

6. A gas station equipment monitoring method based on the Internet of Things according to claim 5, characterized in that: Equipment i The propagation eigenvector of The extraction process is: B1. Extract the basic properties of the link itself: Chain length :From e i The shortest number of propagation steps to the terminal node; Maximum spread width : In any level, the level with the largest number of child nodes; Node aggregation changes : The change in node clustering coefficient along the chain; B2. During the propagation process, extract the risk transfer weight w(e) of each edge, i.e., the connection between devices. i ,e j ), define device e i The cumulative transmission weight of the link : , extract the average single-hop propagation strength : ; Among them, (e u ,e v ) indicates that in the risk chain Chain(e i ), device e u To device v There is a potential risk transfer relationship between them; w(e u ,e v ) indicates the description of the device e u To device v the intensity of transmission when conveying risk; B3. Define device e i The propagation eigenvector of is: .

7. The method for monitoring gas station equipment based on the Internet of Things according to claim 5, characterized in that: The anomaly semantic generation network (SGN) converts multimodal data into anomaly categories with business semantics. The input feature vector includes the current health score, predicted score, trend difference, business key indicators and risk chain propagation characteristics. The multi-layer perceptron (MLP) is used to extract various feature semantics, and the probability distribution of anomaly semantic categories is generated through the Softmax function.

8. The method for monitoring gas station equipment based on the Internet of Things according to claim 5, characterized in that: The optimal operational decision-making process is as follows: S51, construct the current gas station operation state vector S t : ; in, Indicates all devices i A vector of health scores at time t; Represents a business operation indicator vector; represents the current operating cost vector; represents the security risk chain score vector; Represents temporal context information; S52, define the action set A = {a1, a2, ..., a t ,...,a K }, each action a t Represents an operation strategy and constructs an optimization objective function: ; in, represents the expected optimization benefit of strategy π; π represents the strategy function, and the input is the current state S t , output the optimal action a t ; T represents the total number of steps in the decision cycle, that is, the length of the time step considered in this round of optimization plan; represents the discount factor, 0<γ<1; Represents the operating cost function, executing action a t In state S t corresponding operating expenses; represents the security and service benefit function; represents the safety-cost control weight factor; S53. Construct a double-layer reinforcement game optimization strategy generation mechanism, with a double-layer structure of inner device game + outer business game, and generate multi-objective optimization strategies based on the game tension in the strategy evolution process of reinforcement learning modeling; S54, introduce semantic interpretation control unit, abnormal semantic interpretation results, adjust the optimization strategy focus, and correct the action set generated by reinforcement learning through soft constraints; S55. Select the optimal strategy .

9. A gas station equipment monitoring method based on the Internet of Things according to claim 8, characterized in that: The implementation process of the two-layer reinforcement game optimization strategy generation mechanism is as follows: C1. Participating Subjects and State-Action Modeling: Inner game subject, namely device agent: each key device e i Modeled as a reinforcement learning agent i ; The outer game subject is the business service agent: the gas station business process is modeled as a service agent b ; C2. Inner game mechanism, i.e., equipment game: Health-chain risk control tension modeling: Modeling the problem as repeated game reinforcement learning, each Agent i Learning Strategies , participate in the following local benefit game: ; in, Represents the device agent e i The accumulated income during the entire gaming cycle; represents the discount factor, 0<γ<1; Indicates device e i Execute action a t After that, the change in health score; Indicates the risk chain propagation impact of the current equipment failure on other equipment; represents the current action cost; α, β, θ represent dynamic weight coefficients; C3. Outer game mechanism, namely business game: global efficiency-risk control collaborative strategy: business layer agent b Observe the health status and task requirements of all devices and generate scheduling strategies based on the global status , its profit function is: , output strategy guidance factors: α, β, θ, and inject strategy guidance factors: α, β, θ as control parameters into all device agents i In the profit function of in, Represents the cumulative operating income of the business agent during the entire game cycle; Indicates the current business service quality indicator; Indicates the overall equipment risk level of the current system; Represents the operating cost caused by the current scheduling strategy; , , represents the business strategy preference weight; C4. At each moment t, the game process is as follows: Business Intelligence Agent b Observe the global state S t , adjust the weights α, β, θ; each Agent i Execute equipment actions under the guidance of weights ; All device actions affect the system status S t+1 , feedback to the business agent; use the multi-agent reinforcement learning framework MADDPG to iteratively update the strategy; C5. Define the compound profit function of the overall system: ; in, It represents the compound profit index after the overall strategy of the system is executed; It represents the weight control parameter of the business layer revenue in the overall goal; N represents the number of devices participating in the inner layer game; By setting up the Pareto frontier screening mechanism, a multi-objective optimization strategy combination is selected. , }; in, Indicates a combination of device actions; Indicates the scheduling strategy; C6. The final output is a multi-objective optimization strategy: device layer action strategy table, business layer operation strategy, and strategy explanatory structure.

10. A gas station equipment monitoring system based on the Internet of Things, which is used to execute a gas station equipment monitoring method based on the Internet of Things according to any one of claims 1 to 9, characterized in that: include: The IoT collaborative perception module is used to collect real-time operating data of various equipment in the gas station and monitor the dynamics of business process nodes; The equipment-business bidirectional modeling engine is used to model the association between equipment status and business process status as a bidirectional coupling graph structure, using the graph structure to show the physical and logical connections between equipment and define the interaction rules between equipment load and business flow; The abnormal semantic analysis module is used to build a joint reasoning and prediction mechanism for equipment health under dynamic business loads and an abnormal analysis mechanism based on the equipment-business semantic structure. The abnormal semantic analysis module includes: an equipment health joint reasoning unit and an abnormal semantic recognition and interpretation unit; The equipment health joint reasoning unit uses a joint reasoning algorithm to dynamically infer the health status of the equipment, taking into account the current status of the equipment and future changes in business load, and accurately predicts the probability of potential failures, thereby achieving early warning and taking measures in advance; The abnormal semantic recognition and interpretation unit conducts in-depth analysis of predicted abnormalities and provides maintenance personnel with accurate fault diagnosis information by understanding the causes, impact scope, and related business processes behind equipment abnormalities. Dynamic operation optimization module, which is used to comprehensively optimize fault warning, business load and equipment status. Based on the objective function of minimizing fault risk, operating cost and customer waiting time, it generates the optimal resource scheduling plan through optimization algorithm; The intelligent alarm and resource scheduling module is used for real-time alarms, automatic execution of optimization plans, scheduling of equipment resources and arrangement of necessary maintenance work.

Images