Method and System for Monitoring and Early Warning of Oil Tank Data Using Multi-Source Sensors

Abstract

This invention belongs to the field of oil tank monitoring technology and discloses a method and system for oil tank data monitoring and anomaly early warning using multi-source sensors. The method includes the following steps: In the main control device, a multi-source heterogeneous data fusion and monitoring analysis network and a hybrid reinforcement learning anomaly early warning architecture are constructed; multi-source sensors are used to collect real-time multi-source oil tank data from the warehouse's oil tanks, and the real-time multi-source oil tank data is uploaded to the warehouse's main control device; in the main control device, the multi-source heterogeneous data fusion and monitoring analysis network is used to perform oil tank data anomaly detection on the real-time multi-source oil tank data, obtaining real-time anomaly detection results; based on the real-time anomaly detection results, an anomaly early warning strategy is generated using the hybrid reinforcement learning anomaly early warning architecture, obtaining a real-time anomaly early warning strategy, and then sent to the early warning center. This invention solves the problems of data uniformity, lack of spatial correlation, and single early warning model in existing technologies.

Description

Technical Field

[0001] This invention belongs to the field of oil tank monitoring technology, specifically relating to a method and system for monitoring oil tank data and providing early warning of anomalies using multi-source sensors. Background Technology

[0002] With the rapid development of the petrochemical industry, the safe operation of oil tanks, as important facilities for storing petroleum products, is crucial. With the advancement of sensor technology, installing sensors in oil tanks to collect and monitor data has become an important measure to ensure tank safety.

[0003] Existing technologies have the following drawbacks: 1) Data uniformity: Traditional methods usually only use a single type or a few types of sensor data, such as temperature and pressure, which makes it difficult to fully reflect the complex state of the oil tank. This leads to one-sided monitoring and easily overlooks some important abnormal information.

[0004] 2) Lack of spatial correlation: Existing technologies often do not fully consider the spatial relationship between sensors, resulting in insufficient data fusion and an inability to effectively capture the distribution characteristics and spatial correlation of the internal state of the oil tank;

[0005] 3) Single early warning model: Existing technologies usually use a single early warning model, such as threshold method or statistical model. These models are difficult to adapt to the complexity and variability of oil tank conditions, and the early warning effect is limited. Summary of the Invention

[0006] To address the problems of data uniformity, lack of spatial correlation, and single early warning model in existing technologies, the present invention aims to provide a method and system for monitoring and anomaly early warning of oil tank data using multi-source sensors.

[0007] The technical solution adopted in this invention is as follows:

[0008] A method for monitoring and anomaly early warning of oil tank data using multi-source sensors includes the following steps:

[0009] In the main control device, based on the basic information from multiple sources of sensors, a multi-source heterogeneous data fusion and monitoring analysis network and a hybrid reinforcement learning anomaly early warning architecture are constructed.

[0010] Multi-source sensors are used to collect real-time multi-source tank data of the oil tanks in the warehouse and upload the real-time multi-source tank data to the warehouse's main control device.

[0011] In the main control unit, a multi-source heterogeneous data fusion and monitoring and analysis network is used to detect oil tank data anomalies in real-time multi-source oil tank data and obtain real-time anomaly detection results.

[0012] Based on the real-time anomaly detection results, an anomaly warning strategy is generated using a hybrid reinforcement learning anomaly warning architecture, and then sent to the warning center.

[0013] Furthermore, the multi-source heterogeneous data fusion and monitoring analysis network includes a multi-source sensor spatial topology layer, a spatial topology fusion layer, a temporal dynamic fusion layer, a physical constraint fusion layer, and an anomaly detection and identification layer connected in sequence.

[0014] The hybrid reinforcement learning anomaly early warning architecture includes a master control device-level anomaly early warning strategy generation layer, a coordination layer, and an early warning center-level anomaly early warning strategy generation layer, which are connected in sequence.

[0015] Furthermore, in the main control device, based on the basic information from multiple source sensors, a multi-source heterogeneous data fusion and monitoring analysis network and a hybrid reinforcement learning anomaly early warning architecture are constructed, including the following steps:

[0016] In the main control device, a multi-source heterogeneous data fusion and monitoring analysis network is constructed based on the basic information from the multi-source sensors;

[0017] Based on the output parameters of the multi-source heterogeneous data fusion and monitoring analysis network, the control parameters of several warehouse subsystems connected to the main control device, and the control parameters of the early warning center, a hybrid reinforcement learning anomaly early warning architecture is constructed.

[0018] Furthermore, in the main control device, based on the basic information from the multi-source sensors, a multi-source heterogeneous data fusion and monitoring analysis network is constructed, including the following steps:

[0019] Based on the 3D point cloud data of the oil tanks in the warehouse, SLAM technology is used to generate the corresponding 3D point cloud model of the oil tanks.

[0020] Based on the multi-source sensor location information in the basic information of the multi-source sensors, a spatial deployment map of the multi-source sensors is generated in the three-dimensional point cloud model of the oil tank.

[0021] Based on the multi-source sensor type in the basic information of the multi-source sensor, a heterogeneous sensor network is constructed based on the multi-source sensor spatial deployment map to obtain the multi-source sensor spatial topology layer;

[0022] Based on the data structure of historical multi-source oil tank data collected by multi-source sensors, a graph signal generation structure is constructed at the output end of the spatial topology layer of the multi-source sensors.

[0023] A GAT network is set up for the graph signal generation structure, and the attention weight of the GAT network is set according to the calculated importance weight of the multi-source sensors to obtain the spatial topology fusion layer.

[0024] A first and second temporal feature extraction channel are set in parallel at the output end of the spatial topology fusion layer, and a cascaded temporal feature fusion module is set.

[0025] A Causal CNN network is set up for the first temporal feature extraction channel, a BiGRU unit is set up for the second temporal feature extraction channel, and a gating mechanism is set up for the temporal feature fusion module to obtain a temporal dynamic fusion layer;

[0026] At the output end of the time-series dynamic fusion layer, a tank thermodynamics module, a data credibility assessment module, and a physical constraint fusion module are set up in sequence, and a tank thermodynamic state model is set for the tank thermodynamics module to obtain the physical constraint fusion layer.

[0027] An online anomaly detection module is set at the output of the physical constraint fusion layer. An encoder is set in the online anomaly detection module using a contrastive learning framework to obtain the anomaly detection and recognition layer.

[0028] By integrating the spatial topology layer of multi-source sensors, the spatial topology fusion layer, the temporal dynamic fusion layer, the physical constraint fusion layer, and the anomaly detection and identification layer, a multi-source heterogeneous data fusion and monitoring analysis network is obtained.

[0029] Furthermore, based on the output parameters of the multi-source heterogeneous data fusion and monitoring analysis network, the control parameters of several warehouse subsystems connected to the main control device, and the control parameters of the early warning center, a hybrid reinforcement learning anomaly early warning architecture is constructed, including the following steps:

[0030] A master control device-level anomaly early warning strategy generation layer is set at the output end of the multi-source heterogeneous data fusion and monitoring analysis network;

[0031] Based on the output parameters of the multi-source heterogeneous data fusion and monitoring analysis network and the several warehouse subsystems connected to the main control device, several corresponding main control device-level intelligent agents are set in the main control device-level anomaly early warning strategy generation layer.

[0032] A coordination layer is set at the output end of the main control device level anomaly warning strategy generation layer, and the coordination layer is used to connect the output end of the main control device level anomaly warning strategy generation layer to the input end of the warning center level anomaly warning strategy generation layer.

[0033] Based on the control parameters of the early warning center, set up corresponding early warning center-level intelligent agents in the early warning center-level anomaly early warning strategy generation layer;

[0034] By integrating the main control device-level anomaly warning strategy generation layer, the coordination layer, and the early warning center-level anomaly warning strategy generation layer, a hybrid reinforcement learning anomaly warning architecture is obtained.

[0035] Furthermore, in the main control unit, a multi-source heterogeneous data fusion and monitoring analysis network is used to detect tank data anomalies in real-time multi-source tank data, obtaining real-time anomaly detection results, including the following steps:

[0036] In the main control device, a heterogeneous sensor network of the multi-source sensor spatial topology layer of the multi-source heterogeneous data fusion and monitoring analysis network is used to generate a multi-source sensor spatial topology map.

[0037] Using the graph signal generation structure of the multi-source sensor spatial topology layer, real-time multi-source oil tank data is projected onto the multi-source sensor spatial topology graph to obtain real-time graph signals.

[0038] Based on the preset attention weights, the GAT network of the multi-source sensor spatial topology layer is used to extract sensor feature vectors that fuse spatial correlation information from the real-time graph signal.

[0039] Using the first and second temporal feature extraction channels of the temporal dynamic fusion layer, the first and second real-time temporal features of the sensor feature vector with fused spatial correlation information are extracted;

[0040] Based on dynamic gating weights, the temporal feature fusion module of the temporal dynamic fusion layer is used to fuse the first real-time temporal features and the second real-time temporal features to obtain a sensor data representation with fused temporal features.

[0041] The sensor data representation with fused temporal features is input into the tank thermodynamics module of the physical constraint fusion layer, and the corresponding real-time data residual is obtained using the tank thermodynamic state model.

[0042] Based on the real-time data residuals, the data credibility assessment module of the physical constraint fusion layer is used to assess the credibility of the sensor data representation of the fused temporal features and obtain a credibility assessment score.

[0043] Based on the credibility assessment score, the physical constraint fusion module of the physical constraint fusion layer is used to perform physical constraint fusion on the sensor data representation of the fused temporal features to obtain fused physical constraint credible sensor data.

[0044] The online anomaly detection module using the anomaly detection and recognition layer encodes trusted sensor data with fused physical constraints into real-time feature vectors, compares the real-time feature vectors with normal data distributions, and outputs real-time anomaly detection results based on the obtained anomaly detection scores.

[0045] Furthermore, the hybrid reinforcement learning anomaly early warning architecture is built based on the MA-DDPG-PPO algorithm, the master control device-level anomaly early warning strategy generation layer is built based on the MA-DDPG algorithm, and the early warning center-level anomaly early warning strategy generation layer is built based on the PPO algorithm.

[0046] Furthermore, based on the real-time anomaly detection results, an anomaly warning strategy is generated using a hybrid reinforcement learning anomaly warning architecture, resulting in a real-time anomaly warning strategy, which is then sent to the warning center. This process includes the following steps:

[0047] Collect real-time subsystem status data of several warehouse subsystems connected to the main control device, and generate corresponding real-time main control device level anomaly warning strategies using the main control device level anomaly warning strategy generation layer of the hybrid reinforcement learning anomaly warning architecture based on the real-time anomaly detection results and several real-time subsystem status data.

[0048] The coordination layer of the hybrid reinforcement learning anomaly warning architecture inputs the real-time master control device-level anomaly warning strategy into the warning center-level anomaly warning strategy generation layer.

[0049] Collect real-time control status data from the early warning center. Based on the real-time control status data and the real-time master control device-level anomaly early warning strategy, use the early warning center-level anomaly early warning strategy generation layer of the hybrid reinforcement learning anomaly early warning architecture to generate the corresponding real-time early warning center-level anomaly early warning strategy.

[0050] The real-time master control device-level anomaly warning strategy and the real-time warning center-level anomaly warning strategy are integrated to obtain a real-time anomaly warning strategy. The real-time master control device-level anomaly warning strategy is executed using the master control device.

[0051] The real-time anomaly warning strategy will be sent to the warning center, and the warning center will be used to execute the real-time anomaly warning strategy.

[0052] Furthermore, the real-time multi-source oil tank data includes real-time temperature data, real-time humidity data, real-time air pressure data, real-time gas concentration data, real-time liquid level data, and real-time vibration data of the oil tanks in the warehouse.

[0053] A multi-source sensor-based oil tank data monitoring and anomaly early warning system is provided to realize oil tank data monitoring and anomaly early warning methods. The system includes multi-source sensors, a main control device, and an early warning center. The multi-source sensors are installed at the oil tanks in the warehouse and are communicatively connected to the main control device. The main control device is installed in the control room of the warehouse and is communicatively connected to several warehouse subsystems. The early warning center is communicatively connected to the main control device.

[0054] The beneficial effects of this invention are as follows:

[0055] This invention discloses a method and system for monitoring and anomaly warning of oil tank data using multi-source sensors. By integrating multi-source heterogeneous sensor data, including temperature, humidity, air pressure, gas concentration, liquid level, and vibration, it can more comprehensively reflect the real-time status of the oil tank, overcoming the shortcomings of existing technologies in terms of data singularity. A multi-source sensor spatial topology layer is constructed, and a graph attention network is used for spatial topology fusion, effectively capturing the spatial correlation between sensors and making up for the lack of spatial correlation in existing technologies. A hybrid reinforcement learning anomaly warning architecture is adopted, including a master control device-level anomaly warning strategy generation layer and an early warning center-level anomaly warning strategy generation layer, realizing multi-level warning strategy generation. This overcomes the limitation of the single warning model in existing technologies. Through multi-level warning strategies, different warning measures are taken according to the severity of the anomaly, achieving refined management and control.

[0056] Other beneficial effects of the present invention will be further explained in the specific embodiments. Attached Figure Description

[0057] Figure 1 This is a flowchart of the multi-source sensor method for monitoring and anomaly warning of oil tank data in this invention.

[0058] Figure 2 This is a structural block diagram of the multi-source sensor oil tank data monitoring and anomaly early warning system of the present invention. Detailed Implementation

[0059] The present invention will be further explained below with reference to the accompanying drawings and specific embodiments.

[0060] Example 1:

[0061] like Figure 1 As shown, this embodiment provides a method for monitoring and anomaly warning of oil tank data using multi-source sensors, including the following steps:

[0062] S1: In the main control device, based on the basic information of multi-source sensors, a multi-source heterogeneous data fusion and monitoring analysis network and a hybrid reinforcement learning anomaly early warning architecture are constructed.

[0063] The multi-source heterogeneous data fusion and monitoring analysis network includes a multi-source sensor spatial topology layer, a spatial topology fusion layer, a temporal dynamic fusion layer, a physical constraint fusion layer, and an anomaly detection and identification layer connected in sequence.

[0064] The hybrid reinforcement learning anomaly warning architecture includes a master control device-level anomaly warning strategy generation layer, a coordination layer, and an early warning center-level anomaly warning strategy generation layer connected in sequence.

[0065] The hybrid reinforcement learning anomaly warning architecture is built on the multi-agent deep deterministic policy gradient (MA-DDPG) - proximal policy optimization (PPO) algorithm. The master control device level anomaly warning policy generation layer is built on the MA-DDPG algorithm, and the warning center level anomaly warning policy generation layer is built on the PPO algorithm.

[0066] In the main control device, based on the basic information from multiple sensors, a multi-source heterogeneous data fusion and monitoring analysis network and a hybrid reinforcement learning anomaly early warning architecture are constructed, including the following steps:

[0067] S1-1: In the main control device, based on the basic information from the multi-source sensors, a multi-source heterogeneous data fusion and monitoring analysis network is constructed, including the following steps:

[0068] S1-1-1: Based on the 3D point cloud data of the oil tanks in the warehouse, use Simultaneous Localization and Mapping (SLAM) technology to generate the corresponding 3D point cloud model of the oil tanks;

[0069] S1-1-2: Based on the multi-source sensor location information in the basic information of the multi-source sensors, generate a spatial deployment map of the multi-source sensors in the three-dimensional point cloud model of the oil tank;

[0070] S1-1-3: Based on the multi-source sensor type in the basic information of the multi-source sensor, and based on the multi-source sensor spatial deployment map, construct a heterogeneous sensor network to obtain the multi-source sensor spatial topology layer;

[0071] The formula is: ;

[0072] In the formula, A spatial topology diagram of multi-source sensors; A set of nodes (each node represents a sensor); For edge set (representing spatial connections or relationships between sensors); It is an adjacency matrix;

[0073] S1-1-4: Based on the data structure of historical multi-source oil tank data collected by multi-source sensors, construct a graph signal generation structure at the output end of the spatial topology layer of the multi-source sensors;

[0074] S1-1-5: Set up a Graph Attention Network (GAT) for the graph signal generation structure to capture the spatial correlation between sensors, and set the attention weights of the GAT network according to the calculated importance weights of the multi-source sensors to obtain the spatial topology fusion layer.

[0075] The formula is:

[0076] ;

[0077] ;

[0078] In the formula, Let i be the new feature representation of node i at the k-th head in the real-time graph signal; K is the total number of attention heads in the GAT network; k is the attention head indicator; i is the node indicator, where the node is the sensor data component in the real-time multi-source oil tank data. Let be the original feature representation of node i at the k-th head in the real-time graph signal; j is the neighbor node indicator; J is the total number of neighbor nodes; The importance weights of multi-source sensors for node i and its neighboring node j; This represents the node feature matrix after GAT processing; N is the total number of multi-source sensors. The feature dimension is the learnable linear transformation; through To obtain sensor feature vectors that integrate spatial correlation information ; For time step;

[0079] S1-1-6: Set up a first temporal feature extraction channel and a second temporal feature extraction channel in parallel at the output end of the spatial topology fusion layer, and set up a cascaded temporal feature fusion module;

[0080] S1-1-7: For slowly changing data such as pressure and liquid level, a causal convolutional neural network (Causal CNN) is set up for the first temporal feature extraction channel. For abrupt data such as gas concentration and vibration, a bidirectional gated recurrent unit (BiGRU) is set up for the second temporal feature extraction channel. A gating mechanism is set up for the temporal feature fusion module to dynamically adjust the output weights of the two channels according to the data change rate, thus obtaining the temporal dynamic fusion layer.

[0081] The formula for the Causal CNN network is: ;

[0082] In the formula, This is the first real-time time series feature; This is a causal convolution operation function; To fuse sensor feature vectors with spatial correlation information in the th... Sensor data components in the first The first feature vector at the time step; This refers to the sensor data component indication quantity; For time step indication; The kernel size;

[0083] The formula for a BiGRU cell is: ;

[0084] In the formula, This is the second real-time time series feature; For BiGRU operation functions; To fuse sensor feature vectors with spatial correlation information in the th... Sensor data components in the first The second feature vector of the time step;

[0085] The formula for the gating mechanism is: ;

[0086] In the formula, For gating weights; For learnable weights and biases; For the rate of change of data;

[0087] The formula for temporal feature fusion is: ;

[0088] In the formula, To fuse temporal features; through Acquire sensor data representations with fused temporal features ;

[0089] S1-1-8: Set up a tank thermodynamics module, a data credibility assessment module and a physical constraint fusion module in sequence at the output end of the time-series dynamic fusion layer, and set up a tank thermodynamic state model for the tank thermodynamics module to obtain the physical constraint fusion layer.

[0090] The thermodynamic state model of an oil tank includes the ideal gas law, the van der Waals equation, and the hydrostatic equation.

[0091] The ideal gas law is as follows: ;

[0092] In the formula, The pressure inside the liquid in the oil tank (unit: Pascal, Pa). The volume of the liquid in the oil tank (unit: cubic meters, m³). is the ideal gas constant (unit: J / (mol·K)); Temperature (unit: Kelvin, K); The amount of substance of a gas (unit: mole, mol).

[0093] The formula for the van der Waals equation is: ;

[0094] In the formula, It is the van der Waals constant, which depends on the type of gas;

[0095] The formula for the equation of hydrostatics is: ;

[0096] In the formula, The density of the liquid in the oil tank (unit: kg / m³). This is the acceleration due to gravity (unit: m / s²). The liquid level in the oil tank (unit: meter, m);

[0097] S1-1-9: Set up an online anomaly detection module at the output of the physical constraint fusion layer. Use the contrastive learning framework (Simple Contrastive Learning of Visual Representations, SimCLR) to set up an encoder in the online anomaly detection module to obtain the anomaly detection recognition layer.

[0098] S1-1-10: Integrating the multi-source sensor spatial topology layer, spatial topology fusion layer, temporal dynamic fusion layer, physical constraint fusion layer, and anomaly detection and identification layer, a multi-source heterogeneous data fusion and monitoring analysis network is obtained.

[0099] S1-2: Based on the output parameters of the multi-source heterogeneous data fusion and monitoring analysis network, the control parameters of several warehouse subsystems connected to the main control device, and the control parameters of the early warning center, a hybrid reinforcement learning anomaly early warning architecture is constructed, including the following steps:

[0100] S1-2-1: Set a master control device-level anomaly early warning strategy generation layer at the output end of the multi-source heterogeneous data fusion and monitoring analysis network;

[0101] S1-2-2: Based on the output parameters of the multi-source heterogeneous data fusion and monitoring analysis network and several warehouse subsystems connected to the main control device, such as the cooling subsystem, oil tank level control subsystem, valve control subsystem, etc., several corresponding main control device level intelligent agents are set in the main control device level abnormal early warning strategy generation layer.

[0102] The complex task of monitoring and warning oil tanks is decomposed into multiple sub-tasks, which are handled by different master control device-level intelligent agents, thereby improving decision-making efficiency and robustness. Each master control device-level intelligent agent focuses on the status monitoring and warning decision of its corresponding subsystem. There can be information sharing and collaboration mechanisms between intelligent agents. For example, when the temperature is abnormal, the pressure intelligent agent can be notified to pay attention to the pressure change.

[0103] Each master control device-level agent contains two key components: the actor network and the critic network. The input to the actor network is the real-time subsystem state data received by each agent from its corresponding subsystem, as well as partial state information from other agents. The actor network outputs a continuous action value, representing the early warning strategy at the master control center level. For example, for the temperature subsystem, the action could be "start the cooling subsystem" or "adjust the cooling subsystem power". The critic network is used to evaluate the long-term benefits of the policy output by the actor network, providing gradient update signals to the actor network and outputting a Q-value, representing the expected long-term benefits of taking actions in a given state. The goal of the critic network is to minimize the mean square error between the Q-value and the target Q-value. Through interaction with the environment, each agent's actor network and critic network learn the optimal warning policy. The experience replay mechanism stores the experiences (states, actions, rewards, next state) generated by the agent's interaction with the environment in the experience pool and randomly samples them for training to break data correlation and improve sample utilization efficiency. The soft update mechanism periodically updates the parameters of the actor network and critic network to the target network to stabilize the training process. The parallel training mechanism can train multiple agents in a parallel manner to accelerate the training process.

[0104] S1-2-3: Set up a coordination layer at the output end of the main control device level anomaly warning strategy generation layer, and use the coordination layer to connect the output end of the main control device level anomaly warning strategy generation layer to the input end of the warning center level anomaly warning strategy generation layer.

[0105] The coordination layer enables information transmission and coordination between the master control device level and the early warning center level, including defining communication protocols and data formats to ensure that the master control device level and the early warning center level can correctly send and receive information. Middleware technologies such as message queues and publish / subscribe patterns can be used to implement the communication mechanism.

[0106] S1-2-4: Based on the control parameters of the early warning center, set up the corresponding early warning center-level intelligent agent in the early warning center-level abnormal early warning strategy generation layer;

[0107] The early warning center-level agent comprises two key components: an actor network and a critic network. The inputs are real-time control state data and real-time master control device-level anomaly early warning policies, while the output is a real-time early warning center-level anomaly early warning policy. The early warning center-level anomaly early warning policy generation layer updates the policy by maximizing the estimated expected reward. The PPO algorithm is used to train the early warning center-level actor network, enabling it to output the optimal early warning center-level early warning policy given the master control device-level early warning policy. Transfer learning and other techniques can be used to transfer knowledge from the master control device-level policy to the early warning center-level policy.

[0108] S1-2-5: Integrate the main control device-level anomaly warning strategy generation layer, coordination layer and early warning center-level anomaly warning strategy generation layer to obtain a hybrid reinforcement learning anomaly warning architecture;

[0109] S2: Use multi-source sensors to collect real-time multi-source oil tank data of the warehouse's oil tanks and upload the real-time multi-source oil tank data to the warehouse's main control device;

[0110] Real-time multi-source oil tank data includes real-time temperature data, real-time humidity data, real-time air pressure data, real-time gas concentration data, real-time liquid level data, and real-time vibration data of the oil tanks in the warehouse.

[0111] S3: In the main control unit, a multi-source heterogeneous data fusion and monitoring analysis network is used to perform anomaly detection on real-time multi-source oil tank data to obtain real-time anomaly detection results, including the following steps:

[0112] S3-1: In the main control device, a heterogeneous sensor network of the multi-source sensor spatial topology layer of the multi-source heterogeneous data fusion and monitoring analysis network is used to generate a multi-source sensor spatial topology map.

[0113] S3-2: Using the graph signal generation structure of the multi-source sensor spatial topology layer, real-time multi-source oil tank data is projected onto the multi-source sensor spatial topology map to obtain real-time graph signals;

[0114] S3-3: Based on the preset attention weights, the GAT network of the multi-source sensor spatial topology layer is used to extract the sensor feature vectors that fuse spatial correlation information in the real-time graph signal.

[0115] S3-4: Using the first and second time-series feature extraction channels of the time-series dynamic fusion layer, extract the first and second real-time time-series features of the sensor feature vectors that fuse spatial correlation information;

[0116] S3-5: Based on the dynamic gating weights, the temporal feature fusion module of the temporal dynamic fusion layer is used to fuse the first real-time temporal feature and the second real-time temporal feature to obtain the sensor data representation of the fused temporal feature.

[0117] S3-6: Input the sensor data representation with fused temporal features into the tank thermodynamics module of the physical constraint fusion layer, and use the tank thermodynamic state model to obtain the corresponding real-time data residuals;

[0118] S3-7: Based on the real-time data residuals, the data credibility assessment module of the physical constraint fusion layer is used to assess the credibility of the sensor data representation of the fused temporal features and obtain a credibility assessment score.

[0119] S3-8: Based on the credibility assessment score, the physical constraint fusion module of the physical constraint fusion layer is used to perform physical constraint fusion on the sensor data representation of the fused temporal features to obtain fused physical constraint credible sensor data.

[0120] S3-9: The online anomaly detection module using the anomaly detection and recognition layer encodes the trusted sensor data fused with physical constraints into a real-time feature vector, compares the real-time feature vector with the normal data distribution, and outputs the real-time anomaly detection result based on the obtained anomaly detection score.

[0121] In this embodiment, the real-time anomaly detection results are as follows:

[0122] Tank A: Temperature abnormal (current temperature is 5°C higher than the upper limit of the normal range);

[0123] Tank B: No abnormalities found;

[0124] S4: Based on the real-time anomaly detection results, an anomaly warning strategy is generated using a hybrid reinforcement learning anomaly warning architecture. This real-time anomaly warning strategy is then sent to the warning center, including the following steps:

[0125] S4-1: Collect real-time subsystem status data of several warehouse subsystems connected to the main control device, and generate corresponding real-time main control device level anomaly warning strategies based on the real-time anomaly detection results and several real-time subsystem status data, using the main control device level anomaly warning strategy generation layer of the hybrid reinforcement learning anomaly warning architecture.

[0126] Real-time subsystem status data:

[0127] The current operating status (on / off), target cooling temperature, and current outlet water temperature of the cooling subsystem; the current level, target level, and status (running / stopping) of the inlet / outlet pump of the oil tank level control subsystem (for oil tank A); the current opening degree and status (normal / faulty) of each key valve of the valve control subsystem (for related pipelines of oil tank A); and other relevant data such as ambient humidity and air pressure.

[0128] Real-time master control device-level anomaly early warning strategy:

[0129] "Instruction: Start the cooling subsystem corresponding to oil tank A, and adjust the power to 80%", "Instruction: Notify the operator to check the status of relevant valves in oil tank A and confirm that there is no leakage", "Instruction: Adjust the liquid level control strategy of oil tank A and reduce the oil inlet rate".

[0130] S4-2: A coordination layer using a hybrid reinforcement learning anomaly warning architecture to input real-time master control device-level anomaly warning strategies into the warning center-level anomaly warning strategy generation layer;

[0131] S4-3: Collect real-time control status data from the early warning center. Based on the real-time control status data and the real-time master control device-level anomaly early warning strategy, use the early warning center-level anomaly early warning strategy generation layer of the hybrid reinforcement learning anomaly early warning architecture to generate the corresponding real-time early warning center-level anomaly early warning strategy.

[0132] Real-time control status data:

[0133] Current number and level of alarms received;

[0134] Operator status (whether they are available or handling other alarms);

[0135] Status of available remote intervention methods (e.g., whether remote forced valve shut-off authority is available);

[0136] Historical records of similar incidents and their effectiveness evaluations;

[0137] Real-time early warning center-level anomaly warning strategy:

[0138] "Alarm: Oil tank A temperature is abnormal. On-site cooling has been initiated. On-duty personnel are requested to verify the situation immediately and prepare for emergency response." "Instruction: Raise the alert level of the warehouse east area to level two."

[0139] S4-4: Integrate the real-time master control device-level anomaly warning strategy and the real-time warning center-level anomaly warning strategy to obtain a real-time anomaly warning strategy. Use the master control device to execute the real-time anomaly warning strategy at the master control device level.

[0140] S4-5: The real-time anomaly warning strategy will be sent to the warning center, and the warning center will be used to execute the real-time anomaly warning strategy.

[0141] Example 2:

[0142] like Figure 2 As shown, this embodiment provides a multi-source sensor oil tank data monitoring and anomaly early warning system for realizing oil tank data monitoring and anomaly early warning methods. The system includes multi-source sensors, a main control device, and an early warning center. The multi-source sensors are installed at the oil tanks in the warehouse and are communicatively connected to the main control device. The main control device is installed in the control room of the warehouse and is communicatively connected to several warehouse subsystems. The early warning center is communicatively connected to the main control device.

[0143] Multi-source sensors are used to collect real-time multi-source tank data of oil tanks in the warehouse and upload the real-time multi-source tank data to the warehouse's main control device; and to implement a real-time main control device-level anomaly early warning strategy.

[0144] The main control unit is used to perform anomaly detection on real-time multi-source oil tank data using a multi-source heterogeneous data fusion and monitoring analysis network, and obtain real-time anomaly detection results; based on the real-time anomaly detection results, anomaly warning strategy is generated using a hybrid reinforcement learning anomaly warning architecture, and the real-time anomaly warning strategy is sent to the warning center; the real-time main control unit-level anomaly warning strategy is executed.

[0145] The early warning center is used to implement real-time anomaly early warning strategies at the center level.

[0146] This invention discloses a method and system for monitoring and anomaly warning of oil tank data using multi-source sensors. By integrating multi-source heterogeneous sensor data, including temperature, humidity, air pressure, gas concentration, liquid level, and vibration, it can more comprehensively reflect the real-time status of the oil tank, overcoming the shortcomings of existing technologies in terms of data singularity. A multi-source sensor spatial topology layer is constructed, and a graph attention network is used for spatial topology fusion, effectively capturing the spatial correlation between sensors and making up for the lack of spatial correlation in existing technologies. A hybrid reinforcement learning anomaly warning architecture is adopted, including a master control device-level anomaly warning strategy generation layer and an early warning center-level anomaly warning strategy generation layer, realizing multi-level warning strategy generation. This overcomes the limitation of the single warning model in existing technologies. Through multi-level warning strategies, different warning measures are taken according to the severity of the anomaly, achieving refined management and control.

[0147] The specific embodiments described above further illustrate the purpose, technical solution, and beneficial effects of the present invention. It should be understood that the above description is only a specific embodiment of the present invention and is not intended to limit the scope of protection of the present invention. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the scope of protection of the present invention.

Claims

1. A method for monitoring and early warning of abnormality of oil tank data of multi-source sensors, characterized in that: Comprising the following steps: At the master control device, according to the basic information of the multi-source sensor, a multi-source heterogeneous data fusion and monitoring analysis network and a hybrid reinforcement learning abnormal early warning architecture are constructed: At the master control device, according to the basic information of the multi-source sensor, a multi-source heterogeneous data fusion and monitoring analysis network is constructed, comprising the following steps: According to the three-dimensional point cloud data of the oil tank of the warehouse, using SLAM technology, the corresponding oil tank three-dimensional point cloud model is generated; According to the multi-source sensor position information in the basic information of the multi-source sensor, a multi-source sensor spatial deployment graph is generated in the oil tank three-dimensional point cloud model; According to the multi-source sensor type in the basic information of the multi-source sensor, based on the multi-source sensor spatial deployment graph, a heterogeneous sensor network is constructed, and a multi-source sensor spatial topology layer is obtained; According to the data structure of the historical multi-source oil tank data collected by the multi-source sensor, a graph signal generation structure is constructed at the output end of the multi-source sensor spatial topology layer; A GAT network is set for the graph signal generation structure, and the attention weight of the GAT network is set according to the calculated multi-source sensor importance weight, and a spatial topology fusion layer is obtained; A first time sequence feature extraction channel and a second time sequence feature extraction channel are set in parallel at the output end of the spatial topology fusion layer, and a time sequence feature fusion module is set in series; A Causal CNN network is set for the first time sequence feature extraction channel, a BiGRU unit is set for the second time sequence feature extraction channel, and a gating mechanism is set for the time sequence feature fusion module, and a time sequence dynamic fusion layer is obtained; An oil tank thermodynamic module, a data credibility evaluation module and a physical constraint fusion module are sequentially connected at the output end of the time sequence dynamic fusion layer, and an oil tank thermodynamic state model is set for the oil tank thermodynamic module, and a physical constraint fusion layer is obtained; An online anomaly detection module is set at the output end of the physical constraint fusion layer, an encoder is set in the online anomaly detection module using a contrast learning framework, and an anomaly detection identification layer is obtained; The multi-source heterogeneous data fusion and monitoring analysis network is obtained by integrating the multi-source sensor spatial topology layer, the spatial topology fusion layer, the time sequence dynamic fusion layer, the physical constraint fusion layer and the anomaly detection identification layer; According to the output parameters of the constructed multi-source heterogeneous data fusion and monitoring analysis network, the control parameters of the master control device connected to the several warehouse subsystems and the control parameters of the early warning center, a hybrid reinforcement learning abnormal early warning architecture is constructed, comprising the following steps: An abnormal early warning strategy generation layer of the master control device is set at the output end of the multi-source heterogeneous data fusion and monitoring analysis network; According to the output parameters of the constructed multi-source heterogeneous data fusion and monitoring analysis network and the several warehouse subsystems connected to the master control device, a corresponding several master control device level agents are set in the master control device level abnormal early warning strategy generation layer; A coordination layer is set at the output end of the master control device level abnormal early warning strategy generation layer, and the output end of the master control device level abnormal early warning strategy generation layer is connected to the input end of the early warning center level abnormal early warning strategy generation layer using the coordination layer; According to the control parameters of the early warning center, a corresponding early warning center level agent is set in the early warning center level abnormal early warning strategy generation layer; The integrated master control device level abnormality early warning strategy generation layer, the coordination layer, and the early warning center level abnormality early warning strategy generation layer are integrated to obtain a hybrid reinforcement learning abnormality early warning architecture; Real-time multi-source tank data of the oil tank in the warehouse is collected by using a multi-source sensor, and the real-time multi-source tank data is uploaded to the master control device of the warehouse; In the master control device, the multi-source heterogeneous data fusion and monitoring analysis network is used to perform oil tank data anomaly detection on the real-time multi-source tank data, and real-time anomaly detection results are obtained; According to the real-time anomaly detection results, the hybrid reinforcement learning abnormality early warning architecture is used to generate an abnormality early warning strategy, and the real-time abnormality early warning strategy is obtained and sent to the early warning center.

2. The method according to claim 1, wherein the method comprises the following steps: The multi-source heterogeneous data fusion and monitoring analysis network comprises a multi-source sensor spatial topology layer, a spatial topology fusion layer, a time sequence dynamic fusion layer, a physical constraint fusion layer, and an anomaly detection and identification layer connected in sequence; The hybrid reinforcement learning abnormality early warning architecture comprises a master control device level abnormality early warning strategy generation layer, a coordination layer, and an early warning center level abnormality early warning strategy generation layer connected in sequence.

3. The method according to claim 2, wherein the method comprises the following steps: In the master control device, the multi-source heterogeneous data fusion and monitoring analysis network is used to perform oil tank data anomaly detection on the real-time multi-source tank data, and real-time anomaly detection results are obtained, including the following steps: In the master control device, the heterogeneous sensor network of the multi-source sensor spatial topology layer of the multi-source heterogeneous data fusion and monitoring analysis network is used to generate a multi-source sensor spatial topology graph; The real-time multi-source tank data is projected onto the multi-source sensor spatial topology graph by using the graph signal generation structure of the multi-source sensor spatial topology layer to obtain real-time graph signals; According to a preset attention weight, the GAT network of the multi-source sensor spatial topology layer is used to extract a sensor feature vector that fuses spatial correlation information in the real-time graph signals; The first and second time sequence feature extraction channels of the time sequence dynamic fusion layer are used to extract first and second real-time time sequence features of the sensor feature vector that fuses spatial correlation information; According to a dynamic gating weight, the time sequence feature fusion module of the time sequence dynamic fusion layer is used to perform time sequence feature fusion on the first and second real-time time sequence features to obtain a sensor data representation of the fused time sequence features; The sensor data representation of the fused time sequence features is input into the oil tank thermodynamic module of the physical constraint fusion layer, and an oil tank thermodynamic state model is used to obtain corresponding real-time data residuals; According to the real-time data residuals, the data credibility evaluation module of the physical constraint fusion layer is used to perform credibility evaluation on the sensor data representation of the fused time sequence features to obtain a credibility evaluation score; According to the credibility evaluation score, the physical constraint fusion module of the physical constraint fusion layer is used to perform physical constraint fusion on the sensor data representation of the fused time sequence features to obtain physically constrained credible sensor data; The online anomaly detection module of the anomaly detection and identification layer is used to encode the physically constrained credible sensor data into a real-time feature vector, compare the real-time feature vector with a normal data distribution, and output real-time anomaly detection results according to an obtained anomaly detection score.

4. The method according to claim 3, wherein the method comprises the following steps: The hybrid reinforcement learning anomaly early warning architecture is constructed based on the MA-DDPG-PPO algorithm, the main control device level anomaly early warning strategy generation layer is constructed based on the MA-DDPG algorithm, and the early warning center level anomaly early warning strategy generation layer is constructed based on the PPO algorithm.

5. The method for oil tank data monitoring and abnormal early warning of a multi-source sensor according to claim 4, characterized in that: According to the real-time anomaly detection result, the hybrid reinforcement learning anomaly early warning architecture is used to generate an anomaly early warning strategy, obtain a real-time anomaly early warning strategy, and send it to the early warning center, including the following steps: Collect real-time subsystem state data of a plurality of warehouse subsystems connected to the main control device, and use the main control device level anomaly early warning strategy generation layer of the hybrid reinforcement learning anomaly early warning architecture to generate a corresponding real-time main control device level anomaly early warning strategy according to the real-time anomaly detection result and the plurality of real-time subsystem state data; Use the coordination layer of the hybrid reinforcement learning anomaly early warning architecture to input the real-time main control device level anomaly early warning strategy into the early warning center level anomaly early warning strategy generation layer; Collect real-time control state data of the early warning center, and use the early warning center level anomaly early warning strategy generation layer of the hybrid reinforcement learning anomaly early warning architecture to generate a corresponding real-time early warning center level anomaly early warning strategy according to the real-time control state data and the real-time main control device level anomaly early warning strategy; Integrate the real-time main control device level anomaly early warning strategy and the real-time early warning center level anomaly early warning strategy to obtain a real-time anomaly early warning strategy, and use the main control device to execute the real-time main control device level anomaly early warning strategy of the real-time anomaly early warning strategy; The real-time early warning center level anomaly early warning strategy of the real-time anomaly early warning strategy will be sent to the early warning center, and the early warning center will be used to execute the real-time early warning center level anomaly early warning strategy.

6. The method for oil tank data monitoring and abnormal early warning of a multi-source sensor according to claim 5, characterized in that: The real-time multi-source oil tank data includes real-time temperature data, real-time humidity data, real-time air pressure data, real-time gas concentration data, real-time liquid level data, and real-time vibration data of the oil tank of the warehouse.

7. A multi-source sensor-based oil tank data monitoring and abnormality early warning system, used for implementing the oil tank data monitoring and abnormality early warning method according to any one of claims 1-6, characterized in that: The system includes a plurality of source sensors, a main control device, and an early warning center, the plurality of source sensors are arranged at the oil tank of the warehouse, and the plurality of source sensors are in communication connection with the main control device, the main control device is arranged in the control room of the warehouse, and the main control device is in communication connection with a plurality of warehouse subsystems of the warehouse, and the early warning center is in communication connection with the main control device.

Images