A Multi-Source Sensing and Linkage Intelligent Early Warning Method for Oilfield Perimeter Security

Abstract

This invention discloses a multi-source sensing and linkage-based intelligent early warning method for oilfield perimeter security, belonging to the field of oilfield perimeter security technology. To address the problems of unstable multi-source data correlation, inaccurate risk classification, insufficient matching of linkage responses, and inconvenient tracing of early warning events, this invention acquires multi-source sensing data and determines risk level constraints, constructs a standardized multi-source feature dataset, trains a lightweight risk assessment neural network to output real-time risk levels, and generates a set of linkage instructions based on the risk level-linkage action mapping relationship. This method is applicable to intelligent early warning systems for perimeter security in oilfield stations, oil pipelines, and field production areas.

Description

Technical Field

[0001] This invention belongs to the field of oilfield perimeter security technology, specifically involving multi-source sensing data fusion, lightweight risk assessment neural network, hierarchical linkage early warning and early warning event data tracing technology. Background Technology

[0002] Oilfield stations, pipelines, and other perimeter areas are typically wide-ranging, highly open, and often situated in harsh outdoor environments characterized by wind, sand, rain, snow, and significant diurnal variations in light intensity. As oilfield security evolves from manual inspections and single-point alarms to video surveillance, radar detection, vibration sensing, and IoT-based systems, perimeter security systems are shifting from single-sensor devices to multi-source sensing, automatic identification, and coordinated response to improve intrusion detection efficiency and on-site response capabilities.

[0003] Existing perimeter intrusion warning solutions employ a combination of multiple sensing devices, including radar, cameras, and vibration sensors, to jointly detect intrusion risks. These devices assess the presence of intrusion based on target distance, video target area, vibration signals, or environmental conditions. Other solutions combine multi-sensor data with alarm rules, switching between different detection strategies in adverse weather or complex environments to reduce false alarms and missed alarms caused by interference from a single sensor. While these solutions represent improvements over traditional single-camera or single-vibration alarm methods, most still rely primarily on threshold judgments, rule triggers, or simple sensor result aggregation.

[0004] In the context of oilfield perimeter scenarios, the aforementioned solutions still have limitations in adaptability. Firstly, radar sensing data, visual sensing data, and vibration sensing data differ in dimensions, acquisition frequency, and interference methods. Without unified risk level constraints and standardized processing, it's difficult to establish stable correspondences between multi-source data. Secondly, existing solutions typically only output alarm or no alarm results, making it difficult to establish tiered judgments based on general, significant, and emergency risks, and also difficult to further match coordinated actions such as lighting, camera turning, drone takeoff, and audible / visual alarms at different risk levels. Furthermore, obtaining real intrusion samples is difficult, and directly relying on field data to train models can easily lead to insufficient sample coverage and unstable generalization capabilities.

[0005] Therefore, it is necessary to propose an intelligent early warning method for oilfield perimeter scenarios. This method integrates radar perception data, visual perception data, and vibration perception data into a unified processing flow. It constructs a standardized multi-source feature dataset with risk level labels through risk level constraints, and uses a lightweight risk assessment neural network to output real-time risk levels. Then, it generates a set of linkage instructions based on the risk level-linkage action mapping relationship, thereby realizing the hierarchical assessment, linkage response, and early warning event data traceability of oilfield perimeter intrusion risk. Summary of the Invention

[0006] To address the problems of unstable multi-source data correlation, inaccurate risk classification, insufficient matching of linkage responses, and inconvenience in tracing early warning events in existing oilfield perimeter security systems, this invention proposes the following solution: A multi-source sensing and linkage intelligent early warning method for oilfield perimeter security, the method comprising: S1. Acquire multi-source sensing data of the oilfield perimeter and determine the risk level constraint relationship. The multi-source sensing data includes radar sensing data, visual sensing data and vibration sensing data. S2. Based on the risk level constraint relationship, sample construction and standardization processing are performed on the multi-source perception data to obtain a standardized multi-source feature dataset with risk level labels; S3. Construct a lightweight risk assessment neural network, and perform supervised training on the lightweight risk assessment neural network based on the standardized multi-source feature dataset to obtain the trained model weights. S4. Collect real-time multi-source sensing data of the oilfield perimeter according to the preset real-time monitoring refresh interval, and after standardizing the real-time multi-source sensing data, input it into a lightweight risk assessment neural network loaded with the trained model weights to obtain the real-time risk level. S5. Generate a set of linkage instructions based on the real-time risk level and the preset risk level-linkage action mapping relationship, and trigger the corresponding linkage device to execute the set of linkage instructions; S6. Record the early warning event data corresponding to the real-time risk level, the set of linkage instructions, and the execution status of the linkage equipment.

[0007] Furthermore, the multi-source sensing data in S1 includes radar distance, radar speed, area of ​​the moving object from the camera, and vibration intensity, wherein the radar distance and radar speed are obtained from the radar sensing data, the area of ​​the moving object from the camera is obtained from the visual sensing data, and the vibration intensity is obtained from the vibration sensing data.

[0008] Furthermore, the risk level constraint relationship described in S1 includes multi-source sensing data value constraints corresponding to general risk, important risk, and emergency risk, respectively. Among them, general risk is level 0, corresponding to radar distance of 50-100m, radar speed of 0-2m / s, moving object area of ​​camera of 0-50px², and vibration intensity of 0-0.2g; important risk is level 1, corresponding to radar distance of 20-50m, radar speed of 1-5m / s, moving object area of ​​camera of 50-200px², and vibration intensity of 0.2-0.6g; emergency risk is level 2, corresponding to radar distance of 0-20m, radar speed of 3-15m / s, moving object area of ​​camera of 200-1000px², and vibration intensity of 0.6-1.0g.

[0009] Further, the sample construction and standardization processing of the multi-source sensing data described in S2 includes: generating multi-source sensing samples with risk level labels based on the risk level constraint relationship, and normalizing the multi-source sensing samples; when accessing real device data, performing timestamp alignment, missing value filling, outlier filtering, and normalization processing on the real device data, so that the real device data is processed into a standardized feature vector with the same format as the multi-source sensing samples; the real device data includes at least one of CSV format historical data and TCP / IP protocol real-time device data.

[0010] Furthermore, the lightweight risk assessment neural network described in S3 includes an input layer, two hidden layers, and an output layer. The input layer has a dimension of 4, corresponding to radar distance, radar speed, area of ​​the moving object from the camera, and vibration intensity, respectively. Both hidden layers are fully connected layers, with 16 neurons in each hidden layer, and are combined with a ReLU activation function and a Dropout layer. The inactivation probability of the Dropout layer is 0.2. The output layer has a dimension of 3, corresponding to the classification probabilities of general risk, important risk, and emergency risk, respectively.

[0011] Furthermore, the supervised training described in S3 includes: dividing the standardized multi-source feature dataset into a training set and a validation set in an 8:2 ratio; performing forward propagation calculation on the lightweight risk assessment neural network based on the training set; calculating the cross-entropy loss function value based on the real labels and prediction results; updating the network weights through the backpropagation algorithm and the Adam adaptive optimizer; and verifying the risk classification accuracy through the validation set after each round of training.

[0012] Furthermore, obtaining the real-time risk level as described in S4 includes: switching the lightweight risk assessment neural network loaded with the trained model weights to assessment mode, disabling the Dropout layer and gradient calculation, and outputting the real-time risk level using gradient-free forward propagation.

[0013] Furthermore, the risk level-linkage action mapping relationship described in S5 includes: the set of linkage instructions corresponding to general risks includes event log recording instructions and warning area lighting activation instructions; the set of linkage instructions corresponding to important risks includes event log recording instructions, surveillance camera turning to target area instructions, and on-site security personnel notification instructions; the set of linkage instructions corresponding to emergency risks includes event log recording instructions, surveillance camera target real-time tracking instructions, inspection drone emergency take-off instructions, on-site sound and light alarm trigger instructions, and synchronous notification instructions from higher-level management departments.

[0014] Furthermore, the early warning event data mentioned in S6 includes event time, risk level, sensor data, set of linkage instructions, and execution status of linkage equipment; after recording the early warning event data, it also includes displaying and querying the early warning event data in a visual interface, and exporting the early warning event data to a local file.

[0015] Based on the same inventive concept, this invention also proposes a computer storage medium storing a computer program, which executes the above-mentioned multi-source sensing linkage intelligent early warning method for oilfield perimeter security when running on a processor.

[0016] Compared with the prior art, the present invention has the following beneficial effects: By acquiring radar, visual, and vibration sensing data of the oilfield perimeter and determining risk level constraints, multi-source sensing data from different sources can be uniformly correlated around risk levels, thereby improving the problem of unstable correlation of multi-source data.

[0017] By constructing and standardizing samples from multi-source sensing data based on risk level constraints, a standardized multi-source feature dataset with risk level labels is obtained. This allows multi-source sensing data to participate in model training under a unified feature format, thereby improving the consistency and usability of risk assessment input data. Compared with existing technologies, this approach can reduce the impact of insufficient sample coverage caused by relying solely on real-world data.

[0018] By constructing a lightweight risk assessment neural network and conducting supervised training based on a standardized multi-source feature dataset, the risk features corresponding to radar perception data, visual perception data, and vibration perception data can be fused and evaluated, thereby improving the stability of oilfield perimeter risk level judgment.

[0019] By collecting real-time multi-source sensing data at preset real-time monitoring refresh intervals and inputting it into a lightweight risk assessment neural network loaded with trained model weights, the real-time risk level is obtained. This allows the model to output the classification result of the risk of intrusion into the oilfield perimeter in real time, thereby improving the problem of inaccurate risk classification in traditional alarm methods.

[0020] By generating a set of linkage instructions based on real-time risk levels and a preset risk level-linkage action mapping relationship, and triggering the corresponding linkage devices to execute the set of linkage instructions, this approach enables different risk levels to be matched with corresponding linkage actions, thereby improving the problem of insufficient linkage response matching. Compared with existing technologies, this method can avoid mismatches in handling caused by using the same response actions for different risk events.

[0021] By recording early warning event data corresponding to real-time risk levels, sets of linkage instructions, and execution status of linkage equipment, the risk assessment results, linkage instructions, and execution status during the early warning process can be retained, thereby facilitating subsequent data traceability of early warning events at the oilfield perimeter.

[0022] This invention features unified association of multi-source sensing data, real-time risk level assessment, hierarchical matching of linkage actions, and data tracing of early warning events. It can improve the stability, hierarchical accuracy, and linkage response matching of perimeter security early warning in oilfields, and is applicable to intelligent perimeter security early warning in oilfield stations, oil pipelines, and field production areas. Attached Figure Description

[0023] Figure 1 This is a flowchart of the intelligent early warning method for oilfield perimeter security based on multi-source sensing linkage, as described in the implementation method. Detailed Implementation

[0024] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only a part of the embodiments of the present invention, and not all of them. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0025] Implementation Method 1 like Figure 1 As shown, a multi-source sensing linkage intelligent early warning method for oilfield perimeter security includes the following steps: S1, acquiring multi-source sensing data of the oilfield perimeter and determining risk level constraints, wherein the multi-source sensing data includes radar sensing data, visual sensing data, and vibration sensing data; S2, based on the risk level constraints, constructing and standardizing samples from the multi-source sensing data to obtain a standardized multi-source feature dataset labeled with risk levels; S3, constructing a lightweight risk assessment neural network and supervising its training based on the standardized multi-source feature dataset to obtain trained model weights; S4, collecting real-time multi-source sensing data of the oilfield perimeter at a preset real-time monitoring refresh interval, standardizing the real-time multi-source sensing data, and inputting it into the lightweight risk assessment neural network loaded with the trained model weights to obtain a real-time risk level; S5, generating a linkage instruction set based on the real-time risk level and a preset risk level-linkage action mapping relationship, and triggering corresponding linkage devices to execute the linkage instruction set; S6, recording early warning event data corresponding to the real-time risk level, the linkage instruction set, and the execution status of the linkage devices.

[0026] By acquiring multi-source sensing data and determining risk level constraints, radar sensing data, visual sensing data, and vibration sensing data can be used for sample construction, model training, and coordinated decision-making under a unified risk level system, thereby achieving graded assessment and coordinated response effects for oilfield perimeter security early warning.

[0027] Furthermore, the multi-source sensing data in S1 includes radar distance, radar speed, area of ​​the moving object from the camera, and vibration intensity, wherein the radar distance and radar speed are obtained from the radar sensing data, the area of ​​the moving object from the camera is obtained from the visual sensing data, and the vibration intensity is obtained from the vibration sensing data.

[0028] By limiting multi-source sensing data to radar distance, radar speed, area of ​​moving objects from cameras, and vibration intensity, radar sensing data, visual sensing data, and vibration sensing data can be formed into a feature form that can be input into a model, achieving the technical effect of unified expression of multi-source sensing data.

[0029] Furthermore, the risk level constraint relationship described in S1 includes multi-source sensing data value constraints corresponding to general risk, important risk, and emergency risk, respectively. Among them, general risk is level 0, corresponding to radar distance of 50-100m, radar speed of 0-2m / s, moving object area of ​​camera of 0-50px², and vibration intensity of 0-0.2g; important risk is level 1, corresponding to radar distance of 20-50m, radar speed of 1-5m / s, moving object area of ​​camera of 50-200px², and vibration intensity of 0.2-0.6g; emergency risk is level 2, corresponding to radar distance of 0-20m, radar speed of 3-15m / s, moving object area of ​​camera of 200-1000px², and vibration intensity of 0.6-1.0g.

[0030] By setting value constraints for multi-source sensing data corresponding to general risks, important risks, and emergency risks, a correspondence can be established between different risk levels and multi-source sensing data, thus achieving the basic support effect for risk level label construction and risk classification judgment.

[0031] Further, the sample construction and standardization processing of the multi-source sensing data described in S2 includes: generating multi-source sensing samples with risk level labels based on the risk level constraint relationship, and normalizing the multi-source sensing samples; when accessing real device data, performing timestamp alignment, missing value filling, outlier filtering, and normalization processing on the real device data, so that the real device data is processed into a standardized feature vector with the same format as the multi-source sensing samples; the real device data includes at least one of CSV format historical data and TCP / IP protocol real-time device data.

[0032] By generating multi-source sensing samples with risk level labels based on risk level constraints, and standardizing the multi-source sensing samples and real device data, data from different sources can be processed into standardized feature vectors with a unified format, achieving the technical effect of consistent model training input.

[0033] Furthermore, the lightweight risk assessment neural network described in S3 includes an input layer, two hidden layers, and an output layer. The input layer has a dimension of 4, corresponding to radar distance, radar speed, area of ​​the moving object from the camera, and vibration intensity, respectively. Both hidden layers are fully connected layers, with 16 neurons in each hidden layer, and are combined with a ReLU activation function and a Dropout layer. The inactivation probability of the Dropout layer is 0.2. The output layer has a dimension of 3, corresponding to the classification probabilities of general risk, important risk, and emergency risk, respectively.

[0034] By setting up an input layer, two hidden layers, and an output layer, and making the input layer correspond to radar distance, radar speed, the area of ​​the moving object from the camera, and vibration intensity, the lightweight risk assessment neural network can perform fusion calculations on standardized multi-source feature data to achieve the technical effect of outputting the classification probabilities of general risk, important risk, and emergency risk.

[0035] Furthermore, the supervised training described in S3 includes: dividing the standardized multi-source feature dataset into a training set and a validation set in an 8:2 ratio; performing forward propagation calculation on the lightweight risk assessment neural network based on the training set; calculating the cross-entropy loss function value based on the real labels and prediction results; updating the network weights through the backpropagation algorithm and the Adam adaptive optimizer; and verifying the risk classification accuracy through the validation set after each round of training.

[0036] By dividing a standardized multi-source feature dataset into training and validation sets, and updating network weights based on cross-entropy loss function values, backpropagation algorithm, and Adam adaptive optimizer, a lightweight risk assessment neural network can optimize parameters based on data with risk level labels, achieving the technical effect of risk classification ability training.

[0037] Furthermore, obtaining the real-time risk level as described in S4 includes: switching the lightweight risk assessment neural network loaded with the trained model weights to assessment mode, disabling the Dropout layer and gradient calculation, and outputting the real-time risk level using gradient-free forward propagation.

[0038] By switching the lightweight risk assessment neural network to assessment mode, disabling the Dropout layer and gradient calculation, and using gradient-free forward propagation to output real-time risk levels, the risk level output can be completed during the inference stage using real-time multi-source sensing data, achieving the technical effect of real-time risk classification and early warning.

[0039] Furthermore, the risk level-linkage action mapping relationship described in S5 includes: the set of linkage instructions corresponding to general risks includes event log recording instructions and warning area lighting activation instructions; the set of linkage instructions corresponding to important risks includes event log recording instructions, surveillance camera turning to target area instructions, and on-site security personnel notification instructions; the set of linkage instructions corresponding to emergency risks includes event log recording instructions, surveillance camera target real-time tracking instructions, inspection drone emergency take-off instructions, on-site sound and light alarm trigger instructions, and synchronous notification instructions from higher-level management departments.

[0040] By setting up sets of linkage instructions corresponding to general risks, important risks, and emergency risks, the real-time risk level can be matched with the corresponding linkage action, achieving a graded linkage response effect for oilfield perimeter security early warning.

[0041] Furthermore, the early warning event data mentioned in S6 includes event time, risk level, sensor data, set of linkage instructions, and execution status of linkage equipment; after recording the early warning event data, it also includes displaying and querying the early warning event data in a visual interface, and exporting the early warning event data to a local file.

[0042] By recording event time, risk level, sensor data, set of linkage instructions, and execution status of linkage equipment, and displaying, querying, and exporting the early warning event data, the risk assessment results and linkage execution status during the early warning process can be saved, achieving the technical effect of early warning event data traceability.

[0043] The method described in this embodiment can be executed by a processor calling a computer program, which can be stored in a computer storage medium. When the computer program is executed by the processor, the above-mentioned intelligent early warning method for oilfield perimeter security based on multi-source sensing linkage can be realized.

[0044] Implementation Method 2 This embodiment of the multi-source sensing and linkage intelligent early warning method for oilfield perimeter security is implemented in complex environmental security early warning scenarios at oilfield field stations and oil pipeline perimeters. The method uses multi-source data processing, lightweight risk assessment, hierarchical linkage decision-making, real-time inference monitoring, and visualized log management as its main processing lines. It establishes a unified data processing, model training, real-time inference, and linkage response process around radar sensing data, visual sensing data, and vibration sensing data of the oilfield perimeter. This allows intrusion risks at the oilfield perimeter to be output in the form of risk levels and further converted into corresponding linkage command sets. The processing flow of this embodiment can form a closed loop based on the multi-source sensing data layer, risk assessment model layer, linkage decision engine layer, real-time inference and monitoring layer, and visualized interaction and log management layer. Data is transmitted between each layer around multi-source sensing data, risk levels, linkage command sets, and early warning event data.

[0045] In this embodiment, the multi-source sensing data includes radar distance, radar velocity, the area of ​​the moving object detected by the camera, and vibration intensity. Radar distance and radar velocity are obtained from radar sensing data, the area of ​​the moving object detected by the camera is obtained from visual sensing data, and vibration intensity is obtained from vibration sensing data. To enable multi-source sensing data from different sources to participate in subsequent training and inference under a unified risk level system, risk level constraints are first determined. These constraints include multi-source sensing data value constraints corresponding to general risk, significant risk, and emergency risk, respectively. The general risk level is 0, corresponding to minor environmental interference at a long distance, with constraints on radar distance of 50-100m, radar speed of 0-2m / s, moving object area of ​​camera of 0-50px², and vibration intensity of 0-0.2g; the significant risk level is 1, corresponding to suspicious personnel approaching the perimeter, with constraints on radar distance of 20-50m, radar speed of 1-5m / s, moving object area of ​​camera of 50-200px², and vibration intensity of 0.2-0.6g; the emergency risk level is 2, corresponding to vehicle intrusion or facility damage, with constraints on radar distance of 0-20m, radar speed of 3-15m / s, moving object area of ​​camera of 200-1000px², and vibration intensity of 0.6-1.0g.

[0046] In the process of constructing the multi-source dataset, multi-source sensing samples with risk level labels are generated based on risk level constraints. These samples are then normalized to obtain standardized multi-source feature data that can be directly used for model training. The normalization process eliminates dimensional differences between features from different sensors, and its expression is:

[0047] in, These are the original feature values ​​of the sensor. , These represent the upper and lower limits of the corresponding sensor characteristics. These are the normalized standard eigenvalues.

[0048] Through this process, radar distance, radar speed, area of ​​moving objects from the camera, and vibration intensity are converted into standardized features at a uniform scale, and together with the corresponding risk level labels, form a standardized multi-source feature dataset.

[0049] When accessing real device data, the real device data includes at least one of historical data in CSV format and real-time device data via TCP / IP protocol. After loading the real device data, the timestamps of the multi-source sensor data are first aligned to ensure that radar sensing data, visual sensing data, and vibration sensing data can form a correspondence under the same time reference. Then, missing values ​​are filled, outliers are filtered, and normalization is performed to process the real device data into a standardized feature vector that is consistent with the format of the multi-source sensing samples. The standardized feature vector includes at least radar distance, radar speed, area of ​​the moving object from the camera, and vibration intensity, which are used together with the multi-source sensing samples generated based on risk level constraints to serve the training and real-time inference of the lightweight risk assessment neural network.

[0050] In the model construction process, a lightweight risk assessment neural network is constructed, which adopts a three-level structure of input layer, two hidden layers, and output layer. The input layer has a dimension of 4, corresponding to the standardized radar distance, radar speed, area of ​​the moving object from the camera, and vibration intensity, respectively. Both hidden layers are fully connected layers, with 16 neurons in each hidden layer, and are combined with the ReLU activation function and a Dropout layer, with the Dropout layer having an inactivation probability of 0.2. The output layer has a dimension of 3, corresponding to the classification probabilities of general risk, important risk, and emergency risk, respectively. This lightweight risk assessment neural network is used to fuse and calculate standardized multi-source feature data and output the classification probabilities corresponding to the three risk levels.

[0051] During the forward propagation of the lightweight risk assessment neural network, the output of the first hidden layer can be expressed as:

[0052] The output of the second hidden layer can be represented as:

[0053] in, The input is a standardized multi-source feature vector. , The hidden layer weight matrix is... , For hidden layer bias terms, It is a linear rectified activation function. This is a random deactivation regularization operation. , This represents the fused feature output from the two hidden layers. The classification probability of the output layer can be calculated as follows:

[0054] in, This is the output layer weight matrix. For output layer bias terms, For normalized exponential functions, Let be the classification probability vectors corresponding to the three risk levels. Risk level prediction can be expressed as:

[0055] in, The final risk level output by the model is determined by the index corresponding to the maximum classification probability.

[0056] During supervised training, the standardized multi-source feature dataset is divided into training and validation sets in an 8:2 ratio, and a batch data loader is constructed to enable batch reading of training data. The training set data is input into a lightweight risk assessment neural network for forward propagation calculation to obtain the risk level classification probability. The cross-entropy loss function is calculated based on the true labels and prediction results, and the network parameters are updated using a backpropagation algorithm. The cross-entropy loss function can be expressed as:

[0057] in, This represents the number of samples in the training batch. The true label of sample i (one-hot encoded). This represents the classification probability of sample i corresponding to level j in the model output.

[0058] During the network weight update process, the Adam adaptive optimizer is used to iteratively update the model parameters. The weight update formula can be expressed as:

[0059] in, Let be the model weights for the t-th iteration. For learning rate, , These are the bias estimates for the first and second moments of the gradient, respectively. is the numerical stability constant.

[0060] After each round of training, the accuracy of risk classification is verified using a validation set, and the weights of the trained model are saved.

[0061] During real-time inference, the trained model weights are loaded, and real-time multi-source sensing data of the oilfield perimeter is collected at a preset real-time monitoring refresh interval. After standardization, the real-time multi-source sensing data is input into a lightweight risk assessment neural network loaded with the trained model weights. During the inference phase, the lightweight risk assessment neural network is switched to assessment mode, the Dropout layer and gradient calculation are disabled, and gradient-free forward propagation is used to output the real-time risk level. The single-sample inference process includes real-time acquisition of multi-source sensor features, standardization processing, model forward propagation, risk level output, and triggering of linkage commands. When the preset real-time monitoring refresh interval is 500ms, the system refreshes the real-time multi-source sensing data, real-time risk level, and linkage response status at the corresponding interval.

[0062] During the tiered and coordinated decision-making process, a set of coordinated instructions is generated based on the real-time risk level and the preset risk level-coordinated action mapping relationship, and the corresponding coordinated devices are triggered to execute the set of coordinated instructions. The risk level-coordinated action mapping relationship includes the set of coordinated instructions corresponding to general risks, important risks, and emergency risks, respectively. The set of coordinated instructions corresponding to general risks includes an event log recording instruction and a warning area lighting activation instruction; the set of coordinated instructions corresponding to important risks includes an event log recording instruction, a surveillance camera turning to the target area instruction, and an on-site security personnel notification instruction; the set of coordinated instructions corresponding to emergency risks includes an event log recording instruction, a surveillance camera real-time target tracking instruction, an inspection drone emergency take-off instruction, an on-site audible and visual alarm trigger instruction, and a synchronous notification instruction from the superior management department.

[0063] The collaborative decision-making process can be represented as:

[0064] in, The real-time risk level output by the risk assessment model. This is a risk level-linked action mapping function. This is the set of linked instructions for output.

[0065] Once the real-time risk level is determined, the hierarchical multi-device linkage decision engine outputs the corresponding linkage instruction set according to the risk level-linkage action mapping function, and converts the linkage instruction set into control instructions for the corresponding linkage devices to trigger the corresponding devices in the warning area lighting equipment, surveillance cameras, inspection drones, on-site sound and light alarm equipment and notification equipment to perform actions.

[0066] During the recording of early warning event data, data corresponding to the real-time risk level, the set of linkage instructions, and the execution status of the linkage equipment are recorded. This early warning event data includes the event time, risk level, sensor data, the set of linkage instructions, and the execution status of the linkage equipment. After recording the early warning event data, it is displayed and queried in a visualization interface, and can be exported to a local file. The visualization interface also displays real-time sensor data, risk level, and linkage instructions, and supports data source switching, model training, monitoring start / stop, and log export operations.

[0067] In this embodiment, the method can be executed in the following order: system initialization and parameter configuration, multi-source dataset construction and preprocessing, construction and initialization of a lightweight risk assessment neural network, supervised training and optimization of the risk assessment model, real-time perimeter security monitoring and risk inference, hierarchical linkage decision-making and emergency response, and log management and data traceability. System initialization and parameter configuration includes configuring the hardware environment and deep learning framework of the oilfield perimeter security system, determining the computing equipment for model training and inference, and setting risk level classification rules, upper and lower limits for multi-source sensor feature values, model training hyperparameters, and real-time monitoring refresh intervals. Multi-source dataset construction and preprocessing includes selecting training data sources, generating multi-source sensing samples based on risk level constraints, or loading historical sensor data and corresponding risk event labels, and completing timestamp alignment, missing value imputation, outlier filtering, and normalization. Construction and initialization of the lightweight risk assessment neural network includes setting the core parameters of the input layer, double hidden layer, and output layer, initializing network weights and bias terms, and setting the loss function and optimizer core parameters. Supervised training and optimization of the risk assessment model includes inputting training data, forward propagation, loss calculation, backpropagation, updating network weights, validation set verification, and saving model weights. Real-time perimeter security monitoring and risk inference includes loading the trained model weights, initiating real-time monitoring, reading real-time multi-source sensing data, standardization processing, real-time inference, and outputting real-time risk levels. Tiered linkage decision-making and emergency response includes matching linkage command sets according to real-time risk levels, triggering control commands for corresponding linkage devices, and displaying the execution status of linkage commands. Log management and data traceability includes recording system operation logs and early warning event data, and displaying, querying, and exporting early warning event data.

[0068] The above detailed description of the technical solution provided by the present invention is intended to highlight the advantages and benefits of the technical solution provided by the present invention. However, the above detailed embodiments are not intended to limit the scope of protection of the present invention. Any reasonable modifications and improvements to the present invention, recombination of embodiments, and equivalent substitutions based on the spirit and principles of the present invention should be included within the scope of protection of the present invention.

[0069] Those skilled in the art will understand that the above description is merely a preferred embodiment of the present invention, and the features described in the various embodiments and / or claims disclosed in the present invention can be combined or combined in various ways, even if such combinations or combinations are not explicitly described in the disclosure of the present invention. Although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art can still modify the technical solutions described in the foregoing embodiments or make equivalent substitutions for some of the technical features. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principle scope of the present invention should be considered to fall within the protection scope of the present invention.

[0070] Although preferred embodiments of the invention have been described, those skilled in the art, upon learning the basic inventive concept, can make other changes and modifications to these embodiments. Therefore, the appended claims are intended to be interpreted as including both the preferred embodiments and all changes and modifications falling within the scope of the invention. Clearly, those skilled in the art can make various alterations and modifications to the invention without departing from its spirit and scope. Thus, if these modifications and modifications of the invention fall within the scope of the claims and their equivalents, the invention is also intended to include these modifications and modifications.

Claims

1. A multi-source perception linkage oilfield perimeter security intelligent early warning method, characterized in that, The method includes: S1. Acquire multi-source sensing data of the oilfield perimeter and determine the risk level constraint relationship. The multi-source sensing data includes radar sensing data, visual sensing data and vibration sensing data. S2. Based on the risk level constraint relationship, sample construction and standardization processing are performed on the multi-source perception data to obtain a standardized multi-source feature dataset with risk level labels; S3. Construct a lightweight risk assessment neural network, and perform supervised training on the lightweight risk assessment neural network based on the standardized multi-source feature dataset to obtain the trained model weights. S4. Collect real-time multi-source sensing data of the oilfield perimeter according to the preset real-time monitoring refresh interval. After standardizing the real-time multi-source sensing data, input it into a lightweight risk assessment neural network loaded with the trained model weights to obtain the real-time risk level. S5. Generate a set of linkage instructions based on the real-time risk level and the preset risk level-linkage action mapping relationship, and trigger the corresponding linkage device to execute the set of linkage instructions; S6. Record the early warning event data corresponding to the real-time risk level, the set of linkage instructions, and the execution status of the linkage equipment.

2. The method of claim 1, wherein, The multi-source sensing data in S1 includes radar distance, radar speed, area of ​​the moving object from the camera, and vibration intensity. The radar distance and radar speed are obtained from the radar sensing data, the area of ​​the moving object from the camera is obtained from the visual sensing data, and the vibration intensity is obtained from the vibration sensing data.

3. The method of claim 1, wherein, The risk level constraint relationship described in S1 includes multi-source sensing data value constraints corresponding to general risk, important risk, and emergency risk, respectively. Among them, general risk is level 0, corresponding to radar distance of 50-100m, radar speed of 0-2m / s, moving object area of ​​camera of 0-50px², and vibration intensity of 0-0.2g; important risk is level 1, corresponding to radar distance of 20-50m, radar speed of 1-5m / s, moving object area of ​​camera of 50-200px², and vibration intensity of 0.2-0.6g; emergency risk is level 2, corresponding to radar distance of 0-20m, radar speed of 3-15m / s, moving object area of ​​camera of 200-1000px², and vibration intensity of 0.6-1.0g.

4. The method of claim 1, wherein, S2 describes the sample construction and standardization processing of the multi-source sensing data, which includes: generating multi-source sensing samples with risk level labels based on the risk level constraint relationship, and normalizing the multi-source sensing samples; when accessing real device data, performing timestamp alignment, missing value filling, outlier filtering, and normalization processing on the real device data, so that the real device data is processed into a standardized feature vector with the same format as the multi-source sensing samples; the real device data includes at least one of CSV format historical data and TCP / IP protocol real-time device data.

5. The method of claim 1, wherein, The lightweight risk assessment neural network described in S3 includes an input layer, two hidden layers, and an output layer. The input layer has a dimension of 4, corresponding to radar distance, radar speed, area of ​​moving objects near the camera, and vibration intensity, respectively. Both hidden layers are fully connected layers with 16 neurons in each hidden layer, and are combined with the ReLU activation function and a Dropout layer. The inactivation probability of the Dropout layer is 0.

2. The output layer has a dimension of 3, corresponding to the classification probabilities of general risk, important risk, and emergency risk, respectively.

6. The method of claim 1, wherein, S3 The supervised training includes: dividing the standardized multi-source feature dataset into a training set and a validation set in an 8:2 ratio; performing forward propagation calculation on the lightweight risk assessment neural network based on the training set; calculating the cross-entropy loss function value based on the real labels and prediction results; updating the network weights through the backpropagation algorithm and the Adam adaptive optimizer; and verifying the risk classification accuracy through the validation set after each round of training.

7. The method of claim 1, wherein, S4 describes obtaining the real-time risk level by: switching the lightweight risk assessment neural network loaded with the trained model weights to assessment mode, disabling the Dropout layer and gradient calculation, and outputting the real-time risk level using gradient-free forward propagation.

8. The method of claim 1, wherein, The risk level-linkage action mapping relationship described in S5 includes: the set of linkage instructions corresponding to general risks includes event log recording instructions and warning area lighting activation instructions; the set of linkage instructions corresponding to important risks includes event log recording instructions, surveillance camera turning to target area instructions and on-site security personnel notification instructions; the set of linkage instructions corresponding to emergency risks includes event log recording instructions, surveillance camera target real-time tracking instructions, inspection drone emergency take-off instructions, on-site sound and light alarm trigger instructions, and synchronous notification instructions from higher-level management departments.

9. The method of claim 1, wherein, The warning event data mentioned in S6 includes event time, risk level, sensor data, set of linkage instructions, and execution status of linkage equipment; after recording the warning event data, it also includes displaying and querying the warning event data in a visual interface, and exporting the warning event data to a local file.

10. A computer storage medium having stored thereon a computer program, characterized in that The computer program, when running on a processor, performs the method according to any one of claims 1 to 9.

Images