A gas station intelligent supervision method using an AI algorithm and artificial intelligence

Abstract

This invention discloses an intelligent monitoring method for gas stations that combines AI algorithms and artificial intelligence, belonging to the field of gas station monitoring technology. The method includes the following steps: acquiring target behavioral feature values ​​of the gas station to be monitored; identifying the operational scenario based on the target behavioral feature values ​​and activating judgment rules to perform comparative analysis on the target behavioral feature values ​​to drive automatic switching of monitoring mechanisms; acquiring deviation behavioral feature values; wherein, deviation behavioral feature values ​​are feature values ​​obtained by identifying gas pump cheating behavior according to several monitoring mechanisms and performing judgment operations; obtaining violation information based on the deviation behavioral feature values ​​and monitoring mechanisms, acquiring evidence images based on the violation information, and sending early warning reminders through the evidence images; this invention improves the accuracy of identifying gas pump cheating behavior, achieves timely early warning, and improves monitoring efficiency.

Description

Technical Field

[0001] This invention relates to the field of gas station monitoring technology, specifically to an intelligent monitoring method for gas stations that combines AI algorithms and artificial intelligence. Background Technology

[0002] As the core hub of oil product distribution, gas stations mainly handle business processes such as receiving and inspecting goods, storing oil products, and selling oil products. They often introduce Robotic Process Automation (RPA) technology to monitor risks at gas stations and proactively alert staff to the risks of tampering with fuel dispensers.

[0003] However, traditional intelligent monitoring technologies for gas stations still have the following shortcomings: Although existing technologies incorporate artificial intelligence or AI algorithms, such as using basic machine learning models for training and analysis, they fail to build adaptive models tailored to the dynamic characteristics of gas station operating scenarios. This results in insufficient model generalization ability, making the models unable to accurately adapt. In routine inspections, redundant computing power is easily consumed, while in anomaly tracing situations, it is difficult to capture concealed fuel dispenser fraud. For example, AI algorithms using support vector machines to determine fuel dispenser data anomalies, such as training models using historical data on pump code counts and liquid level changes, and setting fixed thresholds to identify deviations, are simple in structure and easy to deploy. However, they can only process linearly correlated data and cannot fully explore the temporal fluctuations and correlational coupling patterns of fuel dispenser fraud. For instance, for tank data and pump code data, it is difficult to eliminate data drift interference caused by oil temperature and tank pressure, leading to a significant reduction in the accuracy of extracted features, low model recognition accuracy, and a significantly increased false negative rate. Furthermore, even if a gas station risk is identified, it is impossible to quickly and accurately pinpoint the time of occurrence, reducing the efficiency of regulatory management. Summary of the Invention

[0004] (a) Technical problems to be solved To address the shortcomings of existing technologies, this invention provides an intelligent gas station monitoring method that combines AI algorithms and artificial intelligence. By accurately acquiring target behavioral feature values ​​of the gas station under monitoring, it achieves a quantitative characterization of the entire operational chain of tankers, nozzles, and vehicles, significantly improving the accuracy and effectiveness of monitoring data. Based on the target behavioral feature values, it automatically identifies different operational scenarios such as standby, acceptance, and sales, and selectively activates corresponding judgment rules to perform comparative analysis on the target behavioral feature values. During this process, the driving monitoring mechanism automatically switches between routine inspection and anomaly tracing states, thereby adaptively adjusting the data collection frequency of the RPA robot and reducing the overall system power consumption and computing cost. Based on several monitoring mechanisms, it obtains deviation behavioral feature values ​​through judgment operations performed by the first and second rules, identifies gas pump fraudulent behavior, and drives cameras to capture key frames to generate evidence images and send early warning reminders, thus solving the problems mentioned in the background technology.

[0005] (II) Technical Solution To achieve the above objectives, the present invention provides the following technical solution: This application provides a method for intelligent supervision of gas stations that combines AI algorithms and artificial intelligence. The method includes: obtaining target behavioral feature values ​​of the gas station to be supervised. The operation scenario is identified based on the target behavior feature value, and the judgment rule is activated to perform comparative analysis on the target behavior feature value to drive the monitoring mechanism to switch automatically; Obtain deviation behavior feature values; wherein, deviation behavior feature values ​​are feature values ​​obtained by identifying fuel dispenser cheating behavior based on several monitoring mechanisms and performing judgment operations; Information on violations is obtained based on the characteristic values ​​of deviant behavior and the monitoring mechanism. Evidence images are then obtained based on the violation information, and early warning reminders are sent through the evidence images.

[0006] Furthermore, the target behavioral characteristic values ​​of the gas stations to be regulated are obtained, including: The system acquires multi-source data from the gas station to be regulated, including tank data, pump code data, and visual images. This multi-source data is then input into a pre-defined behavior analysis model, which outputs target behavior feature values. The pre-defined behavior analysis model is based on a CNN architecture and includes a feature extraction layer, a feature fusion layer, and a behavior evaluation layer. Feature extraction layer: Perform spatial dimension convolution on the visual image to extract visual features. Simultaneously perform three-layer decomposition on the oil tank data and pump code data to obtain the corresponding first and second components. Extract the Mel features of the corresponding components of each layer and perform feature encoding to obtain temporal encoding. Feature fusion layer: Based on the attention mechanism, visual features and temporal codes are correlated in multiple dimensions to construct a behavior correlation matrix between cans, guns, and vehicles; Behavior evaluation layer: Perform eigenvalue decomposition on the behavior correlation matrix, calculate the eigenvector distribution of the behavior correlation matrix, calculate the information entropy of each eigenvector distribution, perform weighted fusion on the information entropy, and obtain the target behavior feature value.

[0007] Furthermore, if the work scenarios include standby, acceptance, and sales, then the activation judgment rule performs comparative analysis on the target behavior feature values, including: Set the initial first comparison threshold, second comparison threshold, and third comparison threshold; Under the condition of triggering the standby scenario: if the target behavior feature value is greater than or equal to the first comparison threshold, an acknowledgment signal is issued; if the target behavior feature value is less than the first comparison threshold, a waiting signal is issued. Under the conditions of triggering the acceptance scenario: if the target behavior feature value is less than or equal to the second comparison threshold, a waiting signal is issued; if the target behavior feature value is greater than the second comparison threshold, a confirmation signal is issued. Under the conditions that trigger a sales scenario: if the target behavior feature value is greater than the third comparison threshold, a waiting signal is issued; if the target behavior feature value is less than or equal to the third comparison threshold, a confirmation signal is issued.

[0008] Furthermore, if it is a waiting signal, the monitoring mechanism enters the anomaly tracing state; if it is a confirmation signal, the monitoring mechanism enters the routine inspection state.

[0009] Furthermore, identifying fuel dispenser fraud includes: Within a preset monitoring period, by combining one-dimensional convolutional analysis of tank data and pump code data, displacement features and pulse features are extracted, and source identifiers are marked to generate feature data sources; among them, displacement features include at least the liquid level fluctuation frequency, and pulse features include at least the pump code jump frequency. A multivariate regression model is established to calculate the physical drift based on the influence of changes in oil temperature and tank pressure. The physical drift is used to perform dynamic compensation on the displacement characteristics. All compensated feature data sources are collected and aggregated into a feature set. N0 feature data sources are randomly selected from the set and aligned to perform task partitioning. C(N1, N0) monitoring interaction tasks are generated through permutation and combination algorithms. Here, N1 represents the number of feature data sources at any time during the monitoring period, and each monitoring interaction task represents a set of tank and gun association mapping pairs. A rules engine is introduced, including a first rule and a second rule; monitoring interaction tasks that meet the first and second rules are filtered, and counterfeit fuel dispenser cheating behavior is marked.

[0010] Furthermore, a multiple regression model was established to calculate the physical drift based on the effects of oil temperature and tank pressure changes, including: Use oil temperature and tank pressure as independent variables; By analyzing the correlation functions between the independent variables and the corresponding displacement features under different scenarios, the dynamic contribution weight of the independent variables to the displacement features is identified, and the physical drift amount matching the current working scenario is obtained. The correlation function is a regression model that uses the independent variable as the input feature and the physical drift as the output target, and uses machine learning to build the input feature and the output target.

[0011] Furthermore, the first rule is a time deviation constraint: by identifying the timestamps between various feature data sources and calculating the corresponding deviations, several monitoring interaction tasks with deviations greater than the preset standard deviation threshold are selected and determined to satisfy the first rule; Under the condition that the first rule is met, it is determined whether the second rule is met. The second rule is the association of effective constraints: extract the liquid level fluctuation slope of the displacement feature and the pump code jump frequency of the pulse feature, and map them to the multi-dimensional feature space. Perform cluster analysis, calculate the Euclidean distance between the feature points in each monitoring interaction task and the cluster center, and filter the monitoring interaction tasks whose distance is greater than or equal to the preset radius threshold. Then, it is determined that the second rule is met.

[0012] Further, the determination operation is performed, including: The system counts the identification times of fuel dispenser cheating behavior and synchronizes the association mapping pairs between the nozzle and the canister in the current scene; it calculates the time interval corresponding to the identification times of two consecutive cheating behaviors under the same association mapping pair, filters out cases where the time interval is greater than a preset time interval threshold, and collects them into a target time set. For the target time set, extract its two time series and define them as cheating time periods. Arrange the cheating time periods in time sequence and divide the cheating time periods into m equal time slices. If there is a remainder in the result of equal division, then it is m+1; where m is a dynamic value and m is a positive integer. Extract the time delay of each time slice and calculate the standard deviation of the time delay of adjacent time slices. Based on the correlation mapping, assign corresponding weight factors to each time slice. Multiply the standard deviation and the corresponding weight factor to generate the deviation behavior feature value.

[0013] Furthermore, obtaining information about violations includes: Obtain all deviation behavior feature values ​​of the fuel dispenser cheating behavior, define them as the character 'r', and edit them as the violation description character. Combine 'r' with the violation description character to generate violation information.

[0014] Furthermore, after identifying the violation information, the following steps are taken: The system identifies the moment of fraudulent activity at refueling machines, establishes a time-travel window, drives the camera to capture key frame images, and generates evidence images.

[0015] (III) Beneficial Effects This invention provides a smart monitoring method for gas stations that combines AI algorithms and artificial intelligence, which has the following beneficial effects: 1. This invention uses a behavior analysis model. The feature extraction layer performs spatial convolution on visual images and wavelet three-layer decomposition and Mel frequency mapping on oil tank and pump code data. The feature fusion layer uses an attention mechanism to quantify the correlation weight between visual features and temporal fluctuations, generating a behavior correlation matrix between tanks, pumps, and vehicles, thus fully restoring the entire process of gas station operations. The behavior evaluation layer uses eigenvalue decomposition to quantify behavioral features, providing an accurate data foundation for subsequent analysis. 2. This invention automatically identifies operational scenarios including standby, acceptance, and sales by using target behavior feature values, and activates corresponding judgment rules accordingly. It drives the monitoring mechanism to dynamically switch between abnormal tracing and regular inspection states, thereby improving the utilization rate of regulatory resources. By adjusting the execution frequency of the RPA robot in different states, it achieves highly reliable automated monitoring and data collection. 3. This invention identifies fuel dispenser cheating behavior based on several monitoring mechanisms, focuses on tank data and pump code data, establishes a correlation mapping pair between the tank and the nozzle, accurately identifies fuel dispenser cheating behavior, and reduces the false negative rate; through the first rule and the second rule, it greatly improves the efficiency and accuracy of fuel dispenser cheating behavior identification. Attached Figure Description

[0016] Figure 1 This is a flowchart illustrating an intelligent monitoring method for gas stations according to an exemplary embodiment. Detailed Implementation

[0017] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0018] The core of this invention lies in acquiring target behavior feature values ​​of gas stations to quantify the operational behavior between tanks, nozzles, and vehicles, thereby improving the accuracy of regulatory data; automatically identifying operational scenarios based on feature values ​​to drive the monitoring mechanism to intelligently switch between routine inspections and abnormal source tracing, adaptively adjusting the collection frequency to reduce system power consumption and computing costs; identifying fuel dispenser cheating behavior through first and second rules, generating deviation behavior feature values ​​by judging operations, triggering capture and evidence collection after locking in violation information, and promptly pushing early warning information.

[0019] Example: This invention provides a method for intelligent supervision of gas stations that combines AI algorithms with artificial intelligence; Figure 1 This is a flowchart illustrating an intelligent monitoring method for gas stations according to an exemplary embodiment; please refer to [link / reference]. Figure 1 The method includes the following steps: S1: Obtain the target behavioral characteristic values ​​of the gas station to be regulated; The process of acquiring target behavioral characteristics of the gas station to be regulated includes: receiving regulatory instructions, driving the RPA robot to conduct periodic patrols, and the RPA robot sending addressing instructions to the edge acquisition nodes through the remote inventory interface of the patrolled gas station according to the set start time, and acquiring multi-source data within the gas station to be regulated, including tank data, pump code data, and visual images. It should be noted that the RPA robot, as the scheduling core, is responsible for coordinating the start clocks of the tank, pump code, and vision sub-units to ensure that multi-source data is collected within the same window period. The tank data includes tank level and oil temperature: a magnetostrictive level probe deployed inside the tank captures the position of the oil level float and water level float, generating a displacement signal reflecting the real-time liquid level height and bottom water height to obtain the tank level; an integrated temperature sensor within the probe simultaneously captures the oil temperature at different tank levels; pump code data includes nozzle lift-up time, nozzle hang-up time, and cumulative pump code count for each nozzle: a non-invasive pulse interceptor extracts pump code data in real-time from the bus circuit of the fuel dispenser control board, generates a level signal through a pulse encoder, and obtains the cumulative pump code count for each nozzle by real-time accumulation of pulse transitions; the nozzle lift-up and hang-up times are accurately recorded by monitoring the potential transitions of the nozzle lift-up and hang-up signals; visual images are monitored in real-time by high-frequency industrial cameras deployed at the gas stations under supervision. When the RPA robot recognizes the nozzle lift-up pulse, it automatically triggers a visual recognition algorithm to extract a sequence of keyframe images containing license plate information, nozzle number, and operational actions from the video stream.

[0020] A linear interpolation algorithm is used to perform time axis correction to ensure that the sampling timestamps of oil tank data, pump code data, and visual images are consistent, and each data is mapped to a unified high-precision time coordinate system to eliminate time misalignment and form strictly aligned multi-source data. The multi-source data is input into a preset behavior analysis model to output target behavior feature values. The behavior analysis model is based on a CNN architecture and includes a feature extraction layer, a feature fusion layer, and a behavior evaluation layer. Feature Extraction Layer: This layer uses differentiated algorithms to extract corresponding features from different types of input data, transforming raw data into analyzable features. It extracts visual features including personnel posture, fuel nozzle mounting status, and vehicle position by performing spatial convolution on the visual image using convolution operators. Specifically, the convolution operators include several kernel sizes. For example, using a 3×3 kernel with a stride of 1, sliding pixel-by-pixel across the image pixel matrix maximizes the preservation of action details and captures visual features of personnel posture, such as lifting, mounting, and bending over while operating the fuel nozzle. Alternatively, using a 5×5 kernel with a stride of 2 allows for precise anchoring of spatial feature location information while reducing redundant computation, thus improving the accuracy of the fuel nozzle mounting status in the visual image. For example, features are extracted for the positioning of the fuel nozzle / lifting it into the air and the vehicle's parking position. By increasing the size and stride of the convolutional kernel, redundant computation is reduced while the receptive field of the features is expanded. This allows for precise anchoring of the spatial position information of the fuel nozzle and the vehicle to capture the visual features of the fuel nozzle's mounting status and the vehicle's position. It should be noted that during the sliding process, each convolutional kernel performs a dot product operation with the image pixels within the coverage area and adds a bias term. The result is used as the pixel value of the corresponding position in the feature map. Through the stacking of multiple convolutional layers, the pixel information of the original image is gradually transformed into high-dimensional visual features. Furthermore, a ReLU activation function is introduced after each convolutional layer to enhance the non-linear expressive power of the features. Then, a pooling layer is used to reduce the feature dimension, ultimately generating visual features that provide a high-quality visual feature foundation for subsequent fusion analysis with temporal coding. Simultaneously, wavelet transform is used to perform a three-level decomposition on the tank data and pump code data. The db4 wavelet basis function is selected, and three-level wavelet decomposition is performed on the tank data and pump code data respectively. Each level of decomposition includes a low-pass filter or a high-pass filter. The approximate component obtained through the low-pass filter is labeled as the first component, and the detail component obtained through the high-pass filter is labeled as the second component. For example: Level 1 decomposition: obtaining the first component λ1 and the second component η1; Level 2 decomposition: extracting the first component λ1 from the first level and performing a second decomposition, i.e., repeating the filtering operation, to obtain the first component λ2 and the second component η2; Level 3 decomposition: extracting the second component η1 from the first level and performing a second decomposition, to obtain the first component λ3 and the second component η3. For each level of decomposition... The solution obtains a set of components {λ1, η1, λ2, η2, λ3, η3}. For each component, Mel frequency mapping is performed to obtain the corresponding Mel features: the time-domain signal of each component is converted into a frequency-domain signal, and the energy features of each frequency band are extracted through a Mel filter bank to generate the corresponding Mel feature matrix. Then, max pooling is performed on each Mel feature matrix to extract significant features from each layer of Mel features, remove redundant noise, and obtain the corresponding feature sequence. Through feature encoding, the feature sequences of the first, second, and third layers are multidimensionally concatenated to obtain a concatenated vector. The concatenated vector is input into a channel convolutional pooling layer for fusion encoding, and then the feature dimension is reduced by a pooling layer to finally output the temporal code. Feature fusion layer: Based on the attention mechanism, visual features and temporal codes are associated in multiple dimensions. It should be noted that before performing multi-dimensional association, the dimensions of the temporal codes are mapped to the same dimensions as the visual features through a fully connected layer, and spatiotemporal alignment is performed. For example, the visual features at a certain moment represent the state of the oil gun and personnel actions at that moment, and the temporal codes at the corresponding moment represent the tank level fluctuations and pump code pulse jumps at that moment. Using visual features as the query vector and temporal codes as the key and value vectors, the dot product similarity between the query vector and the key vector is calculated to quantify the association strength between visual features and temporal codes. This is achieved by multiplying each element of the query vector with the corresponding element of the key vector, summing all the results, and obtaining a scalar similarity value. It should be noted that if the vector dimension is large, the dot product result will be large, affecting gradient stability. Therefore, the similarity value needs to be scaled, i.e., the similarity value obtained from the dot product operation is divided by the square root of the vector dimension. The softmax function is then used to obtain the association weights between visual features and temporal fluctuations at different time sampling points. For example: For the dot product similarity or scaled similarity values ​​at the same sampling point, input the softmax function for normalization so that the sum of all similarity values ​​is 1; the normalized value is the association weight. The higher the weight, the stronger the association between the visual features and the temporal coding. The three types of equipment, tanks, guns, and vehicles, are used as the row dimensions of the matrix, such as: row dimension definition: row 1 = tank A, row 2 = gun 1, row 3 = vehicle 1; the fusion feature dimension of visual features and temporal coding is used as the column dimension of the matrix, such as: multiply the association weights by the value vector and fill each row to obtain the behavioral association matrix between tanks, guns, and vehicles; Behavior Evaluation Layer: This layer performs eigenvalue decomposition on the behavior association matrix, calculates the energy proportion of each eigenvalue, and pairs any eigenvalue with its corresponding energy proportion to form an ordered pair, creating a set of eigenvalue energy distributions. It then calculates the information entropy of each feature's energy distribution. The smaller the entropy, the more concentrated the energy of that feature dimension, and the clearer the corresponding behavioral feature; conversely, the larger the entropy, the more dispersed the energy of that feature dimension, and the more ambiguous the corresponding behavioral feature. Weighted fusion is then performed on the information entropy, using the energy proportion of each feature dimension as the weight. Through weighted summation, the target behavioral feature value is normalized and output. It should be noted that the closer the target behavioral feature value is to 1, the more concentrated the energy distribution and the more significant the behavioral feature; conversely, the closer the target behavioral feature value is to 0, the more dispersed the energy distribution and the less significant the behavioral feature.

[0021] Through the behavior analysis model, the feature extraction layer performs spatial convolution on visual images and wavelet three-layer decomposition and Mel frequency mapping on oil tank and pump code data. The feature fusion layer uses an attention mechanism to quantify the correlation weight between visual features and temporal fluctuations, generating a behavior correlation matrix between tanks, pumps, and vehicles, thus fully reconstructing the entire process of gas station operations. The behavior evaluation layer quantifies behavioral features through eigenvalue decomposition, providing an accurate data foundation for subsequent analysis.

[0022] S2: Identify the work scenario based on the target behavior feature value, and activate the judgment rule to perform comparative analysis on the target behavior feature value to drive the monitoring mechanism to switch automatically; among them, the work scenario includes standby, acceptance and sales; Based on the distribution range of target behavior feature values, the current gas station to be monitored is mapped to a preset operating scenario, and the judgment rules of each scenario are activated simultaneously. The following is an explanation of each operating scenario: Standby scenario: refers to the static state in which there are no external vehicles refueling or tanker trucks unloading oil at the gas station; Acceptance scenario: refers to the operating state in which the gas station receives supplies from the upstream oil depot and tanker trucks unload and fill oil into the underground storage tanks in the station; Sales scenario: refers to the real-time operating state in which the fuel dispenser is on and provides refueling services to customer vehicles. The activation judgment rule configuration includes several monitoring mechanisms, including: under the condition of triggering a standby scenario: if the target behavior feature value is greater than or equal to the first comparison threshold, a confirmation signal is issued and the monitoring mechanism enters the normal inspection state; if the target behavior feature value is less than the first comparison threshold, a waiting signal is issued and the monitoring mechanism enters the abnormal source tracing state, indicating that in the standby scenario, an abnormal situation occurs such as hardware offline, sensor being shielded, or power being disconnected; wherein, the first comparison threshold is the minimum target behavior feature value in the standby scenario; Under the conditions of triggering the acceptance scenario: if the target behavior feature value is less than or equal to the second comparison threshold, the oil unloading operation is determined to be orderly, a waiting signal is issued, and the monitoring mechanism enters the abnormal source tracing state; if the target behavior feature value is greater than the second comparison threshold, a confirmation signal is issued, and the monitoring mechanism enters the routine inspection state; where the second comparison threshold is the target behavior feature value of the last successfully executed acceptance operation, for example: the target behavior feature value when the oil unloading just started. Under the condition of triggering a sales scenario: if the target behavior feature value is less than or equal to the third comparison threshold, the refueling behavior is determined to be valid, a waiting signal is issued, and the monitoring mechanism enters the abnormal source tracing state; if the target behavior feature value is greater than the third comparison threshold, a confirmation signal is issued, and the monitoring mechanism enters the routine inspection state; where the third comparison threshold is the target behavior feature value of the last successfully executed sales operation, for example: the target behavior feature value when the vehicle was just detected picking up the nozzle to refuel. When the monitoring mechanism enters the anomaly tracing state, it drives the RPA robot to perform second-level data acquisition, such as extracting oil tank data every 5 seconds; when the monitoring mechanism enters the routine inspection state, it drives the RPA robot to perform minute-level data acquisition, such as extracting oil tank data every 10 minutes.

[0023] Based on target behavioral feature values, the system automatically identifies operational scenarios including standby, acceptance, and sales, and activates corresponding judgment rules accordingly. This drives the monitoring mechanism to dynamically switch between abnormal tracing and routine inspection states, optimizing the allocation of computing resources and data storage pressure for gas station data supervision, improving the utilization rate of regulatory resources, and achieving highly reliable automated monitoring and data collection by adjusting the execution frequency of the RPA robot in different states.

[0024] S3: Obtain the deviation behavior feature value; wherein, the deviation behavior feature value is the feature value obtained by identifying the fuel dispenser cheating behavior based on several monitoring mechanisms and performing a judgment operation; Several monitoring mechanisms were used to identify fuel dispenser cheating behavior, including: within a preset monitoring period, combining one-dimensional convolutional analysis of tank data and pump code data; for example, the convolutional kernel size was set to 5, and the kernel slid across the temporal feature sequence with a stride of 1. Dot product operations were performed on 5 consecutive sampling points within a window to capture displacement and pulse features. For each displacement and pulse feature, a corresponding source identifier was added, such as: Tank A, Gun 1, Pulse Feature. The displacement and pulse features from all source identifiers were integrated to form a structured feature data source. In this context, displacement characteristics include the liquid level fluctuation slope, and pulse characteristics include the pump code transition frequency and pulse width. The following is an explanation of the relevant terms: Liquid level fluctuation slope: The rate of change of the tank liquid level height per unit time, reflecting how fast the liquid level rises / falls. For example, cheating with a fuel dispenser may result in an abnormal slope. Pump code transition frequency: The number of transitions in the pump code pulse signal per unit time. For example, during normal refueling, it is positively correlated with the amount of fuel dispensed; during cheating, the transition frequency may be abnormally high / low. Pulse width: The duration of a single pulse level signal from its rising edge to its falling edge. A multiple regression model was established to calculate the physical drift based on the effects of changes in oil temperature and tank pressure, specifically including: Extract oil temperature and internal pressure from the oil tank data; for standby, acceptance, and sales operation scenarios, use oil temperature and internal pressure as independent variables in each scenario to identify the dynamic contribution weight of oil temperature and internal pressure to displacement characteristics. For example, the dynamic contribution weight of internal pressure is higher in the acceptance scenario, and the dynamic contribution weight of oil temperature is higher in the standby scenario, ensuring that the weights accurately match the current operation scenario. Calculate the physical drift amount that perfectly matches the current operation scenario; establish a correlation function between the independent variables and the corresponding physical drift amount, where the correlation function uses the independent variables as input features and the physical drift amount as the output target, and uses machine learning to build a regression model between the input features and the output target; after completing the physical drift amount calculation, perform dynamic compensation on the displacement characteristics. Dynamic compensation for displacement features is performed using physical drift, including: acquiring displacement features extracted by a one-dimensional convolutional neural network, quantizing them, and subtracting the physical drift calculated at the corresponding time in real time to obtain compensated displacement features. This eliminates the normal influence of environmental physical factors such as oil temperature and tank pressure on liquid level fluctuations, ultimately obtaining compensated displacement features that truly reflect the actual changes in tank liquid level; collecting all compensated feature data sources and aggregating them into a feature set, randomly selecting N0 feature data sources from which alignment tasks are performed, and generating C(N1, N0) monitoring interaction tasks through permutation and combination algorithms; where N1 represents the number of feature data sources at any time during the monitoring period, and each monitoring interaction task represents a set of potential tank-gun association mapping pairs; For example, within a certain monitoring period, the total number of feature data sources at any given time is N1=4. These 4 feature data sources are: Feature Data Source 1: Tank A, compensated displacement feature: time t1, feature value: actual liquid level fluctuation; Feature Data Source 2: Tank B, compensated displacement feature: time t1, feature value: actual liquid level fluctuation; Feature Data Source 3: Gun 1, pulse feature: time t1, feature value: number of pulse jumps per unit time; Feature Data Source 4: Gun 2, pulse feature: time t1, feature value: duration of a single pulse. Based on this, all dynamically compensated feature data sources are first comprehensively collected. These feature data with unified timestamps and identifiers are systematically sorted and aggregated into a unified feature set. Then, data from this feature set... Two feature data sources (N0=2) are randomly selected. Alignment tasks based on timestamp and device type are performed on these two data sources to ensure they belong to the same monitoring time and cover both tank and nozzle devices. Based on the correspondence between the total number of feature data sources (N1=4) and the number of selected sources (N0=2), the number of combinations (C(4,2)=8) is calculated using a permutation and combination algorithm. Eight corresponding monitoring interaction tasks are then generated based on this number of combinations. Each monitoring interaction task represents a potential tank-nozzle association mapping pair, such as: Tank A, compensated displacement feature + Nozzle 1, pulse feature; Tank B, compensated displacement feature + Nozzle 2, pulse feature, etc. These monitoring interaction tasks provide accurate task carriers for subsequent analysis of fuel dispenser cheating behavior. A rule engine is introduced, including a first rule and a second rule. The first rule is a time deviation constraint: by identifying the timestamps between various feature data sources and calculating the corresponding deviations, several monitoring interaction tasks with deviations greater than a preset standard deviation threshold are selected and determined to meet the first rule; if the deviation is less than or equal to the preset standard deviation threshold, no processing is performed. This process achieves the initial screening of monitoring interaction tasks where fuel dispenser cheating occurs, reducing the data processing volume of subsequent feature association analysis and improving the efficiency of cheating identification. The source of the standard deviation threshold is: collecting historical deviation data that meets the first rule, calculating the mean and standard deviation of the deviation data through statistical analysis, and using the sum of the mean and a multiple of the standard deviation as the standard deviation threshold. It should be noted that the value of a multiple of 3 is just an example, and the specific setting should be based on the actual situation, which will not be elaborated here. Under the condition that the first rule is met, it is determined whether the second rule is met. The second rule is the association of valid constraints: extract the liquid level fluctuation slope of the displacement feature and the pump code jump frequency of the pulse feature, and map them to a multi-dimensional feature space. Perform cluster analysis, calculate the Euclidean distance between the feature points in each monitoring interaction task and the cluster center, and filter the monitoring interaction tasks whose distance is greater than or equal to the preset radius threshold. If the distance is less than the preset radius threshold, it means that the second rule is not met, and the radius threshold is recorded. This process realizes the secondary screening of the monitoring interaction tasks where suspected fuel dispenser cheating behavior occurs. The monitoring interaction tasks that meet both the first and second rules are screened and marked as having fuel dispenser cheating behavior. The source of the radius threshold: By collecting effective feature pairs of displacement and pulse features from historical operation scenarios at gas stations, an operation feature dataset covering the entire scenario is constructed; a clustering analysis process completely consistent with the current monitoring is performed on this dataset, and all distances less than the preset radius threshold are statistically analyzed to obtain the corresponding quantiles. The initial radius threshold is determined using the 95th percentile statistical method, and the corresponding radius thresholds are calibrated for the feature distribution differences of different operation scenarios, which will not be elaborated here.

[0025] The judgment operation includes: counting the identification times of fuel dispenser cheating behavior and synchronizing the association mapping pairs between nozzles and tanks in the current scenario; calculating the time interval corresponding to the identification times of two consecutive cheating behaviors under the same association mapping pair, filtering out cases where the time interval is greater than a preset interval threshold, and aggregating them into a target time set; for the target time set, extracting its two time series, defining them as cheating time periods, and arranging the cheating time periods in time series and labeling them as T1, T2, T3, ..., Tn; where the time series is a dynamic variable, and n represents the number of times fuel dispenser cheating behavior is identified; dividing the cheating time period into m equal time slices, and if there is a remainder in the result of equal division, then it is m+1; where m is a dynamic value, and m is a positive integer; extracting the delay of each time slice, and calculating the standard deviation of the corresponding delay of adjacent time slices, assigning corresponding weight factors based on the association mapping pairs, multiplying the standard deviation and the corresponding weight factors to generate the deviation behavior feature value; delay: the time difference between the corresponding features changing in the same set of tank and nozzle association mapping pairs; Source of weight factors: The target time set is used as the test dataset and input into the deep learning network for training. During training, the mean squared error method is used to measure the linear relationship between the standard deviation and the weights. Larger weights are automatically assigned to high standard deviations and smaller weights are assigned to low priority levels. The larger the standard deviation, the more severe the time delay fluctuations and the more abnormal the canister gun association, and the greater the impact on the identification of fuel dispenser fraud. The assigned weights are used as the initial weight factors under the time segmentation.

[0026] By identifying fuel dispenser fraud through several monitoring mechanisms, focusing on tank data and pump code data, and establishing a correlation mapping between the tank and the nozzle, fraudulent activities can be accurately identified, reducing the false negative rate. In this process, a multivariate regression model is established to calculate the physical drift based on changes in fuel temperature and tank pressure. Dynamic compensation is performed on the displacement characteristics of different operating scenarios. This eliminates the influence of environmental physical factors on liquid level fluctuations, ensuring data authenticity, and avoids misjudgments due to differences in operating scenarios, improving the versatility and robustness of the solution. Through the first and second rules, the efficiency and targeting of fraud detection are significantly improved.

[0027] S4: Obtain violation information based on deviation behavior feature values ​​and monitoring mechanisms, acquire evidence images based on the violation information, and send early warning reminders through the evidence images; Obtaining violation information includes: acquiring all deviation behavior feature values ​​of the fuel dispenser cheating behavior, defining them as the character 'r', and editing them as violation description characters. The 'r' is then combined with the violation description characters to generate violation information. After identifying the violation information, the moment of identification of the fuel dispenser cheating behavior is locked, a time backtracking window is established, and the camera is driven to capture key frame images. The RPA robot retrieves the complete time period containing key actions such as nozzle lifting, nozzle insertion, and refueling start from the visual images to generate evidence images. The evidence images are uploaded to the management platform to send early warning reminders, enabling rapid response and timely handling of violations, effectively stopping fuel dispenser cheating behavior, controlling quality, and continuously improving the intelligent level of gas station supervision by relying on the storage and tracking functions of the management platform.

[0028] In the application, the various formulas mentioned are all calculated by removing dimensions and taking their numerical values. The formulas are derived from the most recent real-world situation by collecting a large amount of data and conducting software simulations. The formulas are set by those skilled in the art according to the actual situation.

[0029] The above embodiments can be implemented, in whole or in part, by software, hardware, firmware, or any other combination thereof. When implemented in software, the above embodiments can be implemented, in whole or in part, as a computer program product. Those skilled in the art will recognize that the units and algorithm steps of the various examples described in conjunction with the embodiments disclosed herein can be implemented in electronic hardware, or a combination of computer software and electronic hardware. Whether these functions are implemented in hardware or software depends on the specific application and design constraints of the technical solution.

[0030] The units described as separate components may or may not be physically separate. The components shown as units may or may not be physical units; they may be located in one place or distributed across multiple network units. Some or all of the units can be selected to achieve the purpose of this embodiment, depending on actual needs.

[0031] The above description is merely a specific embodiment of this application, but the scope of protection of this application is not limited thereto. Any changes or substitutions that can be easily conceived by those skilled in the art within the scope of the technology disclosed in this application should be included within the scope of protection of this application.

Claims

1. A method for intelligent supervision of gas stations that combines AI algorithms and artificial intelligence, characterized in that: This method includes: obtaining target behavioral feature values ​​of the gas stations to be regulated; The operation scenario is identified based on the target behavior feature value, and the judgment rule is activated to perform comparative analysis on the target behavior feature value to drive the monitoring mechanism to switch automatically; Obtain deviation behavior feature values; wherein, deviation behavior feature values ​​are feature values ​​obtained by identifying fuel dispenser cheating behavior based on several monitoring mechanisms and performing judgment operations; Information on violations is obtained based on the characteristic values ​​of deviant behavior and the monitoring mechanism. Evidence images are then obtained based on the violation information, and early warning reminders are sent through the evidence images.

2. The intelligent supervision method for gas stations using AI algorithms and artificial intelligence as described in claim 1, characterized in that, Obtain the target behavioral characteristic values ​​of the gas stations to be regulated, including: The system acquires multi-source data from the gas station to be regulated, including tank data, pump code data, and visual images. This multi-source data is then input into a pre-defined behavior analysis model, which outputs target behavior feature values. The pre-defined behavior analysis model is based on a CNN architecture and includes a feature extraction layer, a feature fusion layer, and a behavior evaluation layer. Feature extraction layer: Perform spatial dimension convolution on the visual image to extract visual features. Simultaneously perform three-layer decomposition on the oil tank data and pump code data to obtain the corresponding first and second components. Extract the Mel features of the corresponding components of each layer and perform feature encoding to obtain temporal encoding. Feature fusion layer: Based on the attention mechanism, visual features and temporal codes are correlated in multiple dimensions to construct a behavior correlation matrix between cans, guns, and vehicles; Behavior evaluation layer: Perform eigenvalue decomposition on the behavior correlation matrix, calculate the eigenvector distribution of the behavior correlation matrix, calculate the information entropy of each eigenvector distribution, perform weighted fusion on the information entropy, and obtain the target behavior feature value.

3. The intelligent supervision method for gas stations using AI algorithms and artificial intelligence as described in claim 1, characterized in that, The work scenarios include standby, acceptance, and sales. The activation judgment rule performs comparative analysis on the target behavior feature values, including: Set the initial first comparison threshold, second comparison threshold, and third comparison threshold; Under the condition of triggering the standby scenario: if the target behavior feature value is greater than or equal to the first comparison threshold, an acknowledgment signal is issued; if the target behavior feature value is less than the first comparison threshold, a waiting signal is issued. Under the conditions of triggering the acceptance scenario: if the target behavior feature value is less than or equal to the second comparison threshold, a waiting signal is issued; if the target behavior feature value is greater than the second comparison threshold, a confirmation signal is issued. Under the conditions that trigger a sales scenario: if the target behavior feature value is greater than the third comparison threshold, a waiting signal is issued; if the target behavior feature value is less than or equal to the third comparison threshold, a confirmation signal is issued.

4. The intelligent supervision method for gas stations using AI algorithms and artificial intelligence as described in claim 3, characterized in that, If it is a waiting signal, the monitoring mechanism enters the anomaly tracing state; if it is a confirmation signal, the monitoring mechanism enters the routine inspection state.

5. A method for intelligent supervision of gas stations using AI algorithms and artificial intelligence as described in claim 2, characterized in that, Identify fraudulent activities at fuel dispensers, including: Within a preset monitoring period, by combining one-dimensional convolutional analysis of tank data and pump code data, displacement features and pulse features are extracted, and source identifiers are marked to generate feature data sources; among them, displacement features include at least the liquid level fluctuation frequency, and pulse features include at least the pump code jump frequency. A multivariate regression model is established to calculate the physical drift based on the influence of changes in oil temperature and tank pressure. The physical drift is used to perform dynamic compensation on the displacement characteristics. All compensated feature data sources are collected and aggregated into a feature set. N0 feature data sources are randomly selected from the set and aligned to perform task partitioning. C(N1, N0) monitoring interaction tasks are generated through permutation and combination algorithms. Here, N1 represents the number of feature data sources at any time during the monitoring period, and each monitoring interaction task represents a set of tank and gun association mapping pairs. A rules engine is introduced, including a first rule and a second rule; monitoring interaction tasks that meet the first and second rules are filtered, and counterfeit fuel dispenser cheating behavior is marked.

6. The intelligent supervision method for gas stations using AI algorithms and artificial intelligence as described in claim 5, characterized in that, A multiple regression model was established to calculate the physical drift based on the effects of oil temperature and tank pressure changes, including: Use oil temperature and tank pressure as independent variables; By analyzing the correlation functions between the independent variables and the corresponding displacement features under different scenarios, the dynamic contribution weight of the independent variables to the displacement features is identified, and the physical drift amount matching the current working scenario is obtained. The correlation function is a regression model that uses the independent variable as the input feature and the physical drift as the output target, and uses machine learning to build the input feature and the output target.

7. The intelligent supervision method for gas stations using AI algorithms and artificial intelligence as described in claim 5, characterized in that, The first rule is a time deviation constraint: by identifying the timestamps between various feature data sources and calculating the corresponding deviations, several monitoring interaction tasks with deviations greater than the preset standard deviation threshold are selected and determined to meet the first rule. Under the condition that the first rule is met, it is determined whether the second rule is met. The second rule is the association of effective constraints: extract the liquid level fluctuation slope of the displacement feature and the pump code jump frequency of the pulse feature, and map them to the multi-dimensional feature space. Perform cluster analysis, calculate the Euclidean distance between the feature points in each monitoring interaction task and the cluster center, and filter the monitoring interaction tasks whose distance is greater than or equal to the preset radius threshold. Then, it is determined that the second rule is met.

8. A method for intelligent supervision of gas stations using AI algorithms and artificial intelligence as described in claim 7, characterized in that, Perform the judgment operation, including: The system counts the identification times of fuel dispenser cheating behavior and synchronizes the association mapping pairs between the nozzle and the canister in the current scene; it calculates the time interval corresponding to the identification times of two consecutive cheating behaviors under the same association mapping pair, filters out cases where the time interval is greater than a preset time interval threshold, and collects them into a target time set. For the target time set, extract its two time series and define them as cheating time periods. Arrange the cheating time periods in time sequence and divide the cheating time periods into m equal time slices. If there is a remainder in the result of equal division, then it is m+1; where m is a dynamic value and m is a positive integer. Extract the time delay of each time slice and calculate the standard deviation of the time delay of adjacent time slices. Based on the correlation mapping, assign corresponding weight factors to each time slice. Multiply the standard deviation and the corresponding weight factor to generate the deviation behavior feature value.

9. A method for intelligent supervision of gas stations using AI algorithms and artificial intelligence as described in claim 1, characterized in that, Obtaining information about violations, including: Obtain all deviation behavior feature values ​​of the fuel dispenser cheating behavior, define them as the character 'r', and edit them as the violation description character. Combine 'r' with the violation description character to generate violation information.

10. A method for intelligent supervision of gas stations using AI algorithms and artificial intelligence as described in claim 9, characterized in that, After identifying the violation information, the following are included: The system identifies the moment of fraudulent activity at refueling machines, establishes a time-travel window, drives the camera to capture key frame images, and generates evidence images.

Images