An automatic inspection method and system for intelligent fire-fighting equipment of an LNG filling station

Abstract

The present application relates to a kind of LNG filling station intelligent fire-fighting equipment automation inspection method and system, specifically related to the field of LNG filling station intelligent fire-fighting equipment, and the dynamic drift model is constructed by real-time acquisition of environmental parameters, and the interference of sensor by adverse conditions is accurately quantified;Drift trend is captured in advance using time series prediction network, and equipment state under extreme scenario is simulated in combination with digital twin technology;Based on unsupervised clustering, abnormal patterns are intelligently identified and multi-source data fusion weights are dynamically optimized;Finally, through reinforcement learning closed-loop iteration drift model and fusion strategy, continuously improve system adaptability, effectively eliminate misjudgment and missed judgment caused by environmental interference throughout the process, significantly enhance the real-time and reliability of fire warning, greatly reduce the risk of safety accident, provide all-weather, high-precision, self-evolution intelligent protection for LNG filling station.

Description

TECHNICAL FIELD

[0001] The present application relates to the field of intelligent fire-fighting equipment for LNG filling stations, and more particularly, to an automatic inspection method and system for intelligent fire-fighting equipment of LNG filling stations. BACKGROUND

[0002] As an important energy infrastructure, the safe operation of LNG filling stations is of great importance. LNG has the characteristics of low temperature and flammability and explosion, and once leakage occurs, it is easy to cause fire, explosion and other serious accidents, resulting in significant casualties, property losses and environmental pollution. Therefore, the implementation of all-weather and high-reliability fire safety monitoring and early warning for LNG filling stations is the core link to ensure their safe operation. In the actual operating environment, especially in the LNG filling stations widely distributed in the coastal areas of China, the fire monitoring system of these stations faces extremely severe challenges. These stations are exposed to high humidity, high salt fog, low temperature and even typhoon and other extreme weather conditions all year round. At the same time, the environment in the station is complex, and there are electromagnetic interference generated by the operation of various electrical equipment, as well as problems such as sensor aging and performance degradation caused by long-term operation of equipment. The combined effects of these internal and external environmental factors have a significant impact on the data acquisition accuracy and reliability of various fire monitoring sensors (such as combustible gas concentration sensors, temperature sensors, pressure sensors, flame detectors, video monitoring and thermal imaging equipment, etc.) deployed in the station. For example, low temperature environment may cause gas sensor response delay or sensitivity decline, high humidity and salt fog may corrode sensor probes or cause optical lens fogging, thereby affecting the clarity of video monitoring and thermal imaging, and electromagnetic interference may introduce signal noise. These factors make the multi-modal data (such as concentration data, temperature data, image / video data) obtained from different physical principle sensors prone to distortion, increased noise or inconsistency between data in harsh environments.

[0003] In the prior art, the intelligent fire-fighting system of the LNG filling station generally adopts a multi-sensor data fusion technology, aiming to improve the accuracy and reliability of state perception, and reduce the false alarm and missed alarm rates by comprehensively processing monitoring information from different sources. Common fusion methods include Kalman filtering algorithm, weighted average method, D-S evidence theory, etc. These methods can play a certain role under the condition that the environment is relatively stable and the sensor performance is good. However, the core problem is that most of the existing fusion algorithms are designed based on static or ideal environment assumptions, and lack the ability to adaptively correct the dynamic changes of the environment and the performance drift of the sensor. When facing the actual existing, continuously changing and extremely harsh operating environment of the coastal LNG filling station, the limitations of the existing fusion technology are exposed. The algorithm cannot effectively distinguish and separate the data noise introduced by environmental interference or sensor state abnormality from the real risk signal, resulting in distorted fusion results. Specifically, under harsh environment, the system may frequently produce false alarms due to distorted fusion data, triggering unnecessary emergency responses such as spraying and shutdown, causing resource waste and operation interruption. More seriously, it may also miss the key opportunity for early disposal due to the failure of the fusion algorithm to effectively identify the real leakage signal masked by noise, and bury the major safety hazards of fire and explosion. Therefore, it is urgent to develop a multi-modal data intelligent fusion method that can adapt to the harsh actual operating environment of the LNG filling station, has environment perception and dynamic correction capability, and significantly improves the perception accuracy, early warning accuracy and overall reliability of the fire-fighting monitoring system under complex conditions. SUMMARY

[0004] The present application provides an automatic inspection method and system for intelligent fire-fighting equipment of LNG filling stations to solve the problems raised in the background art.

[0005] The technical solution of the present application to solve the above technical problems is as follows: an automatic inspection method for intelligent fire-fighting equipment of LNG filling stations, specifically comprising the following steps:

[0006] Step S1, real-time collection of environmental parameter data in the filling station, including temperature, humidity, salt fog concentration and wind speed, construction of an environmental parameter vector based on the collected environmental parameter data, execution of a lightweight neural network training operation, input of historical environmental data and sensor calibration bias records, and output of drift factor model parameters; calculation of the drift amount of each sensor under the current environment according to the drift factor model parameters, and generation of a drift matrix;

[0007] Step S2, inputting the environmental parameter vector and the drift matrix into a double-channel LSTM network in time sequence, executing a time series prediction operation, and outputting drift trend data of each sensor in the next 5 to 10 minutes;

[0008] Step S3, load the drift trend data in the digital twin platform, simulate the sensor monitoring data in the harsh environment, perform unsupervised clustering operation to identify the abnormal pattern in the simulation data, and dynamically allocate the fusion weight of the multi-modal data according to the abnormal pattern recognition result;

[0009] Step S4, obtain the actual fire alarm event data and the digital twin simulation result data, calculate the fusion error value between them, perform reinforcement learning optimization operation, update the drift factor model parameters and the fusion weight allocation strategy, and complete the closed-loop feedback;

[0010] In a preferred embodiment, in the step S1, the specific operation of constructing the environment parameter vector based on the collected environment parameter data is as follows:

[0011] Synchronously collect the real-time environment parameter data of the temperature and humidity sensors, salt fog monitors and anemometers deployed in the gas station tank farm, gas dispensers and pipe corridors, the collection trigger condition is periodic inspection task activation, and the collection period is not more than 10 seconds; the read environment parameter data includes environmental temperature, relative humidity, salt fog concentration and wind speed, wherein the environmental temperature range is-40 degrees Celsius to 60 degrees Celsius, the relative humidity range is 0% to 100%, the salt fog concentration range is 0 milligrams per cubic meter to 500 milligrams per cubic meter, and the wind speed range is 0 meters per second to 60 meters per second; perform sliding window mean filtering operation on the environment parameter data, the window size is 5 sampling points; generate an environment parameter vector based on the filtered environment parameter data, the vector is composed of four components in a fixed order: the first component is the measured value of the environmental temperature, the second component is the measured value of the relative humidity, the third component is the measured value of the salt fog concentration, and the fourth component is the measured value of the wind speed;

[0012] The specific operation of performing lightweight neural network training operation is as follows:

[0013] The input historical environment data is derived from the historical harsh event data set of the gas station, which contains at least three kinds of environment parameter time series data and corresponding sensor calibration bias records recorded during typhoon and high salt fog season;

[0014] The training process adopts the following processing performed independently for each sensor: firstly, reading the current environmental parameter vector, performing a dot product operation with the environmental weight vector in the drift factor model parameter, and superimposing the bias term, inputting the operation result into a linear activation function with leakage correction, outputting the intermediate drift amount, and then multiplying the intermediate drift amount by the sensitivity coefficient of the sensor to generate the final drift amount, wherein the activation function outputs a proportionally reduced negative value when the input value is negative, and outputs the original value when the input value is non-negative; the training target is to minimize the absolute value error sum of the predicted drift amount and the measured drift amount, while constraining the Euclidean norm of the weight vector; the final output drift factor model parameter includes a four-dimensional environmental weight vector corresponding to each sensor and a bias term, wherein the four-dimensional weight corresponds to the environmental influence coefficient of temperature, humidity, salt mist concentration and wind speed.

[0015] In a preferred embodiment, the specific operation of generating the drift matrix is as follows:

[0016] The drift amounts of all sensors deployed at the gas station are arranged in order of number as a drift amount sequence, and are constructed into a drift matrix in the form of a diagonal matrix, and the diagonal line elements of the matrix are the drift amounts of the sensors, and the non-diagonal line elements are set to zero value.

[0017] In a preferred embodiment, the specific steps of performing the time series prediction operation in step S2 are as follows:

[0018] The drift amounts of all sensors are extracted from the drift matrix to generate a drift vector in order of sensor number; at the same time, the environmental parameter vector is organized into an environmental tensor in time sequence; the length of the time window is set to 10 minutes, so that the environmental tensor contains a sequence of environmental parameter vectors for 10 consecutive minutes, and the drift vector contains a sequence of drift amounts for 10 consecutive minutes;

[0019] Two independent processing branches of the double-channel LSTM network are constructed: the first branch is an environmental feature processing channel, which uses a gated recurrent unit structure to process the environmental tensor to capture the slow-changing trend of meteorological parameters; the second branch is a drift feature processing channel, which uses a long short-term memory unit enhanced by a hollow convolution to process the drift vector to capture the fast-changing transient state of the sensor drift; the hidden state outputs of the two branches are fused through a dynamic weight distribution mechanism, wherein the weight coefficient of the environmental channel hidden state is calculated by a coupling function, and the weight coefficient of the drift channel hidden state is 1 minus the weight coefficient of the environmental channel hidden state.

[0020] In a preferred embodiment, the specific operation of outputting the drift trend data of each sensor within the next 5 to 10 minutes is as follows:

[0021] The fusion hidden state sequence output by the dynamic weight allocation mechanism is input into a time domain convolutional neural network to perform multi-step prediction calculation; the fusion hidden state sequence is generated by weighting and superimposing the environment channel hidden state and the drift channel hidden state by a coupling coefficient, the prediction time span is 5 to 10 minutes, and the output dimension is a two-dimensional matrix of time steps multiplied by the number of sensors; each matrix element represents the predicted value of the drift amount of the sensor corresponding to the sensor number at a future time point, wherein the sensor numbers are arranged in the order of the deployment of the gas station, and the time points are continuously distributed in time sequence; the time domain convolutional neural network includes three dilated convolutional layers, each with a kernel size of 3 and a dilation rate of 1, 2, and 4 in turn; the output layer uses a linear activation function, and the loss function is the smooth L1 norm error of the predicted drift amount and the actual drift amount; after the prediction result is processed by inverse normalization, drift trend data that can be directly input into the digital twin platform is generated.

[0022] In a preferred embodiment, in step S3, the specific step of simulating the generation of sensor monitoring data in harsh environments is:

[0023] A virtual sensor network corresponding to the real gas station sensor network is constructed in the digital twin platform, and the drift trend data from step S2, which is a two-dimensional matrix containing the predicted drift amounts of all sensors within the next ten minutes, is loaded. At the same time, a pre-set physical field simulation model is called, which includes a computational fluid dynamics model and a multi-physical field coupling equation. The predicted drift amount and the simulated environmental tensor at the corresponding time are jointly input into the physical field simulation model to calculate the basic readings of each virtual sensor under simulated harsh environments. On this basis, a Gaussian distribution random noise with a mean of zero and a fixed standard deviation is introduced, and the noise is superimposed on the basic readings to simulate the measurement error of the real sensor. Finally, an analog data set is output, which completely contains the simulated monitoring data of all sensors within the next ten minutes under the influence of different harsh environmental factors with noise.

[0024] The specific steps of identifying abnormal patterns in the simulated data are:

[0025] The multi-dimensional feature extraction is performed on the sensor monitoring data generated by the digital twin platform in a harsh environment, and the extracted features include the mean value, variance, first-order difference fluctuation amplitude and energy proportion after principal component analysis of the frequency spectrum of the sensor values; the extracted feature parameters are organized into a feature vector set according to a time window; the improved density peak clustering algorithm is used to process the set, and the local density value of each feature vector is calculated by using an exponential kernel function, wherein the kernel function decay coefficient is associated with a preset cutoff distance; the local density value and the relative distance of all feature vectors are compared, and the feature vectors with a local density lower than a first threshold and a relative distance higher than a second threshold are screened out as abnormal mode points; the abnormal mode points adjacent in space are merged to form an abnormal clustering cluster, which is marked as an abnormal mode set significantly deviating from the normal data distribution.

[0026] In a preferred embodiment, the specific steps of dynamically assigning the multi-modal data fusion weight according to the abnormal mode recognition result are as follows:

[0027] The abnormal mode set recognized by the unsupervised clustering operation and a plurality of preset modal centers are obtained, and each modal center represents a typical operating state mode; for each preset modal, the energy entropy value thereof is calculated, and the calculation process of the energy entropy value is as follows: first, the soft assignment probability of each abnormal mode point in the abnormal mode set to the modal is calculated, and the probability value is negatively correlated with the distance from the abnormal mode point to the modal center; then, according to the soft assignment probability distribution of all abnormal mode points, the energy entropy of the modal is calculated; subsequently, the fusion weight of each modal is dynamically calculated according to the energy entropy value thereof, and the calculation principle is that the higher the energy entropy value of the modal, the lower the fusion weight assigned thereto; finally, the fusion weights of all modes are normalized to output a normalized weight vector for multi-modal data fusion.

[0028] In a preferred embodiment, in the step S4, the specific steps of calculating the fusion error value between the two are as follows:

[0029] The actual fire alarm event data set and the digital twin simulation result data set are obtained, and the data points in the two sets are processed for time sequence alignment; the absolute difference value of the corresponding data points in the two sets is divided by the sum of the larger absolute value in the two sets and a very small constant to obtain the relative error of each data point by using an adaptive relative error calculation method; the arithmetic mean of the relative errors of all data points is calculated to finally obtain the fusion error value.

[0030] In a preferred embodiment, the specific steps of performing the reinforcement learning optimization operation are as follows:

[0031] The state space comprising the environmental parameter vector, the drift matrix and the fusion error value is constructed, a multi-objective reward function comprising the fusion error value and the parameter variation amount is designed, an intelligent agent is trained by using a proximal policy optimization algorithm, the intelligent agent obtains a data sequence of states, actions, rewards and next states by interacting with the environment; the policy gradient is calculated based on the collected data sequence, and the policy network parameters are updated; the drift factor model parameters and the entropy sensitivity factor are continuously adjusted by iterative optimization;

[0032] The optimized parameters are fed back to the drift amount calculation operation of step S1 and the weight distribution operation of step S3 in real time.

[0033] The application also provides an automatic inspection system of an intelligent fire-fighting equipment of an LNG filling station, which specifically comprises:

[0034] An environment perception module, which collects temperature, humidity, salt mist concentration and wind speed data in the filling station in real time, constructs an environmental parameter vector, generates drift factor model parameters through lightweight neural network training, and outputs a drift matrix;

[0035] A time series prediction module, which receives the environmental parameter vector and the drift matrix, and predicts sensor drift trend data in the next 5-10 minutes through a double-channel LSTM network;

[0036] A digital twin module, which loads the drift trend data and simulates sensor monitoring data under harsh environments, dynamically allocates multi-modal data fusion weights by identifying abnormal patterns through unsupervised clustering;

[0037] A closed-loop optimization module, which compares actual alarm events with digital twin simulation results, calculates fusion error values, updates drift factor model parameters and fusion weight strategies through reinforcement learning, and feeds back the updated parameters to the environment perception module.

[0038] The application has the following beneficial effects: a dynamic drift model is constructed by collecting environmental parameters in real time, and the interference of harsh conditions on sensors is accurately quantified; the drift trend is captured in advance by using a time series prediction network, and the device state under extreme scenarios is simulated by combining digital twin technology; abnormal patterns are intelligently identified based on unsupervised clustering, and multi-source data fusion weights are dynamically optimized; finally, the drift model and the fusion strategy are iterated through reinforcement learning in a closed loop, and the system adaptability is continuously improved; the whole process effectively eliminates misjudgment and missed judgment caused by environmental interference, significantly enhances the real-time performance and reliability of fire-fighting early warning, greatly reduces the risk of safety accidents, and provides all-weather, high-precision and self-evolving intelligent protection for LNG filling stations. BRIEF DESCRIPTION OF DRAWINGS

[0039] Figure 1 The application is a method flowchart;

[0040] Figure 2 The application is a system structure block diagram. Detailed Implementation

[0041] The technical solutions of the embodiments of this application will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of this application, and not all embodiments. Based on the embodiments of this application, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of this application.

[0042] In the description of this application, the terms "first" and "second" are used for descriptive purposes only and should not be construed as indicating or implying relative importance or implicitly specifying the number of indicated technical features. Thus, a feature defined as "first" or "second" may explicitly or implicitly include one or more of the stated features. In the description of this application, "multiple" means two or more, unless otherwise explicitly specified.

[0043] In the description of this application, the term "for example" is used to mean "used as an example, illustration, or description." Any embodiment described as "for example" in this application is not necessarily to be construed as being more preferred or advantageous than other embodiments. The following description is provided to enable any person skilled in the art to make and use the invention. Details are set forth in the following description for purposes of explanation. It should be understood that those skilled in the art will recognize that the invention can be made without using these specific details. In other instances, well-known structures and processes will not be described in detail to avoid obscuring the description of the invention with unnecessary detail. Therefore, the invention is not intended to be limited to the embodiments shown, but is consistent with the broadest scope of the principles and features disclosed in this application.

[0044] Example 1

[0045] This embodiment provides, for example Figure 1 The method for automated inspection of intelligent fire-fighting equipment in an LNG refueling station includes the following steps:

[0046] Step S1: Collect environmental parameter data in the LNG refueling station in real time, including temperature, humidity, salt spray concentration, and wind speed. Construct an environmental parameter vector based on the collected environmental parameter data, perform a lightweight neural network training operation, input historical environmental data and sensor calibration deviation records, and output drift factor model parameters. Calculate the drift amount of each sensor in the current environment based on the drift factor model parameters to generate a drift matrix. This step provides accurate input for subsequent drift prediction and dynamic fusion to eliminate the interference of harsh environments (such as high salt spray and low temperature) on fire monitoring sensor data in LNG refueling stations.

[0047] Step S2, input the environmental parameter vector and the drift matrix into the double-channel LSTM network in time sequence, perform time series prediction operation, and output the drift trend data of each sensor in the next 5 to 10 minutes. In order to overcome the dynamic interference of harsh environment on sensor data, the drift trend is predicted based on the drift matrix and the environmental parameter vector generated in step S1 through the double-channel LSTM architecture, providing high-precision input for the digital twin simulation in step S3.

[0048] Step S3, load the drift trend data in the digital twin platform, simulate the sensor monitoring data in harsh environment, perform unsupervised clustering operation to identify abnormal patterns in the simulation data, dynamically allocate fusion weights of multi-modal data according to the abnormal pattern recognition result, based on the drift trend data output in step S2, simulate the sensor state in harsh environment in the digital twin platform, identify abnormal patterns through improved clustering algorithm, and dynamically allocate fusion weights, solve the misjudgment problem caused by fixed weight in traditional method when the environment changes suddenly.

[0049] Step S4, obtain the actual fire alarm event data and the digital twin simulation result data, calculate the fusion error value between them, perform reinforcement learning optimization operation, update the drift factor model parameters and the fusion weight allocation strategy, and complete the closed-loop feedback. The updated drift factor model parameters are fed back to the drift amount calculation operation in step S1 to improve the adaptability in real harsh environment. In this step, the fusion error is calculated by comparing the actual alarm data with the digital twin simulation result, and the reinforcement learning optimization is driven to realize the autonomous iteration of the drift model and the fusion weight, forming a closed-loop feedback.

[0050] In this embodiment, it is specifically pointed out that in step S1, the specific operation of constructing the environmental parameter vector based on the collected environmental parameter data is as follows:

[0051] Synchronously collect real-time environmental parameter data of temperature and humidity sensors, salt fog monitors and anemometers deployed in the gas station tank farm, gas dispensers and pipe corridors. The triggering condition for collection is system startup or periodic patrol task activation, and the collection period is not more than 10 seconds. The read environmental parameter data includes environmental temperature, relative humidity, salt fog concentration and wind speed. The environmental temperature range is -40 degrees Celsius to 60 degrees Celsius, the relative humidity range is 0% to 100%, the salt fog concentration range is 0 milligrams per cubic meter to 500 milligrams per cubic meter, and the wind speed range is 0 meters per second to 60 meters per second. The sliding window mean filtering operation is performed on the environmental parameter data, and the window size is 5 sampling points to suppress transient noise interference. The environmental parameter vector is generated based on the filtered environmental parameter data. The vector is composed of four components in a fixed order: the first component is the measured value of the environmental temperature, the second component is the measured value of the relative humidity, the third component is the measured value of the salt fog concentration, and the fourth component is the measured value of the wind speed. The expression of the environmental parameter vector is:

[0052] ;

[0053] wherein, represents the environmental parameter vector, represents the measured value of ambient temperature, used to monitor the influence of low-temperature environment on sensor sensitivity (such as local low temperature caused by LNG leakage), represents the measured value of relative humidity, used to evaluate the risk of sensor circuit corrosion or optical lens fogging caused by high humidity, represents the measured value of salt mist concentration, used to quantify the interference of salt mist corrosion in coastal areas on the sensor probe (such as the aging of the electrode of an electrochemical sensor), represents the measured value of wind speed, used to judge the gas diffusion speed and sensor reading fluctuation in strong wind environment (such as uneven distribution of combustible gas concentration), represents the current timestamp;

[0054] The specific operation of performing a lightweight neural network training operation is:

[0055] The input historical environmental data is derived from the historical adverse event data set of the gas station, which contains at least three environmental parameter time series data recorded during typhoon and high salt mist season, and the corresponding sensor calibration bias record at that moment;

[0056] The training process independently performs the following processing for each sensor: first, read the current environmental parameter vector, multiply it with the environmental weight vector in the drift factor model parameter and add the bias term, input the operation result into the linear activation function with leakage correction, output the intermediate drift, then multiply the intermediate drift by the sensitivity coefficient of the sensor to generate the final drift , wherein the activation function outputs a proportionally reduced negative value when the input value is negative, and keeps the original value output when the input value is non-negative, the expression is:

[0057] ;

[0058] wherein, represents the sensor number index, represents the sensor The predicted drift at time (unit consistent with the sensor type, such as ppm for combustible gas sensor, ℃ for temperature sensor), quantifies the dynamic interference of the environment on the sensor reading, represents the sensitivity coefficient of the sensor (dimensionless positive real number obtained through factory calibration experiment), representing the reading deviation caused by unit environmental interference, represents the sensor environmental weight vector (four-dimensional trainable parameter vector) for assigning the contribution weight of temperature / humidity / salt fog / wind speed to the drift, represents the environmental parameter vector, represents the sensor bias term (trainable parameter, unit consistent with ) for compensating for sensor baseline drift (such as zero point shift caused by aging), represents a linear activation function with leakage correction, which outputs when the input is , and outputs when the input is (solves the problem of negative gradient disappearance), represents the transpose operator of the environmental weight vector of the sensor

[0059] ; the training target is to minimize the sum of absolute value errors of the predicted drift and the measured drift, while constraining the Euclidean norm of the weight vector, expressed as:

[0060] where, represents the training loss value of the drift factor model, which is used to evaluate the prediction accuracy of the model and control the complexity, represents the predicted drift of the th data, represents the measured drift of the th data (from historical calibration bias records), which is obtained by calibration experiment, represents norm (calculates the sum of absolute errors of predicted values and measured values, resistant to abnormal value interference), represents the regularization coefficient (preset hyperparameter, value range 0.01~0.1, prevents overfitting), represents the norm of the four-dimensional environmental weight vector (Euclidean norm, constrains the weight amplitude, avoids extreme environmental weight distribution); the final output of the drift factor model parameters includes the four-dimensional environmental weight vector corresponding to each sensor and the bias term, where the four-dimensional weight corresponds to the environmental influence coefficient of temperature, humidity, salt fog concentration and wind speed, expressed as:

[0061] ;

[0062] where, represents the temperature influence coefficient, positive weight: temperature rise causes positive bias; negative weight: otherwise, represents the humidity influence coefficient, the weight of optical sensor in high humidity environment is usually negative, This represents the influence coefficient of salt spray concentration, a key parameter in coastal areas. The weight is negative when salt spray corrosion causes a negative drift. This indicates the wind speed influence coefficient; strong winds accelerate heat diffusion, affecting temperature sensor readings.

[0063] The specific operations for generating the drift matrix are as follows:

[0064] The drift of all sensors deployed at the gas station Arranged in numerical order as a drift quantity sequence ( (representing the total number of sensors deployed at the gas station), constructed as a drift matrix in diagonal matrix form. The diagonal elements of the matrix represent the drift values ​​of each sensor, while the off-diagonal elements are set to zero. The expression for the drift matrix is:

[0065] .

[0066] In this embodiment, the specific steps for performing the timing prediction operation in step S2 are as follows:

[0067] From the drift matrix Extract the drift values ​​of all sensors and generate drift vectors in order of sensor number. Simultaneously, the environmental parameter vector Organized into environment tensors based on time series. Set the time window length The time interval is 10 minutes, so that the environment tensor contains a sequence of environment parameter vectors for 10 consecutive minutes, and the drift vector contains a sequence of drift values ​​for 10 consecutive minutes. ;

[0068] Construct a dual-channel LSTM (Long Short-Term Memory) network with two independent processing branches: the first branch is the environment feature processing channel, which uses a gated recurrent unit structure to process the environment tensor. To capture the slow-changing trend of meteorological parameters, the environmental feature processing channel formula is:

[0069] ;

[0070] in, This represents the hidden state of the environmental channel at time j. Indicates a gated loop unit. Let j represent the environment tensor at time j. This represents the hidden state at the previous time step, used to pass on temporal dependencies. j represents the time index, which limits the time window. , This indicates the length of the time window, and its value is set to 10 minutes. The trainable parameters of the gating recurrent unit; the second branch is a drift feature processing channel, which uses a dilated convolution enhanced long short-term memory unit to process the drift vector , which captures the fast-changing transient of sensor drift (such as sudden wind speed), and the drift feature processing channel formula is:

[0071] ;

[0072] Among them, represents the hidden state of the drift channel at time j, represents a dilated convolution long short-term memory unit, which expands the receptive field to 3 times (r = 2) of the traditional, represents the drift vector at time j, represents the hidden state of the last time, which is used to transmit the time sequence dependence, represents the dilated rate, which controls the sparsity of feature extraction (r = 2 is a fixed value), represents the trainable parameters of the dilated convolution long short-term memory unit, represents the current timestamp, j represents the time index, and the time window is limited ; The hidden state outputs of the two branches are fused through a dynamic weight distribution mechanism, in which the weight coefficient of the environment channel hidden state is calculated by a coupling function, and the weight coefficient calculation formula of the environment channel hidden state is:

[0073] ;

[0074] Among them, represents the weight coefficient of the environment channel hidden state, represents the hidden state of the environment channel at time j, represents the hidden state of the drift channel at time j, represents the projection matrix of the environment hidden state, represents the projection matrix of the drift hidden state, represents the attention vector transpose operator (an internal parameter of the dynamic weight distribution mechanism), represents the Sigmoid activation function, the weight coefficient of the drift channel hidden state is 1 minus the weight coefficient of the environment channel hidden state, and the formula of the fused hidden state vector is:

[0075] ;

[0076] Among them, represents the fused hidden state vector, represents the weight coefficient of the drift channel hidden state;

[0077] The specific operation of outputting the drift trend data of each sensor in the next 5 to 10 minutes is:

[0078] The fusion hidden state sequence output by the dynamic weight allocation mechanism is input into a time domain convolutional neural network, and the expression is:

[0079] ;

[0080] wherein, represents the drift trend data (i.e., the drift trend prediction matrix, matrix, represents the total number of sensors deployed at the gas station), used to store the drift of all sensors in the future 10 minutes, represents the time domain convolutional neural network, used to extract the time sequence dependence from the fusion features, represents the network parameters, represents the fusion hidden state sequence (Ht) , represents the length of the time window, which is set to 10 minutes, represents the hidden state dimension, which is set to 64), used to encode the joint features of the environment and the drift, with a dimension of 64, and perform multi-step prediction calculation; the fusion hidden state sequence is generated by coupling and superimposing the environment channel hidden state and the drift channel hidden state through a coupling coefficient , the prediction time span is 5 to 10 minutes, and the output dimension is a two-dimensional matrix of time steps multiplied by the number of sensors; each matrix element represents the drift prediction value of the sensor corresponding to the sensor number at the future time point (in minutes), wherein the sensor numbers are arranged in the order of deployment at the gas station, and the time points are continuously distributed in time sequence; the time domain convolutional neural network includes three dilated convolutional layers, each with a convolution kernel size of 3 and a dilated rate of 1, 2, and 4 in turn; the output layer uses a linear activation function, and the loss function is the smooth L1 norm error of the predicted drift and the actual drift, and the expression of the loss function is:

[0081] ;

[0082] wherein, represents the loss function (the smaller the value, the more accurate the prediction), the error sum of the predicted drift value and the actual value, used to optimize the network parameters of the time domain convolutional neural network , represents the sum of all sensors, represents the sensor number index, represents the sum of each minute in the future 10 minutes, used to accumulate the prediction error of each time step in the future 10 minutes, represents the predicted drift of the sensor in the future minute, represents the predicted drift of the sensor in the future The actual drift amount (obtained through calibration) of each sensor in each minute; after the prediction result is inverse normalized, the drift trend data directly input into the digital twin platform is generated.

[0083] In this embodiment, it is particularly necessary to explain that in step S3, the specific steps of simulating the sensor monitoring data in the harsh environment are as follows:

[0084] In the digital twin platform, a virtual sensor network completely corresponding to the real gas station sensor network is constructed, and the drift trend data from step S2 is loaded The drift trend data is a two-dimensional matrix containing the predicted drift amount of all sensors in the next ten minutes; at the same time, a preset physical field simulation model is called, the physical field simulation model including a computational fluid dynamics model and a multi-physical field coupling equation; the predicted drift amount and the simulated environment tensor at the corresponding time are jointly input into the physical field simulation model, and the basic readings of each virtual sensor in the simulated harsh environment are calculated; on this basis, a Gaussian distribution random noise with a mean of zero and a fixed standard deviation is introduced, and the noise is superimposed on the basic readings to simulate the measurement error of the real sensor, and the expression is as follows:

[0085] ;

[0086] Wherein, represents the simulated reading of the th sensor at the th time step (such as the first to tenth minute in the future) in the digital twin environment, represents a physical model function, representing one or a group of mathematical physical models (such as a computational fluid dynamics (CFD) model, a gas diffusion equation), which accurately simulates the physical influence mechanism of the harsh environment (such as salt spray penetration, low-temperature condensation) on the sensor reading, represents the predicted drift amount of the sensor in the future minute, represents a simulated environment tensor, representing a set of simulated environment parameter vectors in the digital twin environment at the th time step, represents a random noise, representing a random noise (random error) conforming to a Gaussian distribution (normal distribution), with a mean of 0 and a standard deviation of 0.5, used to simulate the inevitable random error of the sensor measurement in the real world, making the simulation data closer to the real situation; finally, a simulated data set ( is output, representing the sensor index range, and this data set contains all the simulated data of the first sensor to the n sensor, a range of time points, i.e. this dataset simulates the data of every minute in the next 10 minutes from the current time), which contains all the simulated monitoring data of all sensors under the influence of different harsh environmental factors with noise in the next 10 minutes;

[0087] The specific steps of identifying abnormal patterns in the simulation data are:

[0088] The sensor monitoring data under harsh environments generated by the digital twin platform Multi-dimensional feature extraction is performed, and the extracted features include the mean, variance, first-order difference fluctuation amplitude of the sensor values, and the energy proportion after principal component analysis of the frequency spectrum; the extracted feature parameters are organized into a feature vector set according to the time window; an improved density peak clustering algorithm is used to process the set, and the local density value of each feature vector is calculated by an exponential kernel function, where the kernel function decay coefficient is associated with the preset cutoff distance, and the local density calculation formula is:

[0089] ;

[0090] wherein, represents the data distribution density around the data point (the local density of the sample ), the greater the value, the more crowded around the point, the more likely it is a normal point, and the smaller the value, the more likely it is an abnormal point, represents the sum of all samples except , when calculating the density, each other point in the dataset is traversed , represents the feature vector of the th sample, represents the feature vector of the th sample (representing another data point for comparison), represents the distance between the two feature vectors, i.e. the Euclidean distance between samples and in the multi-dimensional feature space, which is used to measure their similarity, the closer the distance, the higher the similarity, represents an exponential function that maps the distance to a similarity value that rapidly decays with increasing distance (between 0 and 1), replacing the traditional DPC's simple "counting", which is not sensitive to noise and abnormal values, and has stronger noise resistance, represents the decay coefficient, which is calculated as , so that it is associated with the cutoff distance , achieving adaptive adjustment, represents the cutoff distance, which is a preset distance threshold used to determine whether two points belong to "neighbors", which is the judgment basis for the indicator function , This indicates an indicator function, which is triggered when the condition within the parentheses ( When the condition is met, the function value is 1; when the condition is not met, it is 0. Its function is to limit the range of summation, considering only values ​​falling within the cutoff distance. The contribution of neighboring points within the range to the density; comparing the local density values ​​and relative distances of all feature vectors, feature vectors with local densities below a first threshold and relative distances above a second threshold are selected and identified as anomalous pattern points. The formula for extracting abnormal patterns is:

[0091] ;

[0092] in, Representing relative distance, it appears in subsequent filtering criteria and is defined as a point. The minimum distance to all points with higher density is considered an outlier. A larger value indicates that the point is more distant from the higher-density "mainstream" region and is more likely to be an outlier. Representing data points The density of surrounding data distribution (sample) (local density);

[0093] The filter expression is:

[0094] ;

[0095] in, This represents the density threshold (the median or a certain percentile of the density values ​​of all samples). It appears in subsequent filtering conditions and is used to determine whether the density of a point is low enough. Points below this threshold are considered low-density points (candidate outliers). This represents the distance threshold (taken as the upper quartile or a certain percentile of the relative distances of all samples), used to determine whether the relative distance to a point is large enough. This represents the logical "AND" operator, meaning that the conditions on both sides must be satisfied simultaneously; spatially adjacent anomalous pattern points are merged to form anomalous clusters, which are marked as a set of anomalous patterns that significantly deviate from the normal data distribution. ;

[0096] The specific steps for dynamically allocating multimodal data fusion weights based on the abnormal pattern recognition results are as follows:

[0097] Obtain the set of abnormal patterns identified by unsupervised clustering operations and several preset modality centers. ( Indicates the first A preset modal center (e.g.) Represents the center of the "normal" state. Represents the "leakage" status center. (Represents the "interference" state center). This represents the total number of preset modes, that is, how many typical operating state modes are predefined (e.g., This defines three modes: normal, leakage, and interference. Each mode center represents a typical operating state. For each preset mode, its energy entropy value is calculated. The calculation process is as follows: First, the soft assignment probability of each anomalous mode point in the anomalous mode set belonging to that mode is calculated. This probability value is negatively correlated with the distance from the anomalous mode point to the mode center. Then, based on the soft assignment probability distribution of all anomalous mode points, the energy entropy of that mode is calculated. The higher the energy entropy value, the greater the uncertainty of the anomalous data belonging to that mode. The formula for calculating the energy entropy value is:

[0098] ;

[0099] ;

[0100] in, The energy entropy value of mode l, i.e., the abnormal data set. The degree of disorder belonging to the l-th preset mode (such as "normal" or "leaking") is determined by the energy entropy value. A higher energy entropy value indicates greater uncertainty, meaning that the current abnormal data is less consistent with this mode. Represents a set of abnormal patterns. Indicates the first One abnormal pattern point, This represents the l-th preset modal center (representing a predefined typical operating state, such as the center point of "normal state" or the center point of "minor leakage")). This represents the soft allocation coefficient, which controls the sensitivity of the probability to distance. Its value is usually taken as the reciprocal of the average distance. The greater the distance, the higher the probability. The smaller, Represents the Euclidean distance, i.e., outlier. To the modal center Spatial distance; the greater the distance, the less similar they are. Represents the soft assignment probability, i.e., the th The probability that an anomalous pattern point belongs to the l-th mode is between 0 and 1. Unlike the either-or "hard" assignment, its value is obtained by normalizing the distance using the softmax function. The loop index (dummy variable) representing the summation calculation iterates from 1 to... All modes; then, based on the energy entropy value of each mode, its fusion weight is dynamically calculated. The calculation principle is that the higher the energy entropy value of a mode, the lower its fusion weight is assigned. The formula for calculating the fusion weight is:

[0101] ;

[0102] wherein, represents the fusion weight of the lth modality, represents the adjustment parameter (entropy sensitivity factor), which takes a constant greater than zero, used to adjust the energy entropy The influence of the weight on the entropy, The greater the value, the more sensitive the weight to the change of entropy, represents the unnormalized value of the weight, which is exponentially transformed and taken negative for the energy entropy, and the entropy The higher the value, the smaller the value, which conforms to the core principle of "the higher the uncertainty, the lower the weight", represents the normalization denominator, which adds up the unnormalized weights of all modalities, the purpose is to ensure that the sum of all weights is 1, represents the loop index (dummy variable) of the summation calculation, which traverses all modes from 1 to The purpose is to add up the values of all modalities to achieve normalization, ensuring that the sum of all weights is 1; finally, the fusion weights of all modalities are normalized, and a normalized weight vector (W) is output for multi-modal data fusion.

[0103] In this embodiment, it is specifically necessary to explain that in step S4, the specific steps of calculating the fusion error value between the two are as follows:

[0104] Obtain the actual fire alarm event data set (from field sensors) and the digital twin simulation result data set , perform time series alignment processing on the data points in the two sets to ensure that the compared data points are on the same time reference; adopt an adaptive relative error calculation method, divide the absolute difference value of the corresponding data points in the two sets by the sum of the larger absolute value in the two and a very small constant, to obtain the relative error of each data point; take the arithmetic mean of the relative errors of all data points, and finally obtain the fusion error value; this calculation method can eliminate the influence of different sensor dimension differences, and is robust to extreme measurement values. The calculation formula of the fusion error value is:

[0105] ;

[0106] wherein, represents the fusion error value, which takes a value in the range [0, 1], indicating the average difference between the entire simulation data set and the real data set, represents the total number of sensors deployed at the gas station, represents the total number of time steps (such as ),​ represents the actual measurement value of the i-th sensor at the j-th time point, represents the analog data value of the i-th sensor at the j-th time point, represents a smoothing factor, whose value is set to 1 / (1+exp(-1)) to prevent the denominator from being zero and ensure that the formula is still meaningful when the measurement value is zero, represents a maximum value function, represents an absolute value function.

[0107] The specific steps of performing reinforcement learning optimization operation are as follows:

[0108] A state space containing environmental parameter vectors, drift matrices and fusion error values is constructed, and the expression of the state space is:

[0109] ;

[0110] wherein, represents the state vector at time t, represents the environmental parameter vector at time t (a vector containing real-time environmental data such as temperature, humidity, salt fog concentration and wind speed), represents the drift matrix at time t, represents the fusion error value, a multi-objective reward function containing the fusion error value and the parameter change amount is designed, and the expression of the multi-objective reward function is: ;

[0111] ;

[0112] wherein, represents the multi-objective reward function value, represents the reward weight coefficient (β), represents the adjustment amount of the drift factor model parameter (including a four-dimensional environmental weight vector corresponding to each sensor and a bias term), represents the total number of sensors deployed at the gas station, represents the adjustment parameter (entropy sensitivity factor), a proximal policy optimization algorithm is used to train the agent, and the agent obtains the data sequence of state, action, reward and next state by interacting with the environment; based on the collected data sequence, the policy gradient is calculated, and the policy network parameter is updated, and the expression is:

[0113] ​​​​​​​​​​​;

[0114] wherein, denotes the policy network parameter, denotes the learning rate, which is set to , exponentially decays with the training index, denotes the policy gradient objective function (including the advantage function and the policy clipping); by iteratively optimizing the drift factor model parameter and the entropy sensitivity factor, the reward function value is maximized, so as to realize the minimization of the fusion error and the maximization of the system stability, and the adjustment expression of the drift factor model parameter is:

[0115] ;

[0116] the adjustment expression of the entropy sensitivity factor is:

[0117] ;

[0118] The optimized parameters are fed back to the drift calculation operation of step S1 and the weight distribution operation of step S3 in real time.

[0119] Embodiment 2

[0120] The embodiment provides an automatic inspection system for an LNG filling station intelligent fire-fighting equipment as shown in Figure 2 , and specifically comprises:

[0121] An environment perception module, which collects temperature, humidity, salt mist concentration and wind speed data in the filling station in real time, constructs an environment parameter vector, generates drift factor model parameters through lightweight neural network training, and outputs a drift matrix;

[0122] A time series prediction module, which receives the environment parameter vector and the drift matrix, and predicts sensor drift trend data for the next 5-10 minutes through a double-channel LSTM network;

[0123] A digital twin module, which loads the drift trend data and simulates sensor monitoring data under harsh environments, identifies abnormal patterns through unsupervised clustering, and dynamically allocates multi-modal data fusion weights;

[0124] A closed-loop optimization module, which compares the actual alarm events with the digital twin simulation results, calculates the fusion error value, updates the drift factor model parameters and the fusion weight strategy through reinforcement learning, and feeds back the updated parameters to the environment perception module.

[0125] It should be noted that in the above embodiments, the description of each embodiment has its own emphasis, and the parts not described in detail in a certain embodiment can be referred to the related description of other embodiments.

[0126] Those skilled in the art will appreciate that embodiments of the present application can be devised for a variety of applications. It is intended that the present application be limited only by the scope of the appended claims, and it is intended that various modifications and alterations made by those skilled in the art be considered as within the scope of the present application. The embodiments of the present application will be described with reference to the attached drawings identified below.

[0127] The present application is described in reference to the drawings using a flowchart and / or a block diagram of the method, apparatus (system) and computer program product according to embodiments of the application. It will be understood that each block of the flowchart and / or block diagram, and combinations of blocks in the flowchart and / or block diagram, can be implemented by computer program instructions. These computer program instructions can be provided to a processor of a general purpose computer, special purpose computer, embedded computer, or other programmable data processing apparatus to produce a machine, such that the instructions, which execute via the processor of the computer or other programmable data processing apparatus, create means for implementing the functions specified in the flowchart and / or block diagram block or blocks. Figure 1 one or more functions specified in the flowchart and / or block diagram block or blocks. Figure 1 one or more functions specified in the flowchart and / or block diagram block or blocks.

[0128] These computer program instructions can also be stored in a computer- readable memory that can direct a computer or other programmable data processing apparatus to function in a particular manner, such that the instructions stored in the computer-readable memory produce an article of manufacture including instructions which implement the function specified in the flowchart and / or block diagram block or blocks. Figure 1 one or more functions specified in the flowchart and / or block diagram block or blocks. Figure 1 one or more functions specified in the flowchart and / or block diagram block or blocks.

[0129] These computer program instructions can also be loaded onto a computer or other programmable data processing apparatus to cause a series of operational steps to be performed on the computer or other programmable apparatus to produce a computer-implemented process such that the instructions which execute on the computer or other programmable apparatus provide steps for implementing the functions specified in the flowchart and / or block diagram block or blocks. Figure 1 one or more functions specified in the flowchart and / or block diagram block or blocks. Figure 1 one or more functions specified in the flowchart and / or block diagram block or blocks.

[0130] While the preferred embodiments of the application have been described, additional variations and modifications can be made to the embodiments by those skilled in the art once they learn of the basic inventive concepts. Therefore, the appended claims are intended to cover all such modifications and variations as fall within the scope of the present application.

[0131] Obviously, many modifications and variations of the present application are possible in light of the above teachings. It is, therefore, to be understood that within the scope of the appended claims and their equivalents, the application can be practiced otherwise than as specifically described.

Claims

1. An automated inspection method for intelligent fire-fighting equipment in an LNG refueling station, characterized in that, Specifically, the following steps are included: Step S1: Collect environmental parameter data in the gas station in real time, including temperature, humidity, salt spray concentration and wind speed. Construct an environmental parameter vector based on the collected environmental parameter data, perform a lightweight neural network training operation, input historical environmental data and sensor calibration deviation records, and output drift factor model parameters. Calculate the drift amount of each sensor in the current environment according to the drift factor model parameters and generate a drift matrix. Step S2: Input the environmental parameter vector and the drift matrix into a dual-channel LSTM network in time series, perform time series prediction operation, and output the drift trend data of each sensor in the next 5 to 10 minutes; Step S3: Load the drift trend data into the digital twin platform to simulate sensor monitoring data under harsh conditions; perform unsupervised clustering to identify abnormal patterns in the simulated data, and dynamically allocate the fusion weights of multimodal data based on the abnormal pattern identification results; The specific steps for dynamically allocating multimodal data fusion weights based on the abnormal pattern recognition results are as follows: Obtain the set of abnormal patterns identified by unsupervised clustering operations and several preset mode centers, each mode center representing a typical operating state mode; for each preset mode, calculate its energy entropy value. The calculation process of the energy entropy value is as follows: first, calculate the soft assignment probability of each abnormal pattern point in the abnormal pattern set belonging to the mode. This probability value is negatively correlated with the distance from the abnormal pattern point to the mode center. Then, calculate the energy entropy of the mode based on the soft assignment probability distribution of all abnormal pattern points. Subsequently, based on the energy entropy value of each modality, its fusion weight is dynamically calculated. The calculation principle is that the higher the energy entropy value of a modality, the lower the fusion weight it is assigned. Finally, the fusion weights of all modalities are normalized, and a normalized weight vector is output for multimodal data fusion. Step S4: Obtain actual fire alarm event data and digital twin simulation result data, calculate the fusion error value between the two; perform reinforcement learning optimization operation, update the drift factor model parameters and fusion weight allocation strategy, and complete closed-loop feedback.

2. The automated inspection method for intelligent fire-fighting equipment in an LNG refueling station according to claim 1, characterized in that: In step S1, the specific operation of constructing the environmental parameter vector based on the collected environmental parameter data is as follows: Real-time environmental parameter data from temperature and humidity sensors, salt spray monitors, and anemometers deployed in the gas station's storage tank area, gas dispensers, and pipe corridors are collected synchronously. The data acquisition is triggered by the activation of a periodic inspection task, with a collection cycle not exceeding 10 seconds. The environmental parameter data includes ambient temperature, relative humidity, salt spray concentration, and wind speed. The ambient temperature range is -40°C to 60°C, the relative humidity range is 0% to 100%, the salt spray concentration range is 0 mg / m³ to 500 mg / m³, and the wind speed range is 0 m / s to 60 m / s. A sliding window mean filtering operation is performed on the environmental parameter data, with a window size of 5 sampling points. An environmental parameter vector is generated based on the filtered environmental parameter data. This vector consists of four components in a fixed order: the first component is the measured ambient temperature, the second component is the measured relative humidity, the third component is the measured salt spray concentration, and the fourth component is the measured wind speed. The specific steps for performing lightweight neural network training are as follows: The input historical environmental data comes from the historical adverse event dataset of the gas station. This dataset contains time-series data of at least three environmental parameters recorded during typhoons and high salt spray seasons, as well as sensor calibration deviation records at the corresponding times. The training process involves performing the following independent processing on each sensor: First, the current environmental parameter vector is read, and it is multiplied by the environmental weight vector in the drift factor model parameters, with a bias term added. The result is input into a linear activation function with leakage correction, and the intermediate drift is output. Then, the intermediate drift is multiplied by the sensitivity coefficient of the sensor to generate the final drift. The activation function outputs a proportionally reduced negative value when the input value is negative, and maintains the original value when the input value is non-negative. The training objective is to minimize the sum of the absolute errors between the predicted drift and the measured drift, while constraining the Euclidean norm of the weight vector. The final output drift factor model parameters include a four-dimensional environmental weight vector and bias term for each sensor, where the four-dimensional weights correspond to the environmental influence coefficients of temperature, humidity, salt spray concentration and wind speed, respectively.

3. The automated inspection method for intelligent fire-fighting equipment in an LNG refueling station according to claim 2, characterized in that: The specific operation for generating the drift matrix is ​​as follows: The drift values ​​of all sensors deployed at the gas station are arranged in numerical order to form a drift value sequence, which is then constructed as a diagonal matrix. The diagonal elements of the matrix represent the drift values ​​of each sensor, while the off-diagonal elements are set to zero.

4. The automated inspection method for intelligent fire-fighting equipment in an LNG refueling station according to claim 3, characterized in that: In step S2, the specific steps for performing the time series prediction operation are as follows: Extract the drift values ​​of all sensors from the drift matrix and generate drift vectors in order of sensor number. At the same time, organize the environmental parameter vectors into an environmental tensor according to the time series. Set the time window length to 10 minutes so that the environmental tensor contains a 10-minute sequence of environmental parameter vectors and the drift vectors contain a 10-minute sequence of drift values. Two independent processing branches are constructed in a dual-channel LSTM network: the first branch is the environmental feature processing channel, which uses a gated recurrent unit structure to process the environmental tensor and capture the slow-changing trend of meteorological parameters; the second branch is the drift feature processing channel, which uses dilated convolution-enhanced long short-term memory units to process the drift vector and capture the fast-changing transients of sensor drift; the hidden state outputs of the two branches are fused through a dynamic weight allocation mechanism, where the weight coefficient of the hidden state of the environmental channel is calculated by the coupling function, and the weight coefficient of the hidden state of the drift channel is 1 minus the weight coefficient of the hidden state of the environmental channel.

5. The automated inspection method for intelligent fire-fighting equipment in an LNG refueling station according to claim 4, characterized in that: The specific steps for outputting the drift trend data of each sensor over the next 5 to 10 minutes are as follows: The fused hidden state sequence output by the dynamic weight allocation mechanism is input into a temporal convolutional neural network to perform multi-step prediction calculations. The fused hidden state sequence is generated by weighted superposition of the environmental channel hidden state and the drift channel hidden state through a coupling coefficient. The prediction time span is 5 to 10 minutes, and the output dimension is a two-dimensional matrix with the time step multiplied by the number of sensors. Each matrix element represents the predicted drift value of the sensor corresponding to the sensor number at a future time point. The sensor numbers are arranged according to the deployment order of the gas station, and the time points are continuously distributed in a time series. The temporal convolutional neural network contains three dilated convolutional layers, each with a kernel size of 3 and dilation rates of 1, 2, and 4 respectively. The output layer uses a linear activation function, and the loss function is the smoothed L1 norm error between the predicted drift and the actual drift. After inverse normalization, the prediction results generate drift trend data that can be directly input into the digital twin platform.

6. The automated inspection method for intelligent fire-fighting equipment in an LNG refueling station according to claim 5, characterized in that: In step S3, the specific steps for simulating and generating sensor monitoring data under harsh environments are as follows: A virtual sensor network, completely corresponding to the sensor network of a real gas station, is constructed in a digital twin platform. Drift trend data from step S2 is loaded, which is a two-dimensional matrix containing the predicted drift amounts of all sensors within the next ten minutes. Simultaneously, a pre-set physical field simulation model is invoked, which includes a computational fluid dynamics model and multiphysics coupling equations. The predicted drift amounts and the simulated environment tensor at the corresponding time are input into the physical field simulation model to calculate the basic readings of each virtual sensor under simulated harsh conditions. Based on this, a Gaussian distributed random noise with zero mean and fixed standard deviation is introduced and superimposed on the basic readings to simulate the measurement error of the real sensors. Finally, a simulated dataset is output, which completely contains the simulated monitoring data with noise from all sensors under different harsh environmental factors within the next ten minutes. The specific steps for identifying abnormal patterns in the simulated data are as follows: Multi-dimensional feature extraction is performed on sensor monitoring data generated by a digital twin platform under harsh environments. The extracted features include the mean, variance, first-order difference fluctuation amplitude, and energy proportion after principal component analysis of the sensor values. The extracted feature parameters are organized into a set of feature vectors according to time windows. An improved density peak clustering algorithm is used to process this set. The local density value of each feature vector is calculated using an exponential kernel function, where the kernel function decay coefficient is related to a preset cutoff distance. The local density values ​​and relative distances of all feature vectors are compared. Feature vectors with local densities below a first threshold and relative distances above a second threshold are selected and identified as anomalous pattern points. Spatially adjacent anomalous pattern points are merged to form anomalous clusters, which are marked as a set of anomalous patterns that deviate significantly from the normal data distribution.

7. An automated inspection method for intelligent fire-fighting equipment in an LNG refueling station according to claim 6, characterized in that: In step S4, the specific steps for calculating the fusion error value between the two are as follows: The system acquires a set of actual fire alarm event data and a set of digital twin simulation results data, and performs time-series alignment processing on the data points in the two sets. An adaptive relative error calculation method is used to divide the absolute difference of corresponding data points in the two sets by the sum of the larger of the two absolute values ​​and a very small constant to obtain the relative error of each data point. The arithmetic mean of the relative errors of all data points is then calculated to obtain the final fusion error value.

8. An automated inspection method for intelligent fire-fighting equipment in an LNG refueling station according to claim 7, characterized in that: The specific steps for performing reinforcement learning optimization operations are as follows: A state space containing environmental parameter vectors, drift matrix, and fusion error values ​​is constructed. A multi-objective reward function containing fusion error values ​​and parameter changes is designed. The agent is trained using a proximal policy optimization algorithm. The agent obtains data sequences of state, action, reward, and next state by interacting with the environment. The policy gradient is calculated based on the collected data sequences, and the policy network parameters are updated. The drift factor model parameters and entropy sensitivity factor are continuously adjusted through iterative optimization. The optimized parameters are fed back in real time to the drift calculation operation in step S1 and the weight allocation operation in step S3.

9. An automated inspection system for intelligent fire-fighting equipment at an LNG refueling station is applied to an automated inspection method for intelligent fire-fighting equipment at an LNG refueling station as described in any one of claims 1-8, characterized in that: Specifically, it includes: The environmental perception module collects real-time data on temperature, humidity, salt spray concentration, and wind speed within the gas station, constructs an environmental parameter vector, generates drift factor model parameters through lightweight neural network training, and outputs a drift matrix. The time-series prediction module receives environmental parameter vectors and drift matrices, and uses a dual-channel LSTM network to predict sensor drift trend data for the next 5-10 minutes. The digital twin module loads drift trend data and simulates sensor monitoring data under harsh environments. It identifies abnormal patterns through unsupervised clustering and dynamically allocates multimodal data fusion weights. The closed-loop optimization module compares the actual alarm events with the digital twin simulation results, calculates the fusion error value, updates the drift factor model parameters and fusion weight strategy through reinforcement learning, and feeds the updated parameters back to the environmental perception module.

Images