Concentration measurement and full-chain risk assessment method for combustion and explosion of energy supply composite area

Abstract

The invention discloses a concentration measurement and full-chain risk assessment method for combustion and explosion of an energy supply composite area. The invention provides a risk assessment and concentration measurement method and system for combustion and explosion of an energy supply composite area, and the method comprises the steps: a multi-dimensional data collection unit collects basic information needed in the system, and provides basic information for a subsequent data fusion and modeling unit and a combustion and explosion risk analysis unit; the data fusion and modeling unit performs data preprocessing on the obtained multi-dimensional data information, imports the data of each monitoring point into a geographic information system, and links the data of the monitoring points into an actual urban map layer; the burning and explosion risk analysis unit extracts burning and explosion characteristic information of the observation points based on a dangerous gas burning and explosion characteristic comprehensive database, and calculates a burning and explosion risk level through a burning and explosion risk analysis method; and carrying out comprehensive risk assessment. The technical problem that the timeliness and effectiveness of risk management and control are low due to the fact that gas leakage concentration detection is single, fire blast risk assessment is isolated and real-time monitoring data response is lagged is solved.

Description

Technical Field

[0001] This invention relates to the field of gas and oil safety monitoring and risk assessment, specifically to a method for measuring the concentration of combustion and explosion in energy-supply complex areas and for full-chain risk assessment. Background Technology

[0002] Due to overpopulation density and limited space utilization, various municipal pipelines in cities are inevitably laid out densely in close proximity or overlapping, posing significant challenges and safety hazards to urban underground spaces. In areas where cable tunnels, gas pipelines, and gas stations are intertwined, the risks further overlap. Gas and oil leaks are unavoidable, and these leaks are easily dispersed and flammable / explosive when exposed to an ignition source. Cables in cable tunnels may generate electric arcs or sparks due to other faults or aging with increasing service life. The presence of electric sparks further exacerbates the risk of ignition and explosion of gas / oil mixtures. To mitigate this potential risk and improve emergency response capabilities, rapid measurement of gas and oil mixtures and comprehensive risk assessments are crucial.

[0003] A search revealed Chinese Patent No. CN119918929A, which discloses a risk assessment method, system, and terminal for gas pipelines. The method includes: acquiring sample monitoring data from multiple gas monitoring devices attached to the gas pipeline; establishing a basic risk model for the gas pipeline based on the sample monitoring data; allocating indicator weights to the basic risk model using weighted analysis to obtain a target basic risk model; training the target basic risk model using the sample monitoring data to obtain a risk assessment model; acquiring current sample monitoring data; inputting the current sample monitoring data into the risk assessment model; and outputting the gas risk assessment result for the gas pipeline. This invention establishes a risk model using monitoring data, performs weighted analysis and training, and ultimately assesses the risk of the gas pipeline. By establishing a risk assessment model, potential risks in gas pipelines can be effectively identified and warned of, improving safety management and reducing the probability of accidents.

[0004] A search revealed that Chinese Patent No. CN108846521A discloses a gas safety monitoring system based on cloud-based IoT analysis. This system includes: collecting natural gas concentration data, wind speed data, and wind direction data from each natural gas concentration sensor at each monitoring point at each moment; obtaining the gas diffusion intensity at each monitoring point at each moment based on the differences in different natural gas concentration data; obtaining the degree of wind influence at each monitoring point at each moment based on the relationship between wind speed and leakage direction, thereby obtaining the gas leakage level at each monitoring point; combining the extreme differences in natural gas concentration data to obtain the leakage degree value at each monitoring point; and finally obtaining gas risk monitoring indicators for pipelines to classify gas safety levels.

[0005] Existing technologies for measuring and assessing gas and oil leaks only address one aspect: measuring gas or oil leak concentrations or assessing explosion risks. They cannot provide timely and comprehensive risk assessments after gas leak concentrations are obtained through monitoring. Since gas leak concentrations change in real time, delayed risk assessments render the gas leak concentration data ineffective and untimely, thus compromising the reliability and efficiency of the risk assessment results.

[0006] Existing technologies for gas and oil safety monitoring and risk assessment have significant limitations: on the one hand, these technologies primarily focus on single gas leak concentration detection or independent explosion risk assessment, failing to achieve organic integration of the two; on the other hand, due to the dynamic nature of gas leak concentrations, outdated risk assessment methods struggle to respond promptly to real-time monitoring data. Therefore, the timeliness value of leak concentration data cannot be fully realized, and the current fragmented approach to measurement and assessment restricts the timeliness and effectiveness of risk management.

[0007] The existing invention patent application document with publication number CN111612301A, entitled "A Method and Device for Risk Assessment of Leakage in Buried Gas Pipelines Based on Weight Self-Adjustment," describes a method that includes: acquiring impact characteristic data of leakage risk in urban buried gas pipelines and constructing a feature set; preprocessing the impact characteristic data to obtain impact characteristics; determining the risk weight of each impact characteristic based on the historical event distribution of each impact characteristic; establishing a risk assessment model for leakage in urban buried gas pipelines based on each impact characteristic and its corresponding weight; calculating the risk value of leakage risk based on the risk assessment model; classifying the risk value using a pre-defined classification method; and drawing a thematic map of pipeline status.

[0008] Existing technologies often utilize gas concentration information based on data from a single moment, lacking historical data. Since gas concentrations change continuously over time, the absence of historical information impacts the accuracy and real-time nature of risk assessments. Furthermore, when the acquired gas concentration information is time-series data, traditional expert-weighted labeling methods are insufficient. Therefore, directly inputting this data into a hybrid network will also prevent the network from effectively learning from the data.

[0009] The existing invention patent application document CN115713095A, entitled "A Method and System for Anomaly Detection of Natural Gas Pipelines Based on Hybrid Deep Neural Networks," describes a method that includes: normalizing the feature data of the natural gas pipeline to obtain input feature values; constructing a stacked sparse denoising autoencoder deep neural network model as a first hybrid deep neural network based on the input feature values; constructing a cost function based on the input feature values, and performing unsupervised feature learning on the first hybrid deep neural network using the cost function to obtain a second hybrid deep neural network; adding a supervised classifier to the second hybrid deep neural network to obtain a third hybrid deep neural network; inputting the input feature values ​​into the third hybrid deep neural network to obtain output feature values; calculating the maximum probability value of the output feature values ​​using the supervised classifier; and reducing the difference between the maximum probability value of the output feature values ​​and the label.

[0010] Traditional hybrid neural networks feed all input features into a single network, resulting in the mixing of unrelated features. When the feature dimensionality is too high, this setup prevents the neural network from effectively learning the differences between categories and the correlations within categories, thus reducing prediction accuracy. Furthermore, traditional hybrid neural networks are structurally complex to learn a large number of input features simultaneously. This complex network structure reduces computational efficiency and prevents the use of adaptive optimization algorithms to adjust internal network parameters, significantly increasing the time cost of manual parameter tuning.

[0011] In summary, existing technologies suffer from several technical problems, including limited gas leak concentration detection, isolated combustion and explosion risk assessment, delayed real-time monitoring data response, and consequently, low timeliness and effectiveness of risk management. Summary of the Invention

[0012] The technical problem to be solved by this invention is: how to solve the technical problems of single gas leakage concentration detection, isolated combustion and explosion risk assessment, and delayed response of real-time monitoring data in the prior art, which leads to low timeliness and effectiveness of risk control.

[0013] This invention solves the above-mentioned technical problems by employing the following technical solution: a method for risk assessment and concentration measurement of combustion and explosion in energy-supply composite areas, comprising: S1. Sensors are deployed using a multi-dimensional data acquisition unit to acquire multi-dimensional data information in real time; S2. Concentration spatial distribution modeling is performed through data fusion and modeling units; multidimensional data information is preprocessed, and multidimensional data information from each monitoring point is imported into the geographic information system, and the monitoring point data is linked to the actual urban map layer; spatial expansion modeling is performed based on the existing multidimensional data information from the monitoring points. S3. Conduct combustion and explosion risk analysis based on combustion and explosion risk analysis units and concentration spatial distribution; use combustion and explosion risk analysis units based on a comprehensive database of hazardous gas combustion and explosion characteristics to extract combustion and explosion characteristic information of observation points, and calculate the combustion and explosion risk level through combustion and explosion risk analysis methods; S4. Construct a comprehensive risk prediction model and conduct a comprehensive risk assessment of the complex area based on the risk level of combustion and explosion.

[0014] The gas and oil concentration measurement method and comprehensive risk assessment process proposed in this invention can fully utilize real-time gas and oil concentration monitoring data to conduct immediate comprehensive explosion risk assessments in areas where cables, gas pipelines, and gas stations are laid together. This provides guidance for mitigating potential explosion risks and proactively managing those risks.

[0015] In a more specific technical solution, the multidimensional data information in S1 includes: gas concentration information, weather condition information, and geographical location information.

[0016] In a more specific technical solution, in S2, the data preprocessing module is used to remove outliers using 3σ rules and to denoise the data using wavelet transform, thereby cleaning, completing, and removing outliers from the upper-collected data. Using a geographic information fusion module, three types of information from each monitoring point are bound together; Using the spatial modeling module for concentration distribution, a high-density hazardous gas concentration distribution layer is formed through spatial extension prediction and spatial autocorrelation analysis; Spatial autocorrelation analysis was performed on the high-density hazardous gas distribution layer to calculate the global Moran index and examine whether the concentration of hazardous gases in geographic space exhibits clustering or dispersion characteristics. Spatial expansion prediction is performed; spatial weighted sampling is performed using Bootstrap sampling to cluster spatial regions; spatial random forest expansion algorithm is used to realize spatial modeling and establish a high-density hazardous gas distribution layer based on spatial expansion.

[0017] In a more specific technical solution, in S3, the combustion and explosion characteristic information extraction module is used to extract combustion and explosion characteristic information based on the constructed combustion and explosion characteristic database and gas concentration information and weather condition information. Using the explosion risk level calculation module, with explosion characteristic information as feature input, an explosion risk analysis based on fuzzy comprehensive evaluation method is performed to determine the weight of each explosion characteristic information indicator. The comprehensive score of the explosion risk is calculated by various importance indices. The explosion risk level is divided according to the range of the comprehensive score. A calculation model for the explosion risk level is constructed to calculate the explosion risk level of the observation point.

[0018] In a more specific technical solution, S4 involves inputting and integrating information on the hazard factors of potential combustion and explosion risks, and updating the comprehensive risk assessment index database. Using the comprehensive risk assessment implementation module, with the comprehensive risk assessment index database as input, a comprehensive risk prediction model is constructed based on a deep neural network to optimize the explosion risk level and obtain the comprehensive risk assessment level. A comprehensive risk label was constructed; the Delphi method was used to assign weights to the explosion risk assessment indicators of urban cable channels, gas stations and gas pipeline composite laying areas, and the explosion risk assessment indicators were weighted according to the weights to obtain a comprehensive risk score, and the risk level was divided according to the potential explosion risk status. Data input and preprocessing are performed, a comprehensive risk assessment index database is stored, outlier handling is performed on the data, and the input data is normalized using Min-Max. A comprehensive risk prediction model is constructed; using the weighting rules of the initial comprehensive risk assessment indicators, a comprehensive risk assessment prediction model based on artificial intelligence algorithms is introduced to simulate expert weighting. Based on the hybrid deep neural network HDNN, a comprehensive risk prediction model is constructed, and different types of risk assessment indicators are input into the differential applicability hidden layer. The comprehensive risk prediction model uses the cross-entropy loss function to calculate the distance between the predicted probability distribution and the true probability distribution. The non-dominated sorting genetic algorithm II is adopted. Through fast non-dominated sorting, elitist strategy and crowding comparison mechanism, Pareto optimal solution set is efficiently solved and the number of intermediate layers and neurons in the DNN model in hidden layer 1 and hidden layer 2 is automatically optimized. Randomly generate an initial population P0, and perform a fast non-dominated sort on the current population Pt. i rank Individuals in a population are stratified according to their dominance relationships.

[0019] This invention uses expert opinions to subtly label the risk level of the data and constructs a separate comparative learning model for this data to enhance its feature extraction capabilities.

[0020] This invention classifies risk factors according to data properties and designs optimal models for different categories, making the models more adaptable, traceable in terms of contribution paths of each group, and improving the interpretability of the models. Traditional hybrid neural networks are often structurally complex in order to learn a large number of features at the same time. In the network of this invention, the complexity of the network is simplified by learning feature groups separately through multiple small models, which improves the computational efficiency.

[0021] In a more specific technical solution, the loss is calculated using the cross-entropy loss function:

[0022] The following logic is used to improve the contrast loss based on the difference between positive and negative samples:

[0023] In the formula, For sample i, it is the set of other samples of the same type; The set of all samples.

[0024] The joint loss function based on contrastive learning can be expressed using the following logic: .

[0025] In a more specific technical solution, an objective function is defined; whereby, the accuracy is maximized using the following logic: ; Using the following logic, define a model with minimized complexity: ; in, Number of neurons in each layer The number of fully connected layers. The adjustment coefficient is used to minimize the training time. .

[0026] In a more specific technical solution, the following logic is used to calculate the congestion distance. i d :

[0027] When selecting an operation, prioritize... i rank Smaller individuals, if i rank If they are the same, then select. i d Larger individuals are selected. Finally, the parent and offspring populations are merged, and the best N individuals are selected to be retained for the next generation to avoid losing excellent individuals.

[0028] In a more specific technical solution, an adaptive penalty function is constructed to constrain the objective function:

[0029] In the formula, As constraints, This is the penalty coefficient.

[0030] The hybrid neural network of this invention involves a large number of DNN layers with varied structures, and uses a genetic algorithm to automatically optimize the number of DNN layers and neurons.

[0031] In a more specific technical solution, the risk assessment and concentration measurement system for combustion and explosion in the energy supply complex area includes: A multidimensional data acquisition unit is used to deploy sensors to acquire multidimensional data information in real time; The data fusion and modeling unit is used to model the spatial distribution of concentration; it preprocesses multidimensional data information, imports multidimensional data information from each monitoring point into the geographic information system, and links the monitoring point data to the actual urban map layer; it performs spatial expansion modeling on the existing multidimensional data information from monitoring points, and the data fusion and modeling unit is connected to the multidimensional data acquisition unit. The combustion and explosion risk analysis unit is used to perform combustion and explosion risk analysis in combination with the spatial distribution of concentration. Based on the comprehensive database of the combustion and explosion characteristics of hazardous gases, the combustion and explosion risk analysis unit extracts the combustion and explosion characteristic information of the observation points and calculates the combustion and explosion risk level through the combustion and explosion risk analysis method. The combustion and explosion risk analysis unit is connected to the data fusion and modeling unit. The comprehensive risk assessment unit is used to construct a comprehensive risk prediction model and conduct a comprehensive risk assessment of the complex area based on the level of fire and explosion risk. The comprehensive risk assessment unit is connected to the fire and explosion risk analysis unit.

[0032] The present invention has the following advantages over the prior art: The gas and oil concentration measurement method and comprehensive risk assessment process proposed in this invention can fully utilize real-time gas and oil concentration monitoring data to conduct immediate comprehensive explosion risk assessments in areas where cables, gas pipelines, and gas stations are laid together. This provides guidance for mitigating potential explosion risks and proactively managing those risks.

[0033] This invention uses expert opinions to subtly label the risk level of the data and constructs a separate comparative learning model for this data to enhance its feature extraction capabilities.

[0034] This invention classifies risk factors according to data properties and designs optimal models for different categories, making the models more adaptable, traceable in terms of contribution paths of each group, and improving the interpretability of the models. Traditional hybrid neural networks are often structurally complex in order to learn a large number of features at the same time. In the network of this invention, the complexity of the network is simplified by learning feature groups separately through multiple small models, which improves the computational efficiency.

[0035] The hybrid neural network of this invention involves a large number of DNN layers with varied structures, and uses a genetic algorithm to automatically optimize the number of DNN layers and neurons.

[0036] This invention establishes a full-chain risk assessment process—from rapid concentration measurement to spatial distribution modeling to real-time risk assessment—improving the timeliness and effectiveness of risk management. Existing risk assessment technologies using hybrid neural networks suffer from low feature extraction efficiency and complex network structures, affecting the accuracy of prediction results and computational speed. This patent improves the accuracy and computational efficiency of risk prediction by classifying input features according to their properties and inputting them into corresponding network layers. It employs optimization algorithms to self-adjust the internal parameters of the network layers, thereby optimizing the matching mode between input features and the network layers and adaptively optimizing the network's internal parameters.

[0037] This invention solves the technical problems existing in the prior art, such as single gas leakage concentration detection, isolated combustion and explosion risk assessment, delayed response of real-time monitoring data, and low timeliness and effectiveness of risk management. Attached Figure Description

[0038] Figure 1 This is a schematic diagram of the basic steps of the risk assessment and concentration measurement method for combustion and explosion in the energy supply complex area according to Embodiment 1 of the present invention; Figure 2 This is a schematic diagram of the data stream processing of the risk assessment and concentration measurement system for combustion and explosion in the energy-supply composite area according to Embodiment 1 of the present invention; Figure 3 This is a framework diagram of the HDNN comprehensive risk prediction model in Embodiment 1 of the present invention; Figure 4 This is a schematic diagram of the ResNet layer structure connection form in Embodiment 1 of the present invention. Detailed Implementation

[0039] To make the objectives, technical solutions, and advantages of the embodiments of the present invention clearer, the technical solutions of the embodiments of the present invention will be clearly and completely described below in conjunction with the embodiments of the present invention. Obviously, the described embodiments are only some embodiments of the present invention, 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.

[0040] Example 1 like Figure 1 and Figure 2 As shown, the method for risk assessment and concentration measurement of combustion and explosion in a combined energy supply area provided by the present invention includes the following basic steps: S1. Sensors are deployed using a multi-dimensional data acquisition unit to acquire gas concentration information and weather condition information in real time. In this embodiment, the multidimensional data acquisition unit collects the basic information required by the system and provides foundational information for the subsequent data fusion and modeling unit and the combustion and explosion risk analysis unit. Specifically, the multidimensional data information collected by the multidimensional data acquisition unit includes gas concentration information, such as natural gas concentration and oil and gas concentration; weather condition information, such as temperature, humidity, wind speed and direction; and geographical location information, such as coordinates and elevation. The sensors selected for each type of data information are as follows: Gas concentration information is sampled at a dynamic frequency that is adjusted. For example, it is sampled at 1Hz for normal operation and then increased to 10Hz when the concentration changes abruptly. Specifically, gas concentration information includes, but is not limited to: natural gas (CH4): catalytic combustion sensor (0-100% LEL) or infrared sensor (0-100% VOL); Oil and gas (VOCs) are collected using a photoionization detector, such as a PID sensor, with a range of 0-2000ppm.

[0041] Weather conditions information includes, but is not limited to: Temperature and humidity are collected using digital temperature and humidity sensors, such as the SHT30, with an accuracy of ±0.3℃. Wind speed and direction were collected using an ultrasonic anemometer.

[0042] Geographic location information includes, but is not limited to: Positioning coordinates are collected using GPS / BeiDou modules with an accuracy of ≤1m or RTK positioning (centimeter level). Elevation can be corrected using, for example, a digital elevation model (DEM).

[0043] S2. Concentration spatial distribution modeling is achieved through data fusion and modeling units; In this embodiment, the data fusion and modeling unit preprocesses the obtained multidimensional data information and imports the data from each monitoring point into the geographic information system, linking the monitoring point data to the actual urban map layer. Based on the existing monitoring point data, it performs spatial extension modeling to improve the density and accuracy of the spatial hazardous gas concentration distribution. The data fusion and modeling unit includes a data preprocessing module, a geographic information fusion module, and a concentration distribution spatial modeling module.

[0044] In this embodiment, a data preprocessing module is used to clean, complete, and remove outliers from the data collected above. The main steps of this module are 3σ rule outlier removal and wavelet transform data denoising. The geographic information fusion module binds three types of information from each monitoring point. Using ArcGIS Pro for spatial data management, it overlays multidimensional data from monitoring points with data from municipal gas pipelines and cable channels to create a low-density hazardous gas concentration distribution layer based on the monitoring points.

[0045] The spatial modeling module for concentration distribution is insufficient to meet the risk assessment needs of complex urban areas due to the low density of monitoring point deployment. Therefore, it is necessary to generate a high-density hazardous gas concentration distribution layer through spatial extension prediction and spatial autocorrelation analysis.

[0046] Spatial autocorrelation analysis was performed on the high-density hazardous gas distribution layer to calculate the global Moran index and examine whether the concentration of hazardous gases in geographic space exhibits aggregation or dispersion characteristics.

[0047] In this embodiment, the global Moran index can determine whether the distribution of attributes exhibits clustering and whether outliers exist, thus reflecting the significance of spatial clustering characteristics. As the spatial distribution locations cluster, the correlation becomes increasingly significant. The calculation formula is as follows:

[0048]

[0049]

[0050]

[0051]

[0052] In the formula, I The global Moran exponent, z i and z j w represents the deviation of pollutant values ​​from their average values ​​in spatial units i and j, respectively. ij Let represent the spatial weight between spatial units i and j, n represent the total number of spatial units, and S0 represent the aggregation of all spatial weights. The significance test is passed when z exceeds the critical value of 1.65 (the value set for rejecting the null hypothesis) (P < 0.05). The weights are row-standardized, and the exponent values ​​fall within the range [-1, 1].

[0053] In this embodiment, spatial expansion prediction is performed; specifically, geographical location information and weather condition information are used as inputs, gas concentration information is used as output, and a spatial autocorrelation term is added:

[0054] in, To control spatial weights.

[0055] Subsequently, Bootstrap sampling was used for spatial weighted sampling, and spatial clustering regions were oversampled to ensure that the training set covered spatial heterogeneity. Finally, a spatial random forest expansion algorithm (spatial split weight alpha=0.5, controlling tree complexity max_depth=12, number of trees n_estimators=200) was used to realize spatial modeling, establish a high-density hazardous gas distribution layer based on spatial expansion, and predict the concentration of 1 m×1 m grid after training.

[0056] S3. Conduct combustion and explosion risk analysis based on combustion and explosion risk analysis units and concentration spatial distribution; In this embodiment, the combustion and explosion risk analysis unit extracts combustion and explosion characteristic information of observation points based on a comprehensive database of hazardous gas combustion and explosion characteristics, and calculates the combustion and explosion risk level using combustion and explosion risk analysis methods. The combustion and explosion risk analysis unit includes, but is not limited to, a combustion and explosion characteristic information extraction module and a combustion and explosion risk level calculation module.

[0057] In this embodiment, the combustion and explosion characteristic information extraction module extracts combustion and explosion characteristic information based on the constructed combustion and explosion characteristic database and gas concentration information and weather condition information, including but not limited to: minimum ignition energy, maximum explosion pressure, maximum pressure rise rate and explosion reaction time. The explosion risk level calculation module uses explosion characteristic information as input and performs explosion risk analysis based on the fuzzy comprehensive evaluation method to determine the weight of each explosion characteristic information indicator. It calculates the comprehensive explosion risk score through each importance index and classifies the explosion risk level according to the range of the comprehensive score. This constructs an explosion risk level calculation model to calculate the explosion risk level of the observation point, as detailed in Table 1 below: Table 1 Classification of Fire and Explosion Risk Levels

[0058] S4. Complete the comprehensive risk assessment of the complex area using the comprehensive risk assessment unit.

[0059] In this embodiment, the comprehensive risk assessment data integration module is used to input and integrate information on the hazard factors of potential combustion and explosion risks, establishing a comprehensive risk assessment index database. See Table 2 below; the database factor types mainly include state factors, spatiotemporal factors, human factors, other environmental factors, and gas concentration factors. Table 2 Quantitative Indicator System for Comprehensive Risk Assessment

[0060] In this embodiment, the comprehensive risk assessment implementation module takes the comprehensive risk assessment index database as input, constructs a comprehensive risk prediction model based on a deep neural network, and optimizes the existing combustion and explosion risk level by comprehensively considering state factors, spatiotemporal factors, human factors, other environmental factors and gas concentration factors, so as to obtain the comprehensive risk assessment level.

[0061] In this embodiment, a comprehensive risk label is constructed. Specifically, the Delphi method is used, inviting 10 experts in the field to assign weights to the explosion risk assessment indicators for areas with combined urban cable channels, gas stations, and gas pipelines. The explosion risk assessment indicators are then weighted according to their weights to obtain a comprehensive risk score, and the risk level is classified based on the potential explosion risk. The explosion risk level is used as the label input for the comprehensive risk prediction model. The classification is shown in Table 3 below.

[0062] Table 3 Classification of Fire and Explosion Risk Levels

[0063] Data input and preprocessing are performed, the comprehensive risk assessment index database constructed in the previous stage is stored, outlier handling is performed on the data, and the input data is normalized by Min-Max to avoid differences in units.

[0064] A comprehensive risk prediction model is constructed. Specifically, manually assigning values ​​to the comprehensive risk assessment indicators of fire and explosion by experts would be costly in terms of manpower and time, making it impossible to achieve timely updates, automatic identification, and intelligent prediction of the assessment model. Therefore, after constructing the initial weighting rules for the comprehensive risk assessment indicators, an artificial intelligence-based comprehensive risk assessment prediction model is introduced to simulate expert weighting. This method can adjust in real time through online learning when expert rules are updated late, and can efficiently quantify the interactive effects of a large number of assessment indicators. The constructed comprehensive risk prediction model uses a hybrid deep neural network (HDNN) framework, inputting different types of risk assessment indicators into hidden layers with different applicability. Different risk assessment indicator categories have different data characteristics. Selecting appropriate algorithms for training can improve the feature extraction capability of the hidden layers. In addition, the parallel training of risk assessment indicators by category can further improve the ability to extract significant features within categories and the differences and interpretability between categories.

[0065] In this embodiment, the constructed HDNN comprehensive risk prediction model is as follows: Figure 2As shown. State factors, spatiotemporal factors, human factors, and other environmental factors (including weather conditions from the past 7 days), and gas concentration factors (including 24-hour concentration data sequences of natural gas and oil / gas), with a data collection interval of 1 minute; the data input dimensions for the explosion risk level are 1×12, 1×10, 1×7, 7×1×8, 2×24×60, and 1×1, respectively. In hidden layer 1, the connection form of the convolutional neural network (CNN) layers for state factors, spatiotemporal factors, human factors, and other environmental factors is Conv-ReLU-BN-MaxPool-Conv-ReLU, with a convolutional kernel size of 3, a stride of 1, and a padding layer number of 1. The window size and stride of the MaxPool pooling layer are 2, and the activation function is ReLU. The connection form of the residual neural network (ResNet) layers for other environmental factors is as follows: Figure 3 As shown, the kernel size of the convolutional layers is 3, the stride and the number of padding layers are both 1, and the activation function is tanhLU. The DNN layer in the hidden layer compresses the data to a specified dimension. After passing through hidden layer 1, the feature dimensions of each factor are 5, 5, 5, 5, 5 and 1, respectively. All feature quantities are integrated to form a new feature set with a dimension of 26, which is then input into hidden layer 2. Hidden layer 2 is a DNN layer with a connection form of FC(20)-FC(10)-FC(4)-Softmax. Softmax linear regression transforms the output into a probability distribution, thereby achieving the prediction of the comprehensive risk level. The model uses the cross-entropy loss function to calculate the distance between the predicted probability distribution and the true probability distribution. The cross-entropy loss function calculates the loss as follows:

[0066] In this embodiment, the gas concentration factor includes 24-hour fuel gas / gasoline vapor concentration information, which is time-series data. This type of information not only covers the leakage concentration value at the current moment but also implicitly includes leakage concentration values ​​at past moments and the rate of change of gas concentration. However, traditional expert-weighted risk assessment methods often fail to make reasonable judgments on information that changes over time. To address this characteristic, contrastive learning is introduced to perform unsupervised pre-training on the neural network under the gas concentration factor, improving classification performance through feature enhancement. First, the degree of danger of a small amount of significantly different 24-hour fuel gas / gasoline vapor concentration information is assessed separately based on expert opinions, obtaining a small amount of label information. Contrastive learning uses this small amount of label information to construct positive and negative sample pairs. The encoder extracts features to cluster similar samples in the feature space and separates dissimilar samples. Positive samples are set as all samples of the same category, and negative samples are set as samples of other categories. The encoder extracts the raw data. Features The constructed contrastive learning-based encoder consists of a bidirectional long short-term memory (Bi-LSTM) network and a CNN. The Bi-LSTM layers have 200-300-100-30 neurons each, and a total of two layers. The CNN layers follow the Conv-ReLU-BN-MaxPool-Conv-ReLU configuration. The first convolutional layer has a kernel size of 5, a stride of 3, and 1 padding layer. The second convolutional layer also has a kernel size of 3, a stride of 1, and 1 padding layer. The MaxPool pooling layer has a window size of 2 and a stride of 2, and uses the ReLU activation function. The decoder maps the features to the contrastive space to obtain the predicted data. The constructed encoder / decoder based on contrastive learning restores the data through upsampling. It also improves the contrastive loss based on the difference between positive and negative samples, as shown in the following equation:

[0067] In the formula, For sample i, it is the set of other samples of the same type; The set of all samples.

[0068] The joint loss function based on contrastive learning is now:

[0069] In this embodiment, the Non-Dominated Ranking Genetic Algorithm II (NSGA-II) is used to automatically optimize the number of intermediate layers and neurons in the DNN model in hidden layer 1 and hidden layer 2. NSGA-II is a multi-objective evolutionary algorithm that efficiently solves the Pareto optimal solution set through fast non-dominated ranking, elitist strategy, and crowding comparison mechanism. It is suitable for optimizing DNN structures, which requires balancing multiple objectives (such as accuracy and model complexity).

[0070] In this embodiment, the objective functions are defined as follows: Three objective functions are set: Maximize accuracy: ; Minimize model complexity: ; in, Number of neurons in each layer The number of fully connected layers. The adjustment coefficient is set to 64. Minimize training time:

[0071] The optimal number of fully connected layers and the number of neurons in each layer are obtained by combining three optimization objective functions.

[0072] Set constraints: Set the number of fully connected layers. Number of neurons Set the number of floating-point operations that can be performed per second. .

[0073] Implementation process: First, an initial population P0 is randomly generated, and then a fast non-dominated sort is performed on the current population Pt. i rank The population is stratified according to dominance relationships (e.g., rank1 to rankn), with individuals within each stratum not dominating each other, but higher-ranking individuals dominating lower-ranking individuals. Crowding distance is then calculated. i d The calculation formula is as follows:

[0074] When selecting an operation, prioritize... i rank Smaller individuals, if i rank If they are the same, then select. i d Larger individuals are selected. Finally, the parent and offspring populations (size 2N) are merged, and the best N individuals are selected to be retained for the next generation to avoid losing excellent individuals.

[0075] Constraints are applied; specifically, an adaptive penalty function is constructed to constrain the objective function, calculated as follows:

[0076] in, As constraints, This is the penalty coefficient.

[0077] In summary, the gas and oil concentration measurement method and comprehensive risk assessment process proposed in this invention can fully utilize real-time gas and oil concentration monitoring data to conduct immediate comprehensive explosion risk assessments in areas where cables, gas pipelines, and gas stations are laid together. This provides guidance for mitigating potential explosion risks and proactively managing those risks.

[0078] This invention uses expert opinions to subtly label the risk level of the data and constructs a separate comparative learning model for this data to enhance its feature extraction capabilities.

[0079] This invention classifies risk factors according to data properties and designs optimal models for different categories, making the models more adaptable, traceable in terms of contribution paths of each group, and improving the interpretability of the models. Traditional hybrid neural networks are often structurally complex in order to learn a large number of features at the same time. In the network of this invention, the complexity of the network is simplified by learning feature groups separately through multiple small models, which improves the computational efficiency.

[0080] The hybrid neural network of this invention involves a large number of DNN layers with varied structures, and uses a genetic algorithm to automatically optimize the number of DNN layers and neurons.

[0081] This invention solves the technical problems existing in the prior art, such as single gas leakage concentration detection, isolated combustion and explosion risk assessment, delayed response of real-time monitoring data, and low timeliness and effectiveness of risk management.

[0082] The above embodiments are only used to illustrate the technical solutions of the present invention, and are not intended to limit it. Although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some of the technical features. Such modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the spirit and scope of the technical solutions of the embodiments of the present invention.

Claims

1. A method for risk assessment and concentration measurement of energy-rich composite zone deflagration, characterized in that, The method comprises: S1, arranging sensors with a multi-dimensional data acquisition unit to acquire multi-dimensional data information in real time; S2, modeling the concentration spatial distribution through a data fusion and modeling unit; preprocessing the multi-dimensional data information, importing the multi-dimensional data information of each monitoring point into a geographic information system, linking the monitoring point data to an actual urban land layer, and performing spatial expansion modeling on the existing multi-dimensional data information of the monitoring points; S3, performing explosion risk analysis based on an explosion risk analysis unit and the concentration spatial distribution; using the explosion risk analysis unit to extract explosion characteristic information of the observation point based on a hazardous gas explosion characteristic comprehensive database, and calculating the explosion risk level through an explosion risk analysis method; S4, constructing a comprehensive risk prediction model, and performing composite area comprehensive risk assessment according to the explosion risk level.

2. The method of claim 1, wherein, In S1, the multi-dimensional data information includes gas concentration information, weather condition information, and geographic location information.

3. The method of claim 1, wherein, In S2, 3σ rule outlier rejection and wavelet transform data denoising are performed using a data preprocessing module to clean, complete, and reject the upper collection data; Each of the three types of information in each monitoring point is bound using a geographic information fusion module; A high-density hazardous gas concentration distribution layer is formed through spatial expansion prediction and spatial autocorrelation analysis using a concentration distribution spatial modeling module; Spatial autocorrelation analysis is performed on the high-density hazardous gas distribution layer to calculate the global Moran's index, and the concentration of the hazardous gas in the geographic space is tested for aggregation or dispersion characteristics; Spatial expansion prediction is performed; Spatial weighted sampling is performed using Bootstrap sampling to cluster the space; A spatial random forest expansion algorithm is used to realize spatial modeling, and a high-density hazardous gas distribution layer based on spatial expansion is established.

4. The method of claim 1, wherein the method further comprises: In S3, an explosion characteristic information extraction module is used to extract explosion characteristic information based on the constructed explosion characteristic database according to the gas concentration information and weather condition information; A fuzzy comprehensive evaluation method-based explosion risk analysis is performed using the explosion characteristic information as the characteristic input, the weights of each explosion characteristic information index are determined, the explosion risk comprehensive score is calculated through each importance index, the explosion risk level is divided according to the comprehensive score range, an explosion risk level calculation model is constructed, and the explosion risk level of the observation point is calculated. In S4, the disaster-causing factor information of the explosion potential risk is input and integrated, and the comprehensive risk assessment index database is updated; 5. The method of claim 1, wherein, A comprehensive risk prediction model is constructed based on a deep neural network using the comprehensive risk assessment index database as the input, the explosion risk level is optimized, the comprehensive risk assessment level is obtained, a comprehensive risk label is constructed, the explosion risk assessment indexes of the urban cable channel, gas station, and gas pipeline composite laying area are respectively weighted using the Delphi method, the explosion risk assessment indexes are weighted calculated according to the weights, the comprehensive risk score is obtained, and the risk level is divided according to the explosion potential risk condition; ​ ​ Data input and preprocessing are performed, the comprehensive risk assessment index database is stored, and the data is subjected to outlier processing and Min-Max normalization of the input data; A comprehensive risk prediction model is constructed; An initial comprehensive risk assessment index weight rule is introduced into the comprehensive risk assessment prediction model based on an artificial intelligence algorithm to simulate expert weighting. Based on a hybrid deep neural network (HDNN), the comprehensive risk prediction model is constructed, and different types of risk assessment indexes are respectively input into different applicable hidden layers. The comprehensive risk prediction model uses a cross-entropy loss function to calculate the distance between the predicted probability distribution and the actual probability distribution. A non-dominated sorting genetic algorithm II (NSGA-II) is used to efficiently solve the Pareto optimal solution set and automatically optimize the number of intermediate layers and neurons in the DNN model in hidden layer 1 and hidden layer 2 through fast non-dominated sorting, elitist strategy, and crowding degree comparison mechanism. Randomly generate initial population P0, perform fast non-dominated sorting on current population Pt i rank Sort population individuals by dominance relation.

6. The method of claim 5, wherein the energy-supplied composite zone is a fuel tank of a vehicle. The loss is calculated using a cross-entropy loss function: Based on the difference between positive and negative samples, the contrast loss is improved using the following logic: wherein is a set of other samples of the same class as sample i; is a set of all samples. Based on the contrast learning-based joint loss function, the following logic is used to express it: 。 7. The method of claim 5, wherein the energy-supplied composite zone is a fuel tank of a vehicle. The objective function is defined. The maximum accuracy is defined using the following logic: ； The minimum model complexity is defined using the following logic: ； wherein, number of neurons in each layer, is the number of fully connected layers, is the adjustment coefficient; minimizes training time .

8. The method of claim 5, wherein the energy-supplied composite zone combustion risk assessment and concentration measurement method is characterized by, The crowded distance is calculated using the following logic i d : Selection operation prioritizes selection i rank Smaller individuals, if i rank The same, select i d Larger individuals. Finally, the parent and child populations are merged, and the best N individuals are selected from this pool to carry over to the next generation, avoiding the loss of good individuals.

9. The method of claim 5, wherein the energy-supplied composite zone is a fuel tank of a vehicle. An adaptive penalty function is constructed to constrain the objective function: wherein is a constraint, is a penalty coefficient.

10. A system for risk assessment and concentration measurement of energy-rich composite area deflagration, characterized in that, The system comprises: A multi-dimensional data acquisition unit is used to arrange sensors to acquire multi-dimensional data information in real time. A data fusion and modeling unit is used to model the concentration spatial distribution; the multi-dimensional data information of each monitoring point is imported into a geographic information system, and the monitoring point data is linked to an actual urban land use layer; spatial expansion modeling is performed on the existing multi-dimensional data information of the monitoring points, and the data fusion and modeling unit is connected to the multi-dimensional data acquisition unit. A combustion risk analysis unit is used to analyze the combustion risk in combination with the concentration spatial distribution; the combustion risk analysis unit extracts the combustion characteristic information of the observation point based on a comprehensive database of hazardous gas combustion characteristics, and calculates the combustion risk level through a combustion risk analysis method; the combustion risk analysis unit is connected to the data fusion and modeling unit. A comprehensive risk assessment unit is used to construct a comprehensive risk prediction model and perform comprehensive risk assessment of the composite area according to the combustion risk level; the comprehensive risk assessment unit is connected to the combustion risk analysis unit.

Images