A quick plugging emergency system for leakage of chemical storage tank

Abstract

The present application relates to the technical field of chemical safety emergency, and particularly relates to a quick plugging emergency system for leakage of a chemical storage tank. The present application comprises six modules, namely, leakage sensing data acquisition, leakage feature extraction, plugging decision generation, leakage trend prediction, plugging decision optimization and plugging execution control. The leakage sensing data acquisition module acquires real-time multi-sensor leakage data flow; the leakage feature extraction module analyzes data to determine leakage source and diffusion feature information in combination with preset leakage physical constraint information; the plugging decision generation module generates an initial plugging scheme according to an embedded decision space and fitness constraint conditions; the leakage trend prediction module predicts leakage development to obtain predicted feature information; the plugging decision optimization module optimizes the initial scheme to form an optimized scheme; and the plugging execution control module drives an execution device to complete plugging. The system can improve leakage detection accuracy, decision scientificity and plugging effectiveness, and ensure safe operation of the chemical storage tank.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of chemical safety emergency, in particular to a kind of chemical storage tank leakage rapid plugging emergency system. BACKGROUND

[0002] In the process of chemical production and storage, chemical storage tank as the core storage equipment of various flammable and explosive, toxic and harmful medium, its operation safety is directly related to production order, personnel health and surrounding ecological environment. With the continuous expansion of chemical industry scale, the volume of storage tank gradually increases, the complexity and danger of storage medium also increase, and the risk of leakage accident always exists.

[0003] At present, the emergency treatment means for chemical storage tank leakage still has obvious deficiencies in leakage detection and plugging decision link. In the aspect of leakage detection, the traditional detection method mainly depends on manual inspection or single type sensor monitoring. Manual inspection is not only low in efficiency, but also difficult to guarantee personnel safety in the scene of complex storage tank area and medium with toxicity or corrosiveness. Single sensor monitoring is easily affected by environmental factors such as temperature, humidity change or other gas components, which leads to inaccurate leakage signal identification, cannot obtain comprehensive leakage sensing data in time, and often has the problem of leakage discovery lag, missing the best plugging opportunity.

[0004] In the aspect of plugging decision and execution link, the existing technology often lacks systematic decision mechanism. In most cases, the formulation of plugging scheme depends on the experience of operators, and the experience judgment is difficult to deal with the complex situation of different leakage source characteristics (such as leakage position, leakage aperture size, leakage medium type) and leakage diffusion characteristics (such as diffusion speed, diffusion range). When the leakage condition changes, the pre-prepared plugging scheme cannot be adjusted in time, which leads to poor plugging effect. At the same time, the existing technology lacks effective prediction of leakage trend, cannot optimize the plugging strategy in advance according to the development trend of leakage, often can only passively deal with the current leakage state, is difficult to effectively control the leakage diffusion, and may cause greater property loss, personnel casualty and environmental pollution. In addition, in the process of plugging execution, due to the lack of precise linkage with plugging decision scheme, the action control of execution device is not accurate enough, which further affects the plugging efficiency and effect, and cannot meet the requirements of rapidity, accuracy and effectiveness of chemical storage tank leakage emergency treatment. SUMMARY

[0005] The main purpose of the present application is to provide a kind of chemical storage tank leakage rapid plugging emergency system, to solve the technical problems in the prior art.

[0006] The present application provides a kind of chemical storage tank leakage rapid plugging emergency system, comprising:

[0007] A leakage sensing data collection module is configured to acquire leakage sensing data streams from a plurality of sensors arranged around a chemical storage tank and near a leakage point in real time;

[0008] A leakage feature extraction module is configured to perform leakage feature analysis based on preset leakage physical constraint information and in combination with the leakage sensing data streams, to determine leakage source feature information and leakage diffusion feature information;

[0009] A plugging decision generation module is configured to generate an initial plugging decision scheme based on an embedded plugging decision space and plugging fitness constraint conditions, in combination with the leakage source feature information and the leakage diffusion feature information;

[0010] A leakage trend prediction module is configured to perform leakage trend prediction on the leakage source feature information and the leakage diffusion feature information, to obtain predicted leakage source feature information and predicted leakage diffusion feature information;

[0011] A plugging decision optimization module is configured to perform plugging decision compensation optimization on the initial plugging decision scheme based on the predicted leakage source feature information and the predicted leakage diffusion feature information, to generate an optimized plugging decision scheme;

[0012] A plugging execution control module is configured to drive corresponding plugging execution devices to perform leakage plugging operations based on the optimized plugging decision scheme.

[0013] Preferably, the leakage feature extraction module comprises:

[0014] A storage tank foundation information modeling unit is configured to acquire a set of storage tank foundation information of a target chemical storage tank, and to construct a storage tank physical model based thereon;

[0015] A leakage feature fusion analysis unit is configured to input the leakage sensing data streams into the storage tank physical model, to generate a leakage feature fusion atlas;

[0016] A leakage feature information confirmation unit is configured to finally confirm the leakage source feature information and the leakage diffusion feature information based on the preset leakage physical constraint information and the leakage feature fusion atlas.

[0017] Preferably, the leakage trend prediction module comprises:

[0018] A historical leakage data indexing unit is configured to index historical leakage feature data of a historical similar leakage event according to the leakage source feature information and the leakage diffusion feature information;

[0019] A leakage mode learning unit is configured to perform leakage mode learning analysis according to the historical leakage feature data, to obtain a leakage mode learning rate and a leakage mode degradation rate;

[0020] A prediction feature generation unit is configured to, in combination with the leakage mode learning rate and the leakage mode degradation rate, perform trend extrapolation on the leakage source feature information and the leakage diffusion feature information to generate predicted leakage source feature information and predicted leakage diffusion feature information.

[0021] Preferably, the plugging decision generation module is specifically configured to:

[0022] Traverse the plugging decision space to extract candidate plugging decision records;

[0023] Calculate the similarity depth coefficient of the leakage source feature information and the sample leakage source feature information in the candidate records;

[0024] Calculate the similarity depth coefficient of the leakage diffusion feature information and the sample leakage diffusion feature information in the candidate records;

[0025] According to the preset weight, the two similarity depth coefficients are weighted and fused to generate a decision correlation coefficient;

[0026] The candidate plugging decision records satisfying the correlation threshold form an initial plugging decision scheme.

[0027] Preferably, the plugging decision optimization module is specifically configured to:

[0028] According to the predicted leakage source feature information and the predicted leakage diffusion feature information, the plugging decision space is secondarily associated and calibrated to obtain an optimized decision space;

[0029] Based on the plugging fitness constraint condition and the optimized decision space, iterative decision search is performed to generate a compensation plugging decision scheme;

[0030] The compensation plugging decision scheme and the initial plugging decision scheme are fused and optimized to generate the optimized plugging decision scheme.

[0031] Preferably, the leakage mode learning unit is specifically configured to:

[0032] Classify the leakage event information in the historical leakage feature data to obtain a plurality of classified leakage modes;

[0033] Extract a historical leakage feature parameter sequence under each classified leakage mode;

[0034] Respectively perform feature dimension reduction and time sequence arrangement on each historical leakage feature parameter sequence to obtain a plurality of leakage feature evolution sequences;

[0035] According to the plurality of leakage feature evolution sequences, perform mode evolution analysis to calculate the leakage mode learning rate and the leakage mode degradation rate.

[0036] As preferred, the leakage pattern learning unit performs pattern evolution analysis according to multiple leakage feature evolution sequences, which includes:

[0037] A pattern evolution analyzer is constructed, which takes a set of historical sample leakage feature evolution sequences as input and outputs a set of historical sample leakage pattern learning rates and a set of historical sample leakage pattern degradation rates;

[0038] The multiple leakage feature evolution sequences currently obtained are input into the pattern evolution analyzer for classification;

[0039] The multiple pattern learning rates and the multiple pattern degradation rates output by the classification are weighted and fused according to the similarity between the current leakage event feature and each classified leakage pattern, and finally the leakage pattern learning rate and the leakage pattern degradation rate are obtained.

[0040] As preferred, the prediction feature generation unit combines the leakage pattern learning rate and the leakage pattern degradation rate to perform trend extrapolation, which includes:

[0041] Identify the classified leakage pattern with the highest matching degree with the current leakage event feature as the reference leakage pattern;

[0042] Calculate the feature deviation amount of the current leakage event feature from the standard feature information of the reference leakage pattern;

[0043] Set a prediction correction coefficient based on the feature deviation amount;

[0044] Dynamically adjust the leakage pattern learning rate and the leakage pattern degradation rate using the prediction correction coefficient;

[0045] Use the adjusted learning rate and degradation rate to drive the trend extrapolation process.

[0046] As preferred, the plugging decision generation module weights and fuses the two similarity depth coefficients according to a preset weight, which includes:

[0047] Obtain a set of sample leakage event feature sets and a set of sample decision correlation coefficients;

[0048] Construct a decision correlation weight learner, which takes the sample leakage source feature similarity depth coefficient and the sample leakage diffusion feature similarity depth coefficient as input and outputs the sample decision correlation coefficient;

[0049] Input the leakage source feature information similarity depth coefficient and the leakage diffusion feature information similarity depth coefficient currently obtained into the decision correlation weight learner;

[0050] The decision correlation weight learner outputs the decision correlation coefficient.

[0051] As preferred, the system further includes:

[0052] a real-time data update feedback module, configured to continuously collect updated leakage sensing data streams during the plugging execution process;

[0053] a dynamic decision adjustment module, configured to feed the updated leakage sensing data streams to the leakage feature extraction module, the leakage trend prediction module and the plugging decision optimization module, triggering dynamic iterative adjustment of the optimized plugging decision scheme.

[0054] The present application has the following advantages:

[0055] In the leakage detection link, the leakage sensing data collection module set by the system can obtain the leakage sensing data streams of multiple sensors arranged around the storage tank and near the leakage point in real time. Compared with the traditional single sensor or manual inspection method, the data collection of multiple sensors can cover a wider monitoring range, obtain multiple types of sensing information, effectively reduce the interference of environmental factors on leakage detection, improve the comprehensiveness and accuracy of leakage signal identification, and ensure that the leakage situation can be discovered in time to gain valuable time for subsequent plugging work.

[0056] In terms of plugging decision making, the plugging decision generation module generates an initial plugging decision scheme based on the built-in plugging decision space and plugging fitness constraint conditions, combined with the leakage source feature information and leakage diffusion feature information determined by the leakage feature extraction module, breaking away from the limitations of traditional schemes relying on manual experience. The built-in decision space covers plugging strategies in multiple different leakage scenarios, and the fitness constraint conditions ensure the feasibility and rationality of the initial scheme, which can quickly match the appropriate plugging scheme for specific leakage conditions, avoiding the subjectivity and limitations of experience judgment, and improving the scientificity and efficiency of plugging decision making.

[0057] The leakage trend prediction module can predict the leakage trend of the leakage source feature information and the leakage diffusion feature information, obtain predicted leakage source feature information and predicted leakage diffusion feature information, which compensates for the defect that the prior art cannot predict the development trend of the leakage. Through the prediction of the leakage trend, the possible change direction of the leakage can be known in advance, such as whether the leakage amount will increase or the diffusion range will expand, which provides a key basis for the optimization of the plugging decision, so that the plugging strategy is no longer limited to coping with the current leakage state, but can be laid out in advance to better control the development of the leakage.

[0058] The blocking decision optimization module compensates and optimizes the initial blocking decision scheme based on the predicted leakage feature information, generates an optimized blocking decision scheme, and further improves the adaptability and effectiveness of the blocking scheme. The initial scheme may only be applicable to the current leakage condition, and as the leakage develops, the original scheme may no longer be applicable. Through the adjustment of the optimization module, the blocking scheme can always match the actual development trend of the leakage, ensuring the pertinence of the blocking measures and avoiding the problem of poor blocking effect due to lagging scheme.

[0059] The blocking execution control module drives the corresponding blocking execution device to perform leakage blocking operation based on the optimized blocking decision scheme, realizing accurate linkage of decision and execution. The optimized blocking scheme can provide clear and accurate action instructions for the execution device, ensuring that the execution device can accurately act on the leakage point, improving the accuracy and efficiency of the blocking operation, quickly realizing effective blocking of the leakage, reducing the harm of the leakage medium to the surrounding environment, personnel and property, and reducing the loss caused by the leakage accident. The entire system forms a complete closed loop from leakage detection, decision generation, trend prediction, scheme optimization to execution control, significantly improving the overall ability of the chemical storage tank leakage emergency handling, and ensuring the safety of the chemical production and storage process. BRIEF DESCRIPTION OF DRAWINGS

[0060] Figure 1 a timing diagram of the chemical storage tank leakage rapid blocking emergency system described in the present application;

[0061] Figure 2 a flowchart of the working process of the leakage feature extraction module;

[0062] Figure 3 a flowchart of the working process of the leakage trend prediction module.

[0063] The implementation of the present application, functional features and advantages will be further described with reference to the accompanying drawings. DETAILED DESCRIPTION

[0064] It should be understood that the specific embodiments described herein are only used to explain the present application and not to limit the present application.

[0065] As Figure 1 shown, the present application provides a chemical storage tank leakage rapid blocking emergency system, comprising: a leakage sensing data acquisition module, a leakage feature extraction module, a blocking decision generation module, a leakage trend prediction module, a blocking decision optimization module and a blocking execution control module.

[0066] The leakage sensing data acquisition module acquires a leakage sensing data stream in real time from a plurality of sensors arranged around the chemical storage tank and near the leakage point, the sensor types including but not limited to pressure sensors, temperature sensors, gas concentration sensors and flow sensors, which cover the outer wall of the tank and the adjacent pipeline area in a distributed network layout, and the data acquisition frequency can be dynamically adjusted according to the leakage risk level. The leakage feature extraction module analyzes the leakage features according to the preset leakage physical constraint information and in combination with the real-time incoming leakage sensing data stream, the leakage physical constraint information including the tank material strength parameters, the medium physical property parameters and the environmental temperature and humidity conditions, and determines the leakage source feature information and the leakage diffusion feature information through the feature analysis, the leakage source feature information including the leakage point position coordinates, the leakage aperture size and the leakage flow rate vector, and the leakage diffusion feature information including the diffusion direction angle, the diffusion rate and the concentration gradient distribution. The plugging decision generation module generates an initial plugging decision scheme based on the built-in plugging decision space and plugging fitness constraint conditions in combination with the obtained leakage source feature information and leakage diffusion feature information, the plugging decision space storing a plurality of historical leakage event response schemes and their execution effect records, and the plugging fitness constraint conditions including the plugging material resistance index, the operation time window limit value and the resource availability condition. The leakage trend prediction module predicts the leakage trend of the leakage source feature information and the leakage diffusion feature information, the prediction method based on time series analysis and physical evolution model, obtaining predicted leakage source feature information and predicted leakage diffusion feature information, the predicted leakage source feature information including the expected leakage point change trajectory and the leakage pressure decay curve, and the predicted leakage diffusion feature information including the diffusion range change rate and the hazard area evolution map. The plugging decision optimization module optimizes the initial plugging decision scheme based on the predicted leakage source feature information and the predicted leakage diffusion feature information, the compensation optimization process being realized through a multi-objective decision algorithm to generate an optimized plugging decision scheme, which refines the plugging operation steps, the resource scheduling plan and the timing control logic. The plugging execution control module drives the corresponding plugging execution device to perform the leakage plugging operation based on the optimized plugging decision scheme, the plugging execution device including an automatic injector, a mechanical plugging device and a remote control robot, and the control instructions being issued to the execution terminal through an industrial bus protocol.

[0067] In one embodiment, embodiment 1: refer to Figure 2 and Figure 3, the leakage feature extraction module includes a storage tank basic information modeling unit, a leakage feature fusion analysis unit and a leakage feature information confirmation unit. The storage tank basic information modeling unit calls a complete set of basic information of the target chemical storage tank from the factory database, which covers the cylindrical geometric size parameters of the storage tank, the wall thickness distribution gradient of different height sections, the yield strength data of carbon steel material, the design pressure range value and the viscosity-temperature relationship curve of the current storage medium. When constructing the physical model of the storage tank based on these structured data, the parameterized finite element modeling technology is adopted. The model grid division precision is set to millimeter level according to the leakage monitoring requirements. The unit type includes shell elements and solid elements combined to accurately simulate the stress concentration effect of the weld area. The material constitutive model adopts the elastic-plastic bilinear hardening criterion and embeds the temperature related attribute. The leakage feature fusion analysis unit receives real-time leakage sensing data stream from the distributed sensor network. The data stream includes pressure fluctuation signal, temperature abnormal gradient and gas concentration spatial distribution matrix. These time series data are loaded onto the corresponding grid nodes of the storage tank physical model through coordinate mapping algorithm. The multi-physical field coupling calculation process involves fluid-structure interaction analysis. The calculation engine uses an explicit dynamic solver to handle the transient leakage impact effect. The output result is a leakage feature fusion atlas. The atlas organizes data in the form of four-dimensional tensor (three-dimensional space coordinates + time dimension). The tensor element value comprehensively reflects the spatial diffusion flux of the leaked material, the temperature conduction rate and the material stress concentration coefficient. The leakage feature information confirmation unit calls the preset leakage physical constraint information library, which stores the critical pressure table of medium phase transition, the leakage rate empirical formula under different pore diameters and the diffusion boundary layer calculation model. The leakage feature fusion atlas is input into the constraint optimization solver. The solver minimizes the deviation function of the measured data and the physical constraints based on the sequential quadratic programming algorithm. The output confirmed leakage source feature information includes the accurate three-dimensional position of the leakage point relative to the reference coordinate system of the storage tank, the equivalent diameter value of the equivalent leakage pore diameter and the velocity vector field of the leakage jet. The leakage diffusion feature information includes the diffusion sector angle with the leakage point as the origin, the concentration decay curve at different distances and the grading contour map of the danger area.

[0068] The leakage trend prediction module is composed of a historical leakage data indexing unit, a leakage pattern learning unit, and a prediction feature generation unit. The historical leakage data indexing unit performs multi-dimensional matching retrieval in the factory historical event database. The retrieval conditions include the relative positional relationship between the current leakage source position and the tank weld, the medium type similarity score, and the historical matching interval of the environmental wind speed. The indexing algorithm uses an improved k-d tree space partitioning technique to accelerate the nearest neighbor search process, and introduces a dynamic time warping algorithm to align different time scale leakage event sequences. The retrieval results return three sets of historical similar leakage event feature data, each set containing complete leakage pressure time series records, diffusion cloud infrared thermal imaging sequences, and emergency disposal logs. The leakage pattern learning unit performs pattern mining on the historical leakage feature data obtained by retrieval. First, the spectral clustering algorithm is used to divide the historical events into different pattern categories according to the leakage cause (corrosion perforation / mechanical damage / seal failure). Key feature parameter sequences are extracted under each category, including leakage pressure decay curve, diffusion radius expansion rate matrix, and medium volatility index change value. Then, the time series decomposition technique is used to separate the trend item and the periodic item of each parameter sequence. The trend item is reduced in dimension by principal component analysis to form a low-dimensional feature evolution trajectory. Finally, a parameter estimation framework based on state space model is constructed. This framework models the leakage evolution process as a dynamic system affected by the intensity of external intervention. The system state transition equation contains two core parameters: the leakage pattern learning rate, which represents the decay rate of the leakage intensity affected by the plugging operation per unit time, and the leakage pattern degradation rate, which represents the natural leakage rate increment gradient caused by material fatigue. The parameter estimation uses the maximum likelihood estimation method combined with the EM algorithm for iterative optimization. The prediction feature generation unit uses an adaptive prediction framework to handle the trend extrapolation task. The framework first calculates the Mahalanobis distance between the current leakage event feature vector and each historical pattern category, selects the pattern with the smallest distance as the reference benchmark, and then analyzes the local dynamic time warping distance between the current leakage pressure gradient and the reference mode benchmark curve. According to the distance deviation, a non-linear correction coefficient matrix is generated, which is used to dynamically adjust the weight distribution of the leakage pattern learning rate and the degradation rate. Finally, the extrapolation process uses a physical information neural network architecture. The network input layer injects the adjusted learning rate and degradation rate parameters, the hidden layer embeds the Navier-Stokes equation of fluid mechanics as a physical constraint, and the output layer generates predicted leakage source feature information including the expansion probability distribution of the leakage orifice diameter, the upper and lower bounds of the leakage flow rate confidence interval in the future time window, as well as the predicted leakage diffusion feature information including the diffusion isopleth model at different times, the vapor cloud explosion probability thermodynamic map, and the toxicity influence range spatio-temporal evolution animation.

[0069] The leakage feature fusion analysis unit implements a multi-level verification mechanism in the data processing process. The original sensor data is first filtered to remove abnormal pulse interference through sliding window outlier detection. The timestamp alignment module uses the PTP precise clock protocol to ensure microsecond-level synchronization of distributed sensors. The spatial interpolation algorithm reconstructs the physical field distribution in the area not covered by the sensor based on the radial basis function network. The coupling calculation process implements a dynamic load balancing strategy. When a high concentration leakage signal is detected in a local area, the calculation grid density of that area is automatically encrypted. The leakage feature information confirmation unit is equipped with a constraint conflict resolution mechanism. When the optimization solver detects that the measured leakage rate exceeds the theoretical maximum defined by the physical constraints, the sensor calibration coefficient recalculation process is automatically triggered. The sensitivity parameters of the pressure sensor are adjusted through the back propagation algorithm until the constraint consistency requirement is met.

[0070] The historical leakage data indexing unit establishes a dynamic indexing update strategy. When a new leakage event is handled, the feature vector of the event is automatically extracted and updated to the historical database. The index structure uses the cuckoo hashing algorithm to achieve real-time retrieval capability with O(1) time complexity. The similarity matching function introduces a transfer learning mechanism. When handling new media leakage, the chemical property similarity model is automatically called for cross-media feature mapping. The leakage pattern learning unit implements an online model updating scheme. The pattern classifier uses an incremental learning architecture. After inputting new leakage event data, the classification hyperplane is dynamically adjusted through the passive attack algorithm without global model reconstruction. The parameter estimation process uses the sliding window finite memory least squares method to ensure that the system can track the non-stationary changes of the leakage evolution law.

[0071] The prediction feature generation unit sets a prediction credibility evaluation link. The probability distribution of the prediction result is generated by sampling the neural network output multiple times through the Monte Carlo method. When the coefficient of variation of the predicted leakage pressure exceeds the threshold, it automatically switches to the conservative prediction mode. In this mode, the diffusion range estimation model based on the worst-case scenario is used instead of the regular prediction output. The trend extrapolation process implements a closed-loop correction mechanism. After obtaining new sensor observation data, the residual analysis of the predicted value and the measured value is performed. The residual sequence is estimated through the Kalman filter to estimate the system error parameters and feedback to correct the boundary condition setting of the physical information neural network.

[0072] A feature transfer interface is established between the leakage feature information confirmation unit and the prediction feature generation unit. The leakage source position uncertainty ellipse parameters output by the confirmation unit are directly transferred to the prediction unit. The prediction unit injects this parameter as a spatial constraint condition into the diffusion prediction model. When the leakage point positioning has a ±5 cm position ambiguity, the prediction model automatically generates the diffusion range envelope of the corresponding ambiguous area. The leakage pattern degradation rate parameter output by the pattern learning unit is synchronized in real time to the plugging decision module. The decision module dynamically adjusts the dosage estimation coefficient of the plugging material according to the material degradation rate, realizing the closed-loop flow of the whole system parameters.

[0073] In one embodiment, the occlusion decision generation module traverses the occlusion decision space to extract candidate occlusion decision records, the decision space is organized in a weighted directed graph structure with historical case data, nodes represent occlusion action states (such as glue injection preparation, mechanical clamp positioning), edges represent state transition conditions (such as triggering the next operation after the pressure drops to the threshold), and each edge is associated with historical execution success probability and resource consumption indicators. When calculating the similarity depth coefficient of the leakage source feature information and the candidate record sample leakage source feature information, a multi-scale feature matching strategy is adopted: for the leakage point position feature, the spatial similarity is calculated by the tank surface geodesic distance algorithm; for the leakage hole diameter feature, the dynamic time warping algorithm is applied to align the current diameter change curve and the historical sample curve; for the leakage flow velocity vector, the weighted combination value of the direction cosine similarity and the length relative error is calculated, and the final similarity depth coefficient is the geometric mean of the three sub-indicators. When calculating the similarity depth coefficient of the leakage diffusion feature information and the candidate sample leakage diffusion feature information, a regional comparison analysis is implemented, and the diffusion region is divided into near-field core area, middle-field transition area and far-field edge area, the core area uses the structural similarity index of concentration gradient distribution, the transition area calculates the Fréchet distance of the diffusion front movement trajectory, and the edge area evaluates the Jacard coefficient of the dangerous area overlap rate, and the comprehensive coefficient is aggregated by the regional importance weight. According to the preset weight, the two similarity depth coefficients are weighted and fused to generate the decision correlation coefficient, the weight distribution adopts a dynamic adjustment mechanism based on evidence theory, when the leakage source feature contains high-pressure jet characteristics, the leakage source weight factor is automatically increased, and when the diffusion feature shows large-area cloud diffusion, the diffusion feature weight proportion is increased, and the fusion process introduces the D-S evidence synthesis rule to process uncertain information. The candidate records whose decision correlation coefficients meet the correlation threshold are selected to form the initial occlusion decision scheme, the threshold setting adopts an adaptive mechanism, and the initial value is taken from the median level of the historical decision case library, and when real-time monitoring shows signs of leakage acceleration, the threshold is dynamically adjusted to expand the coverage of the candidate scheme, and the final generated initial scheme contains three groups of alternative decision chains, each decision chain defines the occlusion material injection path, the mechanical clamp installation angle sequence and the operator safety evacuation route in detail.

[0074] The plugging decision optimization module performs secondary correlation calibration on the plugging decision space according to the predicted leakage source feature information, and the calibration process focuses on the dynamic reconstruction of the decision graph structure: the spatial coordinate constraints of the state node are corrected based on the predicted leakage point change trajectory; the pressure threshold condition of state transition is adjusted according to the predicted leakage pressure decay curve; the resource scheduling constraint equation is updated by referring to the predicted diffusion range change rate. After obtaining the optimized decision space, multi-objective iterative decision search is implemented, the search algorithm uses an improved NSGA-III framework, the decision variables include the plugging material injection rate, the clamp locking torque gradient and the operation time allocation proportion, the constraint conditions integrate the material temperature resistance upper limit, the maximum response time window and the available robot quantity limit in the plugging fitness constraint condition, and the objective function optimizes the plugging success rate, the secondary leakage risk index and the emergency resource consumption amount. When generating the compensation plugging decision scheme, the scenario tree analysis method is introduced, the branching decision tree of the benchmark scenario, the optimistic scenario and the pessimistic scenario is established according to the predicted leakage trend, the corresponding parameter adjustment strategy is generated under each scenario node, for example, when the predicted leakage aperture may expand, the deployment scheme of the pre-gluing point is added in the pessimistic scenario branch. When the compensation plugging decision scheme and the initial plugging decision scheme are fused and optimized, the decision sequence splicing technology is used to identify the steady-state operation segment in the initial scheme and the dynamic adjustment segment in the compensation scheme, and the optimal inter-segment connection point is matched through the Hungarian algorithm, and the finally generated optimized plugging decision scheme presents as a time-action-resource three-dimensional decision matrix, the longitudinal dimension of the matrix is divided into operation time slices of fifteen seconds, the horizontal dimension allocates operation instructions of different functional areas, and the depth dimension is associated with resource scheduling lists and safety monitoring points.

[0075] The similar depth coefficient calculation process implements real-time quality monitoring, and automatically triggers a local recalculation mechanism when detecting abnormal data of a certain sensor channel, and only for a subset of reliable sensor data to regenerate part of the similarity index. The decision correlation coefficient fusion link sets up conflict detection logic, and when the similarity index of the leakage source and the diffusion feature appears directional contradiction (such as high source similarity and low diffusion similarity), the conflict resolution program based on case reasoning is automatically started, and the historical conflict case disposal mode is retrieved to guide the current weight allocation. The diversity protection strategy is implemented in the candidate scheme screening stage, in addition to the main scheme that meets the correlation threshold, a backup scheme with low correlation coefficient but significant difference in decision dimension is forced to be reserved, which is stored in the execution queue as an emergency backup scheme. The incremental graph update technique is used in the secondary correlation calibration process, and only the state nodes affected by the predicted features are adjusted locally to avoid the computational overhead of global graph reconstruction. The dynamic population management is implemented in the iterative decision search, the initial population contains the decoded gene sequence of the initial plugging decision scheme, the elite individuals are reserved and new individuals with random disturbance are injected in each generation evolution, and the mutation operator range is automatically expanded when the optimal solution of three consecutive generations is not improved. The probability weight mechanism is established by scenario tree analysis, the benchmark scenario is given a 60% probability of occurrence, the optimistic and pessimistic scenarios each account for 20%, and the probability value is dynamically calibrated according to the confidence interval output by the prediction model. The time sequence conflict detection is implemented in the decision sequence splicing process, and when it is identified that there is a device space interference between the mechanical clamp installation and the glue injection operation, a safety time buffer section is automatically inserted and the robot motion trajectory is adjusted. The logical verification check is performed before the output of the optimized plugging decision scheme, the verifier simulates the whole process and detects whether there is an operation sequence that violates the physical constraints, for example, the close-range manual operation is prohibited when the tank pressure exceeds the safety threshold. The layered encoding format is used for scheme delivery, the execution layer instruction is converted into G code of the robot control system, the monitoring layer instruction generates a list of DCS system alarm trigger conditions, and the resource layer instruction is decomposed into a picking task order of the warehouse. The decision space maintenance module automatically collects the implementation effect data after each scheme execution, and the feature-decision mapping pairs of successful cases are supplemented to the decision space graph structure, and the failure cases are marked as taboo transition paths and the state transition probability matrix is updated. The online parameter update is implemented for the plugging fitness constraint condition, and when the environmental temperature sensor detects abnormal heating, the material tolerance temperature threshold is automatically lowered and the decision optimization cycle is retriggered. The dynamic board mechanism is established for the resource availability condition, and the real-time synchronization of emergency resource inventory state and engineering vehicle GPS position information is implemented, and when the resource change causes the constraint condition to change, a decision scheme revision patch is generated. The elastic adjustment interval is set for the time window constraint, and the standard response time limit is allowed to be broken during the leakage acceleration stage, but additional safety protection measures are required, and the related adjustment strategy is written into the exception handling clause of the decision scheme.

[0076] In one embodiment, embodiment 3: the leakage mode learning unit classifies the leakage event information in the historical leakage feature data, and a density-based spatial clustering method is used for the classification algorithm, which can automatically identify groups of leakage events with similar spatio-temporal evolution patterns. The clustering feature vector contains twelve-dimensional parameters such as initial pressure value, medium volatility level, and diffusion rate change gradient. Each clustering center represents a typical leakage mode category. When extracting the historical leakage feature parameter sequence under each classified leakage mode, time series alignment and standardization preprocessing are performed. The alignment process uses a dynamic time warping algorithm to eliminate the effects of different event duration differences. Standardization converts physical quantities such as pressure and concentration into dimensionless values in the [0, 1] interval, ensuring the comparability of different dimension parameters. Feature dimension reduction is performed on each historical leakage feature parameter sequence. The t-SNE manifold learning algorithm is used to project high-dimensional time series data into a two-dimensional latent space, significantly reducing computational complexity while preserving similarity relationships between sequences. The reduced data points are connected in time order to form leakage feature evolution trajectories. The multi-scale permutation entropy analysis is applied to the time series arrangement process to identify key turning points and stable development stages in the evolution trajectory. Based on the turning points, the continuous trajectory is divided into several evolution stage segments with clear physical meaning.

[0077] When performing mode evolution analysis on multiple leakage feature evolution sequences, a state space representation based on a hidden Markov model is constructed, mapping each evolution stage to a hidden state. The state transition probability matrix describes the evolution relationship between stages. The leakage mode learning rate obtained by calculation represents the response sensitivity of the system to external intervention measures, defined as the percentage change in leakage rate under unit intervention intensity. The leakage mode degradation rate reflects the self-degradation trend caused by material fatigue, obtained by monitoring the natural growth rate of leakage parameters under no intervention state. The following calculation formula is introduced during the mode evolution analysis process to evaluate the evolution dynamics:

[0078]

[0079] wherein: represents the leakage mode learning rate, is the observation time window length, represents the leakage rate change at the time step, is the leakage rate reference value of the previous step, is the external intervention intensity at time, is the maximum intervention intensity allowed by the system. This formula quantifies the leakage change rate caused by unit intervention effect. The higher the value, the more sensitive the system is to intervention.

[0080] When the leakage mode learning unit performs mode evolution analysis according to multiple leakage feature evolution sequences, the mode evolution analyzer constructed adopts a deep recurrent neural network architecture. The network input layer receives a set of standardized historical sample leakage feature evolution sequences, each sequence containing variable-length multidimensional time step data. The hidden layer is designed as a bidirectional LSTM structure, with the forward LSTM capturing historical evolution rules and the backward LSTM learning future trend dependency relationships. The attention mechanism layer automatically focuses on key turning points in the sequence. The output layer generates a set of historical sample leakage mode learning rates and a set of historical sample leakage mode degradation rates in parallel. The learning rate output node uses a Sigmoid activation function to constrain the value to the range [0, 1], and the degradation rate output node uses a Softplus activation function to ensure non-negativity. When multiple leakage feature evolution sequences are input into the mode evolution analyzer, sequence length standardization processing is performed, different length sequences are adjusted to a uniform time step length through an interpolation algorithm, and missing values are filled in based on sequence similarity collaborative filtering. The classification process adopts a hierarchical classification strategy, first performing coarse classification according to the leakage medium type, and then performing fine-grained mode matching within each medium category. The matching degree calculation adopts a weighted combination of dynamic time warping distance and cosine similarity. When the similarity between the current leakage event features and each classified leakage mode is weighted and fused, the Mahalanobis distance is used for similarity measurement to consider the correlation between each dimension of the feature vector, and the weighting coefficient is obtained through Softmax normalization of the similarity value. The final output of the leakage mode learning rate and the leakage mode degradation rate is the expected value of the corresponding parameters of each mode.

[0081] The leakage feature evolution sequence generation process implements a quality control system, which calculates a reconstruction error indicator for each reduced dimension sequence. When the error exceeds the threshold, it automatically switches to the principal component analysis dimension reduction method for reprocessing. The pattern evolution analyzer training uses a transfer learning strategy, pre-training with a large-scale time series prediction dataset, and fine-tuning network parameters with historical leakage data to improve small sample learning ability. Visualization analysis of attention weights helps identify key influencing factors in the evolution process, such as pressure changes being more indicative than concentration changes in some patterns. Similarity calculation introduces uncertainty quantification, automatically reducing the weight proportion of feature dimensions with larger sensor measurement errors, to avoid the negative impact of noisy data on the fusion result. The storage of historical leakage feature parameter sequences uses a time series database architecture, supporting efficient range queries and similarity searches, and each sequence is associated with rich metadata including environmental conditions, device status, and other information. The feature dimension reduction process retains the reverse mapping function, allowing the evolution trajectory in the two-dimensional latent space to be reinterpreted as changes in the original physical quantities, enhancing the interpretability of the results. The pattern evolution analyzer implements an online learning mechanism, automatically adding new leakage event data to the training set after verification, and regularly updating model parameters to adapt to system changes. The similarity weighted fusion link sets up a consistency check, automatically triggering an expert review process when the parameter estimates given by different patterns differ too much, avoiding decision bias caused by incorrect fusion. The calculation of leakage pattern learning rate and degradation rate uses a rolling time window mechanism, with the window length dynamically adjusted according to the leakage evolution speed. In the fast-changing stage, a short window is used to capture details, while in the stable stage, a long window is used to improve stability. The parameter output is accompanied by confidence interval estimation, which calculates the probability distribution of learning rate and degradation rate through Bootstrap resampling method, providing uncertainty information for subsequent decision-making. The pattern evolution analyzer integrates anomaly detection function, when the input sequence has a similarity lower than the threshold with any historical pattern, it is automatically marked as a new type of leakage pattern and a special processing process is started, at the same time detailed data is collected for subsequent model enhancement.

[0082] In one embodiment, embodiment 4: when the prediction feature generation unit identifies the most similar classified leakage mode to the current leakage event features as the reference leakage mode, a multi-dimensional similarity evaluation framework is adopted. Suppose a top flange sealing surface corrosion leakage occurs in a storage tank of a chemical plant, and the current leakage event feature vector contains initial pressure 0.8 MPa, medium ethanol, diffusion angle 120 degrees, and other parameters. The system retrieves three candidate reference modes from the historical mode library: mode A (gasket aging leakage), mode B (bolt fracture leakage), and mode C (flange surface corrosion leakage). The matching degree calculation adopts the feature weighted Hamming distance method, and the distance values are 0.35, 0.58, and 0.12 respectively. Finally, the mode with the smallest distance, mode C, is selected as the reference leakage mode. When calculating the feature deviation of the current leakage event features and the reference mode standard feature information, six key dimensions are quantitatively compared: leakage pressure deviation +0.2 MPa (current value 1.0 MPa minus reference value 0.8 MPa), diffusion angle deviation -15 degrees, temperature gradient deviation +3°C / m, medium concentration deviation -12%, leakage hole diameter change rate deviation +0.1 mm / min, and diffusion rate deviation +0.3 m / s.

[0083] When setting the prediction correction coefficient based on the feature deviation, a deviation-correction coefficient mapping rule library is established. This rule library stores the correction strategies corresponding to different deviation ranges, for example, when the pressure deviation exceeds 0.15 MPa, a nonlinear correction function is triggered, and the diffusion angle deviation uses segmented linear interpolation adjustment. For the pressure deviation +0.2 MPa of the current case, the pressure correction coefficient 1.25 is obtained by looking up the table; the diffusion angle deviation -15 degrees corresponds to the angle correction coefficient 0.85. The final generated prediction correction coefficient matrix contains six dimensions of adjustment parameters, and each parameter is fused into a comprehensive correction coefficient 0.98 by weighted average. When using the prediction correction coefficient to dynamically adjust the leakage mode learning rate and the leakage mode degradation rate, the adjustment algorithm adopts a double-channel adjustment mechanism: the original learning rate value 0.75 multiplied by the comprehensive correction coefficient 0.98 gives the adjusted value 0.735; the original degradation rate value 0.05 divided by the correction coefficient 0.98 gives the adjusted value 0.051. When using the adjusted learning rate and degradation rate to drive the trend extrapolation process, the extrapolation engine uses a time convolutional neural network architecture, the network input layer injects the adjusted parameter value, the hidden layer contains 12 causal convolution modules to handle the time dependence relationship, and the output layer generates the prediction data for the next 30 minutes. For the flange corrosion leakage case, the prediction output shows that the leakage hole diameter will expand to 112% of the baseline prediction value in 15 minutes, and the diffusion range will increase by 18% of the area coverage compared to the baseline prediction.

[0084] Table 1: Feature Deviation and Correction Coefficient Mapping

[0085]

[0086] Referring to Table 1, when the blocking decision generation module obtains the sample leakage event feature set and the sample decision correlation coefficient set, 20 similar leakage event data in the past five years are extracted from the historical database. The sample data set includes the leakage source feature vector (location, aperture, pressure, etc.), the leakage diffusion feature vector (range, concentration, rate, etc.), and the corresponding decision correlation coefficient (0-1 value representing decision applicability) of each event. When constructing the decision correlation weight learner, a support vector regression model is selected as the core algorithm, the input layer of the model receives the standardized sample leakage source feature similarity depth coefficient and the sample leakage diffusion feature similarity depth coefficient, and the output layer corresponds to the sample decision correlation coefficient. The training process uses a radial basis kernel function, the regularization parameter C is set to 5.0, the kernel coefficient gamma is 0.01, and the super parameter combination is optimized through 10-fold cross-validation. Before inputting the current obtained leakage source feature information similarity depth coefficient (0.72) and leakage diffusion feature information similarity depth coefficient (0.65) into the trained weight learner, input data normalization processing is implemented to map the original value to the [-1, 1] interval. When the decision correlation weight learner outputs the decision correlation coefficient, the internal calculation process includes feature space transformation and nonlinear regression prediction. For the input value (0.72, 0.65), the correlation coefficient 0.68 is output, which is converted to the [0, 1] effective interval through the Sigmoid function.

[0087] The prediction correction coefficient setting process implements real-time feedback calibration. When the trend extrapolation result deviates from the actual monitoring data by more than 10%, the update process of the correction rule library is automatically triggered. The new rule generation uses case-based reasoning technology to retrieve the correction strategy of similar deviation situations in historical data, and after expert confirmation, it is supplemented to the rule library. The leakage mode learning rate adjustment mechanism sets a safety boundary limit. When the adjusted learning rate exceeds the threshold value 0.8, the artificial review program is forcibly started to avoid excessive correction leading to prediction distortion. The trend extrapolation network implements architecture optimization. An attention mechanism layer is added after the convolution module to automatically enhance the feature weight of the key time node and improve the prediction sensitivity to the leakage mutation point.

[0088] The decision correlation weight learner establishes a model drift monitoring mechanism, regularly calculates the moving average of the prediction error, and automatically starts the model retraining process when the prediction error exceeds 0.15 for five consecutive times. The retraining data selection adopts an active learning method, which preferentially selects historical samples closest to the current leakage characteristics for incremental training. An anomaly value filtering function is added to the input data processing link, which automatically switches to a backup calculation channel to regenerate the coefficient when a similar depth coefficient deviates from the normal distribution range. The output result explanation module generates a contribution analysis report of the decision correlation coefficient, which intuitively shows the influence weight proportion of the leakage source characteristics and the diffusion characteristics on the final decision, providing a decision basis for the operator. In the practical application of the flange corrosion leakage case, the system performs a full-process update every two minutes: collects the latest sensor data to recalculate the feature deviation, dynamically updates the prediction correction coefficient matrix, refreshes the trend extrapolation result and generates a new decision correlation coefficient. When the leakage pressure suddenly increases to 1.3 MPa at the eighth minute, the recalculated feature deviation causes the comprehensive correction coefficient to drop to 0.92, which in turn causes the predicted leakage hole diameter expansion rate to increase to 125% of the baseline value. The updated decision correlation coefficient is adjusted to 0.62, triggering the dynamic regeneration process of the plugging decision scheme. The entire correction process is completed within 800 milliseconds, ensuring the system's rapid response capability to changes in leakage conditions.

[0089] The training data of the weight learner implements version management, saving historical training set snapshots each time the system is upgraded, allowing rollback to the previous stable version when the prediction performance decreases. The model inference process adopts a lightweight deployment scheme, converting the support vector regression model into an inference graph executable by embedded devices through operator fusion technology, meeting the real-time requirements of on-site edge computing. The decision correlation coefficient output interface provides data verification functions, automatically issuing a system self-check alarm when the calculation result fluctuates by more than 0.2 for three consecutive times, indicating possible sensor data anomalies or model failure risks.

[0090] In one embodiment, the real-time data update feedback module continuously collects updated leakage sensor data streams during the disposal process of the chemical plant storage tank flange corrosion leakage. In the initial stage, the data acquisition frequency is set to 1 per second, and when the leakage pressure change rate exceeds 0.05 MPa / s, it is automatically increased to 5 per second. The sensor network includes 32 pressure sensors, 16 gas concentration detectors, and 8 thermal imagers. The updated data stream transmission uses a time-sensitive network protocol to ensure millisecond-level synchronization accuracy. The dynamic decision adjustment module packages the updated sensor data into data frames and pushes them to the leakage feature extraction module through an event-triggered mechanism, triggering real-time updating when detecting pressure mutations exceeding thresholds or concentration gradients being abnormal.

[0091] The leakage feature extraction module receives the pressure data frame of the 8th minute (containing a sudden pressure surge to 1.3 MPa recorded by the flange area sensor P07), the tank physical model re-executes the finite element calculation, updates the leakage point pressure value in the leakage source feature information and corrects the leakage flow rate vector. The diffusion feature information is updated synchronously: according to the data of the concentration detector C12-C15, the diffusion direction angle is recalculated from 120 degrees to 105 degrees, and the diffusion rate is increased from 1.8 m / s to 2.3 m / s. The feature extraction process implements incremental calculation optimization, only the local grid of the tank area affected by the updated data is recalculated, and the calculation time is compressed from 12 seconds of full model update to 1.8 seconds. The leakage trend prediction module starts the prediction regeneration process based on the updated feature information, the historical leakage data index unit retrieves three high-pressure sub-cases under mode C, and the leakage mode learning unit dynamically adjusts the learning rate parameter to 0.68 (original value 0.75). The prediction feature generation unit shortens the prediction time window from 30 minutes to 15 minutes, and outputs new predicted leakage source feature information: the leakage orifice diameter expansion rate is expected to increase to 0.45 mm per minute (original prediction 0.36 mm), and the leakage pressure decay curve slope decreases by 22%. The updated prediction leakage diffusion feature information shows that the original predicted 15-minute diffusion range radius of 25 meters is corrected to 32 meters, and the dangerous area of the vapor cloud explosion probability thermodynamic map is enlarged by 18%.

[0092] After receiving the updated prediction feature information, the plugging decision optimization module first performs topological reconstruction on the optimization decision space: adds a high-pressure working condition node in the decision graph, and adjusts the state transition conditions of the glue injection operation and the mechanical clamp installation. Based on the model predictive control framework, the multi-objective optimization problem is re-solved in the current decision time domain (8-12 minutes): the decision variable adjusts the glue injection rate from 5 liters per minute to 7 liters per minute, and the mechanical clamp locking torque gradient increases from 50 Nm / s to 70 Nm / s. The generated compensation plugging decision scheme adds two operation instructions: deploying a pre-glue point on the upper side of the flange, and expanding the operator safety distance from 30 meters to 45 meters. The final optimized plugging decision scheme is reorganized through dynamic iterative adjustment: the original scheme's glue injection operation at the 9th minute is advanced to the 8th minute and 30 seconds, and the clamp installation position at the 10th minute is offset by 15 degrees against the wind direction. The plugging execution control module converts the updated decision scheme into equipment control instructions: sends flow adjustment instructions to the automatic glue injection machine, increases the output of the No. 3 glue injection head by 40%; controls the mechanical arm to re-plan the motion trajectory to avoid the newly added dangerous gas concentration area; the remote-controlled robot receives the new path point sequence, the detour distance increases by 8 meters but avoids the predicted high-risk area. The execution process implements dual-channel monitoring: industrial bus transmits main control instructions, and wireless Mesh network transmits safety monitoring instructions synchronously, and automatically activates the backup operation mode when the robot approaches the concentration warning value.

[0093] The real-time data update feedback module continuously collects execution device state data during the implementation phase: actual flow feedback of the glue injection machine (6.8 liters per minute), mechanical clamp pressure sensor reading (68% of the target value), and robot position coordinates (X: 23.7, Y: 15.2). These execution data, together with the sensing data, form a closed-loop feedback flow. When the dynamic decision adjustment module detects that the clamp pressure build-up rate is lower than expected, it immediately triggers a local decision adjustment: under the framework of the overall scheme, a temporary pressure compensation step (increasing the hydraulic auxiliary device) is inserted. This adjustment is completed within 300 milliseconds and is sent to the execution terminal. The leakage feature extraction module implements data fusion verification in continuous updating. When the data difference between pressure sensor P07 and adjacent sensor P08 exceeds 10%, the sensor cross-verification program is automatically started. The verification process calls the thermal imager data to reconstruct the temperature field distribution, combines the material thermal expansion model to invert the real pressure value, and finally uses the weighted average algorithm to generate reliable data input. The trend prediction module establishes a prediction confidence decay model. After each prediction update, the initial confidence value is reset, and then it is automatically decayed by 15% per minute. When the confidence is less than 60%, a risk warning label is added to the prediction output. The rolling optimization process of the decision optimization module uses a sliding time window mechanism. It advances the optimization time domain every 30 seconds, and each optimization retains the feasible solution of the previous optimization as the initial population. The execution control module is equipped with an instruction conflict detector. When there is resource competition between dynamic adjustment instructions and the original plan (such as the intersection of the motion trajectories of two robots), it automatically inserts an avoidance waiting instruction and recalculates the time margin. All adjustment operations are recorded in the decision audit log, including the adjustment trigger reason, change content, and execution effect feedback, forming a complete closed-loop verification chain. When the leakage pressure is monitored to be stable at 1.2 MPa at the 12th minute, the system automatically reduces the data acquisition frequency to 2 times per second, and the decision adjustment mode switches from emergency response to standard optimization. At this time, the cumulative execution decision adjustment is 4 times, and the scheme change involves 8 parameter updates and 3 operation sequence reorganizations, but the core plugging strategy remains consistent. The execution device state data shows that an effective plugging isolation zone has been formed around the leakage point, and the diffusion concentration has dropped below the safety threshold, and the system enters the plugging effect maintenance phase. The dynamic decision adjustment module tracks resource consumption throughout the treatment process and updates the material inventory status in real time: the sealant consumption reaches 85% of the estimated amount, and the number of mechanical clamps used increases by 1 set compared to the initial scheme. When the standby material inventory approaches the warning line, an automatic purchase application is generated and sent to the supply chain system, and the material allocation coefficient in the subsequent maintenance phase is adjusted. The time window constraint dynamic expansion mechanism allows the saved time resources to be reallocated to the safety monitoring link after the leakage control has made progress.

[0094] The above merely describes preferred embodiments of the present application, and is not intended to limit the patent scope of the present application, and any equivalent structure or equivalent process conversion made by using the content of the present application specification and drawings, or directly or indirectly applied in other related technical fields, are also included in the patent protection scope of the present application.

Claims

1. A quick plugging emergency system for chemical storage tank leakage, characterized in that, The system comprises: a leakage sensing data acquisition module configured to acquire a leakage sensing data stream from a plurality of sensors arranged around a chemical storage tank and near a leakage point in real time; a leakage feature extraction module configured to perform leakage feature analysis according to preset leakage physical constraint information and in combination with the leakage sensing data stream, to determine leakage source feature information and leakage diffusion feature information; a plugging decision generation module configured to generate an initial plugging decision scheme based on an embedded plugging decision space and plugging fitness constraint conditions, in combination with the leakage source feature information and the leakage diffusion feature information; a leakage trend prediction module configured to perform leakage trend prediction on the leakage source feature information and the leakage diffusion feature information, to obtain predicted leakage source feature information and predicted leakage diffusion feature information; a plugging decision optimization module configured to perform plugging decision compensation optimization on the initial plugging decision scheme based on the predicted leakage source feature information and the predicted leakage diffusion feature information, to generate an optimized plugging decision scheme; a plugging execution control module configured to drive a corresponding plugging execution device to perform a leakage plugging operation based on the optimized plugging decision scheme; the plugging decision optimization module is specifically configured to: perform secondary correlation calibration on the plugging decision space according to the predicted leakage source feature information and the predicted leakage diffusion feature information, to obtain an optimized decision space; perform iterative decision search based on the plugging fitness constraint conditions and the optimized decision space, to generate a compensation plugging decision scheme; fuse and optimize the compensation plugging decision scheme and the initial plugging decision scheme, to generate the optimized plugging decision scheme; the plugging decision generation module is specifically configured to: traverse the plugging decision space to extract candidate plugging decision records; calculate a similarity depth coefficient of the leakage source feature information and sample leakage source feature information in a candidate record; calculate a similarity depth coefficient of the leakage diffusion feature information and sample leakage diffusion feature information in a candidate record; weight and fuse the two similarity depth coefficients according to preset weights, to generate a decision correlation coefficient; select candidate plugging decision records whose decision correlation coefficients satisfy a correlation threshold to constitute the initial plugging decision scheme; when calculating the similarity depth coefficient of the leakage source feature information and the sample leakage source feature information in a candidate record, a multi-scale feature matching strategy is adopted: for a leakage point position feature, a tank surface geodesic distance algorithm is used to calculate a spatial similarity; for a leakage aperture feature, a dynamic time warping algorithm is applied to align a current aperture change curve with a historical sample curve; for a leakage flow rate vector, a weighted combination value of a direction cosine similarity and a length relative error is calculated; and finally, the similarity depth coefficient is a geometric mean of the spatial similarity of the leakage point position feature, the curve alignment similarity of the leakage aperture feature, and the direction and length combination similarity of the leakage flow rate vector. When the similarity depth coefficient of the leakage diffusion characteristic information and the candidate sample leakage diffusion characteristic information is calculated, a regional comparison analysis is performed, the diffusion region is divided into a near-field core region, a middle-field transition region and a far-field edge region, the core region adopts a structural similarity index of concentration gradient distribution, the transition region calculates a Fréchet distance of a diffusion front motion track, the edge region evaluates a Jaccard coefficient of a dangerous region overlap rate, and the comprehensive coefficient is aggregated through regional importance weights.

2. The chemical tank leakage quick plugging emergency system according to claim 1, characterized in that, The leakage feature extraction module comprises: A storage tank basic information modeling unit is configured to obtain a set of storage tank basic information of a target chemical storage tank, and construct a storage tank physical model based on the set of storage tank basic information; A leakage feature fusion analysis unit is configured to input the leakage sensing data stream into the storage tank physical model to generate a leakage feature fusion atlas; A leakage feature information confirmation unit is configured to finally confirm the leakage source characteristic information and the leakage diffusion characteristic information based on the preset leakage physical constraint information and the leakage feature fusion atlas.

3. The chemical tank leakage quick plugging emergency system according to claim 1, characterized in that, The leakage trend prediction module comprises: A historical leakage data indexing unit is configured to index historical leakage characteristic data of a historical similar leakage event according to the leakage source characteristic information and the leakage diffusion characteristic information; A leakage mode learning unit is configured to perform leakage mode learning analysis according to the historical leakage characteristic data to obtain a leakage mode learning rate and a leakage mode degradation rate; A predicted feature generation unit is configured to combine the leakage mode learning rate and the leakage mode degradation rate to perform trend extrapolation on the leakage source characteristic information and the leakage diffusion characteristic information to generate the predicted leakage source characteristic information and the predicted leakage diffusion characteristic information; The leakage mode learning rate represents an attenuation rate of leakage intensity per unit time affected by plugging operation, and the leakage mode degradation rate represents a natural leakage rate increasing gradient caused by material fatigue; The leakage mode learning unit is specifically configured to: classify leakage event information in the historical leakage characteristic data to obtain a plurality of classified leakage modes; extract historical leakage characteristic parameter sequences in each classified leakage mode; perform feature dimension reduction and time sequence arrangement on each historical leakage characteristic parameter sequence to obtain a plurality of leakage characteristic evolution sequences; perform mode evolution analysis according to the plurality of leakage characteristic evolution sequences to calculate the leakage mode learning rate and the leakage mode degradation rate; The leakage mode learning unit performs mode evolution analysis according to the plurality of leakage characteristic evolution sequences, which comprises: constructing a mode evolution analyzer, wherein an input of the mode evolution analyzer is a set of historical sample leakage characteristic evolution sequences, and an output of the mode evolution analyzer is a set of historical sample leakage mode learning rates and a set of historical sample leakage mode degradation rates; inputting the plurality of currently obtained leakage characteristic evolution sequences into the mode evolution analyzer for classification; weighting and fusing a plurality of mode learning rates and a plurality of mode degradation rates output by the classification according to a similarity between a current leakage event characteristic and each classified leakage mode to finally obtain the leakage mode learning rate and the leakage mode degradation rate; The predicted feature generation unit combines the leakage mode learning rate and the leakage mode degradation rate to perform trend extrapolation, which comprises: Identify the classified leakage mode with the highest matching degree to the current leakage event characteristics as the reference leakage mode; Calculate the feature deviation amount of the current leakage event characteristics and the standard feature information of the reference leakage mode; Set a prediction correction coefficient based on the feature deviation amount; Dynamically adjust the leakage mode learning rate and leakage mode degradation rate using the prediction correction coefficient; Drive the trend extrapolation process using the adjusted learning rate and degradation rate; The driving trend extrapolation includes: the extrapolation engine adopts a time convolution neural network architecture, the network input layer injects the adjusted learning rate and degradation rate, the hidden layer contains 12 causal convolution modules to process the time dependence relationship, and the output layer generates the prediction data of the next 30 minutes.

4. The chemical tank leakage quick plugging emergency system according to claim 1, characterized in that, The plugging decision generation module includes: Obtain a sample leakage event feature set and a sample decision correlation coefficient set; Construct a decision correlation weight learner, whose input is the sample leakage source feature similarity depth coefficient and the sample leakage diffusion feature similarity depth coefficient, and whose output is the sample decision correlation coefficient; Input the current obtained leakage source feature information similarity depth coefficient and leakage diffusion feature information similarity depth coefficient into the decision correlation weight learner; The decision correlation weight learner outputs the decision correlation coefficient.

5. The chemical tank leakage quick plugging emergency system according to claim 1, characterized in that, The system further includes: A real-time data update feedback module for continuously collecting updated leakage sensing data streams during the plugging execution process; A dynamic decision adjustment module for feeding back the updated leakage sensing data streams to the leakage feature extraction module, the leakage trend prediction module and the plugging decision optimization module, triggering dynamic iterative adjustment of the optimized plugging decision scheme.

Images