Methanol leakage prediction system based on artificial intelligence

Abstract

The invention relates to the technical field of methanol safety monitoring, in particular to a methanol leakage prediction system based on artificial intelligence. According to the method, structural extraction of potential leakage paths is realized through dynamic sorting and surge correlation combination of methanol concentration change trends in continuous time periods, and weighted calculation is carried out on conduction characteristics between monitoring nodes in double dimensions of time series and spatial topology by virtue of a graph attention mechanism, so that the detection accuracy of the leakage paths is improved. The method comprises the following steps: establishing a path, forming quantitative expression of node influence factors in the path, capturing synchronous abnormal behaviors under an environment disturbance background through a multi-parameter coupling sudden change identification mode of air pressure and wind speed, further constructing a high-timeliness sudden change event set, and performing index fusion on the influence degree and sudden change frequency of monitoring nodes on the basis to obtain a high-timeliness sudden change event set. And space risk score distribution is formed, and a leakage possibility region prediction map is generated in combination with two-dimensional coordinates, so that boundary fine depiction of a high-risk region is realized.

Description

Technical Field

[0001] The present invention relates to the technical field of methanol safety monitoring, and in particular to a methanol leakage prediction system based on artificial intelligence. Background Art

[0002] The technical field of methanol safety monitoring is a technical category that involves dynamically perceiving and responding to potential hazards caused by methanol during production, storage, transportation, and application. It mainly constructs a continuous perception mechanism for the state of methanol in the working environment, identifies characteristic indicators during the abnormal evolution process, and accordingly conducts risk analysis and disposal judgment.

[0003] Among them, the traditional methanol leakage prediction system refers to a device or method used to judge the possible leakage trend of methanol based on weak changes in relevant environmental parameters when visible leakage of methanol has not occurred. Before methanol reaches the alarm threshold, early identification of potential leakage signs is carried out to achieve pre-judgment of leakage risks.

[0004] The prior art relies on the initial reaction of the change trend of environmental parameters before reaching the alarm threshold for potential leakage pre-judgment. In actual operation, there are problems such as strong dependence and large disturbance interference in the mutation judgment of single-point monitoring data. Especially in scenarios with frequent meteorological disturbances or large fluctuations in background concentration, short-term fluctuations are easily misjudged as leakage signals, resulting in an increase in the false alarm rate. At the same time, it lacks the ability to synergistically analyze the temporal and spatial relationships between multiple monitoring points, fails to effectively construct the correlation structure during the transmission process of leakage events, and also fails to cross-identify multi-source mutations between environmental parameters, resulting in a lag in the response to complex abnormal events. Moreover, its leakage pre-judgment mostly stays in the qualitative stage and cannot form a spatial prediction result of the risk area, making the deployment of emergency response lack a clear positioning basis and affecting the prevention and control efficiency of leakage accidents. Summary of the Invention

[0005] The purpose of the present invention is to solve the drawbacks existing in the prior art, and to propose a methanol leakage prediction system based on artificial intelligence.

[0006] To achieve the above purpose, the present invention adopts the following technical solution: A methanol leakage prediction system based on artificial intelligence includes: Leakage path construction module: Obtain the methanol concentration of each monitoring node within a continuous time period, screen and sort the monitoring nodes whose methanol concentration changes exceed the increase threshold in adjacent time periods, and construct a methanol leakage propagation path diagram; Propagation feature extraction module: Perform monitoring node feature aggregation operations on all paths in the methanol leakage propagation path diagram through a graph attention network to obtain monitoring node propagation influence factors; Environmental response recognition module: Obtain the air pressure and wind speed at each monitoring node location within a continuous time period, screen the time periods when the air pressure and wind speed undergo synchronous mutations, calibrate them as the occurrence points of mutation events, and obtain a set of coupled mutation events; Risk score calculation module: Calculate the leakage risk score of the monitoring node by weighted calculation according to the propagation influence factor of the monitoring node and the occurrence frequency of the corresponding mutation event in the set of coupled mutation events; Spatial prediction generation module: According to the leakage risk score of the monitoring node, extract the boundary of the area with the possibility of leakage with reference to the two-dimensional coordinates of the monitoring node, and construct a methanol leakage prediction map.

[0007] As a further solution of the present invention, the methanol leakage propagation path map includes a sequence of surge monitoring nodes, a directed connection relationship, and a propagation time weight. The propagation influence factor is specifically the importance weight of the monitoring node, the cumulative effect value of the path, and the adjacent influence coefficient. The set of coupled mutation events includes the mutation event location, the synchronous mutation time period, and the event statistical frequency. The leakage risk score is specifically the risk level of the monitoring node, the scoring calculation weight, and the spatial scoring distribution. The methanol leakage prediction map includes a leakage probability field, a leakage boundary area, and leakage spatial coordinates.

[0008] As a further solution of the present invention, the leakage path construction module includes: Concentration acquisition sub-module: Obtain the methanol concentration of each monitoring node within a continuous time period, detect whether the change in the methanol concentration of each monitoring node in the time series exceeds the increase threshold, extract the first occurrence moment and record it, and generate the surge time data of the monitoring node; Surge sorting sub-module: According to the surge time data of the monitoring node, sort all the monitoring nodes in ascending order of time, judge the surge sequence between the monitoring nodes, and screen the adjacent monitoring node combinations in the time sequence relationship to generate the sorted monitoring node pair information; Path generation sub-module: According to the surge time interval of each monitoring node combination in the sorted monitoring node pair information, establish a methanol leakage propagation path map.

[0009] As a further solution of the present invention, the propagation feature extraction module includes: Path reading sub-module: Extract all the monitoring path sequences in the methanol leakage propagation path map, identify the arrangement order and connection direction of the monitoring nodes in each path, and generate the monitoring path topological relationship information; Upstream and downstream aggregation sub-module: According to the monitoring path topological relationship information, call the feature aggregation mechanism of each monitoring node in the graph attention network, integrate the time surge sequence and connection direction of adjacent monitoring nodes, and update according to the monitoring node structure within the aggregation range to generate the monitoring node association features; Weight calculation sub-module: According to the feature vectors of each monitoring node in the monitoring node association features and the connection directions of adjacent connection edges, weighted processing is performed through the graph attention mechanism to calculate the influence degree of each monitoring node in the path, and a monitoring node propagation influence factor is generated.

[0010] As a further solution of the present invention, the environmental response recognition module includes: Data acquisition sub-module: Obtain the air pressure and wind speed at the position of each monitoring node within a continuous time period, divide the two types of environmental parameters by monitoring node respectively, and perform time alignment according to a unified time resolution to generate a monitoring node environmental parameter sequence; Mutation detection sub-module: According to the monitoring node environmental parameter sequence, perform change trend recognition on the air pressure sequence and wind speed sequence of each monitoring node by means of double sliding window change point detection to generate a monitoring node synchronous mutation period; Event calibration sub-module: According to the monitoring node synchronous mutation period, assign a mutation event mark to the corresponding monitoring node position, construct a mutation event item according to the monitoring node position and mutation period, and generate a coupled mutation event set.

[0011] As a further solution of the present invention, the risk score calculation module includes: Location matching sub-module: Extract the monitoring node propagation influence factor and the monitoring node positions in the coupled mutation event set, identify the monitoring nodes with consistent spatial positions, and generate a monitoring node overlapping position list; Index extraction sub-module: According to the monitoring node overlapping position list, extract the propagation influence factor value and the mutation event occurrence frequency of the corresponding monitoring node respectively to generate a monitoring node combined index; Score generation sub-module: According to the monitoring node combined index, perform weighted processing on the propagation influence factor value and the mutation event occurrence frequency of each monitoring node position, and combine the corresponding index of the monitoring node to perform numerical calculation on its leakage risk level to generate a monitoring node leakage risk score.

[0012] As a further solution of the present invention, the spatial prediction generation module includes: Coordinate mapping sub-module: Extract the two-dimensional coordinate positions corresponding to each monitoring node, pair and organize the monitoring node leakage risk score and coordinates, construct a spatial score data set indexed by coordinates, and generate spatial risk score coordinate data; Risk probability estimation sub-module: According to the spatial risk score coordinate data, with the monitoring node coordinates as the center, construct a risk probability distribution layer in the continuous spatial range to generate a spatial risk probability field; Boundary extraction sub-module: Based on the risk probability distribution of each spatial point in the spatial risk probability field, identify the edge positions of the mutation regions and generate a methanol leakage prediction map.

[0013] Compared with the prior art, the advantages and positive effects of the present invention are as follows: In the present invention, through the dynamic sorting and surge correlation combination of the methanol concentration change trends within consecutive time periods, the structured extraction of potential leakage paths is achieved. By means of the graph attention mechanism, the conduction characteristics between monitoring nodes are weighted and calculated in both the time series and spatial topology dimensions, forming a quantitative expression of the influence factors of the nodes within the paths. At the same time, through the multi-parameter coupling mutation identification method of air pressure and wind speed, the synchronous abnormal behaviors under the background of environmental disturbances are captured, and then a high-timeliness mutation event set is constructed. On this basis, the influence degree and mutation frequency of the monitoring nodes are fused as indicators to form a spatial risk score distribution, and combined with two-dimensional coordinates to generate a prediction map of the leakage possibility area, realizing the fine characterization of the boundaries of high-risk areas. This method breaks through the limitations of single-index judgment and static threshold setting, improves the spatial recognition accuracy, risk level discrimination, and response time sensitivity of methanol leakage early warning, and effectively enhances the intelligent recognition ability and spatial prediction accuracy of weak leakage signs under complex working conditions. Description of the Drawings

[0014] Figure 1 is the system flow chart of the present invention; Figure 2 is the flow chart of the leakage path construction module of the present invention; Figure 3 is the flow chart of the propagation feature extraction module of the present invention; Figure 4 is the flow chart of the environmental response recognition module of the present invention; Figure 5 is the flow chart of the risk score calculation module of the present invention; Figure 6 is the flow chart of the spatial prediction generation module of the present invention. Detailed Embodiments

[0015] In order to make the objectives, technical solutions, and advantages of the present invention clearer and more understandable, the present invention will be further described in detail below with reference to the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are only used to explain the present invention and are not used to limit the present invention.

[0016] In the description of the present invention, it should be understood that the orientation or positional relationship indicated by the terms "length", "width", "upper", "lower", "front", "rear", "left", "right", "vertical", "horizontal", "top", "bottom", "inner", "outer", etc. is based on the orientation or positional relationship shown in the drawings. It is only for the convenience of describing the present invention and simplifying the description, rather than indicating or implying that the device or element referred to must have a specific orientation, be constructed and operated in a specific orientation, and therefore should not be construed as a limitation on the present invention. In addition, in the description of the present invention, the meaning of "a plurality of" is two or more, unless otherwise specifically defined.

[0017] Please refer to Figure 1 , a methanol leakage prediction system based on artificial intelligence includes: Leakage path construction module: Obtain the methanol concentration of each monitoring node within a continuous time period, screen and sort the monitoring nodes whose methanol concentration changes exceed the increase threshold in adjacent time periods, and construct a methanol leakage propagation path diagram; Propagation feature extraction module: Perform monitoring node feature aggregation operations on all paths in the methanol leakage propagation path diagram through a graph attention network to obtain the propagation influence factors of the monitoring nodes; Environmental response identification module: Obtain the air pressure and wind speed at the location of each monitoring node within a continuous time period, screen the time periods when the air pressure and wind speed have synchronous mutations, calibrate them as the occurrence points of mutation events, and obtain a set of coupled mutation events; Risk score calculation module: Calculate the leakage risk score of the monitoring node by weighted calculation according to the propagation influence factor of the monitoring node and the occurrence frequency of the corresponding mutation event in the set of coupled mutation events; Spatial prediction generation module: According to the leakage risk score of the monitoring node, extract the boundary of the area where leakage is likely with reference to the two-dimensional coordinates of the monitoring node, and construct a methanol leakage prediction map; The methanol leakage propagation path diagram includes a sequence of surge monitoring nodes, a directed connection relationship, and a propagation time weight. The propagation influence factor is specifically the importance weight of the monitoring node, the cumulative effect value of the path, and the adjacent influence coefficient. The set of coupled mutation events includes the location of the mutation event, the synchronous mutation time period, and the event statistical frequency. The leakage risk score is specifically the risk level of the monitoring node, the weight of the score calculation, and the spatial score distribution. The methanol leakage prediction map includes a leakage probability field, a leakage boundary area, and leakage spatial coordinates.

[0018] Please refer to Figure 2 , the leakage path construction module includes: Concentration acquisition sub-module: Obtain the methanol concentration of each monitoring node within a continuous time period, detect whether the change in the methanol concentration of each monitoring node in the time series exceeds the increase threshold, extract and record the first occurrence time, and generate the surge time data of the monitoring node; Obtain the methanol concentration of each monitoring node within a continuous time period. The obtaining method is to regularly read the data of the gas sensor installed on the monitoring node at a set time frequency. For example, read the methanol concentration value once per minute, and continuously read to form the time series data of the monitoring node. Next, calculate the concentration difference between each time point in this time series and the previous time point. The method is to subtract the concentration of the previous time point from the concentration of the current time point to form a concentration increment sequence. On this basis, set an increase threshold, which can be set by adding twice the standard deviation to the statistical mean of historical data. For example, through the statistical analysis of historical monitoring data in a certain area, it is found that the average value of the concentration change per minute is 3 ppm, and the standard deviation is 2.5 ppm. Then the increase threshold can be set to 3 + 2×2.5 = 8 ppm. In actual operation, it can be rounded up and set to 10 ppm, indicating that when the concentration of a certain monitoring node increases by more than 10 ppm within one minute, it is regarded as a surge event. If the reading of a certain monitoring node at the 15th minute is 38 ppm and at the 14th minute is 26 ppm, then the increment is 12 ppm, which is greater than the set threshold of 10 ppm, so it is determined that the surge first appears, and record this moment as the surge time of this monitoring node. The same processing is performed on all monitoring nodes, and the output is formed into a table. Each record includes the monitoring node number, the surge time, and the methanol concentration value at that time. For example, if the first surge of monitoring node 7 occurs at 08:42 and the concentration is 57 ppm, then the entry written into the table is "Monitoring node 7, 08:42, 57 ppm". In a monitoring system with 10 monitoring nodes deployed in a chemical industrial park, monitoring nodes 1 - 10 read data per minute and perform the above increment judgment, and form the monitoring node surge time data through threshold determination.

[0019] Surge sorting sub-module: According to the monitoring node surge time data, arrange all monitoring nodes in ascending order of time, judge the order of surges between monitoring nodes, and screen adjacent monitoring node combinations in the time sequence relationship to generate the sorted monitoring node pair information; Taking the time data of the surge of monitoring nodes as the input, first sort all the records in ascending order of time. Use the sorted function in Python or the "Ascending" function in Excel to sort by the surge time field, and place the monitoring node with the earliest surge at the front of the list. For example, there are records: Monitoring Node 4 - 08:41, Monitoring Node 2 - 08:38, Monitoring Node 6 - 08:45. Then the sorting result is Monitoring Node 2, Monitoring Node 4, Monitoring Node 6. Next, in the sorting result, combine each pair of adjacent monitoring nodes into a group from front to back, such as (Monitoring Node 2, Monitoring Node 4), (Monitoring Node 4, Monitoring Node 6). The practical significance of this operation is to establish a preliminary judgment combination of possible leakage diffusion paths in terms of time sequence. Then continue to write each pair of combination results into a data structure, which can be saved in formats such as a list of tuples, a CSV file, etc. The content includes the pair number of the monitoring nodes and their respective surge times. For example, one line of the record is: "Monitoring Node Pair (2, 4), Surge Time (08:38, 08:41)". In an actual scenario, if there are 12 monitoring nodes deployed in a petrochemical park and the surge times are recorded in the system respectively. When the surge times of Monitoring Node 5 and Monitoring Node 8 are 09:03 and 09:07 respectively and they are adjacent after sorting, the system will automatically combine them into a monitoring node pair and store it in the dataset for subsequent analysis.

[0020] Path generation sub-module: According to the surge time intervals of each monitoring node combination in the sorted monitoring node pair information, establish a methanol leakage propagation path diagram; For the combination sorted according to the surge time of the monitoring nodes, calculate the surge time interval between each pair of monitoring nodes and construct a leakage propagation path graph. First, read the surge time of each group of monitoring node combinations, calculate the time difference between the two, and record the propagation time weight in minutes. For example, in the monitoring node pair (3, 7), the surge time of monitoring node 3 is 08:40, and that of monitoring node 7 is 08:47, with an interval of 7 minutes. This value is used as the weight of the edge in the graph. The path graph is constructed using a directed graph model. The monitoring nodes represent the monitoring points, and the edges represent the paths through which the leakage may propagate from the monitoring node that surges first to the monitoring node that surges later. The attribute of the edge is the propagation weight. During the graph construction process, the NetworkX tool can be introduced or the edge table can be recorded in CSV format. For example, the edge record: "Source monitoring node: 3, Target monitoring node: 7, Propagation time: 7 minutes". In addition, the setting of the propagation time weight is based on the time difference itself, but a maximum time interval threshold can be set to filter out invalid combinations. For example, if the propagation interval is greater than 15 minutes, it is considered an irrelevant path. The propagation interval threshold can be set to 15 minutes, which is set based on the experience of the diffusion speed. For example, methanol diffuses within a range of 300 meters in about 25 minutes under a natural wind speed of 0.2 m / s. Therefore, the propagation interval in a medium-sized factory area generally does not exceed 15 minutes. Combining the above, if the time interval between the monitoring node pair exceeds this threshold, the edge will not be generated. For example, if the surge time interval between monitoring nodes A and B is 18 minutes, the combination will be directly discarded. Finally, a complete directed graph structure is constructed, representing the potential diffusion path graph formed by the surge order and time interval among multiple monitoring nodes.

[0021] Please refer to Figure 3 , the propagation feature extraction module includes: Path reading sub-module: Extract all the monitoring path sequences in the methanol leakage propagation path graph, identify the arrangement order and connection direction of the monitoring nodes in each path, and generate the monitoring path topology relationship information; Based on the methanol leakage propagation path graph, all directed edges and corresponding monitoring nodes in the graph are read in turn to extract each complete propagation path sequence. The path set can be identified through graph traversal algorithms such as DFS (depth-first search) or BFS (breadth-first search). For example, if there are edges (1→3), (3→5), and (5→7) in the path graph, the path sequence is identified as 1→3→5→7. When traversing, it is necessary to follow the path diffusion logic in ascending time order, read the connection relationship and arrangement order between the monitoring nodes, generate a path sequence list, save it as structured data, and record each path. It includes path number, monitoring node arrangement, and connection direction information. For example, path 001 is recorded as: [monitoring node 1→monitoring node 3→monitoring node 5→monitoring node 7], and path 002 may be [monitoring node 2→monitoring node 4→monitoring node 6]. In addition, during the processing, the adjacency matrix or edge table can be used to assist in storing the graph structure for efficient reading. In the actual park leakage simulation system, if a path graph containing 15 monitoring nodes and 38 edges is constructed, 10 valid paths can be extracted from it. Each path clearly represents the diffusion path of the monitoring node from the leakage source to the perception end point.

[0022] Upstream and downstream aggregation submodule: Based on the topological relationship information of the monitoring path, the feature aggregation mechanism of each monitoring node in the graph attention network is called to integrate the time surge order and connection direction of adjacent monitoring nodes, and the monitoring node structure within the aggregation range is updated to generate monitoring node related features; According to the path topology information, all monitoring nodes in the path graph are aggregated with information of adjacent monitoring nodes. The aggregation operation uses the graph attention network mechanism to fuse the feature vector of each monitoring node with the feature vectors of its upstream and downstream monitoring nodes according to the connection direction. First, the initial feature vector of the monitoring node is set. For example, each monitoring node uses its surge time (converted into a numerical value, such as 08:42 can be converted into 522 minutes), the length of the path, the number of upstream and downstream connections and other attributes to form a vector. For example, the feature of monitoring node 5 is [522, 3, 2]. When aggregating, the adjacent monitoring nodes are classified as upstream and downstream according to the connection direction, and the aggregation function such as weighted average is used for fusion. For example, if the upstream of monitoring node 7 is monitoring node 5 and the vector of monitoring node 5 is [522, 3, 2], the aggregated input of monitoring node 7 is the vector of monitoring node 5. At the same time, the attention weight is adjusted according to the connection direction. Combined with the structural position of the monitoring node, if it is in the middle or end of the path, its aggregation range only includes some neighbors. After aggregation, a new monitoring node representation is generated as the updated monitoring node association feature. In a factory environment, if a path is 1→3→6→8, monitoring node 6 will aggregate information from 3 and 8 and update its feature representation. After executing the aggregation mechanism operation on all monitoring nodes, a set of monitoring node vectors after the entire graph update can be generated.

[0023] Weight calculation sub-module: According to the feature vectors of each monitoring node in the monitoring node association features and the connection directions of the adjacent connection edges, weighted processing is performed through the graph attention mechanism to calculate the influence degree of each monitoring node in the path, and the monitoring node propagation influence factor is generated; According to the direction relationship between the monitoring node association feature vector generated after aggregation by the graph attention network and the connection edges in the path, weighted processing is performed, and finally the influence degree of each monitoring node in the methanol leakage path is calculated, and the propagation influence factor of the monitoring node is generated. The specific steps are as follows: First, extract the feature vector of each monitoring node. Let the monitoring node have a feature vector of , which is a -dimensional vector used to represent the structural state and local graph attributes of the monitoring node, such as surge time, path depth, adjacency degree, etc. The actual vector is like , representing the three-dimensional structure information of monitoring node 7. For each monitoring node and its adjacent monitoring node , it is necessary to judge their upstream and downstream relationships in the path propagation according to the connection direction, and calculate the attention weight of monitoring node to monitoring node , which represents the contribution degree of monitoring node to monitoring node in feature aggregation. The specific calculation uses the following scoring mechanism: ; where: : The original attention score of monitoring node to monitoring node , representing the correlation score before monitoring node aggregates the information of monitoring node ; : The transpose of the scoring vector, with a size of , used to map the concatenated features to a scalar; : The feature concatenation vector of monitoring node and monitoring node , with a dimension of ; : A learnable linear transformation matrix, with a size of , used to transform the concatenated vector; in a simplified scenario, the identity matrix can be taken; : An activation function, defined as when , otherwise , where is the leakage coefficient, usually taking a value of 0.2, used to retain negative information to avoid neuron dead zones.

[0024] After scoring, Softmax normalization is used to obtain the final attention weights: ; Where: : represents the set of adjacent monitoring nodes of the monitoring node ; : the final attention coefficient of the monitoring node to the monitoring node , indicating the proportion of the influence of the monitoring node in the feature aggregation on ; The exponential function ensures that all weights are positive and proportionally distributed, represents the sum operation for each monitoring node in the set of adjacent monitoring nodes belonging to the monitoring node , used to accumulate the attention scores of all neighbor monitoring nodes and the monitoring node .

[0025] Taking the monitoring node 7 as an example, assume its adjacent monitoring nodes are monitoring node 5 and monitoring node 9, and the feature vectors are respectively: , , .

[0026] After concatenation, score: , the score value is approximately ; , the score value is approximately .

[0027] After normalization: ; .

[0028] Subsequently, these weights are used to perform weighted aggregation on the features of neighbor monitoring nodes: ; The component calculation is as follows: The first dimension: ; The second dimension: ; The third dimension: ; The propagated influence feature of the monitoring node 7 is .

[0029] To quantify the influence degree of the monitoring node, the three-dimensional vector is synthesized into a propagated influence value, and the weighted average calculation method is adopted: 。

[0030] Among them, the weights are respectively assigned as [0.4, 0.4, 0.2]. The reason is that the surge time and path depth are more critical than the adjacency topology value during the leakage process, so higher weights are given.

[0031] Finally, this value is classified and judged according to the set level interval. The basis for level setting is as follows: Statistically analyze the distribution of all propagation influence values in the simulation samples. Assume the mean is 0.50 and the standard deviation is 0.12. The classification criteria are: high influence: influence value ≥ mean + 2σ ≈ 0.74; medium influence: mean ≤ influence value < mean + 2σ, that is, 0.50 - 0.74; low influence: influence value < mean ≈ 0.50.

[0032] Therefore, the influence value of monitoring node 7 is 0.547, which falls into the medium influence level. This value will be used to sort the priority processing order of each monitoring node in the leakage path and assist the scheduling system in determining risk monitoring nodes. This method quantifies the propagation role of monitoring nodes under the graph structure through feature fusion and direction attention mechanism, and has a stable structural interpretation ability.

[0033] Please refer to Figure 4 , the environmental response recognition module includes: Data acquisition sub-module: Obtain the air pressure and wind speed at the location of each monitoring node within a continuous time period, divide the two types of environmental parameters by monitoring node respectively, and perform time alignment according to a unified time resolution to generate a monitoring node environmental parameter sequence; Continuously obtain two types of environmental parameters from each monitoring node: air pressure and wind speed. The acquisition period needs to be consistent. For example, set it to record once every 30 seconds. The system will read the sensing data of environmental monitoring instruments on all monitoring nodes at this time frequency. For the data collected from each monitoring node, an air pressure sequence and a wind speed sequence are respectively established. Subsequently, time alignment processing is performed on the two parameters of the same monitoring node. If a certain monitoring node obtains 20 sets of sampling data between 08:00 and 08:10, and there are missing items or recording errors at some sampling time points, the missing data needs to be filled by linear interpolation or time synchronization mechanism to ensure that all sequences are constructed according to a unified time resolution. For example, for the case of missing wind speed value at 08:01, the mean value of 08:00 and 08:02 can be used for filling. After the processing is completed, the system will generate two time sequences for each monitoring node, respectively representing the changes in air pressure and wind speed at the location of the monitoring node in time alignment. In the actual application of the park, if 15 monitoring nodes are deployed and the sampling period is set to one minute, within 10 minutes, 20 air pressure data and 20 wind speed data of each monitoring node will be aggregated, and they will be sorted into an environmental parameter sequence database according to the monitoring node number respectively.

[0034] Mutation Detection Sub-module: Based on the environmental parameter sequences of the monitoring nodes, perform trend identification on the air pressure sequence and wind speed sequence of each monitoring node through the double-sliding window change point detection method, and generate the synchronous mutation time periods of the monitoring nodes. Receive the input of the environmental parameter sequences of each monitoring node, and perform trend analysis on the air pressure and wind speed respectively. Adopt the double-sliding window change point detection method, that is, set two windows to slide along the time axis, one representing historical observations and the other representing current observations, and perform significance analysis on the numerical differences within the two time windows. For example, set the window length to 5 minutes, calculate the average wind speed in the two intervals of 08:00–08:05 and 08:05–08:10 for monitoring node 1 respectively. If the average value of the previous section is 1.2 m / s and the latter section is 3.6 m / s, with an obvious increase, it is judged that a mutation has occurred. The calculation method can be the mean square error difference or the CUSUM algorithm. The mutation time is recorded as the intersection point time of the two windows, that is, 08:05. The system detects mutations in the air pressure sequence and wind speed sequence of each monitoring node respectively in this way. If both of them mutate within a similar time, mark this time period as the synchronous mutation time period. For example, the air pressure mutation time of monitoring node 4 is 08:33, and the wind speed mutation time is 08:34. The difference is less than the threshold of 1 minute, then it is determined that 08:33–08:34 is the synchronous mutation time period, and form the mutation records of the monitoring nodes and time periods.

[0035] Event Labeling Sub-module: According to the synchronous mutation time periods of the monitoring nodes, assign mutation event labels to the corresponding monitoring node positions, construct mutation event items according to the monitoring node positions and mutation time periods, and generate a coupled mutation event set. Based on the synchronous mutation time periods of each monitoring node output by the mutation detection sub-module, perform event-level annotation processing. For each marked time period and monitoring node position, the system constructs a "mutation event item", the content of which includes fields such as monitoring node number, time interval, mutation type, etc. For example, if synchronous mutations of air pressure and wind speed are detected for monitoring node 6 during 08:43–08:44, then generate an event item: "Monitoring node 6, 08:43–08:44, synchronous mutation". The system summarizes all event items to form a coupled mutation event set, which can be further used for spatial analysis, spatio-temporal traceability, etc. Under the industrial site layout conditions, if multiple monitoring nodes mutate simultaneously during a certain time period, multiple event items will be generated to form a dense mutation event set.

[0036] Please refer to Figure 5 The risk score calculation module includes: Location Matching Sub-module: Extract the monitoring node propagation influence factors and the monitoring node positions in the coupled mutation event set, identify the monitoring nodes with consistent spatial positions, and generate a list of overlapping positions of the monitoring nodes. Extract the propagation influence factors of the monitoring nodes and the location information of the monitoring nodes in the set of coupled mutation events. First, read two types of data generated in the previous module: one is the propagation influence factors of each monitoring node. This data is indexed by the monitoring node number and contains the numerical results of the risk propagation of each monitoring node. For example, the influence factor value of monitoring node 7 is 0.547. The other is the location information of the monitoring nodes corresponding to each event in the set of mutation events, such as "monitoring node 7, 08:43–08:44". Then, compare and match the monitoring node location fields in the two types of data one by one to determine whether the spatial coordinates are exactly the same or the monitoring node numbers are the same. If they are the same, mark the monitoring node as an overlapping monitoring node, and construct an overlapping list to record the numbers of such monitoring nodes. In practical applications, if the propagation influence factor set contains monitoring nodes 2, 5, 7, 9, and the set of mutation events contains monitoring nodes 3, 5, 7, 10, then after comparing the location consistency, it is determined that monitoring node 5 and monitoring node 7 are overlapping monitoring nodes, and a monitoring node overlapping location list is recorded: "monitoring node 5, monitoring node 7".

[0037] Index extraction sub-module: According to the monitoring node overlapping location list, extract the propagation influence factor values and the mutation event occurrence frequencies of the corresponding monitoring nodes respectively, and generate joint indicators for the monitoring nodes; According to the overlapping location list, read the two key data of each overlapping monitoring node one by one: the propagation influence factor value and the mutation event occurrence frequency. During the reading process, the two data table structures need to be searched respectively with the monitoring node number as the key to obtain their numerical attributes. The propagation influence factor value is represented by a floating point number indicating the risk level of the monitoring node. For example, for monitoring node 5 it is 0.682, and for monitoring node 7 it is 0.547. The mutation event frequency is the number of times this monitoring node appears in the previous set of mutation events. For example, monitoring node 5 appears 3 times and monitoring node 7 appears 2 times, that is, the frequencies are 3 and 2. Then, assemble these two data items according to the monitoring node number to form a joint indicator. For example, the joint indicator format is "monitoring node number–[influence factor, mutation frequency]". For example, monitoring node 5 is recorded as "monitoring node 5–[0.682, 3]", and monitoring node 7 is "monitoring node 7–[0.547, 2]". All overlapping monitoring nodes are processed in this way to form a complete set of joint indicators.

[0038] Scoring generation sub-module: According to the joint indicators of the monitoring nodes, perform weighted processing on the propagation influence factor values and the mutation event occurrence frequencies at each monitoring node location, and numerically calculate the leakage risk level of the monitoring node in combination with the corresponding indicators of the monitoring node to generate a leakage risk score for the monitoring node; According to the combined index data of each monitoring node, that is, the propagation influence factor value and the mutation event occurrence frequency of each node, perform weighted processing and numerical calculation to generate a leakage risk score and delimit the risk level. First, obtain the combined index provided by the previous module. For example, the index of node 5 is that the propagation influence factor value is 0.682 and the mutation frequency is 3, and the index of node 7 is the influence factor 0.547 and the frequency 2. To unify the data ranges of the two indexes, it is necessary to first standardize the mutation frequency to map its value to the interval [0, 1]. The linear normalization method is used for standardization. Let the minimum value of the mutation frequency among all overlapping nodes be , and the maximum value be . Then, for a certain node , the standardized frequency is . Substitute this value and the propagation influence factor into the risk score calculation formula: ; Among them: : represents the final leakage risk score of the monitoring node , which is the result output; : is the propagation influence factor value of the monitoring node , and the range is [0, 1]; : is the mutation event occurrence frequency of the monitoring node ; : is the weighting coefficient, which determines the contribution ratio of the two indexes in the score; : are respectively the minimum and maximum values of the mutation frequency among all nodes participating in the scoring in this round, which are used for frequency normalization.

[0039] The setting basis of the weighting coefficient is as follows: The larger the propagation influence factor value, the more critical the structural position of the node in the methanol leakage path, and the greater the contribution to the leakage diffusion path. Therefore, the weight should be higher. Although the mutation event frequency can reflect the local disturbance intensity, there are occasional interferences. Therefore, a smaller weight is given. So, the weight ratio is taken as , .

[0040] Suppose the mutation frequency interval of the current overlapping nodes is from 1 to 5, then , . For node 5, , , and its standardized frequency is: .

[0041] Substitute into the scoring formula to get: .

[0042] For node 7, , , and its standardized frequency is: ; 。

[0043] The basis for risk level classification is as follows: First, count all node scores. Assume the average score is 0.55 and the standard deviation is 0.15. Then the classification levels are as follows: High risk: (i.e., the average + 1σ), indicating significant propagation impact and active mutations; Medium risk: , which is the main area where scores are concentrated; Low risk: , with a low impact factor and few mutation frequencies. The final results are: Node 5 has a score of 0.627, which is a medium risk, and Node 7 has a score of 0.458, which is also a medium risk. The complete score result list includes each node number, calculated value, and classification level, such as "Node 5: 0.627 – Medium risk", "Node 7: 0.458 – Medium risk".

[0044] Please refer to Figure 6 , the space prediction generation module includes: Coordinate mapping sub-module: Extract the two-dimensional coordinate positions corresponding to each monitoring node, pair and organize the leakage risk scores of the monitoring nodes with the coordinates, construct a space score data set indexed by coordinates, and generate space risk score coordinate data; Extract the two-dimensional coordinate position data of each monitoring node, usually from the GIS coordinates set during the deployment of the monitoring system or the measured longitude and latitude information. For example, the coordinates of Node 1 are (120.3815, 36.0723), and those of Node 2 are (120.3821, 36.0730). The reading method can obtain the position from the coordinate table based on the node number. Subsequently, pair the leakage risk scores of the monitoring nodes output by the previous module with their corresponding coordinates. For example, if the score of Node 1 is 0.627 and that of Node 2 is 0.458, the pairing results are (120.3815, 36.0723, 0.627), (120.3821, 36.0730, 0.458). After processing all nodes, construct a data set with coordinates as the index field. Each item in this data set contains two types of information: node coordinates and their risk scores. Common structures include three-column CSV or GeoJSON formats for subsequent spatial analysis and processing. In practical applications, if 20 nodes are deployed and the coordinate coverage range is the rectangular boundary of the factory area, the system will generate a space score data table with 20 records for constructing a spatial risk visualization layer or calculating an isogram.

[0045] Risk probability estimation sub-module: Based on the space risk score coordinate data, construct a risk probability distribution layer within a continuous space range centered on the monitoring node coordinates, and generate a space risk probability field; When constructing the spatial risk probability field, with the two-dimensional coordinates of each monitoring node as the center, combined with its leakage risk score, the risk probability of any spatial point is calculated within the entire continuous space range. Specifically, it is implemented through an algorithm based on weighted distance averaging. This method does not use kernel density estimation, but directly weights inversely according to the spatial distance and the node score. The calculation formula is as follows: ; Among them, : represents the estimated value of the risk probability of any point in space, which is the final output result of this point; : represents the th monitoring node, numbered from 1 to ; : is the total number of monitoring nodes, that is, there are nodes participating in the risk valuation of this spatial point; : is the leakage risk score of node ; : is the current valuation point and the th monitoring node The Euclidean distance between them, the calculation formula is: , : is the coordinate of the spatial point for which the risk probability needs to be estimated currently; : is the coordinate of the th monitoring node.

[0046] This method multiplies the risk score value of each node by the influence factor of this node on point , that is, the closer the node is to the current point, the greater its risk contribution to the current point, and the farther away, the smaller the influence. To avoid the division by zero problem caused by the complete coincidence of the node and the valuation point positions, in actual deployment, a distance lower limit processing logic can be added to the program, for example, setting the minimum distance to 0.001 meters.

[0047] Suppose there are 3 monitoring nodes, and their coordinates and risk scores are as follows: Node 1 (number ): coordinate , score ; Node 2 (number ): coordinate , score ; Node 3 (number ): coordinate , score .

[0048] The point to be calculated is ​, calculate its Euclidean distances to three nodes as follows: ; ; .

[0049] Substitute these values into the risk probability estimation formula. The numerator is: ; The denominator is: ; The final risk probability is: .

[0050] This value indicates that the risk probability of the spatial point is 0.504, which is in the medium - risk area. Repeat the above operations for each spatial grid point in the entire monitoring area to generate a complete two - dimensional risk probability distribution layer. This layer forms a grid or matrix with actual coordinates as the index and risk probability as the value, and the data structure can be a 100×100 two - dimensional array.

[0051] Boundary extraction sub - module: According to the risk probability distribution of each spatial point in the spatial risk probability field, identify the edge positions of the mutation regions and generate a methanol leakage prediction map; Read the data of the spatial risk probability field, analyze the risk probability values of each spatial point in this two - dimensional risk layer, identify the edges of the regions where the risk values are concentrated and continuously distributed, which are used to determine the possible methanol leakage diffusion boundaries. The specific process is as follows: First, set the identification threshold of the risk probability. For example, set it to 0.5, which means only the regions with probability values greater than or equal to 0.5 are extracted to participate in the boundary identification. Traverse all points in the layer matrix, mark the qualified points as "high - risk areas", and then perform connectivity determination on these points. That is, if a group of adjacent grid points are all high - risk values, they are regarded as a continuous mutation region. Use contour tracing or edge detection methods (such as based on the Marching Squares algorithm) to extract the boundary curve of this region. The extraction result is recorded in the form of a point set or polygon. Each boundary corresponds to a predicted leakage diffusion range. In practical applications, if the layer is a 100×100 grid and there are 2600 points with risk values greater than 0.5, and these points form 3 continuous regions, then this module will output the prediction map data composed of 3 boundary curves, indicating the positions of the edge regions where methanol leakage may occur and the shapes of their diffusion boundaries.

[0052] The above are only the preferred embodiments of the present invention, and do not limit the present invention in other forms. Any person skilled in the relevant art may use the technical content disclosed above to make changes or modifications into equivalent embodiments with equivalent changes and apply them to other fields. However, as long as it does not depart from the technical solution content of the present invention, any simple modification, equivalent change and modification made to the above embodiments based on the technical essence of the present invention still fall within the protection scope of the technical solution of the present invention.

Claims

1. An artificial intelligence-based methanol leakage prediction system, characterized in that, The system includes: Leakage path construction module: Obtain the methanol concentration of each monitoring node within a continuous time period, screen the monitoring nodes whose methanol concentration changes exceed the increase threshold in adjacent time periods and sort them, and construct a methanol leakage propagation path map; Propagation feature extraction module: Perform a monitoring node feature aggregation operation on all paths in the methanol leakage propagation path map through a graph attention network to obtain the monitoring node propagation influence factor; Environmental response identification module: Obtain the air pressure and wind speed at the location of each monitoring node within a continuous time period, screen the time periods when the air pressure and wind speed undergo synchronous mutations, calibrate them as the occurrence points of mutation events, and obtain a set of coupled mutation events; Risk score calculation module: Calculate the leakage risk score of the monitoring node by weighted calculation according to the monitoring node propagation influence factor and the occurrence frequency of the corresponding mutation events in the set of coupled mutation events; Spatial prediction generation module: According to the leakage risk score of the monitoring node, extract the boundary of the area where leakage is possible with reference to the two-dimensional coordinates of the monitoring node, and construct a methanol leakage prediction map.

2. The methanol leakage prediction system based on artificial intelligence according to claim 1, wherein The methanol leakage propagation path map includes a surge monitoring node sequence, a directed connection relationship, and a propagation time weight. The propagation influence factor is specifically the monitoring node importance weight, the path cumulative effect value, and the adjacent influence coefficient. The set of coupled mutation events includes the mutation event location, the synchronous mutation time period, and the event statistical frequency. The leakage risk score is specifically the monitoring node risk level, the score calculation weight, and the spatial score distribution. The methanol leakage prediction map includes a leakage probability field, a leakage boundary area, and leakage spatial coordinates.

3. The methanol leakage prediction system based on artificial intelligence according to claim 1, wherein, The leakage path construction module includes: Concentration acquisition sub-module: Obtain the methanol concentration of each monitoring node within a continuous time period, detect whether the methanol concentration change of each monitoring node in the time series exceeds the increase threshold, extract the first occurrence time and record it, and generate the monitoring node surge time data; Surge sorting sub-module: According to the monitoring node surge time data, arrange all monitoring nodes in ascending order of time, judge the surge sequence before and after the monitoring nodes, and screen the adjacent monitoring node combinations in the time sequence relationship to generate the sorted monitoring node pair information; Path generation sub-module: According to the surge time interval of each monitoring node combination in the sorted monitoring node pair information, establish a methanol leakage propagation path map.

4. The methanol leakage prediction system based on artificial intelligence according to claim 3, characterized in that, The propagation feature extraction module includes: Path reading sub-module: Extract all monitoring path sequences in the methanol leakage propagation path map, identify the arrangement order and connection direction of the monitoring nodes in each path, and generate monitoring path topology relationship information; Upstream and downstream aggregation sub-module: According to the monitoring path topology relationship information, call the feature aggregation mechanism of each monitoring node in the graph attention network, integrate the time surge order and connection direction of adjacent monitoring nodes, and update according to the monitoring node structure within the aggregation range to generate monitoring node association features; Weight calculation sub-module: According to the feature vectors of each monitoring node and the connection directions of adjacent connection edges in the monitoring node association features, weighted processing is performed through a graph attention mechanism to calculate the influence degree of each monitoring node in the path, and a monitoring node propagation influence factor is generated.

5. The methanol leakage prediction system based on artificial intelligence according to claim 4, wherein, The environmental response recognition module includes: Data acquisition sub-module: Obtain the air pressure and wind speed at the location of each monitoring node within a continuous time period, divide the two types of environmental parameters by monitoring node respectively, and perform time alignment according to a unified time resolution to generate a monitoring node environmental parameter sequence; Mutation detection sub-module: According to the monitoring node environmental parameter sequence, perform change trend recognition on the air pressure sequence and wind speed sequence of each monitoring node by means of double sliding window change point detection, and generate a monitoring node synchronous mutation time period; Event calibration sub-module: According to the monitoring node synchronous mutation time period, assign a mutation event mark to the corresponding monitoring node location, construct a mutation event item according to the monitoring node location and mutation time period, and generate a coupled mutation event set.

6. The methanol leakage prediction system based on artificial intelligence according to claim 5, wherein, The risk score calculation module includes: Location matching sub-module: Extract the monitoring node propagation influence factor and the monitoring node locations in the coupled mutation event set, identify the monitoring nodes with consistent spatial locations, and generate a monitoring node overlapping location list; Index extraction sub-module: According to the monitoring node overlapping location list, extract the propagation influence factor value and the mutation event occurrence frequency of the corresponding monitoring node respectively, and generate a monitoring node combined index; Score generation sub-module: According to the monitoring node combined index, perform weighted processing on the propagation influence factor value and the mutation event occurrence frequency of each monitoring node location, and perform numerical calculation on its leakage risk level in combination with the monitoring node corresponding index to generate a monitoring node leakage risk score.

7. The methanol leakage prediction system based on artificial intelligence according to claim 6, characterized in that, The spatial prediction generation module includes: Coordinate mapping sub-module: Extract the two-dimensional coordinate positions corresponding to each monitoring node, pair and organize the monitoring node leakage risk score and coordinates, construct a spatial score data set indexed by coordinates, and generate spatial risk score coordinate data; Risk probability estimation sub-module: According to the spatial risk score coordinate data, with the monitoring node coordinates as the center, construct a risk probability distribution layer in a continuous spatial range, and generate a spatial risk probability field; Boundary extraction sub-module: According to the risk probability distribution of each spatial point in the spatial risk probability field, identify the edge positions of the mutation area, and generate a methanol leakage prediction map.

Images