A method for monitoring and early warning of a production area

Abstract

The application discloses a kind of production area gas monitoring early warning method, it is related to data processing technical field, comprising the following steps: S1, production workshop is divided into three-dimensional grid, obtain several grid units;S2, utilize the sensor deployed in production workshop to collect multidimensional data sequence for each grid unit, and utilize kriging interpolation method to supplement;S3, according to the multidimensional data sequence of each grid unit, extract several features, and splice, form the feature matrix of production workshop;S4, construct monitoring early warning neural network, input feature matrix into monitoring early warning neural network, determine the gas unqualified early warning position of production workshop.The application fuses the multidimensional data of gas concentration, temperature, wind speed and air pressure, expands risk assessment dimension from single concentration to multi-dimension, avoids the risk misjudgment caused by single data.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of data processing, in particular to a gas monitoring and early warning method for a production area. BACKGROUND

[0002] The chemical production workshop has a complex spatial structure, with densely staggered equipment and horizontally distributed pipelines. After the leakage of hazardous gas, it is easily affected by ventilation conditions and equipment shielding, showing three-dimensional diffusion characteristics. The existing gas monitoring and early warning method mostly adopts a plane point monitoring + threshold comparison mode, which has many defects: first, the spatial coverage is not accurate, and the three-dimensional characteristics of gas diffusion are not considered, which easily forms a spatial blind area due to sparse monitoring points; second, the data correlation is low, and the spatio-temporal evolution law of gas diffusion cannot be captured; third, the risk prediction ability is weak, and traditional algorithms are difficult to process multi-dimensional and strongly coupled monitoring data, and cannot accurately mine the leakage risk characteristics behind the data. SUMMARY

[0003] The present application relates to the technical field of data processing, in particular to a gas monitoring and early warning method for a production area.

[0004] The technical scheme of the present application is as follows: a gas monitoring and early warning method for a production area comprises the following steps:

[0005] S1, three-dimensional grid division is performed on the production workshop to obtain a plurality of grid units;

[0006] S2, a plurality of multi-dimensional data sequences are collected for each grid unit by using sensors deployed in the production workshop, and Kriging interpolation method is used for supplement;

[0007] S3, according to the multi-dimensional data sequence of each grid unit, a plurality of features are extracted and spliced to form a feature matrix of the production workshop;

[0008] S4, a monitoring and early warning neural network is constructed, the feature matrix is input into the monitoring and early warning neural network, and the gas unqualified early warning position of the production workshop is determined.

[0009] Further, in S2, the multi-dimensional data sequence is specifically , wherein, represents the gas concentration of the grid unit at time, represents the temperature of the grid unit at time, represents the wind speed of the grid unit at time, represents the air pressure of the grid unit at time, represents the horizontal coordinate of the grid unit in the production workshop, represents the vertical coordinate of the grid unit in the production workshop, represents the vertical coordinate of the grid cell in the production workshop.

[0010] The beneficial effect of the above further scheme is that in the present application, the gas concentration is a direct indicator of the leakage. The temperature affects the vapor pressure of the gas, and in turn affects the leakage rate. The wind speed affects the diffusion coefficient of the gas, which is positively correlated with the wind speed, and in turn affects the diffusion range. The air pressure affects the diffusion direction of the gas (diffusion from high pressure area to low pressure area), which is a key factor of the diffusion path in a complex workshop.

[0011] Further, S3 comprises the following sub-steps:

[0012] S31, extracting the gas concentration in the multi-dimensional data sequence of each time of the grid cell, and generating a time sequence feature;

[0013] S32, extracting the temperature, wind speed and air pressure in the multi-dimensional data sequence of each time of the grid cell, and generating an environment statistical feature;

[0014] S33, generating a space-time correlation feature according to the gas concentration in the multi-dimensional data sequence of each time of all grid cells in the production workshop and the positional relationship;

[0015] S34, normalizing the time sequence feature, the environment statistical feature and the space-time correlation feature;

[0016] S35, splicing the normalized time sequence feature, the environment statistical feature and the space-time correlation feature of each grid cell to obtain a feature matrix of the production workshop; the number of rows of the feature matrix is the same as the number of grid cells.

[0017] The beneficial effect of the above further scheme is that in the present application, the time sequence feature extracts the time dimension feature of the concentration, the environment statistical feature extracts the statistical feature of the temperature, wind speed and air pressure, and the space-time correlation feature extracts the spatial correlation feature between grids. The feature parameters capable of representing the gas leakage state are extracted from the multi-dimensional time sequence data sequence of each grid, so as to realize dimension reduction and information condensation of high-dimensional time sequence data.

[0018] Further, S31 comprises the following sub-steps:

[0019] S311, extracting the gas concentration in the multi-dimensional data sequence of each time of the grid cell, and describing the gas concentration of all times of the grid cell as a concentration-time curve;

[0020] S312, taking the value in the concentration-time curve satisfying a preset time sequence condition as a first time sequence sub-feature, wherein the preset time sequence condition is specifically and satisfies , represents the vertical coordinate of the grid cell in the production workshop. the gas concentration at the time point, representing the grid cell at the gas concentration at the time point, representing the time difference value, representing taking the maximum value;

[0021] S313, based on the concentration-time curve, taking the gas concentration as the integral object, performing integral from 0- the time point to obtain the total integral area;

[0022] S314, taking the vertical line corresponding to the peak value in the concentration-time curve as the symmetry axis, taking two time points on both sides of the symmetry axis to form a time interval, so that the local integral area of the time interval is a set proportion of the total integral area;

[0023] S315, taking the time interval as the second time sequence sub-feature.

[0024] The beneficial effects of the above further scheme are: in the present application, the maximum time length of continuous rise in the concentration sequence is one of the core indexes of the risk level. Taking the concentration peak value as the symmetry axis, the time interval covering the set proportion of the total integral area is extracted, and according to the concentration curve of the leakage having the peak symmetry, i.e. the core amount of the leakage is concentrated near the peak value, the core leakage period corresponding to 80% of the total integral area can be set.

[0025] On the basis of obtaining the total integral area, taking the time point as the symmetry center, the local time interval is symmetrically divided in front of and behind in the total monitoring time range (satisfying , i.e. is and the midpoint), and then the concentration-time curve in the local interval is integrated to obtain the local integral area with the peak value as the symmetry reference, and the physical meaning is the cumulative concentration of the core period near the peak value.

[0026] Further, S32 includes the following sub-steps:

[0027] S321, extracting the temperature, wind speed and air pressure in the multi-dimensional data sequence of each time point of the grid cell, generating a temperature sequence, a wind speed sequence and an air pressure sequence for the grid cell respectively;

[0028] S322, dividing the temperature sequence into a plurality of temperature subsequences, calculating the mean value of each temperature subsequence, and taking the proportion of the mean value of the temperature subsequence in the sum of all mean values as the weight of the temperature subsequence;

[0029] S323. Based on the weights of the temperature subsequences, the skewness of all temperature subsequences is weighted and summed to obtain the statistical characteristics of the temperature sequence.

[0030] S324. Divide the wind speed sequence into several wind speed subsequences, calculate the mean of each wind speed subsequence, and use the proportion of the mean of the wind speed subsequence to the sum of all the means as the weight of the wind speed subsequence.

[0031] S325. Based on the weights of the wind speed subsequences, the skewness of all wind speed subsequences is weighted and summed to obtain the statistical characteristics of the wind speed sequence.

[0032] S326. Divide the pressure sequence into several pressure subsequences, calculate the mean of each pressure subsequence, and use the proportion of the mean of the pressure subsequence to the sum of all the means as the weight of the pressure subsequence.

[0033] S327. Based on the weights of the pressure subsequences, the skewness of all pressure subsequences is weighted and summed to obtain the statistical characteristics of the pressure sequence.

[0034] The beneficial effects of the above-mentioned further solutions are as follows: In this invention, environmental parameters are generally dynamically changing, and the time series of environmental parameters is divided into several subsequences. The higher the mean of a time period, the stronger the impact of the environment on leakage. Therefore, the proportion of the subsequence mean to the sum of the means of all subsequences is used as a weight, and the time-segmented weighted skewness can better reflect the impact of the dynamic changes of environmental parameters on leakage, such as capturing sudden changes in wind speed caused by the start-up and shutdown of the ventilation system.

[0035] Based on the statistical characteristics of temperature series For example, , ,in, The weights of the first temperature subsequence. The weights for the second temperature subsequence. For the first The weights of each temperature subsequence. For the first The weights of each temperature subsequence. The skewness of the first temperature subsequence. The skewness of the second temperature subsequence. For the first skewness of a temperature subsequence For the first skewness of a temperature subsequence For the first The mean of all temperatures contained in a temperature subsequence. This represents the number of temperature subsequences. The same logic applies to wind speed and air pressure.

[0036] Furthermore, S33 includes the following sub-steps:

[0037] S331, extract the gas concentration in the multi-dimensional data sequence of each time of the grid unit, and generate a gas concentration sequence for the grid unit;

[0038] S332, extract the Pearson correlation coefficient between the gas concentration sequence of the current grid unit and the gas concentration sequence of the adjacent grid unit;

[0039] S333, take the maximum value of all Pearson correlation coefficients corresponding to the current grid unit as the space-time correlation feature.

[0040] The beneficial effects of the above further scheme are: in the present application, the leaked gas diffuses from the current grid to the adjacent grid, and the concentration sequence has strong correlation, and the Pearson correlation coefficient reflects the linear correlation degree of the concentration sequences of the two grids, for example, if the current grid is a leakage source, the concentration sequence of the adjacent grid and its correlation coefficient are close to 1. Among the plurality of adjacent grids, only one is the main diffusion direction (for example, the grid above the leakage source, because of the upward floating of hydrogen, the correlation coefficient is the largest), and taking the maximum value can highlight this core path and avoid the interference of multiple low correlation values.

[0041] Further, in S4, the monitoring and early warning neural network comprises an input module, a time sequence feature encoding module, an environment-space-time feature encoding module, a feature fusion module, a network space encoding module, a risk decoding module and an output module;

[0042] The input end of the input module serves as the input end of the monitoring and early warning neural network, the first output end thereof is connected with the input end of the time sequence feature encoding module, and the second output end thereof is connected with the input end of the environment-space-time feature encoding module; the output end of the time sequence feature encoding module is connected with the first input end of the feature fusion module; the output end of the environment-space-time feature encoding module is connected with the second input end of the feature fusion module; the output end of the feature fusion module, the network space encoding module, the risk decoding module and the input end of the output module are connected in sequence; and the output end of the output module serves as the output end of the monitoring and early warning neural network.

[0043] The beneficial effects of the above further scheme are: in the present application, the conventional network encodes all features together, even if normalized, the feature distribution characteristics of the time dimension feature (numerical range 0~60 seconds) and the dimensionless feature (0~1) are still different, which is easy to cause the network to learn the features with large numerical range; therefore, in the present application, the time dimension time sequence feature and the dimensionless environment-space-time feature are split into two branches, which are processed by time sequence convolution (adapted to time dynamics) and statistical encoding (adapted to dimensionless statistical features) respectively, so as to retain the respective feature characteristics and avoid the interference caused by the dimension.

[0044] The grid space encoding layer performs several neighborhood convolutions based on physically adjacent grids.

[0045] The input module performs feature decomposition on the feature matrix to obtain a time sequence feature sub-tensor, that is, a first time sequence sub-feature and a second time sequence sub-feature are extracted from the first column to the second column of the feature matrix, and only time dimensional features are reserved; and an environment-spatiotemporal feature sub-tensor is obtained, that is, the third column to the Jth column of the feature matrix is extracted, and only dimensionless features are reserved. Wherein, denotes a real set, denotes a batch size, denotes a total number of grids, denotes a feature dimension of a single grid unit. The input module can also perform min-max normalization on the time sequence feature sub-tensor as necessary, so as to avoid learning imbalance caused by mixing of time dimensional features and dimensionless features, and compress the time dimensional features to 0-1, which is aligned with the order of magnitude of the environment-spatiotemporal features.

[0046] In the time sequence feature encoding module, a causal convolution is used in the time sequence convolution layer, and the convolution kernel only acts on the time feature dimension, the convolution kernel size k=3, and the step length s=1. The time sequence attention layer performs channel attention processing on the output of the TCN (time sequence convolution layer), and outputs the encoded time sequence features.

[0047] In the environment-spatiotemporal feature encoding module, a full connection and a batch normalization operation are used in the feature encoding layer, and the enhancement layer can construct an adjacent weight matrix according to the Pearson correlation coefficient, enhance the spatial correlation, and output the encoded environment-spatiotemporal features.

[0048] The feature fusion module performs weighted fusion on the encoded time sequence features and the encoded environment-spatiotemporal features, and obtains the fused features.

[0049] The network space encoding module encodes the fused features by using a 3D grid neighborhood convolution, and obtains the space encoded features. The risk decoding layer uses a full connection dimension reduction, and the output result is input into a gating function to obtain the decoded features. The decoded features are input into an output layer, a risk score of the grid unit is calculated by using a sigmoid function, and the risk level of each grid is determined.

[0050] Further, the time sequence feature encoding module comprises a time sequence convolution layer and a time sequence attention layer connected in sequence;

[0051] The environment-spatiotemporal feature encoding module comprises a feature encoding layer and an enhancement layer connected in sequence.

[0052] The beneficial effects of the present application are:

[0053] (1) The present application fuses multi-dimensional data of gas concentration, temperature, wind speed and air pressure, expands the risk assessment dimension from single concentration to multi-dimension, and avoids risk misjudgment caused by single data;

[0054] (2) The present invention adopts a domain extraction method based on time-series features, environmental statistical features and spatiotemporal correlation features, covering the dynamic time process of leakage, the dynamic influence of environmental factors and the diffusion correlation in three-dimensional space; through spatiotemporal correlation features and spatial coding modules of neural networks, the leakage grid can be accurately located, effectively reducing the risk of personnel being exposed to dangerous gases, and reducing production interruption losses caused by blind handling. Attached Figure Description

[0055] Figure 1 A flowchart for a gas monitoring and early warning method for production areas;

[0056] Figure 2 A schematic diagram of the structure of a monitoring and early warning neural network. Detailed Implementation

[0057] The embodiments of the present invention will be further described below with reference to the accompanying drawings.

[0058] like Figure 1 As shown, the present invention provides a gas monitoring and early warning method for production areas, comprising the following steps:

[0059] S1. Divide the production workshop into three-dimensional meshes to obtain several mesh units;

[0060] S2. Use sensors deployed in the production workshop to collect multi-dimensional data sequences for each grid cell, and supplement them with Kriging interpolation.

[0061] S3. Based on the multi-dimensional data sequence of each grid unit, extract several features and splice them together to form a feature matrix of the production workshop;

[0062] S4. Construct a monitoring and early warning neural network, input the feature matrix into the monitoring and early warning neural network, and determine the location of gas non-compliance warning in the production workshop.

[0063] In this embodiment of the invention, in S2, the multi-dimensional data sequence is specifically as follows: ,in, Indicates that the mesh cell is in The gas concentration at time , Indicates that the mesh cell is in Temperature at any moment Indicates that the mesh cell is in Wind speed at any moment Indicates that the mesh cell is in air pressure at any moment This represents the x-coordinate of the grid cell in the production workshop. This represents the ordinate of the grid cell in the production workshop. represents the vertical coordinate of the grid cell in the production workshop.

[0064] In the present application, the gas concentration is a direct indicator of the leakage. The temperature affects the vapor pressure of the gas, and in turn the leakage rate. The wind speed affects the diffusion coefficient of the gas, which is positively correlated with the wind speed, and in turn the diffusion range. The air pressure affects the diffusion direction of the gas (diffusion from high pressure area to low pressure area), which is a key factor of the diffusion path in a complex workshop.

[0065] In an embodiment of the present application, S3 comprises the following sub-steps:

[0066] S31, extract the gas concentration in the multi-dimensional data sequence of each time of the grid cell, and generate the time sequence feature;

[0067] S32, extract the temperature, wind speed and air pressure in the multi-dimensional data sequence of each time of the grid cell, and generate the environment statistical feature;

[0068] S33, generate the space-time correlation feature according to the gas concentration in the multi-dimensional data sequence of each time of all grid cells in the production workshop and the positional relationship;

[0069] S34, normalize the time sequence feature, the environment statistical feature and the space-time correlation feature;

[0070] S35, splice the normalized time sequence feature, the environment statistical feature and the space-time correlation feature of each grid cell, and obtain the feature matrix of the production workshop; the number of rows of the feature matrix is the same as the number of grid cells.

[0071] In the present application, the time sequence feature extracts the time dimension feature of the concentration, the environment statistical feature extracts the statistical feature of the temperature, wind speed and air pressure, and the space-time correlation feature extracts the spatial correlation feature between grids. The feature parameters capable of representing the gas leakage state are extracted from the multi-dimensional time sequence data sequence of each grid, so as to realize the dimension reduction and information condensation of high-dimensional time sequence data.

[0072] In an embodiment of the present application, S31 comprises the following sub-steps:

[0073] S311, extract the gas concentration in the multi-dimensional data sequence of each time of the grid cell, and depict the gas concentration of all times of the grid cell as a concentration-time curve;

[0074] S312, take the value in the concentration-time curve satisfying the preset time sequence condition as the first time sequence sub-feature, wherein the preset time sequence condition is and satisfies , represents the gas concentration of the grid cell at time, representing the gas concentration of the grid cell at representing the time difference value, representing the maximum value; representing the maximum value;

[0075] S313, based on the concentration-time curve, taking the gas concentration as the integral object, performing integration from 0- to the time, to obtain a total integral area;

[0076] S314, taking the vertical line corresponding to the peak value in the concentration-time curve as the symmetry axis, taking two time points on both sides of the symmetry axis to form a time interval, so that the local integral area of the time interval is a set proportion of the total integral area;

[0077] S315, taking the time interval as a second time sequence sub-feature.

[0078] In the present application, the maximum time length of continuous rise in the concentration sequence is one of the core indicators of the risk level. Taking the concentration peak value as the symmetry axis, the time interval covering the set proportion of the total integral area is extracted, and according to the concentration curve of the leakage having peak symmetry, i.e. the core amount of the leakage is concentrated near the peak value, the total integral area of 80% can be set to correspond to the core leakage period.

[0079] On the basis of obtaining the total integral area, taking the time as the symmetry center, the local time interval is symmetrically determined before and after in the total monitoring time range (satisfying , i.e. is and the midpoint of ), and then the concentration-time curve in the local interval is integrated to obtain the local integral area with the peak value as the symmetry reference, and the physical meaning is the cumulative concentration of the core period near the peak value.

[0080] In the embodiment of the present application, S32 includes the following sub-steps:

[0081] S321, extracting the temperature, wind speed and air pressure in the multi-dimensional data sequence of each time of the grid cell, to generate a temperature sequence, a wind speed sequence and an air pressure sequence for the grid cell respectively;

[0082] S322, dividing the temperature sequence into a plurality of temperature subsequences, calculating the mean value of each temperature subsequence, and taking the proportion of the mean value of the temperature subsequence in the sum of all mean values as the weight of the temperature subsequence;

[0083] S323, based on the weight of the temperature subsequence, performing weighted summation on the skewness of all temperature subsequences to obtain the statistical feature of the temperature sequence;

[0084] S324, divide the wind speed sequence into several wind speed subsequences, calculate the mean value of each wind speed subsequence, and take the proportion of the mean value of the wind speed subsequence in the sum of all mean values as the weight of the wind speed subsequence;

[0085] S325, based on the weight of the wind speed subsequence, the skewness of all wind speed subsequences is weighted and summed to obtain the statistical feature of the wind speed sequence;

[0086] S326, divide the air pressure sequence into several air pressure subsequences, calculate the mean value of each air pressure subsequence, and take the proportion of the mean value of the air pressure subsequence in the sum of all mean values as the weight of the air pressure subsequence;

[0087] S327, based on the weight of the air pressure subsequence, the skewness of all air pressure subsequences is weighted and summed to obtain the statistical feature of the air pressure sequence.

[0088] In the present application, the time sequence of the environmental parameter is generally dynamically changed, and the time sequence of the environmental parameter is divided into several subsequences. The higher the mean value of the period, the stronger the influence of the environment on the leakage, so the proportion of the mean value of the subsequence in the sum of all subsequence mean values is taken as the weight, and the weighted skewness in time period can better reflect the influence of the dynamic change of the environmental parameter on the leakage, such as capturing the wind speed mutation caused by the start and stop of the ventilation system.

[0089] Taking the statistical feature of the temperature sequence as an example, , wherein, is the weight of the first temperature subsequence, is the weight of the second temperature subsequence, is the weight of the n-th temperature subsequence, is the weight of the (n+1)th temperature subsequence, is the skewness of the first temperature subsequence, is the skewness of the second temperature subsequence, is the skewness of the n-th temperature subsequence, is the skewness of the (n+1)th temperature subsequence, is the skewness of the (n+1)th temperature subsequence, is the skewness of the (n+1)th temperature subsequence, is the skewness of the (n+1)th temperature subsequence, is the skewness of the (n+1)th temperature subsequence, is the skewness of the (n+1)th temperature subsequence, is the skewness of the (n+1)th temperature subsequence, is the number of temperature subsequences. The same applies to wind speed and air pressure.

[0090] In the embodiment of the present application, S33 comprises the following sub-steps:

[0091] S331, extract the gas concentration in the multi-dimensional data sequence of each time of the grid unit, and generate a gas concentration sequence for the grid unit;

[0092] S332, extract the Pearson correlation coefficient between the gas concentration sequence of the current grid unit and the gas concentration sequence of the adjacent grid unit;

[0093] S333, taking the maximum value of all Pearson correlation coefficients corresponding to the current grid unit as the space-time correlation feature.

[0094] In the present application, the leaked gas diffuses from the current grid to the adjacent grid, and the concentration sequence has strong correlation, and the Pearson correlation coefficient reflects the linear correlation degree of the concentration sequence of the two grids, such as the current grid being the leakage source, the concentration sequence of the adjacent grid and its correlation coefficient close to 1. Among the plurality of adjacent grids, only one is the main diffusion direction (such as the grid above the leakage source, because of the upward floating of hydrogen, the correlation coefficient is the largest), and the maximum value is taken to highlight this core path and avoid the interference of multiple low correlation values.

[0095] In the embodiment of the present application, as shown in S4, Figure 2 The monitoring and early warning neural network comprises an input module, a time sequence feature encoding module, an environment-space-time feature encoding module, a feature fusion module, a network space encoding module, a risk decoding module and an output module.

[0096] The input end of the input module serves as the input end of the monitoring and early warning neural network, the first output end thereof is connected with the input end of the time sequence feature encoding module, and the second output end thereof is connected with the input end of the environment-space-time feature encoding module; the output end of the time sequence feature encoding module is connected with the first input end of the feature fusion module; the output end of the environment-space-time feature encoding module is connected with the second input end of the feature fusion module; the output end of the feature fusion module, the network space encoding module, the risk decoding module and the input end of the output module are connected in sequence; and the output end of the output module serves as the output end of the monitoring and early warning neural network.

[0097] In the present application, the conventional network encodes all features together, even if normalized, the feature distribution characteristics of the time dimension feature (numerical range 0~60 seconds) and the dimensionless feature (0~1) are still different, which is easy to cause the network to learn the features with large numerical range; therefore, the time sequence feature of the time dimension and the environment-space-time feature of the dimensionless feature are split into two branches in the present application, which are processed by time sequence convolution (adapted to time dynamics) and statistical encoding (adapted to dimensionless statistical features) respectively, so as to retain the respective feature characteristics and avoid the interference caused by the dimension.

[0098] The grid space encoding layer performs a plurality of neighborhood convolutions based on the physically adjacent grids.

[0099] The input module processes the feature matrix Feature splitting is performed to obtain a time sequence feature sub-tensor, that is, a first time sequence sub-feature and a second time sequence sub-feature are extracted from the first column and the second column of the feature matrix, and only time dimension features are reserved; an environment-spatiotemporal feature sub-tensor is obtained, that is, the third column to the Jth column of the feature matrix is extracted, and only dimensionless features are reserved. Wherein, denotes a real set, denotes a batch size, denotes a total number of grids, denotes a feature dimension of a single grid unit. The input module can also perform min-max normalization on the time sequence feature sub-tensor as needed, so as to avoid learning imbalance caused by mixing of time dimension features and dimensionless features, and compress the time dimension features to 0-1, aligning with the order of magnitude of the environment-spatiotemporal features.

[0100] In the time sequence feature encoding module, a causal convolution is used in the time sequence convolution layer, and the convolution kernel only acts on the time feature dimension, the convolution kernel size k=3, and the step length s=1. The time sequence attention layer performs channel attention processing on the output of the TCN (time sequence convolution layer), and outputs the encoded time sequence features.

[0101] In the environment-spatiotemporal feature encoding module, a full connection and a batch normalization operation are used in the feature encoding layer, and the enhancement layer can construct an adjacent weight matrix according to the Pearson correlation coefficient, enhance the spatial correlation, and output the encoded environment-spatiotemporal features.

[0102] The feature fusion module performs weighted fusion on the encoded time sequence features and the encoded environment-spatiotemporal features, and obtains the fused features.

[0103] The network space encoding module encodes the fused features by using a 3D grid neighborhood convolution, and obtains the spatial encoding features. The risk decoding layer uses a full connection dimension reduction, and the output result is input into a gating function to obtain the decoded features. The decoded features are input into an output layer, a risk score of the grid unit is calculated by using a sigmoid function, and the risk level of each grid is determined. The grid with an unqualified risk level is taken as a warning point, and an audible and visual alarm is sent.

[0104] In the embodiment of the application, the time sequence feature encoding module comprises a time sequence convolution layer and a time sequence attention layer connected in sequence;

[0105] The environment-spatiotemporal feature encoding module comprises a feature encoding layer and an enhancement layer connected in sequence.

[0106] Those skilled in the art will appreciate that the embodiments described herein are presented for purposes of illustration and that the inventive principles are not limited to these particular embodiments. Other variations and modifications can be made to the embodiments without departing from the spirit and scope of the inventive principles.

Claims

1. A gas monitoring and early warning method for a production area, characterized in that, Includes the following steps: S1. Divide the production workshop into three-dimensional meshes to obtain several mesh units; S2. Use sensors deployed in the production workshop to collect multi-dimensional data sequences for each grid cell, and supplement them with Kriging interpolation. S3. Based on the multi-dimensional data sequence of each grid unit, extract several features and splice them together to form a feature matrix of the production workshop; S4. Construct a monitoring and early warning neural network, input the feature matrix into the monitoring and early warning neural network, and determine the location of gas non-compliance warning in the production workshop; S3 includes the following sub-steps: S31. Extract the gas concentration from the multi-dimensional data sequence of each grid cell at each time step to generate time-series features; S32. Extract temperature, wind speed and air pressure from the multi-dimensional data sequence of each grid cell at each time point to generate environmental statistical features; S33. Generate spatiotemporal correlation features based on the gas concentration and positional relationships in the multi-dimensional data sequences of all grid cells in the production workshop at various times. S34. Normalize the temporal characteristics, environmental statistical characteristics, and spatiotemporal correlation characteristics; S35. The time-series features, environmental statistical features, and spatiotemporal correlation features of each grid cell after normalization are concatenated to obtain the feature matrix of the production workshop; the number of rows in the feature matrix is ​​the same as the number of grid cells. S31 includes the following sub-steps: S311. Extract the gas concentration from the multi-dimensional data sequence of each grid cell at each time step, and plot the gas concentration of the grid cell at all times as a concentration-time curve. S312. The values ​​in the concentration-time curve that satisfy the preset time series conditions are taken as the first time series sub-features, wherein the preset time series conditions are specifically as follows: and satisfy , Indicates that the mesh cell is in The gas concentration at time , Indicates that the mesh cell is in The gas concentration at time , Indicates the time difference. This indicates taking the maximum value; S313. Based on the concentration-time curve, with gas concentration as the integrand, perform an integration from 0- The integral at each time step yields the total area of ​​the integral. S314. Using the vertical line corresponding to the peak value in the concentration-time curve as the axis of symmetry, take two symmetrical time points on both sides of the axis of symmetry to form a time interval, such that the local integral area of ​​the time interval is a set proportion of the total integral area. S315. Use the time interval as the second time series sub-feature.

2. The gas monitoring and early warning method for production areas according to claim 1, characterized in that, In S2, the multi-dimensional data sequence is specifically as follows: ,in, Indicates that the mesh cell is in The gas concentration at time , Indicates that the mesh cell is in Temperature at any moment Indicates that the mesh cell is in Wind speed at any moment Indicates that the mesh cell is in air pressure at any moment This represents the x-coordinate of the grid cell in the production workshop. This represents the ordinate of the grid cell in the production workshop. This represents the vertical coordinate of the grid cell in the production workshop.

3. The gas monitoring and early warning method for production areas according to claim 1, characterized in that, S32 includes the following sub-steps: S321. Extract the temperature, wind speed and air pressure from the multi-dimensional data sequence of each grid cell at each time moment, and generate temperature sequence, wind speed sequence and air pressure sequence for the grid cell respectively; S322. Divide the temperature sequence into several temperature subsequences, calculate the mean of each temperature subsequence, and use the proportion of the mean of the temperature subsequence to the sum of all the means as the weight of that temperature subsequence. S323. Based on the weights of the temperature subsequences, the skewness of all temperature subsequences is weighted and summed to obtain the statistical characteristics of the temperature sequence. S324. Divide the wind speed sequence into several wind speed subsequences, calculate the mean of each wind speed subsequence, and use the proportion of the mean of the wind speed subsequence to the sum of all the means as the weight of the wind speed subsequence. S325. Based on the weights of the wind speed subsequences, the skewness of all wind speed subsequences is weighted and summed to obtain the statistical characteristics of the wind speed sequence. S326. Divide the pressure sequence into several pressure subsequences, calculate the mean of each pressure subsequence, and use the proportion of the mean of the pressure subsequence to the sum of all the means as the weight of the pressure subsequence. S327. Based on the weights of the pressure subsequences, the skewness of all pressure subsequences is weighted and summed to obtain the statistical characteristics of the pressure sequence.

4. The gas monitoring and early warning method for production areas according to claim 1, characterized in that, S33 includes the following sub-steps: S331. Extract the gas concentration from the multi-dimensional data sequence of each grid cell at each time step, and generate a gas concentration sequence for each grid cell. S332. Extract the Pearson correlation coefficient between the gas concentration sequence of the current grid cell and the gas concentration sequences of adjacent grid cells; S333. Take the maximum value of all Pearson correlation coefficients corresponding to the current grid cell as the spatiotemporal correlation feature.

5. The gas monitoring and early warning method for production areas according to claim 1, characterized in that, In S4, the monitoring and early warning neural network includes an input module, a temporal feature encoding module, an environment-spatiotemporal feature encoding module, a feature fusion module, a network spatial encoding module, a risk decoding module, and an output module; The input terminal of the input module serves as the input terminal of the monitoring and early warning neural network. Its first output terminal is connected to the input terminal of the temporal feature encoding module, and its second output terminal is connected to the input terminal of the environment-spatiotemporal feature encoding module. The output terminal of the temporal feature encoding module is connected to the first input terminal of the feature fusion module. The output terminal of the environment-spatiotemporal feature encoding module is connected to the second input terminal of the feature fusion module. The output terminal of the feature fusion module, the network spatial encoding module, the risk decoding module, and the input terminal of the output module are connected in sequence. The output terminal of the output module serves as the output terminal of the monitoring and early warning neural network.

6. The gas monitoring and early warning method for production areas according to claim 5, characterized in that, The temporal feature encoding module includes a temporal convolutional layer and a temporal attention layer connected in sequence; The environment-spatiotemporal feature encoding module includes a feature encoding layer and an enhancement layer connected in sequence.

Images