Harmful gas diffusion process rapid prediction method based on 3DCNN-LSTM coupling model
By applying the 3DCNN-LSTM coupling model in the underground factory of hydropower stations, combined with CFD numerical simulation and deep learning, the real-time and integrity problems of pollutant concentration field prediction in traditional methods are solved, and the pollutant concentration field in the underground factory of hydropower stations is realized. Fast real-time prediction to ensure the safety of construction workers.
Patent Information
- Application Number
- CN202410878423.7
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2024-07-02
- Publication Date
- 2024-11-15
AI Technical Summary
The existing technology method used to predict the concentration field of toxic and harmful gases and dust pollutants in underground factories of hydropower stations after blasting construction has real-time and integrity problems. Traditional CFD numerical simulation computing resources are consumed and time-consuming, while deep learning models only It can predict the future pollutant concentration data of a single point, and it is impossible to obtain the complete pollutant concentration field in the cave room.
Using a fast prediction method for harmful gas diffusion process based on 3DCNN-LSTM coupled model, the tensor data at grid nodes were obtained through CFD numerical simulation, and the 3DCNN-LSTM coupled neural network model was constructed, and deep learning was used to perform it in time and space dimensions. Real-time prediction of pollutant concentration field.
It has achieved rapid real-time prediction of the complete pollutant concentration field in the underground factory of the hydropower station at any time in the future, guided construction personnel to enter the site and regulated fan, and ensured the health and safety of construction personnel.
Smart Images

Figure CN118969132A_ABST
Abstract
Description
Technical Field
[0001] The present invention relates to the field of data prediction, and in particular to a method for quickly predicting a harmful gas diffusion process based on a 3DCNN-LSTM coupling model. Background Art
[0002] The excavation of underground powerhouse caverns of hydropower stations is often carried out by blasting. After the blasting construction of the caverns, a large amount of toxic and harmful gases and dust will be generated, which will cause serious harm to the health of construction workers, and a large amount of dust will aggravate mechanical wear. At present, the judgment of whether the concentration of toxic and harmful gases and dust and other pollutants before blasting work has dropped to a level that construction workers can enter mainly relies on the detection values of fixed and mobile sensors, or CFD numerical simulation to view the concentration distribution of various pollutants in the cavern at different times. However, the sensor can only detect the concentration values at a small number of measuring points, and cannot obtain the complete concentration field of pollutants. There may be a phenomenon of higher pollutant concentrations at non-sensor points; although the CFD numerical simulation method can obtain the complete concentration field of pollutants, it requires more computing resources and takes a long time, and cannot achieve the purpose of real-time rapid prediction.
[0003] There are two main methods for predicting pollutant concentrations in front of the working surface of the underground powerhouse of a traditional hydropower station: (1) Plan the construction conditions in advance, establish a three-dimensional model for ventilation calculation during the construction period of the underground powerhouse of the hydropower station, mesh the three-dimensional model, and then use CFD numerical simulation technologies such as multiphase flow models and discrete phase models to simulate the entire process of the migration and diffusion of various pollutants. Based on the calculation results, the concentration distribution of various pollutants in the cavern at a specific time is predicted in advance. (2) For each sensor monitoring point, collect historical time series data to construct a sensor historical monitoring data set. Then build a deep learning algorithm such as a bp neural network model or an LSTM long short-term memory network model, use the constructed sensor historical monitoring data set to train the model, input the real-time monitoring pollutant concentration sensor data into the deep learning model, and predict the pollutant concentration at a future time point.
[0004] The two methods for predicting pollutant concentrations in front of the working surface of the underground powerhouse of the hydropower station mentioned above have the following problems: (1) Although the pollutant concentration prediction method based on CFD numerical simulation technology can obtain the complete concentration field of pollutants, it requires more computing resources and takes a long time, and cannot achieve the purpose of real-time rapid prediction. In addition, the boundary conditions and initial conditions of its calculation are initial proposed values, and the calculation results cannot be coupled with the real-time monitoring data on site, and the reliability of the results is questionable. (2) The prediction method based on deep learning algorithms such as bp neural network and LSTM long short-term memory network model can couple with the real-time monitoring data on site, but it can only predict the pollutant concentration data at multiple single points in the future, and cannot predict the data of the complete pollutant concentration field in the cavern. Summary of the invention
[0005] The present invention aims to solve one of the technical problems in the related art at least to a certain extent.
[0006] To this end, the present invention proposes a method for quickly predicting the diffusion process of harmful gases based on a 3DCNN-LSTM coupling model. Based on the real-time monitoring data of sensors and by means of deep learning, the real-time prediction of the complete pollutant concentration field distribution in the cave at future moments is performed in time and space dimensions.
[0007] Another object of the present invention is to propose a rapid prediction system for the harmful gas diffusion process based on a 3DCNN-LSTM coupling model.
[0008] To achieve the above objectives, the present invention proposes a method for quickly predicting the diffusion process of harmful gases based on a 3DCNN-LSTM coupling model, comprising:
[0009] Based on the working condition data of the project site, the CFD numerical simulation calculation of pollutant migration and diffusion is carried out to obtain the numerical simulation calculation results;
[0010] Obtaining a preprocessed tensor data set at each grid node in the numerical simulation calculation result;
[0011] Building a coupled neural network model based on the 3DCNN model and the LSTM model, and inputting the tensor data set into the coupled neural network model to perform model training based on the extracted features to obtain a trained coupled neural network model;
[0012] The measured data at the engineering site is input into the trained coupled neural network model to predict and output a three-dimensional tensor of the spatial distribution of pollutant concentration.
[0013] In addition, the method for quickly predicting the harmful gas diffusion process based on the 3DCNN-LSTM coupling model according to the above embodiment of the present invention may also have the following additional technical features:
[0014] Furthermore, in one embodiment of the present invention, a numerical simulation calculation result of pollutant migration and diffusion is obtained by performing CFD numerical simulation calculation based on the working condition data of the engineering site, including:
[0015] A working condition table is constructed based on working condition data of the engineering site; wherein the working condition data includes the distance to the blasting smoke throwing area, the initial harmful gas concentration, the diameter of the wind tube, the distance between the wind tube outlet and the construction face, the wind speed at the wind tube outlet, the construction face area and the dust mass flow rate;
[0016] Based on the operating condition table, three-dimensional calculation models of the main powerhouse, main transformer tunnel, tail gate chamber and each connecting tunnel are established for each operating condition;
[0017] Based on the three-dimensional calculation model, common node operation is performed, and grid division with structural grid as the main component is performed to output a file in a preset format;
[0018] Based on the file in the preset format, CFD numerical simulation calculation of the migration and diffusion process of pollutants is performed to obtain numerical simulation calculation results; wherein the numerical simulation calculation results include the X-direction wind speed value, Y-direction wind speed value, Z-direction wind speed value, CO concentration value and dust mass percentage on all grid nodes.
[0019] Further, in one embodiment of the present invention, obtaining a preprocessed tensor data set at each grid node in the numerical simulation calculation result includes:
[0020] Convert the output file of CFD numerical simulation calculation into five-channel three-dimensional tensor data for model input; each channel tensor is respectively the X-axis velocity, Y-axis velocity, Z-axis velocity, harmful gas concentration and dust concentration tensor that conforms to the actual physical space distribution position of each sensor;
[0021] Convert each file output by CFD numerical simulation calculation into two-channel three-dimensional tensor data output by the model, where the two channel tensors are the concentration of harmful gases and the dust concentration tensor that conform to the actual physical space distribution points;
[0022] According to the preset input data time series length and the predicted data time series length, the five-channel three-dimensional tensor data and the two-channel three-dimensional tensor data are spliced in the time dimension to obtain the final tensor data set.
[0023] Further, in one embodiment of the present invention, a coupled neural network model is constructed based on the 3DCNN model and the LSTM model, and the tensor data set is input into the coupled neural network model to perform model training based on the extracted features to obtain model parameters, including:
[0024] A 3DCNN-LSTM coupled neural network model is established based on the 3DCNN model and the LSTM model; wherein the 3DCNN-LSTM coupled neural network model includes a spatial feature extraction network and a temporal feature extraction network; the spatial feature extraction network includes a 3DCNN layer, a BatchNorm3d layer, a Relu layer, a max_pool3d maximum pooling layer, and a fully connected layer; the temporal feature extraction network includes an encoder model and a decoder model;
[0025] Inputting the tensor data set into the 3DCNN-LSTM coupled neural network model, and performing point-by-point multiplication operations on the local voxel blocks of the input tensor data using the three-dimensional convolution kernel of the 3DCNN layer to output a first feature map;
[0026] The feature map is normalized using the BatchNorm3d layer to calculate the mean and variance of the feature map to obtain a second feature map; the second feature map is nonlinearly transformed using the Relu activation function to output a third feature map; and the third feature map is downsampled using the max_pool3d maximum pooling layer to obtain a fourth feature map of the maximum value in the pooling domain;
[0027] The first feature map, the second feature map, the third feature map and the fourth feature map are integrated by using a fully connected layer, mapped into an output vector to input into the temporal feature extraction network to output a feature vector, and the feature vector is input into the encoder model of the temporal feature extraction network to extract time series features;
[0028] The time series features are input into the decoder model for decoding and the output predicted one-dimensional vector is obtained, so as to train the 3DCNN-LSTM coupled neural network model based on the predicted one-dimensional vector to obtain a trained 3DCNN-LSTM coupled neural network model.
[0029] Furthermore, in one embodiment of the present invention, the measured data at the engineering site is input into the trained coupled neural network model to predict and output a three-dimensional tensor of the spatial distribution of pollutant concentration, including:
[0030] Hazardous gas concentration sensors and wind speed sensors are deployed at set points on the project site, and are converted to USB interfaces through conversion interfaces and connected to the host computer outside the tunnel;
[0031] The host computer outside the cave sends polling instructions to the sensor at a preset interval, and sends the acquired messages to the specified IP and port of the cloud server based on the UDP transmission protocol;
[0032] The cloud server listens to the specified port, parses each message after receiving it, and stores it in the specified database;
[0033] The parsed message in the database is converted into tensor data in a preset format, and the tensor data in the preset format is input into a trained 3DCNN-LSTM coupled neural network model to output a prediction result; wherein the prediction result includes a three-dimensional tensor of the predicted spatial distribution of pollutant concentration.
[0034] Furthermore, in one embodiment of the present invention, after outputting the three-dimensional tensor of the spatial distribution of pollutant concentration, the method further includes:
[0035] At preset intervals, concatenate each row of data in the ASCII file of the prediction result with the corresponding header into a key-value pair, and then concatenate each row of data into a json data format file;
[0036] On the Web side, an empty page is created based on the Vue.js architecture, and a three-dimensional cloud map is constructed based on the THREE.Points particle system API of THREE.js to complete the dynamic three-dimensional cloud map visualization display.
[0037] To achieve the above object, the present invention proposes a rapid prediction system for the diffusion process of harmful gases based on a 3DCNN-LSTM coupling model, comprising:
[0038] Numerical simulation calculation module, used to perform CFD numerical simulation calculation of pollutant migration and diffusion based on the working condition data of the engineering site to obtain numerical simulation calculation results;
[0039] A data preprocessing module, used to obtain a preprocessed tensor data set at each grid node in the numerical simulation calculation result;
[0040] A network model training module, used for building a coupled neural network model based on the 3DCNN model and the LSTM model, and inputting the tensor data set into the coupled neural network model to perform model training based on the extracted features to obtain a trained coupled neural network model;
[0041] The data prediction output module is used to input the measured data of the engineering site into the trained coupled neural network model to predict and output the three-dimensional tensor of the spatial distribution of pollutant concentration.
[0042] The harmful gas diffusion process rapid prediction method and system based on the 3DCNN-LSTM coupling model of the embodiment of the present invention can predict the complete pollutant concentration field distribution in the factory building at the future moment through the limited sensor point monitoring values at the current moment, and can guide the construction personnel's entry into the site and guide the advance frequency conversion control of the indoor air supply fan in the factory building, which helps to protect the health and life safety of the construction personnel.
[0043] Additional aspects and advantages of the present invention will be given in part in the following description and in part will be obvious from the following description, or will be learned through practice of the present invention. BRIEF DESCRIPTION OF THE DRAWINGS
[0044] The above and / or additional aspects and advantages of the present invention will become apparent and easily understood from the following description of the embodiments in conjunction with the accompanying drawings, in which:
[0045] Figure 1is a flowchart of a method for quickly predicting a harmful gas diffusion process based on a 3DCNN-LSTM coupling model according to an embodiment of the present invention;
[0046] Figure 2 It is a flowchart of building and using a 3DCNN-LSTM model according to an embodiment of the present invention;
[0047] Figure 3 It is a structural diagram of a harmful gas diffusion process rapid prediction system based on a 3DCNN-LSTM coupling model according to an embodiment of the present invention. DETAILED DESCRIPTION
[0048] It should be noted that, in the absence of conflict, the embodiments of the present invention and the features in the embodiments can be combined with each other. The present invention will be described in detail below with reference to the accompanying drawings and in combination with the embodiments.
[0049] In order to enable those skilled in the art to better understand the scheme of the present invention, the technical scheme in the embodiments of the present invention will be clearly and completely described below in conjunction with the drawings in the embodiments of the present invention. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work should fall within the scope of protection of the present invention.
[0050] The following describes a method and system for quickly predicting the harmful gas diffusion process based on a 3DCNN-LSTM coupling model according to an embodiment of the present invention with reference to the accompanying drawings.
[0051] Figure 1 It is a flowchart of a method for quickly predicting the diffusion process of harmful gases based on a 3DCNN-LSTM coupling model in an embodiment of the present invention.
[0052] like Figure 1 As shown, the method includes but is not limited to the following steps:
[0053] S1, based on the working data of the engineering site, the CFD numerical simulation calculation of pollutant migration and diffusion is carried out to obtain the numerical simulation calculation results.
[0054] Specifically, the present invention proposes different working conditions to perform CFD numerical simulation calculations on pollutant migration and diffusion and produce a data set. Specifically, the steps include:
[0055] S101, formulate different parameters such as distance of blasting smoke throwing area, initial harmful gas concentration, wind tube diameter, distance between wind tube outlet and construction face, wind tube outlet wind speed, construction face area, dust mass flow rate, etc., and construct a working condition table;
[0056] S102, based on the working condition table prepared in S101, according to the geometric space parameters such as the distance of the blast smoke throwing area, the diameter of the wind tube, the distance between the wind tube outlet and the construction face in each working condition, a three-dimensional calculation model of the main powerhouse, the main transformer tunnel, the tail gate chamber and each connecting tunnel is established for each working condition;
[0057] S103, based on the three-dimensional models obtained in S102, using the SpaceClaim module of the Ansys platform to perform common node operations, and using the Mesh module to perform grid division mainly based on the structural grid, and exporting them to .msh format files for storage;
[0058] S104, based on the .msh format files obtained in S103, the Fluent module of the Ansys platform is used to perform numerical simulation calculations on the migration and diffusion process of pollutants, and the values of the X-direction wind speed value, the Y-direction wind speed value, the Z-direction wind speed value, the CO concentration value, the dust mass percentage, etc. on all grid nodes are output once every certain time step. Specifically, the S104 step specifically includes:
[0059] S1041, import the .msh file into the Fluent module, add the harmful gas material and start the component transport model, set the velocity inlet for the wind tube outlet according to the wind speed under each proposed scheme, and set the pressure outlet at the outlet of each shaft and cavern, initialize the blasting smoke production area according to the initial concentration of harmful gases in the initial blasting smoke throwing area under each proposed scheme and perform calculations;
[0060] S1042, set to automatically save an ASCII file at a fixed time interval, the content of which should include the X coordinate, Y coordinate, Z coordinate, X-direction speed, Y-direction speed, Z-direction speed and harmful gas concentration of each node;
[0061] S1043, import the .msh file into the Fluent module, start the DPM discrete phase model, set the velocity inlet at the wind tube outlet according to the wind speed under each proposed scheme, and set the pressure outlet at the outlet of each shaft and cavern, set the DPM related parameters for the model tunnel face according to the dust particle size distribution form and mass flow rate under each proposed scheme, initialize and execute the calculation;
[0062] S1044, set to automatically save an ASCII file at a fixed time interval, and the content should include the X coordinate, Y coordinate, Z coordinate, X-axis speed, Y-axis speed, Z-axis speed and dust concentration of each node.
[0063] S2, obtaining the preprocessed tensor data set at each grid node in the numerical simulation calculation results.
[0064] Specifically, this step performs preprocessing based on the gas concentration, wind speed in all directions and other values at each grid node in the CFD numerical simulation calculation results under different working conditions in S1. It includes:
[0065] S201, convert each ASCII file output by the CFD calculation into a five-channel three-dimensional tensor data format for model input, where each channel tensor is an X-axis velocity, Y-axis velocity, Z-axis velocity, harmful gas concentration, and dust concentration tensor that conforms to the actual physical spatial distribution position of each sensor.
[0066] Furthermore, the specific operation of step S201 includes the following steps:
[0067] S2011: Divide the node data of the model calculation result at a certain time point into several XOY sections, and write the 5 channel data of each node into a two-dimensional matrix according to the size of the X coordinate and the Y coordinate;
[0068] S2012: retain the data corresponding to the X and Y coordinates of the sensor layout points in the matrix, and write all other matrix data to 0;
[0069] S2013: splice the two-dimensional matrices of each cross section of each channel data obtained in S2012 into a three-dimensional tensor according to the size of the Z coordinate.
[0070] S202, converting each ASCII file output by the CFD calculation into a two-channel three-dimensional tensor data format output by the model, wherein the two channel tensors are respectively the concentration of harmful gases and the dust concentration tensor that conform to the actual physical space distribution points.
[0071] Furthermore, the specific operation of step S202 includes the following steps:
[0072] S2021, dividing the node data of the model calculation result at a certain time point into a number of XOY sections, and writing the two channel data of each node into a two-dimensional matrix according to the size of the X coordinate and the Y coordinate respectively;
[0073] S2022, splicing the two-dimensional matrices of each cross section of each channel data obtained in S2021 into a three-dimensional tensor according to the size of the Z coordinate.
[0074] S203, according to the set input data time series length and predicted data time series length, the multi-channel tensors obtained in steps S201 and S202 are spliced in the time dimension to obtain the final format of the data set.
[0075] S3, builds a coupled neural network model based on the 3DCNN model and the LSTM model, and inputs the tensor data set into the coupled neural network model to perform model training based on the extracted features to obtain a trained coupled neural network model.
[0076] For example, Figure 2 This is a flowchart of the construction and use of the 3DCNN-LSTM model in the present invention. Specifically, S3 includes the following steps:
[0077] S301, establish a 3DCNN-LSTM coupled neural network model. The spatial feature extraction part of the network includes two 3DCNN layers, two BatchNorm3d layers, two Relu layers, two max_poo13d maximum pooling layers and one fully connected layer; the temporal feature extraction part of the network is an encoder-decoder architecture. The encoder encoder model includes an LSTM layer and a fully connected layer, and the decoder decoder model includes two CNN layers, two BatchNorm2d layers, two Relu layers and two fully connected layers.
[0078] S302, input the tensor data set obtained in S2 into the 3DCNN-LSTM coupled neural network model established in S301, and the 3DCNN layer uses a three-dimensional convolution kernel to align the local voxel blocks of the input tensor data Perform point-by-point multiplication operations. Here, different convolution kernels repeat the above operations to extract the three-dimensional spatial composite feature y l(i,j,z) ,in, is the i-th convolution kernel in the l-th layer. The formula is as follows:
[0079]
[0080] S303, using the BatchNorm3d layer to normalize the feature map output in step S302 to accelerate the training process and improve the stability of the model. This layer calculates the mean and variance of the feature map on the batch data to normalize each voxel, and introduces learnable parameters for further scaling and translation transformations to optimize the training process. The formula is as follows, where μ (l) and (σ (l) ) 2 are the data mean and variance, y (l) and β (l) are learnable scaling and offset parameters, and ∈ is a small parameter that prevents the denominator from being zero:
[0081]
[0082] S304, using Rel u The activation function is used to calculate the feature value z obtained in step S303 l ( i ,j,z) to perform nonlinear transformation. is the eigenvalue output after nonlinear transformation. The formula is as follows:
[0083]
[0084] S305, using the max_pool3d three-dimensional maximum pooling layer to downsample the feature values obtained in step S304, and obtain the maximum value m in the pooling domain l(i,j,z) . B ijz represents the pooling domain, α l(i,j,z) is the activation value of any neuron in the lth layer. The formula is as follows:
[0085]
[0086] S306, repeating steps S303, S304, and S305 to extract deeper features;
[0087] S307, expand the feature map data into a one-dimensional vector, use the fully connected layer to integrate the features extracted from the above layers, map them into the output vector, and input them into the time feature extraction network. The formula is as follows, where o l is the output of the fully connected layer, W l is the weight matrix, b l is the bias vector, m l Is the feature value passed into the fully connected layer:
[0088] o l =W l m l +b l
[0089] S308, input the data output from S307 into the encoder of the temporal feature extraction network, e The ncoder encoder consists of an LSTM layer and a fully connected layer. The LSTM layer extracts and encodes the time series features of the data based on different gating operations.
[0090] Furthermore, the specific operation of step S308 includes the following steps:
[0091] S3081, input the data of each time step output from step S307 into the input gate, where o l is the input one-dimensional vector data, i t is the output of the input gate, which determines how much new information will be added to the cell state. is the new candidate value, which represents the combination of the current input and the previous state. The formula is as follows:
[0092] i t =σ(W i ·[h t-1 , o l ]+bi )
[0093]
[0094] S3082, input the data of each time step outputted in step S307 into the forget gate, where o l is the input one-dimensional vector data, f t is the output of the forget gate, which controls the cell state c at the previous time step t-1 How much information should be forgotten? The formula is as follows:
[0095] f t =σ(W f ·[h t-1 , o l ]+b f )
[0096] S3083, update the unit state, forget the old information and add the new information, the formula is as follows:
[0097]
[0098] S3084, based on the updated unit state, a time step prediction output is performed, where o t is the control signal of the output gate, h t is the output of this time step and the hidden state of the next time step. The formula is as follows:
[0099] o t =σ(W o ·[h t-1 , o l ]+b o )
[0100] h t =o t *tanh(c t )
[0101] S309, input the data obtained in step S308 into the decoder decoder. The decoder consists of two CNN layers, two BatchNorm2d layers, two Relu layers and two fully connected layers. The encoded data obtained in S308 is decoded, and the output one-dimensional vector data is the final predicted data.
[0102] S310, modify the data shape of the one-dimensional vector data obtained in step S309 to the data shape of the input tensor in step S302, and concatenate the x, y, z three-dimensional spatial coordinates of each point data into an ASCII file to facilitate subsequent visualization operations.
[0103] S4, the measured data at the engineering site are input into the trained coupled neural network model to predict the three-dimensional tensor of the spatial distribution of pollutant concentration.
[0104] Specifically, based on the model parameters trained in step S3, the measured data at the engineering site is converted into the input format required by the model to obtain a three-dimensional tensor of the spatial distribution of pollutant concentration. S4 includes the following steps:
[0105] S401, hazardous gas concentration sensors and wind speed sensors are deployed at set points on the project site, connected to the host computer outside the cave in wired form through the RS485 interface of the sensor, and converted to USB interface form through the conversion interface to connect to the host computer outside the cave;
[0106] S402, the host computer outside the cave sends polling instructions to the sensor according to the set interval time, and sends the obtained message to the specified IP and port of the cloud server based on the UDP transmission protocol;
[0107] S403, the cloud server listens to the specified port, parses each message after receiving it, and stores it in the specified MySQL database;
[0108] S404, based on the steps in S201, converting into a tensor data format, and calling the 3DCNN-LSTM prediction model constructed in S3, outputting an ASCII file of the prediction results.
[0109] Furthermore, the method for rapid prediction of harmful gas diffusion process based on 3DCNN-LSTM coupling model in the embodiment of the present invention further includes: S5 performs a WEB-side visualization and intuitive display based on the three-dimensional tensor of pollutant concentration obtained in S4. Specifically, S5 includes the following steps:
[0110] S501, at a set interval, each row of data in the new ASCII file obtained in S404 is concatenated with the corresponding header into a key-value pair, and then the rows of data are concatenated and converted into a json data format file for easy visual reading;
[0111] S502, creating an empty page based on the Vue.js architecture on the Web side, constructing a three-dimensional cloud map based on the THREE.Points particle system API of THREE.js, and completing the dynamic three-dimensional cloud map visualization display.
[0112] Further, S502 includes the following steps:
[0113] S5021, using THREE.Scene scene creation API to initialize the creation of a visual display scene, and recreate it at the time interval set in S501;
[0114] S5022, four thresholds are set for the concentration of harmful gases, and the points in the five intervals divided by the four thresholds are assigned blue, green, yellow, orange, and red color attributes using the THREE.PointsMaterial point material API of THREE.js;
[0115] S5023, using THREE.WebGLRenderer rendering API, the initialized scene and the set particle system are used to complete the visualization of the concentration distribution of harmful gases during the construction period of the underground powerhouse cavern.
[0116] According to the method for rapid prediction of harmful gas diffusion process based on 3DCNN-LSTM coupling model of the embodiment of the present invention, the real-time prediction of the complete pollutant concentration field distribution in the cavern at the future moment is carried out in the time dimension and the space dimension by means of deep learning. On the basis of the existing methods, the real-time and completeness of the pollutant concentration field prediction are improved. Based on the wind speed in each direction and the concentration data of each pollutant monitored in real time by the sensor, it is processed into a multidimensional tensor with geometric meaning of spatial distribution, and the 3D convolutional neural network is used to extract the spatial features implicit in the data of each dimension. At the same time, based on the encoder-decoder structure, the convolutional neural network is used as the encoder and the LSTM long short-term memory neural network is used as the decoder to extract the time features of the time series data of the dimension. This method can effectively predict the concentration field in the cavern at the future moment, which is helpful to guide the construction personnel's entry operation time, and guide the advance frequency conversion control of the air supply fan in the factory cavern, so as to ensure the health and life safety of the construction personnel.
[0117] In order to implement the above embodiment, Figure 3 As shown, this embodiment also provides a harmful gas diffusion process rapid prediction system 10 based on a 3DCNN-LSTM coupling model, the system 10 includes a numerical simulation calculation module 100, a data preprocessing module 200, a network model training module 300 and a data prediction output module 400;
[0118] The numerical simulation calculation module 100 is used to perform CFD numerical simulation calculation of pollutant migration and diffusion based on the working condition data of the engineering site to obtain numerical simulation calculation results;
[0119] A data preprocessing module 200 is used to obtain a preprocessed tensor data set at each grid node in the numerical simulation calculation result;
[0120] A network model training module 300 is used to construct a coupled neural network model based on the 3DCNN model and the LSTM model, and input the tensor data set into the coupled neural network model to perform model training based on the extracted features to obtain a trained coupled neural network model;
[0121] The data prediction output module 400 is used to input the measured data of the engineering site into the trained coupled neural network model to predict and output the three-dimensional tensor of the spatial distribution of pollutant concentration.
[0122] Furthermore, the numerical simulation calculation module 100 includes:
[0123] A working condition table is constructed based on working condition data of the engineering site; wherein the working condition data includes the distance to the blasting smoke throwing area, the initial harmful gas concentration, the diameter of the wind tube, the distance between the wind tube outlet and the construction face, the wind speed at the wind tube outlet, the construction face area and the dust mass flow rate;
[0124] Based on the operating condition table, three-dimensional calculation models of the main powerhouse, main transformer tunnel, tail gate chamber and each connecting tunnel are established for each operating condition;
[0125] Based on the three-dimensional calculation model, common node operation is performed, and grid division with structural grid as the main component is performed to output a file in a preset format;
[0126] Based on the file in the preset format, CFD numerical simulation calculation of the migration and diffusion process of pollutants is performed to obtain numerical simulation calculation results; wherein the numerical simulation calculation results include the X-direction wind speed value, Y-direction wind speed value, Z-direction wind speed value, CO concentration value and dust mass percentage on all grid nodes.
[0127] Furthermore, the data preprocessing module 200 includes:
[0128] Convert the output file of CFD numerical simulation calculation into five-channel three-dimensional tensor data for model input; each channel tensor is respectively the X-axis velocity, Y-axis velocity, Z-axis velocity, harmful gas concentration and dust concentration tensor that conforms to the actual physical space distribution position of each sensor;
[0129] Convert each file output by CFD numerical simulation calculation into two-channel three-dimensional tensor data output by the model, where the two channel tensors are the concentration of harmful gases and the dust concentration tensor that conform to the actual physical space distribution points;
[0130] According to the preset input data time series length and the predicted data time series length, the five-channel three-dimensional tensor data and the two-channel three-dimensional tensor data are spliced in the time dimension to obtain the final tensor data set.
[0131] Furthermore, the network model training module 300 includes:
[0132] A 3DCNN-LSTM coupled neural network model is established based on the 3DCNN model and the LSTM model; wherein the 3DCNN-LSTM coupled neural network model includes a spatial feature extraction network and a temporal feature extraction network; the spatial feature extraction network includes a 3DCNN layer, a BatchNorm3d layer, a Relu layer, a max_pool3d maximum pooling layer, and a fully connected layer; the temporal feature extraction network includes an encoder model and a decoder model;
[0133] Inputting the tensor data set into the 3DCNN-LSTM coupled neural network model, and performing point-by-point multiplication operations on the local voxel blocks of the input tensor data using the three-dimensional convolution kernel of the 3DCNN layer to output a first feature map;
[0134] The feature map is normalized using the BatchNorm3d layer to calculate the mean and variance of the feature map to obtain a second feature map; the second feature map is nonlinearly transformed using the Relu activation function to output a third feature map; and the third feature map is downsampled using the max_pool3d maximum pooling layer to obtain a fourth feature map of the maximum value in the pooling domain;
[0135] The first feature map, the second feature map, the third feature map and the fourth feature map are integrated by using a fully connected layer, mapped into an output vector to input into the temporal feature extraction network to output a feature vector, and the feature vector is input into the encoder model of the temporal feature extraction network to extract time series features;
[0136] The time series features are input into the decoder model for decoding and the output predicted one-dimensional vector is obtained, so as to train the 3DCNN-LSTM coupled neural network model based on the predicted one-dimensional vector to obtain a trained 3DCNN-LSTM coupled neural network model.
[0137] According to the harmful gas diffusion process rapid prediction system based on the 3DCNN-LSTM coupling model of the embodiment of the present invention, the real-time prediction of the complete pollutant concentration field distribution in the cavern at the future moment is carried out in the time dimension and the space dimension by means of deep learning. On the basis of the existing methods, the real-time and completeness of the pollutant concentration field prediction are mainly improved. Based on the wind speed in each direction and the concentration data of each pollutant monitored in real time by the sensor, it is processed into a multidimensional tensor with geometric meaning of spatial distribution, and the 3D convolutional neural network is used to extract the spatial features implicit in the data of each dimension. At the same time, based on the encoder-decoder structure, the convolutional neural network is used as the encoder and the LSTM long short-term memory neural network is used as the decoder to extract the time features of the time series data of the dimension. This method can effectively predict the concentration field in the cavern at the future moment, which is helpful to guide the construction personnel's entry operation time, and guide the advance frequency conversion control of the air supply fan in the factory cavern, so as to ensure the health and life safety of the construction personnel.
[0138] In the description of this specification, the description with reference to the terms "one embodiment", "some embodiments", "example", "specific example", or "some examples" etc. means that the specific features, structures, materials or characteristics described in conjunction with the embodiment or example are included in at least one embodiment or example of the present invention. In this specification, the schematic representations of the above terms do not necessarily refer to the same embodiment or example. Moreover, the specific features, structures, materials or characteristics described may be combined in any one or more embodiments or examples in a suitable manner. In addition, those skilled in the art may combine and combine the different embodiments or examples described in this specification and the features of the different embodiments or examples, without contradiction.
[0139] In addition, the terms "first" and "second" are used for descriptive purposes only and should not be understood as indicating or implying relative importance or implicitly indicating the number of the indicated technical features. Therefore, the features defined as "first" and "second" may explicitly or implicitly include at least one of the features. In the description of the present invention, the meaning of "plurality" is at least two, such as two, three, etc., unless otherwise clearly and specifically defined.
Claims
1. A rapid prediction method for harmful gas diffusion process based on 3DCNN-LSTM coupling model, characterized in that: include: Based on the working condition data of the project site, the CFD numerical simulation calculation of pollutant migration and diffusion is carried out to obtain the numerical simulation calculation results; Obtaining a preprocessed tensor data set at each grid node in the numerical simulation calculation result; Building a coupled neural network model based on the 3DCNN model and the LSTM model, and inputting the tensor data set into the coupled neural network model to perform model training based on the extracted features to obtain a trained coupled neural network model; The measured data at the engineering site is input into the trained coupled neural network model to predict and output a three-dimensional tensor of the spatial distribution of pollutant concentration.
2. The method according to claim 1, characterized in that Based on the working data of the project site, the CFD numerical simulation calculation of pollutant migration and diffusion is carried out to obtain the numerical simulation calculation results, including: A working condition table is constructed based on working condition data of the engineering site; wherein the working condition data includes the distance to the blasting smoke throwing area, the initial harmful gas concentration, the diameter of the wind tube, the distance between the wind tube outlet and the construction face, the wind speed at the wind tube outlet, the construction face area and the dust mass flow rate; Based on the operating condition table, three-dimensional calculation models of the main powerhouse, main transformer tunnel, tail gate chamber and each connecting tunnel are established for each operating condition; Based on the three-dimensional calculation model, common node operation is performed, and grid division with structural grid as the main component is performed to output a file in a preset format; Based on the file in the preset format, CFD numerical simulation calculation of the migration and diffusion process of pollutants is performed to obtain numerical simulation calculation results; wherein the numerical simulation calculation results include the X-direction wind speed value, Y-direction wind speed value, Z-direction wind speed value, CO concentration value and dust mass percentage on all grid nodes.
3. The method according to claim 1, characterized in that Obtaining a preprocessed tensor data set at each grid node in the numerical simulation calculation result, including: Convert the output file of CFD numerical simulation calculation into five-channel three-dimensional tensor data for model input; each channel tensor is respectively the X-axis velocity, Y-axis velocity, Z-axis velocity, harmful gas concentration and dust concentration tensor that conforms to the actual physical space distribution position of each sensor; Convert each file output by CFD numerical simulation calculation into two-channel three-dimensional tensor data output by the model, where the two channel tensors are the concentration of harmful gases and the dust concentration tensor that conform to the actual physical space distribution points; According to the preset input data time series length and the predicted data time series length, the five-channel three-dimensional tensor data and the two-channel three-dimensional tensor data are spliced in the time dimension to obtain the final tensor data set.
4. The method according to claim 3, characterized in that A coupled neural network model is constructed based on the 3DCNN model and the LSTM model, and the tensor data set is input into the coupled neural network model to perform model training based on the extracted features to obtain model parameters, including: A 3DCNN-LSTM coupled neural network model is established based on the 3DCNN model and the LSTM model; wherein the 3DCNN-LSTM coupled neural network model includes a spatial feature extraction network and a temporal feature extraction network; the spatial feature extraction network includes a 3DCNN layer, a BatchNorm3d layer, a Relu layer, a max_pool3d maximum pooling layer, and a fully connected layer; the temporal feature extraction network includes an encoder model and a decoder model; Inputting the tensor data set into the 3DCNN-LSTM coupled neural network model, and performing point-by-point multiplication operations on the local voxel blocks of the input tensor data using the three-dimensional convolution kernel of the 3DCNN layer to output a first feature map; The feature map is normalized using the BatchNorm3d layer to calculate the mean and variance of the feature map to obtain a second feature map; the second feature map is nonlinearly transformed using the Relu activation function to output a third feature map; and the third feature map is downsampled using the max_pool3d maximum pooling layer to obtain a fourth feature map of the maximum value in the pooling domain; The first feature map, the second feature map, the third feature map and the fourth feature map are integrated by using a fully connected layer, mapped into an output vector to input into the temporal feature extraction network to output a feature vector, and the feature vector is input into the encoder model of the temporal feature extraction network to extract time series features; The time series features are input into the decoder model for decoding and the output predicted one-dimensional vector is obtained, so as to train the 3DCNN-LSTM coupled neural network model based on the predicted one-dimensional vector to obtain a trained 3DCNN-LSTM coupled neural network model.
5. The method according to claim 4, characterized in that Inputting the measured data of the engineering site into the trained coupled neural network model to predict and output a three-dimensional tensor of the spatial distribution of pollutant concentration, including: Hazardous gas concentration sensors and wind speed sensors are deployed at set points on the project site, and are converted to USB interfaces through conversion interfaces and connected to the host computer outside the tunnel; The host computer outside the cave sends polling instructions to the sensor at a preset interval, and sends the acquired messages to the specified IP and port of the cloud server based on the UDP transmission protocol; The cloud server listens to the specified port, parses each message after receiving it, and stores it in the specified database; The parsed message in the database is converted into tensor data in a preset format, and the tensor data in the preset format is input into a trained 3DCNN-LSTM coupled neural network model to output a prediction result; wherein the prediction result includes a three-dimensional tensor of the predicted spatial distribution of pollutant concentration.
6. The method according to claim 5, characterized in that After outputting the three-dimensional tensor of the spatial distribution of pollutant concentration, the method further includes: At preset intervals, concatenate each row of data in the ASCII file of the prediction result with the corresponding header into a key-value pair, and then concatenate each row of data into a json data format file; On the Web side, an empty page is created based on the Vue.js architecture, and a three-dimensional cloud map is constructed based on the THREE.Points particle system API of THREE.js to complete the dynamic three-dimensional cloud map visualization display.
7. A rapid prediction system for harmful gas diffusion process based on 3DCNN-LSTM coupling model, characterized in that: include: Numerical simulation calculation module, used to perform CFD numerical simulation calculation of pollutant migration and diffusion based on the working condition data of the engineering site to obtain numerical simulation calculation results; A data preprocessing module, used to obtain a preprocessed tensor data set at each grid node in the numerical simulation calculation result; A network model training module, used for building a coupled neural network model based on the 3DCNN model and the LSTM model, and inputting the tensor data set into the coupled neural network model to perform model training based on the extracted features to obtain a trained coupled neural network model; The data prediction output module is used to input the measured data of the engineering site into the trained coupled neural network model to predict and output the three-dimensional tensor of the spatial distribution of pollutant concentration.
8. The system according to claim 7, characterized in that The numerical simulation calculation module comprises: A working condition table is constructed based on working condition data of the engineering site; wherein the working condition data includes the distance to the blasting smoke throwing area, the initial harmful gas concentration, the diameter of the wind tube, the distance between the wind tube outlet and the construction face, the wind speed at the wind tube outlet, the construction face area and the dust mass flow rate; Based on the operating condition table, three-dimensional calculation models of the main powerhouse, main transformer tunnel, tail gate chamber and each connecting tunnel are established for each operating condition; Based on the three-dimensional calculation model, common node operation is performed, and grid division with structural grid as the main component is performed to output a file in a preset format; Based on the file in the preset format, CFD numerical simulation calculation of the migration and diffusion process of pollutants is performed to obtain numerical simulation calculation results; wherein the numerical simulation calculation results include the X-direction wind speed value, Y-direction wind speed value, Z-direction wind speed value, CO concentration value and dust mass percentage on all grid nodes.
9. The system according to claim 8, characterized in that The data preprocessing module comprises: Convert the output file of CFD numerical simulation calculation into five-channel three-dimensional tensor data for model input; each channel tensor is respectively the X-axis velocity, Y-axis velocity, Z-axis velocity, harmful gas concentration and dust concentration tensor that conforms to the actual physical space distribution position of each sensor; Convert each file output by CFD numerical simulation calculation into two-channel three-dimensional tensor data output by the model, where the two channel tensors are the concentration of harmful gases and the dust concentration tensor that conform to the actual physical space distribution points; According to the preset input data time series length and the predicted data time series length, the five-channel three-dimensional tensor data and the two-channel three-dimensional tensor data are spliced in the time dimension to obtain the final tensor data set.
10. The system according to claim 9, characterized in that The network model training module includes: A 3DCNN-LSTM coupled neural network model is established based on the 3DCNN model and the LSTM model; wherein the 3DCNN-LSTM coupled neural network model includes a spatial feature extraction network and a temporal feature extraction network; the spatial feature extraction network includes a 3DCNN layer, a BatchNorm3d layer, a Relu layer, a max_pool3d maximum pooling layer, and a fully connected layer; the temporal feature extraction network includes an encoder model and a decoder model; Inputting the tensor data set into the 3DCNN-LSTM coupled neural network model, and performing point-by-point multiplication operations on the local voxel blocks of the input tensor data using the three-dimensional convolution kernel of the 3DCNN layer to output a first feature map; The feature map is normalized using the BatchNorm3d layer to calculate the mean and variance of the feature map to obtain a second feature map; the second feature map is nonlinearly transformed using the Relu activation function to output a third feature map; and the third feature map is downsampled using the max_pool3d maximum pooling layer to obtain a fourth feature map of the maximum value in the pooling domain; The first feature map, the second feature map, the third feature map and the fourth feature map are integrated by using a fully connected layer, mapped into an output vector to input into the temporal feature extraction network to output a feature vector, and the feature vector is input into the encoder model of the temporal feature extraction network to extract time series features; The time series features are input into the decoder model for decoding and the output predicted one-dimensional vector is obtained, so as to train the 3DCNN-LSTM coupled neural network model based on the predicted one-dimensional vector to obtain a trained 3DCNN-LSTM coupled neural network model.
Citation Information
Patent Citations
GAT-LSTM-based radionuclide diffusion prediction method and system
CN117313551A
Mine wind environment building pollutant diffusion prediction method and device based on Informer and WRF
CN117852399A
Method of manufacturing emulsified tofu with extended shelf life
KR1020250132600A
Cited By
Environment pollution online monitoring method and system
CN119224238A
Petrochemical park gas leakage diffusion prediction method and system, electronic equipment and storage medium
CN119623358A
Petrochemical Park Gas Leakage Diffusion Prediction Method, System, Electronic Device and Storage Medium
CN119623358B
Hazardous chemical substance loading and unloading leakage diffusion area intelligent prediction method and system
CN119692539A
Gas pollution comprehensive detection method based on selection state space model
CN119804772A
