CN119357659A

CN119357659A - A nuclide diffusion model training, prediction method, system and readable medium

Info

Publication number: CN119357659A
Application number: CN202411322238.6A
Authority: CN
Inventors: 汪建业; 李夏娟; 张子恒; 陈春花; 阮方
Original assignee: Hefei Institutes of Physical Science of CAS
Current assignee: Hefei Institutes of Physical Science of CAS
Priority date: 2024-09-23
Filing date: 2024-09-23
Publication date: 2025-01-24
Anticipated expiration: 2044-09-23
Also published as: CN119357659B

Abstract

The invention relates to the technical field of nuclide diffusion prediction, in particular to a nuclide diffusion model training and prediction method, a nuclide diffusion model prediction system and a readable medium. The nuclide diffusion prediction method comprises the steps of firstly converting acquired time sequence data into Markov transfer field images, then reconstructing the Markov transfer field images through a preprocessing module to obtain reconstruction data capable of more highlighting characteristic information, inputting the reconstruction data into a DDPM network to process the reconstruction data to generate radionuclide diffusion images in a prediction time period, and outputting the radionuclide diffusion images by the DDPM network, wherein the radionuclide diffusion images are actually Markov transfer field images corresponding to time sequence data prediction values in the prediction time period, and obtaining the radionuclide diffusion prediction method through Markov inverse coding of the radionuclide diffusion images. In the invention, the nuclide diffusion model predicts based on the Markov transfer field image, so that the characteristic of the input data is more comprehensively considered, and the prediction result is more accurate.

Description

Nuclide diffusion model training and predicting method, system and readable medium

Technical Field

The invention relates to the technical field of nuclide diffusion prediction, in particular to a nuclide diffusion model training and prediction method, a nuclide diffusion model prediction system and a readable medium.

Background

The radioactive contamination causes irreversible damage to humans and the natural environment. Therefore, predictive research of radionuclide diffusion is of great importance, helping to scientifically guide nuclear emergency decisions.

The research of nuclide diffusion prediction relates to a plurality of subjects such as nuclear physics, atmospheric environment science, computer science and the like, and mature theoretical basis and application modes have been developed. The traditional nuclide diffusion prediction method comprises an atmospheric tracing experiment, a wind tunnel experiment and numerical simulation prediction. The atmospheric tracing experiment refers to the release of the tracer in a specific position, and the atmospheric flow characteristics and the tracer diffusion rule in the regional environment are obtained through sampling and monitoring. Common gas tracers are SF6 gas, SF6 is stable in nature, and the concentration of SF6 can be detected by gas chromatography. The wind tunnel experiment is similar to the atmospheric tracing experiment principle, and the regional atmospheric flow characteristics and the diffusion rule are reflected through the tracer. Unlike the atmospheric tracing experiment, which is performed in a real scene, the wind tunnel test is performed in a model scaled up and down to a certain scale for the real scene. The wind tunnel experiment has the advantages that the cost is reduced, the control of meteorological conditions and the underlying surface environment is realized, and the diffusion prediction can be performed by a quantitative research mode.

The traditional atmospheric diffusion numerical model mainly comprises a Gaussian smoke plume model, a Lagrangian smoke mass model, an Euler model and computational fluid dynamics. However, these conventional nuclide diffusion prediction methods all rely on source item information of radionuclides, which makes it difficult for conventional nuclide simulation methods to be quickly applied to sudden nuclide diffusion scenes with uncertain places, complex terrains and changeable weather.

Disclosure of Invention

In order to overcome the defect that the nuclide diffusion prediction method in the prior art is too dependent on source item information of radionuclides and is difficult to adapt to sudden nuclide diffusion scenes, the invention provides a training method of a nuclide diffusion model, which is applicable to any scene and can predict nuclide diffusion conditions at high speed and accurately.

The invention provides a training method of a nuclide diffusion model, which comprises the following steps:

SA1, constructing a basic model, a first data set and a second data set;

The system comprises a first data set, a second data set, a first data set, a second data set and a third data set, wherein the first data set is used for storing time sequence data of a missing part of time points, the second data set is used for storing time sequence data which is complete and marked with radionuclide diffusion images, and the time sequence data comprises monitoring objects which are acquired by a monitoring area at a plurality of continuous monitoring time points;

The basic model comprises a preprocessing module and a DDPM network, wherein the input of the preprocessing module is the input of the nuclide diffusion model, and the output of the DDPM network is the output of the nuclide diffusion model;

SA2, processing time sequence data in the first data set and the second data set into Markov transition field images;

SA3, training a preprocessing module on the first data set until convergence;

SA4, extracting a second training sample from the second data set, performing deletion processing on time sequence data of the second training sample, and converting the time sequence data into a Markov transfer field image serving as a deletion sample;

SA5, processing the missing sample by a preprocessing module to obtain a blurred image, and processing the blurred image by a DDPM network to obtain a predicted radionuclide diffusion image serving as a predicted value;

SA6, taking the radionuclide diffusion image associated with the second training sample as a true value, and calculating loss by combining the predicted value and the true value;

And SA7, judging whether the loss converges, if not, updating DDPM the network through back propagation of the loss, and returning to the step SA4, and if so, fixing the basic model as a nuclide diffusion model.

Preferably, in SA2, the method for processing the time series data into the Markov transition field image comprises the steps of processing the time series data into the time series of each monitoring object, processing the time series of each monitoring object into the Markov code image by adopting a Markov transition matrix, and then carrying out weighted superposition on the Markov code images of each monitoring object to obtain the Markov transition field image.

Preferably, in SA3, the method for training the preprocessing module on the first data set includes that the preprocessing module processes the Markov transfer field image in the first data set and outputs a reconstructed image as a blurred image, DDPM generates a radionuclide diffusion image according to the blurred image, updates parameters of the preprocessing module according to an optimization target until the preprocessing module converges, and the optimization target is:

A markov transfer field image representing the input of the preprocessing module, X represents the radionuclide diffusion image output by DDPM; The DDPM network is used for guiding generation; representing the basic model under the control of parameter eta AndThe probability of generating x, i.e. the preprocessing module input isThe output isThe DDPM network outputs a probability of x.

Preferably, the monitoring object further includes one or more of atmospheric pressure, wind speed, temperature, and humidity.

Preferably, the wind speed takes the average of the X-axis wind speed, the Y-axis wind speed and the Z-axis wind speed.

Preferably, the pre-treatment module is a beta-VAE module.

The invention provides a nuclide diffusion prediction method adopting the training method of a nuclide diffusion model, which comprises the following steps:

S1, acquiring a nuclide diffusion model by adopting the training method of the nuclide diffusion model;

S2, randomly sampling the time sequence data to form new time sequence data of a missing part time point, and converting the new time sequence data into a Markov transfer field image;

s3, inputting the Markov transfer field image into a nuclide diffusion model, and carrying out reconstruction data on the Markov transfer field image by a preprocessing module, processing the reconstruction data by a DDPM network and outputting a radionuclide diffusion image;

S4, carrying out Markov inverse coding on the radionuclide diffusion image to obtain a predicted value of time sequence data formed by monitoring objects in a monitoring area in a predicted time period, and extracting a radionuclide concentration predicted sequence from the predicted value to serve as a radionuclide diffusion predicted result.

Preferably, in S2, the method for converting the new time sequence data into the Markov transition field image comprises the steps of processing the new time sequence data into time sequences of all monitoring objects, processing the time sequences of all the monitoring objects into Markov code images by adopting a Markov transition matrix, and then carrying out weighted superposition on the Markov code images of all the monitoring objects to obtain the Markov transition field image.

The invention provides a nuclide diffusion prediction system, which comprises a memory and a processor, wherein a computer program is stored in the memory, the processor is connected with the memory, and the processor is used for executing the computer program to realize the nuclide diffusion prediction method.

The invention provides a readable medium storing a computer program which is used for realizing the nuclide diffusion prediction method when being executed.

The invention has the advantages that:

(1) The nuclide diffusion prediction method comprises the steps of firstly converting acquired time sequence data into Markov transfer field images, then reconstructing the Markov transfer field images through a preprocessing module to obtain reconstruction data capable of more highlighting characteristic information, inputting the reconstruction data into a DDPM network to process the reconstruction data to generate radionuclide diffusion images in a prediction time period, and outputting the radionuclide diffusion images by the DDPM network, wherein the radionuclide diffusion images are actually Markov transfer field images corresponding to time sequence data prediction values in the prediction time period, and obtaining the radionuclide diffusion prediction method through Markov inverse coding of the radionuclide diffusion images. In the invention, the nuclide diffusion model predicts based on the Markov transfer field image, so that the characteristic of the input data is more comprehensively considered, and the prediction result is more accurate.

(2) In the invention, the preprocessing module adopts the beta-VAE module, and the generated image is taken as a priori condition to be brought into DDPM networks, so that DDPM conditionally generates a corresponding nuclide diffusion Markov transfer field image, and the reliability of prediction is further improved.

(3) In the invention, firstly, the time series Markov code image of each monitoring object is obtained, then the multi-feature coupled Markov transfer field image is obtained through weighted superposition, so that the input of the nuclide diffusion model deeply expresses the monitoring features in a simple and image mode, thereby facilitating the feature mining in the model processing process.

(4) According to the training method of the nuclide diffusion model, in the training process, the supervised training of the preprocessing module and the unsupervised training of DDPM are mutually independent, so that the overall convergence speed in the model training process is improved. The result of the preprocessing module, as one of the inputs to DDPM, directs DDPM to generate the final image, further improving the training accuracy and efficiency of the DDPM network.

Drawings

FIG. 1 is a flow chart of a training method of a nuclide diffusion model;

FIG. 2 is a schematic diagram of a training process of a species diffusion model;

FIG. 3 is a flow chart of a method for predicting the diffusion of a species;

FIG. 4 is a complex underlying geometric model composed of three different vegetation;

FIG. 5 is a schematic view of a monitoring area;

FIG. 6 shows radionuclide concentration variation for each node;

FIG. 7 shows the result of radionuclide concentration prediction at node A1.

Detailed Description

The following description of the embodiments of the present invention will be made clearly and completely with reference to the accompanying drawings, in which it is apparent that the embodiments described are only some embodiments of the present invention, but not all embodiments. All other embodiments, which can be made by those skilled in the art based on the embodiments of the invention without making any inventive effort, are intended to be within the scope of the invention.

DDPM (conditional diffusion model) is a generation model for learning the distribution of target data from a sample, which consists of two markov processes, one is a fixed forward process simulating from a real picture to random gaussian noise, and the other is a reverse process based on learning, which generates an image with similar characteristics as the original picture. The form of the forward process in DDPM can be summarized as follows:

Wherein c represents the control condition added in DDPM, q (x _1：T|x₀, c) represents the joint conditional probability distribution from the initial state x ₀ to the time step T under the guidance of the condition c, describing the whole process of gradually adding noise from the original data x ₀ to x _T, T being the set maximum time step, T being the time step number;

q (x _t|x_t-1, c) represents a conditional probability distribution of state x _t-1 for time step t-1 to state x _t for time step t under the direction of condition c, this process typically involves adding random noise to x _t-1 to obtain a noise addition strategy in the forward process denoted by x _t;δ₁、δ₂、...、δ_t、...、δ_T; Mean value is expressed as The standard deviation is Gaussian distribution of delta _t I, wherein I represents an identity matrix;

q (x _t|x₀, c) represents the conditional probability distribution of state x _t from initial state x ₀ to time step t under the guidance of condition c; Mean value is expressed as Standard deviation isIs of Gaussian distribution, and hasγ_t＝1-δ_t。

DDPM the inverse process can also be parameterized with a gaussian transfer distribution obtained by learning, as follows:

Wherein p (x _0：T |c) represents a joint probability distribution of backtracking from the final noise state x _T to the original state x ₀ under the influence of the condition c, describing a process of gradually recovering data from noise;

p _λ(x_t-1|x_t, c) represents the conditional probability distribution of state x _t of time step t back to state x _t-1 of time step t-1 under the influence of condition c, this process being parameterized by a learnable model, λ being the parameter of DDPM.

Let q (x _T|x₀, c) represent a conditional probability distribution from the initial state x ₀ to x _T under the guidance of condition c, when T is sufficiently large, the distribution q (x _T|x₀, c) may be approximated as an isotropic gaussian distribution and λ may be trained based on the variance inference. However, as can be seen from the generation process DDPM in equation 1 and equation 1, adding the auxiliary condition c to DDPM can improve the accuracy of sample generation and reduce the instability of DDPM. To this end, the present invention proposes a DDPM framework that can be coupled to any preprocessing module. By usingRepresenting a generic preprocessing module, its optimization objective function can be expressed as:

η is a parameter of an integral model formed by connecting the pretreatment module with the DDPM framework front and back; Representing the original time sequence data, i.e. the input data of the preprocessing module, x representing the original time sequence data An image reconstructed through DDPM after being processed by the preprocessing module; Representing prior information, namely data output by the preprocessing module, and also being a condition c in the formula 1; the DDPM network is used for guiding x generation; Representing the model parameters eta under control of AndThe probability of generating x, i.e. DDPM at the input isAndThe probability of x is output.

To obtain the original state x ₀, x in equation (3) is replaced with x ₀ so that DDPM is according toAndThe output is x ₀ and the following inequality is obtained from the lower bound of evidence:

Wherein phi is a parameter of DDPM inverse processes, Representation DDPM at input asAndThe probability of the time output being x ₀; Representation of Is used as a means for controlling the speed of the vehicle,Representing raw time series dataThe probability of the data x ₀ obtained through the preprocessing module and DDPM reconstruction plus noise to generate a data sequence x _1:T is represented by p (x _0:T) which represents the joint probability distribution of the backtracking from the final noise state x _T to the original state x ₀; Output data representing DPPM combined with pre-processing module And the original time sequence dataObtaining the probability of a data sequence x _0:T; Representing the preprocessing module according to GeneratingIs a function of the probability of (1),Representation ofAnd ψ represents the parameters of the encoder in the preprocessing model.

The inequality (4) shows that in the controllable DDPM training guided by the preprocessing module, the supervised training of the preprocessing module and the unsupervised training of DDPM are independent of each other, the result of the generation of the preprocessing moduleAs one of the inputs to DDPM, guide DDPM generates the final image.

Based on the derivation of the theory, the invention adopts a beta-VAE module (beta-Variational Autoencoder, beta-variation self-encoder) as a preprocessing module, and the data input by the preprocessing moduleUsing Markov transfer field images, using images generated by a preprocessing module as a priori conditionsIs brought into DDPM model, so that DDPM conditionally generates corresponding nuclide diffusion markov transfer field images.

In the invention, an integral model formed by a beta-VAE module and a DDPM network is named as a beta-VDRDPF model, the input of the beta-VAE module is used as the input of a beta-VDRDPF model, the input data of the beta-VDRDPF model adopts a Markov transfer field image, and the DDPM network is combined with the Markov transfer field image input by the beta-VDRDPF model and the output data of the beta-VAE module to generate a radionuclide diffusion image.

The following describes the data processing of the β -VAE module, DDPM model, and markov transfer field, respectively.

The beta-VAE module inputs a Markov transfer field image serving as an initial state X ₀, wherein the Markov transfer field image is obtained by converting acquired time sequence data, and the time sequence data comprises acquisition time(s), acquisition point acquisition radionuclide I-131 concentration (KBq/m ³), acquisition point atmospheric pressure (Pa), X-axis wind speed (m/s), Y-axis wind speed (m/s), Z-axis wind speed (m/s) and three-dimensional coordinates of the acquisition point.

The beta-VAE module comprises an encoder and a decoder, input data x ₀ is encoded by the encoder, encoding characteristic data output by the encoder is weighted into potential codes z, the potential codes z do not directly participate in DDPM training and are not output, and the potential codes z are only used in the beta-VAE module training and are specifically expressed in the following formulas (5) - (8).

In the training process of the beta-VAE module, an auxiliary signal y is needed as an input, the beta-VAE module processes the auxiliary signal y to generate output data, and the beta-VAE module generates potential codes z, DDPM generates a time sequence x _1:T according to the output data of the beta-VAE module, T is the total number of time steps, and thus, the joint distribution p (x _0：T, y, z) of the beta-VDRDPF model can be expressed as follows:

p(x_0：T,y,z)＝p(z)p_θ(y|z)pφ(x_0:T|y,z) (5)

Wherein θ represents the parameters of the decoder in the β -VAE module, φ represents the parameters of the DPPM network inverse process, p (z) represents the probability that the β -VAE module generates the potential code z, p _θ (y|z) represents the probability that the β -VAE module generates the potential code z according to the input data y when the decoder parameters in the β -VAE module are θ, and p _φ(x_0：T |y, z) represents the probability that the β -VDRDPF model generates the time series data x _0：T according to the input data y when the inverse process parameters of the DPPM network are φ and the potential code generated by the β -VAE module is z.

Since the true joint posterior distribution p (x _1：T,z|y,x₀) is difficult to calculate, using the approximate proxy distribution q (x _1：T,z|y,x₀) to approximate substitution, q (x _1：T,z|y,x₀) can also be described as the following conditional distribution:

q(x_1：T,z|y,x₀)＝q_ψ(z|y,x₀)q(x_1：T|y,z,z₀) (6)

Where p (x _1：T,z|y,x₀) represents the true joint posterior distribution of the β -VDRDPF model, i.e., the probability that the β -VDRDPF model generates the latent code z and the time series data x ₁:_T when the input is y and the output of the β -VAE module is x ₀, ψ represents the parameters of the β -VAE module, q _ψ(z|y,x₀) represents the probability that the β -VAE module outputs the latent code z when the input is y and the output is x ₀, and q (x _1：T|y,z,x₀) represents the probability that the β -VDRDPF model generates the time series x _1:T when the input of the β -VAE module is y, the output is x ₀, and the latent code is z.

During the training process, the forward process of DDPM does not perform parameter updating or training, but rather keeps its parameters fixed. The log likelihood of the markov transfer field data for training can be expressed as:

lgp(x₀,y)＝lg∫p(x_0：T,y,z)dx_1：Tdz (7)

Wherein p (x ₀, y) represents the probability that the β -VAE module outputs x ₀ when the input is y, p (x _0:T, y, z) represents the probability that the β -VDRDPF model generates a time series x _0:T when the β -VAE module inputs y and the generated potential code is z;

since equation (7) is difficult to estimate analytically, the present invention optimizes the lower bound of evidence corresponding to the log of likelihood

Wherein, psi is the parameter of the encoder in the beta-VAE module, theta is the parameter of the decoder in the beta-VAE module, phi is the parameter of the DDPM inverse process, p _θ (y|z) represents the probability that the decoder of the beta-VAE module generates the input data y according to the potential code z, q _ψ(z|y,x₀) represents the probability that the encoder of the beta-VAE module generates the potential code z when the input is y and the output is x ₀; Representing the expectation of p _θ (y|z) calculated on probability distribution q _ψ(z|y,x₀);

KL divergence between q _ψ(z|y,x₀) and p (z), p (z) representing the probability of the β -VAE module generating a potential code z, β representing the hyper-parameters of the β -VAE module controlling the KL divergence weight, determining the regularization degree of the β -VAE module;

p _φ(x_0:T |y, z) represents the probability that the DDPM inverse generates x _0:T based on the input data y under the direction of the potential code z, q (x _1:T|y,z,x₀) represents the probability of x _1:T generated by DDPM in the β -VDRDPF model when the β -VAE module input data is x ₀ for y output and the potential code z is generated; A desired ratio of p _φ(x_0：T |y, z to q (x _1：T|y,z,x₀) on q (x _1：T|y,z,x₀); Representing the desire to calculate over a z-distribution that specifies the input of the β -VAE module as the output x ₀.

In the above-mentioned (8),A loss term for the β -VAE module that represents the expectation that the potential code z generated by its encoder can reconstruct the conditional signal y; A loss term of DDPM, representing the desired ratio between the back diffusion process p _φ(x_0：T y, z) and the forward diffusion process q (x _1：T|y,z,x₀); Representing a desired calculation of a latent variable z output by the VAE encoder network; Representing the loss function and contribution of the beta-VAE module, A loss function and contribution representing DDPM;

in order to enable the beta-VDRDPF model provided by the invention to quickly converge in the training process, the following design mode is used for simplifying part of network parameters when parameter design is carried out, so that the formula (8) is rewritten into the following formula (9) and formula (10).

1) To ensure the mapping relationship between y and x ₀, it is assumed that y is the Markov transfer field image x ₀ itself, and the back diffusion process of the diffusion model is not performed on the condition of y in the training process, but is regarded as p _φ(x_0：T |z in formula (5). At this time, the formula (8) can be rewritten as:

Wherein q _ψ(z|x₀) represents the posterior distribution of the latent variable z generated by the encoder of the β -VAE module from the input image x ₀, and p _θ(x₀ |z) represents the probability of the decoder of the β -VAE module generating the input image x ₀ from the latent code z; Representing the expectation of calculating p _θ(x₀ |z) over a q _ψ(z|x₀) distribution, for characterizing the expected reconstruction loss of the a posteriori distribution q _ψ(z|x₀) of z given x ₀;

p (x ₀) represents the probability that the β -VAE module generates x ₀ and p (z) represents the probability that the β -VAE module generates z; The beta represents the super-parameter for controlling the weight of the KL divergence, and determines the regularization degree of the model;

p _φ(x_0:T |z) represents the probability that the DDPM inverse generates x _0:T based on the input data y under the direction of the potential code z, i.e., the inverse distribution of image x _0：T is gradually recovered from noise at a given z, q (x _1:T|z,x₀) represents the probability that the β -VAE module input data is x ₀ and generates x _1:T generated by DDPM in the β -VDRDPF model at a potential code z, i.e., the forward distribution of noise diffusion process x _1：T at a given z and x ₀; Is the desired ratio of p _φ(x_0:T |z to q (x _1:T|z,x₀) on q (x _1:T|z,x₀), i.e., the desired ratio between the back diffusion process p _φ(x_0：T |z and the forward diffusion process q (x _1：T|z,x₀); Representing the desire to calculate over the z-distribution of the input image x ₀, i.e., the potential variable z of the encoder output in the β -VAE module given x ₀.

2) In performing back diffusion of the model, the potential code z of x ₀ generated by the β -VAE module is not directly used as input of the diffusion model in the invention, but reconstructed from the β -VAE moduleTraining to guide the reverse transfer process of diffusion model, and reconstructing beta-VAEIs a deterministic function of z. At this time, the formula (8) can be rewritten as:

representing the output of the β -VAE module when x ₀ is the input; Representing a given And x ₀, the forward process distribution of noise diffusion process x _1:T; Representing a given Gradually recovering the inverse process distribution of the time sequence x _0：T from the noise; Representing the expectation of the ratio of the inverse process distribution to the forward process distribution of DDPM; representing, for a given input image x ₀, the output from Reconstruction of images by in-distribution samplingIs not limited to the desired one; Reconstructed image representing beta-VAE module Is used for the calculation of the expected calculation of (a).

The lower bound expression of equation (10) includes two parts, the first part being ELBO (lower bound of variation) of the β -VAE module, measuring reconstruction loss and KL divergence between the distribution of latent variable z and the prior distribution, and the second part being DDPM loss, measuring reconstruction image fromThe departure resumes the degree of matching of the inverse process of the original image x ₀ with the forward process.

3) According to formula (8), it can be seen that the present invention uses a two-step training strategy in the training process, the first step of optimizationSecond step of optimizationThe second training phase fixes the parameters θ and ψ simultaneously and freezes the parameters of the β -VAE module in the first step, so its final joint distribution can be defined as:

In the formula, Representing the beta-VDRDPF model at a given pointThe conditional probability of x _0：T is generated under the condition of (1); For generation by beta-VAE The probability that the current will be the same,Representation DDPM givenThe conditional probability of x _0：T is generated.

And the A sensors acquire A Markov transfer field images, and the A Markov transfer field images are weighted and overlapped to acquire a final Markov transfer field image.

Based on the deduction, a nuclide diffusion model formed by a beta-VAE module and a DDPM network is provided in the embodiment, wherein the input of the beta-VAE module is the input of the nuclide diffusion model, the output of the beta-VAE module is connected with the input of the DDPM network, and the output of the DDPM network is the output of the nuclide diffusion model.

The input of the nuclide diffusion model is Markov transfer field image y, beta-VAE module which is obtained by converting time sequence data collected by a monitoring area, the Markov transfer field image is encoded and decoded by the beta-VAE module, and the decoded data is output by the beta-VAE moduleAnd generating a radionuclide diffusion image x ₀ as a radionuclide diffusion prediction result through DDPM network processing, namely a Markov transfer field image corresponding to time sequence data in a prediction time period.

In this embodiment, the time series data includes monitoring objects acquired by the monitoring area at a plurality of consecutive monitoring time points, and the monitoring objects include radionuclide concentration (kBq/m ³), atmospheric pressure (Pa), wind speed (m/s), and temperature and humidity. The wind speed adopts the average value of X-axis wind speed (m/s), Y-axis wind speed (m/s) and Z-axis wind speed (m/s) in the monitoring area.

The time series data is essentially composed of time series of each monitoring object, and the method for processing the time series data into Markov transition field images comprises the steps of processing the time series of each monitoring object into Markov code images by adopting a Markov transition matrix, and then carrying out weighted superposition on the Markov code images of each monitoring object to obtain the Markov transition field images with multi-feature coupling as the input of a nuclide diffusion model.

When the Markov code images of all the monitoring objects are subjected to weighted superposition, the weight of the radionuclide concentration is not less than 0.95, and the sum of the weight of the temperature and the humidity is not more than 0.01.

It is noted that the number of monitoring time points included in the input of the nuclide diffusion model is equal to the data amount of time points included in the output of the nuclide diffusion model, and assuming that the time sequence data formed by the monitoring objects collected at 1-M monitoring time points in the monitoring area are converted into Markov transfer field images to be input into the nuclide diffusion model, the radionuclide diffusion images output by the nuclide diffusion model are subjected to Markov inverse coding, the predicted value of the time sequence data formed by the monitoring objects in the monitoring area at M+1 to 2M monitoring time points can be obtained, and then the predicted time sequence of the radionuclide concentration can be extracted from the predicted value.

Referring to fig. 1 and 2, the training method of the nuclide diffusion model provided in the present embodiment includes the following steps:

SA1, constructing a basic model, a first data set and a second data set, wherein the basic model comprises a preprocessing module and DDPM networks, and the preprocessing module specifically adopts a beta-VAE module.

The preprocessing module comprises an encoder and a decoder, wherein the encoder encodes an input Markov transfer field image and generates a latent feature z, the decoder decodes the latent feature z and outputs a reconstructed image, and the reconstructed image is processed into a radionuclide diffusion image through DDPM networks and is output.

The second data set is used for storing the time sequence data which is complete and marked with the radionuclide diffusion image;

SA3, training a preprocessing module on a first data set until convergence, wherein in the training process, the preprocessing module processes a Markov transfer field image in the first data set and outputs a reconstructed image as a blurred image, DDPM generates a radionuclide diffusion image according to the blurred image, in the training process, parameters of the preprocessing module are updated according to the following optimization targets, parameters of DDPM are fixed, and the optimization targets are as follows:

Specifically, step SA3 includes the following sub-steps:

SA31, extracting a first training sample from the first data set, substituting a Markov transfer field image of the first training sample into a basic model, and optimizing parameters of the preprocessing module by combining the optimization targets;

SA32, judging whether the iteration times of the preprocessing module reach a set value, if so, finishing the pre-training of the preprocessing module and executing step S4, and if not, returning to step SA31.

Specifically, in the step, firstly, random sampling is performed on time sequence data in a second training sample to form a data sequence of a missing part time point, and then the data sequence of the missing part time point is converted into a Markov transfer field image serving as a missing sample;

And SA7, judging whether the loss converges, if not, updating DDPM the network through the back propagation of the loss, and returning to the step SA4, if so, fixing parameters of the preprocessing module and the DDPM network, substituting the parameters into a basic model to obtain a nuclide diffusion model which is input into a Markov transfer field image and is output into a radionuclide diffusion image.

Referring to fig. 3, when the trained nuclide diffusion image is used for nuclide diffusion prediction, the following steps are performed:

s1, adopting the steps SA1-SA7 to obtain a nuclide diffusion model, and collecting time sequence data in a set monitoring area;

S3, inputting the Markov transfer field image into a nuclide diffusion model, encoding and decoding the Markov transfer field image by a preprocessing module, and outputting decoding data, and processing the decoding data by a DDPM network and outputting a radionuclide diffusion image;

The above-described nuclide diffusion prediction method is described below with reference to specific embodiments.

In the embodiment, radionuclide diffusion original time sequence data is generated through CFD software OpenFOAM simulation. The data simulation scene is used for establishing a complex underlying surface geometric model consisting of three different vegetation for leakage accidents of spent fuel in the transportation process, and the model is shown in fig. 4. The model calculation area is a three-dimensional space of 1000 meters multiplied by 100 meters, the left plane of the cube is an air inlet, the right plane is an air outlet, the airflow direction is set to be positive along the x-axis, and the wind speed is set to be 4 meters/s. The red solid dots in FIG. 4 represent the source of radionuclide I-131 released, with coordinates (400 meters, 500 meters, 1 meter), assuming a continuous release of the source for 1 hour with a release rate of 3.3 (kBq/cubic meter)/second. The half-life of radionuclide I-131 is 8.02 days. kBq is kilo-belleville and is a unit indicating radioactivity.

The underlying pattern comprises three different vegetation groups, different vegetation having different capacities to adsorb and change the flow field, which is known as the vegetation effect. The essence of the vegetation effect is that the vegetation is used as a porous medium, so that pressure loss can be generated on the diffusion fluid passing through the surface of the vegetation, and the pressure loss generated by different vegetation types and different density degrees is different. Therefore, the invention represents different vegetation undersides by setting different pressure loss coefficients, wherein the pressure loss coefficient of a vegetation group A is set to 0.5/m, the space size is 120m multiplied by 60 m multiplied by 15 m, the pressure loss coefficient of a vegetation group B is set to 2/m, the space size is 60 m multiplied by 140 m multiplied by 15 m, the pressure loss coefficient of a vegetation group C is set to 8/m, and the space size is 60 m multiplied by 140 m multiplied by 15 m.

According to the above setting, the diffusion data of I-131 are obtained through simulation by a CFD method, wherein the time step of diffusion data collection is 10 seconds, the total simulated diffusion time is 120 minutes, and finally 720 groups of original data are obtained. The raw data includes time(s), radionuclide I-131 concentration (KBq/m ³), atmospheric pressure (Pa), X-axis wind speed (m/s), Y-axis wind speed (m/s), Z-axis wind speed (m/s), and three-dimensional coordinates. Table 1 shows a partial data sample generated by CFD software OpenFOAM.

Table 1 diffusion 60 min part of raw data presentation

Table 2 shows three monitoring areas selected among three different vegetation, the monitoring points being given by fig. 5, wherein red solid dots represent radionuclide leakage sources and blue solid dots represent monitoring points.

Table 2 monitoring point coordinates

In the invention, a monitoring area is abstracted into nodes, radionuclide concentration, wind speed, atmospheric pressure and underlying surface characteristics in the monitoring area are abstracted into node characteristics, image coding is realized by using a Markov transfer field according to different node characteristics, and then weighted superposition is carried out to obtain a multi-characteristic coupled Markov transfer field image. Fig. 6 shows the variation of radionuclide concentration at each node in fig. 5.

Specifically, in the embodiment, firstly, time series of the concentration of the radionuclide I-131 is subjected to Markov coding to obtain a Markov coded image of the concentration, and time series of atmospheric pressure is subjected to Markov coding to obtain a Markov coded image of the pressure;

Then, carrying out weighted superposition on the concentrated Markov code image, the pressure Markov code image and the wind speed Markov code image to form a Markov transfer field image;

And finally, inputting the Markov transfer field image into a nuclide diffusion model to obtain a predicted radionuclide diffusion image, and carrying out Markov inverse coding on the radionuclide diffusion image to obtain the radionuclide I-131 concentration at M continuous time points.

The invention takes LSTM, RNN and GAT-LSTM as comparison models, and takes Markov transfer field images of nodes as inputs, so that radionuclide diffusion images of the nodes are predicted, and the radionuclide concentration change condition of the nodes is obtained.

In this experiment, data from 600s to 3310s were all selected as model training sets, and data from 3320s to 3620s were selected as model test sets. Table 3 shows a quantitative analysis of the accuracy of predictions for each of the three detection regions in FIG. 5. FIG. 7 shows the results of the method of the present invention, LSTM, RNN and GAT-LSTM on the prediction of the concentration of radioactivity at node A1. It can be seen from fig. 7 that the predicted result of β -VDRDPF proposed by the present invention is closest to the true value. Meanwhile, according to quantitative result analysis in the table 3, it can be seen that when the radionuclide concentration change of different nodes in different areas is predicted, the method provided by the invention keeps higher accuracy, and is superior to other three methods, and the algorithm provided by the invention has higher accuracy on the radionuclide concentration change prediction.

TABLE 1 quantitative analysis of radionuclide concentration prediction accuracy at different nodes for different methods

It will be understood by those skilled in the art that the present invention is not limited to the details of the foregoing exemplary embodiments, but includes other specific forms of the same or similar structures that may be embodied without departing from the spirit or essential characteristics thereof. The present embodiments are, therefore, to be considered in all respects as illustrative and not restrictive, the scope of the invention being indicated by the appended claims rather than by the foregoing description, and all changes which come within the meaning and range of equivalency of the claims are therefore intended to be embraced therein. Any reference sign in a claim should not be construed as limiting the claim concerned.

Furthermore, it should be understood that although the present disclosure describes embodiments, not every embodiment is provided with a separate embodiment, and that this description is provided for clarity only, and that the disclosure is not limited to the embodiments described in detail below, and that the embodiments described in the examples may be combined as appropriate to form other embodiments that will be apparent to those skilled in the art.

The technology, shape, and construction parts of the present invention, which are not described in detail, are known in the art.

Claims

1. The method for training the nuclide diffusion model is characterized by comprising the following steps of:

SA1, constructing a basic model, a first data set and a second data set;

SA3, training a preprocessing module on the first data set until convergence;

2. The method for training a nuclide diffusion model according to claim 1, wherein SA2 is a method for processing time series data into Markov transition field images, wherein the time series data are processed into time series of each monitoring object, the time series of each monitoring object are processed into Markov code images by using a Markov transition matrix, and then the Markov code images of each monitoring object are weighted and overlapped to obtain the Markov transition field images.

3. The method for training a radionuclide diffusion model according to claim 1, wherein in SA3, the method for training the preprocessing module on the first data set is that the preprocessing module processes the Markov transfer field image in the first data set and outputs a reconstructed image as a blurred image, DDPM generates a radionuclide diffusion image according to the blurred image, updates parameters of the preprocessing module according to an optimization target until the preprocessing module converges, and the optimization target is:

4. The method of training a species diffusion model of claim 1, wherein the monitored subject comprises a model further comprising one or more of atmospheric pressure, wind speed, temperature, and humidity.

5. The method of claim 4, wherein the wind speed is the mean of the X-axis wind speed, the Y-axis wind speed, and the Z-axis wind speed.

6. The method of claim 4, wherein the preprocessing module is a β -VAE module.

7. A nuclide diffusion prediction method using the training method of the nuclide diffusion model according to any one of claims 1 to 5, comprising the steps of:

S1, acquiring time sequence data in a set monitoring area by adopting the training method of the nuclide diffusion model as claimed in any one of claims 1 to 5 to obtain the nuclide diffusion model;

8. The nuclide diffusion prediction method according to claim 7, wherein in S2, the method of converting the nascent time series data into the Markov transition field image is that the nascent time series data is processed into the time series of each monitoring object, the time series of each monitoring object is processed into the Markov encoded image by using the Markov transition matrix, and then the Markov encoded images of each monitoring object are weighted and overlapped to obtain the Markov transition field image.

9. A species diffusion prediction system comprising a memory and a processor, the memory having stored therein a computer program, the processor being coupled to the memory, the processor being configured to execute the computer program to implement the species diffusion prediction method of claim 7.

10. A readable medium, characterized in that a computer program is stored, which computer program, when executed, is adapted to implement the nuclide diffusion prediction method of claim 7.