Nuclide diffusion model training method and system, nuclide diffusion model prediction method and system and readable medium

Abstract

The invention relates to the technical field of nuclide diffusion prediction, in particular to a nuclide diffusion model training method, a nuclide diffusion model prediction method, a nuclide diffusion model training system and a nuclide diffusion model prediction system. The nuclide diffusion prediction method provided by the invention comprises the following steps: firstly, converting collected time sequence data into a Markov transfer field image, and then reconstructing the Markov transfer field image through a preprocessing module to obtain reconstructed data which can better manifest feature information; inputting the reconstruction data into a DDPM network for processing, and generating a radionuclide diffusion image in a prediction time period; a radionuclide diffusion image output by the DDPM network is actually a Markov transfer field image corresponding to a time sequence data prediction value in a prediction time period, and the image can be obtained through Markov inverse coding of the radionuclide diffusion image. According to the method, the nuclide diffusion model is predicted based on the Markov transfer field image, the characteristics of the input data are considered more comprehensively, and the prediction result is more accurate.

Description

Technical Field

[0001] The present invention relates to the technical field of nuclide diffusion prediction, and in particular to a nuclide diffusion model training, prediction method, system and readable medium. Background Art

[0002] Radioactive pollution causes irreversible damage to humans and the natural environment. Therefore, the prediction of radionuclide diffusion is of vital importance and helps to scientifically guide nuclear emergency decision-making.

[0003] The research on radionuclide diffusion prediction involves multiple disciplines such as nuclear physics, atmospheric environmental science and computer science, and has developed a mature theoretical basis and application model. Traditional radionuclide diffusion prediction methods include atmospheric tracer experiments, wind tunnel experiments and numerical simulation predictions. Atmospheric tracer experiments refer to the release of tracers in a specific location, and the atmospheric flow characteristics and tracer diffusion laws in the regional environment are obtained through sampling and monitoring. Common gas tracers include SF6 gas. SF6 is stable in nature and its concentration can be detected by gas chromatography. The wind tunnel experiment is similar to the atmospheric tracer experiment in principle, and tracers are also used to reflect the atmospheric flow characteristics and diffusion laws of the region. Unlike the atmospheric tracer experiment, which is carried out in a real scene, the wind tunnel test is carried out in a model that is scaled to a certain scale of the real scene. The advantage of the wind tunnel experiment is that it reduces costs while achieving control of meteorological conditions and underlying surface environment, and diffusion prediction can be carried out through quantitative research.

[0004] Traditional numerical models of atmospheric diffusion mainly include Gaussian plume model, Lagrangian puff model, Euler model and computational fluid dynamics. However, these traditional nuclide diffusion prediction methods are very dependent on the source information of radionuclides, which makes it difficult for traditional nuclide simulation methods to be quickly applied to sudden nuclide diffusion scenarios with uncertain locations, complex terrain and changeable weather. Summary of the invention

[0005] In order to overcome the defects of the above-mentioned prior art nuclide diffusion prediction methods that are too dependent on the source information of radioactive nuclides and difficult to apply to sudden nuclide diffusion scenarios, the present invention proposes a nuclide diffusion model training method that is applicable to any scenario and can predict nuclide diffusion situations with high speed and accuracy.

[0006] The present invention provides a method for training a nuclide diffusion model, comprising the following steps:

[0007] SA1, build the basic model, the first data set and the second data set;

[0008] The first data set is used to store time series data of missing time points; the second data set is used to store complete time series data annotated with radionuclide diffusion images; the time series data includes monitoring objects collected at multiple consecutive monitoring time points in the monitoring area; the monitoring objects include radionuclide concentrations;

[0009] The basic model includes a preprocessing module and a DDPM network. 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. The preprocessing module is used to reconstruct the input Markov transfer field image, and the DDPM network generates a Markov transfer field image in the prediction time period based on the reconstructed Markov transfer field image.

[0010] SA2, processing the time series data in the first data set and the second data set into Markov transition field images;

[0011] SA3, training the preprocessing module on the first data set until convergence;

[0012] SA4, extracting a second training sample from the second data set, performing missing processing on the time series data of the second training sample and then converting it into a Markov transition field image as a missing sample;

[0013] SA5, the missing samples are processed by the preprocessing module to obtain a blurred image, and the DDPM network processes the blurred image to obtain a predicted radionuclide diffusion image as the predicted value;

[0014] SA6, taking the radionuclide diffusion image associated with the second training sample as the true value, and calculating the loss by combining the predicted value and the true value;

[0015] SA7. Determine whether the loss converges; if not, update the DDPM network through loss back propagation, and then return to step SA4; if yes, fix the basic model as the nuclide diffusion model.

[0016] Preferably, in SA2, the method for processing time series data into a Markov transition field image is: processing the time series data into a time series of each monitored object, using a Markov transfer matrix to process the time series of each monitored object into a Markov coded image, and then performing weighted superposition on the Markov coded images of each monitored object to obtain a Markov transition field image.

[0017] Preferably, in SA3, the method for training the preprocessing module on the first data set is: the preprocessing module processes the Markov transfer field image in the first data set and outputs the reconstructed image as a blurred image, and the DDPM generates a radionuclide diffusion image based on the blurred image; the parameters of the preprocessing module are updated according to the optimization target until the preprocessing module converges; the optimization target is:

[0018]

[0019] The Markov transition field image representing the input of the preprocessing module, represents the output data of the preprocessing module; x represents the radionuclide diffusion image output by DDPM; Feed into the DDPM network for guided generation; It means that the basic model is controlled by the parameter η. and The probability of generating x, that is, the input of the preprocessing module is The output is The probability that the DDPM network outputs x when .

[0020] Preferably, the monitoring objects also include one or more of atmospheric pressure, wind speed, temperature and humidity.

[0021] Preferably, the wind speed is the average of the X-axis wind speed, the Y-axis wind speed and the Z-axis wind speed.

[0022] Preferably, the preprocessing module uses a β-VAE module.

[0023] The present invention proposes a method for predicting nuclide diffusion using the training method of the nuclide diffusion model, comprising the following steps:

[0024] S1. Using the training method of the nuclide diffusion model, a nuclide diffusion model is obtained; collecting time series data in a set monitoring area;

[0025] S2, randomly sampling the time series data to form new time series data of the missing time points, and converting the new time series data into a Markov transition field image;

[0026] S3, inputting the Markov transfer field image into the nuclide diffusion model; the preprocessing module reconstructs the Markov transfer field image; the DDPM network processes the reconstructed data and outputs the radionuclide diffusion image;

[0027] S4. Perform Markov inverse coding on the radionuclide diffusion image to obtain the predicted value of the time series data consisting of the monitoring objects in the monitoring area in the predicted time period, and extract the radionuclide concentration prediction sequence from the predicted value as the nuclide diffusion prediction result.

[0028] Preferably, in S2, the method for converting the new time series data into a Markov transition field image is: processing the new time series data into a time series of each monitored object, using a Markov transfer matrix to process the time series of each monitored object into a Markov coded image, and then performing weighted superposition on the Markov coded images of each monitored object to obtain a Markov transition field image.

[0029] The present invention provides a nuclide diffusion prediction system, which includes a memory and a processor. The memory stores a computer program. The processor is connected to the memory and is used to execute the computer program to implement the nuclide diffusion prediction method.

[0030] The present invention provides a readable medium storing a computer program, which is used to implement the nuclide diffusion prediction method when executed.

[0031] The advantages of the present invention are:

[0032] (1) The present invention proposes a method for predicting radionuclide diffusion. First, the collected time series data is converted into a Markov transfer field image. Then, the Markov transfer field image is reconstructed through a preprocessing module to obtain reconstructed data that can better highlight the characteristic information. The reconstructed data is then input into a DDPM network for processing to generate a radionuclide diffusion image over a prediction time period. The radionuclide diffusion image output by the DDPM network is actually a Markov transfer field image corresponding to the predicted value of the time series data over the prediction time period, which can be obtained through the Markov inverse coding of the radionuclide diffusion image. In the present invention, the radionuclide diffusion model is predicted based on the Markov transfer field image, and the characteristics of the input data are considered more comprehensively, and the prediction results are more accurate.

[0033] (2) In the present invention, the preprocessing module adopts the β-VAE module, and uses the image generated by it as a priori condition to be introduced into the DDPM network, so that DDPM conditionally generates the corresponding nuclide diffusion Markov transfer field image, further improving the reliability of the prediction.

[0034] (3) In the present invention, the Markov coded image of the time series of each monitored object is first obtained, and then the Markov transfer field image of multi-feature coupling is obtained by weighted superposition; so that the input of the nuclide diffusion model can deeply express the monitoring characteristics in a simple and vivid image way, thereby facilitating feature mining in the model processing process.

[0035] (4) In the training method of a nuclide diffusion model proposed by the present invention, the supervised training of the preprocessing module and the unsupervised training of the DDPM are independent of each other during the training process, thereby improving the overall convergence speed during the model training process. The generated result of the preprocessing module is used as one of the inputs of the DDPM to guide the DDPM to generate the final image, further improving the training accuracy and efficiency of the DDPM network. BRIEF DESCRIPTION OF THE DRAWINGS

[0036] Figure 1 A flow chart of a training method for a nuclide diffusion model;

[0037] Figure 2A schematic diagram of the training process of a nuclide diffusion model;

[0038] Figure 3 It is a flow chart of a method for predicting nuclide diffusion;

[0039] Figure 4 It is a complex underlying surface geometry model composed of three different vegetation types;

[0040] Figure 5 This is a schematic diagram of the monitoring area;

[0041] Figure 6 is the change of radionuclide concentration at each node;

[0042] Figure 7 This is the predicted result of radionuclide concentration of node A1. DETAILED DESCRIPTION

[0043] The following will be combined with the drawings in the embodiments of the present invention to clearly and completely describe the technical solutions 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 are within the scope of protection of the present invention.

[0044] DDPM (conditional diffusion model) is a generative model that learns the target data distribution from samples. It consists of two Markov processes: one is a fixed forward process that simulates the transformation from real images to random Gaussian noise, and the other is a reverse process based on learning to generate images with similar features to the original image. The form of the forward process in DDPM can be summarized as follows:

[0045]

[0046] Where: c represents the control condition added in DDPM; q(x 1：T |x0, c) represents the joint conditional probability distribution from the initial state x0 to time step T under the guidance of condition c, which describes the gradual addition of noise from the original data x0 to x T The whole process; T is the maximum time step set; t is the time step sequence number;

[0047] q(x t |x t-1 , c) represents the state x at time step t-1 under the guidance of condition c t-1 The state x at time step t t The conditional probability distribution of ; this process usually involves adding random noise to x t-1 To obtain x t ; δ1, δ2, ..., δ t, ..., δ T represents the noise adding strategy in the forward process; It means the mean The standard deviation is δ t Gaussian distribution of I; I represents the identity matrix;

[0048] q(x t |x0, c) means that under the guidance of condition c, from the initial state x0 to the state x at time step t t The conditional probability distribution of ; It means the mean The standard deviation is Gaussian distribution, and there is γ t =1-δ t .

[0049] The reverse process of DDPM can also be parameterized by a learned Gaussian transfer distribution as follows:

[0050]

[0051] Where: p(x 0：T |c) indicates that under the influence of condition c, from the final noise state x T The joint probability distribution of the original state x0 describes the process of gradually recovering data from noise; p(x T |c) indicates that under condition c, the final noise state x T The prior probability distribution of ;

[0052] p λ (x t-1 |x t , c) represents the state x at time step t under the influence of condition c t Go back to the state x at time step t-1 t-1 The conditional probability distribution of this process is parameterized by a learnable model, and λ is the parameter of DDPM.

[0053] Let q(x T |x0, c) means that under the guidance of condition c, from the initial state x0 to x T The conditional probability distribution of ; when T is large enough, the distribution q(x T |x0, c) can be approximated as an isotropic Gaussian distribution, and λ can be trained based on variational inference. However, according to Formula 1 and the generation process of DDPM in Formula 1, it can be seen that adding 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 with any preprocessing module. represents a general preprocessing module, and its optimization objective function can be expressed as:

[0054]

[0055] η is the parameter of the overall model formed by connecting the preprocessing module and the DDPM framework front and back; represents the original time series data, that is, the input data of the preprocessing module; x represents the original time series data The image processed by the preprocessing module and then reconstructed by DDPM; represents the prior information, that is, the data output by the preprocessing module, which is also the condition c in formula 1; Send it to the DDPM network to guide x generation; It means that the model parameter η is controlled by and The probability of generating x, that is, the DDPM when the input is and The probability that the output is x.

[0056] To obtain the original state x0, replace x with x0 in formula (3), so that DDPM can be calculated according to and The output is x0, and based on the evidence lower bound, we get the following inequality:

[0057]

[0058] Where φ is the parameter of the DDPM inverse process, It means that DDPM has and The probability of outputting x0 when ; express expectations, Represents the original time series data The data x0 obtained by the preprocessing module and DDPM reconstruction is denoised to generate a data sequence x 1:T The probability of p(x 0:T ) represents the final noise state x T Back to the joint probability distribution of the original state x0; Represents the output data of DPPM combined with the preprocessing module and the original time series data Get the data sequence x 0:T probability; Indicates that the preprocessing module is based on generate The probability of express distribution of ; ψ represents the parameters of the encoder in the preprocessing model.

[0059] This 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 the DDPM are independent of each other; the generated result of the preprocessing module is As one of the inputs of DDPM, guide DDPM to generate the final image.

[0060] Based on the derivation of the above theory, the present invention adopts the β-VAE module (β-Variational Autoencoder) as the preprocessing module. Using the Markov transition field image, the image generated by the preprocessing module is used as a priori condition Bringing it into the DDPM model enables DDPM to conditionally generate the corresponding nuclide diffusion Markov transfer field image.

[0061] In the present invention, the overall model composed of the β-VAE module and the DDPM network is recorded as a β-VDRDPF model, the input of the β-VAE module is used as the input of the β-VDRDPF model, the input data of the β-VDRDPF model adopts the Markov transfer field image, and the DDPM network combines the Markov transfer field image input by the β-VDRDPF model and the output data of the β-VAE module to generate a radionuclide diffusion image.

[0062] The following introduces the data processing process of the β-VAE module, DDPM model and Markov transfer field respectively.

[0063] The input of the β-VAE module is the Markov transition field image of the initial state x0. The Markov transition field image is converted from the collected time series data. The time series data includes: acquisition time (s); radionuclide I-131 concentration (kBq / m 3 ); atmospheric pressure at the collection point (Pa); wind speed on the X axis (m / s), wind speed on the Y axis (m / s), wind speed on the Z axis (m / s) and three-dimensional coordinates of the collection point.

[0064] The β-VAE module includes an encoder and a decoder. The input data x0 is encoded by the encoder, and the encoded feature data output by the encoder is weighted to form a latent code z. The latent code z does not directly participate in the training of DDPM and is not output. The latent code z is only used in the training of the β-VAE module, which is specifically expressed in the following formulas (5)-(8).

[0065] During the training process of the β-VAE module, an auxiliary signal y is required as input. The β-VAE module processes the auxiliary signal y to generate output data, and the β-VAE module generates a potential code z; DDPM generates a time series x based on the output data of the β-VAE module. 1:T , T is the total number of time steps; thus, the joint distribution p(x 0：T , y, z) can be expressed as:

[0066] p(x 0：T ,y,z)=p(z)p θ (y|z)pφ(x 0:T |y,z) (5)

[0067] Where θ represents the parameters of the decoder in the β-VAE module, φ represents the parameters of the inverse process of the DPPM network; p(z) represents the probability of generating the potential code z by the β-VAE module; 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 parameter in the β-VAE module is θ; p φ (x 0：T |y, z) indicates that when the inverse process parameter of the DPPM network is φ and the potential code generated by the β-VAE module is z, the β-VDRDPF model generates time series data x according to the input data y 0：T probability.

[0068] Since the true joint posterior distribution p(x 1：T , z|y, x0) is difficult to calculate, so an approximate proxy distribution q(x 1：T , z|y, x0) to approximate, q(x 1：T , z|y,x0) can also be described as the following conditional distribution:

[0069] q(x 1：T , z|y,x0)=q ψ (z|y,x0)q(x 1：T |y,z,z0) (6)

[0070] Among them, p(x 1：T , z|y, x0) represents the true joint posterior distribution of the β-VDRDPF model, that is, the β-VDRDPF model generates the latent code z and time series data x1 when the input is y and the output of the β-VAE module is x0: T The probability of; ψ represents the parameters of the β-VAE module; q ψ (z|y, x0) represents the probability of outputting the latent code z when the β-VAE module input is y and the output is x0, when the encoder parameter of the β-VAE module is ψ; q(x 1：T|y,z,x0) means that when the β-VAE module input is y, the output is x0, and the potential code is z, the β-VDRDPF model generates the time series x 1:T probability.

[0071] During the training process, the forward process of DDPM does not perform parameter updates or training, but keeps its parameters fixed. Therefore, the logarithm of the likelihood of the Markov transition field data used for training can be expressed as:

[0072] lgp(x0,y)=lg∫p(x 0：T ,y,z)dx 1：T dz (7)

[0073] Among them, p(x0, y) represents the probability that the output of the β-VAE module is x0 when the input is y; p(x 0:T , y, z) means that when the β-VAE module takes input y and generates a latent code z, the β-VDRDPF model generates a time series x 0:T probability;

[0074] Since formula (7) is difficult to estimate analytically, the present invention optimizes the lower bound of evidence corresponding to the log-likelihood

[0075]

[0076] Among them, ψ is the parameter of the encoder in the β-VAE module, θ is the parameter of the decoder in the β-VAE module, and φ is the parameter of the inverse process of DDPM; p θ (y|z) represents the probability that the decoder of the β-VAE module generates the input data y according to the latent code z; q ψ (z|y, x0) represents the probability that the encoder of the β-VAE module generates the latent code z when the input is y and the output is x0; In the probability distribution q ψ p calculated on (z|y, x0) θ Expectation of (y|z);

[0077] for q ψ The KL divergence between (z|y, x0) and p(z); p(z) represents the probability of the β-VAE module generating the latent code z; β represents the hyperparameter that controls the KL divergence weight in the β-VAE module, which determines the degree of regularization of the β-VAE module;

[0078] p φ (x 0:T |y,z) represents the inverse process of DDPM to generate x based on input data y under the guidance of latent code z 0:T The probability of q(x 1:T|y,z,x0) means that when the β-VAE module input data is y and the output is x0 and the generated potential code is z, the x generated by DDPM in the β-VDRDPF model 1:T probability; is q(x 1：T |y,z,x0) on p φ (x 0：T |y,z) and q(x 1：T |y, z, x0); represents the expectation computed over the z distribution with input x0 as output for a given β-VAE module.

[0079] In the above formula (8), is the loss term of the β-VAE module, indicating that the latent code z generated by its encoder can reconstruct the expectation of the conditional signal y; is the loss term of DDPM, representing the reverse diffusion process p φ (x 0：T |y,z) and the forward diffusion process q(x 1：T |The expected ratio between y, z, x0); represents the expected calculation of the latent variable z output by the VAE encoder network; represents the loss function and contribution of the β-VAE module, Represents the loss function and contribution of DDPM;

[0080] In order to enable the β-VDRDPF model proposed in the present invention to converge quickly during the training process, the following design method is used to simplify some network parameters during parameter design, thereby rewriting formula (8) into the following formulas (9) and (10).

[0081] 1) In order to ensure the mapping relationship between y and x0, it is assumed that y is the Markov transfer field image x0 itself, and during the training process, the reverse diffusion process of the diffusion model is not conditioned on y, but is regarded as p in formula (5). φ (x 0：T |z). At this time, formula (8) can be rewritten as:

[0082]

[0083] Among them, q ψ (z|x0) represents the encoder of the β-VAE module generating the posterior distribution of the latent variable z according to the input image x0, p θ (x0|z) represents the probability that the decoder of the β-VAE module generates the input image x0 according to the latent code z; Indicates that in q ψ Compute p on the (z|x0) distributionθ The expectation of (x0|z) is used to characterize the posterior distribution q of z given x0. ψ The expected reconstruction loss of (z|x0);

[0084] p(x0) represents the probability that the β-VAE module generates x0, and p(z) represents the probability that the β-VAE module generates z; Denotes the posterior distribution q ψ The KL divergence between (z|x0) and the prior distribution p(z); β represents the hyperparameter that controls the weight of the KL divergence and determines the degree of regularization of the model;

[0085] p φ (x 0:T |z) represents the inverse process of DDPM, which generates x based on the input data y under the guidance of the latent code z. 0:T The probability of gradually recovering the image x from the noise given z 0：T The inverse process distribution of q(x 1:T |z,x0) represents the x generated by DDPM in the β-VDRDPF model when the input data of the β-VAE module is x0 and the generated latent code is z. 1:T The probability that the noise diffusion process x is given z and x0. 1：T The forward process distribution of is q(x 1:T |z,x0) on p φ (x 0:T |z) and q(x 1:T |z,x0), that is, the expectation of the reverse diffusion process p φ (x 0：T |z) and the forward diffusion process q(x 1：T |z, x0) between the expected ratio; represents the expectation calculated on the z distribution of the input image x0, that is, the expected calculation of the latent variable z output by the encoder in the β-VAE module given x0.

[0086] 2) When performing the reverse diffusion of the model, the invention does not directly use the latent code z of x0 generated by the β-VAE module as the input of the diffusion model, but instead uses the latent code z reconstructed by the β-VAE module Guide the training of the reverse transfer process of the diffusion model, at this time, β-VAE reconstructs is a deterministic function of z. At this time, formula (8) can be rewritten as:

[0087]

[0088] represents the output of the β-VAE module when x0 is input; Indicates a given When x0 and x0, the noise diffusion process x 1:T The forward process distribution of Indicates a given When , the time series x is gradually recovered from the noise 0：T The inverse process distribution of It represents the expectation of the ratio of the inverse process distribution of DDPM to the forward process distribution; It means that for a given input image x0, from Reconstructing images by sampling from a distribution expectations; Reconstructed image representing the β-VAE module Expected calculation.

[0089] The evidence lower bound expression of formula (10) consists of two parts: the first part is the ELBO (variational lower bound) of the β-VAE module, which measures the reconstruction loss and the KL divergence between the distribution of the latent variable z and the prior distribution; the second part is the loss of DDPM, which measures the loss from the reconstructed image The degree of match between the inverse process and the forward process of restoring the original image x0.

[0090] 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 is to optimize Second step optimization The second training phase fixes the parameters θ and ψ at the same time, and freezes the parameters of the β-VAE module in the first step; therefore, its final joint distribution can be defined as:

[0091]

[0092] In the formula, It means that the β-VDRDPF model is given Generate x under the condition 0：T The conditional probability of is generated by β-VAE Probability, Indicates that DDPM is given The conditional generation of x 0：T The conditional probability of .

[0093] A sensors, data from the A sensors are used to obtain A Markov transfer field images, and the A Markov transfer field images are weightedly superimposed to obtain a final Markov transfer field image.

[0094] Based on the above derivation, this embodiment proposes a nuclide diffusion model composed of a β-VAE module and a DDPM network; the input of the β-VAE module is the input of the nuclide diffusion model, the output of the β-VAE module is connected to the input of the DDPM network, and the output of the DDPM network is the output of the nuclide diffusion model.

[0095] The input of the radionuclide diffusion model is the Markov transition field image y converted from the time series data collected in the monitoring area. The β-VAE module encodes and decodes the Markov transition field image, and the decoded data output by the β-VAE module is After DDPM network processing, the radionuclide diffusion image x0 is generated as the nuclide diffusion prediction result, that is, the Markov transfer field image corresponding to the time series data in the prediction time period.

[0096] In this embodiment, the time series data includes monitoring objects collected at multiple consecutive monitoring time points in the monitoring area, and the monitoring objects include radionuclide concentrations (kBq / m 3 ), atmospheric pressure (Pa), wind speed (m / s), and temperature and humidity. The wind speed is the average of the X-axis wind speed (m / s), Y-axis wind speed (m / s), and Z-axis wind speed (m / s) in the monitoring area.

[0097] The time series data is essentially composed of the time series of each monitored object. The method for processing the time series data into a Markov transfer field image is: using the Markov transfer matrix to process the time series of each monitored object into a Markov coded image, and then weighted superposition of the Markov coded images of each monitored object is performed to obtain a multi-feature coupled Markov transfer field image as the input of the nuclide diffusion model.

[0098] When weighted superposition is performed on the Markov coded images of each monitored object, the weight of the radionuclide concentration shall not be less than 0.95; the sum of the weights of temperature and humidity shall not be greater than 0.01.

[0099] It is worth noting that the number of monitoring time points included in the input of the nuclide diffusion model is equal to the amount of time point data included in the output of the nuclide diffusion model; assuming that the time series data consisting of the monitoring objects collected at 1-M monitoring time points in the monitoring area are converted into Markov transfer field images and input into the nuclide diffusion model, the radioactive nuclide diffusion image output by the nuclide diffusion model is Markov inverse encoded to obtain the predicted values ​​of the time series data consisting of the monitoring objects in the monitoring area at M+1 to 2M monitoring time points, and then the predicted time series of the radioactive nuclide concentration can be extracted from the predicted values.

[0100] Reference Figure 1 , Figure 2 , a training method for a nuclide diffusion model proposed in this embodiment includes the following steps:

[0101] SA1. Construct a basic model, a first data set, and a second data set. The basic model includes a preprocessing module and a DDPM network. The preprocessing module specifically uses a β-VAE module.

[0102] The preprocessing module includes an encoder and a decoder. The encoder encodes the input Markov transfer field image and generates a potential feature z. The decoder decodes the potential feature z and outputs a reconstructed image. The reconstructed image is processed by the DDPM network into a radionuclide diffusion image and output.

[0103] The first data set is used to store the time series data of some missing time points; the second data set is used to store the complete time series data annotated with radionuclide diffusion images;

[0104] SA2, processing the time series data in the first data set and the second data set into Markov transition field images;

[0105] SA3. Train the preprocessing module on the first data set until convergence; during the training process, the preprocessing module processes the Markov transfer field image in the first data set and outputs the reconstructed image as a blurred image, and the DDPM generates a radionuclide diffusion image based on the blurred image; during the training process, update the parameters of the preprocessing module according to the following optimization objectives, and the parameters of the DDPM are fixed; the optimization objectives are:

[0106]

[0107] The Markov transition field image representing the input of the preprocessing module, represents the output data of the preprocessing module; x represents the radionuclide diffusion image output by DDPM; Feed into the DDPM network for guided generation; It means that the basic model is controlled by the parameter η. and The probability of generating x, that is, the input of the preprocessing module is The output is The probability that the DDPM network outputs x when .

[0108] Specifically, step SA3 includes the following sub-steps:

[0109] SA31, extracting a first training sample from the first data set, substituting the Markov transition field image of the first training sample into the basic model, and optimizing the parameters of the preprocessing module in combination with the above optimization objectives;

[0110] SA32, determine whether the number of iterations of the preprocessing module reaches the set value; if yes, complete the pre-training of the preprocessing module and execute step S4; if no, return to step SA31.

[0111] SA4, extracting a second training sample from the second data set, performing missing processing on the time series data of the second training sample and then converting it into a Markov transition field image as a missing sample;

[0112] Specifically, in this step, firstly, the time series data in the second training sample is randomly sampled to form a data sequence of the missing time points, and then the data sequence of the missing time points is converted into a Markov transition field image as the missing sample;

[0113] SA5, the missing samples are processed by the preprocessing module to obtain a blurred image, and the DDPM network processes the blurred image to obtain a predicted radionuclide diffusion image as the predicted value;

[0114] SA6, taking the radionuclide diffusion image associated with the second training sample as the true value, and calculating the loss by combining the predicted value and the true value;

[0115] SA7. Determine whether the loss converges; if not, update the DDPM network through loss back propagation, and then return to step SA4; if yes, fix the parameters of the preprocessing module and the DDPM network, and substitute them into the basic model to obtain a nuclide diffusion model with the input as the Markov transfer field image and the output as the radionuclide diffusion image.

[0116] Reference Figure 3 , when using the trained nuclide diffusion image to predict nuclide diffusion, the steps are as follows:

[0117] S1. Using the above steps SA1-SA7, a nuclide diffusion model is obtained; and time series data is collected in a set monitoring area;

[0118] S2, randomly sampling the time series data to form new time series data of the missing time points, and converting the new time series data into a Markov transition field image;

[0119] S3, inputting the Markov transfer field image into the nuclide diffusion model; the preprocessing module encodes and decodes the Markov transfer field image and outputs the decoded data; the DDPM network processes the decoded data and outputs the radionuclide diffusion image;

[0120] S4. Perform Markov inverse coding on the radionuclide diffusion image to obtain the predicted value of the time series data consisting of the monitoring objects in the monitoring area in the predicted time period, and extract the radionuclide concentration prediction sequence from the predicted value as the nuclide diffusion prediction result.

[0121] The above-mentioned nuclide diffusion prediction method is described below in conjunction with specific embodiments.

[0122] In this example, the original time series data of radionuclide diffusion is generated by simulating with the CFD software OpenFOAM. The scenario simulated by the data is a leakage accident of spent fuel during transportation. A complex underlying surface geometry model consisting of three different vegetation types is established. Figure 4 As shown in the figure, the model calculation area is a three-dimensional space of 1000m×1000m×100m, the left plane of the cube is the air inlet, the right plane is the air outlet, the airflow direction is set to be along the positive direction of the x-axis, and the wind speed is set to 4m / s. Figure 4 The red solid point in the middle represents the release source of the radionuclide I-131, with coordinates of (400 meters, 500 meters, 1 meter). The present invention assumes that the release source releases continuously for 1 hour, and the release rate is 3.3 (kBq / cubic meter) / second. The half-life of the radionuclide I-131 is 8.02 days. kBq is kilobecquerel, which is a unit for indicating radioactivity.

[0123] The underlying surface model contains three different vegetation groups. Different vegetation has different abilities to adsorb and change the flow field, which is called the vegetation effect. The essence of the vegetation effect is that vegetation, as a porous medium, will produce pressure loss to the diffused fluid passing through its surface. Different vegetation types and different densities produce different pressure losses. Therefore, the present invention represents different vegetation underlying surfaces by setting different pressure loss coefficients, wherein the pressure loss coefficient of vegetation group A is set to 0.5 / m, and the spatial dimensions are 120m×60m×15m; the pressure loss coefficient of vegetation group B is set to 2 / m, and the spatial dimensions are 60m×140m×15m; the pressure loss coefficient of vegetation group C is set to 8 / m, and the spatial dimensions are 60m×140m×15m.

[0124] According to the above settings, the diffusion data of I-131 was obtained by CFD simulation. The time step of diffusion data collection was 10 seconds, the total simulated diffusion time was 120 minutes, and 720 sets of raw data were finally obtained. The raw data included 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 some data samples generated by the CFD software OpenFOAM.

[0125] Table 1: Part of the original data of the 60th minute of diffusion

[0126]

[0127] Table 2 shows the three monitoring areas selected in three different vegetation types. Figure 5 The red solid dots represent the radionuclide leakage sources and the blue solid dots represent the monitoring points.

[0128] Table 2 Coordinates of monitoring points

[0129]

[0130]

[0131] In this invention, the monitoring area is abstracted as a node, and the radionuclide concentration, wind speed, atmospheric pressure and underlying surface characteristics in the monitoring area are abstracted as node features. Image encoding is achieved using the Markov transfer field according to different node features, and then weighted superposition is performed to obtain a multi-feature coupled Markov transfer field image. Figure 6 Shown Figure 5 The changes in the concentration of radionuclides at each node.

[0132] Specifically, in this embodiment, the time series of the radionuclide I-131 concentration is firstly Markov coded to obtain a Markov coded image of the concentration; the time series of the atmospheric pressure is Markov coded to obtain a Markov coded image of the pressure; the average of the X-axis wind speed (m / s), the Y-axis wind speed (m / s), and the Z-axis wind speed at each time point is calculated as the wind speed value, and the time series of the wind speed value is Markov coded to obtain a Markov coded image of the wind speed;

[0133] Then, the Markov coding image of concentration, the Markov coding image of pressure and the Markov coding image of wind speed are weightedly superimposed to form a Markov transfer field image;

[0134] Finally, the Markov transfer field image is input into the nuclide diffusion model to obtain the predicted radionuclide diffusion image, and the radionuclide diffusion image is subjected to Markov inverse encoding to obtain the radionuclide I-131 concentration at M consecutive time points.

[0135] The present invention uses LSTM, RNN and GAT-LSTM as comparison models. The nuclide diffusion model in the method of the present invention and the LSTM, RNN and GAT-LSTM used as comparison models all use the Markov transition field image of the node as input, thereby predicting the radionuclide diffusion image of the node to obtain the change of the radionuclide concentration of the node.

[0136] In this experiment, the data from the 600th to the 3310th seconds are selected as the model training set, and the data from the 3320th to the 3620th seconds are selected as the model test set. Figure 5 Quantitative analysis of the prediction accuracy of each point in the three detection areas. Figure 7 The results of the method of the present invention, LSTM, RNN and GAT-LSTM for predicting the radioactivity concentration of node A1 are shown. Figure 7It can be seen that the predicted result of the β-VDRDPF proposed in the present invention is closest to the true value. At the same time, according to the quantitative results analysis in Table 3, it can be seen that when predicting the changes in the concentration of radionuclides at different nodes in different regions, the method proposed in the present invention maintains a high accuracy, which is better than the other three methods, proving that the algorithm proposed in the present invention has a high accuracy in predicting the changes in the concentration of radionuclides.

[0137] Table 1 Quantitative analysis of the prediction accuracy of radionuclide concentrations at different nodes by different methods

[0138]

[0139] Of course, it is obvious to those skilled in the art that the present invention is not limited to the details of the exemplary embodiments described above, but also includes the same or similar structures that can be implemented in other specific forms without departing from the spirit or essential features of the present invention. Therefore, no matter from which point of view, the embodiments should be regarded as exemplary and non-limiting, and the scope of the present invention is defined by the appended claims rather than the above description, and it is intended that all changes falling within the meaning and scope of the equivalent elements of the claims are included in the present invention. Any reference numerals in the claims should not be regarded as limiting the claims involved.

[0140] In addition, it should be understood that although the present specification is described according to implementation modes, not every implementation mode contains only one independent technical solution. This description of the specification is only for the sake of clarity. Those skilled in the art should regard the specification as a whole. The technical solutions in each embodiment may also be appropriately combined to form other implementation modes that can be understood by those skilled in the art.

[0141] The techniques, shapes, and structural parts not described in detail in the present invention are all well-known techniques.

Claims

1. A method for training a nuclide diffusion model, characterized in that: The following steps are involved: SA1, build the basic model, the first data set and the second data set; The first data set is used to store time series data with some missing time points; The second data set is used to store complete time series data annotated with radionuclide diffusion images; the time series data includes monitoring objects collected at multiple consecutive monitoring time points in the monitoring area; the monitoring objects include radionuclide concentrations; The basic model includes a preprocessing module and a DDPM network. 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. The preprocessing module is used to reconstruct the input Markov transfer field image, and the DDPM network generates a Markov transfer field image in the prediction time period based on the reconstructed Markov transfer field image. SA2, processing the time series data in the first data set and the second data set into Markov transition field images; SA3, training the preprocessing module on the first data set until convergence; SA4, extracting a second training sample from the second data set, performing missing processing on the time series data of the second training sample and then converting it into a Markov transition field image as a missing sample; SA5, the missing samples are processed by the preprocessing module to obtain a blurred image, and the DDPM network processes the blurred image to obtain a predicted radionuclide diffusion image as the predicted value; SA6, taking the radionuclide diffusion image associated with the second training sample as the true value, and calculating the loss by combining the predicted value and the true value; SA7. Determine whether the loss converges; if not, update the DDPM network through loss back propagation and then return to step SA4; if yes, fix the basic model as the nuclide diffusion model.

2. The method for training a nuclide diffusion model according to claim 1, wherein: In SA2, the method for processing time series data into Markov transition field images is as follows: processing the time series data into the time series of each monitored object, using the Markov transfer matrix to process the time series of each monitored object into a Markov coded image, and then weighted superposition of the Markov coded images of each monitored object to obtain a Markov transition field image.

3. The method for training a nuclide diffusion model according to claim 1, wherein: In SA3, the method for training the preprocessing module on the first data set is: the preprocessing module processes the Markov transfer field image in the first data set and outputs the reconstructed image as a blurred image, and the DDPM generates a radionuclide diffusion image based on the blurred image; the parameters of the preprocessing module are updated according to the optimization target until the preprocessing module converges; the optimization target is: The Markov transition field image representing the input of the preprocessing module, represents the output data of the preprocessing module; x represents the radionuclide diffusion image output by DDPM; Feed into the DDPM network for guided generation; It means that the basic model is controlled by the parameter η. and The probability of generating x, that is, the input of the preprocessing module is The output is The probability that the DDPM network outputs x when .

4. The method for training a nuclide diffusion model according to claim 1, wherein: The monitoring objects also include one or more of atmospheric pressure, wind speed, temperature and humidity.

5. The method for training a nuclide diffusion model according to claim 4, wherein: The wind speed value is the average of the X-axis wind speed, Y-axis wind speed, and Z-axis wind speed.

6. The method for training a nuclide diffusion model according to claim 4, wherein: The preprocessing module uses the β-VAE module.

7. A method for predicting nuclide diffusion using the training method of the nuclide diffusion model according to any one of claims 1 to 5, characterized in that: The following steps are involved: S1. Using the training method of the nuclide diffusion model as described in any one of claims 1 to 5 to obtain a nuclide diffusion model; collecting time series data in a set monitoring area; S2, randomly sampling the time series data to form new time series data of the missing time points, and converting the new time series data into a Markov transition field image; S3, inputting the Markov transfer field image into the nuclide diffusion model; The preprocessing module reconstructs the Markov transfer field image; the DDPM network processes the reconstructed data and outputs the radionuclide diffusion image; S4. Perform Markov inverse coding on the radionuclide diffusion image to obtain the predicted value of the time series data consisting of the monitoring objects in the monitoring area in the predicted time period, and extract the radionuclide concentration prediction sequence from the predicted value as the nuclide diffusion prediction result.

8. The method for predicting nuclide diffusion according to claim 7, characterized in that: In S2, the method for converting the new time series data into a Markov transition field image is as follows: processing the new time series data into a time series of each monitored object, using a Markov transfer matrix to process the time series of each monitored object into a Markov coded image, and then performing weighted superposition on the Markov coded images of each monitored object to obtain a Markov transition field image.

9. A nuclide diffusion prediction system, characterized in that: The method comprises a memory and a processor, wherein a computer program is stored in the memory, and the processor is connected to the memory, and the processor is used to execute the computer program to implement the nuclide diffusion prediction method as claimed in claim 7.

10. A readable medium, characterized in that: A computer program is stored, and when the computer program is executed, it is used to implement the nuclide diffusion prediction method according to claim 7.

Images