A key nuclide prediction device based on a multi-channel hollow convolution attention model

Abstract

The application discloses a kind of key nuclide prediction devices based on multi-channel hollow convolution attention model, the device is successively composed of sodium iodide spectrometer data acquisition module, database and host computer;The host computer includes data division module, multi-channel hollow convolution attention prediction model modeling module, multi-channel hollow convolution attention prediction module and prediction result output module.The present application realizes the intelligence of strong, high-precision, high-accuracy key nuclide ionization chamber dose rate online prediction, solves the problem of low precision and low efficiency of artificial prediction of key nuclide ionization chamber dose rate, and the feature extraction is not perfect.

Description

TECHNICAL FIELD

[0001] The application belongs to the field of key nuclide prediction, and particularly relates to a key nuclide prediction device based on a multi-channel hollow convolution attention model. BACKGROUND

[0002] With the rapid development of China's national economy, China's large demand for energy makes the vigorous development of nuclear energy an inevitable requirement. In the field of nuclear power, nuclear safety is the top priority of nuclear power operation. For a long time, China's radiation environment monitoring automatic station has adopted manual detection and analysis method. This method is affected by expert human factors, and the manual method is not accurate and needs a long prediction time. Therefore, high-precision prediction of key nuclides is a problem and challenge that needs to be solved at present.

[0003] In recent years, intelligent monitoring in the industrial field has begun to attract attention. Nuclear detectors are important devices in radiation safety protection monitoring and nuclear power device safety monitoring, and they work in high temperature and humidity and high radiation intensity environment for a long time, which can easily lead to aging, performance degradation or partial functional failure, and ultimately reduce the accuracy of measurement data in the monitoring place. Instrument and meter technology is widely used in industrial processes. The gradual progress of science and technology has improved the automation level of industrial process production management. The actual industrial process instrument and meter measurement data are stored in the database through field devices. Large-scale data contains a lot of process information, which can be used for industrial process monitoring. Due to the accumulation of massive historical data in the field of nuclear power, machine learning and deep learning methods have become possible in this field. Therefore, it is of great significance to realize fast and accurate prediction of key nuclide ionization chamber dose rate according to sodium iodide spectrometer data, so as to carry out predictive maintenance of nuclear power process. SUMMARY

[0004] The purpose of the present application is to solve the problems of the prior art, and provide a key nuclide prediction device based on a multi-channel hollow convolution attention model.

[0005] The purpose of the present application is achieved by the following technical solution: a key nuclide prediction device based on a multi-channel hollow convolution attention model, which is composed of a sodium iodide spectrometer data acquisition module, a database and an upper computer in sequence; the upper computer includes a data division module, a multi-channel hollow convolution attention prediction model modeling module, a multi-channel hollow convolution attention prediction module and a prediction result output module;

[0006] The sodium iodide spectrometer data acquisition module is used to acquire sodium iodide spectrometer data E and upload it to the database;

[0007] The database is used to save the sodium iodide spectrometer data E collected by the sodium iodide spectrometer data acquisition module and upload it to the upper computer.

[0008] The host computer is used to train the multi-channel dilated convolutional attention prediction model with the sodium iodide spectrometer data E to obtain an optimized multi-channel dilated convolutional attention prediction model. Then, the optimized multi-channel dilated convolutional attention prediction model is used to predict the sodium iodide spectrometer data to be tested to obtain the predicted value.

[0009] Furthermore, the sodium iodide spectrometer data E is E = {e1, e2, ..., e}. g ,…,e G}, where eg is the g-th spectrometer vector in the sodium iodide spectrometer data E, and each spectrometer vector e g The dimension of each is d, and G is the length of the sodium iodide spectrometer data E, where g = 1, 2, ..., g, ..., G.

[0010] Furthermore, the host computer trains the multi-channel dilated convolutional attention prediction model using the sodium iodide spectrometer data E, obtaining an optimized multi-channel dilated convolutional attention prediction model. Subsequently, based on the optimized multi-channel dilated convolutional attention prediction model, the predicted value is obtained by predicting the sodium iodide spectrometer data to be tested, specifically as follows:

[0011] (a.1) First, the uploaded sodium iodide spectrometer data E is divided into training set E1, validation set E2 and test set E3 by the data partitioning module in the host computer. The training set E1 and validation set E2 are uploaded to the multi-channel dilated convolutional attention prediction model modeling module, and the test set E3 is uploaded to the multi-channel dilated convolutional attention prediction module.

[0012] (a.2) Subsequently, the multi-channel dilated convolutional attention prediction model modeling module trains the multi-channel dilated convolutional attention prediction model using the training set E1, and obtains the optimized multi-channel dilated convolutional attention prediction model and uploads it to the multi-channel dilated convolutional attention prediction module.

[0013] (a.3) The multi-channel hollow convolutional attention prediction module uses the optimized multi-channel hollow convolutional attention prediction model to predict the sodium iodide spectrometer data to be tested, and obtains the predicted value of the dose rate of the key nuclide ionization chamber corresponding to the data.

[0014] (a.4) The predicted values ​​of the dose rates of the key nuclide ionization chambers corresponding to the data are output through the prediction result output module.

[0015] Furthermore, step (a.2) specifically includes the following sub-steps:

[0016] (a.2.1) The multi-channel dilated convolutional attention prediction model includes n dilated convolutional feature extraction channels with different dilation rates, a bidirectional recurrent feature extraction module, an attention mechanism module, and a reverse random deactivation module;

[0017] (a.2.2) First, feature channels are extracted using dilated convolutions with different dilation rates to extract n output vectors from the training set E1: X1, X2, ..., X s ,…,X n , where X s This indicates that the s-th dilated convolution extracts the output vector from the feature channels, where s = 1, 2, ..., s, ..., n;

[0018] (a.2.3) Subsequently, any output vector X s The output B is obtained through a bidirectional recurrent feature extraction module with l hidden neurons. s ;

[0019] (a.2.4) Then output B. s Output A is obtained through the attention mechanism module. s ;

[0020] (a.2.5) The reverse random deactivation module will deactivate each output vector X s Repeat steps (a.2.3)-(a.2.4) to obtain n outputs: A1, A2, ..., A s ,…,A n ;

[0021] Then output the n outputs A1, A2, ..., A s ,…,A n By concatenating the vectors, we obtain A. multiple A multiple =concat(A1,A2,…,A) s ,…,A n );

[0022] Some units are temporarily dropped from the network with a certain probability, and the output D after the reverse random deactivation is... multiple for:

[0023] D multiple =DropOut(A multiple ,dr);

[0024] Where dr represents the probability of discarding a neural network unit;

[0025] (a.2.6) The training set E1 is processed through a bidirectional cyclic feature extraction module, an attention mechanism module, and a reverse random deactivation module to obtain the output D. original ;

[0026] (a.2.7) will output D multiple and output D original After integration, the output F is obtained: F = D multiple +D original ;

[0027] (a.2.8) Finally, the output F and the learnable transformation weights W will be calculated. O Multiplying them yields the prediction vector for the training set E1.

[0028] (a.2.9) Using the training set E1 and the predicted vectors A loss function is constructed, and the multi-channel dilated convolutional attention prediction model is trained using the loss function. The optimized multi-channel dilated convolutional attention prediction model is then uploaded to the multi-channel dilated convolutional attention prediction module.

[0029] Furthermore, the output vector X s The extraction process is as follows:

[0030] The s-th dilated convolution extracts feature channels, first using a dilation rate of r. s The dilated convolution extracts features from the training set E1, and uses hyperbolic units as activation functions. This process can be expressed as:

[0031]

[0032] Among them, X s W represents the output vector. c ξ represents the weight of the dilated convolution kernel; ξ(·) represents the hyperbolic unit; α is the hyperparameter in the hyperbolic unit;

[0033] The training set E1 is in, Let X be the t-th spectrometer vector in the training set E1, where k is the length of the training set E1, and t = 1, 2, ..., t, ..., k; and let X be the output vector. s for in, Represents the output vector X s The output of the t-th dilated convolutional module.

[0034] Further, step (a.2.3) specifically includes:

[0035] The bidirectional cyclic feature extraction module includes a forward unidirectional cyclic feature extraction module and a backward unidirectional cyclic feature extraction module, wherein the forward unidirectional cyclic feature extraction module and the backward unidirectional cyclic feature extraction module are respectively composed of a forget gate, an input gate and an output gate;

[0036] The output f of the forget gatet It can be represented as:

[0037] f t =σ(W f [h t-1 ,x t ]+b f );

[0038] Among them, f t W represents the output of the t-th forget gate. f h represents the weight of the forget gate. t-1 x represents the output of the (t-1)th unidirectional loop feature extraction module unit. t b represents the input of the t-th unidirectional loop feature extraction module unit. f The term represents the bias term of the forget gate, and σ(·) represents the sigmoid function; the forget gate passes through h t-1 and x t The information that needs to be forgotten can be identified;

[0039] The output i of the input gate t It can be represented as:

[0040] i t =σ(W i [h t-1 ,x t ]+b i );

[0041] Among them, i t W represents the output of the t-th input gate. i The weight of the input gate, b i This represents the bias term of the input gate;

[0042] The input gate simultaneously updates the temporary cell state.

[0043]

[0044] in, W represents a temporary cell state. c b represents the cell state update weight. c C represents the bias term for cell state updates. t-1 This represents the state of the (t-1)th cell; the input gate passes through h. t-1 and x t We can determine which information needs to be updated through the input gate, and then obtain the new cell state based on the output of the forget gate and the output of the input gate.

[0045] The output of the output gate can be represented as:

[0046] o t =σ(Wo [h t-1 ,x t ]+b o );

[0047] wherein o t represents the output of the output gate, W o represents the weight of the output gate, b o represents the bias term of the output gate; the output gate can obtain the judgment condition of the output through h t-1 and x t , the output h t of the one-way recurrent feature extraction module unit is:

[0048] h t = o t *tanh(C t );

[0049] The output B s of the bidirectional recurrent feature extraction module is represented as:

[0050]

[0051] wherein represents the output of the forward one-way recurrent feature extraction module unit to the output vector X s , represents the output of the backward one-way recurrent feature extraction module unit to the output vector X s .

[0052] Further, the step (a.2.4) is specifically:

[0053] The attention mechanism module assigns different weights to the output A s , obtaining three matrices Q s , K s and V s ; then according to the matrices Q s , K s and V s , the output A s is obtained, and the calculation formula is as follows:

[0054] Q s =B s W Q ;

[0055] K s =B s W K ;

[0056] V s =B s W V ;

[0057]

[0058] wherein, W Q , W K and W V represent different weight matrices, d B represents the dimension of B s , and softmax(·) converts the input value into a probability distribution with a range of [0, 1] and a sum of 1, and the calculation formula is as follows:

[0059]

[0060] wherein z i represents the i-th input value, and C represents the number of input nodes.

[0061] The beneficial effects of the present application are:

[0062] 1) The multi-channel hollow convolution method can efficiently extract the features of the sodium iodide spectrometer data;

[0063] 2) The multi-channel hollow convolution attention prediction model can be updated in real time using newly input sodium iodide spectrometer data;

[0064] 3) The key nuclide prediction device based on the multi-channel hollow convolution attention model can automatically learn from training data, has strong intelligence, and is less affected by human factors. BRIEF DESCRIPTION OF DRAWINGS

[0065] Figure 1 is a structural diagram of a key nuclide prediction device based on a multi-channel hollow convolution attention model;

[0066] Figure 2 is a structural diagram of a host computer;

[0067] In the figure, 1 is a sodium iodide spectrometer data acquisition module; 2 is a database; 3 is a host computer; 4 is a data division module; 5 is a multi-channel hollow convolution attention prediction model modeling module; 6 is a multi-channel hollow convolution attention prediction module; and 7 is a prediction result output module. DETAILED DESCRIPTION

[0068] In order to make the purpose, technical scheme and advantages of the present application more clear and clear, the present application is further described in detail in combination with the drawings and examples. It should be understood that the specific examples described herein are only used to explain the present application, but not all examples. Based on the examples in the present application, all other examples obtained by those skilled in the art without creative labor are within the scope of protection of the present application.

[0069] Example 1

[0070] As shown in Figure 1 The application provides a key nuclide prediction device based on a multi-channel hollow convolution attention model, which is composed of a sodium iodide spectrometer data acquisition module 1, a database 2 and an upper computer 3 in sequence. The upper computer 3 includes a data division module 4, a multi-channel hollow convolution attention prediction model modeling module 5, a multi-channel hollow convolution attention prediction module 6 and a prediction result output module 7.

[0071] The sodium iodide spectrometer data acquisition module 1 is used to acquire sodium iodide spectrometer data E and upload it to the database 2.

[0072] The sodium iodide spectrometer data E is E={e1,e2,…,e g ,…,e G}, wherein eg is the gth spectrometer vector in the sodium iodide spectrometer data E, the dimension of each spectrometer vector e g is d, G is the length of the sodium iodide spectrometer data E, and g=1,2,…,g,…,G.

[0073] The database 2 is used to save the sodium iodide spectrometer data E collected by the sodium iodide spectrometer data acquisition module 1 and upload it to the upper computer 3.

[0074] The upper computer 3 is used to train the multi-channel hollow convolution attention prediction model with the sodium iodide spectrometer data E. The multi-channel hollow convolution method can efficiently extract the features of the sodium iodide spectrometer data, and an optimized multi-channel hollow convolution attention prediction model is trained. Then, the sodium iodide spectrometer data to be measured is predicted according to the optimized multi-channel hollow convolution attention prediction model to obtain the prediction value, specifically:

[0075] (a.1) First, the data division module 4 in the upper computer 3 divides the uploaded sodium iodide spectrometer data E into a training set E1, a validation set E2 and a test set E3 according to a ratio of 7:1:2, uploads the training set E1 and the validation set E2 to the multi-channel hollow convolution attention prediction model modeling module 5, and uploads the test set E3 to the multi-channel hollow convolution attention prediction module 6.

[0076] (a.2) Then, the multi-channel hollow convolution attention prediction model modeling module 5 trains the multi-channel hollow convolution attention prediction model with the training set E1, trains an optimized multi-channel hollow convolution attention prediction model and uploads it to the multi-channel hollow convolution attention prediction module 6.

[0077] The step (a.2) specifically includes the following sub-steps:

[0078] (a.2.1) The multi-channel dilated convolutional attention prediction model includes n dilated convolutional feature extraction channels with different dilation rates, a bidirectional cyclic feature extraction module, an attention mechanism module, and a reverse random deactivation module.

[0079] (a.2.2) First, feature channels are extracted using dilated convolutions with different dilation rates to extract n output vectors from the training set E1: X1, X2, ..., X s ,…,X n , where X s This indicates that the output vector is obtained by extracting the feature channels through the s-th dilated convolution, where s = 1, 2, ..., s, ..., n.

[0080] The output vector X s The extraction process is as follows:

[0081] The s-th dilated convolution extracts feature channels, first using a dilation rate of r. s The dilated convolution extracts features from the training set E1, and uses hyperbolic units as activation functions. This process can be expressed as:

[0082]

[0083] Among them, X s W represents the output vector. c ξ represents the weight of the dilated convolution kernel; ξ(·) represents the hyperbolic unit; α is the hyperparameter in the hyperbolic unit;

[0084] The training set E1 is in, Let X be the t-th spectrometer vector in the training set E1, where k is the length of the training set E1, and t = 1, 2, ..., t, ..., k; and let X be the output vector. s for in, Represents the output vector X s The output of the t-th dilated convolutional module.

[0085] (a.2.3) Subsequently, any output vector X s The output B is obtained through a bidirectional recurrent feature extraction module with l hidden neurons. s .

[0086] The specific steps (a.2.3) are as follows:

[0087] The bidirectional recurrent feature extraction module comprises a forward unidirectional recurrent feature extraction module and a backward unidirectional recurrent feature extraction module, and the forward unidirectional recurrent feature extraction module and the backward unidirectional recurrent feature extraction module are respectively provided with a forget gate, an input gate and an output gate, can automatically learn according to training data, have strong intelligence, and are less affected by human factors;

[0088] The output f of the forget gate t can be expressed as:

[0089] The output f of the forget gate t = σ (W f [h t-1 , x t ]+b f ) ;

[0090] Wherein, f t represents the output of the tth forget gate, W f represents the weight of the forget gate, h t-1 represents the output of the (t-1) th unidirectional recurrent feature extraction module unit, x t represents the input of the tth unidirectional recurrent feature extraction module unit, b f represents the bias term of the forget gate, and sigma (·) represents the Sigmoid function; The forget gate can determine the information to be forgotten through h t-1 and x t ;

[0091] The output i of the input gate t can be expressed as:

[0092] The output i of the input gate t = σ (W i [h t-1 , x t ]+b i ) ;

[0093] Wherein, i t represents the output of the tth input gate, W i represents the weight of the input gate, and b i represents the bias term of the input gate;

[0094] The input gate simultaneously updates the temporary cell state

[0095]

[0096] Wherein, represents the temporary cell state, W c represents the cell state update weight, b c represents the bias term of the cell state update, and C t-1 represents the (t-1) th cell state; The input gate can update the temporary cell state through h t-1and x t It can be determined which information needs to be updated through the input gate, and then a new cell state is obtained according to the output of the forget gate and the output of the input gate;

[0097] The output of the output gate can be expressed as:

[0098] o t = σ (W o [h t-1 , x t ]+b o );

[0099] wherein o t represents the output of the output gate, W o represents the weight of the output gate, and b o represents the bias term of the output gate; the output gate can obtain the judgment condition of the output through h t-1 and x t The output h t of the unidirectional recurrent feature extraction module unit is:

[0100] h t = o t *tanh (C t );

[0101] The output B s of the bidirectional recurrent feature extraction module is expressed as:

[0102]

[0103] wherein represents the output of the forward unidirectional recurrent feature extraction module unit to the output vector X s , represents the output of the backward unidirectional recurrent feature extraction module unit to the output vector X s .

[0104] (a.2.4) The output B s is then obtained through the attention mechanism module, and the output A s is obtained.

[0105] The step (a.2.4) is specifically:

[0106] The attention mechanism module assigns different weights to the output A s to obtain three matrices Q s , K s and V s ; then, according to the matrices Q s , K s and V s , the output A s, the calculation formula is as follows:

[0107] Q s = B s W Q ;

[0108] K s = B s W K ;

[0109] V s = B s W V ;

[0110]

[0111] wherein W Q , W K and W V represent different weight matrices, d B represents the dimension of B s , softmax(·) converts the input value into a probability distribution with a range of [0, 1] and a sum of 1, and the calculation formula is as follows:

[0112]

[0113] wherein z i represents the i-th input value, and C represents the number of input nodes.

[0114] The reverse random inactivation module in (a.2.5) temporarily discards each output vector X s from the network according to a certain probability, and the output D s after reverse random inactivation is obtained by repeating steps (a.2.3)-(a.2.4) n times: n .

[0115] Then, the n outputs A1, A2, …, An are spliced into a vector A s : n . multiple multiple A s = concat(A1, A2, …, An). n

[0116] According to a certain probability, some units are temporarily discarded from the network, and the output D multiple after reverse random inactivation is:

[0117] D multiple = DropOut(A multiple , dr);

[0118] wherein dr represents the probability of discarding the neural network unit.​​

[0119] (a.2.6) and the training set E1 passes through the bidirectional recurrent feature extraction module, the attention mechanism module and the reverse random inactivation module, and the output D is obtained original .

[0120] (a.2.7) the output D multiple and the output D original are integrated to obtain the output F: F = D multiple + D original .

[0121] (a.2.8) finally, the output F and the learnable transformation weight W O are multiplied to obtain the prediction vector of the training set E1

[0122] (a.2.9) a loss function is constructed by the training set E1 and the prediction vector , and the multi-channel hollow convolution attention prediction model is trained through the loss function, and the optimized multi-channel hollow convolution attention prediction model is obtained and uploaded to the multi-channel hollow convolution attention prediction module 6, and the multi-channel hollow convolution attention prediction model can be updated in real time by using the new input sodium iodide spectrometer data.

[0123] The above only describes the preferred embodiments of the present application and is not intended to limit the present application. Any modification, equivalent replacement, improvement, etc. made within the spirit and principles of the present application shall be included in the protection scope of the present application.

Claims

1. A key nuclide prediction device based on a multi-channel dilated convolutional attention model, characterized in that, The device consists of a sodium iodide spectrometer data acquisition module, a database, and a host computer. The host computer includes a data partitioning module, a multi-channel dilated convolutional attention prediction modeling module, a multi-channel dilated convolutional attention prediction module, and a prediction result output module. The sodium iodide spectrometer data acquisition module is used to acquire sodium iodide spectrometer data. And upload it to the database; The database is used to store the sodium iodide spectrometer data collected by the sodium iodide spectrometer data acquisition module. And upload it to the host computer; The host computer is used to transfer sodium iodide spectrometer data. A multi-channel dilated convolutional attention prediction model was trained to obtain an optimized model. Subsequently, the optimized model was used to predict the sodium iodide spectrometer data to be tested, yielding the predicted values. Specifically: (a.1) First, the uploaded sodium iodide spectrometer data is divided by the data partitioning module in the host computer. Divided into training set Validation set and test set and the training set and verification set Upload the data to the multi-channel dilated convolutional attention prediction model modeling module and use the test set. Upload to the multi-channel dilated convolutional attention prediction module; (a.2) Subsequently, the multi-channel dilated convolutional attention prediction model modeling module is trained on the training set. The multi-channel dilated convolutional attention prediction model is trained to obtain an optimized multi-channel dilated convolutional attention prediction model, which is then uploaded to the multi-channel dilated convolutional attention prediction module. Step (a.2) specifically includes the following sub-steps: (a.2.1) The multi-channel dilated convolutional attention prediction model includes The module includes a dilated convolution with different dilation rates to extract feature channels, a bidirectional recurrent feature extraction module, an attention mechanism module, and a reverse random deactivation module. (a.2.2) First use Different dilatational convolutions with varying dilatation rates extract feature channels on the training set. Extracted Output vectors: ,in, Indicates the first Each dilated convolutional layer extracts feature channels to obtain the output vector. ; (a.2.3) Subsequently, any output vector The output is obtained through a bidirectional recurrent feature extraction module with l hidden neurons. ; (a.2.4) Then output The output is obtained through the attention mechanism module. ; (a.2.5) The reverse random deactivation module will deactivate each output vector Repeat steps (a.2.3)-(a.2.4) to obtain... One output: ; Then Output By concatenating the vectors, we obtain : ; Some units are temporarily dropped from the network with a certain probability, and the output is then randomly deactivated in reverse order. for: ； in, This represents the probability of discarding a neural network unit; (a.2.6) and the training set The output is obtained through a bidirectional cyclic feature extraction module, an attention mechanism module, and a reverse random deactivation module. ; (a.2.7) will output and output Integrate to obtain output : ; (a.2.8) Finally, the output will be... and learnable transform weights Multiply to obtain the training set. Prediction vector ; (a.2.9) Through the training set and prediction vector Construct a loss function and train the multi-channel dilated convolutional attention prediction model using the loss function. The optimized multi-channel dilated convolutional attention prediction model is then obtained and uploaded to the multi-channel dilated convolutional attention prediction module. (a.3) The multi-channel hollow convolutional attention prediction module uses the optimized multi-channel hollow convolutional attention prediction model to predict the sodium iodide spectrometer data to be tested, and obtains the predicted value of the dose rate of the key nuclide ionization chamber corresponding to the data. (a.4) The predicted value of the dose rate of the key nuclide ionization chamber corresponding to the data is output through the prediction result output module.

2. The key nuclide prediction device based on a multi-channel dilated convolutional attention model according to claim 1, characterized in that, The sodium iodide spectrometer data for ,in, Data from sodium iodide spectrometer The Middle Each spectrometer vector is a spectrometer vector. All dimensions are , Data from sodium iodide spectrometer Length, .

3. A key nuclide prediction device based on a multi-channel dilated convolutional attention model according to claim 2, characterized in that, The output vector The extraction process is as follows: No. Each dilated convolutional layer extracts feature channels, first using a dilation rate of [value missing]. dilated convolution pairs on the training set Feature extraction, and the use of hyperbolic units as activation functions, can be represented as follows: ； in, Indicates the output vector; The weights of the dilated convolution kernel are indicated. Represents hyperbolic units; These are hyperparameters in a hyperbolic element; The training set for ,in, For training set The Middle spectrometer vectors For training set Length, The output vector for ,in, Represents the output vector The Middle The output of a dilated convolution module.

4. A key nuclide prediction device based on a multi-channel dilated convolutional attention model according to claim 3, characterized in that, The specific steps (a.2.3) are as follows: The bidirectional cyclic feature extraction module includes a forward unidirectional cyclic feature extraction module and a backward unidirectional cyclic feature extraction module, wherein the forward unidirectional cyclic feature extraction module and the backward unidirectional cyclic feature extraction module are respectively composed of a forget gate, an input gate and an output gate; Output of the Forgot Gate It can be represented as: ； in, Indicates the first The output of the forget gate, Indicates the weight of the forget gate. Indicates the first The output of a unidirectional cyclic feature extraction module unit Indicates the first The input of a unidirectional cyclic feature extraction module unit, The bias term representing the forget gate. This represents the Sigmoid function; the forget gate is used. and The information that needs to be forgotten can be identified; Input gate output It can be represented as: ； in, Indicates the first The output of an input gate Indicates the weights of the input gates. This represents the bias term of the input gate; The input gate simultaneously updates the temporary cell state. : ； ； in, Indicates a temporary cell state. Indicates the cell state update weights. A bias term representing cell state updates. Indicates the first Individual cell states; input gate passes and We can determine which information needs to be updated through the input gate, and then obtain the new cell state based on the output of the forget gate and the output of the input gate. The output of the output gate can be represented as: ； in, This indicates the output of the output gate. Indicates the weight of the output gate. This represents the bias term of the output gate; the output gate can... and The output condition is used to determine the output of the unidirectional loop feature extraction module unit. for: ； Output of the bidirectional cyclic feature extraction module Represented as: ； in , This indicates that the forward unidirectional loop feature extraction module unit outputs the vector. The output, This indicates that the backward unidirectional loop feature extraction module unit outputs the vector. The output.

5. A key nuclide prediction device based on a multi-channel dilated convolutional attention model according to claim 4, characterized in that, The specific steps (a.2.4) are as follows: The attention mechanism module is an output. By assigning different weights, three matrices are obtained. , and ; then according to the matrix , and , get output The calculation formula is as follows: ； ； ； ； in, , and Representing different weight matrices, express dimensionality The input values ​​are converted into a probability distribution in the range [0, 1] with a sum of 1. The calculation formula is as follows: ； in Indicates the first There are 1 input value, where C represents the number of input nodes.

Images