Nuclear power key equipment measuring point optimization method based on interpretable neural network

Abstract

The invention relates to the field of nuclear power, in particular to a nuclear power key equipment measuring point optimization method based on an interpretable neural network. The method comprises the following steps: acquiring original data and a diagnosis result of equipment operation by utilizing an intelligent diagnosis system on the equipment; establishing an equipment fault mode recognition model through a sparse neural network; calculating the contribution degree of input of each measuring point to an output fault type through a Grad-CAM + + technology; and evaluating the contribution degree level of each measuring point to the output fault type, and carrying out measuring point optimization according to an evaluation result. According to the method, the measuring points set by the nuclear power equipment are evaluated, the measuring points with low contribution degree to the evaluation result are optimized, and the economical efficiency and efficiency of operation and maintenance of key equipment can be improved.

Description

Technical Field

[0001] The present invention relates to the field of nuclear power, and in particular to a method for optimizing measuring points of key nuclear power equipment based on an interpretable neural network. Background Art

[0002] In recent years, the share of nuclear power in the global energy market has continued to rise, and the safety and reliability requirements for the operation of key equipment in nuclear power plants have also been raised to a new level. With the introduction of Industry 4.0 and the development of artificial intelligence and big data technology in recent years, higher requirements have been put forward for the intelligent operation and maintenance of key equipment in nuclear power plants, and there are endless cases of design, development and deployment of intelligent diagnosis systems for key equipment in nuclear power plants. When conducting intelligent diagnosis research on key equipment in nuclear power plants, it is often necessary to arrange multiple measurement points. Generally speaking, the more measurement points there are, the more we can understand the equipment operating parameters and obtain more accurate equipment operation status and health assessment results. However, the more measurement points there are, the more instruments and equipment are required, the more spare parts need to be prepared for each maintenance, the operation and maintenance costs will be greatly increased, and the difficulty of data processing will also increase. Therefore, the research on the optimization arrangement method of measurement points in the intelligent diagnosis system of key equipment in nuclear power plants is of great significance, and it also provides direction for the overall planning of subsequent digital nuclear power. Summary of the invention

[0003] The technical problem to be solved by the present invention is: to provide a method for optimizing measuring points of key nuclear power equipment based on an interpretable neural network, to evaluate the measuring points set on the nuclear power equipment, to optimize the measuring points with low contribution to the evaluation results, and to improve the economy and efficiency of the operation and maintenance of key equipment.

[0004] The present invention provides a method for optimizing measuring points of key nuclear power equipment based on an interpretable neural network, comprising:

[0005] Step 1: Use the intelligent diagnosis system on the equipment to obtain the original data and diagnosis results of the equipment operation;

[0006] Step 2: Establish equipment failure mode recognition model through sparse neural network;

[0007] Step 3: Calculate the contribution of each measurement point input to the output fault type through Grad-CAM++ technology;

[0008] Step 4: Evaluate the contribution level of each measurement point to the output fault type and optimize the measurement points based on the evaluation results.

[0009] In a specific embodiment of the present invention, step 1 is specifically as follows: for multiple intelligent diagnostic systems arranged on the equipment, according to the measuring points and signal conditioning and acquisition units of the intelligent systems, the original data of the equipment operation is obtained, and the original data is integrated according to time, and saved as multi-parameter original files at different times, and the current diagnostic results of the equipment are given according to the intelligent diagnostic system.

[0010] In a specific embodiment of the present invention, the step 2 is specifically:

[0011] Using the neural network algorithm, a mapping function structure from equipment vibration state data X to fault mode Y is established, and the function structure parameters are adjusted using the stochastic gradient descent algorithm to finally obtain the optimal mapping function in the hypothesis space:

[0012]

[0013] In the formula, f θ is the hypothesis space containing the optimal mapping function with parameter θ, and R is the risk function. In order to make the output result as relevant as possible to the values ​​measured by some of the channels, it is necessary to impose constraints on the model parameters in combination with the regularization operator, so that the model parameters are as sparse as possible, that is:

[0014] R(f θ )=L(f θ (X),Y)+λJ(θ)

[0015] In the formula, L(f θ (X),Y) represents the theoretical model f θ The degree of inconsistency between (X) and the true value Y in the average sense of the joint distribution P(X,Y); J(θ) represents the sparsity of the parameter θ in the theoretical model, and λ is the penalty coefficient. The sparser the parameter θ, the more concentrated the attention on the input X.

[0016] In a specific embodiment of the present invention, in step 3, the calculation process of the Grad-CAM++ technology is:

[0017]

[0018] The normalization process is for the cth fault type, all The maximum value is 1, is the mapping coefficient from the i-th data in the d-th measurement point to the k-th convolution kernel output, It represents the weight of the model's kth convolution kernel output to fault type c, and its value is expressed as:

[0019]

[0020] In the formula represents the gradient of the target feature map, It represents the weight corresponding to the fault type c output by the kth convolution kernel after the i-th data in the d-th measuring point is processed.

[0021] In a specific embodiment of the present invention, step 4 is specifically as follows:

[0022] Contribution of the i-th data in the d-th measurement point to the c-th fault type The importance of each measuring point is calculated for the maximum value, and the total contribution of all data of the dth measuring point to the fault type c is obtained;

[0023]

[0024] In order to determine the contribution level of the dth measuring point to the output fault type, the total maximum contribution Co of all fault types is calculated by using the dth measuring point data. d To indicate the level of contribution;

[0025]

[0026] Finally, when the total maximum contribution is less than the threshold 1, it is determined that the contribution of the measuring point to the establishment of the equipment fault diagnosis model is not large, and the measuring point is optimized.

[0027] In a specific embodiment of the present invention, in step 2, when using a neural network algorithm, each measurement point passes through a set of convolutional layers and pooling layers separately.

[0028] Compared with the prior art, the present invention is a nuclear power key equipment measurement point optimization method based on interpretable neural network, combined with the existing intelligent diagnosis system of the power station, the data collected by the system is integrated, combined with sparse neural network, the equipment fault model in the operation process is trained twice, and the Grad-CAM++ technology is used to mine the contribution of each sensor to the output fault type, and the total maximum contribution Co of the dth measurement point data to all fault types is calculated. d To indicate the contribution level, the measuring points with low contribution can be optimized, and finally the number of measuring points can be optimized and the operation and maintenance cost can be reduced. BRIEF DESCRIPTION OF THE DRAWINGS

[0029] Figure 1 It represents the optimization flow chart of measuring points of key nuclear power equipment based on interpretable neural network;

[0030] Figure 2 A schematic diagram showing the arrangement of equipment sensors;

[0031] Figure 3 A schematic diagram showing a sparse neural network used in the present invention;

[0033] Figure 4 Indicates the contribution of the input data of measuring point 5 to fault type 3 calculated based on Grad-CAM++ technology. DETAILED DESCRIPTION

[0034] In order to further understand the present invention, the embodiments of the present invention are described below in conjunction with examples, but it should be understood that these descriptions are only for further illustrating the features and advantages of the present invention, rather than limiting the present invention.

[0035] The embodiment of the present invention discloses a method for optimizing measurement points of key nuclear power equipment based on an interpretable neural network, such as Figure 1 As shown, including:

[0036] Step 1: Use the intelligent diagnosis system on the equipment to obtain the original data and diagnosis results of the equipment operation;

[0037] Specifically: for multiple intelligent diagnosis systems deployed on the equipment, the original data of the equipment operation is obtained according to the measurement points and signal conditioning and acquisition units of the intelligent systems, and the original data is fused according to time and saved as multi-parameter original files at different times, and the current diagnosis results of the equipment are given according to the intelligent diagnosis system.

[0038] Step 2: Establish equipment failure mode recognition model through sparse neural network;

[0039] Specifically:

[0040] Using the neural network algorithm, each measuring point passes through a set of convolution layers and pooling layers to establish a mapping function structure from the equipment vibration state data X to the fault mode Y, and the stochastic gradient descent algorithm is used to adjust the function structure parameters to finally obtain the optimal mapping function in the hypothesis space:

[0041] In the formula, f θ is the hypothesis space containing the optimal mapping function with parameter θ, and R is the risk function. In order to make the output result as relevant as possible to the values ​​measured by some of the channels, it is necessary to impose constraints on the model parameters in combination with the regularization operator, so that the model parameters are as sparse as possible, that is:

[0042] R(f θ )=L(f θ (X),Y)+λJ(θ)

[0043] In the formula, L(f θ (X),Y) represents the theoretical model f θThe degree of inconsistency between (X) and the true value Y in the average sense of the joint distribution P(X,Y); J(θ) represents the sparsity of the parameter θ in the theoretical model, and λ is the penalty coefficient. The sparser the parameter θ, the more concentrated the attention on the input X.

[0044] Step 3: Calculate the contribution of each measurement point input to the output fault type through Grad-CAM++ technology;

[0045] The calculation process of the Grad-CAM++ technology is:

[0046]

[0047] The normalization process is for the cth fault type, all The maximum value is 1, is the mapping coefficient from the i-th data in the d-th measurement point to the k-th convolution kernel output, It represents the weight of the model's kth convolution kernel output to fault type c, and its value is expressed as:

[0048]

[0049] In the formula represents the gradient of the target feature map, It represents the weight corresponding to the fault type c output by the kth convolution kernel after the i-th data in the d-th measuring point is processed.

[0050] Step 4: Evaluate the contribution level of each measurement point to the output fault type and optimize the measurement points based on the evaluation results.

[0051] 4 Specifically:

[0052] Contribution of the i-th data in the d-th measurement point to the c-th fault type The importance of each measuring point is calculated for the maximum value, and the total contribution of all data of the dth measuring point to the fault type c is obtained;

[0053]

[0054] In order to determine the contribution level of the dth measuring point to the output fault type, the total maximum contribution Co of all fault types is calculated by using the dth measuring point data. d To indicate the level of contribution;

[0055]

[0056] Finally, when the total maximum contribution is less than the threshold 1, it is determined that the contribution of the measuring point to the establishment of the equipment fault diagnosis model is not large, and the measuring point is optimized.

[0057] In order to further understand the present invention, the present invention is described in detail below in conjunction with the embodiments, and the protection scope of the present invention is not limited by the following embodiments.

[0058] Example 1

[0059] Step 1: Obtain the vibration signal of the pump from the nuclear power water pump intelligent diagnosis system. Six vibration sensors are evenly arranged on the outer pressure shell near the upper and lower guide bearings of the water pump. At the same time, the inlet and outlet flow, pressure, temperature and other parameters of the nuclear power water pump are obtained. The original data X is fused according to the time when the data is generated, and saved as multi-parameter original files at different times. According to the current diagnosis result Y of the equipment given by the nuclear power water pump intelligent diagnosis system, the original file is labeled to lay the foundation for subsequent model training. The state parameters obtained by the water pump are shown in the attached figure. Figure 2 shown.

[0060] Step 2: Convert the vibration signal in the original data X to the frequency domain, and use the small convolution kernel convolutional neural network to learn the data characteristics of each channel. The state value does not need to be learned. Use the attention network to measure the similarity of the convolution results. After the measurement results are transformed by the fully connected layer, they are sent to the softmax network for classification. Finally, the mapping function structure from the pump vibration state data X to the fault mode Y is established. The network structure is as follows: Figure 3 As shown in the figure, the loss function selects the sum of the cross entropy of the classification result and the L1 norm of the model parameter θ, the penalty coefficient λ is 0.05, and the stochastic gradient descent algorithm is used to adjust the function structure parameters. When the loss function error is less than 1e-6, the iteration ends, and finally the optimal mapping function in the hypothesis space is obtained:

[0061] Y=f(X)

[0062] Step 3: For the water pump fault pattern recognition model, the Grad-CAM++ technology is used to calculate the contribution of the model input data to the output result, and the contribution of the i-th data in the d-th measurement point (the state data includes inlet and outlet pressure difference, outlet temperature, flow rate and other parameters, which are input into the convolutional network as an array) to the c-th fault type can be obtained. The contribution of the input data of measuring point 5 to fault type 3 is shown in the attached figure. Figure 4 As shown, the color depth indicates the degree of contribution.

[0063] Contribution of the i-th data in the d-th measurement point to the c-th fault type The importance of each measuring point is maximized separately to obtain the total contribution of all data of the dth measuring point to fault type c. The specific total contribution values ​​are shown in Table 1.

[0064]

[0065] Table 1 Total contribution of all data of the dth measurement point to fault type c

[0066]

[0067] In order to determine the contribution level of the dth measuring point to the output fault type, the total maximum contribution Co of all fault types is calculated by using the dth measuring point data. d To indicate the level of contribution.

[0068]

[0069] Step 4: When the total maximum contribution is less than the threshold value 1, it is determined that the measurement point does not contribute much to the establishment of the equipment fault diagnosis model, and the measurement point can be optimized. According to the elimination principle, in this water pump fault diagnosis system, the measurement point 6 and the state quantity 1 can be optimized. According to engineering experience, the measurement point 6 represents the axial vibration, which is similar to the function of the measurement point 5. The function of state 1 (inlet and outlet temperature difference) is similar to the parameters of state 3 (flow rate), which is similar to engineering experience, indicating the feasibility of the patent of this invention.

[0070] The above embodiments are only used to help understand the method and core idea of ​​the present invention. It should be noted that, for those skilled in the art, several improvements and modifications can be made to the present invention without departing from the principles of the present invention, and these improvements and modifications also fall within the scope of protection of the claims of the present invention.

[0071] The above description of the disclosed embodiments enables one skilled in the art to implement or use the present invention. Various modifications to these embodiments will be apparent to one skilled in the art, and the general principles defined herein may be implemented in other embodiments without departing from the spirit or scope of the present invention. Therefore, the present invention will not be limited to the embodiments shown herein, but rather to the widest scope consistent with the principles and novel features disclosed herein.

Claims

1. A method for optimizing measurement points of key nuclear power equipment based on interpretable neural networks, characterized in that: include: Step 1: Use the intelligent diagnosis system on the equipment to obtain the original data and diagnosis results of the equipment operation; Step 2: Establish equipment failure mode recognition model through sparse neural network; Step 3: Calculate the contribution of each measurement point input to the output fault type through Grad-CAM++ technology; Step 4: Evaluate the contribution level of each measurement point to the output fault type and optimize the measurement points based on the evaluation results.

2. The method for optimizing measuring points of key nuclear power equipment based on interpretable neural network according to claim 1 is characterized in that: The step 1 is specifically as follows: for multiple intelligent diagnosis systems arranged on the equipment, according to the measurement points and signal conditioning and acquisition units of the intelligent systems, the original data of the equipment operation is obtained, and the original data is integrated according to time, and saved as multi-parameter original files at different times, and the current diagnosis result of the equipment is given according to the intelligent diagnosis system.

3. The method for optimizing measuring points of key nuclear power equipment based on interpretable neural network according to claim 1 is characterized in that: The step 2 is specifically as follows: Using the neural network algorithm, a mapping function structure from equipment vibration state data X to fault mode Y is established, and the function structure parameters are adjusted using the stochastic gradient descent algorithm to finally obtain the optimal mapping function in the hypothesis space: In the formula, f θ is the hypothesis space containing the optimal mapping function with parameter θ, and R is the risk function. In order to make the output result as relevant as possible to the values ​​measured by some of the channels, it is necessary to impose constraints on the model parameters in combination with the regularization operator, so that the model parameters are as sparse as possible, that is: R(f θ )=L(f θ (X),Y)+λJ(θ) In the formula, L(f θ (X),Y) represents the theoretical model f θ The degree of inconsistency between (X) and the true value Y in the average sense of the joint distribution P(X,Y); J(θ) represents the sparsity of the parameter θ in the theoretical model, and λ is the penalty coefficient. The sparser the parameter θ, the more concentrated the attention on the input X.

4. The method for optimizing measuring points of key nuclear power equipment based on interpretable neural network according to claim 1 is characterized in that: In step 3, the calculation process of the Grad-CAM++ technology is: The normalization process is for the cth fault type, all The maximum value is 1, is the mapping coefficient from the i-th data in the d-th measurement point to the k-th convolution kernel output, It represents the weight of the model's kth convolution kernel output to fault type c, and its value is expressed as: In the formula represents the gradient of the target feature map, It represents the weight corresponding to the fault type c output by the kth convolution kernel after the i-th data in the d-th measuring point is processed.

5. The method for optimizing measuring points of key nuclear power equipment based on interpretable neural network according to claim 1 is characterized in that: The step 4 is specifically as follows: Contribution of the i-th data in the d-th measurement point to the c-th fault type The importance of each measuring point is calculated for the maximum value, and the total contribution of all data of the dth measuring point to the fault type c is obtained; In order to determine the contribution level of the dth measuring point to the output fault type, the total maximum contribution Co of all fault types is calculated by using the dth measuring point data. d To indicate the level of contribution; Finally, when the total maximum contribution is less than the threshold 1, it is determined that the contribution of the measuring point to the establishment of the equipment fault diagnosis model is not large, and the measuring point is optimized.

6. The method for optimizing measuring points of key nuclear power equipment based on interpretable neural network according to claim 1 is characterized in that: In step 2, when using the neural network algorithm, each measurement point passes through a set of convolutional layers and pooling layers separately.

Images