Nuclear power valve fault diagnosis method and system

Abstract

The invention provides a nuclear power valve fault diagnosis method and system, and the method comprises the steps: collecting vibration signal data of different types of valves of a nuclear power plant, carrying out the normalization of the vibration signal data, and carrying out the drawing storage of a target picture; wherein the normal vibration signal data pictures in the target pictures are used as a training set, and the normal and abnormal vibration signal data pictures are used as a test set; a loss function of the ProGAN model is improved based on a Wasserstein distance, and a threshold determination module is introduced; and training the improved model based on the target picture. According to the nuclear power valve fault diagnosis method, the ProGAN model is improved, the nuclear power plant valve data imbalance problem and the valve diversity problem are effectively adapted, meanwhile, the loss function of the ProGAN is improved based on the Wasserstein distance, and the model training stability and the generated sample quality are improved.

Description

Technical Field

[0001] The present invention relates to the technical field of valve fault monitoring, and in particular to a nuclear power valve fault diagnosis method and system. Background Art

[0002] The safe operation of nuclear power plants depends on the good condition of their equipment. Valves, as one of the key components, play a vital role in controlling fluid flow and ensuring system pressure and temperature stability. Regular maintenance of valves and timely detection of potential faults can significantly reduce unexpected downtime and reduce maintenance costs.

[0003] However, nuclear power plant valve fault diagnosis faces the following problems:

[0004] (1) Data imbalance: The amount of data on normal operation far exceeds the amount of data on failure, which makes it difficult for traditional supervised learning methods to train effective models.

[0005] (2) Complexity and diversity: Valve failure may be caused by a variety of reasons and manifest in various forms, which makes it very difficult to identify the fault mode.

[0006] Based on this, the inventor of the present application proposes a nuclear power valve fault diagnosis method and system in order to solve the above technical problems. Summary of the invention

[0007] The technical problem to be solved by the present invention is to overcome the defect of great difficulty in valve fault diagnosis in the prior art and to provide a nuclear power valve fault diagnosis method and system.

[0008] The present invention solves the above technical problems through the following technical solutions:

[0009] The present invention provides a method for diagnosing nuclear power valve faults, which is characterized by comprising:

[0010] Step 1, collecting vibration signal data of different types of valves in a nuclear power plant, and normalizing the vibration signal data and then drawing and saving it as a target image; wherein the normal vibration signal data image in the target image is used as a training set, and the vibration signal data image containing normal and abnormal vibration signal data is used as a test set;

[0011] Step 2: Improve the loss function of the ProGAN model based on the Wasserstein distance and introduce a threshold determination module;

[0012] Step 3: training the improved model based on the target image.

[0013] According to one embodiment of the present invention, step 2 comprises:

[0014] Step 21, the original ProGAN uses the target image with the minimum resolution as the training object when initializing the network;

[0015] Step 22: Improve the loss function of the ProGAN model and introduce Wasserstein distance as a measure of the difference between the generated data distribution and the real data distribution;

[0016] Step 23: introduce a threshold determination module, determine the threshold based on the output distribution of the training set, calculate the reconstruction error of all samples on the training set, and then select a 0-100 percentile to determine a suitable threshold based on the error.

[0017] According to one embodiment of the present invention, in step 22, the calculation formula of the Wasserstein distance is as follows:

[0018]

[0019] Among them, E x,y～γ [||xy||] represents the expected value of the distance between x and y under the gamma distribution.

[0020] According to one embodiment of the present invention, step 3 includes in sequence: training the discriminator, training the generator, and progressive transition training.

[0021] According to one embodiment of the present invention, the process of training the discriminator is:

[0022] Step 411, randomly taking a tensor from a standard normal distribution, the tensor is passed to the generator to generate a fake image, and the generator network is used to convert the noise into a fake image;

[0023] Step 412: Create a first target tensor and a second target tensor; the first target tensor contains all labels of real images, and the second target tensor contains all labels of fake images;

[0024] Step 413: Use the discriminator to evaluate the real image to obtain a first output tensor, where the first output tensor represents the probability that the discriminator considers the real image to be real; use the discriminator to evaluate the generated fake image to obtain a second output tensor;

[0025] Step 414: Calculate a first loss between the discriminator's prediction of the real image and the real label, and calculate a second loss between the discriminator's prediction of the fake image and the fake label, and combine the first loss and the second loss into a total discriminator loss;

[0026] Step 415, perform back propagation to calculate the gradient of the loss with respect to all learnable parameters in the discriminator network;

[0027] Step 416: Use the optimizer to update the weights of the discriminator.

[0028] According to one embodiment of the present invention, the process of training the generator is:

[0029] Step 421: Generate a new random noise tensor, and use the noise tensor as input of the generator;

[0030] Step 422, using a generator network to convert the noise tensor into a batch of fake images;

[0031] Step 423: Send the generated fake images to the discriminator network to obtain the output value of the discriminator for the fake images; the output value represents the probability that the discriminator considers these images to be real;

[0032] Step 424, calculating the loss of the generator;

[0033] Step 425, perform back propagation on the generator loss and calculate the gradient of the loss with respect to all learnable parameters in the generator network;

[0034] Step 426: Use the optimizer to update the weights of the generator.

[0035] According to an embodiment of the present invention, the progressive training is divided into an initialization phase and a transition phase.

[0036] According to an embodiment of the present invention, in the initialization stage, alpha is set as a parameter for controlling the degree of mixing of the new and old images;

[0037] During the transition period, fade_in_epochs is set to represent the epoch value required for the transition from the old image to the new image.

[0038] According to one embodiment of the present invention, step 3 further includes: testing the test set data based on the trained model.

[0039] The present invention also provides a nuclear power valve fault diagnosis system, which is implemented by the nuclear power valve fault diagnosis method as described above, and the diagnosis system comprises:

[0040] The acquisition module is used to collect vibration signal data of different types of valves in the nuclear power plant, and normalize the vibration signal data and save it as a target image; wherein the normal vibration signal data image in the target image is used as a training set, and the vibration signal data image containing normal and abnormal vibration signal data is used as a test set;

[0041] Improvement module, used to improve the loss function of the ProGAN model based on Wasserstein distance and introduce a threshold determination module;

[0042] The training module is used to train the improved model based on the target image.

[0043] The positive and progressive effects of the present invention are:

[0044] The nuclear power valve fault diagnosis method of the present invention effectively adapts to the valve data imbalance problem and valve diversity problem of nuclear power plants by improving the ProGAN model. At the same time, the loss function of ProGAN is improved based on the Wasserstein distance, which improves the stability of model training and the quality of generated samples. BRIEF DESCRIPTION OF THE DRAWINGS

[0045] The above and other features, properties and advantages of the present invention will become more apparent through the following description in conjunction with the accompanying drawings and embodiments, in which:

[0046] Figure 1 This is a flow chart of the nuclear power valve fault diagnosis method of the present invention;

[0047] Figure 2 This is a diagram of the model training process of the present invention;

[0048] Figure 3 This is a flow chart of the discriminator training of the present invention;

[0049] Figure 4 This is a flow chart of generator training of the present invention;

[0050] Figure 5 This is a flow chart of the progressive transition training of the present invention. DETAILED DESCRIPTION

[0051] In order to more clearly illustrate the technical solutions of the embodiments of the present application, the following is a brief introduction to the drawings required for the description of the embodiments. Obviously, the drawings described below are only some examples or embodiments of the present application. For ordinary technicians in this field, the present application can also be applied to other similar scenarios based on these drawings without creative work. Unless it is obvious from the language environment or otherwise explained, the same reference numerals in the figures represent the same structure or operation.

[0052] As shown in this application and claims, unless the context clearly indicates an exception, the words "a", "an", "an" and / or "the" do not refer to the singular and may also include the plural. Generally speaking, the terms "include" and "comprise" only indicate the inclusion of the steps and elements that have been clearly identified, and these steps and elements do not constitute an exclusive list. The method or device may also include other steps or elements.

[0053] Unless otherwise specifically stated, the relative arrangement, numerical expressions and numerical values ​​of the parts and steps set forth in these embodiments do not limit the scope of the present application. At the same time, it should be understood that, for ease of description, the sizes of the various parts shown in the accompanying drawings are not drawn according to the actual proportional relationship. The technology, method and equipment known to those of ordinary skill in the relevant field may not be discussed in detail, but in appropriate cases, the technology, method and equipment should be considered as a part of the authorization specification. In all examples shown and discussed here, any specific value should be interpreted as being merely exemplary, rather than as a limitation. Therefore, other examples of exemplary embodiments may have different values. It should be noted that similar reference numerals and letters represent similar items in the following drawings, so that once a certain item is defined in an accompanying drawing, it does not need to be further discussed in subsequent drawings.

[0054] It should be noted that the present invention refers to and improves the ProGAN model to make it closer to the application scenario of nuclear power plant valves. Generative adversarial network (GAN) is a relatively new concept in machine learning. The goal of GAN is to synthesize artificial samples, such as images, that are indistinguishable from real images. The key innovation of ProGAN lies in progressive training. On the basis of classic GAN, it first creates the basic parts of the image by learning the basic features that can also be displayed in low-resolution images, and learns more and more details as the resolution increases and time goes by.

[0055] However, since the field of nuclear power valve fault diagnosis is more about the diversity of valve types, the present invention will improve the training process of the ProGAN model. That is, during progressive training, instead of inputting valve images from low resolution to high resolution, normal images of valves of different types are input. As normal images of valves of different types are input, network layers are gradually added to increase the model's feature learning of valves of different categories.

[0056] At the same time, a threshold determination module is introduced to enable the model to learn the characteristics of normal valve images, which can also be used to identify abnormal valve images.

[0057] Secondly, ProGAN often encounters the following problems during training:

[0058] 1. Mode collapse: The generator may only learn to generate a limited number of samples and ignore other modes.

[0059] 2. Unstable training: The capabilities of the generator and the discriminator need to be carefully balanced during training, otherwise training failure may occur.

[0060] Based on this, the present invention improves the loss function of ProGAN based on the Wasserstein distance, thereby improving the stability of training and the quality of generated samples.

[0061] Please refer to Figure 1 The nuclear power valve fault diagnosis method proposed by the present invention is described in detail as follows. The diagnosis method includes:

[0062] Step 1: collect vibration signal data of different types of valves in a nuclear power plant, normalize the vibration signal data, and save them as target images; wherein, the normal vibration signal data images in the target images are used as training sets, and the images containing normal and abnormal vibration signal data are used as test sets.

[0063] Specifically, a large amount of vibration signal data of different types of valves in nuclear power plants is first collected, and the collected quantity signals are normalized and plotted and saved as pictures. The training set is a normal vibration signal data picture, and the test set includes normal vibration signal data pictures and abnormal vibration signal data pictures.

[0064] Step 2: Based on the Wasserstein distance, the loss function of the ProGAN model is improved and a threshold determination module is introduced.

[0065] Specifically, step 2 includes:

[0066] Step 21: When the original ProGAN initializes the network, the target image with the minimum resolution is used as the training object.

[0067] That is, the original ProGAN uses the minimum resolution image as the training object when initializing the network. First, a normal image of a certain type of valve is used as the training object. Different valves have different fault manifestations. Therefore, when the training is stable at this stage, new normal images of valve categories are added and the discriminator model and generator model are improved by adding new network layers to adapt to the changes in valve categories. The dynamic model can learn different valve characteristics.

[0068] Step 22: Improve the loss function of the ProGAN model and introduce Wasserstein distance as a measure of the difference between the generated data distribution and the real data distribution.

[0069] Wasserstein distance is used to indicate the similarity between two distributions, as shown in the following formula:

[0070]

[0071] Where Π(p, q) represents the set of all possible joint distributions of distributions p and q. For each possible joint distribution γ, we can sample (x, y)~γ to get a sample x and y, and calculate the distance between the pair of samples (x and y) ||xy||, E x,y～γ[||xy||] represents the expected value of the distance between x and y under the γ distribution. The lower bound of this expected value in all possible joint distributions is the Wasserstein distance.

[0072] Step 23: introduce a threshold determination module, determine the threshold based on the output distribution of the training set, calculate the reconstruction error of all samples on the training set, and then select the 0-100 percentile to determine the appropriate threshold based on the error.

[0073] The present invention determines a threshold based on the output distribution of a training set, calculates reconstruction errors of all samples on the training set, and then selects a 0-100 percentile based on the errors to determine a suitable threshold.

[0074] It should be noted that a high threshold means that only very high outputs will be considered true, which will reduce the number of false positives of the model, but it will also cause some normal images to be misclassified as abnormal images.

[0075] Step 3: Train the improved model based on the target image.

[0076] Reference Figure 2 ,The model training process is divided into three stages: training the discriminator, training the generator, and progressive transition training.

[0077] Please refer to Figure 3 , the process of training the discriminator is:

[0078] Step 411, randomly take a tensor from the standard normal distribution, the tensor is passed to the generator to generate a fake image, and the generator network is used to convert the noise into a fake image.

[0079] Step 412: Create a first target tensor and a second target tensor; the first target tensor contains all labels of real images, and the second target tensor contains all labels of fake images.

[0080] The first target tensor is used to train the discriminator to recognize real images, and the second target tensor is used to train the discriminator to recognize fake images.

[0081] Step 413: Use the discriminator to evaluate the real image to obtain a first output tensor, where the first output tensor represents the probability that the discriminator believes that the real image is real; use the discriminator to evaluate the generated fake image to obtain a second output tensor.

[0082] Step 414, calculate a first loss between the discriminator's prediction of the real image and the real label, and calculate a second loss between the discriminator's prediction of the fake image and the fake label, and combine the first loss and the second loss into a total discriminator loss.

[0083] Step 415: Perform back propagation to calculate the gradient of the loss with respect to all learnable parameters in the discriminator network.

[0084] Step 416: Use the optimizer to update the weights of the discriminator. The optimizer adjusts the weights based on the previous gradient to minimize the loss function.

[0085] In this way, the discriminator tries to maximize the probability of giving a “true” label to real images, while minimizing the probability of giving a “true” label to fake images.

[0086] Please refer to Figure 4 , the process of training the generator is:

[0087] Step 421: Generate a new random noise tensor and use the noise tensor as the input of the generator.

[0088] Step 422: Use the generator network to convert the noise tensor into a batch of fake images.

[0089] Step 423: Send the generated fake images to the discriminator network to obtain the output value of the discriminator for the fake images; the output value represents the probability that the discriminator considers these images to be real.

[0090] Step 424: Calculate the loss of the generator.

[0091] The goal of the generator is to make the discriminator believe as much as possible that the generated image is real. Therefore, the goal of the generator is to maximize the probability that the discriminator gives a "real" result for a fake image, which is equivalent to minimizing the generator loss function.

[0092] Step 425: Perform backpropagation on the generator loss to calculate the gradient of the loss with respect to all learnable parameters in the generator network.

[0093] Step 426: Use the optimizer to update the weights of the generator.

[0094] The optimizer adjusts the weights based on previous gradients, allowing the generator to generate more realistic images and thus deceive the discriminator.

[0095] Please refer to Figure 5 ,The progressive training is divided into the initialization phase and the transition phase.

[0096] During the initialization phase, alpha is set as a parameter to control the degree of blending of the new and old images;

[0097] During the transition phase, set fade_in_epochs to indicate the epoch value required for the old image to transition to the new image.

[0098] Increasing the value of fade_in_epochs means extending the length of the transition period. This is mainly to allow the model to adapt more smoothly to more categories of valve images.

[0099] That is, in this process, the generator tries to generate more realistic images, while the discriminator tries to distinguish between real images and generated images. Through such adversarial training, the generator gradually learns to generate high-quality images.

[0100] After step 3, the method further includes: testing the test set data based on the trained model.

[0101] That is, if the result of the model test is greater than the threshold, the output is normal; otherwise, the output is abnormal.

[0102] The present invention also proposes a nuclear power valve fault diagnosis system, which is implemented by the above nuclear power valve fault diagnosis method, and is characterized in that the diagnosis system includes:

[0103] The acquisition module is used to collect vibration signal data of different types of valves in nuclear power plants, and normalize the vibration signal data and save them as target images; wherein the normal vibration signal data images in the target images are used as training sets, and the images containing normal and abnormal vibration signal data are used as test sets;

[0104] Improvement module, used to improve the loss function of the ProGAN model based on Wasserstein distance and introduce a threshold determination module;

[0105] The training module is used to train the improved model based on the target image.

[0106] In summary, the improvements of the present invention include at least the following parts:

[0107] First: the original ProGAN uses the minimum resolution image as the training object when initializing the network. However, the present invention uses a normal image of a certain type of valve as the training object. The fault manifestations of different valves may be different. Therefore, when the training is sufficiently stable at this stage, new normal images of valve categories are added and the discriminator model and generator model are improved and new network layers are added to adapt to the changes in valve categories. The dynamic model can learn different valve features to avoid using the same network structure for all types of valves, resulting in some network nodes failing and becoming 0 on different valves.

[0108] Second, the loss function of the ProGAN model is improved. The traditional GAN ​​model may have a pattern collapse problem, that is, the generator only learns part of the patterns in the training data set, resulting in a decrease in the diversity of the generated samples and the possible loss of some important valve features.

[0109] The present invention introduces Wasserstein distance as a method to measure the difference between the generated data distribution and the real data distribution. It has a more intuitive geometric interpretation and measures the minimum amount of work required to transform one distribution into another in space. This feature makes it more effective in image processing and other data. Wasserstein also provides a more stable gradient, making the training process more stable.

[0110] Third, introduce the threshold determination module.

[0111] Traditional GAN ​​models or ProGAN models do not have a threshold determination module, which results in the model using a single threshold for abnormality judgment. For example, in a binary or multi-classification model, the value with the highest probability is output as the predicted value (the sum of all class probabilities is 1), even if the probability is very low, such as 0.5. However, in general, if the predicted value is relatively accurate, the predicted probability will be close to 1.

[0112] Therefore, the present invention introduces a threshold determination module to determine a threshold to distinguish normal and abnormal samples. Specifically, the threshold is determined based on the output distribution of the training set, the reconstruction error of all samples is calculated on the training set, and then a suitable threshold between 0 and 100 is selected based on these error values. A high threshold means that only those very high outputs will be considered true, which may reduce the number of false positives of the model, but it may also cause some normal images to be misjudged as abnormal images. Determining a suitable threshold helps to improve the accuracy of the model output.

[0113] In the description of the embodiments of the present application, unless otherwise clearly specified and limited, technical terms such as "install", "connect", "connect", "fixed" and the like should be understood in a broad sense. For example, it can be a fixed connection, a detachable connection, or an integral connection; it can also be a mechanical connection. For ordinary technicians in this field, the specific meanings of the above terms in the embodiments of the present application can be understood according to the specific circumstances.

[0114] In addition, it should be noted that the use of words such as "first" and "second" to define components is only for the convenience of distinguishing the corresponding components. If not otherwise stated, the above words have no special meaning and cannot be understood as limiting the scope of protection of this application. In addition, although the terms used in this application are selected from well-known and commonly used terms, some terms mentioned in the specification of this application may be selected by the applicant at his or her discretion, and their detailed meanings are explained in the relevant parts of the description of this article. In addition, it is required to understand this application not only by the actual terms used, but also by the meaning implied by each term.

[0115] Flowcharts are used in the present application to illustrate the operations performed by the system according to the embodiments of the present application. It should be understood that the preceding or following operations are not necessarily performed accurately in order. On the contrary, various steps may be processed in reverse order or simultaneously. At the same time, other operations may be added to these processes, or one or more operations may be removed from these processes.

[0116] The present application uses specific words to describe the embodiments of the present application. For example, "one embodiment", "an embodiment", and / or "some embodiments" refer to a certain feature, structure or characteristic related to at least one embodiment of the present application. Therefore, it should be emphasized and noted that "one embodiment" or "an embodiment" or "an alternative embodiment" mentioned twice or multiple times in different positions in this specification does not necessarily refer to the same embodiment. In addition, some features, structures or characteristics in one or more embodiments of the present application can be appropriately combined.

[0117] Although the present invention is disclosed as above with preferred embodiments, it is not intended to limit the present invention. Any person skilled in the art may make possible changes and modifications without departing from the spirit and scope of the present invention. Therefore, any modification, equivalent change and modification made to the above embodiments according to the technical essence of the present invention without departing from the content of the technical solution of the present invention shall fall within the protection scope defined by the claims of the present invention.

Claims

1. A method for diagnosing nuclear power valve faults, characterized in that: include: Step 1, collecting vibration signal data of different types of valves in a nuclear power plant, and normalizing the vibration signal data and then drawing and saving it as a target image; wherein the normal vibration signal data image in the target image is used as a training set, and the vibration signal data image containing normal and abnormal vibration signal data is used as a test set; Step 2: Improve the loss function of the ProGAN model based on the Wasserstein distance and introduce a threshold determination module; Step 3: training the improved model based on the target image.

2. The nuclear power valve fault diagnosis method according to claim 1, characterized in that: The step 2 comprises: Step 21, the original ProGAN uses the target image with the minimum resolution as the training object when initializing the network; Step 22: Improve the loss function of the ProGAN model and introduce Wasserstein distance as a measure of the difference between the generated data distribution and the real data distribution; Step 23: introduce a threshold determination module, determine the threshold based on the output distribution of the training set, calculate the reconstruction error of all samples on the training set, and then select a 0-100 percentile to determine a suitable threshold based on the error.

3. The nuclear power valve fault diagnosis method according to claim 2, characterized in that: In step 22, the calculation formula of the Wasserstein distance is as follows: Among them, E x,y～γ [||xy||] represents the expected value of the distance between x and y under the gamma distribution.

4. The nuclear power valve fault diagnosis method according to claim 1, characterized in that: The step 3 includes in sequence: training the discriminator, training the generator, and progressive transition training.

5. The nuclear power valve fault diagnosis method according to claim 4 is characterized in that: The process of training the discriminator is: Step 411, randomly taking a tensor from a standard normal distribution, the tensor is passed to the generator to generate a fake image, and the generator network is used to convert the noise into a fake image; Step 412: Create a first target tensor and a second target tensor; the first target tensor contains all labels of real images, and the second target tensor contains all labels of fake images; Step 413: Use the discriminator to evaluate the real image to obtain a first output tensor, where the first output tensor represents the probability that the discriminator considers the real image to be real; Use the discriminator to evaluate the generated fake image and obtain a second output tensor; Step 414: Calculate a first loss between the discriminator's prediction of the real image and the real label, and calculate a second loss between the discriminator's prediction of the fake image and the fake label, and combine the first loss and the second loss into a total discriminator loss; Step 415, perform back propagation to calculate the gradient of the loss with respect to all learnable parameters in the discriminator network; Step 416: Use the optimizer to update the weights of the discriminator.

6. The nuclear power valve fault diagnosis method according to claim 4, characterized in that: The process of training the generator is: Step 421: Generate a new random noise tensor, and use the noise tensor as input of the generator; Step 422, using a generator network to convert the noise tensor into a batch of fake images; Step 423: Send the generated fake image to the discriminator network to obtain an output value of the discriminator for the fake image; The output values ​​represent the probability that the discriminator considers these images to be real; Step 424, calculating the loss of the generator; Step 425, perform back propagation on the generator loss and calculate the gradient of the loss with respect to all learnable parameters in the generator network; Step 426: Use the optimizer to update the weights of the generator.

7. The nuclear power valve fault diagnosis method according to claim 4, characterized in that: The progressive training is divided into an initialization phase and a transition phase.

8. The method for diagnosing nuclear power valve faults according to claim 7, characterized in that: In the initialization stage, alpha is set as a parameter to control the degree of mixing of the new and old images; During the transition period, fade_in_epochs is set to represent the epoch value required for the transition from the old image to the new image.

9. The nuclear power valve fault diagnosis method according to claim 1, characterized in that: The step 3 also includes: testing the test set data based on the trained model.

10. A nuclear power valve fault diagnosis system, implemented by the nuclear power valve fault diagnosis method according to any one of claims 1 to 9, characterized in that: The diagnostic system comprises: The acquisition module is used to collect vibration signal data of different types of valves in the nuclear power plant, and normalize the vibration signal data and save it as a target image; wherein the normal vibration signal data image in the target image is used as a training set, and the vibration signal data image containing normal and abnormal vibration signal data is used as a test set; Improvement module, used to improve the loss function of the ProGAN model based on Wasserstein distance and introduce a threshold determination module; The training module is used to train the improved model based on the target image.

Images