A method and system for diagnosing faults of nuclear power feedwater pumps

Abstract

The present invention provides a nuclear power feed water pump fault diagnosis method and system, which relate to the field of water pump operation status analysis. To address the problem of low accuracy in current nuclear power feed water pump fault diagnosis, the present invention obtains water pump operation data in other scenarios of long-term operation, extracts fault features, and combines transfer learning and deep convolutional networks to transfer the fault features in other working conditions to the health features of the nuclear power feed water pump as samples to update the fault diagnosis model, thereby improving the diagnosis accuracy under the condition of missing fault samples.

Description

Technical Field

[0001] The present invention relates to the field of water pump operation status analysis, and in particular to a nuclear power feed water pump fault diagnosis method and system. Background Art

[0002] As an important component of the reactor cooling system of a nuclear power plant, the high-speed electric feedwater pump can keep the temperature of the nuclear reactor within the normal range. Its normal operation is the key to ensuring the safe operation of the nuclear power plant. The data-driven feedwater pump fault diagnosis has the advantages of high accuracy and wide adaptability. This method does not require the establishment of detailed knowledge or complex physical models for the system. It uses a large amount of system status data, combined with machine learning and deep learning technologies to realize the automated fault diagnosis process, reducing manual intervention and labor costs.

[0003] However, newly built nuclear power plants only record normal data during short-term operation, lacking fault data accumulation. Moreover, the equipment is in a normal state during long-term service. Even if a fault occurs, the fault type is difficult to cover all the states of the equipment, resulting in reduced diagnostic accuracy. A Chinese patent (publication number CN113669246 B) discloses an intelligent diagnosis method for water pump faults under cross-operating conditions. Through fault classification and operating domain classification model learning and joint optimization, the source domain vibration information and the target domain vibration information are aligned in the high-order vector feature space, thereby realizing the diagnosis of water pump faults under cross-operating conditions. A large amount of target domain vibration information is required to match and align with the source domain vibration information, but the nuclear power feed water pump has high precision and stability. It is difficult to provide sufficient fault data during long-term service. It is difficult to meet the diagnosis needs under the condition of missing fault samples, and the fault diagnosis accuracy is low. Summary of the invention

[0004] The purpose of the present invention is to provide a nuclear power feed water pump fault diagnosis method and system to address the defects of the prior art, obtain the pump operation data under other scenarios of long-term operation, extract fault features, and combine transfer learning and deep convolutional networks to migrate the fault features under other working conditions to the health features of the nuclear power feed water pump as samples to update the fault diagnosis model, thereby improving the diagnosis accuracy under the condition of missing fault samples.

[0005] The first object of the present invention is to provide a method for diagnosing a nuclear power feedwater pump fault, which adopts the following scheme:

[0006] include:

[0007] The operation data of the nuclear power feed water pump is input into the nuclear power feed water pump fault diagnosis model, and the nuclear power feed water pump fault diagnosis model is used to diagnose the feed water pump fault;

[0008] Among them, the update of the nuclear power feed water pump fault diagnosis model includes:

[0009] Obtain and process the health status data of nuclear power feed water pumps and the operation data of thermal power feed water pumps;

[0010] Extract the data features of thermal power feed water pumps and use them as the source domain, and extract the health features of nuclear power feed water pumps as the target domain;

[0011] According to the distribution similarity of the health features in the source domain and the health features in the target domain, the corresponding thermal power feed water pump fault features are extracted and combined with the health features of the nuclear power feed water pump to obtain samples containing the nuclear power feed water pump fault features;

[0012] The samples are used as training data to update the nuclear power feedwater pump fault diagnosis model.

[0013] Furthermore, the health status data of nuclear power feedwater pumps and the operation data of thermal power feedwater pumps are processed to capture the continuity and change trend between the characteristic data and retain the key data features of the original data.

[0014] Furthermore, capturing the continuity and change trend between the feature data includes:

[0015] The input data is mapped to the mean and variance potential representation in the latent space. The latent space is a Gaussian distribution. The parameters of the Gaussian distribution are determined according to the mean and variance to capture the continuity and change trend between features.

[0016] Furthermore, the key data features of the original data retained include:

[0017] The latent representation is decoded into reconstructed features in the original data space, preserving the key data features of the original data.

[0018] Furthermore, before calculating the distribution similarity of the health features, the feature extraction module is used to extract the distribution features of the thermal power feed water pump row data and the distribution features of the health status data of the nuclear power feed water pump.

[0019] Furthermore, the similarity is measured by the difference between the healthy features in the source and target domains, and the parameters of the feature extraction module are updated through backpropagation based on the maximum similarity.

[0020] Furthermore, through parameter sharing, the extracted healthy features in the source domain and the healthy features in the target domain are made close to each other, thereby improving the fault features of the thermal power feed water pump extracted in the source domain to be close to the fault features of the nuclear power feed water pump.

[0021] Furthermore, through transfer learning, the fault features of thermal power feed water pumps extracted from the source domain are migrated to the operation data of nuclear power feed water pumps in the target domain, and input into the deep convolutional network to generate samples containing the fault features of nuclear power feed water pumps.

[0022] Furthermore, a nuclear feed water pump fault diagnosis model is established using the initial operation data of the nuclear feed water pump, the nuclear feed water pump fault diagnosis model is updated with samples, the updated nuclear feed water pump fault diagnosis model is tested, and the diagnosis accuracy is calculated and evaluated.

[0023] A second object of the present invention is to provide a nuclear power feedwater pump fault diagnosis system, comprising:

[0024] The diagnosis module is configured as follows: the operation data of the nuclear power feed water pump is input into the nuclear power feed water pump fault diagnosis model, and the nuclear power feed water pump fault diagnosis model is used to diagnose the feed water pump fault;

[0025] The model update module is configured as follows: The update of the nuclear power feed water pump fault diagnosis model includes:

[0026] Obtain and process the health status data of nuclear power feed water pumps and the operation data of thermal power feed water pumps;

[0027] Extract the data features of thermal power feed water pumps and use them as the source domain, and extract the health features of nuclear power feed water pumps as the target domain;

[0028] According to the distribution similarity of the health features in the source domain and the health features in the target domain, the corresponding thermal power feed water pump fault features are extracted and combined with the health features of the nuclear power feed water pump to obtain samples containing the nuclear power feed water pump fault features;

[0029] The samples are used as training data to update the nuclear power feedwater pump fault diagnosis model.

[0030] Compared with the prior art, the present invention has the following advantages and positive effects:

[0031] (1) To address the problem of low fault diagnosis accuracy of nuclear power feed water pumps, the operation data of the pumps in other scenarios with long-term operation are obtained to extract fault features. The fault features under other working conditions are transferred to the healthy features of the nuclear power feed water pumps as samples by combining transfer learning and deep convolutional networks to update the fault diagnosis model, thereby improving the diagnosis accuracy under the condition of missing fault samples.

[0032] (2) The decoder retains the key data features of the original data, improving the quality of extracted features. BRIEF DESCRIPTION OF THE DRAWINGS

[0033] The accompanying drawings in the specification, which constitute a part of the present invention, are used to provide a further understanding of the present invention. The exemplary embodiments of the present invention and their descriptions are used to explain the present invention and do not constitute improper limitations on the present invention.

[0034] Figure 1 It is a flow chart of the nuclear power feed water pump fault diagnosis method in embodiments 1 and 2 of the present invention.

[0035] Figure 2Schematic diagram of the structure of the feature extraction network in Examples 1 and 2 of the present invention.

[0036] Figure 3 Schematic diagram of the transfer learning diagnostic network in Examples 1 and 2 of the present invention. DETAILED DESCRIPTION

[0037] Example 1

[0038] In a typical embodiment of the present invention, Figure 1-Figure 3 As shown, a method for fault diagnosis of nuclear power feed water pump is given.

[0039] In this embodiment, a nuclear power feed water pump fault diagnosis method based on transfer learning and variational autoencoder is proposed. The method relies on the data acquisition system to obtain equipment status information, and extracts fault features based on the variational autoencoder. The method combines transfer learning and deep convolutional network to realize the fault diagnosis of nuclear power feed water pump, thereby improving the diagnosis accuracy under the condition of missing fault samples.

[0040] See also Figure 1 , the fault diagnosis methods for nuclear power feed water pumps include:

[0041] The operation data of the nuclear power feed water pump is input into the nuclear power feed water pump fault diagnosis model, and the nuclear power feed water pump fault diagnosis model is used to diagnose the feed water pump fault;

[0042] Among them, the update of the nuclear power feed water pump fault diagnosis model includes:

[0043] Obtain and process the health status data of nuclear power feed water pumps and the operation data of thermal power feed water pumps;

[0044] Extract the data features of thermal power feed water pumps and use them as the source domain, and extract the health features of nuclear power feed water pumps as the target domain;

[0045] According to the distribution similarity of the health features in the source domain and the health features in the target domain, the corresponding thermal power feed water pump fault features are extracted and combined with the health features of the nuclear power feed water pump to obtain samples containing the nuclear power feed water pump fault features;

[0046] The samples are used as training data to update the nuclear power feedwater pump fault diagnosis model.

[0047] In this embodiment, a nuclear power feedwater pump fault diagnosis method based on transfer learning and variational autoencoder relies on equipment status data to achieve accurate fault diagnosis of the equipment.

[0048] like Figure 1As shown in the figure, firstly, the health status data of nuclear power feed water pumps are collected through the data acquisition system, and the thermal power feed water pump data and the health status data of nuclear power feed water pumps are denoised; the key features of the time series signal are extracted by the variational autoencoder, and the processed data are divided into the source domain and the target domain; the fault features of the thermal power feed water pumps in the source domain are transferred to the nuclear power feed water pump status in the target domain through transfer learning, and the model parameters are updated in combination with the health data features of the nuclear power feed water pumps in the target domain, and the diagnosis of nuclear power feed water pump faults is realized through the deep convolutional network.

[0049] As a high-speed rotating machine, the water pump's operating status information is mostly manifested in the form of vibration, such as blade deformation, blade weld cracking, blade shedding, bearing fatigue wear, loose bearing fit, bearing damage, coupling misalignment, coupling pin hole wear, uneven coupling quality, and loose foot connection. Therefore, the operating data and fault data in this embodiment include the vibration data of the vibrating water pump.

[0050] Process the health status data of nuclear power feed water pumps and the operation data of thermal power feed water pumps, capture the continuity and change trends between feature data, and retain the key data features of the original data.

[0051] like Figure 2 As shown, the variational autoencoder is used to extract the data of the water pump when processing the data. The time series signal of the thermal power water pump and the healthy time series signal of the nuclear power water pump are subjected to wavelet denoising. The variational autoencoder is composed of an encoder and a decoder.

[0052] Among them, for the wavelet denoising process, the Haar wavelet basis function is used to perform wavelet transform on the timing signal of the thermal power feed water pump and the healthy timing signal of the nuclear power feed water pump, and the signal is decomposed into a series of wavelet coefficients of different scales. These coefficients describe the characteristics of the signal at different frequencies and times, and then the soft threshold function is selected to denoise the wavelet coefficients, and the processed wavelet coefficients are subjected to wavelet inverse transform to obtain the denoised signal.

[0053] The encoder is responsible for mapping the input data to the mean and variance potential representation in the latent space, assuming the latent space to be a Gaussian distribution, and determining the parameters of the Gaussian distribution according to the mean and variance output by the encoder, thereby capturing the continuity and change trend between features.

[0054] In the decoder, the activation layer uses the ReLU activation function. The decoder uses the deconvolution layer and the ReLU activation function to decode the potential representation into the reconstructed features in the original data space, retaining the key data features of the original data and improving the quality of the extracted features.

[0055] In time series data, the continuity of features usually reflects the natural change trend of the data. By capturing the continuity between features, it can be ensured that the smooth transition of the data will not be disrupted during signal processing, thereby avoiding the introduction of unnecessary noise or artifacts. Capturing continuity can ensure that while removing noise, the useful information in the original signal is retained as much as possible.

[0056] The changing trend in the time series signal often reveals the long-term behavior pattern of the data. In the health monitoring of the water supply pump, the changing trend indicates the performance degradation or impending failure of the equipment. Therefore, in this embodiment, these trends are captured to facilitate prediction and early warning.

[0057] The decoder is responsible for decoding the potential representation into reconstructed features in the original data space, retaining the key data features of the original data and improving the quality of the extracted features.

[0058] The encoder assumes that the latent space is Gaussian and uses mean and variance parameters to describe the distribution of data, allowing the points in the latent space to remain continuous and structured, making it easier for the decoder to recover the key features of the original data from the latent representation. The continuity of the latent space allows similar input data to have similar representations in the latent space, which helps the decoder generate more accurate reconstructions.

[0059] like Figure 3 As shown in the figure, before calculating the distribution similarity of health features, the feature extraction module is used to extract the distribution features of the thermal power feed water pump row data and the distribution features of the health status data of the nuclear power feed water pump. The similarity is measured by the difference between the health features in the source domain and the target domain, and the parameters of the feature extraction module are updated through reverse propagation based on the maximum similarity.

[0060] Specifically, in this embodiment, a fault diagnosis network based on transfer learning is proposed to realize fault data migration and fault diagnosis. First, the data features of the thermal power feed water pump are divided into the source domain, and the health features of the nuclear power feed water pump are divided into the target domain. Since the two samples have different data distributions, a convolutional network is used to further extract the distribution features of the two data, thereby reducing the model calculation cost and improving the generalization ability of the model in the target domain.

[0061] MMD (Maximum Mean Discrepancy) is used to measure the difference between two feature distributions. Its expression formula is: Where f is a mapping function on the sample space, X represents the key features extracted from the source domain, Y represents the key features extracted from the target domain, n represents the n feature samples extracted from X, and x i represents the feature samples in X, m represents the m feature samples extracted from Y, y irepresents the feature samples in Y, Represents the norm in the feature space H (Hilbert). The similarity between the two distributions is measured by calculating the difference between the healthy features in their feature space.

[0062] Minimizing MMD continuously updates the parameters of the feature extraction module through reverse propagation, and makes the health feature distributions of the two fields as close as possible through parameter sharing, so that the model can better reflect the distribution of fault features of nuclear power feed water pumps when extracting thermal power fault data.

[0063] The processed features are input into the deep convolutional network, and the data features extracted by the deconvolution layer and the ReLU activation function are combined to generate samples containing the fault characteristics of the nuclear power feed water pump. This solves the problem of low diagnostic accuracy caused by missing and insufficient nuclear power feed water pump fault samples, and the fully connected layer is used to diagnose the fault status of the nuclear power feed water pump.

[0064] The initial operation data of nuclear feed water pumps was used to establish a fault diagnosis model for nuclear feed water pumps. The fault diagnosis model for nuclear feed water pumps was updated with samples. The updated fault diagnosis model for nuclear feed water pumps was tested, and the diagnostic accuracy was calculated and evaluated.

[0065] Finally, the qualified nuclear power feed water pump fault diagnosis model is used to predict the operating status of the nuclear power feed water pump and perform fault diagnosis.

[0066] Example 2

[0067] In another typical embodiment of the present invention, Figure 1-Figure 3 As shown, a nuclear power feed water pump fault diagnosis system is given.

[0068] The nuclear power feed water pump fault diagnosis system includes:

[0069] The diagnosis module is configured as follows: the operation data of the nuclear power feed water pump is input into the nuclear power feed water pump fault diagnosis model, and the nuclear power feed water pump fault diagnosis model is used to diagnose the feed water pump fault;

[0070] The model update module is configured as follows: The update of the nuclear power feed water pump fault diagnosis model includes:

[0071] Obtain and process the health status data of nuclear power feed water pumps and the operation data of thermal power feed water pumps;

[0072] Extract the data features of thermal power feed water pumps and use them as the source domain, and extract the health features of nuclear power feed water pumps as the target domain;

[0073] According to the distribution similarity of the health features in the source domain and the health features in the target domain, the corresponding thermal power feed water pump fault features are extracted and combined with the health features of the nuclear power feed water pump to obtain samples containing the nuclear power feed water pump fault features;

[0074] The samples are used as training data to update the nuclear power feedwater pump fault diagnosis model.

[0075] The working process of the nuclear power feed water pump fault diagnosis system is shown in Example 1 and will not be repeated here.

[0076] The above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. For those skilled in the art, the present invention may have various modifications and variations. Any modification, equivalent replacement, improvement, etc. made within the spirit and principle of the present invention shall be included in the protection scope of the present invention.

Claims

1. A method for diagnosing a nuclear power feedwater pump fault, characterized in that: include: The operation data of the nuclear power feed water pump is input into the nuclear power feed water pump fault diagnosis model, and the nuclear power feed water pump fault diagnosis model diagnoses the feed water pump fault and obtains the fault data. The operation data and the fault data both include the vibration data of the vibration feed water pump. Among them, the update of the nuclear power feed water pump fault diagnosis model includes: The health status data of nuclear power feed water pump and the operation data of thermal power feed water pump are obtained and processed, specifically, the health status data of nuclear power feed water pump and the operation data of thermal power feed water pump are processed, the continuity and change trend between the characteristic data are captured, and the key data features of the original data are retained. When processing the data, the thermal power feed water pump time series signal and the nuclear power feed water pump health time series signal are subjected to wavelet denoising. The capturing of the continuity and change trend between the characteristic data includes: When processing the data, a variational autoencoder feedwater pump data extraction network is used to perform wavelet denoising on the thermal power feedwater pump timing signal and the nuclear power feedwater pump health timing signal. The variational autoencoder is composed of an encoder and a decoder. The encoder maps the input data to the mean and variance potential representation in the latent space. The latent space is a Gaussian distribution. The parameters of the Gaussian distribution are determined according to the mean and variance of the encoder output, thereby capturing the continuity and change trend between features. The key data features of the original data retained include: In the decoder, the activation layer uses the ReLU activation function. The decoder uses the deconvolution layer and the ReLU activation function to decode the potential representation into the reconstructed features in the original data space, retaining the key data features of the original data; Extract the data features of thermal power feed water pumps and use them as the source domain, and extract the health features of nuclear power feed water pumps as the target domain; According to the distribution similarity of the health features in the source domain and the health features in the target domain, the corresponding thermal power feed water pump fault features are extracted and combined with the health features of the nuclear power feed water pump to obtain samples containing the nuclear power feed water pump fault features; the specific processing is as follows: Before calculating the distribution similarity of health features, the feature extraction module is used to extract the distribution features of the thermal power feed water pump row data and the distribution features of the health status data of the nuclear power feed water pump; the similarity is measured by the difference between the health features in the source domain and the target domain, and the parameters of the feature extraction module are updated through reverse propagation based on the maximum similarity; the extracted health features in the source domain and the health features in the target domain are made close by parameter sharing, thereby improving the extracted fault features of the thermal power feed water pump in the source domain to be close to the fault features of the nuclear power feed water pump; Through transfer learning, the fault features of thermal power feedwater pumps extracted from the source domain are transferred to the operating data of nuclear power feedwater pumps in the target domain, and then input into the deep convolutional network. The data features extracted by the deconvolution layer and the ReLU activation function are combined to generate samples containing the fault features of nuclear power feedwater pumps. The samples are used as training data to update the nuclear power feedwater pump fault diagnosis model.

2. The method for diagnosing a nuclear power feedwater pump fault according to claim 1, characterized in that: The initial operation data of nuclear feed water pumps was used to establish a fault diagnosis model for nuclear feed water pumps. The fault diagnosis model for nuclear feed water pumps was updated with samples. The updated fault diagnosis model for nuclear feed water pumps was tested, and the diagnostic accuracy was calculated and evaluated.

3. A nuclear power feed water pump fault diagnosis system, using a nuclear power feed water pump fault diagnosis method as claimed in claim 1, characterized in that: include: The diagnosis module is configured to: input the operation data of the nuclear power feed water pump into the nuclear power feed water pump fault diagnosis model, diagnose the feed water pump fault by the nuclear power feed water pump fault diagnosis model, and obtain fault data, wherein the operation data and the fault data both include the vibration data of the vibration feed water pump; The model update module is configured as follows: The update of the nuclear power feed water pump fault diagnosis model includes: The health status data of nuclear power feed water pump and the operation data of thermal power feed water pump are obtained and processed, specifically, the health status data of nuclear power feed water pump and the operation data of thermal power feed water pump are processed, the continuity and change trend between the characteristic data are captured, and the key data features of the original data are retained. When the above data is processed, the thermal power feed water pump time series signal and the nuclear power feed water pump health time series signal are subjected to wavelet denoising. The continuity and change trend between the characteristic data are captured, including: Map the input data to the mean and variance potential representation in the latent space. The latent space is a Gaussian distribution. The parameters of the Gaussian distribution are determined according to the mean and variance to capture the continuity and change trend between features. The key data features of the original data retained include: Decode the latent representation into the reconstructed features in the original data space, preserving the key data features of the original data; Extract the data features of thermal power feed water pumps and use them as the source domain, and extract the health features of nuclear power feed water pumps as the target domain; According to the distribution similarity of the health features in the source domain and the health features in the target domain, the corresponding thermal power feed water pump fault features are extracted and combined with the health features of the nuclear power feed water pump to obtain samples containing the nuclear power feed water pump fault features; The samples are used as training data to update the nuclear power feedwater pump fault diagnosis model.

Images