An intelligent fault diagnosis method and system based on unbalanced learning

Abstract

This invention provides a fault diagnosis method for nuclear power plant circulating water pumps based on imbalanced learning, comprising: acquiring signals from fault-prone components of the circulating water pump to obtain raw vibration signals; performing data augmentation using a sliding window method to obtain augmented signal samples; extracting time-frequency domain features and generating wavelet time-frequency maps to construct an imbalanced dataset that meets the requirements of real-world operation, and dividing the imbalanced dataset into training and testing sets; building and training a neural network model using the training set until the model meets the usage requirements, then saving the parameters of the neural network model to obtain the trained neural network model; testing the fault diagnosis effect of the trained neural network model and visualizing the results. This invention significantly improves the model's generalization ability on minority class samples and effectively solves the impact of data imbalance on the model's decision boundary.

Description

Technical Field

[0001] This invention relates to the field of intelligent maintenance of nuclear power plant equipment, and in particular to a method and system for diagnosing faults in nuclear power plant circulating water pumps based on unbalanced learning. Background Technology

[0002] Circulating water pumps are crucial auxiliary equipment in the conventional island of a nuclear power plant. A failure in a circulating water pump can reduce condenser cooling capacity, forcing the unit to operate at reduced load or even shut down. While a circulating water pump failure does not directly threaten reactor safety, it can cause unplanned shutdowns of the nuclear power plant, affecting the continuity and stability of power supply and resulting in significant economic losses. Therefore, ensuring the reliable operation of circulating water pumps is of paramount importance for the stable and economical operation of a nuclear power plant.

[0003] Currently, there are two main types of fault diagnosis methods commonly used for mechanical equipment: signal processing-based methods and artificial intelligence-based methods. Signal processing-based fault diagnosis methods typically require expert experience and prior knowledge to select appropriate analysis methods and parameters, lacking adaptability and flexibility. These methods have high computational complexity when processing large-scale data, making it difficult to process massive amounts of data in real time. In addition, signal processing methods have limited performance when dealing with nonlinear and non-stationary signals, making it difficult to capture complex fault modes.

[0004] Artificial intelligence-based fault diagnosis methods, by automatically extracting and learning features and patterns from data, can handle large-scale and high-dimensional data, exhibiting greater robustness and generalization ability. AI methods require minimal expert intervention and can adaptively adjust the model, improving the accuracy and efficiency of fault diagnosis. Therefore, with increasing data volume and signal complexity, AI-based fault diagnosis methods show greater advantages and application potential in the fault diagnosis of modern mechanical equipment. However, these methods rely on high-quality training data. If the training data is unbalanced or contains significant noise interference, it can lead to deviations in the model's decision boundaries, resulting in misdiagnosis and missed diagnosis. Therefore, there is an urgent need in this field for a fault diagnosis method for nuclear power plant circulating water pumps to improve the accuracy of fault diagnosis. Summary of the Invention

[0005] To address the aforementioned technical problems, the objective of this invention is to provide a method and system for fault diagnosis of nuclear power plant circulating water pumps based on imbalanced learning. This method combines traditional signal processing-based fault diagnosis methods with deep learning methods, and uses imbalanced learning techniques to address the uneven distribution of fault and normal samples in the dataset, thereby improving the model's diagnostic accuracy for a smaller number of faulty samples.

[0006] To achieve the above objectives, the present invention adopts the following approach.

[0007] A fault diagnosis method for nuclear power plant circulating water pumps based on imbalanced learning, the method comprising:

[0008] S100: For components in nuclear power plant circulating water pumps that are prone to failure, the system collects signals during their healthy state and when a failure occurs, and obtains the corresponding raw vibration signals.

[0009] S200: The original vibration signal is preprocessed, and data augmentation is performed using a sliding window method to obtain a data-augmented signal sample;

[0010] S300: Perform continuous wavelet transform on the data-enhanced signal samples to extract time-frequency domain features and generate wavelet time-frequency maps. Use Zipf distribution to construct an imbalanced dataset that meets the real-world operating scenario and divide the imbalanced dataset into training and test sets.

[0011] S400: Build and train a neural network model; define the training model. During the training phase of the model In this phase, a marginal regularization algorithm is used to apply stronger regularization to the minority class than to the majority class;

[0012] S500: Defines the training model stage, in In this phase, a cost-sensitive reweighting algorithm is used, employing the concepts of random coverage and effective sample number, and the decision boundary is moved by allocation class penalty to reduce bias caused by data imbalance;

[0013] S600: In In this stage, the influence function is used to reweight the samples based on their impact on the decision boundary, forming a generalized decision boundary on the imbalanced data;

[0014] S700: Train the neural network model using the training set until it meets the usage requirements, then save the parameters of the neural network model to obtain the neural network model that has been initially trained.

[0015] S800: Use the test set to test the fault diagnosis effect of the initially trained neural network model, and visualize the results through confusion matrix and ROC curve. The test meets the usage requirements and the trained neural network model is obtained.

[0016] S900: Uses a trained neural network model to diagnose faults in nuclear power plant circulating water pumps.

[0017] Optionally, in step S100, a triaxial accelerometer is installed at a component of the nuclear power plant's circulating water pump that is prone to failure. After setting the sampling frequency and sampling time, vibration signals of the component are collected under healthy conditions and when different types of failures occur, respectively, to obtain the original vibration signal, i.e., the time domain signal. ,in, This indicates that the signal has not been preprocessed. Indicates the number of collected data. There are 1 signal segments, where t represents time.

[0018] Optionally, in step S200, the time-domain signal is... Perform DC component removal processing on the time-domain signal. The j-th data point The formula for removing the DC component is as follows:

[0019] ,

[0020] in, Indicates the first in the signal The value corresponding to the j-th data point of a signal segment; This represents the summation operation. Indicates signal The CCP Data points, express The value after removing the DC component.

[0021] Optionally, in step S400, by training the model... The stage uses a marginally regularized cross-entropy loss function It applies stronger regularization to the minority class than to the majority class, improving the generalization ability of the minority class without sacrificing the model's ability to fit the majority class.

[0022] Marginally regularized cross-entropy loss function for: .

[0023] in, Represents logarithmic operations. This indicates exponentiation. This represents the summation operation. This represents the model output logit value for category c, which is the raw score before softmax processing. This represents the model output logit value corresponding to the true class y, where y represents the true label class, and c represents iterating through all possible class labels. This represents the marginal parameter for the true category y, used to adjust the regularization strength. c≠y means "category c is not equal to the true category y", that is, c represents all incorrect categories.

[0024] Optionally, in step S500, the number of valid samples... The calculation formula is:

[0025] ,

[0026] in, N represents the number of valid samples, N represents the total number of samples, and n represents the number of samplings or repetitions.

[0027] Optionally, in step S500, a quasi-balance factor is introduced. To balance the loss, it is related to the first The number of valid samples in a class is inversely proportional to its number of samples: .

[0028] right Normalize to make Normalization The formula is:

[0029] ,

[0030] in, The normalized equilibrium factor, The total number of categories in the dataset. Let i be the number of valid samples in the category to which the i-th sample belongs. For the first The number of valid samples in the class This formula represents the summation of the reciprocals of the effective sample counts for all categories. It achieves normalization adjustment for category balance by weighting the reciprocals of the effective sample counts.

[0031] Optionally, in step S600, the impact on balance loss... The formula is:

[0032] ,

[0033] Where x represents the input sample, This represents the model parameters, where m is the training dataset. The total number of samples in the sample, This represents the sample-label pairs in the dataset. Let be the weight coefficient of the i-th sample. Let n be the standard loss function for the nth sample. Let be the prediction function of the model for the nth sample. For the true label of the sample, This represents the L1 norm distance between the model's predicted value for the target sample x and its true label y. This function represents the L1 norm of the influence vector h. It balances the influence of the samples by adjusting the ratio of the weighted loss to the product of the prediction error and the influence vector.

[0034] A computer-readable storage medium for storing a computer program configured to implement the method when invoked by a processor.

[0035] An electronic device includes: a memory, a processor, and a computer program stored in the memory and executable on the processor; wherein the processor implements the method when executing the program.

[0036] Compared with the prior art, the present invention has the following beneficial technical effects:

[0037] The method proposed in this invention reduces the impact of data imbalance from a more comprehensive perspective.

[0038] 1. The method proposed in this invention performs well in confusion matrix and ROC curve evaluation, with high diagnostic accuracy for all categories, low variance, and strong robustness.

[0039] 2. This invention significantly improves the model's generalization ability on minority class samples by using imbalanced learning techniques, especially marginal regularization, cost-sensitive weighting based on random coverage and effective sample number, and weighting based on influence function, effectively solving the impact of data imbalance on the model's decision boundary. Attached Figure Description

[0040] The accompanying drawings illustrate exemplary embodiments of the invention and, together with the description thereof, serve to explain the principles of the invention. These drawings are included to provide a further understanding of the invention and are incorporated in and constitute a part of this specification.

[0041] Figure 1 This is a structural diagram of a nuclear power plant circulating water pump test bench according to one embodiment of the present invention;

[0042] Figure 2 This is a schematic diagram of the sensor arrangement on the pump guide bearing in one embodiment of the present invention;

[0043] Figure 3 This is a schematic diagram of pump guide bearing failure types in one embodiment of the present invention;

[0044] Figure 4 This is a schematic diagram of sliding window data enhancement in one embodiment of the present invention;

[0045] Figure 5This is a wavelet time-frequency diagram obtained by performing continuous wavelet transform on the time-domain signal of the pump guide bearing in one embodiment of the present invention;

[0046] Figure 6 This is a schematic diagram of the overall structure of a deep migration network constructed in one embodiment of the present invention;

[0047] Figure 7 This is a schematic diagram of the confusion matrix and a recall and precision curve diagram of using the method of the present invention and other comparison methods in one embodiment of the present invention;

[0048] Figure 8 This is an overall flowchart of fault diagnosis of a circulating water pump in one embodiment of the present invention. Detailed Implementation

[0049] The following is in conjunction with the appendix Figures 1 to 8 The present invention will be further described in detail below with reference to the embodiments. It is to be understood that the specific embodiments described herein are for illustrative purposes only and are not intended to limit the scope of the invention. Furthermore, it should be noted that, for ease of description, only the parts relevant to the present invention are shown in the accompanying drawings.

[0050] It should be noted that, unless otherwise specified, the embodiments and features described in this invention can be combined with each other. The technical solution of this invention will now be described in detail with reference to the accompanying drawings and embodiments.

[0051] Unless otherwise stated, the exemplary embodiments / exemplifications shown are to be understood as providing exemplary features of various details that provide ways in which the technical concept of the invention can be implemented in practice. Therefore, unless otherwise stated, the features of the various embodiments / exemplifications may be additionally combined, separated, interchanged and / or rearranged without departing from the technical concept of the invention.

[0052] The use of crosshairs and / or shading in the accompanying drawings is generally used to clarify the boundaries between adjacent components. Thus, unless otherwise stated, the presence or absence of crosshairs or shading does not convey or indicate any preference or requirement for the specific material, material properties, dimensions, proportions, commonalities between the illustrated components, or any other characteristics, properties, etc., of the components. Furthermore, in the accompanying drawings, the dimensions and relative dimensions of components may be exaggerated for clarity and / or descriptive purposes. When exemplary embodiments can be implemented differently, a specific process sequence may be performed in a different order than that described. For example, two consecutively described processes may be performed substantially simultaneously or in the reverse order of their description. Furthermore, the same reference numerals denote the same components.

[0053] When a component is referred to as being "on" or "above" another component, "connected to," or "joined to" another component, the component may be directly on, directly connected to, or directly joined to the other component, or there may be intermediate components. However, when a component is referred to as being "directly on" another component, "directly connected to," or "directly joined to" another component, there are no intermediate components. Therefore, the term "connection" can refer to a physical connection, an electrical connection, etc., and may or may not have intermediate components.

[0054] For descriptive purposes, the present invention may use spatial relative terms such as “below,” “under,” “below,” “down,” “above,” “above,” “higher,” and “side (e.g., in a “sidewall”)” to describe the relationship between one component and another component as shown in the accompanying drawings. In addition to the orientations depicted in the drawings, the spatial relative terms are also intended to encompass different orientations of the device during use, operation, and / or manufacture. For example, if the device in the drawings is flipped, a component described as “below” or “under” another component or feature would subsequently be positioned “above” said other component or feature. Thus, the exemplary term “below” can encompass both “above” and “below” orientations. Furthermore, the device may be otherwise positioned (e.g., rotated 90 degrees or in other orientations), thus interpreting the spatial relative descriptive terms used herein accordingly.

[0055] The terminology used herein is for the purpose of describing particular embodiments and is not intended to be limiting. As used herein, unless the context clearly indicates otherwise, the singular forms “a” and “the” are intended to include the plural forms as well. Furthermore, when the terms “comprising” and / or “including” and variations thereof are used in this specification, it indicates the presence of the stated features, integrals, steps, operations, parts, components, and / or groups thereof, but does not exclude the presence or addition of one or more other features, integrals, steps, operations, parts, components, and / or groups thereof. It should also be noted that, as used herein, the terms “substantially,” “about,” and other similar terms are used as approximate terms rather than as terms of degree, thus explaining the inherent biases in measurements, calculated values, and / or provided values ​​that would be recognized by one of ordinary skill in the art.

[0056] In one embodiment, reference is made to Figure 8 This invention provides a method for fault diagnosis of nuclear power plant circulating water pumps based on imbalanced learning, comprising:

[0057] S100: For components in nuclear power plant circulating water pumps that are prone to failure, the system collects signals during their healthy state and when a failure occurs, and obtains the corresponding raw vibration signals.

[0058] S200: The original vibration signal is preprocessed, and data augmentation is performed using a sliding window method to obtain a data-augmented signal sample;

[0059] S300: Perform continuous wavelet transform on the data-enhanced signal samples to extract time-frequency domain features and generate wavelet time-frequency maps. Use Zipf distribution to construct an imbalanced dataset that meets the real-world operating scenario and divide the imbalanced dataset into training and test sets.

[0060] S400: Build and train a neural network model; define the training model. During the training phase of the model In this phase, a marginal regularization algorithm is used to apply stronger regularization to the minority class than to the majority class;

[0061] S500: Defines the training model stage, in In this phase, a cost-sensitive reweighting algorithm is used, employing the concepts of random coverage and effective sample number, and the decision boundary is moved by allocation class penalty to reduce bias caused by data imbalance;

[0062] S600: In In this stage, the influence function is used to reweight the samples based on their impact on the decision boundary, forming a generalized decision boundary on the imbalanced data;

[0063] S700: Train the neural network model using the training set until it meets the usage requirements, then save the parameters of the neural network model to obtain the neural network model that has been initially trained.

[0064] S800: Use the test set to test the fault diagnosis effect of the initially trained neural network model, and visualize the results through confusion matrix and ROC curve to obtain the trained neural network model;

[0065] S900: Uses a trained neural network model to diagnose faults in nuclear power plant circulating water pumps.

[0066] In step S100, a triaxial accelerometer is installed at the components of the circulating water pump prone to failure. After setting the sampling frequency and sampling time, the triaxial accelerometer is used to collect the vibration signals of the components under healthy conditions and when a failure occurs, respectively, to obtain the time-domain signal. ,in, This indicates that the signal has not been preprocessed. Indicates the number of collected data. There are 1 signal segments, where t represents time.

[0067] In step S200, the acquired time-domain signal is processed. DC component removal is performed to eliminate the influence of DC components on subsequent processing; the formula for DC component removal is as follows:

[0068] ,

[0069] in, Indicates the first in the signal The value corresponding to the j-th data point of a signal segment; This represents the summation operation. Indicates signal The CCP Data points, express The value after removing the DC component.

[0070] To increase the number of training samples, a sliding window method is used to process the signal. Perform data augmentation. Utilize length... To sample the sliding window pairs, the sliding window moves a distance of [distance to be filled in] each time. The signal obtained by over-sampling partially overlaps with the subsequent signals. The signal samples obtained by sliding window sampling are represented as follows: , Indicates the th after sliding window oversampling One sample.

[0071] In step S300, the signal samples after overlapping sampling of the sliding window are processed. Continuous wavelet transform is performed to extract the time-frequency features of the signal. The PyWavelets wavelet analysis toolkit in Python is used to analyze the signal samples. A continuous wavelet transform is performed to obtain the wavelet coefficient matrix. The specific process of the continuous wavelet transform is as follows:

[0072] ,

[0073] in, The wavelet coefficient matrix, It is a scale factor. It is the translation factor. Describing wavelet basis functions The complex conjugate function.

[0074] This invention selects the complex Morlet wavelet as the wavelet basis function. The complex Morlet wavelet is a commonly used wavelet basis for continuous wavelet transform, capable of achieving good localization characteristics in both the time and frequency domains. This allows it to accurately locate transient and periodic features in a signal in both time and frequency, and its expression is as follows:

[0075] ,

[0076] in, For the bandwidth of the complex Morlet wavelet, Let be the center frequency of the complex Morlet wavelet, and exp be the exponential operation.

[0077] Actual frequency and scale factor of the signal The following relationship exists:

[0078] ,

[0079] in, Indicates the frequency corresponding to the signal scale. This represents the center frequency of the wavelet basis function. This represents the sampling frequency of the original signal. The wavelet coefficient matrix is ​​obtained using continuous wavelet transform. Then, use the `imshow()` function from Python's Matplotlib plotting library to plot the wavelet coefficient matrix. Generate a wavelet time-frequency plot and save it as a file with a pixel size of [value missing]. Three-channel RGB image. Then use wavelet coefficient matrix. Wavelet time-frequency plots are generated to form a balanced dataset for training the neural network model. The balanced dataset is divided into training and test sets in an 8:2 ratio. Then, using the probability-quality distribution of each sample in the balanced training set generated by the Zipf distribution, samples are drawn from the balanced dataset without replacement. This is represented as... A random variable has parameters. and The Zipf distribution. The parameters are... and The Zipf distribution has a probability mass function:

[0080] ,

[0081] in, Let x be a positive integer, and let x be a random variable with x = 1, 2, ..., n. It is power. , The parameters control the degree of imbalance. The larger the size, the more uneven the distribution.

[0082] In step S400, a neural network model is constructed, which consists of three parts: a backbone network ResNet18, a bottleneck layer, and a classifier.

[0083] In step S500, during the training of the model... In this phase, a marginal regularization algorithm is used to apply stronger regularization to the minority class than to the majority class, thereby improving the generalization ability of the minority class without sacrificing the model's ability to fit the majority class.

[0084] As a simple and intuitive data dependency attribute, we start with the margins of the training samples. For a binary linear classifier, the margins... Defined as the first The minimum distance from the class sample to the decision boundary. The test error of a uniform label distribution is constrained by the following formula:

[0085] .

[0086] By fixing the marginal distance to The model is encouraged to find the optimal trade-off between each margin, and a larger margin is given to the minority class. When The generalization error is minimized when the time is shortest, where For the first The number of samples per class. This indicates that reducing the distance from the decision boundary to the samples belonging to the majority class achieves the best classification results. For category Example index set, This represents the total number of categories in the dataset. For the model... A given sample The margin can be defined as:

[0087] .

[0088] category The trainable margin can be defined as:

[0089] .

[0090] Based on the principle of marginal maximization, the minimum margin for all categories is defined as:

[0091] .

[0092] Then, for multi-class classification problems, an optimized label margin is introduced. ,in For hyperparameters. To simplify the representation, use... The representation model for the th The output is divided into several categories. The marginal loss is then designed as follows to encourage the model to have the aforementioned marginal loss:

[0093] .

[0094] The non-smoothness of the marginal ranking loss poses a challenge to optimization. Therefore, by performing log-exponential smoothing on it, we obtain the following marginally regularized cross-entropy loss function. :

[0095] ,

[0096] in, Represents logarithmic operations. This indicates exponentiation. This represents the summation operation. This represents the model output logit value (the raw score without softmax processing) for category c. This represents the model output logit value corresponding to the true class y, where y represents the true label class, and c represents iterating through all possible class labels. This represents the marginal parameter for the true category y, used to adjust the regularization strength. c≠y means "category c is not equal to the true category y", that is, c represents all incorrect categories.

[0097] Through the training phase Using a marginally regularized cross-entropy loss function It can apply stronger regularization to the minority class than to the majority class, improving the generalization ability of the minority class without sacrificing the model's ability to fit the majority class.

[0098] In step S600, during the training of the model... In this phase, the concepts of random coverage and effective sample number are used to assign weights to samples of different classes to adjust their importance. Due to the highly imbalanced data, directly training the model or reweighting the loss function using the inverse of the sample number is unlikely to achieve ideal results. Intuitively, more data usually makes it easier to train a better-performing model. However, due to information overlap between data, the marginal benefit of the information extracted by the model from the data diminishes as the number of samples increases. Based on this, the theoretical framework of random coverage is introduced to characterize data overlap and to calculate the effective sample number without depending on the specific model and loss function. A class-balanced reweighting term inversely proportional to the effective sample number is added to the loss function. This method associates each sample with a small neighboring region, rather than a single point. As the number of samples increases, newly added samples are more likely to be similar to existing samples. In other words, all samples can be augmented by small neighboring regions of some samples, and these regions are defined as effective samples.

[0099] Valid sample size With the expected volume of the sample The relevant definition is as follows:

[0100] ,

[0101] in, N represents the number of valid samples, N represents the total number of samples, and n represents the number of samplings or repetitions.

[0102] However, it is difficult to find a good set of hyperparameters that is applicable to all categories empirically. Therefore, in practice, we assume Only relevant to the dataset, and set to apply to all samples in the dataset. To balance the losses, a quasi-balance factor was introduced. It is related to the first The number of valid samples in a class is inversely proportional to its number of samples: To ensure the total loss is accounted for in the application At roughly the same scale, for Normalize to make Normalization The formula is:

[0103] ,

[0104] in, The normalized equilibrium factor, The total number of categories in the dataset. Let i be the number of valid samples in the category to which the i-th sample belongs. For the first The number of valid samples in the class This formula represents the summation of the reciprocals of the effective sample counts for all categories. It achieves normalization adjustment for category balance by weighting the reciprocals of the effective sample counts.

[0105] In summary, for a sample size of Category For the loss function Add a weight term that is inversely proportional to the number of valid samples. To balance the impact of sample size.

[0106] Reweighted loss It can be represented as: .

[0107] This invention uses the concepts of random coverage and effective sample number to optimize sample weight allocation, ensuring that the model makes efficient use of each sample while reducing data overlap and information redundancy. This invention uses an influence function to reduce the weight of samples with greater influence, and obtains a smoother decision boundary through fine-tuning.

[0108] In step S700, directly training the neural network model using imbalanced data can lead to overfitting to the class with a larger number of samples. Samples from the larger class infiltrate the sparse minority class samples and dominate the overlapping region near the decision boundary, thus complicating the decision boundary and causing it to shift towards the minority class region. Therefore, a smoother decision boundary can be obtained by reducing the weights of the more influential samples and fine-tuning the model. Thus, in training the model... In this phase, an influence function is used to measure the degree to which each sample influences the complex and biased decision boundary. This function estimates the change in model parameters when a sample is removed without actually removing data and retraining the model. The samples are then reweighted based on their influence, which helps to form a good generalized decision boundary on imbalanced data.

[0109] For input ,definition The parameters are The model output. After... The optimal parameters of the model after the training phase are:

[0110] ,

[0111] Here, def is a definition symbol, meaning "defined as". This indicates that the optimization of parameter δ aims to find the parameter value that minimizes the objective function. It is a loss function that measures the true label. With model output The differences between them. During the training phase, in order to address the imbalance problem, a quasi-balance factor is used in step S400. Reweighted loss. If for a certain point... By slightly modifying the distribution of the training data, the expression can be approximated by changing the parameters. The impact of removing training points. The new parameters are:

[0112] ,

[0113] in This represents the disturbance parameter or influence coefficient. Let the disturbance term be represented, and assume... , exist Nearby, the sample loss during the fine-tuning phase is reweighted using an influence function, defined as follows:

[0114] ,

[0115] in, This is the hidden layer feature vector of the last fully connected layer in the input model. The inverse of the influence function can be used as a reweighting factor to reduce the weight of influential samples during fine-tuning, adjusting the decision boundary and enhancing learning from imbalanced data. Influence-balanced loss. The formula is:

[0116] ,

[0117] in, This refers to the class balance factor in step S600. x represents the input sample. This represents the model parameters, where m is the training dataset. The total number of samples in the sample, This represents the sample-label pairs in the dataset. Let n be the standard loss function for the nth sample. Let be the prediction function of the model for the nth sample. For the true label of the sample, This represents the L1 norm distance between the model's predicted value for the target sample x and its true label y. This function represents the L1 norm of the influence vector h. It balances the influence of the samples by adjusting the ratio of the weighted loss to the product of the prediction error and the influence vector.

[0118] In one embodiment, the present invention provides a method for fault diagnosis of nuclear power plant circulating water pumps based on imbalanced learning, comprising:

[0119] Step S100: On the nuclear power plant circulating water pump test bench, vibration signals are collected from the components of the circulating water pump that are prone to failure. In this embodiment, the pump guide bearing is taken as an example.

[0120] Pump guide bearings play a crucial role in the entire circulating water pump system of a nuclear power plant. Their primary function is to support and stabilize the pump shaft, ensuring proper alignment and balance during operation, while minimizing friction and wear between the shaft and other components. Because nuclear power plant circulating water pumps require continuous operation for extended periods, the performance of the pump guide bearings directly impacts the pump's stability and efficiency. Proper operation of the pump guide bearings guarantees smooth circulation of cooling water, maintaining a safe temperature for the nuclear reactor and ensuring the normal operation of the nuclear power plant. Failure of the pump guide bearings can lead to misalignment or wobble of the pump shaft, increasing friction and wear, and potentially causing mechanical failure of the circulating water pump. For these reasons, monitoring the condition and diagnosing faults of the pump guide bearings are of paramount importance.

[0121] The nuclear power plant circulating water pump test bench used in this embodiment is as follows: Figure 1As shown, the circulating pump test bench consists of a motor, planetary gearbox, impeller, and pump guide bearing. The motor speed is 1480 r / min, and the gearbox output shaft speed is 336 r / min (there may be slight deviations in reality). To ensure the effectiveness of the test bench, the main structure and principle of the planetary gearbox are consistent with the real circulating pump, with only the lubrication station simplified. The sensor arrangement is as follows. Figure 2 As shown, vibration acceleration sensors are installed in different directions (x, y, z) of the pump guide bearing to monitor its vibration characteristics in real time, facilitating the evaluation of equipment vibration in various directions. Sound pressure sensors are installed on top of the pump guide bearing to measure the noise level during equipment operation, thereby analyzing the sound pressure characteristics of the bearing and gears. Displacement sensors are mainly installed in the x and y directions of the pump guide bearing to monitor its displacement and determine its displacement characteristics under different load conditions. The pump guide bearing fault types used in the experiment are as follows: Figure 3 As shown, the fault types include: peeling, pitting, and scratching. The sampling frequency of the data acquisition equipment used in the experiment was 10240Hz, and the number of sampling points was 32768. Signal acquisition was performed on the pump guide bearings with different fault types to obtain the original vibration signals. .

[0122] In step S200, the original vibration signal is processed. DC component removal processing is performed to eliminate the influence of DC components on subsequent processing.

[0123] To increase the number of training samples, use Figure 4 The sliding window method shown applies to signals Perform data augmentation. Utilize length... To sample the sliding window pairs, the sliding window moves a distance of [distance to be filled in] each time. 300 signal samples were obtained for each fault category, for a total of 1200 signal samples. The signal samples obtained through sliding window sampling are represented as follows: , Indicates the th after sliding window oversampling One sample.

[0124] In step S300, the signal samples after overlapping sampling of the sliding window are processed. Continuous wavelet transform is performed to extract the time-frequency features of the signal. The PyWavelets wavelet analysis toolkit in Python is used to analyze the signal samples. A continuous wavelet transform is performed to obtain the wavelet coefficient matrix. In this example, the total scale of the continuous wavelet transform is set to 256, the complex Morlet wavelet is chosen as the wavelet basis function, and the center frequency and bandwidth of the complex Morlet wavelet are set to 5Hz. Then, the `imshow()` function from Python's Matplotlib plotting library is used to plot the wavelet coefficient matrix. Generate a wavelet time-frequency plot and save it as follows: Figure 5 The pixel size shown is Three-channel RGB images. The balanced dataset was divided into training and test sets in an 8:2 ratio. Then... The probability quality distribution of each sample in the balanced training set is generated by sampling without replacement from the balanced dataset to construct an imbalanced dataset. In the imbalanced dataset, there are 225 healthy samples, 129 peeling samples, 30 pitting samples, and 16 scratch samples.

[0125] In step S400, a neural network model is constructed, which consists of a backbone network ResNet18, a bottleneck layer, and a classifier. The specific structure of the network is as follows: Figure 6 As shown.

[0126] The basic unit of ResNet18 is the residual block, which contains two convolutional layers, each followed by batch normalization and a ReLU activation function. Within the residual block, the input data is not only processed by these layers but also directly added to the output via skip connections. This design allows the network to be stacked at greater depths without sacrificing training stability and accuracy. Specifically, ResNet18 consists of eight stacked residual blocks: five regular residual blocks and three residual blocks with downsampling.

[0127] The specific structure is as follows: First, there is a 7x7 convolutional layer with a stride of 2, followed immediately by a 3x3 max-pooling layer. This design effectively reduces the size of the input image. Then, the network enters four sequentially connected residual layers, each containing two residual blocks. Within each residual block, the input passes through two 3x3 convolutional layers, and skip connections directly add the input to the output. This residual learning approach allows ResNet18 to train deeper networks without significantly increasing computational complexity, improving the model's representational power and classification accuracy. Simultaneously, the introduction of residual blocks also gives the network better gradient flow characteristics during training, effectively avoiding the problems of vanishing or exploding gradients.

[0128] The bottleneck layer consists of a weight matrix of... The classification layer consists of fully connected layers and batch normalization (BN) layers. The bottleneck layer is used to reduce feature dimensionality and extract important features. It is commonly used for dimensionality reduction, feature extraction, and improving the computational efficiency of the network. The classification layer consists of a weight matrix... It consists of fully connected layers.

[0129] In step S500, the first 20 epochs are defined as... Training phase. During the model training process... During this phase, the loss function used to train the model is: marginally regularized cross-entropy loss function. , represented as:

[0130] .

[0131] Based on experience, the training phases are divided, and the number of iterations in each phase can be flexibly adjusted according to the actual situation. The purpose of dividing it into two phases is to optimize the decision boundary margin in the first phase, and to reweight in the second phase to improve the model's attention to minority class samples and adjust the decision boundary with finer granularity.

[0132] In step S600, the last 30 epochs are defined as follows: Training phase. First, the class balance factor is calculated and normalized, and then combined with the influence function loss from step S600 as... Loss function during training:

[0133] .

[0134] In step S700, the number of training iterations (epochs) is set to 50, the learning rate is 0.001, and the weight decay is 0.001. Set the size to 32 and use the SGD optimizer with a momentum of 0.9. Convergence should occur within 50 epochs, depending on the actual situation. If the results are unsatisfactory, the number of iterations can be increased. There is no fixed requirement for the number of iterations; it needs to be adjusted flexibly according to the actual situation.

[0135] To prevent overfitting during training, an early stopping method with an interval of 20 iterations was used. Specifically, the early stopping method uses the model's accuracy on the test set as the criterion. After each iteration, the model updates the optimal weight parameters. If the model's accuracy on the test set does not increase for 20 consecutive iterations, the model is considered to have converged, and training is stopped.

[0136] The ResNet18 backbone network is loaded with a pre-trained model from ImageNet. To accelerate model convergence and reduce model oscillations, this embodiment uses learning rate decay. ,in Indicates the current iteration number. The maximum number of iterations, This represents the initial learning rate.

[0137] In step S800, to demonstrate the superiority of the method of the present invention, inverse square root loss (INV), class balanced loss, and label margin regularization loss were selected as comparison methods. To reduce the impact of randomness, all methods used a random number seed. The method was trained five times, and the mean accuracy and variance were reported. Experimental results show that the method of this invention has the highest mean diagnostic accuracy.

[0138]

[0139] The proposed method demonstrates excellent performance in confusion matrix and ROC curve evaluations, achieving diagnostic accuracy exceeding 95% for all categories. Experimental results show that the average accuracy using this method is 97.33 ± 0.87%, significantly higher than other comparative methods. This method achieves the highest average accuracy while exhibiting the lowest variance, indicating its robustness. This is because the proposed method mitigates the impact of imbalanced datasets on the network decision boundary from multiple perspectives: marginal regularization, cost-sensitive weighting based on random coverage and effective sample size, and weighting based on the influence function. These three methods combined enable the model to learn a generalized and comprehensive decision boundary.

[0140] To visually demonstrate the performance of the method proposed in this invention, this embodiment uses confusion matrices and ROC curves to evaluate the fault diagnosis capabilities of different methods. The visualization results are as follows: Figure 7 As shown in the diagram. The confusion matrix results indicate that the method proposed in this invention has the highest fault diagnosis accuracy, exceeding 95% across all categories. In the visualization of the ROC curves, the area under the curves of multiple ROC curves for this invention is close to 1, indicating that the method has a very low false positive rate and a very high true positive rate for each category.

[0141] In one embodiment, the present invention provides a nuclear power plant circulating water pump fault diagnosis system based on imbalanced learning, comprising:

[0142] Data acquisition and preprocessing module: Uses an accelerometer to collect vibration signals from components of the circulating water pump that are prone to failure, and performs DC component removal and sliding window processing on the collected signals, and then converts them into wavelet time-frequency graphs.

[0143] Model training module: Input wavelet time-frequency plot data into the network model and train the model using an imbalanced learning method. When the model accuracy meets the requirements, save the model parameters to prepare for deployment.

[0144] Fault diagnosis module: After the model is deployed to the computer system, the signal to be detected is processed by the data acquisition and preprocessing module and then input into the network model for fault diagnosis, and the fault category is output.

[0145] In one embodiment, the present invention provides a computer device including a controller, an arithmetic logic unit, a memory, an input device, an output device, and a computer program stored in the memory and executable on a processor, wherein the processor executes the program to implement the above-described method.

[0146] In one embodiment, the present invention provides a readable storage medium having a computer program stored thereon, which performs the above-described method when executed by a processor.

[0147] In the description of this specification, the references to terms such as "one embodiment / mode," "some embodiments / modes," "example," "specific example," or "some examples," etc., indicate that a specific feature, structure, material, or characteristic described in connection with that embodiment / mode or example is included in at least one embodiment / mode or example of this application. In this specification, the illustrative expressions of the above terms do not necessarily refer to the same embodiment / mode or example. Moreover, the specific features, structures, materials, or characteristics described may be combined in any suitable manner in one or more embodiments / modes or examples. Furthermore, without contradiction, those skilled in the art can combine and integrate the different embodiments / modes or examples described in this specification, as well as the features of different embodiments / modes or examples.

[0148] Furthermore, the terms "first" and "second" are used for descriptive purposes only and should not be construed as indicating or implying relative importance or implicitly specifying the number of technical features indicated. Thus, a feature defined as "first" or "second" may explicitly or implicitly include at least one of that feature. In the description of this application, "multiple" means at least two, such as two, three, etc., unless otherwise explicitly specified.

[0149] Those skilled in the art should understand that the above embodiments are merely for illustrating the present invention and are not intended to limit the scope of the invention. Those skilled in the art can make other changes or modifications based on the above disclosure, and these changes or modifications still fall within the scope of the present invention.

Claims

1. A fault diagnosis method for nuclear power plant circulating water pumps based on imbalanced learning, characterized in that, The method includes: S100: For components in nuclear power plant circulating water pumps that are prone to failure, the system collects signals during their healthy state and when a failure occurs, and obtains the corresponding raw vibration signals. S200: The original vibration signal is preprocessed, and data augmentation is performed using a sliding window method to obtain a data-augmented signal sample; S300: Perform continuous wavelet transform on the data-enhanced signal samples to extract time-frequency domain features and generate wavelet time-frequency maps. Use Zipf distribution to construct an imbalanced dataset that meets the real-world operating scenario and divide the imbalanced dataset into training and test sets. S400: Build and train a neural network model; define the training model. During the training phase of the model In this phase, a marginal regularization algorithm is used to apply stronger regularization to the minority class than to the majority class; S500: Defines the training model stage, in In this phase, a cost-sensitive reweighting algorithm is used, employing the concepts of random coverage and effective sample number, and the decision boundary is moved by allocation class penalty to reduce bias caused by data imbalance; S600: In In this stage, the influence function is used to reweight the samples based on their impact on the decision boundary, forming a generalized decision boundary on the imbalanced data; S700: Train the neural network model using the training set until it meets the usage requirements, then save the parameters of the neural network model to obtain the neural network model that has been initially trained. S800: Use the test set to test the fault diagnosis effect of the initially trained neural network model, and visualize the results through confusion matrix and ROC curve. The test meets the usage requirements and the trained neural network model is obtained. S900: Uses a trained neural network model to diagnose faults in nuclear power plant circulating water pumps.

2. The method according to claim 1, characterized in that, Preferably, in step S100, a triaxial accelerometer is installed at the components of the nuclear power plant circulating water pump that are prone to failure. After setting the sampling frequency and sampling time, vibration signals of the components are collected under healthy conditions and when different types of failures occur, respectively, to obtain the original vibration signals, i.e., time domain signals. ,in, This indicates that the signal has not been preprocessed. Indicates the number of collected data. There are 1 signal segments, where t represents time.

3. The method according to claim 2, characterized in that, In step S200, the time-domain signal is... Perform DC component removal processing on the time-domain signal. The j-th data point The formula for removing the DC component is as follows: ， in, Indicates the first in the signal The value corresponding to the j-th data point of a signal segment; This represents the summation operation. Indicates signal The CCP Data points, express The value after removing the DC component.

4. The method according to claim 1, characterized in that, In step S400, by training the model... The stage uses a marginally regularized cross-entropy loss function It applies stronger regularization to the minority class than to the majority class, improving the generalization ability of the minority class without sacrificing the model's ability to fit the majority class. Marginally regularized cross-entropy loss function for: , in, Represents logarithmic operations. This indicates exponentiation. This represents the summation operation. This represents the logit value output by the model corresponding to category c. This represents the model output logit value corresponding to the true class y, where y represents the true label class, and c represents iterating through all possible class labels. This represents the marginal parameter for the true category y, used to adjust the regularization strength. c≠y means "category c is not equal to the true category y", that is, c represents all incorrect categories.

5. The method according to claim 1, characterized in that, In step S500, the number of valid samples The calculation formula is: ， in, N represents the number of valid samples, N represents the total number of samples, and n represents the number of samplings or repetitions.

6. The method according to claim 5, characterized in that, In step S500, a quasi-balance factor is introduced. To balance the loss, it is related to the first The number of valid samples in a class is inversely proportional to its number of samples: ； right Normalize to make Normalization The formula is: ， in, The normalized equilibrium factor, The total number of categories in the dataset. Let i be the number of valid samples in the category to which the i-th sample belongs. For the first The number of valid samples in the class This formula represents the summation of the reciprocals of the effective sample counts for all categories. It achieves normalization adjustment for category balance by weighting the reciprocals of the effective sample counts.

7. The method according to claim 1, characterized in that, In step S600, the impact on balance loss The formula is: ， Where x represents the input sample, This represents the model parameters, where m is the training dataset. The total number of samples in the sample, This represents the sample-label pairs in the dataset. Let be the weight coefficient of the i-th sample. Let n be the standard loss function for the nth sample. Let be the prediction function of the model for the nth sample. For the true label of the sample, This represents the L1 norm distance between the model's predicted value for the target sample x and its true label y. This function represents the L1 norm of the influence vector h. It balances the influence of the samples by adjusting the ratio of the weighted loss to the product of the prediction error and the influence vector.

8. A computer-readable storage medium, characterized in that, The computer-readable storage medium is used to store a computer program configured to implement the method of any one of claims 1-7 when invoked by a processor.

9. An electronic device, characterized in that, The electronic device includes: a memory, a processor, and a computer program stored in the memory and executable on the processor; wherein, when the processor executes the program, it implements the method of any one of claims 1-7.

Images