A Fault Diagnosis Method for Rotating Components in Nuclear Power Plants Based on MOCO Siamese Neural Network

Abstract

This invention relates to a fault diagnosis method for rotating components in nuclear power plants based on a MoCo-SINN twin neural network, comprising: acquiring sensor signals from the rotating component; preprocessing the sensor signals to obtain time-frequency features containing rich spatial location information; inputting the time-frequency features into a MoCo-SINN model to obtain input features; training the twin neural network model using a small sample training set and updating the model parameters using a FocalLoss loss function; training the twin neural network model using a small sample training set includes optimizing the process of training the twin neural network model using a dictionary lookup task and momentum encoding; calculating the similarity of the input features; and finally obtaining the fault diagnosis result of the rotating component through a decoder. This invention effectively solves the fault diagnosis problem under small sample conditions through momentum transformation and dictionary lookup.

Description

Technical Field

[0001] This invention relates to the field of fault diagnosis technology for rotating components in nuclear power plants, and in particular to a fault diagnosis method for rotating components in nuclear power plants based on MOCO twin neural networks. Background Technology

[0002] Rotating components of various rotating mechanical equipment in nuclear power units are prone to failure. Monitoring their operating status and promptly identifying and diagnosing faults is of great significance to ensuring the safe and stable operation of nuclear power units.

[0003] Currently, data-driven methods that combine sensor monitoring data with machine learning algorithms are a research hotspot in the field of bearing fault diagnosis. However, they rely heavily on signal processing techniques and expert experience, and it is difficult to efficiently extract feature information from massive amounts of data.

[0004] Deep learning theory can adaptively extract deep features from input signals, effectively solving the problem of efficient mining of a large number of signals. However, these methods require a sufficient number of samples to train the model. During the operation of nuclear power plants, the complex environment often makes it impossible to obtain enough samples of various fault states. Under small sample conditions, the accuracy of these models will be affected.

[0005] For fault diagnosis problems with limited data samples, Siamese neural networks are usually used for fault diagnosis. However, after the encoder is updated, the feature extraction of Siamese networks may change abruptly from the previous batch, which will affect the diagnostic accuracy. Summary of the Invention

[0006] To address the problems existing in the prior art, the present invention aims to provide a fault diagnosis method for rotating components in nuclear power plants based on MOCO twin neural networks. By using momentum transformation and dictionary lookup, the method reduces the problem of abrupt changes in extracted features caused by encoder updates, significantly improves the robustness of the model, and can effectively solve the fault diagnosis problem under small sample conditions.

[0007] To achieve the above objectives, the present invention provides the following solution:

[0008] A fault diagnosis method for rotating components in nuclear power plants based on MOCO twin neural networks includes:

[0009] The sensor signals of the rotating component are acquired, and the sensor signals are preprocessed to obtain time-frequency features containing rich spatial position information;

[0010] The time-frequency features are input into the MoCo-SINN model to obtain the input features; the MoCo-SINN model uses a small sample training set to train a Siamese neural network model and updates the model parameters using the FocalLoss loss function; the small sample training set includes: the original time-frequency plot;

[0011] Training a Siamese neural network model using a few-shot training set includes optimizing the process of training the Siamese neural network model using a dictionary lookup task and momentum encoding.

[0012] The encoder's deep feature extraction capability is achieved by performing similarity measurement calculation on the input features, and finally the fault diagnosis result of the rotating component is obtained through the decoder.

[0013] Optionally, obtaining the time-frequency features includes: performing data reconstruction, zero-mean normalization, and continuous wavelet transform on the sensor signal to extract the time-frequency features.

[0014] Optionally, zero-mean normalization of the sensor signal includes:

[0015]

[0016] Where x is the original data, x * For the normalized data, μ and σ are the mean and variance of x, respectively.

[0017] Optionally, the process of optimizing the training of the Siamese neural network model using a dictionary lookup task on a small training set includes:

[0018] Comparative learning is used as a dictionary lookup task. A dictionary queue is created to store key samples, so that query samples can get more comparisons.

[0019] A sample from a small training set is randomly selected as a query sample. The query sample is encoded using a separate encoder, and the reconstructed query sample data is loaded into the dictionary queue as a key sample for storage. The data reconstruction involves randomly selecting the clipped size of the query sample as the key sample at a ratio between 0.2 and 1.0.

[0020] The key sample is encoded using another encoder, wherein the key sample has a 50% probability of being randomly applied Gaussian blur with a blur radius between 0.1 and 2.0, an 80% probability of being randomly applied color jitter with a range of ±0.4 for brightness, contrast, and saturation, and ±0.1 for hue, and a 20% probability of the sample being randomly converted to a grayscale image.

[0021] During the dictionary queue storage process, the sample with the key input this time is used as a positive sample, and all other samples in the queue are used as negative samples to train the Siamese neural network model. When the queue reaches its storage limit, the earliest sample will be removed from the queue.

[0022] Optionally, the process of optimizing the training of a Siamese neural network model using momentum encoding on a small training set includes:

[0023] The encoder is updated using momentum encoding:

[0024] y t =m·y t-1 +(1-m)·y t

[0025] Among them, y t This represents the encoder's current state, where m represents the momentum update coefficient, the magnitude of which affects the encoder's update speed, and y... t-1 The encoder that represents the previous state.

[0026] Optionally, encoding using the encoder includes: inputting input data into a convolutional layer and residual blocks for feature extraction, and inputting the feature extraction results into a fully connected layer for integration.

[0027] Optionally, the FocalLoss loss function expression is:

[0028] FL(p t )=-α t (1-p t ) γ log(p t )

[0029] Where, p t To predict the similarity probability between two samples, α t γ is the weighting coefficient, and γ is the adjustment parameter.

[0030] Optionally, calculating the similarity measure for the input features includes:

[0031] Cosine similarity is expressed as:

[0032]

[0033] The cosine distance is:

[0034]

[0035] Among them, D w(X1,X2) represents the cosine distance, ||X1|| represents the features encoded from the query sample, and ||X2|| represents the features encoded from the key sample. Finally, the Sigmoid function is used to output the relative distance between 0 and 1. The closer to 0, the more similar the samples are, and vice versa.

[0036] The mathematical form of the Sigmoid function is:

[0037]

[0038] Where x is the relative distance.

[0039] Optionally, obtaining the fault diagnosis results includes:

[0040] The query encoder trained by the Siamese neural network is used to extract features from the samples, and the obtained deep feature information is input into the decoder to complete the fault diagnosis. The decoder is implemented using the Softmax method.

[0041] The mathematical form of the Softmax function is:

[0042]

[0043] Where x is the output vector of the fully connected layer, i is the index of the corresponding element of the fully connected layer's output vector, and N is the number of elements in the fully connected layer's output vector.

[0044] The beneficial effects of this invention are as follows:

[0045] This invention employs a MOCO-based learning strategy, creates a key queue, compares samples using dictionary lookups, and slowly updates the key encoder using momentum encoding. This reduces the problem of abrupt changes in extracted features caused by encoder updates, significantly improving the model's robustness. It can effectively solve the fault diagnosis problem with small samples, achieving an accuracy of 75% even with 50 samples.

[0046] To address the problem of large measurement errors, this invention uses cosine distance to measure the similarity between the features of the key and the query, which can better distinguish the differences between the key and the query and improve the accuracy of the model.

[0047] To address the issue that the output of difficult-to-distinguish sample pairs hovers around 0.5 in the measurement method, this invention uses the FocalLoss loss function to reduce the relative weight of the easy-to-distinguish part, enhance the focus on difficult sample pairs, better train on difficult-to-distinguish sample pairs, accelerate the calculation process, and improve the accuracy by about 5% compared with traditional methods. Attached Figure Description

[0048] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the drawings used in the embodiments will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0049] Figure 1 This is a flowchart of a fault diagnosis method for rotating components in a nuclear power plant based on a MOCO twin neural network, according to an embodiment of the present invention.

[0050] Figure 2 This is a schematic diagram of the structure of a Siamese neural network according to an embodiment of the present invention;

[0051] Figure 3 This is a structural diagram of the MoCo-SINN model according to an embodiment of the present invention;

[0052] Figure 4 This is a schematic diagram of the encoder principle according to an embodiment of the present invention;

[0053] Figure 5 This is a schematic diagram of the decoder principle according to an embodiment of the present invention;

[0054] Figure 6 This is a schematic diagram illustrating the accuracy of the MoCo-SINN model in an embodiment of the present invention;

[0055] Figure 7 This is a t-SNE visualization image of the unextracted sample features in an embodiment of the present invention;

[0056] Figure 8 This is a t-SNE visualization image of the sample features extracted by the encoder in an embodiment of the present invention. Detailed Implementation

[0057] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0058] To make the above-mentioned objects, features and advantages of the present invention more apparent and understandable, the present invention will be further described in detail below with reference to the accompanying drawings and specific embodiments.

[0059] like Figure 1As shown, this embodiment discloses a fault diagnosis method for rotating components in nuclear power plants based on a MoCo-SINN neural network, including: acquiring sensor signals from the rotating component; preprocessing the sensor signals to obtain time-frequency features containing rich spatial location information; inputting the time-frequency features into a MoCo-SINN model to obtain input features; training the MoCo-SINN model using a small sample training set and updating the model parameters using a FocalLoss loss function; the small sample training set includes: the original time-frequency graph; training the Siamese neural network model using the small sample training set includes: optimizing the process of training the Siamese neural network model using a dictionary lookup task and momentum encoding; performing similarity measurement calculation on the input features to achieve the deep feature extraction capability of the encoder; and finally obtaining the fault diagnosis result of the rotating component through a decoder.

[0060] Specifically, this invention provides a method for fault diagnosis of rotating components in nuclear power plants based on MOCO twin neural networks, comprising the following steps: data preprocessing: performing wavelet transform on the sensor signals of the rotating components to denoise and extract time-frequency features; MOCO-based twin neural network model construction: encoding training samples using the twin neural network method and updating the encoder of the twin network using the momentum comparison method; distance measurement: using the cosine distance strategy to measure the similarity of the encoded features of the samples to complete the fault diagnosis of rotating components in nuclear power plants.

[0061] Furthermore, obtaining time-frequency features includes: performing zero-mean normalization and continuous wavelet transform on the sensor signal to extract time-frequency features.

[0062] Furthermore, zero-mean normalization of the sensor signal includes:

[0063]

[0064] Where x is the original data, x * For the normalized data, μ and σ are the mean and variance of x, respectively.

[0065] As a preferred embodiment of the present invention, in the wavelet transform, the cmor wavelet is selected as the wavelet basis function of the continuous wavelet transform, and the window size parameter is an array from 1, 1 / 2, ..., 1 / 127. Finally, a time-frequency diagram with a size of 228×228 is obtained after continuous wavelet transform.

[0066] Furthermore, contrastive learning is used as a dictionary lookup task. A dictionary queue is created to store key samples, allowing query samples to obtain more contrast. Samples from a small training set are randomly selected as query samples. Each query sample is encoded using a separate encoder, and the reconstructed query sample data is loaded into the dictionary queue as a key sample for storage. Data reconstruction involves randomly selecting the cropped size of the query sample as the key sample at a ratio between 0.2 and 1.0. The key sample is then encoded using another encoder, with a 50% probability of randomly applying Gaussian blur (blur radius between 0.1 and 2.0), an 80% probability of randomly applying color jitter (range: ±0.4 for brightness, contrast, and saturation, ±0.1 for hue), and a 20% probability of randomly converting the image to grayscale. During dictionary queue storage, samples with the current input key sample are treated as positive samples, and all other samples in the queue are treated as negative samples. This is used to train the Siamese neural network model. When the queue reaches its storage limit, the oldest sample is removed from the queue.

[0067] Furthermore, the process of optimizing the training of the Siamese neural network model using momentum encoding on a small training set includes: updating the encoder using momentum encoding.

[0068] y t =m·y t-1 +(1-m)·y t

[0069] Among them, y t This represents the encoder's current state, where m represents the momentum update coefficient, the magnitude of which affects the encoder's update speed, and y... t-1 The encoder that represents the previous state.

[0070] Furthermore, encoding using an encoder includes: inputting the input data into convolutional layers and residual blocks for feature extraction, and inputting the feature extraction results into fully connected layers for integration.

[0071] Furthermore, the FocalLoss loss function is expressed as follows:

[0072] FL(p t )=-α t (1-p t ) γ log(p t )

[0073] Where, p t To predict the similarity probability between two samples, α t γ is the weighting coefficient, and γ is the adjustment parameter.

[0074] Furthermore, the similarity measurement calculation for the input features includes:

[0075] Cosine similarity is expressed as:

[0076]

[0077] The cosine distance is:

[0078]

[0079] Among them, D w (X1,X2) represents the cosine distance, ||X1|| represents the features encoded from the query sample, and ||X2|| represents the features encoded from the key sample. Finally, the Sigmoid function is used to output the relative distance between 0 and 1. The closer to 0, the more similar the samples are, and vice versa.

[0080] The mathematical form of the Sigmoid function is:

[0081]

[0082] Where x is the relative distance.

[0083] Furthermore, obtaining fault diagnosis results includes:

[0084] A query encoder trained using a Siamese neural network extracts features from samples. The acquired deep feature information is then input into a decoder to complete fault diagnosis. The decoder is implemented using the Softmax method. Figure 5 As shown.

[0085] The mathematical form of the Softmax function is:

[0086]

[0087] Where x is the output vector of the fully connected layer, i is the index of the corresponding element of the fully connected layer's output vector, and N is the number of elements in the fully connected layer's output vector.

[0088] As a preferred embodiment of the present invention, in the Siamese neural network model based on MOCO, the Siamese neural network requires paired samples (x1, x2, y) as input, where y is the label value. The encoder extracts features from the data, measures the extracted features, obtains the similarity of the sample pairs, and finally makes a judgment based on the similarity. After each batch of inputs is completed, the encoder is updated. The two encoders of the Siamese neural network have the same structure and equal weights.

[0089] As a preferred embodiment of the present invention, the improved method of the Siamese neural network model based on MOCO is as follows: Treating contrastive learning as a dictionary lookup task, creating a dictionary queue, and randomly selecting a sample... Using a standalone encoder for the query, the sample The data is loaded into a queue for storage and used as a key in another encoder. Positive samples in the queue represent the current input. All others are The negative samples are used to increase the number of training iterations for small samples; momentum encoding is used. t =m·y t-1 +(1-m)·y t , where y t This represents the encoder's current state, where m represents the momentum update coefficient, the magnitude of which affects the encoder's update speed, and y... t-1 Represents the encoder of the previous state; ensures that the features of all keys in the dictionary are extracted by similar encoders, and updates the encoder slowly while maintaining feature consistency as much as possible.

[0090] As a preferred embodiment of the present invention, the encoder uses the ResNet-152 model. The data is first passed through a convolutional layer with a kernel size of 7x7, a span of 2x2, and a pooling of 3x3. Then, the four residual blocks of the ResNet model are used to extract features from the graphic data. The first and fourth layers use 3 bottleneck modules, the second layer uses 4 bottleneck blocks, the third layer uses 23 bottleneck modules, and finally, a fully connected layer is used to integrate the feature data.

[0091] As a preferred embodiment of the present invention, the loss function of the Siamese neural network model based on MOCO is FocalLoss, i.e., FL(p t )=-α t (1-p t ) γ log(p t This allows the model to focus on optimizing difficult-to-distinguish sample pairs with differences hovering around 0.5.

[0092] As a preferred embodiment of the present invention, in the distance metric method, cosine similarity can be expressed as: The cosine distance is then expressed as: The smaller the angle between two vectors (the closer it is to 0), the smaller the cosine distance, and the closer their feature distances are, meaning they are more similar; conversely, the larger the angle, the less similar they are.

[0093] Many specific details are set forth in the following description in order to provide a full understanding of the invention. However, the invention may also be practiced in other ways different from those described herein, and those skilled in the art can make similar extensions without departing from the spirit of the invention. Therefore, the invention is not limited to the specific embodiments disclosed below.

[0094] Secondly, the term "one embodiment" or "embodiment" as used herein refers to a specific feature, structure, or characteristic that may be included in at least one implementation of the present invention. The phrase "in one embodiment" appearing in different places in this specification does not necessarily refer to the same embodiment, nor is it a single or selective embodiment that is mutually exclusive with other embodiments.

[0095] This invention employs a fault diagnosis method for rotating components in nuclear power plants based on MOCO twin neural networks, the process of which is as follows: Figure 1 As shown, the diagnostic method consists of three parts: data preprocessing, a MoCo-SINN based Siamese neural network model, and a distance metric.

[0096] Example 1

[0097] Figure 2 The diagram shows the structure of the Siamese neural network. To verify the effectiveness of the MoCo-SINN model in fault diagnosis, the Mechanical Fault Prevention Technology Society (MFPT) rotating component dataset was selected for testing. The test bearing model in the MFPT rotating component dataset is NICE. Single-point faults were set on the inner and outer rings of the rotating component, and data measurements were taken using different loads. Seven states of the dataset were selected, as shown in Table 1. Table 1 provides a detailed description of the MFPT rotating component.

[0098] Table 1

[0099]

[0100] Example 2

[0101] As a preferred embodiment of the present invention, data preprocessing includes data reconstruction, data normalization, and wavelet transform.

[0102] As a preferred embodiment of the present invention, data reconstruction in data preprocessing: Each sample contains a one-dimensional time series signal of length 1024, and each sample is converted into a two-dimensional feature map with a length and width of 64 using wavelet transform. Sample. Then, a sample is randomly selected at a ratio between 0.2 and 1.0. The cutting dimensions are used as a sample, where There is a 50% probability that a Gaussian blur will be randomly applied to the sample, with a blur radius between 0.1 and 2.0. There is an 80% probability that color dithering will be randomly applied, with the color dithering range being ±0.4 for brightness, contrast, and saturation, and ±0.1 for hue. There is a 20% probability that the sample will be randomly converted to a grayscale image.

[0103] As a preferred embodiment of the present invention, data normalization in data preprocessing involves: to eliminate the effects of gradient vanishing or gradient exploding during fault diagnosis, zero-mean normalization is performed on the signal data acquired by the rotating component equipment, i.e.:

[0104]

[0105] In the above formula, x represents the original data; x * The data is normalized, and μ and σ are the mean and variance of x, respectively. 70% of the samples are selected as the training set, and the remainder as the test dataset.

[0106] In a preferred embodiment of this invention, wavelet transform in data preprocessing can be used to address spatial location differences in faults caused by varying signal periods. Therefore, continuous wavelet transform (CWT) is introduced to convert sample data into a time-frequency graph containing rich spatial location information, thus preserving most of the fault's characteristic information. Simultaneously, CWT can effectively remove noise during data acquisition. By convolving the original data with wavelet basis functions, the wavelet transform coefficients of the signal are obtained, which measure the degree of scale matching between the signal and the wavelet basis functions. Convolution at different window scales yields local feature information of the signal within different frequency ranges, thereby achieving time-frequency feature extraction of the signal.

[0107] As a preferred embodiment of the present invention, the parameters selected in the wavelet transform during data preprocessing are as follows:

[0108] The cmor wavelet is selected as the wavelet basis function for continuous wavelet transform; the window size parameter is an array from 1, 1 / 2, ..., 1 / 127, and finally a time-frequency plot with a size of 228×228 is obtained after continuous wavelet transform.

[0109] As a preferred embodiment of the present invention, in the Siamese neural network model based on MOCO, to address the problem that the transition between samples before and after encoder updates during feature extraction can affect diagnostic accuracy, and to solve the problem that encoder parameters change after each backpropagation, resulting in abrupt changes in the sample features extracted by the encoder between two iterations, the present invention uses the MOCO method to improve the Siamese neural network, such as... Figure 3As shown. The MOCO method treats contrastive learning as a dictionary lookup task, creating a dictionary queue, randomly selecting a sample as the query, and using a single encoder. The data is loaded into a queue for storage and used as a key in another encoder. Positive samples in the queue represent the current input. All others are The negative samples are used to increase the number of training iterations for small samples. Each time a positive or negative sample is determined, it can be viewed as searching for a key in the dictionary queue that matches the query. If the key in the current batch obtains features from the current encoder, and subsequent keys obtain features from the updated encoder, the consistency of keys in the dictionary cannot be guaranteed. Therefore, the momentum encoding method in Equation 2 is used.

[0110] y t =m·y t-1 +(1-m)·y t

[0111] In the formula, y t This represents the encoder's current state, where m represents the momentum update coefficient, the magnitude of which affects the encoder's update speed, and y... t-1 This represents the encoder of the previous state. It ensures that the features of all keys in the dictionary are extracted by similar encoders, maintaining feature consistency as much as possible while slowly updating the encoder.

[0112] As a preferred embodiment of the present invention, the encoder in the MOCO-based Siamese neural network model uses the ResNet-152 model to prevent the gradient vanishing problem caused by excessive network depth. Data is first passed through a convolutional layer with a kernel size of 7x7, a stride of 2x2, and padding of 3x3. Then, a four-layer residual block architecture of the ResNet model is used to extract features from the image data. The first and fourth layers use three bottleneck modules, the second layer uses four bottleneck modules, and the third layer uses 23 bottleneck modules. Finally, a fully connected layer is used to integrate the feature data. Figure 4 As shown.

[0113] As a preferred embodiment of the present invention, the loss function of the model is selected as follows: The SINN model outputs a binary classification, and the training objective is to make the difference between samples of the same class as close to 0 as possible, and the difference between samples of different classes as close to 1 as possible. Therefore, FocalLoss is chosen as the model's loss function, allowing the model to focus on optimizing difficult-to-distinguish sample pairs with differences hovering around 0.5.

[0114] FL(p t )=-α t (1-p t ) γlog(p t )

[0115] In the formula, p t α represents the probability that the model predicts the similarity between two samples. t γ is a weighting coefficient used to balance the contributions between different classes. γ is an adjustment parameter used to reduce the relative weight of simple samples and enhance attention to difficult samples, accelerating the computation process and increasing model accuracy. After determining the loss function, the MoCo-SINN encoder is trained using the gradient descent algorithm to achieve backpropagation of the error.

[0116] As a preferred embodiment of this invention, the MoCo-SINN model uses distance metrics to determine whether samples belong to the same category. Commonly used methods include Euclidean distance, cosine similarity, and Manhattan distance. This invention selects cosine similarity to determine the differences between samples because cosine similarity measures the differences between samples by calculating the cosine of the angle between two vectors. It focuses more on the difference in direction between the two vectors rather than distance or length, and can better distinguish the differences in output features. Therefore, using cosine similarity to determine the differences between samples can better measure the similarity of output features and improve the accuracy of classification. Cosine similarity can be expressed as:

[0117]

[0118] The cosine distance is then expressed as:

[0119] D w (X1,X2)=1-cos(X1,X2)

[0120]

[0121] Therefore, the smaller the angle between two vectors (the closer it is to 0), the smaller the cosine distance, and the closer their feature distances are, meaning they are more similar; conversely, the larger the angle, the less similar they are.

[0122] Example 3

[0123] First, the number of iterations and batch size were set to 100 and 32 respectively. The optimizer was SGD (Stochastic Gradient Descent), and the learning rate was dynamically adjusted during training. The MoCo-SINN model was pre-trained using the training dataset. Then, deep features were extracted from the training dataset samples using the pre-trained query encoder, and these deep features were then used to train the model through fully connected layers. After training, the MoCo-SINN model was evaluated using test set data. To reduce random effects, the model was tested 10 times. The results on the test set are as follows: Figure 6As shown, the highest accuracy rate in multiple tests was 98.45%, and the average accuracy rate was 98.21%, indicating that the MoCo-SINN model has stable diagnostic performance and can accurately identify the state type of rotating parts.

[0124] To more vividly demonstrate the feature extraction capabilities of the MoCo-SINN model, the t-distributed random neighborhood embedding (t-SNE) method is used for dimensionality reduction, thus visualizing the sample features. Figure 7 It is a t-SNE image without feature extraction from the query encoding layer, indistinguishable between categories and sparsely distributed within the same category. Figure 8 The image shows the result after feature extraction via the query encoding layer. It can be observed that the relative distance between categories increases, indicating a significant improvement in classification performance, and the distribution of samples within the same category becomes more focused. This further demonstrates that the constructed model possesses excellent feature extraction capabilities.

[0125] The embodiments described above are merely preferred embodiments of the present invention and are not intended to limit the scope of the present invention. Various modifications and improvements made to the technical solutions of the present invention by those skilled in the art without departing from the spirit of the present invention should fall within the protection scope defined by the claims of the present invention.

Claims

1. A fault diagnosis method for rotating components in nuclear power plants based on MOCO twin neural networks, characterized in that, include: The sensor signals of the rotating component are acquired, and the sensor signals are preprocessed to obtain time-frequency features containing rich spatial position information; The time-frequency features are input into the MoCo-SINN model to obtain the input features; the MoCo-SINN model uses a small sample training set to train the Siamese neural network model and combines the FocalLoss loss function to update the model parameters. The small sample training set includes: the original time-frequency plot; Training a Siamese neural network model using a few-shot training set includes optimizing the process of training the Siamese neural network model using a dictionary lookup task and momentum encoding. The process of optimizing the training of a Siamese neural network model using a dictionary lookup task on a small training set includes: Comparative learning is used as a dictionary lookup task. A dictionary queue is created to store key samples, so that query samples can get more comparisons. A sample from a small training set is randomly selected as a query sample. The query sample is encoded using a separate encoder, and the reconstructed query sample data is loaded into the dictionary queue as a key sample for storage. The data reconstruction involves randomly selecting the clipped size of the query sample as the key sample at a ratio between 0.2 and 1.

0. The key sample is encoded using another encoder, wherein the key sample has a 50% probability of being randomly coated with Gaussian blur with a blur radius between 0.1 and 2.0, an 80% probability of being randomly coated with color jitter with a range of ±0.4 for brightness, contrast, and saturation, and ±0.1 for hue, and a 20% probability of the sample being randomly converted to a grayscale image. During the dictionary queue storage process, the sample with the key input this time is used as a positive sample, and all other samples in the queue are used as negative samples to train the Siamese neural network model. When the queue reaches its storage limit, the earliest sample will be removed from the queue. The process of optimizing a twin neural network model using momentum encoding on a small training set includes: The encoder is updated using momentum encoding: in, The encoder representing the current state. This represents the momentum update coefficient, the magnitude of which affects the encoder update speed. The encoder representing the previous state; Encoding using the encoder includes: inputting input data into a convolutional layer and residual blocks for feature extraction, and inputting the feature extraction results into a fully connected layer for integration; The encoder's deep feature extraction capability is achieved by performing similarity measurement calculation on the input features, and finally the fault diagnosis result of the rotating component is obtained through the decoder.

2. The method for fault diagnosis of rotating components in nuclear power plants based on MOCO twin neural networks according to claim 1, characterized in that, Obtaining the time-frequency features includes: performing data reconstruction, zero-mean normalization, and continuous wavelet transform on the sensor signal to extract the time-frequency features.

3. The method for fault diagnosis of rotating components in nuclear power plants based on MOCO twin neural networks according to claim 2, characterized in that, The zero-mean normalization process for the sensor signal includes: in, The original data, For the normalized data, , They are respectively The mean and variance of.

4. The method for fault diagnosis of rotating components in nuclear power plants based on MOCO twin neural networks according to claim 1, characterized in that, The FocalLoss loss function expression is as follows: in, To predict the similarity probability between two samples for the model. These are the weighting coefficients. To adjust the parameters.

5. The method for fault diagnosis of rotating components in nuclear power plants based on MOCO twin neural networks according to claim 1, characterized in that, The similarity measurement calculation for the input features includes: Cosine similarity is expressed as: The cosine distance is: in, Cosine distance Features encoded from query samples The features encoded for the key samples are then used to output the relative distance between 0 and 1 using the Sigmoid function. The closer the distance is to 0, the more similar the samples are, and vice versa. The mathematical form of the Sigmoid function is: Where x is the relative distance.

6. The method for fault diagnosis of rotating components in nuclear power plants based on MOCO twin neural networks according to claim 1, characterized in that, Obtaining the fault diagnosis results includes: The query encoder trained by the Siamese neural network is used to extract features from the samples, and the obtained deep feature information is input into the decoder to complete the fault diagnosis. The decoder is implemented using the Softmax method. The mathematical form of the Softmax function is: Where x is the output vector of the fully connected layer, i is the index of the corresponding element of the fully connected layer's output vector, and N is the number of elements in the fully connected layer's output vector.

Images