A method and system for extracting weak rubbing noise from loose components inside a reactor based on signal processing and machine learning

Abstract

This invention relates to the field of nuclear reactor fault detection technology, specifically to a method and system for extracting weak impact noise from loose components within a reactor based on signal processing and machine learning. The method includes: acquiring a reactor vibration signal; decomposing the vibration signal into multiple sub-signals using empirical mode decomposition; performing Hilbert transform on each sub-signal to extract its instantaneous frequency and amplitude, and constructing a two-dimensional time-frequency matrix for each sub-signal; superimposing the two-dimensional time-frequency matrices along the sub-signal dimensions to generate a three-dimensional instantaneous time-frequency matrix; and inputting the three-dimensional instantaneous time-frequency matrix into a pre-trained convolutional neural network model to suppress reactor operating noise and extract the denoised weak impact noise signal. The purpose is to extract the weak impact noise signal from loose components within the reactor, thereby achieving the identification of impact noise faults in loose components within the reactor.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of nuclear reactor fault detection, and in particular to a reactor in-vessel loose part weak impact and friction noise extraction method and system based on signal processing and machine learning. BACKGROUND

[0002] During long-term operation of the reactor, internal connecting parts are prone to fatigue deterioration due to the coupling effect of thermal stress, mechanical vibration and irradiation effect, etc., and loose parts are formed. The scattered loose parts will move in the reactor under the driving of the coolant and collide and rub with the structure of the in-vessel component, exciting transient mechanical vibration response; due to the small mass and low collision energy of the loose parts, the impact and friction signals generated have the characteristics of weak amplitude, dispersed frequency band and non-stationary. When the signal is transmitted to the outer wall of the pressure vessel through the complex structure in the reactor, it is affected by the superposition of structural damping attenuation and background noise, and the signal-to-noise ratio is significantly reduced, making it difficult for traditional vibration monitoring methods to effectively extract and identify the weak impact and friction signals of the in-vessel loose parts. SUMMARY

[0003] In order to extract the weak impact and friction noise signals of the in-vessel loose parts of the reactor and realize the impact and friction fault identification of the in-vessel loose parts, the present application provides a reactor in-vessel loose part weak impact and friction noise extraction method and system based on signal processing and machine learning, and the technical scheme adopted is as follows:

[0004] The technical scheme of the first aspect of the present application provides a reactor in-vessel loose part weak impact and friction noise extraction method based on signal processing and machine learning, which comprises:

[0005] Obtaining a reactor vibration signal, the vibration signal is decomposed into a plurality of sub-signals by empirical mode decomposition;

[0006] Performing Hilbert transform processing on each sub-signal, extracting the instantaneous frequency and instantaneous amplitude of each sub-signal, and constructing a time-frequency two-dimensional matrix of each sub-signal;

[0007] Stacking the time-frequency two-dimensional matrix of each sub-signal along the sub-signal dimension to generate a three-dimensional instantaneous time-frequency matrix;

[0008] Inputting the three-dimensional instantaneous time-frequency matrix into a pre-trained convolutional neural network model to suppress reactor operating noise and extract weak impact and friction noise signals after noise reduction.

[0009] Further, obtaining a reactor vibration signal, the vibration signal is decomposed into a plurality of sub-signals by empirical mode decomposition, comprising:

[0010] The vibration signal is sequentially decomposed into a plurality of intrinsic modal functions and a residual signal, wherein each intrinsic modal function satisfies a matching condition of a local extreme point number and a zero-crossing point number through iterative screening, and the residual signal is a low-frequency component after decomposition.

[0011] Further, Hilbert transform processing is performed on each sub-signal to extract the instantaneous frequency and instantaneous amplitude of each sub-signal, including:

[0012] An analytic signal is constructed for each sub-signal, and the analytic signal includes the sub-signal and the imaginary part of the Hilbert transform of the sub-signal;

[0013] The instantaneous amplitude at each time point is calculated according to the modulus value of the analytic signal;

[0014] The instantaneous frequency at each time point is calculated according to the phase change rate of the analytic signal.

[0015] Further, a time-frequency two-dimensional matrix of each sub-signal is constructed, including:

[0016] The instantaneous amplitude corresponding to each time point is filled into the matrix element with time sequence as column dimension and instantaneous frequency as row dimension, to generate a time-frequency two-dimensional matrix of time-frequency energy distribution, which can be represented as:

[0017]

[0018] In the formula, represents the time-frequency two-dimensional matrix; represents the instantaneous amplitude at frequency and time ; represents the first frequency point; represents the first time point.

[0019] Further, the time-frequency two-dimensional matrices of the sub-signals are stacked along the sub-signal dimension to generate a three-dimensional instantaneous time-frequency matrix, including:

[0020] The time-frequency two-dimensional matrices of the sub-signals are subjected to interpolation processing in the frequency and time dimensions;

[0021] The interpolated time-frequency two-dimensional matrices are sequentially stacked along the sub-signal dimension to generate a three-dimensional instantaneous time-frequency matrix including the sub-signals, frequencies, and times;

[0022] The three-dimensional instantaneous time-frequency matrix is subjected to smoothing filtering to remove abnormal data points introduced by interpolation.

[0023] Further, the three-dimensional instantaneous time-frequency matrix is input into a pre-trained convolutional neural network model to suppress reactor operation noise and extract a weak rub noise signal after noise reduction, including:

[0024] extracting three-dimensional local features by sliding a three-dimensional convolution kernel in the three-dimensional instantaneous time-frequency matrix;

[0025] down-sampling the three-dimensional local feature matrix output by the three-dimensional convolution kernel using a max-pooling layer;

[0026] mapping the down-sampled features to a one-dimensional time series signal through a fully connected layer, and outputting a denoised weak impact-rubbing noise signal.

[0027] Further, the training process of the convolutional neural network model comprises the following sub-steps:

[0028] constructing a training data set, including a three-dimensional instantaneous time-frequency matrix with noise and a corresponding noise-free target signal;

[0029] quantifying the difference between the model output and the target signal through a loss function;

[0030] updating the convolution kernel weights and the fully connected layer parameters using a backpropagation algorithm until the convolutional neural network model converges.

[0031] Further, the loss function is a weighted mean square error function, and the weight of the loss function is dynamically determined by the energy distribution of the intrinsic mode function of the target signal. The expression of the loss function is:

[0032]

[0033] In the formula, represents the loss function value; represents the total number of time points of the signal; represents the total number of intrinsic mode functions obtained after empirical mode decomposition; represents the square value of the th intrinsic mode function of the target signal at time point ; represents the value of the denoised signal output by the model at time point ; represents the true value of the noise-free target signal at time point .

[0034] Further, the method further comprises:

[0035] reconstructing the time-domain waveform of the denoised weak impact-rubbing noise signal to generate a one-dimensional output signal that is time-aligned with the original vibration signal, for identifying the impact-rubbing event of the loosening part.

[0036] The technical scheme of the second aspect of the present application provides a reactor in-core loose part weak impact and grinding noise extraction system based on signal processing and machine learning, which adopts the reactor in-core loose part weak impact and grinding noise extraction method based on signal processing and machine learning in the technical scheme of the first aspect of the present application, and the system comprises:

[0037] A data acquisition module is configured to acquire a reactor vibration signal, and decompose the vibration signal into a plurality of sub-signals through empirical mode decomposition;

[0038] A time-frequency analysis module is configured to perform Hilbert transform processing on each sub-signal, extract the instantaneous frequency and instantaneous amplitude of each sub-signal, and construct a time-frequency two-dimensional matrix of each sub-signal;

[0039] A three-dimensional matrix generation module is configured to stack the time-frequency two-dimensional matrix of each sub-signal along the sub-signal dimension to generate a three-dimensional instantaneous time-frequency matrix;

[0040] A noise suppression module is configured to input the three-dimensional instantaneous time-frequency matrix into a pre-trained convolutional neural network model, suppress reactor operation noise, and extract a weak impact and grinding noise signal after noise reduction;

[0041] A fault identification module is configured to identify the loose part impact and grinding time and judge the fault type based on the weak impact and grinding noise signal after noise reduction.

[0042] The present application has the following beneficial effects:

[0043] The reactor in-core loose part weak impact and grinding noise extraction method based on signal processing and machine learning provided by the present application first adaptively decomposes a non-stationary vibration signal into a plurality of sub-signals through empirical mode decomposition, achieving preliminary separation of noise and target signal; then extracts the instantaneous frequency and instantaneous amplitude of each sub-signal through Hilbert transform, constructs a time-frequency two-dimensional matrix to enhance the time-frequency localization representation ability of the signal; further stacks the two-dimensional time-frequency matrices of multiple sub-signals into a three-dimensional structure to provide multi-dimensional feature input for the deep learning model; finally, a pre-trained CNN model extracts high-distinguishability features in the time-frequency space, suppresses background noise, and reconstructs the weak impact and grinding signal. This method solves the limitations of traditional single-dimensional signal processing techniques in extracting reactor in-core loose part noise, significantly improves the detection sensitivity and reliability of weak impact and grinding events under low signal-to-noise ratio conditions, and realizes the rapid extraction and accurate reconstruction of in-core loose part impact and grinding signals. This method can quickly and effectively extract the weak vibration signal generated by the in-core loose part impact and grinding from the reactor operation noise, supporting subsequent reactor in-core loose part identification and avoiding safety hazards. BRIEF DESCRIPTION OF DRAWINGS

[0044] In order to more clearly illustrate the technical solutions in the embodiments of the present application or the prior art and the advantages thereof, the following will briefly introduce the drawings needed in the embodiments or prior art description. Obviously, the drawings in the following description are only some embodiments of the present application, and for those skilled in the art, other drawings can also be obtained from these drawings without creative effort.

[0045] Figure 1 A method flowchart of the reactor in-vessel loose part weak impact grinding noise extraction method based on signal processing and machine learning provided by an embodiment of the present application;

[0046] Figure 2 A schematic diagram of a three-dimensional instantaneous time-frequency matrix provided by an embodiment of the present application;

[0047] Figure 3 A structural schematic diagram of the reactor in-vessel loose part weak impact grinding noise extraction system based on signal processing and machine learning provided by an embodiment of the present application. DETAILED DESCRIPTION

[0048] In order to further illustrate the technical means and effects adopted by the present application to achieve the predetermined purposes, the following describes the method and system for extracting weak impact grinding noise of reactor in-vessel loose parts based on signal processing and machine learning according to the present application, the specific implementation, structure, features and effects thereof in detail in combination with the drawings and preferred embodiments. In the following description, different "one embodiment" or "another embodiment" do not necessarily refer to the same embodiment. In addition, the specific features, structures or characteristics in one or more embodiments can be combined in any suitable form.

[0049] Unless otherwise defined, all technical and scientific terms used herein have the same meaning as commonly understood by one of ordinary skill in the art to which the present application belongs.

[0050] The following specifically describes the specific scheme of the method and system for extracting weak impact grinding noise of reactor in-vessel loose parts based on signal processing and machine learning provided by the present application in combination with the drawings.

[0051] Please refer to Figure 1 which shows the method flowchart of the method for extracting weak impact grinding noise of reactor in-vessel loose parts based on signal processing and machine learning provided by an embodiment of the present application, the method comprising:

[0052] Step S100: acquiring a reactor vibration signal, and decomposing the vibration signal into a plurality of sub-signals through empirical mode decomposition; specifically, an acceleration vibration sensor can be installed at a key position such as a coolant inlet, a mechanical connection point, etc. of a reactor pressure vessel head to collect a one-dimensional time sequence vibration signal at a high frequency sampling rate The signal is a mixed signal, which contains a narrow-band impact signal, i.e., a target signal, generated by the in-pile loose part rubbing, and an operating noise of the reactor;

[0053] The step S100 specifically comprises:

[0054] Step S110: sequentially decomposing the vibration signal into a plurality of intrinsic modal functions and a residual signal, wherein each intrinsic modal function satisfies a matching condition of a number of local extreme points and a number of zero-crossing points through iterative screening, and the residual signal is a low-frequency component after decomposition; specifically, first, the vibration signal is subjected to iterative screening, and an intrinsic modal function satisfying the matching condition of the number of local extreme points and the number of zero-crossing points is sequentially extracted; that is, the signal is traversed , all local maximum points and local minimum points are marked; then, the maximum points and the minimum points are respectively fitted using a cubic spline interpolation method to generate an upper envelope line and a lower envelope line, and then the mean values of the upper envelope line and the lower envelope line are calculated, and the mean values are subtracted from the original signal to obtain a preliminary component, and whether the preliminary component satisfies the matching condition of the number of local extreme points and the number of zero-crossing points is checked, so that the intrinsic modal function is extracted; after each intrinsic modal function is extracted, the residual signal is taken as a new input signal for continuous decomposition; when the residual signal is a low-frequency component that is monotonic or contains only a single extreme value, the decomposition process is terminated, and all intrinsic modal functions and a final residual signal are obtained. The expression of the intrinsic modal function is:

[0055]

[0056] In the formula, y(t) is the vibration signal input at the moment t; is the intrinsic modal function of the jth order at the moment t, i.e., a sub-signal after decomposition, , , is the total number of intrinsic modal functions; is the residual signal at the moment t, reflecting a low-frequency trend of the signal;

[0057] In this embodiment, the mixed vibration signal of the reactor is adaptively decomposed into a plurality of intrinsic modal functions (IMF) and a residual signal through empirical mode decomposition, so that multi-scale separation of the noise and the target signal is realized. This process does not need to preset a base function, automatically adapts to the non-stationary characteristics of the signal, enriches the high-frequency characteristics of the weak rubbing signal in a specific IMF component, and disperses the wide-frequency noise to a plurality of IMF and residual signals. This decomposition method not only retains the integrity of the time-frequency characteristics of the rubbing signal, but also avoids the feature loss caused by the fixed frequency band assumption in the traditional filtering method, and can effectively improve the recognition ability of the subsequent time-frequency analysis and machine learning model for the weak rubbing signal.

[0058] ​​​​Step S200: Hilbert transform is performed on each sub-signal, the instantaneous frequency and the instantaneous amplitude of each sub-signal are extracted, and a time-frequency two-dimensional matrix of each sub-signal is constructed;

[0059] Step S200 specifically comprises:

[0060] Step S210: an analytic signal is constructed for each sub-signal, the analytic signal comprising the sub-signal and the imaginary part of the Hilbert transform of the sub-signal; specifically, based on the intrinsic mode function obtained by the decomposition in step S100 ; first, frequency domain conversion is performed based on the frequency domain sequence obtained by fast Fourier transform , and then phase offset is applied to the frequency domain sequence to obtain the frequency domain sequence after Hilbert transform; finally, the frequency domain sequence after Hilbert transform is converted back to time domain through inverse Fourier transform, that is, the imaginary part of the Hilbert transform is obtained; the analytic signal can be expressed as:

[0061]

[0062] In the formula, is the analytic signal, the real part is the original sub-signal, and the imaginary part is the Hilbert transform result; is the imaginary part of the Hilbert transform; represents the imaginary unit;

[0063] Step S220: the instantaneous amplitude at each time point is calculated according to the modulus value of the analytic signal; the instantaneous amplitude can be expressed as:

[0064]

[0065] In the formula, represents the instantaneous amplitude at the time point, and the greater the instantaneous amplitude, the stronger the vibration energy at the time point, and the more likely it is to correspond to the loose part rubbing event;

[0066] Step S230: the instantaneous frequency at each time point is calculated according to the phase change rate of the analytic signal; specifically, the instantaneous phase at each time point is calculated according to the phase of the analytic signal of each sub-signal, and the instantaneous frequency is calculated based on the time sequence change rate of the instantaneous phase, and the instantaneous phase can be expressed as:

[0067]

[0068] In the formula, represents the instantaneous phase at the time point; represents the inverse tangent function;

[0069] The instantaneous frequency can be expressed as:

[0070]

[0071] In the formula, express Instantaneous frequency at any given moment; Represents the differential operator;

[0072] Step S240: Using the time series as the column dimension and the instantaneous frequency as the row dimension, fill the matrix elements with the instantaneous amplitude corresponding to each time point to generate a two-dimensional time-frequency matrix of time-frequency energy distribution, which can be represented as:

[0073]

[0074] In the formula, Represents a two-dimensional time-frequency matrix; Indicates frequency and time Instantaneous amplitude at the location; Indicates the first One frequency point; Indicates the first There are 1 time points; the matrix size is 1 In this embodiment, if the frequency or time resolution is insufficient, such as if there are high-frequency details in the signal that are not covered by discrete frequency points, bilinear interpolation can be used to generate new values ​​between adjacent matrix elements to improve the smoothness of the matrix. For sparsely distributed energy points, two-dimensional Gaussian filtering can be applied to suppress noise, enhance the continuity of time-frequency features, and improve the subsequent noise reduction effect.

[0075] In summary, this embodiment uses Hilbert transform to convert the intrinsic mode functions obtained from step S100 into a two-dimensional time-frequency matrix, achieving a refined time-frequency characterization of the vibration signal. This process extracts instantaneous amplitude and frequency by constructing an analytical signal, mapping the one-dimensional time-domain signal into a three-dimensional feature space. This allows the high-frequency impact characteristics of weak impact signals to appear as local energy accumulation regions in the time-frequency matrix, while broadband noise is dispersed as a low-energy background. This time-frequency separation characteristic not only enhances the distinguishability between the target signal and noise but also provides structured input for subsequent convolutional neural networks. This enables the convolutional neural network to achieve adaptive noise suppression through spatial feature extraction, further improving the extraction accuracy of weak impact signals from loose components within the reactor and enhancing early fault detection capabilities.

[0076] Step S300: Superimpose the two-dimensional time-frequency matrices of each sub-signal along the sub-signal dimensions to generate a three-dimensional instantaneous time-frequency matrix;

[0077] Step S300 specifically includes:

[0078] Step S310: Interpolation processing is performed on the time-frequency two-dimensional matrix of each sub-signal in the frequency and time dimensions; specifically, due to the insufficient sampling rate or frequency resolution of the original signal, the time-frequency two-dimensional matrix of each sub-signal generated in step S200 may have sparse grid points or jagged edges in the matrix, affecting the continuity of the three-dimensional instantaneous time-frequency matrix and the model training effect. The embodiment fills in the missing time-frequency points, improves the spatial resolution and smoothness of the matrix, and ensures that the characteristics of different sub-signals are accurately aligned in the three-dimensional space. Two-dimensional interpolation is performed on the time axis and the frequency axis of each two-dimensional matrix to ensure that the time-frequency matrices of all sub-signals have the same frequency division, and the frequency axis grid points of different sub-signals are completely consistent through interpolation;

[0079] Step S320: The interpolated time-frequency two-dimensional matrix is sequentially stacked along the sub-signal dimension to generate a three-dimensional instantaneous time-frequency matrix including sub-signals, frequencies, and times; specifically, the interpolated time-frequency two-dimensional matrix of each IMF is sequentially stacked along the depth direction to form a three-dimensional instantaneous time-frequency matrix, which can be seen from Figure 2 The three-dimensional instantaneous time-frequency matrix has a width of time , a height of instantaneous frequency , and a depth of sub-signal samples , corresponding to the total number of intrinsic mode functions obtained in step S100;

[0080] Step S330: Smooth filtering is performed on the three-dimensional instantaneous time-frequency matrix to remove abnormal data points introduced by interpolation; specifically, the interpolation process may introduce high-frequency noise or abnormal values, such as amplitude mutations caused by signal edge effects. The embodiment suppresses such interference through filtering to ensure the continuity of the characteristics of the three-dimensional instantaneous time-frequency matrix; a three-dimensional Gaussian filter can be used to convolve the three-dimensional instantaneous time-frequency matrix with a three-dimensional Gaussian kernel, and then threshold pruning is performed on the filtered three-dimensional instantaneous time-frequency matrix to remove abnormal points with an amplitude lower than a preset proportion of the global mean, further purifying the data.

[0081] The embodiment fuses the dispersed one-dimensional sub-signal time-frequency characteristics into a structured three-dimensional instantaneous time-frequency matrix by interpolation processing, stacking along the sub-signal dimension and smoothing filtering of each sub-signal time-frequency two-dimensional matrix, realizes the three-dimensional representation and enhancement of multi-dimensional characteristics. The process ensures the accurate alignment of the time-frequency characteristics of different sub-signals by unifying the resolution of the frequency and time dimensions, forms a three-dimensional data structure of "frequency-time-sub-signal", not only retains the independent time-frequency characteristics of each sub-signal, but also establishes the cross-scale feature association through the depth dimension. The interpolation and smoothing processing eliminates the discretization noise and edge distortion, so that the energy characteristics of the rubbing signal present continuous and concentrated distribution in the three-dimensional space, while the noise is suppressed as a low-amplitude smooth background. This three-dimensional data structure is highly adapted to the input characteristics of the convolutional neural network, providing the model with multi-level features containing time series, frequency distribution and sub-signal association, and further improving the feature extraction ability and generalization adaptability of the subsequent noise suppression algorithm for complex vibration signals.

[0082] Step S400: inputting the three-dimensional instantaneous time-frequency matrix into a pre-trained convolutional neural network model, suppressing the reactor operation noise and extracting the weak rubbing noise signal after noise reduction;

[0083] Step S400 specifically includes:

[0084] Step S410: extracting three-dimensional local features by sliding a three-dimensional convolution kernel in the three-dimensional instantaneous time-frequency matrix; based on the three-dimensional instantaneous time-frequency matrix finally generated in step S300, a three-dimensional convolution kernel is used to perform convolution operation in three dimensions of depth, height and width of the three-dimensional matrix, and the size of the convolution kernel is wherein controlling the perception range in the depth dimension, controlling the frequency perception range in the time-frequency plane, controlling the local window on the time axis; the depth dimension model can capture the correlation characteristics between different IMFs. For example, the rubbing signal of the loose part is manifested as a sudden change in instantaneous amplitude in the high-frequency IMF, and at the same time causes frequency modulation in the adjacent medium-frequency IMF, and the three-dimensional convolution kernel can detect such cross-layer characteristics; secondly, for time-frequency local pattern recognition, in the frequency-time plane (f x t window), the convolution kernel can be used to identify specific time-frequency patterns, such as a sudden increase in energy of a preset frequency interval within a short time window, corresponding to the typical time-frequency characteristics of the rubbing impact.

[0085] ​Step S411: The three-dimensional local feature matrix output by the three-dimensional convolution kernel is down-sampled using a max-pooling layer; specifically, the pooling layer down-samples the three-dimensional feature matrix output by the convolution layer through a three-dimensional pooling operation to reduce the dimension size and computational complexity while retaining the main features; this embodiment adopts a multi-convolution layer stacking strategy to gradually extract abstract features through alternating convolution layers and pooling layers, and the three-dimensional feature matrix output after each convolution and pooling operation will be used as the operation of the next convolution and pooling operation. After multiple convolution and pooling processes, a three-dimensional feature matrix of the original data features is obtained, and the three-dimensional feature matrix is flattened into a one-dimensional vector using a Flatten layer.

[0086] Step S412: The down-sampled features are mapped into a one-dimensional time series signal through a fully connected layer, and a denoised weak rubbing noise signal is output; specifically, the fully connected layer includes a hidden layer and an output layer, and is used to map the flattened features into a feature vector of higher or lower dimension size. By continuously stacking multiple fully connected layers and introducing a nonlinear activation function between the layers, further feature extraction is achieved; finally, the last fully connected layer is used as the output layer, and the dimension size is configured to be consistent with the number of collection points of the weak rubbing signal of the in-core loose part submerged in the reactor operating noise, so as to obtain the time series data of the denoised signal.

[0087] Step S410: The sliding convolution operation is performed on the depth, height and width dimensions of the three-dimensional instantaneous time-frequency matrix through the three-dimensional convolution kernel, realizing multi-dimensional local feature extraction and enhancement of the reactor vibration signal. The three-dimensional convolution kernel is used to capture the depth, frequency and time dimensions among the sub-signals at the same time, and the cross-layer energy mutation and time-frequency local clustering pattern of the rubbing signal in different intrinsic mode functions are converted into identifiable feature mapping. Through the stacking of multiple convolution kernels, the model can abstract complex composite features from the basic time-frequency point features layer by layer, thereby effectively suppressing the dispersion interference of the reactor operating noise and enhancing the feature saliency of the weak rubbing signal. This method not only retains the integrity of the signal in the three dimensions of sub-signal, frequency and time, but also reduces the model parameter complexity and improves the adaptability to non-stationary vibration signals through the weight sharing and local connection characteristics of the convolution operation.

[0088] The training process of the convolutional neural network model includes the following sub-steps:

[0089] Step S420: Construct a training data set including a three-dimensional instantaneous time-frequency matrix with noise and a corresponding noise-free target signal, which is used to train the convolutional neural network to learn the mapping relationship for recovering the true rubbing signal from the noisy signal; the noisy mixed noise is subjected to empirical mode decomposition, Hilbert transform and matrix difference and stacking operations; the noise-free target signal is subjected to empirical mode decomposition, Hilbert transform and matrix difference and stacking operations Similarly, empirical mode decomposition is performed to obtain noise-free IMF, which can be directly used as labeled time-series data. This embodiment can also address random noise and local occlusion in the three-dimensional instantaneous time-frequency matrix, such as Gaussian white noise, to improve the robustness of the model.

[0090] Step S421: Quantify the difference between the model output and the target signal using a loss function; wherein, the loss function is a weighted mean square error function, the weights of the loss function are dynamically determined by the energy distribution of the intrinsic mode functions of the target signal, and the expression of the loss function is:

[0091]

[0092] In the formula, This represents the value of the loss function; Indicates the total number of time points of the signal; This represents the total number of intrinsic mode functions obtained after empirical mode decomposition; The first one represents the target signal The intrinsic mode functions at time point The square value; This indicates the denoised signal output by the model at time point The value; Indicates the noise-free target signal at time point The true value; Indicates the target signal at time point The total energy is used as a weighting term. For example, the error weight corresponding to the impact of rubbing and grinding is larger, so that the model can prioritize the recovery of signal details in high-energy regions.

[0093] Step S422: Update the convolutional kernel weights and fully connected layer parameters using the backpropagation algorithm until the convolutional neural network model converges. Specifically, the Adam optimizer can be used to adaptively adjust the learning stride of each parameter. For hyperparameter tuning, the size of each dimension of the convolutional kernel can be manually adjusted, such as the receptive range in the depth direction, the size of the local receptive region in the time-frequency plane, the number of stacked convolutional pooling layers, and the number of fully connected layers. Simultaneously, different types of loss functions can be tried, such as mean squared error loss function, hinge loss function, cross-entropy loss function, etc., and different types of activation functions, such as those simulating biological neurons. The optimal parameter combination is selected by comparing the model's performance on the training and validation sets using sigmoid functions, hyperbolic tangent functions, and modified linear unit functions, which are used to activate neurons. To prevent overfitting, this embodiment employs two strategies for regularization: first, a random deactivation mechanism is introduced into the neural network layers, which randomly ignores some neurons during training at a preset ratio to avoid excessive dependence between neurons; second, a weight decay term is added to the loss function to constrain the weight parameters of the neural network layers, suppressing overfitting caused by an excessively large weight matrix, thereby improving the model's generalization ability and stability.

[0094] In summary, the embodiment provides a multi-dimensional feature and label accurately aligned learning sample for the convolutional neural network by constructing a training data set containing a noisy three-dimensional instantaneous time-frequency matrix and a corresponding noise-free target signal. By simulating real rubbing scenarios and injecting diversified noise, complex interference conditions that may occur during reactor operation are covered. At the same time, the empirical mode decomposition processing of the noise-free target signal ensures that the label data and the input feature are consistent in physical meaning, so that the model can learn the nonlinear mapping relationship from the "sub-signal-frequency-time" three-dimensional structure to the noise-free time series signal. The data set generated thereby not only provides the energy distribution basis for the dynamic weight of the loss function, but also can be connected to the real-time monitoring system through the standardization preprocessing process, ultimately improving the recognition accuracy of the model for weak rubbing signals in actual application, and enhancing the reliability of early monitoring of reactor internal loose parts.

[0095] As a preferred technical solution, the reactor internal loose part weak rubbing noise extraction based on signal processing and machine learning further comprises:

[0096] Step S500: reconstructing the denoised weak rubbing noise signal in time domain waveform to generate a one-dimensional output signal aligned with the time sequence of the original vibration signal for loose part rubbing event identification; specifically, align the one-dimensional time sequence data generated by the full connection layer of the denoised weak rubbing noise signal output by step S400 with the time sequence reference of the original vibration signal. Since the time dimension is compressed by the pooling operation in the convolutional neural network processing process, linear interpolation or cubic spline interpolation method is used to restore the time sequence length of the denoised signal to the same number of collection points as the original signal, eliminating the time sequence deviation caused by downsampling.

[0097] Step S500 specifically comprises:

[0098] Step S510: based on the instantaneous phase information extracted in step S200 , phase compensation is performed on the reconstructed signal. Specifically, by calculating the phase difference between the denoised signal and the original signal at the key frequency points in the main frequency band of the loose part rubbing event, for example, the denoised signal is phase corrected by using a digital filter to ensure that its time domain waveform is consistent with the phase evolution characteristics of the original signal;

[0099] Step S520: arrange the interpolated and phase calibrated signal in time sequence to generate the final one-dimensional output signal; if the CNN output is a multi-channel feature vector, such as a multi-sub-signal fusion result, it needs to be mapped to a single channel by weighted summation or principal component analysis to retain the core rubbing features while suppressing redundant noise;

[0100] Step S530: Ensure that the reconstructed signal does not introduce artificial distortion by calculating the correlation coefficient of the reconstructed signal and the original signal in the noise-free time period, or verifying the integrity of the rub-impact event using short-time energy analysis.

[0101] The embodiment reconstructs the denoised weak rub-impact noise signal into a one-dimensional time sequence signal strictly aligned with the original vibration signal, eliminates the time sequence offset caused by the down-sampling of the convolutional neural network, ensures the strict synchronization of the denoised signal and the original signal, and provides high-precision input for subsequent time-domain positioning of the loose part rub-impact event; on the other hand, based on the calibration mechanism of instantaneous phase compensation, the interference of phase distortion on frequency analysis in the denoising process can be avoided, and the extraction reliability of the rub-impact event frequency feature is improved; finally, the one-dimensional output signal can be directly adapted to traditional loose part recognition algorithms, such as envelope demodulation and support vector machine classifier, without the need for additional design of feature conversion interface, thereby reducing the system integration complexity; the reconstructed signal completely retains the transient impact waveform of the rub-impact event, which is convenient for operation and maintenance personnel to intuitively judge the mechanical behavior mode of the loose part and assist in fault diagnosis decision.

[0102] To sum up, the weak rub-impact noise extraction method for the in-core loose part of the reactor based on signal processing and machine learning provided by the present application uses empirical mode decomposition to adaptively decompose the non-stationary vibration signal into multiple sub-signals, realizes the preliminary separation of noise and target signal, then extracts the instantaneous frequency and instantaneous amplitude of each sub-signal through Hilbert transform, constructs a two-dimensional time-frequency matrix to enhance the time-frequency localization representation ability of the signal, further stacks the two-dimensional time-frequency matrices of multiple sub-signals into a three-dimensional structure to provide multi-dimensional feature input for the deep learning model, and finally extracts high-distinguishability features in the time-frequency space through the pre-trained CNN model, suppresses background noise and reconstructs the weak rub-impact signal. This method solves the limitations of traditional single-dimensional signal processing techniques in extracting the noise of the in-core loose part of the reactor, significantly improves the detection sensitivity and reliability of weak rub-impact events under low signal-to-noise ratio conditions, and realizes the rapid extraction and accurate reconstruction of the rub-impact signal of the in-core loose part of the reactor. This method can quickly and effectively extract the weak vibration signal generated by the rub-impact of the in-core loose part of the reactor, which is submerged in the operation noise of the reactor, to support the subsequent identification of the in-core loose part of the reactor and avoid safety hazards.

[0103] Referring to Figure 3 which shows the structure of the weak rub-impact noise extraction system for the in-core loose part of the reactor based on signal processing and machine learning provided by an embodiment of the present application, the system comprises:

[0104] The data acquisition module is configured to acquire the vibration signal of the reactor and decompose the vibration signal into multiple sub-signals through empirical mode decomposition;

[0105] The time-frequency analysis module is configured to perform Hilbert transform on each sub-signal, extract the instantaneous frequency and instantaneous amplitude of each sub-signal, and construct a time-frequency two-dimensional matrix of each sub-signal;

[0106] The three-dimensional matrix generation module is configured to stack the time-frequency two-dimensional matrix of each sub-signal along the sub-signal dimension to generate a three-dimensional instantaneous time-frequency matrix;

[0107] The noise suppression module is configured to input the three-dimensional instantaneous time-frequency matrix into a pre-trained convolutional neural network model, suppress the reactor operation noise, and extract a weak rubbing noise signal after noise reduction;

[0108] The fault identification module is configured to identify the rubbing time of the loose part and determine the fault type based on the weak rubbing noise signal after noise reduction.

[0109] It should be noted that the above-mentioned embodiments of the present application are only for description, and do not represent the advantages and disadvantages of the embodiments. The processes depicted in the drawings do not necessarily require the specific order or continuous order shown to achieve the desired results. In some embodiments, multi-task processing and parallel processing are also possible or may be advantageous.

[0110] Each embodiment in the specification is described in a progressive manner, and the same or similar parts between each embodiment can be referred to each other. Each embodiment focuses on the difference from other embodiments.

Claims

1. A method for extracting weak impact noise of in-core loose parts of a reactor based on signal processing and machine learning, characterized in that, The method comprises: acquiring a reactor vibration signal, decomposing the vibration signal into a plurality of sub-signals through empirical mode decomposition; performing Hilbert transform processing on each sub-signal, extracting the instantaneous frequency and instantaneous amplitude of each sub-signal, and constructing a time-frequency two-dimensional matrix of each sub-signal; stacking the time-frequency two-dimensional matrices of each sub-signal along the sub-signal dimension to generate a three-dimensional instantaneous time-frequency matrix; inputting the three-dimensional instantaneous time-frequency matrix into a pre-trained convolutional neural network model to suppress reactor operation noise and extract a weak rub-impact noise signal after noise reduction.

2. The signal processing and machine learning based reactor incore loose part weak rub noise extraction method of claim 1, wherein, The method for acquiring a reactor vibration signal and decomposing the vibration signal into a plurality of sub-signals through empirical mode decomposition comprises: decomposing the vibration signal into a plurality of intrinsic mode functions and a residual signal in sequence, wherein each intrinsic mode function meets the matching condition of the number of local extreme points and the number of zero-crossing points through iterative screening, and the residual signal is a low-frequency component after decomposition.

3. The signal processing and machine learning based reactor incore loose part weak rub noise extraction method of claim 1, wherein, The method for performing Hilbert transform processing on each sub-signal and extracting the instantaneous frequency and instantaneous amplitude of each sub-signal comprises: constructing an analytic signal for each sub-signal, wherein the analytic signal comprises the sub-signal and the imaginary part of the Hilbert transform of the sub-signal; calculating the instantaneous amplitude of each time point according to the modulus of the analytic signal; calculating the instantaneous frequency of each time point according to the phase change rate of the analytic signal.

4. The signal processing and machine learning based reactor incore loose part rub noise extraction method of claim 3, wherein, The method for constructing a time-frequency two-dimensional matrix of each sub-signal comprises: filling the instantaneous amplitude corresponding to each time point into the matrix elements with time series as the column dimension and instantaneous frequency as the row dimension to generate a time-frequency two-dimensional matrix of time-frequency energy distribution, which can be represented as: wherein represents a time-frequency two-dimensional matrix; represents an instantaneous amplitude at frequency and time ; represents the th frequency point; represents the th time point.

5. The signal processing and machine learning based reactor incore loose part weak rub noise extraction method of claim 1, wherein, The method for stacking the time-frequency two-dimensional matrices of each sub-signal along the sub-signal dimension to generate a three-dimensional instantaneous time-frequency matrix comprises: performing interpolation processing on the frequency and time dimensions of the time-frequency two-dimensional matrix of each sub-signal; sequentially stacking the interpolated time-frequency two-dimensional matrix along the sub-signal dimension to generate a three-dimensional instantaneous time-frequency matrix comprising sub-signals, frequencies, and times; performing smoothing filtering on the three-dimensional instantaneous time-frequency matrix to remove abnormal data points introduced by interpolation.

6. The signal processing and machine learning based reactor incore loose part weak rub noise extraction method of claim 1, wherein, The method for inputting the three-dimensional instantaneous time-frequency matrix into a pre-trained convolutional neural network model to suppress reactor operation noise and extract a weak rub-impact noise signal after noise reduction comprises: extracting three-dimensional local features by sliding a three-dimensional convolution kernel in the three-dimensional instantaneous time-frequency matrix; down-sampling the three-dimensional local feature matrix output by the three-dimensional convolution kernel using a max-pooling layer; mapping the down-sampled features to a one-dimensional time series signal through a fully connected layer to output a weak rub-impact noise signal after noise reduction.

7. The signal processing and machine learning based reactor incore loose part weak rub noise extraction method of claim 6, wherein, The training process of the convolutional neural network model comprises the following sub-steps: constructing a training data set comprising a three-dimensional instantaneous time-frequency matrix with noise and a corresponding noise-free target signal; quantifying the difference between the model output and the target signal by a loss function; updating the convolution kernel weights and fully connected layer parameters using a backpropagation algorithm until the convolutional neural network model converges.

8. The signal processing and machine learning based reactor incore loose part weak rub noise extraction method of claim 7, wherein, The loss function is a weighted mean square error function, and the weights of the loss function are dynamically determined by the intrinsic mode function energy distribution of the target signal. The expression of the loss function is: In the formula, denotes the loss function value; denotes the total number of time points of the signal; denotes the total number of intrinsic mode functions obtained after empirical mode decomposition; denotes the square value of the th intrinsic mode function of the target signal at the time point ; denotes the value of the denoised signal output by the model at the time point ; denotes the true value of the noise-free target signal at the time point ; denotes the true value of the noise-free target signal at the time point .

9. The signal processing and machine learning based loose part weak rub noise extraction method of any one of claims 1 to 8, wherein, The method further comprises: The weak rubbing noise signal after noise reduction is subjected to time domain waveform reconstruction to generate a one-dimensional output signal in time sequence alignment with the original vibration signal, which is used for loose part rubbing event identification.

10. A signal processing and machine learning based reactor incore loose part weak rub noise extraction system, characterized in that, The system comprises the weak rubbing noise extraction method based on signal processing and machine learning according to any one of claims 1 to 9. The data acquisition module is configured to acquire a reactor vibration signal, and decompose the vibration signal into a plurality of sub-signals through empirical mode decomposition; The time-frequency analysis module is configured to perform Hilbert transform processing on each sub-signal, extract the instantaneous frequency and instantaneous amplitude of each sub-signal, and construct a two-dimensional time-frequency matrix of each sub-signal; The three-dimensional matrix generation module is configured to stack the two-dimensional time-frequency matrices of each sub-signal along the sub-signal dimension to generate a three-dimensional instantaneous time-frequency matrix; The noise suppression module is configured to input the three-dimensional instantaneous time-frequency matrix into a pre-trained convolutional neural network model to suppress reactor operating noise and extract a weak rubbing noise signal after noise reduction; The fault identification module is configured to identify the loose part rubbing time and determine the fault type based on the weak rubbing noise signal after noise reduction.

Images