A method and device for identifying a nuclear power plant pipeline blockage based on vibration monitoring

Abstract

The application provides a kind of nuclear power station pipeline blockage identification method and device based on vibration monitoring, it is related to artificial intelligence technical field, by collecting the vibration signal of pipeline in the process of fluid state change, can directly reflect the flow field and structural vibration change caused by blockage. Its vibration component set can cover the vibration modal information under different frequency bands, different excitation sources, after discriminant component reorganization, can obtain noise and irrelevant interference, highlight the sensitive features related to blockage degradation. Corresponding multi-view time-invariant features can depict the stable change law of signal from multiple dimensions, and the fusion can construct a global feature vector that reflects the correlation between fluid state and blockage degradation. The global feature has the ability to accurately identify different blockage levels and degradation levels, can effectively capture the differences and gradual change trend of blockage state, and finally realizes the efficient, stable and reliable identification of nuclear power station pipeline blockage state.

Description

Technical Field

[0001] This invention relates to the field of artificial intelligence technology, and in particular to a method and apparatus for identifying pipeline blockages in nuclear power plants based on vibration monitoring. Background Technology

[0002] Nuclear power plants are a crucial component of national energy production, and their piping systems, as critical infrastructure, are responsible for transporting essential media such as coolant and feedwater. The operational stability of the piping system directly impacts the safety and efficiency of the nuclear power plant. However, due to the long-term exposure to fluid impacts, temperature variations, and mechanical wear, pipes can experience varying degrees of blockage, particularly at bends, valves, and downstream of pumps. Pipeline blockage not only increases resistance to fluid transport but can also lead to system pressure imbalances, equipment overload, and even safety accidents. Therefore, timely and accurate identification and early warning of pipeline blockage are of paramount importance for ensuring the safe operation of nuclear power plants.

[0003] Existing methods for monitoring pipeline blockage based on vibration signals typically employ time-frequency analysis to decompose and extract features from the acquired vibration signals. However, the vibration signals from nuclear power plant pipelines are essentially composite signals resulting from the coupling and superposition of multiple excitation sources, such as the main pump frequency, fluid turbulence, and structural modes. The weak feature components caused by blockage highly overlap with strong interference components in the frequency band. Existing methods, based on time-frequency decomposition strategies with a pre-set number of decomposition layers, struggle to adaptively adjust to the frequency band distribution characteristics of the signal itself. Furthermore, directly using all components after decomposition introduces a large amount of noise into the feature space, resulting in the obscuring of blockage-related sensitive features and an excessively high feature space dimension, making it difficult to accurately separate effective blockage-related information from the vibration signal.

[0004] In terms of feature extraction, existing methods are mostly limited to a single perspective, such as focusing only on time-domain statistics or frequency-domain amplitude spectra. However, changes in blockage state are often reflected in multiple dimensions simultaneously, such as impact response features in the time domain, spectral centroid shift in the frequency domain, and energy ridge migration in the time-frequency domain. Existing methods, which only characterize features from a single perspective, cannot fully capture the dynamic patterns related to blockage state from different transform domains, resulting in the loss of some key information in the feature extraction stage and affecting recognition accuracy. Summary of the Invention

[0005] The purpose of this invention is to provide a method and device for identifying pipeline blockage in nuclear power plants based on vibration monitoring. This method can accurately extract blockage-sensitive features under complex nuclear power operating conditions, achieve efficient and reliable identification of pipeline blockage status and its progressive deterioration trend, and improve monitoring accuracy and engineering practicality.

[0006] In a first aspect, embodiments of the present invention provide a method for identifying pipeline blockage in a nuclear power plant based on vibration monitoring. The method includes: acquiring pipeline vibration signals during fluid state changes in a target pipeline; performing discriminative component recombination on the pipeline vibration signals based on the set of vibration components contained in the pipeline vibration signals to construct discriminative features corresponding to the pipeline vibration signals; performing feature fusion on the time-invariant features of the discriminative features from multiple perspectives to obtain a global feature vector characterizing the fluid state changes of the target pipeline; and identifying and judging the blockage state of the target pipeline based on the global feature vector to determine the blockage identification result of the target pipeline.

[0007] In conjunction with the first aspect, the present invention also provides a first implementation of the first aspect, wherein the step of discriminatively recombining the pipeline vibration signal based on the set of vibration components contained in the pipeline vibration signal to construct discriminative features corresponding to the pipeline vibration signal includes: determining an adaptive weight corresponding to each sub-component according to a preset discriminative index corresponding to each sub-component in the set of vibration components; wherein the preset discriminative index is determined based on the kurtosis characteristics of the corresponding sub-component under different blockage states; and using the adaptive weights, adaptively weightedly recombining the signal components of the set of vibration components to construct discriminative features corresponding to the pipeline vibration signal.

[0008] In conjunction with the first aspect, the present invention also provides a second implementation of the first aspect, wherein the method further includes: performing a first variational mode decomposition on the pipeline vibration signal to obtain multiple intrinsic mode components; obtaining key mode components from the intrinsic mode components, and performing a second variational mode decomposition within the center frequency range of the key mode components to obtain corresponding key sub-components; and summing the intrinsic mode components and key sub-components to form a vibration component set corresponding to the pipeline vibration signal.

[0009] In conjunction with the first aspect, this embodiment of the invention also provides a third implementation of the first aspect, wherein the step of fusing discriminative features with time-invariant features from multiple perspectives to obtain a global feature vector characterizing the fluid state changes of the target pipeline includes: concatenating the discriminative features with time-invariant features from multiple perspectives along the feature dimension to generate a multi-perspective fusion matrix; performing attention enhancement processing on the multi-perspective fusion matrix using a multi-head self-attention mechanism to obtain a target feature matrix; and performing adaptive pooling aggregation on the target feature matrix to obtain a global feature vector characterizing the fluid state changes of the target pipeline.

[0010] In conjunction with the first aspect, this invention also provides a fourth implementation of the first aspect, wherein the discriminative features in the time-invariant features of the multi-viewpoint discriminative features include: time-domain waveform features of the discriminative features in the time domain, frequency-domain envelope features in the frequency domain, and time-frequency ridge features in the time-frequency domain; the time-domain waveform features are determined based on the local difference and integration results of the discriminative features; the frequency-domain envelope features are determined based on the energy center of the multi-resolution frequency domain analysis; and the time-frequency ridge features are determined based on the energy concentration trajectory of the high-resolution time-frequency transform.

[0011] In conjunction with the first aspect, this invention also provides a fifth implementation of the first aspect, wherein the step of identifying and judging the blockage state of a target pipeline based on a global feature vector and determining the blockage identification result of the target pipeline includes: inputting the global feature vector into a trained nonlinear projection network to map the corresponding continuous state latent vector; determining the blockage state and its continuous deterioration degree of the target pipeline based on the position of the continuous state latent vector in the continuous latent space constructed by the nonlinear projection network; performing adaptive gating fusion on the continuous state latent vector in conjunction with the pipeline operating parameters corresponding to the fluid state change process to generate a fused feature vector; performing multi-task prediction on the fused feature vector to output the blockage category, state continuity value, and prediction uncertainty of the target pipeline, thereby forming the blockage identification result of the target pipeline.

[0012] In conjunction with the first aspect, this embodiment of the invention also provides a sixth implementation of the first aspect, wherein the blockage identification result is determined based on a pre-constructed pipeline blockage identification model; the method further includes: dynamically updating the contribution weight of the training samples in a preset total loss function according to the classification difficulty and category of the training samples; wherein the total loss function is composed of a weighted sum of classification loss, regression loss, manifold regularization loss and uncertainty calibration loss; and training the pipeline blockage identification model based on the total loss function after the contribution weight adjustment.

[0013] Secondly, embodiments of the present invention provide a nuclear power plant pipeline blockage identification device based on vibration monitoring. The device includes: a signal acquisition module for acquiring pipeline vibration signals during fluid state changes in the target pipeline; a preprocessing module for performing discriminative component recombination on the pipeline vibration signals based on the set of vibration components contained in the pipeline vibration signals to construct discriminative features corresponding to the pipeline vibration signals; a data processing module for performing feature fusion on the time-invariant features of the discriminative features from multiple perspectives to obtain a global feature vector characterizing the fluid state changes of the target pipeline; and a prediction module for identifying and judging the blockage state of the target pipeline based on the global feature vector to determine the blockage identification result of the target pipeline.

[0014] Thirdly, embodiments of the present invention provide an electronic device, characterized in that the electronic device includes a processor and a memory, the memory storing machine-executable instructions that can be executed by the processor, and the processor executing the machine-executable instructions to implement the nuclear power plant pipeline blockage identification method based on vibration monitoring in any of the above embodiments.

[0015] Fourthly, embodiments of the present invention provide a machine-readable storage medium, characterized in that the machine-readable storage medium stores machine-executable instructions, which, when called and executed by a processor, cause the processor to implement the vibration monitoring-based nuclear power plant pipeline blockage identification method of any of the above embodiments.

[0016] The embodiments of this invention bring the following beneficial effects: This invention provides a method and apparatus for identifying pipeline blockage in nuclear power plants based on vibration monitoring. By acquiring vibration signals from the pipeline during fluid state changes, dynamic response information coupling flow-induced vibration and the internal flow state of the pipeline can be obtained, truly reflecting the flow field distortion, local eddies, and structural vibration changes caused by blockage. The corresponding vibration component set can cover vibration mode information under different frequency bands and different excitation sources. Based on this, discriminative component recombination of the original data can eliminate operating condition interference and noise-dominant components, retaining and strengthening feature components highly correlated with blockage deterioration, forming a highly discriminative feature expression. Furthermore, based on the discriminative feature extraction of multi-view time-invariant features, the stable change law of the signal can be characterized from multiple dimensions. The global feature vector obtained by fusion can comprehensively characterize the intrinsic correlation between fluid state changes and pipeline blockage deterioration, possessing the ability to accurately identify different blockage levels and different degrees of deterioration. Therefore, based on this global feature vector, the differences in blockage state and continuous deterioration trends can be accurately captured, achieving efficient and reliable pipeline blockage identification.

[0017] Other features and advantages of the invention will be set forth in the following description, or some features and advantages may be inferred from the description or determined without doubt, or may be learned by practicing the techniques described above.

[0018] To make the above-mentioned objects, features and advantages of the present invention more apparent and understandable, preferred embodiments are described below in detail with reference to the accompanying drawings. Attached Figure Description

[0019] To more clearly illustrate the specific embodiments of the present invention or the technical solutions in the prior art, the drawings used in the description of the specific embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are some embodiments of the present invention. For those skilled in the art, other drawings can be obtained from these drawings without creative effort.

[0020] Figure 1 A flowchart illustrating a method for identifying pipeline blockages in a nuclear power plant based on vibration monitoring, provided in an embodiment of the present invention; Figure 2 A flowchart illustrating the determination of a discriminative signal provided in an embodiment of the present invention; Figure 3 A schematic diagram illustrating the construction process of a global feature vector provided in an embodiment of the present invention; Figure 4 A schematic diagram of a pipeline blockage identification and multi-task prediction based on global feature vectors provided in an embodiment of the present invention; Figure 5 This is a schematic diagram comparing the performance stability of the embodiments of the present invention under different noise levels with various comparative methods; Figure 6 A schematic diagram of a nuclear power plant pipeline blockage identification device based on vibration monitoring provided in an embodiment of the present invention; Figure 7 This is a schematic diagram of the structure of an electronic device provided in an embodiment of the present invention. Detailed Implementation

[0021] To make the objectives, technical solutions, and advantages of the embodiments of the present invention clearer, 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. The components of the embodiments of the present invention described and shown in the accompanying drawings can generally be arranged and designed in various different configurations.

[0022] Therefore, the following detailed description of the embodiments of the invention provided in the accompanying drawings is not intended to limit the scope of the claimed invention, but merely to illustrate selected embodiments of the invention. All other embodiments obtained by those skilled in the art based on the embodiments of the invention without inventive effort are within the scope of protection of the invention.

[0023] To facilitate understanding, the method for identifying pipeline blockage in nuclear power plants based on vibration monitoring, provided in an embodiment of the present invention, will first be described, referring to... Figure 1 The method includes the following steps: Step S102: Obtain the pipe vibration signal of the target pipe during the fluid state change process.

[0024] In a nuclear power plant setting, a large amount of fluid needs to continuously circulate in a dense piping system to achieve heat transfer and conversion. The entire nuclear power plant is equipped with multiple piping systems, including the primary loop that delivers cooling water to the reactor core for cooling, and the secondary loop that delivers feedwater to the steam generator to produce steam. The fluids flowing through these pipes are all water-based media operating under high temperature and high pressure conditions, and are subjected to relatively harsh operating environments for extended periods.

[0025] As pipelines age, blockages become increasingly apparent and pose a significant threat to pipeline safety. In primary loop pipelines, the high-temperature, high-pressure water environment causes continuous corrosion on the inner walls of pipes and equipment, releasing trace amounts of corrosion products or metal ions into the fluid. These corrosion products circulate with the coolant and tend to deposit at locations where the flow cross-section abruptly changes, such as bends, valves, or pump outlets, due to localized flow turbulence. In secondary loop pipelines, trace impurities and corrosion products introduced by the feed water also gradually deposit and scale on the inner walls of the pipes as the water vapor circulates. Over time, these deposits gradually reduce the effective flow area of ​​the pipeline, leading to blockages.

[0026] To promptly detect such potential blockages, this invention acquires pipe vibration signals during fluid state changes in the target pipeline. Analysis of these vibration signals determines the presence and severity of blockages within the pipeline. Target pipelines can be selected from bends in the main coolant pipeline, downstream sections of the main feedwater regulating valve, or turbulent areas near the main pump outlet. These locations are prone to blockage due to abrupt changes in flow cross-section or flow field reversals, making them high-risk areas for impurity particle deposition and adhesion. Fluid state changes refer to transient phases during nuclear power plant operation where flow velocity, pressure, or temperature undergo dynamic transitions. Examples include the brief window of a step change in flow rate during main pump speed adjustments, the transition range of coolant temperature fluctuations caused by control rod movements, or the brief period of sudden pressure changes downstream of the feedwater regulating valve in response to load commands.

[0027] When an external transient event occurs, pressure waves and sudden changes in flow velocity naturally arise within the pipeline and propagate along the pipe. This broadband excitation, when applied to a blocked area, induces local flow field disturbances such as jet impact and eddy shedding due to the narrowing of the flow cross-section. Acquiring the vibration signal during this transient phase allows for the capture of the pipe wall's vibration response at the instant the pressure wave impacts the blocked area. By extracting the flow disturbance characteristics caused by the blockage from the waveform changes, information on fluid state changes due to the blockage can be effectively captured.

[0028] Step S104: Based on the set of vibration components contained in the pipeline vibration signal, perform discriminative component recombination on the pipeline vibration signal to construct the discriminative features corresponding to the pipeline vibration signal.

[0029] Since the pipeline vibration signal is a pipeline vibration response obtained during fluid state changes, it includes various vibration components such as the periodic mechanical vibration corresponding to the main pump frequency and its harmonics, broadband random pulsations caused by fluid turbulence, and the response induced by local flow changes at the blockage site. These components together constitute the vibration component set of the signal. In this set, the low-frequency components represent the rigid body motion of the main pump frequency and pipeline support structure, the mid-frequency components represent the bending vibration modes of the pipeline shell, and the high-frequency components represent the response induced by local flow changes at the blockage edge. Among these components, only a few high-frequency components are related to the blockage state, while the pump frequency component and turbulence noise component, which have a large energy proportion, are unrelated to the blockage. The physical information related to the blockage is suppressed and difficult to distinguish in these strong interference components.

[0030] To address this issue, embodiments of the present invention perform discriminative recombination on the vibration component set. The aim is to identify components sensitive to changes in blockage from the set, enhance the selected sensitive components, and then recombine them to construct a new time-domain sequence. This sequence represents the discriminative characteristics of the pipeline vibration signal. The pump frequency-dominant component unrelated to blockage can be effectively suppressed in its waveform, while the sensitive components reflecting changes in the pipeline flow cross-section are highlighted, thus forming a characteristic expression that can represent the current pipeline flow state. In one embodiment, sensitive components can be selected from the vibration component set based on the degree of correlation between each vibration component and changes in blockage state. These sensitive components are then merged through recombination processing to generate a characteristic sequence characterizing the pipeline state.

[0031] Step S106: Perform feature fusion on the discriminative features and the time-invariant features from multiple perspectives to obtain a global feature vector characterizing the fluid state changes of the target pipeline.

[0032] Although the discriminative features have been sensitively screened and enhanced, the flow rate and temperature of a nuclear power plant are frequently adjusted according to load commands during operation. The characteristic amplitudes of the same degree of blockage still differ under different operating conditions, which makes it difficult to determine stability.

[0033] To address this issue, this invention also analyzes the feature from multiple perspectives (such as statistical characteristics, frequency domain distribution, and time domain variation trends) to determine time-invariant features that reflect the blockage state from different perspectives and are unaffected by fluctuations in operating conditions. Based on this, the extracted multi-perspective time-invariant features are fused to obtain a compact and comprehensive global feature vector. This vector characterizes the blockage state of the target pipeline under changing fluid conditions. Different degrees of blockage correspond to separate projection positions in this vector space, and the more severe the blockage, the more monotonically increasing the magnitude of the vector along a specific direction.

[0034] Step S108: Based on the global feature vector, identify and judge the blockage status of the target pipeline, and determine the blockage identification result of the target pipeline.

[0035] The global feature vector determined by the above steps is a comprehensive data expression obtained by fusing stable information that is sensitive to congestion from multiple perspectives and is not affected by operating condition fluctuations. The multi-perspective time-invariant features it contains have had operating condition fluctuation factors such as flow rate and temperature removed during the construction process, so that the vector itself is a stable description of the congestion state under cross-operating condition conditions.

[0036] Based on this, the congestion status can be determined by the mapping relationship between the numerical combination of global feature vectors and the degree of congestion. Since the influence of operating condition fluctuations has been removed during the construction of this vector, even if the congestion is in its early, weak stage or the change process is extremely slow, the characteristic changes caused by the congestion will not be masked by amplitude drift. Furthermore, by tracking the evolution of this vector within a continuous time window, the dynamic trend of the congestion degree as the operation progresses can be captured, thus effectively supporting continuous monitoring and early warning throughout the entire process from early, weak congestion to severe congestion.

[0037] Furthermore, based on the above embodiments, this invention also provides another method for identifying pipeline blockages in nuclear power plants based on vibration monitoring, to provide a detailed description of the above embodiments. Specifically, refer to the following steps: Step S202: Obtain the pipe vibration signal of the target pipe during the fluid state change process.

[0038] In one implementation, high-sensitivity vibration acceleration sensors can be installed at specific locations on critical pipelines within the nuclear power plant. Critical pipelines may include sections prone to blockage, such as main coolant pipelines and main feedwater pipelines. Specific locations may include bends, pump outlets, and areas near valves where the flow cross-section changes abruptly. The installation direction is ensured to align with the radial or axial vibration direction of the pipeline to effectively capture pipeline vibration signals caused by changes in fluid state. This monitoring can be continuously performed during nuclear power plant operation via a data acquisition system, focusing on fluid state changes, specifically transient conditions such as main pump speed adjustments, control rod movements, or valve opening changes. During such transient events, naturally occurring pressure waves and velocity changes within the pipeline can effectively excite local flow disturbances caused by the narrowing of the flow cross-section at the blockage site. The pipeline vibration signals captured by the sensors during this phase exhibit non-stationary time-domain waveforms with fluctuating amplitude envelopes and dynamically evolving frequency components. The original vibration signals can be acquired at a fixed sampling frequency, with each sample containing a fixed length of time-series data as the obtained pipeline vibration signal sample. In one implementation, the sampling frequency can be set to 1000Hz, and the length of each sample is 8192 data points.

[0039] Step S204: Determine the adaptive weight corresponding to each sub-component based on the preset discriminant index corresponding to each sub-component in the set of vibration components contained in the pipeline vibration signal.

[0040] Step S206: Using adaptive weights, the signal components of the vibration component set are adaptively weighted and recombined to construct the discriminative features corresponding to the pipeline vibration signal.

[0041] In nuclear power plant pipeline vibration monitoring scenarios, the original vibration signal is subjected to coupling interference from strong background noise and multiple sources of excitation such as main pump frequency and fluid turbulence. The weak characteristic signal frequency bands representing blockage overlap with the frequency bands of interference components, making them difficult to identify directly. If all frequency band components are analyzed uniformly without distinction, the weak information related to blockage will be masked by strong interference components and cannot be effectively extracted. Conventional methods such as wavelet transform or empirical mode decomposition can achieve time-frequency decomposition of the signal (to determine the included vibration components), but they have limitations such as requiring pre-setting the number of decomposition levels or modes. Furthermore, directly using all decomposed components will introduce a large number of noise components unrelated to blockage, and will also result in excessively high feature space dimensionality in subsequent analysis, affecting identification efficiency and accuracy.

[0042] To address the aforementioned issues, this invention employs a dual variational mode decomposition strategy to determine the set of vibration components in the pipeline vibration signal, enabling precise separation of signal frequency bands and focused extraction of key frequency bands. Furthermore, a discriminant index is introduced to adaptively weight and recombine the components in the vibration component set, thereby enhancing blockage-related sensitive components and effectively suppressing interference components, thus constructing a discriminant feature that reflects the pipeline blockage state.

[0043] 1) Calculate the comprehensive discriminant index of all components and the congestion category label. This index measures the correlation strength of each component in distinguishing different congestion states. Adaptively recombining all components using a weighted method based on this index generates an enhanced preprocessing signal (i.e., discriminant features), expressed as:

[0044] In the formula, The enhanced preprocessed signal (i.e. discriminative feature) is the preprocessed signal output after discriminative reconstruction. It is an enhanced vibration signal in which the signal components related to blockage are enhanced, while noise and irrelevant modes are effectively suppressed. Its dimension is 8192. This represents the set of indices for all vibration components. In this embodiment of the invention, the set of vibration components is determined based on a dual variational mode decomposition strategy, corresponding to all components included in the first decomposition. Each intrinsic mode function and all sub-components obtained from the second decomposition. Indicates that the index is The vibrational components can be the eigenmode functions of the first-level decomposition. Or the subcomponents of the second decomposition . A global index representing a component; Indicates the first The adaptive weights of each component in the reorganization process, with larger weight values ​​indicating greater importance for distinguishing congestion states, are calculated as follows: ; This represents the sharpening factor, used to amplify the weights of high-discriminative components and enhance the selectivity of the reconstructed signal. It is a positive hyperparameter, and examples of its values ​​are shown below. ; Indicates the first The comprehensive discriminant index of each component indicates that the higher the value, the stronger the class discrimination ability of that component. Specifically, it is obtained by calculating the Pearson correlation coefficient between the kurtosis feature of that component and the class label on all training samples, expressed as: .

[0045] Indicates the first The kurtosis value of each component sequence, i.e. Kurtosis is highly sensitive to the impact components and transient changes in the signal. Local flow changes such as eddy shedding caused by blockage and jet impact will be reflected in the corresponding components as significant differences in kurtosis characteristics. Indicates the first The training sample data corresponding to the th training sample data The kurtosis value of each component sequence is calculated as follows:

[0046] in, Represents the Hilbert transform operator; Indicates the first A comprehensive discriminant index for each component; Indicates the total number of samples; Indicates the sample index; Indicates the first The training sample data corresponding to the th training sample data The component sequence is the set of components belonging to the sample obtained from the dual variational mode decomposition. That is, it can be the eigenmode functions obtained from the first decomposition of the sample. Or, the subcomponents obtained from the second decomposition. ; Represents the category label vector, i.e. ; Indicates the first The congestion status category label corresponding to each training sample is numerically encoded, for example, 0 represents normal, 1 represents slight congestion, and 2 represents severe congestion; This represents the Pearson correlation coefficient calculation function, used to measure the degree of linear correlation between two sequences. The larger the absolute value, the stronger the linear correlation.

[0047] 2) The set of vibration components in the embodiments of the present invention can be determined by the following steps: a- Perform first-stage variational mode decomposition on the original vibration signal to obtain multiple intrinsic mode functions and their corresponding center frequencies.

[0048] In one implementation, the original vibration signal can be adaptively decomposed into K1 intrinsic mode functions (EMFs) by constructing and solving a variational optimization problem. The core idea of ​​variational mode decomposition is to transform the signal decomposition problem into a constrained variational optimization problem, aiming to achieve effective signal separation and reconstruction while minimizing the sum of estimated bandwidths for each mode. Specifically, for each EMF, its analytic signal can be calculated using the Hilbert transform to obtain a one-sided spectrum. Then, the spectrum of each mode is shifted to the baseband using an exponential term, and the bandwidth of each mode is estimated using Gaussian smoothing, i.e., the squared L2 norm of the shifted gradient is used as the bandwidth estimate. By introducing a quadratic penalty term and a Lagrange multiplier term, the constrained variational problem is transformed into an unconstrained augmented Lagrange function form. The alternating direction multiplier method is used to alternately optimize and iteratively update each EMF and its corresponding center frequency until the convergence condition is met, thus decomposing the original signal into K1 EMFs and their corresponding center frequencies.

[0049] Each intrinsic mode function (EMF) is represented as an amplitude-frequency modulated (AM-FM) signal, distributed around its center frequency in the frequency domain. It characterizes the oscillation mode within a specific frequency band of the original vibration signal, corresponding to different physical sources. For example, the low-frequency EMF characterizes the main pump rotation frequency and the rigid body motion of the pipeline support structure; the mid-frequency EMF characterizes the bending vibration mode of the pipeline shell; and the high-frequency EMF covers the response components induced by local flow changes at the blockage edge and various noise interferences. The center frequency corresponding to each EMF represents the dominant oscillation frequency where the energy of that component is most concentrated in the frequency domain.

[0050] Among them, can be defined The first variational mode decomposition yields the... The time series of the intrinsic mode functions is a time... A function. To show how it changes over time, we define... The first variational mode decomposition yields the... Each intrinsic mode function characterizes the oscillation mode of a specific frequency band in the original vibration signal; This represents a discrete-time index.

[0051] b - Obtain key mode components from intrinsic mode components, and perform second variational mode decomposition within the center frequency range of the key mode components to obtain the corresponding key sub-components.

[0052] After the first variational mode decomposition is completed, key modal components to guide the second decomposition need to be selected from the obtained K1 eigenmode functions to obtain more refined vibrational sub-components. To achieve this selection, embodiments of the present invention quantify the characterization ability of each component for pipeline blockage characteristics by calculating a discriminant score. This score comprehensively considers two aspects: on the one hand, it measures the amount of shared information between the component and the original vibration signal through mutual information to assess the degree to which the component retains the main information of the original signal; on the other hand, it assesses the response sensitivity to blockage-related impact characteristics by calculating the kurtosis of the component's instantaneous amplitude: local flow changes such as eddy shedding and jet impact induced by blockage usually produce transient impact responses with high kurtosis values.

[0053] Among them, the key modal components of the second decomposition guided by discriminative scoring can be expressed as:

[0054] In the formula, Indicates the first The discriminant score of each component is used to quantify the component's ability to represent the characteristics of pipe blockage; the higher the score, the stronger the representation ability. This represents the mutual information calculation function, used to calculate the amount of shared information between two sequences. The larger the value, the stronger the statistical dependence between the two sequences. This represents the kurtosis calculation function, used to measure the peak level of the instantaneous amplitude distribution of a signal. Impact vibrations caused by blockages typically produce high kurtosis values. This represents the Hilbert transform operator, used to obtain the envelope of the analyzed signal, thereby extracting the instantaneous amplitude information of the signal. This represents the balancing weights used to adjust the mutual information and kurtosis terms in the discriminative score. The relative importance of is a hyperparameter, and its value range is . Example of a value: . The first variational mode decomposition yields the... Time series of intrinsic mode functions. The index of the first-level decomposition component represents the sequential number of the eigenmode functions generated by the first-level variational mode decomposition, and its value range is [value range missing]. ; This represents the total number of eigenmode functions obtained from the first variational mode decomposition. The original vibration signal is a one-dimensional time-series data acquired by a vibration sensor, with a length of [missing information]. It contains vibration information of the pipeline under various conditions, with a dimension of 8192.

[0055] Furthermore, after obtaining the discriminative scores for each component, the highest score is selected. Each component is considered a key component, forming a key component set, represented as:

[0056] In the formula, Represents the set of key components, containing An intrinsic mode function selected from the first decomposition result is used to guide the second decomposition; The index represents the key component's sequential number within the key component set, and its value ranges from 1 to 2. ; This indicates the number of key components selected; it is a preset hyperparameter, and examples of possible values ​​are provided. .

[0057] Furthermore, a frequency band is defined with the center frequency corresponding to each key component as the center, expressed as:

[0058] In the formula, Indicates the first The center frequency corresponding to the r-th key component in the first variational mode decomposition, in Hertz, represents the dominant oscillation frequency of the component. It is located in the frequency band where the signal energy is concentrated and is related to the vibration mode caused by blockage. Indicates the first The center frequency range corresponding to each key component will be used as the focusing frequency band for the second decomposition, in Hertz. This represents the half-width of the frequency range, used to define the extended range around the center frequency. It is a preset hyperparameter, measured in Hertz (Hz). Example values ​​are provided. hertz.

[0059] Based on the above definitions, for each key component, a second variational mode decomposition is performed within its corresponding center frequency range to further remove potentially aliased subtle modes within that frequency band, thereby obtaining more refined vibrator sub-components. This represents the set of subcomponents obtained after performing a second-level variational mode decomposition on the r-th critical component (i.e., the critical subcomponents of the second-level variational mode decomposition). Each subcomponent is a function of time t. This indicates that the second decomposition yields a value belonging to the first level. The first key component The quantity is time. The function characterizes more refined vibration modes within the key frequency band; This represents the index of the subcomponent in the second decomposition result, indicating the relationship to the first subcomponent. The sequential numbering of the sub-components generated after the second decomposition of the key components, with values ​​ranging from 1 to 2. ; Indicates the first The total number of subcomponents obtained after the second decomposition of the key components.

[0060] c- The intrinsic modal components and key sub-components are combined to form the set of vibration components corresponding to the pipeline vibration signal.

[0061] Figure 2 A schematic flowchart illustrating the process of determining the discriminative signal according to an embodiment of the present invention is shown. Figure 2 As shown, the original vibration signal is first decomposed into multiple intrinsic mode functions by the first variational mode decomposition (VMD); then the discriminant score of each component is calculated, and key components are selected based on the score results; the center frequency range of the key components is used as the band of interest to perform the second variational mode decomposition to obtain more refined vibration sub-components; on this basis, the comprehensive discriminant index of all components is calculated and adaptive weighted recombination is performed, and finally the enhanced preprocessed signal (i.e. discriminant features) is output.

[0062] Step S208: Perform feature fusion on the discriminative features and the time-invariant features from multiple perspectives to obtain a global feature vector characterizing the fluid state changes of the target pipeline.

[0063] After constructing the discriminative features, directly inputting the high-dimensional enhanced preprocessed signal into the network would result in a heavy computational burden and easily lead to overfitting. Conversely, using conventional manual statistical features would result in the loss of a significant amount of temporal structure information and make it difficult to capture the dynamic patterns of congestion-related signals in different transform domains. This invention also extracts local features with time-translation invariance from three perspectives: time-domain waveform, frequency-domain envelope, and time-frequency ridge. These features are then fused and aggregated across perspectives using a multi-head self-attention mechanism, encoding the high-dimensional time-series signal into a compact global feature vector. The specific steps are as follows: 1) The discriminative features are concatenated along the feature dimension of the time-invariant features from multiple perspectives to generate a multi-perspective fusion matrix.

[0064] Among them, the time-invariant features of the discriminative features in the multi-view perspective include: time-domain waveform features, frequency-domain envelope features, and time-frequency ridge features in the time-frequency domain; the time-domain waveform features are determined based on the local difference and integration results of the discriminative features; the frequency-domain envelope features are determined based on the energy center of the multi-resolution frequency domain analysis; and the time-frequency ridge features are determined based on the energy concentration trajectory of the high-resolution time-frequency transform.

[0065] a- Regarding time-domain waveform characteristics, the discriminative features mentioned above (i.e.) can be calculated. The first-order difference sequence and the local window absolute value integral sequence are concatenated along the feature channel dimension to form a time-domain feature matrix, which captures the local rate of change and cumulative energy information of the vibration waveform. This is defined as... Represents the time-domain feature matrix, with dimension . Its composition is .

[0066] in, Represents a first-order difference sequence with dimension . The first difference sequence element It characterizes the instantaneous rate of change of a signal at each discrete time point, which helps to capture transient features such as shocks; Indicates enhanced preprocessing signal At discrete time points The signal amplitude at that location; Indicates enhanced preprocessing signal At discrete time points The signal amplitude at that location; This represents the discrete-time index, indicating the position of a data point in the signal sequence, with a value range of [value range missing]. ; This indicates the length of the enhanced preprocessed signal, determined based on the original vibration signal. ; Represents a local integral sequence with dimension . The first local integral sequence element , used to characterize the signal at the first The accumulated energy within a local time window centered on a point; Indicates enhanced preprocessing signal At discrete time points The signal amplitude at that location; This represents the size of the integration window, a preset positive integer hyperparameter, with examples of possible values. ; Indicates difference from Discrete time point index.

[0067] b - Regarding the frequency domain envelope characteristics in the frequency domain, the signal can be analyzed... Multi-resolution short-time Fourier transform is performed, and the spectral centroid of each time frame is calculated independently for each resolution. The spectral centroid sequences of all resolutions are aligned to the same length through linear interpolation and then concatenated to form a frequency domain feature matrix, which characterizes the change in the distribution center of signal energy along the frequency axis over time. This matrix is ​​defined as follows: Represents the frequency domain feature matrix, with dimension . Each feature channel is a sequence of spectral centroids at a given resolution, reflecting the time-varying characteristics of the signal's dominant frequency. Its composition is as follows: .

[0068] in, This represents the set of window lengths, specifically the set of window lengths used for the short-time Fourier transform. It is a set of hyperparameters, for example... To cover multi-resolution analysis, Represents the set of window lengths The number of elements in the middle; Indicates the use of the window length set. element Calculated spectral centroid sequence Aligned to length via linear interpolation The result afterward; Indicates the use of the window length set. element The calculated spectral centroid sequence, its t-th element Characterization in the The spectral centroid calculated over each time frame, physically representing the center of the frequency distribution of signal energy within that time frame, is calculated as follows: ; This represents the index in the window length set, and signifies the window length set. The sequential numbering of the elements, with a value range of 100. ; Indicates the use of the window length set. element Enhanced preprocessing of signals After performing a short-time Fourier transform, in time frames and frequency Complex spectrum value at that location, express The amplitude; Indicates frequency index; This indicates the lower limit of the frequency range considered when calculating the spectral centroid; This indicates the upper limit of the frequency range considered when calculating the spectral centroid; This represents an extremely small positive number, used to prevent the denominator from being zero and to ensure numerical stability. Examples of its values ​​are shown below. In practical implementation, and These are preset hyperparameters; for example, if the sampling frequency is 1000Hz, then... , .

[0069] c- For the time-frequency ridge characteristics in the time-frequency domain, the signal can be... Continuous wavelet transform and synchronous compression transform are performed to extract several frequency points with the highest energy and their corresponding energy values ​​at each time point, constructing a time-frequency ridge feature matrix to accurately locate the energy-concentrated ridge patterns on the time-frequency plane, defining... This represents the time-frequency ridge feature matrix, with dimension 1. The column vectors are arranged alternately. A frequency point sequence and A sequence of energy values ​​characterizes the energy concentration trajectory of the signal in the time-frequency domain, and its composition is as follows: .

[0070] in, This represents the number of frequency-energy pairs to be extracted at each time point. It is a preset hyperparameter, and examples of its values ​​are provided. ; This represents the index of the frequency-energy pair extracted at each time point, indicating the order sorted by energy size, with a value range of [value range missing]. ; Represents the extraction of the first time point at all time points. The sequence consisting of the nth frequency points, its nth frequency point element Characterized at time points Based on high-resolution time spectrum After sorting by energy size, the first Large frequency points; Represents the extraction of the first time point at all time points. A sequence consisting of energy values, the first of which is... element Indicates a point in time At frequency point The corresponding energy value, i.e. ; This indicates the enhancement of preprocessed signals. The high-resolution time spectrum obtained by performing continuous wavelet transform followed by synchronous compression transform has better time-frequency focusing than ordinary wavelet transform. Specifically, the enhanced preprocessed signal is first calculated. The wavelet coefficients are obtained by continuous wavelet transform, and then rearranged along the frequency axis through synchronous compression to obtain a high-resolution time spectrum with better focusing. .

[0071] This can be achieved by concatenating the feature matrices from the time domain, frequency domain, and time-frequency domain along the feature channel dimension to obtain a comprehensive feature matrix. For example, ,Right now , dimension ,in, Represents the comprehensive feature matrix The total feature dimension is calculated as follows: .

[0072] 2) The multi-head self-attention mechanism is used to enhance the attention of the multi-view fusion matrix to obtain the target feature matrix.

[0073] 3) Adaptive pooling aggregation is performed on the target feature matrix to obtain a global feature vector characterizing the fluid state changes in the target pipeline.

[0074] Multi-head self-attention mechanisms can capture long-range dependencies between features at different time points and from different perspectives in a comprehensive feature matrix. By computing multiple attention heads in parallel and focusing on the correlation patterns between features from different subspaces, the response of key features is enhanced while suppressing interference from irrelevant information, resulting in an attention-enhanced target feature matrix. Since this target feature matrix is ​​still a variable-length sequence, it needs to be aggregated into a fixed-length vector for subsequent recognition. This embodiment of the invention employs an adaptive generalized mean pooling operator to aggregate the target feature matrix. This pooling method includes a learnable pooling parameter that can be optimized end-to-end through backpropagation. It adaptively adjusts the pooling strategy between average pooling and max pooling, ensuring that the aggregation process preserves the overall distribution information of the features while highlighting local salient responses.

[0075] The above process can be represented as:

[0076] In the formula, The global feature vector is obtained after pooling, which integrates multi-view information and has time-shift invariance. This represents a multi-head self-attention mechanism, the processing of which involves synthesizing the feature matrix. By projecting the learnable weight matrix into a query, key, and value matrix, attention scores and outputs are calculated. After parallel computation by multiple heads, the matrices are concatenated and projected to obtain a multi-head self-attention feature matrix. , dimension , its first row element Characterizing the multi-head self-attention feature matrix In the Feature row vectors at each time point; The feature dimension of the multi-head self-attention feature matrix is ​​a network structure hyperparameter. This represents the Adaptive-Generalized-Mean-Pooling operator, which, for each feature channel of the input sequence, computes the sum of all elements in that channel. The arithmetic mean of powers, then take the average of those powers. The power root, where These are learnable scalar parameters that are trained and optimized through backpropagation. It degenerates into average pooling when The time approximates max pooling. Through the above processing, high-dimensional time-series signals can be encoded into a compact global feature vector, which incorporates time-invariant features from multiple perspectives.

[0077] Figure 3 A schematic diagram illustrating the construction process of the global feature vector according to an embodiment of the present invention is shown. Figure 3 As shown, the enhanced preprocessed signal (discriminative features) is input into three parallel processing branches: a time-domain perspective, a frequency-domain perspective, and a time-frequency-domain perspective. The time-domain perspective extracts local rate of change and cumulative energy information through time-domain difference and integral features; the frequency-domain perspective characterizes the distribution center change of signal energy on the frequency axis through multi-resolution spectral centroid features; and the time-frequency-domain perspective accurately locates energy-concentrated ridge patterns on the time-frequency plane through time-frequency ridges and synchronization compression features. The feature vectors output from the three perspectives are concatenated to form a multi-view fusion matrix, which is then subjected to a multi-head self-attention mechanism for cross-view dependency modeling and feature enhancement. Finally, adaptive generalized mean pooling is used to aggregate the matrix into a fixed-length global feature vector.

[0078] Step S210: Based on the global feature vector, identify and judge the blockage status of the target pipeline, and determine the blockage identification result of the target pipeline.

[0079] After constructing the global feature vector, it needs to be further mapped to an interpretable blockage state determination result. The blockage state of nuclear power plant pipelines is essentially a continuous deterioration process; there are no clear boundaries between different blockage degrees, and samples with similar states in the feature space should have geometric continuity. However, standard classifiers treat each state as an independent category, ignoring this inherent manifold constraint. Furthermore, pipeline vibration characteristics are affected by real-time operating conditions (such as flow rate, pressure, and speed), and relying solely on vibration signal features may lead to a decrease in the model's generalization ability when operating conditions drift. To address these issues, this invention constructs a recognition network that integrates state continuity learning and multi-source information adaptive gating. This network uses the global feature vector as its core input, maps it to a latent space with state continuity through manifold regularization, and simultaneously introduces real-time operating parameters and performs adaptive fusion through a gating mechanism. It synchronously predicts the blockage category, state continuity value, and prediction uncertainty through a multi-task output head. The specific steps are as follows: 1) Input the global feature vector into the trained nonlinear projection network to obtain the corresponding continuous state hidden vector.

[0080] To ensure that the geometry of the latent space accurately reflects the physical nature of the continuous evolution of congestion, this embodiment of the invention applies a manifold regularization constraint. This involves randomly sampling training pairs and calculating the proportional relationship between the distance between each pair of samples in the latent space and the difference in their state labels, thus constructing a regularization loss term. This loss term enables the projection network to learn a latent space with state continuity: samples with similar levels of congestion move closer to each other, while samples with significantly different levels of congestion move further apart. Furthermore, the increment of distance is monotonically positively correlated with the increment of state difference, allowing the latent space to accurately depict the continuous evolution trajectory of congestion from normal to mild and from mild to severe. The projection process is represented as follows:

[0081] In the formula, The continuous state latent vector is a low-dimensional representation obtained by projecting the global feature vector through two fully connected layers. Its geometric structure is constrained to reflect the continuous changing trend of the congestion state, and its dimension is [missing information]. ; This represents the weight matrix of the first layer of the projection network, which are trainable parameters with dimension 1. It is used to linearly project high-dimensional global features into an intermediate feature space. This represents the dimension of the output of the first layer of the projection network. It is a hyperparameter of the network structure, with an example value of 64. This represents the bias vector of the first layer of the projection network, which are trainable parameters with dimension . ; This indicates that the activation function of the linear unit is modified to introduce nonlinear transformation capability; This represents the weight matrix of the second layer of the projection network, which are trainable parameters with dimension 1. This is used to map intermediate features to the final continuous state latent space; This represents the bias vector of the second layer of the projection network, which are trainable parameters with dimension . ; Represents the hidden vector of continuous state The dimension is a hyperparameter of the network structure, with an example value of 32.

[0082] This invention employs a manifold regularization loss designed for the latent space, which ensures that the distance between the latent vectors of any pair of samples in the training set is positively correlated with the absolute difference between their state labels. The calculation method is expressed as follows:

[0083] In the formula, This represents the manifold regularization loss, used to constrain the geometry of the latent space of continuous states. This represents the set of sample pairs randomly sampled from all training sample pairs, used to calculate the regularization loss; This represents the index of the first training sample in the sample pair; This represents the index of the second training sample in the sample pair, and ; Indicates the first The continuous hidden state vectors corresponding to each training sample; Indicates the first The continuous hidden state vectors corresponding to each training sample; The L2 norm of a vector is represented, which is the Euclidean distance. Indicates the first Numerical congestion status labels corresponding to each training sample, for example, 0 represents normal, 1 represents slight congestion, and 2 represents severe congestion; Indicates the first Numericalized blocking status labels corresponding to each training sample; This represents a very small positive number, used to prevent the denominator from being zero and to ensure numerical stability. Examples of possible values ​​are shown below. ; represents the target distance scaling factor, which is a trainable positive scalar parameter used to learn the expected ideal distance between two sample points in the latent space when the state labels differ by one unit. Represents a set The number of sample pairs in the middle.

[0084] 2) Determine the blockage state and continuous deterioration degree of the target pipeline based on the position of the continuous latent vector in the continuous latent space constructed by the nonlinear projection network.

[0085] Since the distance between any two points in this latent space is proportional to the difference in the congestion state label, the region where the latent vector of the current sample is located in the space corresponds to its congestion level, and the displacement of this vector along the deterioration direction in the space reflects the continuous evolution of the congestion degree. Through the above mapping and constraints, the global feature vector can be transformed from the original feature space to a continuous state latent space with clear physical meaning.

[0086] 3) Combining the pipeline operating parameters corresponding to the fluid state change process, adaptive gating fusion is performed on the continuous state latent vector to generate a fused feature vector.

[0087] After the continuous-state latent vector is constructed, to further improve the model's robustness under varying operating conditions, this embodiment of the invention introduces a pipeline operating condition parameter vector collected in real time by external sensors, including pipeline speed ratio, flow ratio, pressure ratio, etc., and fuses it with the continuous-state latent vector through an adaptive gating mechanism. The core of this gating mechanism is to avoid the operating condition information from simply interfering with or overwhelming the vibration characteristics. Instead, it adopts an attention-based adaptive gating mechanism to dynamically adjust the fusion ratio of vibration characteristics and operating condition characteristics in each feature dimension.

[0088] The continuous-state latent vector is concatenated with the linearly transformed working condition parameter vector. An adaptive gating vector is then generated through a fully connected layer and a sigmoid activation function. Each element of this vector takes a value between 0 and 1, used to finely control the contribution weights of vibration and working condition information in each dimension. The fusion process is represented as follows:

[0089] In the formula, The fused feature vector is a feature representation that adaptively fuses vibration state information and operating condition information through a gating mechanism, with a dimension of [missing information]. ; This represents an adaptive gating vector, where each element has a value between 0 and 1, used to finely control vibration features along the feature dimension. Characteristics of the transformed working condition The fusion weight, with dimensions as follows: It is determined by both vibration characteristics and operating condition characteristics, and the calculation method is expressed as follows: ; This represents element-wise multiplication, i.e., the Hadamard product; The projected weight matrix for the operating condition parameters is used to linearly transform the vector of operating condition parameters to a weight matrix with the same dimension as the latent vector of the continuous state. It consists of trainable parameters with dimension O(log n). ; The operating condition parameter vector is a vector composed of real-time operating parameters collected by external sensors and normalized, with dimensions of [missing information]. Examples of its elements are pipe speed ratio, flow ratio, and pressure ratio. Optionally, the pipe speed ratio, flow ratio, and pressure ratio can be normalized to eliminate dimensional differences. This represents the dimension of the operating condition parameter vector, with an example value of 3. Examples include pipe speed ratio, flow ratio, and pressure ratio. This represents the Sigmoid activation function, which maps the input to the (0,1) interval; The gating vector weight matrix represents the weight matrix used to generate the gating vector. It consists of trainable parameters with dimension O(n). ; This represents a vector concatenation operation; The gating vector bias vector is the bias vector used to generate the gating vector. It is a trainable parameter with dimension O(n). .

[0090] 4) Perform multi-task prediction on the fused feature vector, output the blockage category, state continuity value and prediction uncertainty of the target pipeline, and form the blockage identification result of the target pipeline.

[0091] After obtaining the fused feature vector, this embodiment of the invention inputs it into the multi-task output head and executes three related but different prediction tasks simultaneously, including congestion state classification, state continuity value regression, and prediction uncertainty estimation, in order to enhance the interpretability of the model and its reliability in safety-critical scenarios. Figure 4 This diagram illustrates a process flow for pipeline blockage identification and multi-task prediction based on global feature vectors, as shown in an embodiment of the present invention. Figure 4 As shown, the global feature vector is first input into a nonlinear projection network consisting of two fully connected layers, mapped to a continuous-state latent vector. The geometry of this latent space is constrained by manifold regularization, making the distance between samples positively correlated with the difference in blockage state labels. Subsequently, the continuous-state latent vector and the real-time operating condition parameter vector are jointly input into an adaptive gating mechanism. The Sigmoid gating vector dynamically adjusts the fusion ratio of vibration features and operating condition features in each dimension to generate a fused feature vector. This fused feature vector is input in parallel into three task output heads: the classification head outputs the probability distribution of normal, slightly blocked, and severely blocked categories after Softmax normalization; the regression head outputs a scalar value reflecting the continuous severity of blockage; and the uncertainty head outputs the model's uncertainty score for the current prediction, used for risk warning or rejection decisions. The calculation methods for each task are as follows: a) The probability distribution of each category output by the congestion status classification header is calculated as follows:

[0092] In the formula, Let the classification probability vector be denoted by the first digit. element Indicates that the sample belongs to the first The predicted probability of a blockage-like state, where the sum of all elements is 1; This represents the Softmax normalization function; The weight matrix representing the classification head is a trainable parameter with dimension 1. ; The bias vector representing the classifier head is a trainable parameter with dimension 1. ; This represents the category index, indicating different levels of congestion status, with a value range of [value range missing]. ; This represents the total number of congestion categories. For example, a value of 3 corresponds to normal, minor congestion, and severe congestion.

[0093] b) The state continuity regression head outputs a continuous scalar, reflecting the continuous severity of the blockage. The calculation method is expressed as follows:

[0094] In the formula, The value representing the state continuity regression is a continuous scalar, theoretically analogous to the class label vector. It is relevant but smoother, and can be used to monitor trends of state degradation; The weight vector representing the regression head consists of trainable parameters of dimension 1. ; The bias scalar representing the regression head is a trainable parameter.

[0095] c) The uncertainty estimation head outputs an uncertainty score for classification prediction, calculated as follows:

[0096] In the formula, The prediction uncertainty score is a scalar between 0 and 1. The larger the value, the more uncertain the model's classification prediction of the sample is. It can be used for risk warning or rejection decisions. The weight vector representing the uncertainty estimation head is a trainable parameter with dimension O(n). ; The bias scalar representing the uncertainty estimation head is a trainable parameter.

[0097] In summary, the classification head outputs the probability distribution of the target pipeline belonging to the normal, slightly blocked, or severely blocked category through Softmax normalization; the regression head outputs a scalar value reflecting the continuous severity of the blockage, which is related to the category label but more smooth and can be used to monitor the trend of state deterioration; the uncertainty estimation head outputs the model's uncertainty score for the current classification prediction, which ranges from 0 to 1, with a higher value indicating greater uncertainty in the model's judgment of the sample. In safety-critical scenarios, when the uncertainty score exceeds a preset threshold, a rejection decision or a manual review warning can be triggered to prevent the model from making incorrect judgments under low confidence. Through the outputs of the above three tasks, the final blockage identification result includes not only blockage category information but also continuous deterioration degree and prediction confidence information, enabling a multi-dimensional comprehensive judgment of the pipeline blockage state.

[0098] Furthermore, the blockage identification result in this embodiment of the invention can be determined based on a pre-built pipeline blockage identification model. Specifically, this embodiment of the invention also performs the following steps to construct the corresponding model: 1) Dynamically update the contribution weight of the training samples in the preset total loss function based on the classification difficulty and category of the training samples.

[0099] In the training process of a nuclear power plant pipeline blockage identification model, a single loss function is insufficient to simultaneously meet the requirements of multi-task learning, state continuity modeling, and focusing on difficult samples. To collaboratively optimize the multi-task objectives in the identification network and effectively address the potential class imbalance and differences in difficulty between blockage samples, this invention employs a hybrid loss function. While supervising the accuracy of classification and regression tasks, it constrains the feature space structure through manifold regularization and introduces an adaptive weighting strategy to focus on difficult-to-classify samples and minority class samples. The total loss is composed of a weighted sum of four parts: classification loss, regression loss, manifold regularization loss, and uncertainty calibration loss, calculated as follows:

[0100] In the formula, The total loss is the objective function used to guide the overall optimization of all trainable parameters of the network. Represents classification loss. The fixed weight coefficient representing the classification loss is a hyperparameter used to balance the relative importance of the losses of different tasks; the example value is 1.0. Indicates regression loss, The fixed weighting coefficients representing the regression loss are hyperparameters, with an example value of 0.5. This represents the manifold regularization loss. The fixed weight coefficient representing the manifold regularization loss is a hyperparameter, with an example value of 0.1. Indicates the calibration loss due to uncertainty. The fixed weighting coefficient representing the uncertainty calibration loss is a hyperparameter, with an example value of 0.05.

[0101] Specifically, the calculation methods for each sub-loss are as follows: a) The classification loss employs a label-smoothed focus loss, which alleviates class imbalance while reducing overfitting and making the training process more focused on difficult-to-classify samples. Specifically, this is achieved by smoothing the true labels (allocating some confidence to non-target classes with a predetermined smoothing factor) to prevent the model from overconfident in the training labels; simultaneously, a focus parameter is introduced to reduce the weight of the loss on easily classified samples, making the model training more focused on difficult-to-classify samples; furthermore, a weight factor is assigned to each class to balance class imbalance. The calculation method is expressed as follows:

[0102] In the formula, Indicates for the first The model predicts the training sample as the nth training sample. The probability of class is the first The classification probability vector of each training sample The Each element is based on the classification probability vector. The calculation method is used to obtain it; Indicates the first The training sample about the th The smoothed label of a class is calculated as follows: ; It is an indicator function that takes the value 1 when the internal condition is true, and 0 otherwise; This represents the label smoothing factor, a hyperparameter used to prevent the model from becoming overconfident in the training labels; an example value is 0.1. Indicates the first The class weight factor, used to balance class imbalance, can be set as the reciprocal of the training sample frequency of that class or adjusted through hyperparameters; This represents the focusing parameter, which is a hyperparameter. It degenerates into cross-entropy loss. The loss of easily classified samples is reduced to make the training focus more on difficult-to-classify samples. The example value is 2.0.

[0103] b) The regression loss uses Huber loss to calculate the prediction error of the state continuity regression value. This loss function is more robust to outliers, and its calculation method is expressed as follows:

[0104] In the formula, Indicates the first The predicted state continuity regression value of each training sample, based on the state continuity regression value The calculation method is used to obtain it; Indicates the first Numericalized blocking status labels corresponding to each training sample; Represents the Huber loss function; This is the threshold parameter for Huber loss, a hyperparameter used to control the switching point between squared loss and absolute loss, with an example value of 1.0.

[0105] c) The manifold regularization loss, as described in the above embodiment, is used to constrain the geometry of the latent space of continuous states, so that the distance between samples in the latent space is positively correlated with the difference between their clogging state labels.

[0106] d) Uncertainty calibration loss encourages the model to assign higher uncertainty scores when misclassifying and lower uncertainty scores when correctly classifying, thereby enhancing the reliability of uncertainty estimation. The calculation method is expressed as follows:

[0107] In the formula, Indicates the first The predicted class of each training sample, i.e. ; Indicates the first The prediction uncertainty score of each training sample, based on the prediction uncertainty score The calculation method is used to obtain it; This represents a logarithmic function, with the default base being the natural constant.

[0108] Furthermore, in each training iteration, this embodiment of the invention dynamically adjusts the contribution weight of a sample to the total loss based on its current classification difficulty and class, thereby further optimizing the training process. Specifically, the adaptive weight of each sample is updated based on its classification difficulty and class. For example, the lower the model's prediction probability of a sample's true class (i.e., the greater the classification difficulty), the greater the weight of that sample. Simultaneously, samples belonging to minority classes receive higher base weights to alleviate class imbalance. Through this adaptive weighting strategy, the training process can dynamically focus on samples that are difficult to classify at the current stage and minority class samples, further improving the model's recognition performance in class imbalance scenarios. This strategy is expressed as:

[0109] In the formula, This represents the total loss function weighted by the sample weights, used to perform backpropagation to optimize the network parameters; Indicates the first The sample at the th In the training rounds, the adaptive weights for each sample are updated based on their classification difficulty and category. The calculation method is expressed as follows: and initial value ; Indicates the index of the current training round; The momentum coefficient is a hyperparameter used to smooth the weight update process, and its value ranges from [value missing]. The example value is 0.9; Indicates the first The sample at the th Adaptive weights in round training; Indicates the first The basic class weights are used to alleviate class imbalance and can be inversely proportional to class frequency or set as adjustable hyperparameters. The model represents the first Each sample's true category The predicted probability, This reflects the classification difficulty of the sample under the current model; the larger the value, the more difficult it is to classify. Indicates the first The unweighted total loss value calculated from each sample.

[0110] In one embodiment, the present invention also evaluates the performance robustness of different identification methods under different noise levels. Vibration signals from nuclear power plant sites typically contain significant background noise, therefore the method's anti-interference capability is crucial. (Refer to...) Figure 5 This diagram illustrates a performance stability comparison of the present invention's embodiment with various comparative methods under different noise levels. The experiment compares the conventional Fast Fourier Transform (FFT), wavelet transform (WFT), support vector machine (SVM), and the method of this invention. In the experimental configuration, different intensities of random noise were added to the vibration signal to construct four test conditions: high SNR, medium SNR, low SNR, and extremely low SNR. The evaluation metric is the F1 score, and the experimental results are presented as grouped bar charts, with the horizontal axis representing the four SNR levels and the vertical axis representing the F1 score. The experimental results show that the bar representing the method of this invention has the highest height at all SNR levels. The experimental results demonstrate that the signal preprocessing technique based on discriminative component reconstruction in the method of this invention can effectively suppress noise and enhance useful signals. Simultaneously, the multi-source information adaptive gating mechanism can help the model make more robust decisions under harsh conditions, thereby significantly improving the model's generalization ability and robustness in high-noise environments.

[0111] 2) The pipeline blockage identification model is trained based on the total loss function after adjusting the contribution weights.

[0112] Training the nuclear power plant pipeline blockage identification model is an iterative optimization process to converge the model parameters. Training is based on the training and validation sets constructed using the aforementioned data. Before training the nuclear power plant pipeline blockage identification model, a dataset for model learning and evaluation needs to be pre-constructed. This dataset is based on vibration signals collected from nuclear power plant pipelines under known and defined conditions. Specifically, these conditions include normal, unblocked states, as well as simulated blockage states of varying degrees, artificially set. For example, throttle valves with different openings or baffles of different sizes are installed in the pipeline to simulate different levels of blockage, such as minor and severe blockages. For each data acquisition, real-time pipeline operating parameters measured by other sensors at the corresponding time are simultaneously recorded, including pipeline speed ratio, flow ratio, and pressure ratio. These parameters will be used as multi-source information for model training.

[0113] After data collection is completed, the large number of raw samples collected are labeled, with the labeled categories corresponding to the actual physical state of the pipeline. In one implementation, the pipeline state is divided into three categories: normal state (no blockage), slightly blocked state, and severely blocked state. Each sample is assigned a category label, and the category label is numerically encoded, for example, normal is encoded as 0, slightly blocked as 1, and severely blocked as 2, thereby forming a category label vector.

[0114] Subsequently, all labeled samples can be randomly divided into three mutually exclusive sets: a training set, a validation set, and a test set. The training set is used for learning and training the model parameters; the validation set is used to monitor model performance, adjust hyperparameters, and implement early stopping strategies to prevent overfitting during training; and the test set is used to finally evaluate the generalization ability of the trained model. Based on the constructed training and validation sets, the training process for the nuclear power plant pipeline blockage identification model can be executed.

[0115] At the start of training, all trainable parameters in the model are initialized. In each training iteration, samples are drawn in batches from the training set, and the original vibration signal of each sample, along with the corresponding real-time operating condition parameter vector, is input into the model. The forward propagation process of the model performs the following operations in sequence: signal preprocessing based on dual variational mode decomposition and discriminative component reconstruction to obtain an enhanced preprocessed signal; then, a global feature vector is extracted through a multi-view time-invariant feature encoding module; finally, a recognition network that integrates state continuity learning and multi-source information adaptive gating outputs a classification probability vector, a state continuity regression value, and a prediction uncertainty score. Then, based on the model output and the true labels of the samples, the defined total loss function is calculated, and the gradient of the total loss with respect to all trainable parameters is calculated using the backpropagation algorithm. The Adam optimizer is then used to update the parameters based on the gradient to minimize the total loss.

[0116] After each training round, the current model is evaluated using a validation set, and performance metrics such as loss and classification accuracy on the validation set are calculated. Training continues until a preset stopping condition is met. The stopping condition is typically set as follows: the loss on the validation set no longer decreases over several consecutive training rounds (e.g., 10 rounds), or the preset maximum number of training rounds is reached. When training stops, the model parameter state with the best performance on the validation set (e.g., lowest overall loss or highest classification accuracy) is selected and saved as the final trained nuclear power plant pipeline blockage identification model for use in actual identification tasks.

[0117] In summary, this invention proposes a method for identifying pipeline blockages in nuclear power plants based on vibration monitoring, which is innovative in the following aspects compared to existing technologies: 1. By employing dual variational mode decomposition to perform two rounds of fine decomposition and adaptive reconstruction of pipeline vibration signals, the signal components related to blockage can be effectively enhanced, while background noise and irrelevant modes are suppressed, thereby improving the accuracy and robustness of signal preprocessing.

[0118] 2. A multi-view temporally invariant feature encoding module is adopted to extract local features from three dimensions: time domain, frequency domain, and time-frequency domain. A multi-head self-attention mechanism is used for cross-view fusion, which ensures the compactness of global features and the accurate capture of dynamic patterns.

[0119] 3. By using manifold regularization constraints, the continuity of the pipeline blockage state in the latent space is ensured, and subtle changes in the blockage process are captured, thereby improving the model's ability to identify the evolution of the blockage state.

[0120] 4. A multi-task learning network is adopted, which not only classifies the congestion state, but also predicts the continuity of the congestion state through regression tasks, and outputs the prediction uncertainty score, thereby improving the interpretability of the model and its reliability in safety-critical scenarios.

[0121] Furthermore, based on the above embodiments, this invention also provides a nuclear power plant pipeline blockage identification device based on vibration monitoring, referring to... Figure 6 The device includes: a signal acquisition module 10 for acquiring the pipe vibration signal of the target pipe during the fluid state change process; a preprocessing module 20 for performing discriminative component reconstruction on the pipe vibration signal based on the set of vibration components contained in the pipe vibration signal to construct the discriminative features corresponding to the pipe vibration signal; a data processing module 30 for performing feature fusion on the time-invariant features of the discriminative features from multiple perspectives to obtain a global feature vector characterizing the fluid state change of the target pipe; and a prediction module 40 for identifying and judging the blockage state of the target pipe based on the global feature vector to determine the blockage identification result of the target pipe.

[0122] The nuclear power plant pipeline blockage identification device based on vibration monitoring provided in this embodiment of the invention has the same implementation principle and technical effect as the aforementioned method embodiment. For the sake of brevity, any parts not mentioned in the device embodiment can be referred to the corresponding content in the aforementioned method embodiment.

[0123] This invention also provides an electronic device, including a memory, a processor, and a computer program stored in the memory and executable on the processor. When the processor executes the computer program, it implements the above-described... Figure 1 The steps of the method are shown. Embodiments of the present invention also provide a computer-readable storage medium storing a computer program, which, when executed by a processor, performs the above-described steps. Figure 1 The steps of the method are shown. This invention also provides a schematic diagram of the structure of an electronic device, as shown. Figure 7 The diagram shows the structure of the electronic device, which includes a processor 101 and a memory 100. The memory 100 stores computer-executable instructions that can be executed by the processor 101. The processor 101 executes the computer-executable instructions to implement the above-mentioned... Figure 1 The method shown.

[0124] exist Figure 7In the illustrated embodiment, the electronic device further includes a bus 102 and a communication interface 103, wherein the processor 101, the communication interface 103, and the memory 100 are connected via the bus 102. The memory 100 may include high-speed random access memory (RAM) or non-volatile memory, such as at least one disk drive. Communication between this system network element and at least one other network element is achieved through at least one communication interface 103 (which can be wired or wireless), using the Internet, wide area network, local area network, metropolitan area network, etc. Bus 102 can be an ISA (Industry Standard Architecture) bus, a PCI (Peripheral Component Interconnect) bus, or an EISA (Extended Industry Standard Architecture) bus, or an AMBA (Advanced Microcontroller Bus Architecture) bus. AMBA defines three types of buses: APB (Advanced Peripheral Bus), AHB (Advanced High-performance Bus), and AXI (Advanced eXtensible Interface). Bus 102 can be divided into address bus, data bus, control bus, etc. For ease of representation, Figure 7The diagram uses only a single double-headed arrow, but this does not imply a single bus or a single type of bus. Processor 101 may be an integrated circuit chip with signal processing capabilities. In implementation, each step of the above method can be completed by the integrated logic circuitry in the hardware of processor 101 or by instructions in software form. Processor 101 can be a general-purpose processor, including a Central Processing Unit (CPU), a Network Processor (NP), etc.; it can also be a Digital Signal Processor (DSP), an Application Specific Integrated Circuit (ASIC), a Field-Programmable Gate Array (FPGA), or other programmable logic devices, discrete gate or transistor logic devices, or discrete hardware components. The general-purpose processor can be a microprocessor or any conventional processor. The steps of the method disclosed in the embodiments of this invention can be directly manifested as execution by a hardware decoding processor, or execution by a combination of hardware and software modules in the decoding processor. The software module can reside in a mature storage medium in the field, such as random access memory, flash memory, read-only memory, programmable read-only memory, electrically erasable programmable memory, or registers. This storage medium is located in memory, and the processor 101 reads information from the memory and, in conjunction with its hardware, completes the aforementioned tasks. Figure 1 The method shown.

[0125] The computer program product of the method and device for identifying pipeline blockage in nuclear power plants based on vibration monitoring provided in this invention includes a computer-readable storage medium storing program code. The instructions included in the program code can be used to execute the methods described in the preceding method embodiments. For specific implementation, please refer to the method embodiments, which will not be repeated here. Those skilled in the art will clearly understand that, for the sake of convenience and brevity, the specific working process of the system described above can be referred to the corresponding process in the preceding method embodiments, which will not be repeated here. If the function is implemented in the form of a software functional unit and sold or used as an independent product, it can be stored in a computer-readable storage medium. Based on this understanding, the technical solution of this invention, in essence, or the part that contributes to the prior art, or part of the technical solution, can be embodied in the form of a software product. This computer software product is stored in a storage medium and includes several instructions to cause a computer device (which may be a personal computer, a server, or a network device, etc.) to execute all or part of the steps of the methods described in the various embodiments of this invention. The aforementioned storage medium includes various media capable of storing program code, such as USB flash drives, mobile hard drives, read-only memory (ROM), random access memory (RAM), magnetic disks, or optical disks.

[0126] In the description of this invention, it should be noted that the terms "first," "second," and "third" are used for descriptive purposes only and should not be construed as indicating or implying relative importance. Finally, it should be noted that the above embodiments are merely specific implementations of this invention, used to illustrate the technical solutions of this invention, and not to limit it. The scope of protection of this invention is not limited thereto. Although this invention has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that any person skilled in the art can still modify or easily conceive of changes to the technical solutions described in the foregoing embodiments within the technical scope disclosed in this invention, or make equivalent substitutions for some of the technical features; and these modifications, changes, or substitutions do not cause the essence of the corresponding technical solutions to deviate from the spirit and scope of the technical solutions of the embodiments of this invention, and should all be covered within the scope of protection of this invention. Therefore, the scope of protection of this invention should be determined by the scope of the claims.

Claims

1. A method for identifying pipeline blockage in nuclear power plants based on vibration monitoring, characterized in that, The method includes: Acquire the pipe vibration signal of the target pipe during the fluid state change process; Based on the set of vibration components contained in the pipeline vibration signal, discriminative component reconstruction is performed on the pipeline vibration signal to construct the discriminative features corresponding to the pipeline vibration signal; The discriminative features are fused across multiple perspectives to obtain a global feature vector characterizing the fluid state changes of the target pipeline. This process includes: concatenating the time-invariant features of the discriminative features along the feature dimension to generate a multi-perspective fusion matrix; applying a multi-head self-attention mechanism to the multi-perspective fusion matrix to enhance attention, resulting in a target feature matrix; and performing adaptive pooling aggregation on the target feature matrix to obtain a global feature vector characterizing the fluid state changes of the target pipeline. The time-invariant features of the discriminative features across multiple perspectives include: time-domain waveform features, frequency-domain envelope features, and time-frequency ridge features in the time-frequency domain. Based on the global feature vector, the blockage state of the target pipeline is identified and judged to determine the blockage identification result of the target pipeline, including: inputting the global feature vector into a trained nonlinear projection network to map the corresponding continuous state latent vector; determining the blockage state and its continuous deterioration degree of the target pipeline according to the position of the continuous state latent vector in the continuous latent space constructed by the nonlinear projection network; combining the pipeline operating parameters corresponding to the fluid state change process, performing adaptive gating fusion on the continuous state latent vector to generate a fused feature vector; performing multi-task prediction on the fused feature vector, and outputting the blockage category, state continuity value and prediction uncertainty of the target pipeline to form the blockage identification result of the target pipeline. The method further includes: The pipeline vibration signal is subjected to first variational mode decomposition to obtain multiple intrinsic mode components; Key mode components are obtained from the intrinsic mode components, and a second variational mode decomposition is performed within the center frequency range of the key mode components to obtain the corresponding key sub-components. The intrinsic modal components and the key sub-components are combined to form the vibration component set corresponding to the pipeline vibration signal.

2. The method according to claim 1, characterized in that, The step of constructing discriminative features corresponding to the pipeline vibration signal by performing discriminative component reconstruction on the set of vibration components contained in the pipeline vibration signal includes: Based on the preset discriminant index corresponding to each sub-component in the vibration component set, the adaptive weight corresponding to each sub-component is determined; wherein, the preset discriminant index is determined based on the kurtosis characteristics of the corresponding sub-component under different blockage states; Using the adaptive weights, the signal components of the vibration component set are adaptively weighted and recombined to construct the discriminative features corresponding to the pipeline vibration signal.

3. The method according to claim 1, characterized in that, The time-domain waveform features are determined based on the local difference and integration results of the discriminative features; The frequency domain envelope features are determined based on the energy center of multi-resolution frequency domain analysis; The time-frequency ridge features are determined based on the energy concentration trajectory of high-resolution time-frequency transformation.

4. The method according to claim 1, characterized in that, The blockage identification result is determined based on a pre-built pipeline blockage identification model; The method further includes: The contribution weight of the training sample in the preset total loss function is dynamically updated based on the classification difficulty and category of the training sample; wherein, the total loss function is composed of a weighted sum of classification loss, regression loss, manifold regularization loss and uncertainty calibration loss; The pipeline blockage identification model is trained based on the total loss function adjusted by contribution weights.

5. A nuclear power plant pipeline blockage identification device based on vibration monitoring, characterized in that, The device is used to implement the nuclear power plant pipeline blockage identification method based on vibration monitoring as described in any one of claims 1 to 4, and the device comprises: The signal acquisition module is used to acquire the pipe vibration signal of the target pipe during the process of fluid state change; The preprocessing module is used to perform discriminative component reconstruction on the pipeline vibration signal based on the set of vibration components contained in the pipeline vibration signal, and construct the discriminative features corresponding to the pipeline vibration signal. The data processing module is used to perform feature fusion on the time-invariant features of the discriminative features from multiple perspectives to obtain a global feature vector characterizing the fluid state changes of the target pipeline. This includes: concatenating the time-invariant features of the discriminative features from multiple perspectives along the feature dimension to generate a multi-view fusion matrix; performing attention enhancement processing on the multi-view fusion matrix using a multi-head self-attention mechanism to obtain a target feature matrix; and performing adaptive pooling aggregation on the target feature matrix to obtain a global feature vector characterizing the fluid state changes of the target pipeline. The time-invariant features of the discriminative features from multiple perspectives include: time-domain waveform features in the time domain, frequency-domain envelope features in the frequency domain, and time-frequency ridge features in the time-frequency domain. The prediction module is used to identify and judge the blockage state of the target pipeline based on the global feature vector, and determine the blockage identification result of the target pipeline. This includes: inputting the global feature vector into a trained nonlinear projection network to map it into a corresponding continuous state latent vector; determining the blockage state and its continuous degradation degree of the target pipeline based on the position of the continuous state latent vector in the continuous latent space constructed by the nonlinear projection network; performing adaptive gating fusion on the continuous state latent vector in conjunction with pipeline operating parameters corresponding to the fluid state change process to generate a fused feature vector; performing multi-task prediction on the fused feature vector, and outputting the blockage category, state continuity value, and prediction uncertainty of the target pipeline to form the blockage identification result of the target pipeline.

6. An electronic device, characterized in that, The electronic device includes a processor and a memory, the memory storing machine-executable instructions that can be executed by the processor, the processor executing the machine-executable instructions to implement the nuclear power plant pipeline blockage identification method based on vibration monitoring as described in any one of claims 1 to 4.

7. A machine-readable storage medium, characterized in that, The machine-readable storage medium stores machine-executable instructions, which, when invoked and executed by a processor, cause the processor to implement the nuclear power plant pipeline blockage identification method based on vibration monitoring as described in any one of claims 1 to 4.

Images