A method and system for predicting defects in steam generator heat transfer tubes
By establishing a defect propagation dynamic equation and a leakage rate calculation model, and constructing a competitive expert prediction network, the problems of accuracy and real-time performance in predicting defects in heat transfer tubes of steam generators were solved, and high-precision joint inversion of defect type, location, and geometric parameters was achieved.
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Patents(China)
- Current Assignee / Owner
- SUN YAT SEN UNIV
- Filing Date
- 2026-03-20
- Publication Date
- 2026-06-19
AI Technical Summary
Existing technologies for predicting defects in heat transfer tubes of steam generators suffer from low accuracy, lack of physical interpretability, and inability to achieve high-precision, real-time joint inversion, especially when the leakage rate curves caused by different defect types have subtle and overlapping characteristics.
We establish defect propagation dynamic equations and leakage rate calculation models for various defect types, construct a competitive expert prediction network, and achieve high-precision prediction of defect type, location, and geometric parameters through physical mechanism expert network branches and recoupled decoders. We also combine a deep learning architecture for real-time monitoring.
It achieves high-precision real-time prediction of the type, location, and geometric parameters of defects in the heat transfer tubes of steam generators, conforms to the basic laws of fracture mechanics and tribology, and improves the accuracy and applicability of the prediction.
Smart Images
Figure CN121901653B_ABST
Abstract
Description
Technical Field
[0001] This application relates to the field of structural integrity monitoring, and in particular to a method and system for predicting defects in heat transfer tubes of steam generators. Background Technology
[0002] The heat transfer tubes of the steam generator in a pressurized water reactor nuclear power plant are the main pressure boundary of the primary loop, and they have been operating under complex conditions such as high temperature, high pressure, and fluid vibration for a long time. Engineering practice shows that the failure of heat transfer tubes is mainly driven by three types of typical defects: stress corrosion cracking (SCC), fretting wear, and fatigue cracks, and the evolution dynamics of these defects are significantly different.
[0003] Currently, nuclear power plants primarily acquire time-series data through leakage rate monitoring systems. However, existing technologies have significant shortcomings in predicting heat transfer tube defects: First, the leakage rate curves caused by different defect types exhibit subtle and overlapping characteristics, making it difficult to accurately rely solely on operator experience. Second, existing machine learning methods are mostly data-driven, lacking physical interpretability and prone to producing unauthorized predictions that do not conform to material damage mechanisms. Third, current monitoring schemes cannot achieve high-precision, real-time joint inversion from leakage signals to defect types, locations, and geometric parameters (such as crack length and wear depth). Therefore, there is an urgent need for a prediction scheme that can deeply integrate complex physical dynamics mechanisms with deep learning architectures to improve the safe operation and maintenance of nuclear power plants. Summary of the Invention
[0004] Based on this, it is necessary to provide a method and system for predicting defects in heat transfer tubes of steam generators that can deeply integrate complex physical dynamics mechanisms with deep learning architecture, and which has the characteristics of accurate identification, engineering applicability and strong real-time performance.
[0005] To achieve the above-mentioned objectives of this invention, the technical solution adopted is as follows:
[0006] A method for predicting defects in heat transfer tubes of a steam generator includes the following specific steps:
[0007] For various preset defect types in the heat transfer tubes of steam generators, defect propagation dynamic equations describing the evolution of defect size with operating time and leakage rate calculation models based on defect size are established for each preset defect type.
[0008] Based on the aforementioned defect propagation dynamics equation and leakage rate calculation model, a forward simulation calculation is performed on the defect propagation process of each preset defect type under different initial defect sizes and operating conditions to obtain a time series data training set of leakage rates corresponding to each preset defect type.
[0009] A competitive expert prediction network is constructed. The network includes multiple physical mechanism expert network branches corresponding one-to-one with each preset defect type. Each physical mechanism expert network branch embeds a physical feature encoding layer associated with the defect expansion dynamics equation of the corresponding defect type. Each physical mechanism expert network branch explicitly encodes the physical mechanism feature parameters of the corresponding defect type into intermediate feature representations through the physical feature encoding layer, and outputs its own branch feature vector. The network outputs a defect prediction value based on the branch feature vector.
[0010] The competitive expert prediction network is trained using the aforementioned leakage rate time series data training set as training input.
[0011] Real-time leakage rate monitoring data of the steam generator is acquired and input into the trained competitive expert prediction network, which outputs defect prediction values including defect type and geometric parameters.
[0012] Preferably, the various preset defect types include fatigue cracks, stress corrosion cracks, and fretting wear defects;
[0013] The defect propagation kinetic equation established for fatigue cracks is as follows:
[0014]
[0015] in The length of the crack is half its length. For the cumulative number of cyclic loads, This is a temperature-dependent material coefficient. As an accelerating factor for the water environment, For the range of stress intensity factors, The crack propagation index is... This is the load ratio correction index. The stress ratio; the range of the stress intensity factor. Calculate using the following formula:
[0016]
[0017] in For geometric correction factor, This refers to the range of cyclic load stress.
[0018] The defect propagation kinetic equation established for stress corrosion cracking is as follows:
[0019]
[0020] in For runtime, The power-law baseline coefficient, Crack tip stress intensity factor Influence function, For temperature Influence function, Dissolved hydrogen concentration Influence function, Crack propagation direction angle Influence function;
[0021] The crack tip stress intensity factor Calculate using the following formula:
[0022]
[0023] in For geometric correction factor, It is static stress;
[0024] The kinetic equation for the propagation of fretting wear defects is as follows:
[0025]
[0026] in For the volume of the wear groove, The wear coefficient is... For contact force, For micro-motion speed, Material hardness;
[0027] Width of wear groove opening and the axial length of the wear groove The relationship is:
[0028]
[0029] The physical feature encoding layer encodes the physical mechanism feature parameters corresponding to the above formulas into components of the corresponding branch feature vectors.
[0030] Furthermore, the leakage rate calculation model includes a defect opening area model and an orifice flow model; the defect opening area model calculates the current defect opening area according to the defect type using the following formulas. :
[0031] For fatigue cracks, their opening area for:
[0032]
[0033] in For elastic modulus, This is the shell geometry correction factor;
[0034] For stress corrosion cracking, its effective opening area for:
[0035]
[0036] in This is a semi-closed correction factor;
[0037] For fretting wear defects, the equivalent leakage area for:
[0038]
[0039] Orifice flow model for calculating instantaneous leakage rate The formula is:
[0040]
[0041] in The flow coefficient, The voltage difference between the primary and secondary loops. This refers to the density of the primary coolant.
[0042] Furthermore, the competitive expert prediction network also includes a shared temporal feature extractor, a gating network, and a recoupled decoder;
[0043] The shared temporal feature extractor takes the leakage rate time series data in the obtained leakage rate time series data training set as input, extracts the shared feature vector, and simultaneously inputs the shared feature vector into the multiple physical mechanism expert network branches and the gating network;
[0044] The gated network takes the shared feature vector as input and outputs the competitive activation weights of each physical mechanism expert network branch;
[0045] The recoupled decoder takes the branch feature vectors output by each of the physical mechanism expert network branches as input, performs weighted fusion of the branch feature vectors according to the competitive activation weights, and outputs the defect type classification result, defect location classification result, and defect prediction value of geometric parameters.
[0046] Furthermore, the loss function used to train the network is a composite physical constraint loss function:
[0047] :
[0048] in The data fitting loss includes the cross-entropy loss calculated on the defect type classification result output by the recoupled decoder, the cross-entropy loss calculated on the defect location classification result, and the normalized mean square error loss calculated on the defect prediction value of the geometric parameters. The physical residual loss is obtained by substituting the defect prediction values of the geometric parameters output by the recoupled decoder into the corresponding defect propagation dynamics equation and calculating the differential residual. For physical consistency loss; weighting coefficients It adapts and adjusts itself during training.
[0049] Furthermore, the physical residual loss Calculated in the following way:
[0050] Substitute the predicted values of the geometric parameters into the finite difference form of the defect propagation dynamics equation for the corresponding defect type, and calculate the difference margin at each time step.
[0051] For fatigue cracks, the differential margin for:
[0052]
[0053] in for The crack half-length prediction value output by the recoupled decoder at time [time]. This represents the increment of the number of cyclic loads corresponding to adjacent time points. These are predicted values for the stress intensity factor range. This serves as a reference value for the normalized crack length.
[0054] For stress corrosion cracking, the differential allowance for:
[0055]
[0056] in For time step, This is the predicted value of the crack tip stress intensity factor. This is a normalized reference value for the stress corrosion cracking propagation rate.
[0057] For fretting wear defects, the differential margin for:
[0058]
[0059] in This refers to the predicted wear volume change rate output by the physical mechanism expert network branch corresponding to the fretting wear defect. The normalized reference value for the wear volume change rate; the physical residual loss The remainder is the mean of each difference over all training samples and all time points.
[0060] Furthermore, the competitive expert prediction network also includes a dynamic flow coefficient prediction subnetwork that works in conjunction with the recoupled decoder;
[0061] The dynamic flow coefficient prediction subnetwork takes the current leakage rate time series data and its time derivative, the defect type classification result currently output by the recoupled decoder, the defect prediction value of geometric parameters, and the operating condition parameters as input, and outputs the dynamic flow coefficient prediction value. The predicted value of the dynamic flow coefficient The calculation method is as follows:
[0062]
[0063] in The linear output of the dynamic flow coefficient prediction subnetwork is given. and The lower and upper bounds of the physical constraint intervals of the flow coefficient corresponding to the defect type classification results are determined by physical experience based on the defect type.
[0064] The dynamic flow coefficient prediction subnetwork and the competing expert prediction network are separated by physical consistency loss. Joint optimization, its formula is:
[0065]
[0066] in The number of samples in the training batch. To predict the leakage rate, The voltage difference between the primary and secondary loops. The current defect opening area is predicted by the network. This is a normalized reference value for the leakage rate.
[0067] Furthermore, the shared temporal feature extractor integrates a crack propagation stage perception attention mechanism: the stage division auxiliary classifier divides the defect propagation process into the initiation stage, stable propagation stage, rapid propagation stage, and through-stage stage based on the first and second derivative features of the leakage rate temporal data; the stage adaptive attention module configures different attention bias priors based on the division results and the magnitude of the defect propagation rate of each stage estimated by the corresponding physical mechanism model.
[0068] Furthermore, the specific architecture of the competitive expert prediction network is as follows: the shared temporal feature extractor adopts a temporal convolutional network with dilated causal convolution structure; the gated network obtains the competitive activation weights of each branch after calculation by a multilayer perceptron and processing by a Softmax function. The training objective of the gated network includes gated regularization loss:
[0069]
[0070] This causes the gating network to tend to assign higher weights to the branches that minimize the physical residual loss; The total number of branches in the physical mechanism expert network; the recoupled decoder uses the feature vectors of each branch and their corresponding competitive activation weights as input to calculate the weighted fusion feature vector. This leads to the output of various prediction results.
[0071] A defect prediction system for heat transfer tubes in a steam generator, comprising:
[0072] The physical mechanism modeling module is used to establish, for various preset defect types in the heat transfer tubes of steam generators, defect propagation kinetic equations describing the evolution of defect size over operating time and leakage rate calculation models based on defect size; and based on the aforementioned defect propagation kinetic equations and leakage rate calculation models, forward simulation calculations are performed on the defect propagation process of each preset defect type under different initial defect sizes and operating parameters to obtain a time series data training set of leakage rates corresponding to each preset defect type.
[0073] The prediction network construction module is used to construct a competitive expert prediction network comprising multiple physical mechanism expert network branches corresponding one-to-one with each preset defect type; each physical mechanism expert network branch embeds a physical feature encoding layer associated with the defect expansion dynamics equation of the corresponding defect type; each physical mechanism expert network branch explicitly encodes the physical mechanism feature parameters of the corresponding defect type into intermediate feature representations through the physical feature encoding layer, and outputs its own branch feature vector; the network outputs the defect prediction value based on the branch feature vector;
[0074] Training module: Using the aforementioned leakage rate time series data training set as training input, train the competitive expert prediction network;
[0075] Prediction module: acquires real-time leakage rate monitoring data of the steam generator, inputs it into the trained competitive expert prediction network, and outputs defect prediction values including defect type and geometric parameters.
[0076] The beneficial effects of this invention are as follows:
[0077] This invention establishes corresponding defect propagation kinetic equations for various defect types in heat transfer tubes; generates a training sample set including leakage rate time series; and constructs a competitive expert prediction network comprising multiple expert network branches. By dynamically activating expert branches with different physical backgrounds, it accurately captures subtle feature differences in SCC, fatigue, and wear defects, achieving high-precision joint inversion of defect type, location, and geometric parameters. This yields prediction results for defect type, location, and geometric parameters, realizing deep decoupling and recoupling of physical mechanisms and data-driven approaches. It changes the prediction logic of traditional black-box models, ensuring from an architectural perspective that the prediction results conform to the fundamental laws of fracture mechanics and tribology. It features real-time accuracy and applicability to complex environments. Attached Figure Description
[0078] Figure 1 This is a flowchart of the steam generator heat transfer tube defect prediction method of the present invention in one embodiment.
[0079] Figure 2 This is a block diagram of the steam generator heat transfer tube defect prediction system of the present invention in one embodiment. Detailed Implementation
[0080] To make the objectives, technical solutions, and advantages of this application clearer, the following detailed description is provided in conjunction with the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are merely illustrative and not intended to limit the scope of this application.
[0081] Example 1
[0082] like Figure 1 As shown, this embodiment provides a method for predicting defects in the heat transfer tubes of a steam generator, including the following specific steps:
[0083] For various preset defect types in the heat transfer tubes of steam generators, defect propagation dynamic equations describing the evolution of defect size with operating time and leakage rate calculation models based on defect size are established for each preset defect type.
[0084] The various preset defect types include fatigue cracks, stress corrosion cracks, and fretting wear defects;
[0085] The defect propagation kinetic equation established for fatigue cracks is as follows:
[0086]
[0087] in The length of the crack is half its length. For the cumulative number of cyclic loads, This is a temperature-dependent material coefficient. As an accelerating factor for the water environment, For the range of stress intensity factors, The crack propagation index is... This is the load ratio correction index. The stress ratio; the range of the stress intensity factor. Calculate using the following formula:
[0088]
[0089] in For geometric correction factor, This refers to the range of cyclic load stress.
[0090] The defect propagation kinetic equation established for stress corrosion cracking is as follows:
[0091]
[0092] in For runtime, The power-law baseline coefficient, Crack tip stress intensity factor Influence function, For temperature Influence function, Dissolved hydrogen concentration Influence function, Crack propagation direction angle Influence function;
[0093] The crack tip stress intensity factor Calculate using the following formula:
[0094]
[0095] in For geometric correction factor, It is static stress;
[0096] The kinetic equation for the propagation of fretting wear defects is as follows:
[0097]
[0098] in For the volume of the wear groove, The wear coefficient is... For contact force, For micro-motion speed, Material hardness;
[0099] Width of wear groove opening and the axial length of the wear groove The relationship is:
[0100]
[0101] The physical feature encoding layer encodes the physical mechanism feature parameters corresponding to the above formulas into components of the corresponding branch feature vectors.
[0102] In one specific embodiment, the leakage rate calculation model includes a defect opening area model and an orifice flow model; the defect opening area model calculates the current defect opening area according to the defect type using the following formulas. :
[0103] For fatigue cracks, their opening area for:
[0104]
[0105] in For elastic modulus, This is the shell geometry correction factor;
[0106]
[0107] in The average radius of the heat transfer tube. For the thickness of the heat transfer tube wall;
[0108] For stress corrosion cracking, its effective opening area for:
[0109]
[0110] in This is a semi-closed correction factor;
[0111] For fretting wear defects, the equivalent leakage area for:
[0112]
[0113] Orifice flow model for calculating instantaneous leakage rate The formula is:
[0114]
[0115] in The flow coefficient, The voltage difference between the primary and secondary loops. This refers to the density of the primary coolant.
[0116] In one specific embodiment, the forward simulation calculation is performed separately in multiple preset structural location regions of the steam generator heat transfer tube within the sampling parameter space, with each preset structural location region corresponding to different combinations of operating condition parameters. During this process, the defect size at each time stepwise deduces the defect size based on the defect propagation kinetic equation, and the leakage rate at each time step is calculated from the defect size using the leakage rate calculation model to obtain leakage rate time-series data. The input to the shared temporal feature extractor is the leakage rate temporal data. The first and second time derivatives of the time series data.
[0117] Each sample in the time-series leakage rate training set carries a corresponding defect type label, defect location label, and defect geometric parameter truth value label; the defect location label identifies the preset structural location region corresponding to the sample.
[0118] In one specific embodiment, the pre-defined structural location regions of the steam generator heat transfer tubes and the corresponding differences in their operating parameters are as follows:
[0119] Tube sheet expansion transition zone: This area has a high local stress level due to the superposition of expansion residual stress and operating stress, and there is a crevice corrosion environment on the outside of the tube, making it a high-incidence area for stress corrosion cracking; a high static stress and a low dissolved hydrogen concentration are configured in the simulation.
[0120] The top region of the U-bend: Due to the residual stress generated during pipe bending and the high operating temperature, this region is also prone to stress corrosion cracking, but its geometric correction factor is different from that of the straight pipe section; the corresponding geometric correction factor and temperature parameters are configured during simulation.
[0121] Vibration damping strip contact area: This area is the main location where fretting wear defects occur due to the gap fit between the heat transfer tube and the vibration damping strip and the flow-induced vibration; larger fretting speed and contact force parameters are configured during simulation.
[0122] Straight pipe section free crossing zone: The alternating stress generated by flow-induced vibration in this region is a potential location for fatigue cracks; a large range of cyclic load stress and an incremental number of cycles are configured during simulation.
[0123] By configuring the aforementioned location-related differential chemical condition parameters, the leakage rate time series data generated in each preset structural location region exhibits distinguishable differences in the rise rate, steady-state value level, and fluctuation characteristics of the time series waveform, thus providing an effective training data foundation for the defect location classification task of the competitive expert prediction network.
[0124] Based on the aforementioned defect propagation dynamics equation and leakage rate calculation model, a forward simulation calculation is performed on the defect propagation process of each preset defect type under different initial defect sizes and operating conditions to obtain the time series data training set of leakage rate and the true value label of defect geometric parameters corresponding to each preset defect type.
[0125] Construct a competitive expert prediction network; the network includes a shared temporal feature extractor, multiple physical mechanism expert network branches corresponding one-to-one with the various preset defect types, a gating network, and a recoupled decoder;
[0126] The shared temporal feature extractor takes the leakage rate time series data in the obtained leakage rate time series data training set as input, extracts the shared feature vector, and simultaneously inputs the shared feature vector into the multiple physical mechanism expert network branches and the gating network;
[0127] Each physical mechanism expert network branch embeds a physical feature encoding layer associated with the defect expansion dynamics equation of the corresponding defect type; each physical mechanism expert network branch takes the shared feature vector as input, and through the physical feature encoding layer, explicitly encodes the physical mechanism feature parameters of the corresponding defect type into intermediate feature representations, and outputs its own branch feature vector;
[0128] In this embodiment, the physical feature encoding layer is implemented as follows:
[0129] The physical feature coding layer of each physical mechanism expert network branch contains two functional units: differentiable physical operator units and feature splicing units.
[0130] The differentiable physical operator unit performs differentiable operations on the components corresponding to physical parameters in the shared feature vector according to the physical equations for the corresponding defect types, generating intermediate physical quantities. The differentiable physical operator units of each branch implement the following operations:
[0131] For fatigue crack branches, the differentiable physical operator element extracts components related to the crack half-length a, the cyclic load stress range, and the geometric correction factor F from the shared eigenvector h, and calculates the predicted value of the stress intensity factor range:
[0132]
[0133] Furthermore, the physical intermediate quantity of crack propagation rate was calculated:
[0134]
[0135] For stress corrosion cracking branches, the differentiable physical operator element extracts components related to the crack half-length a, static stress, and geometric correction factor from the shared eigenvector h, and calculates the predicted value of the crack tip stress intensity factor:
[0136]
[0137] Furthermore, the physical intermediate quantity of stress corrosion crack propagation rate was calculated:
[0138]
[0139] For the fretting wear defect branch, the differentiable physical operator unit extracts components related to the contact force W, fretting velocity v, and material hardness H from the shared eigenvector h, and calculates the physical intermediate quantity of the wear volume change rate:
[0140]
[0141] The feature concatenation unit concatenates the physical intermediate vector output by the differentiable physical operator unit with the shared feature vector h along the feature dimension. After linear transformation and activation function processing, it outputs the branch feature vector of that branch:
[0142]
[0143] Through the design of the above physical feature encoding layer, each expert network branch explicitly embeds the physical dynamics knowledge of the corresponding defect type during the feature extraction stage. This makes the branch feature vector not only contain data-driven temporal features, but also structured physical information consistent with the physical equations, thus providing a physically interpretable feature foundation for subsequent gating competition selection and recoupled decoding.
[0144] The gated network takes the shared feature vector as input and outputs the competitive activation weights of each physical mechanism expert network branch;
[0145] The recoupled decoder takes the branch feature vectors output by each of the physical mechanism expert network branches as input, performs weighted fusion of the branch feature vectors according to the competitive activation weights, and outputs defect type classification results, defect location classification results, and defect prediction values of geometric parameters.
[0146] The competitive expert prediction network also includes a dynamic flow coefficient prediction subnetwork that works in conjunction with the recoupled decoder.
[0147] The dynamic flow coefficient prediction subnetwork takes the current leakage rate time series data and its time derivative, the defect type classification result currently output by the recoupled decoder, the defect prediction value of geometric parameters, and the operating condition parameters as input, and outputs the dynamic flow coefficient prediction value. The predicted value of the dynamic flow coefficient The calculation method is as follows:
[0148]
[0149] in The linear output of the dynamic flow coefficient prediction subnetwork is given. and The lower and upper bounds of the physical constraint intervals of the flow coefficient corresponding to the defect type classification results are determined by physical experience based on the defect type.
[0150] Real-time leakage rate monitoring data of the steam generator is acquired and input into the trained competitive expert prediction network. The recoupled decoder outputs defect prediction values including defect type and geometric parameters.
[0151] The competitive expert prediction network is trained using the aforementioned leakage rate time series data training set as training input and the corresponding defect type label, defect location label, and defect geometric parameter true value label as training objectives.
[0152] In one specific embodiment, the loss function used to train the network is a composite physical constraint loss function:
[0153] :
[0154] in The data fitting loss includes the cross-entropy loss calculated on the defect type classification result output by the recoupled decoder, the cross-entropy loss calculated on the defect location classification result, and the normalized mean square error loss calculated on the defect prediction value of the geometric parameters. The physical residual loss is obtained by substituting the defect prediction values of the geometric parameters output by the recoupled decoder into the corresponding defect propagation dynamics equation and calculating the differential residual. The physical consistency loss is the deviation between the predicted leakage rate and the calculated leakage rate; weighting coefficient. It adapts and adjusts itself during training.
[0155] In one specific embodiment, the physical residual loss Calculated in the following way:
[0156] Substitute the predicted values of the geometric parameters into the finite difference form of the defect propagation dynamics equation for the corresponding defect type, and calculate the difference margin at each time step.
[0157] For fatigue cracks, the differential margin for:
[0158]
[0159] in for The crack half-length prediction value output by the recoupled decoder at time [time]. This represents the increment of the number of cyclic loads corresponding to adjacent time points. These are predicted values for the stress intensity factor range. This serves as a reference value for the normalized crack length.
[0160] For stress corrosion cracking, the differential allowance for:
[0161]
[0162] in For time step, This is the predicted value of the crack tip stress intensity factor. This is a normalized reference value for the stress corrosion cracking propagation rate.
[0163] For fretting wear defects, the differential margin for:
[0164]
[0165] in This refers to the predicted wear volume change rate output by the physical mechanism expert network branch corresponding to the fretting wear defect. The normalized reference value for the wear volume change rate; the physical residual loss The remainder is the mean of each difference over all training samples and all time points.
[0166] In one specific implementation, the dynamic flow coefficient prediction subnetwork and the competing expert prediction network are separated by physical consistency loss. Joint optimization, its formula is:
[0167]
[0168] in The number of samples in the training batch. To predict the leakage rate, The voltage difference between the primary and secondary loops. The current defect opening area is predicted by the network. This is a normalized reference value for the leakage rate.
[0169] In one specific implementation, the shared temporal feature extractor integrates a crack propagation stage perception attention mechanism: the stage division auxiliary classifier divides the defect propagation process into the initiation stage, stable propagation stage, rapid propagation stage, and through-stage stage based on the first and second derivative features of the leakage rate temporal data; the stage adaptive attention module configures different attention bias priors based on the division results and the magnitude of the defect propagation rate of each stage estimated by the corresponding physical mechanism model.
[0170] In one specific implementation, the architecture of the competitive expert prediction network is as follows: the shared temporal feature extractor employs a temporal convolutional network with dilated causal convolution structure; the gated network obtains the competitive activation weights of each branch after computation by a multilayer perceptron and processing by a Softmax function. The training objective of the gated network includes gated regularization loss:
[0171]
[0172] This causes the gating network to tend to assign higher weights to the branches that minimize the physical residual loss; The total number of branches in the physical mechanism expert network; the recoupled decoder uses the feature vectors of each branch and their corresponding competitive activation weights as input to calculate the weighted fusion feature vector. ,in For the first Competitive activation weights output by each branch of the physical mechanism expert network For the first The branch feature vectors output by each branch of the physical mechanism expert network are then used to output the prediction results for each item.
[0173] In this embodiment, the training of the competitive expert prediction network adopts a phased strategy:
[0174] Warm-up phase: During the initial training phase, the gating network parameters are frozen to ensure a uniform distribution of competitive activation weights across branches, i.e., each weight is 1 / This ensures that each branch of the physical mechanism expert network receives sufficient training signals, and avoids premature suppression of some branches due to the instability of physical residuals in the early stages of training.
[0175] Joint training phase: After warm-up, the gated network parameters are unfrozen, and gating regularization loss is introduced. The parameter update rule for the gated network is as follows:
[0176]
[0177] This phased training strategy ensures that each expert network branch participates in the competition only after it has been fully pre-trained, thus making the physical guidance of the gating regularization loss based on the fact that each branch already has basic prediction capabilities, and avoiding the problem of training instability.
[0178] Example 2
[0179] This embodiment simulates the entire process of the steam generator heat transfer tube defect prediction method in a real-world operating scenario and demonstrates the network prediction mechanism to verify the technical effectiveness of the competitive expert prediction network in this invention.
[0180] The heat transfer tubes of a steam generator in a simulated pressurized water reactor nuclear power plant, made of Inconel 690, developed a hidden defect during long-term service. For simplicity, this description only shows the defect at two representative cycle points. and The results of real-time monitoring and inversion prediction using this method.
[0181] Set initial parameters: primary and secondary loop pressure difference 10MPa (i.e.) Pa). Primary coolant density. 700 kg / m³. Elastic modulus of heat transfer pipe material. 200 GPa (i.e.) Pa). The heat transfer tube is subjected to static stress. 80MPa (i.e., 8.0×10) 7 Pa). Flow coefficient 0.6. Half-closure correction factor. 0.5. Time interval. : =1000 hours.
[0182] To verify the network's predictive capabilities, this embodiment presets the actual defect type as stress corrosion cracking (SCC) in the background forward simulation. At what moment, the true value of half the crack length =2mm; driven by complex physical mechanisms, in At time 1, the true value of the crack half-length extends to =3mm.
[0183] Based on the aforementioned leakage rate calculation model, including the defect opening area model and the orifice flow model, the system calculates respectively. and The actual leakage rate at any given moment: =0.002m into the formula:
[0184]
[0185] Will Substitute into the orifice flow model formula:
[0186]
[0187] Will Substituting 0.003m into the model:
[0188]
[0189]
[0190] Thus, a real-time leakage rate monitoring data sequence containing feature evolution was obtained. .
[0191] The real-time leakage rate monitoring data obtained above is input into a trained competitive expert prediction network for joint inversion: a shared temporal feature extractor is used to extract leakage rate time-series data. Using its first and second time derivatives as input, the shared feature vector is extracted.
[0192] Perform branch-independent decoupling decoding and physical residual verification; in the fatigue crack branch: if the fit is a fatigue crack, due to the increase in the number of cyclic loads within the time period in this embodiment... Minimal, the predicted value derived by reverse substitution This will result in a huge loss of physical residuals during calculation. In the fretting wear branch: if the fit is a wear defect, the calculated... and physical residual loss Also at a high level. SCC branch: Fits the time series characteristics to stress corrosion cracking. The time step derived from it is... The physical residual loss is calculated based on the equation constraints corresponding to the highly accurate predicted values. It approaches the minimum value.
[0193] The gating network outputs competitive activation weights based on the physical residual signals fed back from each branch: , , The recoupled decoder performs weighted fusion of the branch feature vectors:
[0194]
[0195] The recoupled decoder outputs the defect type prediction result for the heat transfer tube of the steam generator as "stress corrosion cracking (SCC)," and outputs... Defect prediction values of geometric parameters at time points mm, close to the true value of 3mm.
[0196] The predicted defect geometry parameter a = 3.01 mm was substituted back into the leakage rate calculation model for cross-validation:
[0197]
[0198]
[0199] The predicted dynamic flow coefficient output by the dynamic flow coefficient prediction subnetwork is 0.601, which is consistent with the initially set flow coefficient C. d =0.6 High consistency, after physical consistency loss The calculation shows that the residual of this test term approaches zero, further confirming the inherent consistency between the network prediction results and the physical model.
[0200] As can be seen from the full-process simulation in this embodiment, the present invention achieves deep decoupling and recoupling of physical mechanisms and data-driven approaches, changes the prediction logic of traditional black-box models, and ensures that the prediction results conform to the basic laws of fracture mechanics and tribology from the architectural level.
[0201] Example 3
[0202] like Figure 2 As shown, a steam generator heat transfer tube defect prediction system includes:
[0203] Physical mechanism modeling module: For various preset defect types in the heat transfer tubes of steam generators, it is used to establish defect propagation dynamic equations that describe the evolution of defect size with operating time and leakage rate calculation models based on defect size to calculate the corresponding leakage rate for each preset defect type.
[0204] Based on the aforementioned defect propagation dynamics equation and leakage rate calculation model, a forward simulation calculation is performed on the defect propagation process of each preset defect type under different initial defect sizes and operating conditions to obtain the time series data training set of leakage rate and the true value label of defect geometric parameters corresponding to each preset defect type.
[0205] Based on the aforementioned defect propagation dynamics equation and leakage rate model, forward simulation is performed to generate a training sample set containing leakage rate time series.
[0206] Prediction network construction module: used to construct a competitive expert prediction network, including a shared temporal feature extractor, multiple physical mechanism expert network branches corresponding one-to-one with the various preset defect types, a gating network, and a recoupled decoder;
[0207] The shared temporal feature extractor takes the leakage rate time series data in the obtained leakage rate time series data training set as input, extracts the shared feature vector, and simultaneously inputs the shared feature vector into the multiple physical mechanism expert network branches and the gating network;
[0208] Each physical mechanism expert network branch embeds a physical feature encoding layer associated with the defect expansion dynamics equation of the corresponding defect type; each physical mechanism expert network branch takes the shared feature vector as input, and through the physical feature encoding layer, explicitly encodes the physical mechanism feature parameters of the corresponding defect type into intermediate feature representations, and outputs its own branch feature vector;
[0209] The gated network takes the shared feature vector as input and outputs the competitive activation weights of each physical mechanism expert network branch;
[0210] The recoupled decoder takes the branch feature vectors output by each of the physical mechanism expert network branches as input, performs weighted fusion of the branch feature vectors according to the competitive activation weights, and outputs defect type classification results, defect location classification results, and defect prediction values of geometric parameters.
[0211] Training module: Using the aforementioned leakage rate time series data training set as training input, and the corresponding defect type label, defect location label, and defect geometric parameter truth value label as training target, the competitive expert prediction network is trained.
[0212] Prediction module: acquires real-time leakage rate monitoring data of the steam generator, inputs it into the trained competitive expert prediction network, and outputs defect prediction values including defect type and geometric parameters by the recoupled decoder.
Claims
1. A method for predicting a defect of a heat transfer tube of a steam generator, characterized by: The specific steps include the following: For various preset defect types in the heat transfer tubes of steam generators, defect propagation dynamic equations describing the evolution of defect size with operating time and leakage rate calculation models based on defect size are established for each preset defect type. Based on the aforementioned defect propagation dynamics equation and leakage rate calculation model, a forward simulation calculation is performed on the defect propagation process of each preset defect type under different initial defect sizes and operating conditions to obtain a time series data training set of leakage rates corresponding to each preset defect type. A competitive expert prediction network is constructed. The network includes multiple physical mechanism expert network branches corresponding one-to-one with each preset defect type. Each physical mechanism expert network branch embeds a physical feature encoding layer associated with the defect expansion dynamics equation of the corresponding defect type. Each physical mechanism expert network branch explicitly encodes the physical mechanism feature parameters of the corresponding defect type into intermediate feature representations through the physical feature encoding layer, and outputs its own branch feature vector. The network outputs a defect prediction value based on the branch feature vector. The competitive expert prediction network is trained using the aforementioned leakage rate time series data training set as training input. The loss function employed by the training network is a composite physics-constrained loss function that includes a physical residual loss This is calculated by: Substitute the predicted values of the geometric parameters into the finite difference form of the defect propagation dynamics equation for the corresponding defect type, and calculate the difference margin at each time step. For the fatigue crack, the differential margin is: wherein is the crack half-length prediction value of the momentary recoupling decoder output, is the corresponding cyclic load number increment of adjacent moments, is the prediction value of the stress intensity factor range, is the crack length normalized reference value, is the temperature-dependent material coefficient; For stress corrosion cracking cracks, the differential margin is: in For runtime, The power-law baseline coefficient, Crack tip stress intensity factor Influence function, For temperature Influence function, Dissolved hydrogen concentration Influence function, Crack propagation direction angle Influence function, For time step, This is the predicted value of the crack tip stress intensity factor. This is a normalized reference value for the stress corrosion cracking propagation rate. For fretting wear defects, the differential margin is: in The wear coefficient is... For contact force, For micro-motion speed, For material hardness, This refers to the predicted wear volume change rate output by the physical mechanism expert network branch corresponding to the fretting wear defect. The normalized reference value for the wear volume change rate; the physical residual loss The mean of each difference residual over all training samples and all time steps; Real-time leakage rate monitoring data of the steam generator is acquired and input into the trained competitive expert prediction network, which outputs defect prediction values including defect type and geometric parameters.
2. The method according to claim 1, characterized in that: The various preset defect types include fatigue cracks, stress corrosion cracks, and fretting wear defects; The defect propagation kinetic equation established for fatigue cracks is as follows: in The length of the crack is half its length. For the cumulative number of cyclic loads, As an accelerating factor for the water environment, For the range of stress intensity factors, The crack propagation index is... This is the load ratio correction index. The stress ratio; the range of the stress intensity factor. Calculate using the following formula: in For geometric correction factor, This refers to the range of cyclic load stress. The defect propagation kinetic equation established for stress corrosion cracking is as follows: The crack tip stress intensity factor Calculate using the following formula: in For geometric correction factor, It is static stress; The kinetic equation for the propagation of fretting wear defects is as follows: in The volume of the wear groove; Width of wear groove opening and the axial length of the wear groove The relationship is: The physical feature encoding layer encodes the physical mechanism feature parameters corresponding to the above formulas into components of the corresponding branch feature vectors.
3. The method according to claim 2, characterized in that: The leakage rate calculation model includes a defect opening area model and an orifice flow model; the defect opening area model calculates the current defect opening area according to the defect type using the following formulas. : For fatigue cracks, their opening area for: in For elastic modulus, This is the shell geometry correction factor; For stress corrosion cracking, its effective opening area for: in This is a semi-closed correction factor; For fretting wear defects, the equivalent leakage area for: Orifice flow model for calculating instantaneous leakage rate The formula is: in The flow coefficient, The voltage difference between the primary and secondary loops. This refers to the density of the primary coolant.
4. The method according to claim 3, characterized in that: The competitive expert prediction network also includes a shared temporal feature extractor, a gating network, and a recoupled decoder; The shared temporal feature extractor takes the leakage rate time series data in the obtained leakage rate time series data training set as input, extracts the shared feature vector, and simultaneously inputs the shared feature vector into the multiple physical mechanism expert network branches and the gating network; The gated network takes the shared feature vector as input and outputs the competitive activation weights of each physical mechanism expert network branch; The recoupled decoder takes the branch feature vectors output by each of the physical mechanism expert network branches as input, performs weighted fusion of the branch feature vectors according to the competitive activation weights, and outputs the defect type classification result, defect location classification result, and defect prediction value of geometric parameters.
5. The method according to claim 4, characterized in that: The loss function used to train the network is the composite physical constraint loss function: ： in The data fitting loss includes the cross-entropy loss calculated on the defect type classification result output by the recoupled decoder, the cross-entropy loss calculated on the defect location classification result, and the normalized mean square error loss calculated on the defect prediction value of the geometric parameters. The physical residual loss is obtained by substituting the defect prediction values of the geometric parameters output by the recoupled decoder into the corresponding defect propagation dynamics equation and calculating the differential residual. The physical consistency loss is the deviation between the predicted leakage rate and the calculated leakage rate; weighting coefficient. It adapts and adjusts itself during training.
6. The method according to claim 5, characterized in that: The competitive expert prediction network also includes a dynamic flow coefficient prediction subnetwork that works in conjunction with the recoupled decoder. The dynamic flow coefficient prediction subnetwork takes the current leakage rate time series data and its time derivative, the defect type classification result currently output by the recoupled decoder, the defect prediction value of geometric parameters, and the operating condition parameters as input, and outputs the dynamic flow coefficient prediction value. The predicted value of the dynamic flow coefficient The calculation method is as follows: in The linear output of the dynamic flow coefficient prediction subnetwork is given. and The lower and upper bounds of the physical constraint intervals of the flow coefficient corresponding to the defect type classification results are determined by physical experience based on the defect type. The dynamic flow coefficient prediction subnetwork and the competing expert prediction network are separated by physical consistency loss. Joint optimization, its formula is: in The number of samples in the training batch. To predict the leakage rate, The voltage difference between the primary and secondary loops. The current defect opening area is predicted by the network. This is a normalized reference value for the leakage rate.
7. The method according to claim 4, characterized in that: The shared temporal feature extractor integrates a crack propagation stage perception attention mechanism: the stage division auxiliary classifier divides the defect propagation process into the initiation stage, stable propagation stage, rapid propagation stage, and through-stage stage based on the first and second derivative features of the leakage rate temporal data; the stage adaptive attention module configures different attention bias priors based on the division results and the magnitude of the defect propagation rate of each stage estimated by the corresponding physical mechanism model.
8. The method according to claim 5, characterized in that: The specific architecture of the competitive expert prediction network is as follows: the shared temporal feature extractor adopts a temporal convolutional network with dilated causal convolution structure; the gated network obtains the competitive activation weights of each branch after calculation by a multilayer perceptron and processing by a Softmax function. The training objective of the gated network includes gated regularization loss: This causes the gating network to tend to assign higher weights to the branches that minimize the physical residual loss; The total number of branches in the physical mechanism expert network; the recoupled decoder uses the feature vectors of each branch and their corresponding competitive activation weights as input to calculate the weighted fusion feature vector. This leads to the output of various prediction results.
9. A defect prediction system for heat transfer tubes in a steam generator, characterized in that: include: The physical mechanism modeling module is used to establish, for various preset defect types in the heat transfer tubes of steam generators, defect propagation kinetic equations describing the evolution of defect size over operating time and leakage rate calculation models based on defect size; and based on the aforementioned defect propagation kinetic equations and leakage rate calculation models, forward simulation calculations are performed on the defect propagation process of each preset defect type under different initial defect sizes and operating parameters to obtain a time series data training set of leakage rates corresponding to each preset defect type. The prediction network construction module is used to construct a competitive expert prediction network comprising multiple physical mechanism expert network branches corresponding one-to-one with each preset defect type; each physical mechanism expert network branch embeds a physical feature encoding layer associated with the defect expansion dynamics equation of the corresponding defect type; each physical mechanism expert network branch explicitly encodes the physical mechanism feature parameters of the corresponding defect type into intermediate feature representations through the physical feature encoding layer, and outputs its own branch feature vector; the network outputs the defect prediction value based on the branch feature vector; Training module: Using the aforementioned leakage rate time series data training set as training input, train the competitive expert prediction network; The loss function used to train the network is a composite physical constraint loss function that includes physical residual loss. Calculated in the following way: Substitute the predicted values of the geometric parameters into the finite difference form of the defect propagation dynamics equation for the corresponding defect type, and calculate the difference margin at each time step. For fatigue cracks, the differential margin for: in for The crack half-length prediction value output by the time-recoupled decoder. This represents the increment of the number of cyclic loads corresponding to adjacent time points. These are predicted values for the stress intensity factor range. This is a normalized reference value for crack length. This refers to the temperature-dependent material coefficient; For stress corrosion cracking, the differential allowance for: in For runtime, The power-law baseline coefficient, Crack tip stress intensity factor Influence function, For temperature Influence function, Dissolved hydrogen concentration Influence function, Crack propagation direction angle Influence function, For time step, This is the predicted value of the crack tip stress intensity factor. This is a normalized reference value for the stress corrosion cracking propagation rate. For fretting wear defects, the differential margin for: in The wear coefficient is... For contact force, For micro-motion speed, For material hardness, This refers to the predicted wear volume change rate output by the physical mechanism expert network branch corresponding to the fretting wear defect. The normalized reference value for the wear volume change rate; the physical residual loss The mean of each difference residual over all training samples and all time steps; Prediction module: acquires real-time leakage rate monitoring data of the steam generator, inputs it into the trained competitive expert prediction network, and outputs defect prediction values including defect type and geometric parameters.