A physical constraint and large model fusion nuclear power cooling system diagnosis method

Abstract

The application discloses a nuclear power cooling system diagnosis method based on physical constraints and large models, and belongs to the technical field of intelligent monitoring and fault diagnosis of nuclear power important equipment. The method comprises the following steps: S1, collecting multi-source physical quantity time sequence data, constructing a time sequence tensor and a multi-modal time sequence data set; S2, constructing a physical constraint loss function; S3, constructing an enhanced sample; S4, extracting system dynamic characteristics and fusing multi-modal characteristics; S5, performing diagnosis through a large language model to obtain a fault diagnosis result; and S6, performing confidence evaluation and outputting a diagnosis conclusion with confidence evaluation. In the application, basic thermal-hydraulic laws are embedded in model training and reasoning in the form of physical constraints, the feature learning space is constrained, and diagnosis results that violate physical laws are avoided in pure data-driven models. Through physical Jacobian matrix guided disturbance generation and physical consistency discrimination condition screening, physically reasonable enhanced samples are generated under few real fault samples, and the model generalization ability is improved.

Description

Technical Field

[0001] This invention relates to the field of intelligent monitoring and fault diagnosis technology for critical nuclear power equipment, specifically a diagnostic method for nuclear power cooling systems that integrates physical constraints and large-scale models. Background Technology

[0002] Nuclear power plant cooling systems operate under high temperature, high pressure, and strong radiation environments for extended periods, and their operational reliability directly impacts reactor safety. Failure to detect critical equipment malfunctions such as bearing wear or seal failure in a timely manner could lead to serious accidents. Therefore, conducting high-precision, highly reliable condition diagnostics of critical nuclear power equipment has significant engineering value.

[0003] Currently, diagnostic methods for key nuclear power equipment mainly include the following categories: (I) Analysis methods based on thermal-hydraulic mechanisms. This type of method establishes physical equations such as mass, energy, momentum conservation, and point reactor dynamics, and uses residual analysis to achieve fault detection. Its advantages are strong interpretability and independence from fault samples. However, nuclear power systems are highly complex and strongly nonlinear, making accurate modeling extremely difficult. Even after model simplification, the accuracy is limited, making it difficult to detect gradual or localized faults. Furthermore, it is sensitive to changes in boundary conditions, easily leading to false alarms or missed alarms.

[0004] (II) Data-driven machine learning diagnostic methods. These methods utilize models such as deep neural networks to directly learn fault mapping relationships from historical data, achieving good results in the diagnosis of general rotating machinery. However, applying them to nuclear main pumps faces three major bottlenecks: First, real fault samples are extremely scarce, making the model prone to overfitting and resulting in poor generalization ability; second, multi-source sensor signals (pressure, temperature, flow rate, neutron flux, etc.) exhibit strong physical heterogeneity, and existing methods lack physical causal modeling, making effective fusion difficult; third, the black-box nature of the diagnostic results leads to a lack of physical consistency, often violating basic conservation laws, which fundamentally contradicts the stringent requirements for credibility and interpretability in the field of nuclear safety.

[0005] (III) Diagnostic methods based on physical information neural networks. This type of method introduces physical residual constraints into the loss function, enabling the network to learn in accordance with physical laws. However, existing physical information neural networks are mainly geared towards single physical fields or forward and inverse problems, lacking a unified representation capability for multimodal heterogeneous physical constraints, lacking a physical consistency enhancement mechanism for few-sample faults, and not integrated with new inference architectures such as large language models, making it difficult to directly meet the needs of engineering diagnosis of nuclear power cooling systems.

[0006] In summary, existing technologies for diagnosing the operational status of nuclear main pumps generally suffer from problems such as scarce fault samples, difficulty in fusing multi-source heterogeneous information, and insufficient physical consistency and interpretability of diagnostic results. Therefore, there is an urgent need to propose a diagnostic method that can achieve high accuracy, physical consistency, and interpretability under conditions of very few fault samples, in order to improve the safety assurance capabilities of critical nuclear power equipment. Summary of the Invention

[0007] To address the problems existing in the prior art, this invention provides a diagnostic method for nuclear power cooling systems that integrates physical constraints with a large model.

[0008] To achieve the above objectives, the technical solution of the present invention is as follows: A diagnostic method for nuclear power plant cooling systems that integrates physical constraints and large-scale models is provided, comprising the following steps: S1. Collect time-series data of multi-source physical quantities in the nuclear power water reactor cooling system, and combine the system operating parameters to perform time alignment and resampling processing on data with different sampling frequencies, and construct a time-series tensor and multimodal time-series data set under a unified time scale; S2. Based on the time series tensor of S1, the basic physical conservation relationships of mass, energy and momentum of the nuclear power cooling system are introduced as soft constraints to construct the physical constraint loss function and the total loss function. S3. Based on the multimodal time series data set of the normal operation samples generated by S1 and the physical constraint loss function constructed by S2, plus physical perturbation, an enhanced sample is constructed. S4. Extract dynamic features at different time scales from the enhanced samples generated in S3, introduce the physical correlation matrix to guide cross-modal feature fusion, and generate a physically consistent multimodal comprehensive feature vector. S5. Construct a physical knowledge graph of the nuclear power cooling system, map the multimodal fusion features output by S4 into symbolic representations with physical semantics, and embed them with the physical knowledge graph. Then, use a large language model for diagnosis to obtain fault diagnosis results. S6, based on the physical constraint loss function constructed by S2, performs uncertainty assessment and confidence calibration on the diagnostic results output by S5 through physical consistency, and outputs a diagnostic conclusion with uncertainty and confidence assessment.

[0009] Preferably, in S1, the multi-source physical quantity time-series data includes pressure, temperature, flow rate, and neutron flux data; the system operating parameters are one or more of the following: control rod position, main pump speed, main pump frequency command, pressure regulator spray valve opening, pressure regulator heater power, and makeup water flow rate. A running sample represents all multimodal data collected by the system within a continuous time window. The first running sample is defined as... A physical mode in time The observed values ​​are as follows Time alignment is handled through interpolation, and resampling is performed according to the following formula: In the formula, To unify the first on the timeline At a certain point in time, Indicates a unified point in time Upper m Estimation data for physical modes; For the first sample A point in time; Indicates in Time of the first m The actual measured values ​​of the physical modes; Represents the first of the original samples A point in time; The unified multimodal feature vector is represented as : Construct the temporal tensor for each running sample for: in, The total number of time steps under a unified time scale; The number of physical modes; Indicates the first A physical mode in time The observed values, the physical modes include at least one of the following: pressure, temperature, mass flow rate, neutron flux, and control parameters; Indicates time The joint observation vector at any given time is composed of multiple physical modes from various sources; Based on the above and , No. A unified multimodal time series dataset of each running sample. Represented as: Among them, X (i) Indicates the first The multimodal feature matrix of each running sample; The dimension of the multimodal features; N For the number of independently run samples, For the first The feature dimension of a physical mode, compared to the temporal tensor. , Emphasizing is the first A collection of multimodal time-series data from a set of running samples.

[0010] Preferably, S2 includes the following steps: S2.1, Definition The system state variables at time t are: in, This is the system state vector; This refers to the pressure state in the nuclear power plant's cooling circuit. This refers to the coolant temperature state; This refers to the coolant mass flow rate status; This refers to the reactor neutron flux state; Further define the system control input vector for: in, The first part represents the nuclear power cooling system. System operating parameters; The changes in system state variables over time follow the dynamic evolution law under the coupling effect of thermal-hydraulic and control systems, and its state evolution relationship is described as follows: in, The rate of change of the state of the nuclear power system over time; It is a nonlinear physical evolution function composed of the mass conservation equation, the energy conservation equation, and the momentum conservation equation; This is the system control input vector; S2.2. Based on the system state variables, the thermo-hydraulic conservation equations are introduced as constraints. The thermo-hydraulic conservation equations include mass, energy, and momentum conservation equations: The mass conservation equation is: in, This represents the rate of change of pressure with respect to time. Indicates the density of the coolant; Represents the velocity field; For leakage mass flow rate; It was Dirac function; r Let be the position coordinate vector of any point; r l The spatial coordinates of the leak point; For replenishing water mass flow rate; The energy conservation equation is: in, This refers to the specific heat capacity of the coolant under constant pressure. The temperature field is represented by k, and the thermal conductivity is represented by k. Represents the gradient; Represents the gradient of the temperature field; For volumetric heat source term, This refers to the system boundary heat loss rate; This indicates the rate of energy loss due to leakage; The momentum conservation equation is: in, Indicates a pressure field; μ Represents dynamic viscosity; g It is the vector of gravitational acceleration; This is the frictional resistance term; This refers to the momentum loss due to leakage; S2.3 Constructing a computable physical constraint loss function for: In the formula, Indicates the first A physical conservation residual function; Indicates the total number of physical constraints; Index for physical constraints; S2.4 Embed physical constraints into the total loss function : In the formula, These are the physical constraint weighting coefficients; For data-driven cross-entropy loss.

[0011] Preferably, S3 includes the following steps: S3.1 The constructed enhanced sample is: in, For the generated enhanced samples; For physically guided perturbation terms; A coefficient used to control the amplitude of the disturbance; S3.2, Disturbance Term The formula is: In the formula, Indicates standard Gaussian noise. This indicates a disturbance that is physically constrained. It is the physical model relative to the input parameters Jacobian matrix; It follows a standard normal distribution.

[0012] Preferably, S4 includes the following steps: S4.1 Multi-scale Feature Extraction Features at three scales are extracted using convolutional or recurrent networks with different receptive fields: , and Representing the time series tensor Results after extracting short-term, medium-term, and long-term features; , and These represent the feature extraction functions for the short, medium, and long term, respectively. S4.2 Multi-scale feature splicing The features from the three scales of S4.1 are concatenated into a complete feature vector. for: In the formula, ‖ represents the vector concatenation operation; S4.3, Physically Guided Cross-Modal Feature Fusion After extracting multi-scale features, attention fusion is performed based on the physical correlation matrix: When calculating cross-modal attention weights, the physical correlation matrix is ​​used. As a bias term, it is added to define the physical correlation matrix. for: In the formula, for and The strength of the physical relationship; and Both are physical mode indexes. For source mode; For the target mode; This is the expected value; For modality When changing, mode Sensitivity to changes; The formula for calculating cross-modal attention weights is: in, For attention weights, and These are the transformation matrices of the query vector and the key vector, respectively. Indicates the dimension of the key vector; It is a normalized exponential function; The weighted fusion formula is: In the formula, This is the multimodal comprehensive feature vector obtained after multi-scale feature extraction and cross-modal fusion; For the first A value vector of modalities.

[0013] Preferably, S5 includes the following steps: S5.1 Mapping of Multimodal Fusion Features to Physical Semantic Symbols Multimodal fusion features The mapping is to a set of symbols with physical semantics, as shown in the formula: in, It is a set of physical semantic symbol states; No. A semantic state description; It is a feature-to-semantic mapping function; Number of semantic states; S5.2 Embedding of Physics Knowledge Graph A graph neural network is used to embed the physical knowledge graph into a vector space, as shown in the formula: In the formula, Representing entities In the Layer embedding vectors, Representing entities In the Layer embedding vector; Representing entities In the Layer embedding vector; For layer index; For activation functions; A collection of relation types; It can be any type of relation; To pass through relationships With entity A set of connected neighboring entities; For neighboring entities; For the first Layer Relationship The transformation matrix; For the first Self-connected transformation matrix in the layer; S5.3, Joint Reasoning Diagnosis Combining the physical semantic symbols of S5.1 with the knowledge graph embedding of S5.2, the diagnostic results are obtained through a neural symbolic inference engine: the neural symbolic inference engine combines neural network features with knowledge graph embedding, as shown in the formula: In the formula, The diagnostic result is represented by Re(·), which is the inference function. For nuclear-related physical knowledge or parameters; These are the learnable parameters of the inference engine; After neural network After layer propagation, each entity receives its final embedding. Aggregate related entities ,get ; S5.4 Output Constraints In the reasoning and diagnostic process, a physical consistency discriminant function is introduced. for: In the formula, This represents the first step in the generation process of the large language model. The words output in the step, This refers to words or symbols generated by a large language model. This represents the positional step in the sequence generated by the large language model. The corrected language model generation probability is: In the formula, Generate probabilities for large language models; This is a physical consistency discrimination function; This is a set of candidate semantics.

[0014] Preferably, S6 includes the following steps: S6.1 Uncertainty Calculation Based on noise in the measurement data, model errors, and uncertainties in physical modeling, a comprehensive uncertainty is constructed. The metric is used to calibrate the reliability of diagnostic results, and the formula is: in, Due to noise uncertainty in measurement data; For model parameter uncertainty; To model the uncertainty of physical errors; S6.2 Confidence Calibration Confidence assessment is corrected using a physical consistency calibration function, the formula of which is: The calibrated confidence level; The confidence level of the model's original output; For calibration coefficients; S6.3, Output Diagnostic Output The final output, combining the diagnostic results from S5 and the confidence calibration from S6, includes a diagnostic conclusion encompassing fault type, severity, evolution trend, reasoning path, confidence level, and uncertainty assessment. This invention provides a diagnostic method for nuclear power plant cooling systems that integrates physical constraints and large-scale models, which has the following beneficial effects: (1) This invention embeds the basic laws of thermal hydraulics, such as mass conservation, energy conservation, and momentum conservation, into the model training and reasoning process in the form of physical constraints, thereby constraining the feature learning space and avoiding the generation of diagnostic results that violate physical laws by purely data-driven models.

[0015] (2) This invention proposes a physical consistency-guided few-sample augmentation, which generates perturbations by guiding the physical Jacobian matrix and is supplemented by physical consistency discrimination criteria. It can still generate physically reasonable augmented samples under the condition of very few real fault samples, which greatly improves the generalization ability of the model.

[0016] (3) This invention constructs a physical-guided multi-scale feature extraction and cross-modal attention fusion network, and uses the physical correlation matrix to guide the calculation of cross-modal attention weights to achieve deep fusion of heterogeneous physical information such as pressure, temperature, flow rate, neutron flux, and control parameters, and make full use of the physical causal relationship of multiple sources.

[0017] (4) This invention uses a large language model to perform neural and symbolic joint reasoning, maps multimodal fusion features to a set of physical semantic symbols and embeds them with a physical knowledge graph, and generates a natural language diagnostic reasoning chain that conforms to physical laws and an uncertainty quantification assessment through physical mask constraints. This enables interpretable and reliable intelligent diagnosis. Attached Figure Description

[0018] Figure 1 This is an overall flowchart of a diagnostic method for nuclear power cooling systems that integrates physical constraints and large models, according to the present invention.

[0019] Figure 2 This is a flowchart of steps S1-S4 of the present invention.

[0020] Figure 3 This is a flowchart of steps S5-S6 of the present invention. Detailed Implementation

[0021] The present invention will be further described in detail below with reference to specific embodiments and accompanying drawings, but the scope of protection of the present invention is not limited to the content described.

[0022] To address the challenges of extremely scarce fault samples, drastic changes in operating conditions, strong physical heterogeneity of multi-source sensor information, and insufficient physical consistency and reliability of traditional data-driven diagnostic methods in nuclear power cooling systems operating under long-term high temperature, high pressure, and high radiation conditions, this invention provides a diagnostic method for nuclear power cooling systems that integrates physical constraints and large-scale models. By introducing constraints based on thermal-hydraulic physical mechanisms, this invention deeply integrates multi-source sensor signals, physical models, and the reasoning capabilities of large-scale linguistic models. Under conditions of zero or very few fault samples, it achieves high-precision perception, fault prediction, and reliable diagnostic reasoning of the operating status of nuclear power cooling systems, thereby improving the operational safety and intelligent operation and maintenance level of key nuclear power equipment.

[0023] like Figure 1-3 As shown, the present invention specifically includes the following steps: S1. Unified representation of multi-source data acquisition and physical consistency Time-series data of multiple physical quantities, including pressure, temperature, flow rate, and neutron flux, are collected from the nuclear power water reactor cooling system. Combined with control parameters, time alignment and resampling are performed on data from different sampling frequencies to construct a multimodal time-series data set with a unified time scale.

[0024] Pressure, temperature, flow rate, neutron flux and other status monitoring sensors are installed at key locations in the nuclear power cooling system. At the same time, the system operating parameters are collected, including one or more of the following: control rod position, main pump speed, main pump frequency command, pressurizer spray valve opening, pressurizer heater power and makeup water flow rate.

[0025] Because different acquisition devices have inconsistent sampling frequencies and units, it is necessary to map multi-source data to a unified time axis to construct a multimodal time series representation with clear physical meaning. A running sample represents all multimodal data acquired by the system within a continuous time window, and the first running sample is defined as... A physical mode in time The observed values ​​are as follows Time alignment is handled through interpolation, and resampling is performed according to the following formula: In the formula, To unify the first on the timeline At a certain point in time, Indicates a unified point in time Above, the first m Estimation data for physical modes; For the first sample A point in time; Indicates in Time of the first mThe actual measured values ​​of the physical modes; Represents the first of the original samples A specific point in time.

[0026] The unified multimodal feature vector is represented as : Further construct the temporal tensor for each running sample for: in, The total number of time steps under a unified time scale; The number of physical modes; Indicates the first A physical mode in time The observed values, the physical modes include at least one of the following: pressure, temperature, mass flow rate, neutron flux, and control parameters; Indicates time The time-series joint observation vector is composed of multiple physical modes.

[0027] Based on the above and , No. A unified multimodal time series dataset of running samples. Represented as: Among them, X (i) Indicates the first The multimodal feature matrix of each running sample; The dimension of the multimodal features; N For the number of independently run samples, For the first The feature dimension of each physical mode; compared to the temporal tensor. , Emphasizing is the first A collection of multimodal time-series data from a set of running samples.

[0028] S2. Constructing a physical constraint model for a nuclear power plant cooling system. S1-based temporal tensors We introduce the basic physical conservation relationships of mass, energy, and momentum conservation in nuclear power cooling systems as soft constraints and construct a physical soft constraint loss function.

[0029] This invention introduces the thermal-hydraulic mechanism constraint of nuclear power cooling systems, which can avoid the deep model from producing physically unrealizable state prediction results during the learning process.

[0030] S2.1, Definition The system state variables at time t are: in, This is the system state vector; This refers to the pressure state in the nuclear power plant's cooling circuit. This refers to the coolant temperature state; This refers to the coolant mass flow rate status; It represents the neutron flux state of the reactor, or a power-related state quantity used to characterize the core power level.

[0031] Further define the system control input vector for: in, The first part represents the nuclear power cooling system. System operating parameters; The changes in system state variables over time follow the dynamic evolution law under the coupling effect of thermal-hydraulic and control systems, and its state evolution relationship is described as follows: in, The rate of change of the state of the nuclear power system over time; It is a nonlinear physical evolution function composed of the mass conservation equation, the energy conservation equation, and the momentum conservation equation; This is the system control input vector, used to characterize the control system's adjustment effect on the operating state of the cooling circuit.

[0032] S2.2. Based on the system state variables, the thermo-hydraulic conservation equations are introduced as constraints. The thermo-hydraulic conservation equations include mass, energy, and momentum conservation equations: mass conservation equation: in, This represents the rate of change of pressure with respect to time. This indicates the density of the coolant, which varies with time and space. This represents the velocity field, which is a three-dimensional vector field. Leakage mass flow rate is a scalar function that may vary over time; It was Dirac The function represents the leak point. r l Leakage at the location; r These are spatial coordinates, representing the position vector of any point in the system; r l These are the spatial coordinates of the leak point, indicating the specific location where the leak occurred; The mass flow rate of the makeup water is a scalar function, usually determined by the control system.

[0033] Energy conservation equation: in, This refers to the specific heat capacity of the coolant under constant pressure. The temperature field varies with time and space; k This refers to thermal conductivity, which may vary with temperature. Represents the gradient; This represents the gradient of the temperature field, that is, the rate of change of temperature in all directions in space; This is a volumetric heat source term (including nuclear fission heat, decay heat, etc.). This refers to the system boundary heat loss rate; This indicates the rate of energy loss due to leakage.

[0034] Momentum conservation equation: in, This represents the pressure field, which varies with time and space. μ Represents dynamic viscosity; g It is the vector of gravitational acceleration; This is the frictional resistance term; This is the momentum loss term caused by leakage.

[0035] The mass, energy, and momentum conservation equations are primarily used to constrain the thermo-hydraulic state variables of pressure, temperature, and flow rate in a system. Essentially, they correspond to the pressure, heat transfer, and flow variations of the coolant within the system. Neutron flux, while also a system state variable... It is part of the equation, but its variation is mainly controlled by reactor neutron dynamics and power evolution mechanism. It belongs to a different physical level of constraint object from the thermal-hydraulic conservation equation, and therefore is not directly constrained by the mass, energy and momentum conservation equation.

[0036] In this invention, the neutron flux is not unconstrained, but rather constrained indirectly or jointly in the following ways: (1) There is a thermo-mechanical coupling relationship between neutron flux and variables such as temperature and flow rate. For example, changes in coolant temperature can cause reactive feedback, which in turn affects neutron flux.

[0037] (2) In the subsequent steps (S5) of physical knowledge graph and joint reasoning, neutron flux participates in physical association modeling as a key physical semantic state.

[0038] Therefore, the present invention currently adopts the method of "thermal-hydraulic conservation constraint + physical correlation reasoning constraint" to indirectly constrain neutron flux. This design ensures physical rationality, thereby maintaining the engineering feasibility of the scheme and the stability of the patent protection scope.

[0039] S2.3, Construct a computable physical constraint loss function as follows: : Indicates the first The physical conservation residual function is the residual obtained after applying the above three conservation equations to the model input X; Indicates the total number of physical constraints; Index for physical constraints; S2.4 Embed physical constraints into the total loss function : In the formula, These are the physical constraint weighting coefficients; Data-driven cross-entropy loss enables the model to learn to identify fault types. S3, Physical Consistency-Guided Small Sample Fault Data Augmentation To address the scarcity of real fault samples in nuclear power cooling systems, physically constrained perturbations are introduced into the normal operation samples to generate enhanced fault samples that satisfy the laws of thermo-hydraulic dynamics.

[0040] S3.1, Normal operating sample X generated based on S1 (i) The physical constraint loss function constructed with S2, plus physical perturbations, yields the augmented samples. The enhanced sample is constructed as follows: in, The generated augmented sample is A ×D matrix represents fault data over a continuous time period; For physically guided perturbation terms; A coefficient used to control the amplitude of the disturbance.

[0041] S3.2, Disturbance Term The formula is: In the formula, Indicates standard Gaussian noise. This indicates a disturbance that is physically constrained. It is the physical model relative to the input parameters The Jacobian matrix ensures that the direction of the perturbation conforms to physical rationality; It follows a standard normal distribution.

[0042] S3.3 Enhanced Sample Validation The generated enhanced samples are not all accepted; they still need to go through a verification step to ensure that they do indeed conform to physical laws. The verification includes two constraint verification methods: dynamic evolution constraint and conservation residual constraint. In actual use, one can be used or used in combination for constraint. Combination means that both constraints are verified separately.

[0043] S3.3.1 Dynamic Evolution Constraints: in, To enhance the system state variables corresponding to the samples; It is the L2 norm, used to measure the size of physical residuals; The physical consistency tolerance threshold is used to limit the extent to which augmented samples deviate from the physical model.

[0044] S3.3.2, Conservation of residual constraints: In the formula, For the first j Each physical constraint allows for an error threshold.

[0045] S4. Physically Guided Multiscale Feature Extraction and Cross-Modal Fusion In the process of multimodal feature fusion, there are clear physical coupling relationships between different modalities. This invention introduces a physical constraint matrix to guide the multimodal fusion process.

[0046] Dynamic features at different time scales (short-term, medium-term, and long-term) are extracted from the enhanced samples generated by S3, and attention fusion between different physical modalities (pressure, temperature, flow rate, control parameters, etc.) is guided by the physical correlation matrix to generate a physically consistent multimodal comprehensive feature vector.

[0047] The dynamic behavior of nuclear power plant cooling systems varies across different time scales, as shown in Table 1: The relationship between time windows at different time scales is as follows: .

[0048] S4.1 Multi-scale Feature Extraction Enhanced samples generated based on S3 By extracting system dynamic features at multiple time scales and introducing a physical correlation matrix to guide cross-modal feature fusion, multi-scale fused features that satisfy physical consistency are obtained.

[0049] Features at three scales are extracted using convolutional or recurrent networks with different receptive fields: , and They represent respectively to Results after extracting short-term, medium-term, and long-term features; , and These represent the feature extraction functions for the short, medium, and long term, respectively. For short-term sequences in small receptive field convolutional kernels (e.g., kernel_size=3) or Long Short-Term Memory (LSTM) networks; For medium receptive field convolutional kernels (such as kernel_size=7) or stacked short time sequences; Use large receptive field convolutional kernels (e.g., kernel_size=15) or multi-layer temporal attention.

[0050] S4.2 Multi-scale feature splicing The features from the three scales of S4.1 are concatenated into a complete feature vector. for: In the formula, ‖ represents the vector concatenation operation.

[0051] S4.3, Physically Guided Cross-Modal Feature Fusion After extracting multi-scale features, it is necessary to fuse features from different physical modes (pressure, temperature, flow rate, etc.). This invention uses attention fusion based on the physical correlation matrix. When calculating cross-modal attention weights, the physical correlation matrix is ​​used. As a bias term, it is added to define the physical correlation matrix. for: In the formula, for and The strength of the physical relationship is calculated from a physical knowledge graph or a physical model; and Both are physical mode indexes. The source mode is the independent variable. is the target mode, and is the dependent variable; This is the expected value (statistical average). For modality When changing, mode Sensitivity to changes; The formula for calculating cross-modal attention weights is: in, For attention weights, and These are the transformation matrices of the query vector and the key vector, respectively. Indicates the dimension of the key vector; It is a normalized exponential function.

[0052] The weighted fusion formula is: In the formula, This is the multimodal comprehensive feature vector obtained after multi-scale feature extraction and cross-modal fusion; For the first A value vector of modalities.

[0053] S5, Physically Constrained Large Language Model-Driven Diagnostics Construct a physical knowledge graph of the nuclear power plant cooling system and fuse the multimodal features output by S4. The data is mapped to symbolic representations with physical semantics, then combined with physical knowledge graph embedding, and diagnosed through a large language model to obtain interpretable fault diagnosis results.

[0054] S5.1 Mapping of Multimodal Fusion Features to Physical Semantic Symbols Multimodal fusion features The mapping is to a set of symbols with physical semantics, as shown in the formula: in, It is a set of physical semantic symbol states; No. A semantic state description (such as) The pressure is too high. (Temperature is normal). It is a feature-to-semantic mapping function; Number of semantic states. S5.1 provides the semantic symbols (transient states) of the current observation.

[0055] S5.2 Embedding of Physics Knowledge Graph Symbolic physical knowledge (such as "increased pressure leads to increased temperature") is converted into vector representations for fusion with neural network features. A graph neural network is used to embed the physical knowledge graph into the vector space, as shown in the formula: In the formula, Representing entities In the Layer embedding vectors, Representing entities In the Layer embedding vector; Representing entities In the Layer embedding vector; For layer index; For activation functions; A set of relation types (such as including cause, like about, equal to); For a certain type of relationship (e.g., "cause"); To pass through relationships With entity A set of connected neighboring entities; For neighboring entities; For the first Layer Relationship The transformation matrix; For the first Self-connected transformation matrices in layers. S5.2 provides vector representations of physical laws (long-term knowledge).

[0056] S5.3, Joint Reasoning Diagnosis Combining the physical semantic symbols of S5.1 with the knowledge graph embedding of S5.2, the diagnostic results are obtained through a neural symbolic inference engine: the neural symbolic inference engine combines neural network features with knowledge graph embedding, as shown in the formula: In the formula, The diagnostic results include fault type, severity, and evolution trend; Re(·) is the inference function; For nuclear-related physical knowledge or parameters; These are the learnable parameters of the inference engine; MLP It is a multilayer perceptron.

[0057] After neural network After layer propagation, each entity receives its final embedding. Aggregate related entities ,get The more layers there are, the farther information travels, but too many layers may lead to over-smoothing. This invention... Option 2. Physically constrained Large Language Model (LLM) joint inference algorithm, such as... Figure 3 As shown.

[0058] S5.2 and splicing, making MLPSimultaneously access the current state and physical knowledge. When the current semantic state is temperature rise, joint reasoning is performed by combining the correlation between temperature and control rod position and leakage fault in the knowledge graph.

[0059] S5.4 Output Constraints During the reasoning diagnosis process, physical masks are used to constrain the output of the large language model to ensure that the generated reasoning chain conforms to physical laws, and a physical consistency discriminant function is introduced. for: In the formula, This represents the first step in the generation process of the large language model. The words output in the step, This refers to words or symbols generated by a large language model. This represents the position step in the sequence generated by the large language model.

[0060] Correcting the language model generation probability: In the formula, Generate probabilities for large language models; This is a physical consistency discrimination function; This is a set of candidate semantics. The language model's generation probability is adjusted to automatically filter out statements that violate physical laws when generating diagnostic reports.

[0061] S6. Diagnostic uncertainty assessment and credibility calibration Diagnostic results based on S5 output and the physical constraint loss constructed by S2 This step quantifies the various uncertainties in the diagnostic process and calibrates the confidence level using physical consistency, ultimately outputting a diagnostic conclusion with a confidence assessment.

[0062] S6.1 Uncertainty Calculation Considering the noise in the measurement data, model errors, and uncertainties in physical modeling, a comprehensive uncertainty is constructed. The final conclusion is output after the metric performs a credibility calibration on the diagnostic results: in, Due to noise uncertainty in measurement data; For model parameter uncertainty; To model the uncertainty of physical errors.

[0063] S6.2 Confidence Calibration Confidence assessment is corrected using a physical consistency calibration function, the formula of which is: The calibrated confidence level; The confidence level of the model's original output; This is the calibration coefficient.

[0064] S6.3, Output Diagnostic Output The final output, which combines the diagnostic results of S5 and the confidence calibration of S6, includes: fault type, severity, evolution trend, reasoning path, confidence level, and uncertainty assessment.

[0065] The present invention also provides a nuclear power cooling system diagnostic system that integrates physical constraints and large models, including a data acquisition module, a physical constraint modeling module, a sample augmentation module, a multi-scale feature fusion module, a large model inference module, and an uncertainty assessment module.

[0066] The present invention also provides a computer-readable storage medium having a computer program stored thereon, wherein the computer program, when run on a processor, provides the aforementioned method for diagnosing nuclear power cooling systems that integrates physical constraints and large models.

[0067] This invention achieves accurate modeling of the complex dynamic behavior of nuclear power plant cooling systems through a comprehensive technical approach: unified representation of multimodal data → physical mechanism constraint modeling → physical consistency data enhancement → multi-scale cross-modal feature fusion → large-scale joint inference → uncertainty assessment. The method of this invention does not simply rely on historical fault samples, but rather introduces physical constraints and the knowledge reasoning capabilities of a large model, enabling the model to have generalized diagnostic capabilities under unknown operating conditions and extreme conditions.

[0068] Although embodiments of the invention have been shown and described, it will be understood by those skilled in the art that various changes, modifications, substitutions and alterations can be made to these embodiments without departing from the principles and spirit of the invention, the scope of which is defined by the appended claims and their equivalents.

Claims

1. A physically constrained and large model fused nuclear power cooling system diagnosis method, characterized in that: Includes the following steps: S1. Collect time-series data of multi-source physical quantities in the nuclear power water reactor cooling system, and combine the system operating parameters to perform time alignment and resampling processing on data with different sampling frequencies, and construct a time-series tensor and multimodal time-series data set under a unified time scale; S2. Based on the time series tensor of S1, the basic physical conservation relationships of mass, energy and momentum of the nuclear power cooling system are introduced as soft constraints to construct the physical constraint loss function and the total loss function. S3. Based on the multimodal time series data set of the normal operation samples generated by S1 and the physical constraint loss function constructed by S2, plus physical perturbation, an enhanced sample is constructed. S4. Extract dynamic features at different time scales from the enhanced samples generated in S3, introduce the physical correlation matrix to guide cross-modal feature fusion, and generate a physically consistent multimodal comprehensive feature vector. S5. Construct a physical knowledge graph of the nuclear power cooling system, map the multimodal fusion features output by S4 into symbolic representations with physical semantics, and embed them with the physical knowledge graph. Then, use a large language model for diagnosis to obtain fault diagnosis results. S6, based on the physical constraint loss function constructed by S2, performs uncertainty assessment and confidence calibration on the diagnostic results output by S5 through physical consistency, and outputs a diagnostic conclusion with uncertainty and confidence assessment.

2. The diagnostic method for nuclear power plant cooling systems that integrates physical constraints and large-scale models according to claim 1, characterized in that: In S1, the multi-source physical quantity time series data includes pressure, temperature, flow rate, and neutron flux data; the system operating parameters are one or more of the following: control rod position, main pump speed, main pump frequency command, pressure regulator spray valve opening, pressure regulator heater power, and makeup water flow rate. A running sample represents all multimodal data collected by the system within a continuous time window. The first running sample is defined as... A physical mode in time The observed values ​​are as follows Time alignment is handled through interpolation, and resampling is performed according to the following formula: In the formula, To unify the first on the timeline At a certain point in time, Indicates a unified point in time Upper m Estimation data for physical modes; For the first sample A point in time; Indicates in Time of the first m The actual measured values ​​of the physical modes; Represents the first of the original samples A point in time; The unified multimodal feature vector is represented as : Construct the temporal tensor for each running sample for: in, The total number of time steps under a unified time scale; The number of physical modes; Indicates the first A physical mode in time The observed values, the physical modes include at least one of the following: pressure, temperature, mass flow rate, neutron flux, and control parameters; Indicates time The joint observation vector at any given time is composed of multiple physical modes from various sources; Based on the above and , No. A unified multimodal time series dataset of each running sample. Represented as: Among them, X (i) Indicates the first The multimodal feature matrix of each running sample; The dimension of the multimodal features; N For the number of independently run samples, For the first The feature dimension of a physical mode, compared to the temporal tensor. , Emphasizing is the first A collection of multimodal time-series data from a set of running samples.

3. The diagnostic method for nuclear power plant cooling systems based on the fusion of physical constraints and large models according to claim 2, characterized in that: S2 includes the following steps: S2.1, Definition The system state variables at time t are: in, This is the system state vector; This refers to the pressure state in the nuclear power plant's cooling circuit. This refers to the coolant temperature state; This refers to the coolant mass flow rate status; This refers to the reactor neutron flux state; Further define the system control input vector for: in, The first part represents the nuclear power cooling system. System operating parameters; The changes in system state variables over time follow the dynamic evolution law under the coupling effect of thermal-hydraulic and control systems, and its state evolution relationship is described as follows: in, The rate of change of the state of the nuclear power system over time; It is a nonlinear physical evolution function composed of the mass conservation equation, the energy conservation equation, and the momentum conservation equation; This is the system control input vector; S2.

2. Based on the system state variables, the thermo-hydraulic conservation equations are introduced as constraints. The thermo-hydraulic conservation equations include mass, energy, and momentum conservation equations: The mass conservation equation is: in, This represents the rate of change of pressure with respect to time. Indicates the density of the coolant; Represents the velocity field; For leakage mass flow rate; It was Dirac function; r Let be the position coordinate vector of any point; r l The spatial coordinates of the leak point; For replenishing water mass flow rate; The energy conservation equation is: in, This refers to the specific heat capacity of the coolant under constant pressure. For temperature field; k Thermal conductivity; Represents the gradient; Represents the gradient of the temperature field; For volumetric heat source term, This refers to the system boundary heat loss rate; This indicates the rate of energy loss due to leakage; The momentum conservation equation is: in, Indicates a pressure field; μ Represents dynamic viscosity; g It is the vector of gravitational acceleration; This is the frictional resistance term; This refers to the momentum loss due to leakage. S2.3 Constructing a computable physical constraint loss function for: In the formula, Indicates the first A physical conservation residual function; Indicates the total number of physical constraints; Index for physical constraints; S2.4 Embed physical constraints into the total loss function : In the formula, These are the physical constraint weighting coefficients; For data-driven cross-entropy loss.

4. The diagnostic method for nuclear power plant cooling systems that integrates physical constraints and large-scale models according to claim 3, characterized in that: S3 includes the following steps: S3.1 The constructed enhanced sample is: in, For the generated enhanced samples; For physically guided perturbation terms; A coefficient used to control the amplitude of the disturbance; S3.2, Disturbance Term The formula is: In the formula, Indicates standard Gaussian noise. Indicates a disturbance that is physically constrained; It is the physical model relative to the input parameters The Jacobian matrix; It follows a standard normal distribution.

5. The diagnostic method for nuclear power plant cooling systems that integrates physical constraints and large models according to claim 4, characterized in that: S4 includes the following steps: S4.1 Multi-scale Feature Extraction Features at three scales are extracted using convolutional or recurrent networks with different receptive fields: , and Representing the time series tensor Results after extracting short-term, medium-term, and long-term features; , and These represent the feature extraction functions for the short, medium, and long term, respectively. S4.2 Multi-scale feature splicing The features from the three scales of S4.1 are concatenated into a complete feature vector. for: In the formula, ‖ represents the vector concatenation operation; S4.3, Physically Guided Cross-Modal Feature Fusion After extracting multi-scale features, attention fusion is performed based on the physical correlation matrix: When calculating cross-modal attention weights, the physical correlation matrix is ​​used. As a bias term, it is added to define the physical correlation matrix. for: In the formula, for and The strength of the physical relationship; and Both are physical mode indexes. For source mode; For the target mode; This is the expected value; For modality When changing, mode Sensitivity to changes; The formula for calculating cross-modal attention weights is: in, For attention weights, and These are the transformation matrices of the query vector and the key vector, respectively. Indicates the dimension of the key vector; It is a normalized exponential function; The weighted fusion formula is: In the formula, This is the multimodal comprehensive feature vector obtained after multi-scale feature extraction and cross-modal fusion; For the first A value vector of modalities.

6. The diagnostic method for nuclear power plant cooling systems that integrates physical constraints and large models according to claim 5, characterized in that: S5 includes the following steps: S5.1 Mapping of Multimodal Fusion Features to Physical Semantic Symbols Multimodal fusion features The mapping is to a set of symbols with physical semantics, as shown in the formula: in, It is a set of physical semantic symbol states; No. A semantic state description; It is a feature-to-semantic mapping function; Number of semantic states; S5.2 Embedding of Physics Knowledge Graph A graph neural network is used to embed the physical knowledge graph into a vector space, as shown in the formula: In the formula, Representing entities In the Layer embedding vectors, Representing entities In the Layer embedding vector; Representing entities In the Layer embedding vector; For layer index; For activation functions; A collection of relation types; It can be any type of relation; To pass through relationships With entity A set of connected neighboring entities; For neighboring entities; For the first Layer Relationship The transformation matrix; For the first Self-connected transformation matrix in the layer; S5.3, Joint Reasoning Diagnosis Combining the physical semantic symbols of S5.1 with the knowledge graph embedding of S5.2, the diagnostic results are obtained through a neural symbolic inference engine: the neural symbolic inference engine combines neural network features with knowledge graph embedding, as shown in the formula: In the formula, The diagnostic result is represented by Re(·), which is the inference function. For nuclear-related physical knowledge or parameters; These are the learnable parameters of the inference engine; MLP It is a multilayer perceptron; After neural network After layer propagation, each entity receives its final embedding. Aggregate related entities ,get ; S5.4 Output Constraints In the reasoning and diagnostic process, a physical consistency discriminant function is introduced. for: In the formula, This represents the first step in the generation process of the large language model. The words output in the step, This refers to words or symbols generated by a large language model. This represents the positional step in the sequence generated by the large language model. The corrected language model generation probability is: In the formula, Generate probabilities for large language models; This is a physical consistency discrimination function; This is a set of candidate semantics.

7. The diagnostic method for nuclear power plant cooling systems based on the fusion of physical constraints and large models according to claim 6, characterized in that: S6 includes the following steps: S6.1 Uncertainty Calculation Based on noise in the measurement data, model errors, and uncertainties in physical modeling, a comprehensive uncertainty is constructed. The metric is used to calibrate the reliability of diagnostic results, and the formula is: in, Due to noise uncertainty in measurement data; For model parameter uncertainty; To model the uncertainty of physical errors; S6.2 Confidence Calibration Confidence assessment is corrected using a physical consistency calibration function, the formula of which is: The calibrated confidence level; The confidence level of the model's original output; For calibration coefficients; S6.3, Output Diagnostic Output By combining the diagnostic results of S5 and the confidence calibration of S6, the final output includes diagnostic conclusions that include fault type, severity, evolution trend, reasoning path, confidence level, and uncertainty assessment.

Images