Turbine fault diagnosis system based on multi-parameter fusion monitoring

Abstract

The invention relates to the technical field of control and adjustment, in particular to a steam turbine fault diagnosis system based on multi-parameter fusion monitoring, which is characterized in that a quantum hybrid sensing module is used for establishing a picosecond-level space-time reference through entangled photon pair distribution, and unifying a time sequence label of multi-source heterogeneous data; the hypergraph feature extraction module loads the coefficient to a vibration measurement point vertex attribute channel, dynamically calculates a heat conduction hyperedge weight according to a rotor eccentricity inversion value, aggregates temperature field, vibration spectrum and magnetic domain features in a space-time hypergraph which retains physical topology, outputs a cross-scale aligned fusion feature tensor, and outputs a vibration measurement point vertex attribute channel; generating a diagnosis feature vector to drive a multi-scale twin experience certificate; and the federation decision module triggers a load reduction instruction according to the microscopic crack cloud picture to inhibit thermal stress sudden change, and meanwhile, the feature representation precision is improved through cloud federation optimization hyperedge weight parameters. According to the method, the problem of characterization distortion caused by time sequence asynchronization and scale mismatch of heat and machine coupling data is solved.

Description

Technical Field

[0001] The present invention relates to the technical field of control and regulation, and in particular to a steam turbine fault diagnosis system based on multi-parameter fusion monitoring. Background Art

[0002] Steam turbine fault diagnosis, as the core technology for key equipment condition monitoring, relies on high-precision sensors to continuously collect operating parameters, including vibration amplitude, spectrum, phase, as well as steam pressure, temperature, flow, cylinder efficiency, axial displacement, etc. Through time series analysis, pattern recognition and expert systems, it performs logical reasoning and characterization on abnormal signals that represent equipment health, builds a fault probability model, locates potential failed components and assesses the degree of degradation, and ultimately forms maintenance decision recommendations to optimize maintenance plans and maintain unit reliability.

[0003] In a turbine fault diagnosis system based on multi-parameter fusion monitoring, fault signature distortion results from time-series synchronization errors and inconsistent feature scales in heterogeneous multi-source data. This is specifically manifested by differences in sampling frequency and physical dimensions inherent in different sensors, such as the 10kHz high-frequency sampling of a vibration accelerometer and the 1Hz low-frequency update of a fiber Bragg grating temperature sensor. When the system forces a uniform time axis through linear interpolation and concatenates PCA-reduced feature vectors, high-frequency transient signals, such as temperature gradients during rotor thermal shock, may be averaged or phase-shifted. Furthermore, the scales of oil monitoring spectral data and the Hilbert envelope of the vibration spectrum are difficult to directly correlate. Consequently, the fused multimodal features cannot accurately represent the thermal-mechanical coupling effects under real-world operating conditions. For example, when assessing rotor misalignment, this distortion weakens the correlation between thermal deformation and mechanical vibration anomalies. Consequently, the calculated fault probability, when input into a fault tree analysis model, deviates from the actual degree of degradation, increasing the risk of misjudgment and significantly impacting diagnostic reliability. These defects stem from the conflict between the data acquisition mechanism under physical constraints and the ideal fusion algorithm. Summary of the Invention

[0004] In view of the shortcomings of the existing technology, the present invention provides a turbine fault diagnosis system based on multi-parameter fusion monitoring to solve the problem of thermal and mechanical coupling characterization distortion caused by timing asynchrony and feature scale mismatch in multi-source heterogeneous data fusion of supercritical turbine rotor systems.

[0005] In order to solve the above technical problems, the specific technical solutions of the present invention are as follows:

[0006] The steam turbine fault diagnosis system based on multi-parameter fusion monitoring provided by the present invention includes:

[0007] The quantum hybrid sensing module deploys diamond NV color center sensors and fiber quantum gyroscopes to form a hybrid sensing network that collects and outputs multi-source heterogeneous data of the rotor system;

[0008] a causal alignment module, receiving the multi-source heterogeneous data, establishing a thermal and mechanical coupling causal graph model, performing a timing alignment operation to generate and output a thermally induced vibration contribution coefficient;

[0009] A hypergraph feature extraction module inputs the thermally induced vibration contribution coefficient, constructs a spatiotemporal hypergraph structure, and aggregates the heat conduction and force and magnetic coupling features to obtain a fused feature tensor;

[0010] a neural symbolic fusion module that receives the fused feature tensor, compresses it into a core tensor that preserves thermal-magnetic-resonance correlation, and injects symbolic rules to constrain feature fusion to generate a diagnostic feature vector;

[0011] The digital twin diagnosis module inputs the diagnostic feature vector, calls the multi-scale twin to compare the feature vector with the simulation prediction to output the fault location and degradation level;

[0012] a federated decision module that receives the degradation level, generates a maintenance strategy, and dynamically optimizes load distribution to the turbine control system;

[0013] Among them, the quantum hybrid sensing module forms a closed-loop data flow to the federal decision-making module.

[0014] Furthermore, in the steam turbine fault diagnosis system based on multi-parameter fusion monitoring of the present invention, the quantum hybrid sensing module includes:

[0015] The quantum sensor array is embedded in the bearing seat surface to collect and output magnetic domain thermal fluctuation signals, and non-contactly monitors the rotor angular velocity quantum noise to generate raw quantum data;

[0016] A photon entanglement synchronization unit inputs the original quantum data, distributes entangled photon pairs to the sensor receiving end, and adds a picosecond time stamp tag to the data to generate synchronized data;

[0017] The data encapsulation unit receives the synchronization data, generates a quantum frame carrying a quantum state verification tag, partitions and stores the magnetic domain intensity, quantum noise spectrum and vibration waveform, and outputs the multi-source heterogeneous data.

[0018] Furthermore, in the steam turbine fault diagnosis system based on multi-parameter fusion monitoring of the present invention, the causal alignment module performs:

[0019] Thermodynamic index extraction: Analyze the steam temperature gradient from the quantum frame as the input of the thermodynamic layer; Counterfactual alignment operation: When the temperature rise rate of the steam temperature gradient exceeds the threshold, generate an adversarial spectrum based on the real-time vibration spectrum, calculate the Wasserstein distance between the actual spectrum and the adversarial spectrum, and output the thermally induced vibration contribution coefficient;

[0020] Quantum noise correction: When the quantum gyro noise in the quantum frame suddenly increases in the 50-150Hz frequency band, the pure angular velocity signal is reconstructed to correct the calculation results of the cylinder thermal expansion.

[0021] Furthermore, in the steam turbine fault diagnosis system based on multi-parameter fusion monitoring of the present invention, in the hypergraph feature extraction module:

[0022] A vertex attribute injection unit is used to load the thermally induced vibration contribution coefficient into the attribute channel of the vibration measurement point vertex;

[0023] The super-edge construction unit is based on the physical topological structure: the heat conduction super-edge connects the axially adjacent temperature measurement points, and the force and magnetic coupling super-edge connects the vibration measurement points of the same bearing seat and the magnetic domain unit;

[0024] The graph convolution engine aggregates the multimodal features of temperature, vibration, and magnetic domains within the hyperedge, calculates and outputs the fused feature tensor, including the coupling coefficient between the rotor eccentricity field induced by thermal expansion and the eddy current intensity of the worn magnetic domain.

[0025] Furthermore, in the steam turbine fault diagnosis system based on multi-parameter fusion monitoring of the present invention, in the neural symbolic fusion module:

[0026] A tensor construction unit receives the fused feature tensor and constructs a fourth-order feature tensor according to time, quantum state, physical quantity and spatial dimension;

[0027] Neural-Symbolic Interaction Unit, which injects symbolic rules: dynamically associates tensor components with fault predicate confidence, and reversely prunes modes that are not related to real-time faults;

[0028] A tensor compression engine decomposes the fourth-order eigentensor to generate a core tensor that retains thermal-magnetic-vibration correlation and outputs a diagnostic eigenvector.

[0029] Furthermore, in the steam turbine fault diagnosis system based on multi-parameter fusion monitoring described in the present invention, the digital twin diagnosis module includes:

[0030] Input the diagnostic feature vector and call the multi-scale twin cluster:

[0031] Molecular dynamics twinning analyzes magnetic domain fluctuation signals to simulate dislocation evolution, discrete element twinning analyzes oil wear debris transport trajectory, and finite element twinning calculates friction dynamics based on thermal expansion eccentric fields;

[0032] Dynamic verification unit: Compares the simulated prediction of the twin output with the measured vibration envelope energy. When the deviation exceeds the threshold, it adaptively reconstructs the fault tree nodes and generates fault location and degradation level.

[0033] Furthermore, in the steam turbine fault diagnosis system based on multi-parameter fusion monitoring of the present invention, the federal decision module performs:

[0034] receiving the degradation level, calling the neural symbolic engine at the edge layer to load the core tensor output by the neural symbolic fusion module, performing real-time reasoning and generating a primary diagnosis report;

[0035] The cloud-based federated aggregation module aggregates the encrypted fault case gradients of multiple power plants and uses homomorphic encryption weighted averaging to update the diagnostic model parameters;

[0036] The dynamic strategy generation unit integrates the primary diagnostic report and the updated model parameters, calculates the remaining life distribution according to the degradation level, generates a maintenance strategy and triggers steam parameter adjustment instructions to the turbine control system.

[0037] Furthermore, in the steam turbine fault diagnosis system based on multi-parameter fusion monitoring of the present invention, the symbol rule injection unit in the neural symbol fusion module performs:

[0038] The static rule loading unit loads the judgment criteria from the fault knowledge base: when the axial vibration 2× increase is greater than 30%, it is considered rotor misalignment, and when the temperature is greater than 120°C, it is considered bearing overheating;

[0039] The dynamic rule generation unit receives the crack cloud map output by the digital twin diagnosis module and generates a three-level warning rule when the eddy current dispersion degree is greater than 0.7;

[0040] The rule activation engine injects the judgment criteria and warning rules into the neural symbolic interaction unit, and constrains the association of tensor components with the confidence of the fault predicate.

[0041] Furthermore, in the steam turbine fault diagnosis system based on multi-parameter fusion monitoring of the present invention, the thermally induced vibration contribution coefficient is input into the vibration measurement point vertex attribute channel of the hypergraph feature extraction module to quantify the contribution intensity of thermal deformation to vibration;

[0042] The inversion value of the rotor eccentricity output by the quantum noise correction mechanism is input into the hyperedge construction unit to dynamically calculate the thermal conductivity weight of the heat conduction hyperedge;

[0043] Among them, the collaborative correlation between the vibration measurement point attributes and the heat conduction hyperedge weights eliminates the feature distortion caused by the scale mismatch of thermal-vibration data.

[0044] Furthermore, in the steam turbine fault diagnosis system based on multi-parameter fusion monitoring of the present invention, the dynamic strategy generation unit of the federal decision module receives the microcrack initiation probability cloud map output by the micro-twin, generates a load reduction instruction based on the degradation level, and sends it to the steam turbine control system;

[0045] The cloud-based federation aggregation module of the federated decision-making module generates hyperedge optimization parameters based on the gradient of encrypted fault cases of multiple power plants, outputs them to the hyperedge construction unit of the hypergraph feature extraction module, and dynamically updates the weight coefficients of the force and magnetic coupling hyperedges;

[0046] Among them, the load reduction instruction reduces the thermal stress mutation, the hyperedge weight optimization improves the accuracy of thermal and vibration correlation characterization, and synergistically suppresses thermal and mechanical coupling distortion.

[0047] Beneficial effects of the present invention:

[0048] The quantum hybrid sensing module of the present invention establishes a picosecond-level spatiotemporal benchmark through the distribution of entangled photon pairs, eliminating the inherent time scale difference between high-frequency vibration sampling and low-frequency temperature updates. The thermally induced vibration contribution coefficient and rotor eccentricity inversion value generated by the causal alignment module are respectively injected into the vertex attribute channel and hyperedge weight calculation unit of the hypergraph structure. By quantifying the thermal contribution of vibration characteristics and mapping the deformation response of the heat conduction path, the physical correlation calibration of the thermodynamic scale and the mechanical response scale is achieved. The federal decision-making layer suppresses the heat source distortion amplitude based on the load reduction instruction triggered by the microscopic crack cloud map. The hyperedge weight parameters optimized by the cloud-side federation continuously improve the feature fusion accuracy and collaboratively maintain the phase alignment state between thermal expansion and mechanical vibration. This multi-level collaborative mechanism constructs a complete closed loop from data synchronization, feature fusion to decision optimization, systematically suppressing the risk of cross-scale drift caused by thermal hysteresis and supporting the diagnostic accuracy of the fault tree model under strong thermal-mechanical coupling conditions. BRIEF DESCRIPTION OF THE DRAWINGS

[0049] In order to more clearly illustrate the technical solution of the present invention, the following is a brief introduction to the drawings required for use in the embodiments. Obviously, for ordinary technicians in this field, other drawings can be obtained based on the drawings without paying any creative labor.

[0050] Figure 1 This is a system architecture diagram of a steam turbine fault diagnosis system based on multi-parameter fusion monitoring provided by an embodiment of the present invention. DETAILED DESCRIPTION

[0051] In order to make the purpose, technical solutions and advantages of the present invention clearer, the technical solutions of the present invention will be clearly and completely described below in conjunction with the specific embodiments of the present invention and the corresponding drawings. Obviously, the embodiments described are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without making creative work are within the scope of protection of the present invention. The technical solutions provided by each embodiment of the present invention are described in detail below in conjunction with the drawings. In order to better understand the purpose of the present invention, the present invention is further described in detail below.

[0052] See also Figure 1 The steam turbine fault diagnosis system based on multi-parameter fusion monitoring provided by the present invention includes:

[0053] The quantum hybrid sensing module deploys diamond NV color center sensors and fiber quantum gyroscopes to form a hybrid sensing network that collects and outputs multi-source heterogeneous data of the rotor system;

[0054] a causal alignment module, receiving the multi-source heterogeneous data, establishing a thermal and mechanical coupling causal graph model, performing a timing alignment operation to generate and output a thermally induced vibration contribution coefficient;

[0055] A hypergraph feature extraction module inputs the thermally induced vibration contribution coefficient, constructs a spatiotemporal hypergraph structure, and aggregates the heat conduction and force and magnetic coupling features to obtain a fused feature tensor;

[0056] a neural symbolic fusion module that receives the fused feature tensor, compresses it into a core tensor that preserves thermal-magnetic-resonance correlation, and injects symbolic rules to constrain feature fusion to generate a diagnostic feature vector;

[0057] The digital twin diagnosis module inputs the diagnostic feature vector, calls the multi-scale twin to compare the feature vector with the simulation prediction to output the fault location and degradation level;

[0058] a federated decision module that receives the degradation level, generates a maintenance strategy, and dynamically optimizes load distribution to the turbine control system;

[0059] Among them, the quantum hybrid sensing module forms a closed-loop data flow to the federal decision-making module.

[0060] The quantum hybrid sensing module deploys a diamond NV color center sensor array on the surface of the turbine bearing seat to capture the thermal fluctuation signals of the microscopic magnetic domains of the bearing alloy in real time. At the same time, a fiber optic quantum gyroscope is installed at the rotor shaft end to monitor the angular velocity quantum noise. The two types of quantum sensors form a hybrid sensing network with traditional MEMS vibration sensors and fiber grating thermometers. The photon entanglement synchronization unit distributes entangled photon pairs to each sensor, establishes a unified picosecond time reference through Bell state measurement, and eliminates the timing drift caused by the inherent sampling frequency differences of the sensors. The data encapsulation unit encodes the synchronized magnetic domain intensity, quantum noise spectrum and vibration waveform partitions to generate quantum frames carrying quantum state verification tags as multi-source heterogeneous data output carriers.

[0061] The causal alignment module parses the steam temperature gradient time series from the quantum frame to establish a thermodynamic layer input benchmark. When it is detected that the temperature gradient temperature rise rate exceeds a critical threshold, the real-time temperature distribution state is frozen, and the pre-trained generative adversarial network is called to construct a virtual vibration spectrum without thermal distortion. The thermally induced vibration contribution coefficient is quantified by calculating the Wasserstein distance between the actual vibration spectrum and the virtual spectrum in the power frequency band. When the gyroscope noise in the quantum frame mutates in a specific frequency band, the quantum state tomography filtering mechanism is triggered to reconstruct the pure angular velocity signal and correct the rotor eccentricity inversion value output by the thermal expansion calculation model.

[0062] The hypergraph feature extraction module loads the thermally induced vibration contribution coefficient into the vertex attribute channel of the vibration measurement point, marking the contribution intensity of thermal deformation to the mechanical response; constructs a spatiotemporal hypergraph structure according to the physical topology of the rotor system: axially adjacent temperature measurement points are connected by heat conduction hyperedges, and vibration measurement points in the same bearing seat area are associated with magnetic domain units through force and magnetic coupling hyperedges; the graph convolution engine aggregates the temperature gradient, vibration spectrum and magnetic domain eddy current characteristics within the hyperedge, and outputs a fused feature tensor containing the thermal expansion eccentricity field and the wear magnetic domain coupling coefficient.

[0063] The neural symbolic fusion module constructs a fourth-order tensor by fusion feature tensors according to the four dimensions of time slice, quantum state, physical quantity type, and spatial position; the symbolic rule injection unit loads the static judgment rules of rotor misalignment and bearing overheating in the fault knowledge base, and at the same time receives the crack cloud map output by the digital twin module to generate dynamic warning rules; the neural symbolic interaction unit maps the symbolic rules to tensor components and prunes the characteristic modes that are not related to real-time faults; the tensor chain decomposition engine compresses the fourth-order tensor dimension, retains the core features related to heat, magnetism and vibration, and generates a diagnostic feature vector.

[0064] The digital twin diagnosis module inputs the diagnostic feature vector into the multi-scale twin cluster: molecular dynamics twins analyze the magnetic domain fluctuation signal to simulate the migration path of grain boundary dislocations, discrete element twins reconstruct the carrying trajectory of wear debris in the oil film, and finite element twins calculate the dynamic response of rotor and seal friction based on the thermal expansion eccentric field; the dynamic verification unit compares the twin simulation prediction with the actual vibration envelope energy. When the energy deviation in a specific frequency band exceeds the tolerance threshold, the fault tree node logic is adaptively reconstructed and the fault location coordinates and degradation level indicators are output.

[0065] The federated decision-making module loads core tensors at the edge layer to perform real-time neural symbolic reasoning, generating a primary diagnostic report based on degradation levels. The cloud-based federated aggregation module uses homomorphic encryption to weighted average the gradients of multiple power plant fault cases and update diagnostic model parameters. The dynamic strategy generation unit integrates real-time diagnostic results with federated knowledge to calculate the probability distribution of equipment remaining life, generate a load distribution optimization matrix, and trigger steam parameter adjustment instructions, forming a closed-loop decision chain from quantum sensing to control execution. This data flow permeates the thermodynamic-mechanical coupling relationship and, through a multi-level collaborative mechanism of quantum benchmark alignment, hypergraph topological constraints, symbolic rule guidance, twin verification feedback, and federated continuous optimization, eliminates the characteristic distortion caused by the scale mismatch between thermal and mechanical quantities in traditional methods.

[0066] Specifically, the steam turbine fault diagnosis system based on multi-parameter fusion monitoring of the present invention, the quantum hybrid sensing module includes:

[0067] The quantum sensor array is embedded in the bearing seat surface to collect and output magnetic domain thermal fluctuation signals, and non-contactly monitors the rotor angular velocity quantum noise to generate raw quantum data;

[0068] A photon entanglement synchronization unit inputs the original quantum data, distributes entangled photon pairs to the sensor receiving end, and adds a picosecond time stamp tag to the data to generate synchronized data;

[0069] The data encapsulation unit receives the synchronization data, generates a quantum frame carrying a quantum state verification tag, partitions and stores the magnetic domain intensity, quantum noise spectrum and vibration waveform, and outputs the multi-source heterogeneous data.

[0070] The quantum sensing array arranges diamond NV color center sensor units axially on the surface of the high-voltage rotor bearing seat, directly contacting the bearing alloy to capture the microscopic magnetic domain thermal fluctuation signal. At the same time, a fiber optic quantum gyroscope is coaxially installed at the free end of the rotor to non-contact measure the angular velocity quantum noise spectrum; the magnetic field pulse sequence and phase fluctuation output by the two types of quantum sensors constitute the original quantum data stream, in which the magnetic domain signal reflects the microscopic thermal stress state of the material, and the quantum noise spectrum characterizes the dynamic behavior of the rotor.

[0071] The photon entanglement synchronization unit receives the original quantum data stream, and the quantum master node emits polarization-entangled photon pairs, which are transmitted to the local receiving end of each sensor through polarization-maintaining optical fiber beam splitting; each receiving end performs a joint Bell state measurement, compares the arrival time difference between the local clock and the entangled photons, and generates a space-time reference tag with picosecond-level accuracy; this tag is bound to the original quantum data at the corresponding moment to generate synchronization data, eliminating the inherent time scale difference between high-frequency vibration sampling and low-frequency temperature updates.

[0072] The data encapsulation unit reads the synchronized data stream and allocates storage partitions based on the sensor's physical coordinate encoding rules: the magnetic domain intensity sequence is stored in the first data area, the quantum noise spectrum occupies the second data area, and the synchronized vibration waveform is stored in the third data area. Each partition header embeds a quantum state verification tag generated by hashing the polarization state of the entangled photons. Finally, the data is encapsulated into a quantum frame with spatiotemporal verifiable properties, which is output as multi-source heterogeneous data to downstream modules. The quantum frame's data structure establishes a mapping between physical dimensions and spatiotemporal positions, providing a hardware-level benchmark for thermal-mechanical coupling analysis.

[0073] The magnetic domain intensity region continuously records changes in the bearing alloy's microscopic magnetization state, the quantum noise spectrum region monitors the rotor's angular momentum fluctuations, and the vibration waveform region preserves the mechanical response time-domain signal; all three maintain physical and logical alignment at a unified timestamp. When the quantum frame is transmitted to the causal alignment module, the steam temperature gradient can be directly derived from the temperature sensor data associated with the vibration waveform region. The magnetic domain intensity and quantum noise spectrum contribute to the calculation of the causal strength at the thermodynamic level, forming a complete data supply chain from raw signal acquisition to multi-physical quantity fusion.

[0074] Specifically, in the steam turbine fault diagnosis system based on multi-parameter fusion monitoring of the present invention, the causal alignment module performs:

[0075] Thermodynamic index extraction: Analyzing the steam temperature gradient from the quantum frame as the thermodynamic layer input;

[0076] Counterfactual alignment operation: when the temperature rise rate of the steam temperature gradient exceeds a threshold, an adversarial spectrum is generated based on the real-time vibration spectrum, and the Wasserstein distance between the actual spectrum and the adversarial spectrum is calculated to output a thermally induced vibration contribution coefficient;

[0077] Quantum noise correction: When the quantum gyro noise in the quantum frame suddenly increases in the 50-150Hz frequency band, the pure angular velocity signal is reconstructed to correct the calculation results of the cylinder thermal expansion.

[0078] The temperature sensor time series in the third data zone of the quantum frame is analyzed using spectral analysis to analyze the axial steam temperature gradient distribution. This gradient serves as the input variable for the thermodynamic layer. The temperature gradient is divided into discrete segments along the rotor axis, and the differences between these segments form a sequence of thermal stress indicators for thermal expansion of the drive cylinder. The analysis process synchronously reads the quantum frame's spatiotemporal tags to ensure that the temperature data and vibration spectrum data are within the same observation period.

[0079] When the rate of rise of the steam temperature gradient in a specific axial section exceeds a preset threshold, the real-time temperature distribution is frozen. A pre-trained generative adversarial network is used to construct a virtual vibration spectrum. This process maintains other operating parameters unchanged, eliminates only the temperature mutation factor, and outputs the vibration spectrum under ideal operating conditions as an adversarial sample. The amplitude envelopes of the actual operating vibration spectrum and the adversarial spectrum in the operating frequency band (1× / 2×) are extracted, and the Wasserstein distance between the two envelopes is calculated. This distance maps the contribution of thermal stress to vibration, and after normalization, the thermally induced vibration contribution coefficient is output.

[0080] If the quantum gyro noise power spectrum in the second data area of ​​the quantum frame shows an abnormal peak in the 50-150 Hz frequency range, it is determined that external electromagnetic interference has caused angular velocity signal distortion. The quantum state tomography reconstruction program is triggered: at least 100 sets of orthogonal projection measurement data are collected, and a pure angular velocity fluctuation curve is reconstructed through maximum likelihood estimation. The corrected angular velocity signal is input into the thermal expansion calculation model to update the inverted value of the rotor radial eccentricity. The updated thermal expansion replaces the original calculation result in the subsequent causal diagram construction.

[0081] The temperature gradient sequence in the thermodynamic layer initiates a counterfactual alignment operation, and the output thermally induced vibration contribution coefficient transmits the strength of the thermal-mechanical correlation to downstream hypergraph modules. The angular velocity signal reconstructed by the quantum noise correction step improves the reliability of the thermal expansion calculation. The correction value participates in the causal topology update of the thermodynamic layer in the new period. The temperature gradient, thermal expansion, and vibration contribution form a closed-loop verification chain, eliminating coupling distortion caused by sensor drift and electromagnetic interference at the source of the data.

[0082] Specifically, in the steam turbine fault diagnosis system based on multi-parameter fusion monitoring of the present invention, in the hypergraph feature extraction module:

[0083] A vertex attribute injection unit is used to load the thermally induced vibration contribution coefficient into the attribute channel of the vibration measurement point vertex;

[0084] The super-edge construction unit is based on the physical topological structure: the heat conduction super-edge connects the axially adjacent temperature measurement points, and the force and magnetic coupling super-edge connects the vibration measurement points of the same bearing seat and the magnetic domain unit;

[0085] The graph convolution engine aggregates the multimodal features of temperature, vibration, and magnetic domains within the hyperedge, calculates and outputs the fused feature tensor, including the coupling coefficient between the rotor eccentricity field induced by thermal expansion and the eddy current intensity of the worn magnetic domain.

[0086] The causal alignment module receives the sequence of thermally induced vibration contribution coefficients transmitted and maps these coefficients to the corresponding vibration measurement point vertices in the bearing seat area based on the spatial coordinate encoding of the measurement points recorded within the quantum frame. A dedicated attribute channel is created for each vibration measurement point vertex, and the thermally induced vibration contribution coefficient value at that measurement point is loaded as an additional vertex attribute. This coefficient attribute quantifies the influence of thermal deformation at the real-time spatial location on mechanical vibration, establishing a feature association between the thermodynamic layer and the mechanical response layer.

[0087] The hyperedge network is deployed based on the actual spatial structure of the rotor and bearing system. Axially adjacent temperature measurement points are connected via a heat conduction hyperedge, extending along the steam flow direction. The transmission path reflects the direction of thermal expansion and conduction of the cylinder body. Vibration measurement points and magnetic domain units in the same bearing seat coordinate system are interconnected via force and magnetic coupling hyperedges. The hyperedge coverage area corresponds to the metal friction interface, and the transmission path expresses the energy transfer relationship between mechanical vibration and magnetic domain reversal. Hyperedge weights are dynamically initialized based on the physical distance between the sensors.

[0088] By traversing each heat conduction hyperedge unit in the hyperedge structure, the axial gradient characteristics of the connected temperature measurement point vertices are aggregated. Simultaneously traversing the force and magnetic coupling hyperedge units, the power frequency spectrum characteristics of the vibration measurement points are integrated with the eddy current intensity characteristics of the magnetic domain units. A gated graph convolutional network performs message passing within the hyperedge domain, with the output channels generating a thermal expansion eccentricity field tensor and a wear magnetic domain coupling coefficient matrix. These two matrices are stacked along the time dimension to form a fused feature tensor. The eccentricity field tensor indicates the distribution of the rotor's axis offset caused by thermal deformation, while the coupling coefficient matrix represents the efficiency of converting vibration energy into magnetic field under wear conditions.

[0089] The thermally induced vibration contribution coefficients loaded by the vertex attribute injection unit enhance the physical interpretability of vibration measurement points, providing feature priors for subsequent hyperedge aggregation. The spatial topological constraints preserved by the hyperedge construction unit avoid the loss of physical location information caused by data flattening in traditional methods. The fused feature tensor output by the graph convolution engine integrates the heat conduction path and the correlation between force and magnetic coupling into a unified representation framework, forming a parameterized representation of the spatiotemporal alignment of the thermal expansion eccentric field and the wear magnetic domain eddy current. When this feature tensor is transmitted downstream, the temperature distribution feature dimension it carries supports thermodynamic analysis, and the vibration and magnetic domain correlation dimension carries mechanical failure assessment, fully preserving the physical nature of thermal and mechanical coupling.

[0090] Specifically, in the steam turbine fault diagnosis system based on multi-parameter fusion monitoring of the present invention, in the neural symbolic fusion module:

[0091] A tensor construction unit receives the fused feature tensor and constructs a fourth-order feature tensor according to time, quantum state, physical quantity and spatial dimension;

[0092] Neural-Symbolic Interaction Unit, which injects symbolic rules: dynamically associates tensor components with fault predicate confidence, and reversely prunes modes that are not related to real-time faults;

[0093] A tensor compression engine decomposes the fourth-order eigentensor to generate a core tensor that retains thermal-magnetic-vibration correlation and outputs a diagnostic eigenvector.

[0094] The tensor construction unit receives the fused feature tensor transmitted by the hypergraph feature extraction module and slices it along the time axis to extract feature snapshots for continuous monitoring periods. The observation sequence is divided according to the quantum state evolution stage, and the temperature field, vibration spectrum, and magnetic domain eddy current channel are separated according to the type of physical quantity. A four-dimensional index is established based on the sensor spatial topological coordinates. Feature components are arranged along the time, quantum state, physical quantity, and spatial dimensions to construct a fourth-order feature tensor data structure, achieving structured organization of multi-period and multi-physical quantity observation data.

[0095] The neural-symbolic interaction unit loads a static rule set from the fault knowledge base, including rotor misalignment determination logic (activated when the amplitude growth rate of 2x the power frequency exceeds 30%) and bearing overheating identification conditions (triggered by a temperature value >120°C for 5 seconds). It dynamically receives microcrack cloud images transmitted by the digital twin module and generates eddy current dispersion warning rules when the crack density distribution exceeds a critical threshold. The static determination conditions and dynamic warning rules are mapped to corresponding components of a fourth-order eigenvalue tensor: the temperature channel is associated with the overheating rule confidence, the vibration component is bound to the rotor misalignment detection probability, and the magnetic domain channel is linked to the crack warning weight. Eigenmodes unrelated to the real-time active rules are back-pruned, removing redundant data channels not associated with any fault predicates.

[0096] The tensor compression engine performs iterative rank-adaptive decomposition on the reduced fourth-order feature tensor, retaining a core feature subset consisting of temperature gradient amplitude, vibration power frequency energy, and magnetic domain eddy current intensity. This subset enforces constraints on the associated eigenmodes of heat conduction (temperature gradient), mechanical response (vibration energy), and material wear (magnetic domain eddy current), generating a core tensor that preserves the essence of thermal-magnetic-vibration physics. The core tensor is flattened into a diagnostic feature vector, whose dimensional order corresponds to quantitative features such as temperature field anomaly indicators, vibration spectrum distortion, and magnetic domain fluctuation intensity. This vector provides the digital twin module with a standard input structure that simultaneously carries physical rule constraints and state quantification features.

[0097] The temperature-dependent components of the diagnostic feature vector are input into the molecular dynamics twin to simulate thermal stress distribution. The vibration feature components drive the finite element twin to calculate dynamic response. Magnetic domain data guides the discrete element twin to analyze wear evolution. Quantitative metrics from the vector output enable the twin to accurately replicate the thermal-magnetic-vibration coupling process. The rule engine prunes uncorrelated features to avoid wasting twin simulation resources, while core features retain physical correlations to ensure simulation fidelity, forming a closed technical loop from data fusion to simulation verification.

[0098] Specifically, the steam turbine fault diagnosis system based on multi-parameter fusion monitoring described in the present invention, the digital twin diagnosis module includes:

[0099] Input the diagnostic feature vector and call the multi-scale twin cluster:

[0100] Molecular dynamics twinning analyzes magnetic domain fluctuation signals to simulate dislocation evolution, discrete element twinning analyzes oil wear debris transport trajectory, and finite element twinning calculates friction dynamics based on thermal expansion eccentric fields;

[0101] Dynamic verification unit: Compares the simulated prediction of the twin output with the measured vibration envelope energy. When the deviation exceeds the threshold, it adaptively reconstructs the fault tree nodes and generates fault location and degradation level.

[0102] A diagnostic feature vector is fed into a multi-scale twin cluster. The temperature field components, vibration spectrum characteristics, and magnetic domain eddy current data within this vector activate simulation processes at different scales. Molecular dynamics twins interpret the magnetic domain fluctuation signal components, simulate the migration paths of atomic dislocations at the bearing alloy grain boundaries, and output a probability cloud map of microcrack initiation locations, quantifying the extent of material microscopic damage. Discrete element twins receive parameters for metal particle concentration in the oil and reconstruct the transport trajectory and deposition hotspot distribution of wear debris in the lubricant film. Finite element twins load thermal expansion eccentricity field parameters to calculate the axis trajectory offset caused by rotor thermal deformation and predict the dynamic response of seal impact.

[0103] The dynamic verification unit collects simulation prediction data output by the twin cluster: molecular dynamics crack cloud maps mark microscopic damage coordinates, discrete element hotspot distributions reflect wear particle deposition areas, and finite element axis trajectory offsets predict macroscopic deformation. Simultaneously, the measured vibration envelope energy—derived from the power frequency spectrum characteristics processed by the causal alignment module—is acquired. The simulated axis trajectory offset is time-series compared with the actual vibration envelope energy distribution in specific frequency bands (specifically 2× the power frequency), and the root mean square deviation of the energy values ​​in each frequency band is calculated.

[0104] When the energy deviation in a specific frequency band exceeds the preset tolerance threshold, the fault tree adaptive reconstruction process is triggered: the vibration frequency band component with the largest deviation is extracted, and the micro-crack probability value and wear debris deposition density of the corresponding spatial position are reversely retrieved; the fault tree node logic is updated, and when the crack probability cloud map shows local aggregation and wear debris deposition exceeds the limit, a new bearing fatigue crack fault branch is added; the degradation level index is generated by integrating the three factors of micro-crack density, wear debris deposition and macro-vibration deviation, and the fault location coordinates are output to mark the specific quadrant position of the bearing seat and the axial coordinates of the rotor.

[0105] The microscopic crack location in the fault location coordinates is fed back to the neural-symbolic fusion module, which dynamically generates eddy current dispersion warning rules. Wear debris deposition information synchronously updates the discrete element twin boundary conditions. Macroscopic vibration characteristics trigger the finite element model re-meshing. This closed-loop verification mechanism ensures that simulation predictions consistently approximate the real physical state, suppressing model drift caused by thermal and mechanical coupling distortion through cross-scale data interaction.

[0106] Specifically, in the steam turbine fault diagnosis system based on multi-parameter fusion monitoring of the present invention, the federal decision module performs:

[0107] receiving the degradation level, calling the neural symbolic engine at the edge layer to load the core tensor output by the neural symbolic fusion module, performing real-time reasoning and generating a primary diagnosis report;

[0108] The cloud-based federated aggregation module aggregates the encrypted fault case gradients of multiple power plants and uses homomorphic encryption weighted averaging to update the diagnostic model parameters;

[0109] The dynamic strategy generation unit integrates the primary diagnostic report and the updated model parameters, calculates the remaining life distribution according to the degradation level, generates a maintenance strategy and triggers steam parameter adjustment instructions to the turbine control system.

[0110] The edge layer's neural symbolic engine receives the degradation level signal and loads the core tensor data transmitted by the neural symbolic fusion module. This tensor, which retains three-dimensional features such as temperature gradient amplitude, vibration power frequency energy, and magnetic domain eddy current intensity, serves as input. It then performs real-time reasoning based on symbolic rule constraints. It analyzes the rotor misalignment logic to activate the axial vibration alarm, identifies bearing overheating rules to trigger the temperature limit warning, and outputs a preliminary diagnostic report containing the fault location and confidence level. The reasoning process utilizes a local fault knowledge base, ensuring response time within the decision cycle.

[0111] The cloud-based federated aggregation module establishes an encrypted communication channel to receive diagnostic model gradient data uploaded by edge nodes across multiple power plants. This gradient information contains characteristic patterns from historical failure cases of similar units, and homomorphic encryption is used to blind the gradient matrix. The cloud server performs weighted average aggregation, dynamically assigning weights based on the similarity of power plant unit parameters to generate global model update parameters. These updated parameters are distributed to each edge node via a secure channel, simultaneously refreshing the rule confidence thresholds of the local neural symbolic engine.

[0112] The dynamic strategy generation unit integrates the primary diagnostic report from the edge layer with the model parameters updated in the cloud, and calculates the probability distribution of the remaining life of the equipment based on the degradation level index. This calculation introduces a wear accumulation model and a thermal fatigue damage algorithm: the bearing area integrates the crack initiation probability and wear debris deposition rate, and the rotor system integrates the eccentricity increase and vibration energy trend. The generated maintenance strategy matrix includes downtime window recommendations and component replacement priority ranking, and simultaneously triggers steam parameter adjustment instructions—reducing the main steam temperature change rate when thermal deformation exceeds the limit and limiting the power ramp rate when load fluctuations exceed the limit. The instructions are transmitted to the turbine DCS controller via hard wiring, forming a decision-making closed loop from state assessment to execution control.

[0113] The thermal-magnetic-vibration correlation provided by the core tensor supports accurate lifespan calculations, while cross-plant case studies introduced by federated aggregation enhance model generalization. Steam parameter commands dynamically suppress sudden thermal stress fluctuations, fundamentally reducing the risk of timing distortion in subsequent data collection cycles. This creates a negative feedback loop of "local diagnosis - cloud optimization - control execution," maintaining stable system operation through decision-making interventions.

[0114] Specifically, in the steam turbine fault diagnosis system based on multi-parameter fusion monitoring of the present invention, the symbol rule injection unit in the neural symbol fusion module performs:

[0115] The static rule loading unit loads the judgment criteria from the fault knowledge base: when the axial vibration 2× increase is greater than 30%, it is considered rotor misalignment, and when the temperature is greater than 120°C, it is considered bearing overheating;

[0116] The dynamic rule generation unit receives the crack cloud map output by the digital twin diagnosis module and generates a three-level warning rule when the eddy current dispersion degree is greater than 0.7;

[0117] The rule activation engine injects the judgment criteria and warning rules into the neural symbolic interaction unit, and constrains the association of tensor components with the confidence of the fault predicate.

[0118] The static rule loading unit accesses the local fault knowledge base and retrieves predefined physical failure criteria. The rotor misalignment determination rule is defined as: a fault condition is triggered when the axial vibration monitoring point's spectral energy increase by twice the power frequency exceeds the baseline value by 30%. The bearing overheating identification rule is set as: an alarm condition is activated when the bearing seat temperature sensor reading exceeds 120°C for 5 consecutive seconds. The loaded rules are converted into predicate logic structures, where the vibration increase threshold is associated with the output channel of the spectrum analysis module, and the temperature threshold is bound to the temperature sensor data stream.

[0119] The dynamic rule generation unit receives the crack cloud map data transmitted by the digital twin diagnosis module in real time. The cloud map is output through molecular dynamics twin simulation and contains the density distribution and position coordinates of the micro cracks in the bearing alloy. The discrete distribution value of the crack density in the radial quadrant of the bearing seat is analyzed. When the crack density concentration in a specific quadrant exceeds the critical threshold (for example, the density in the third quadrant accounts for more than 70%), a three-level warning rule with an eddy current dispersion greater than 0.7 is generated. The warning rule marks the trend characteristics of the fault development stage, and its parameters are strongly correlated with the spatial distribution of the crack cloud map.

[0120] The rule activation engine logically integrates statically loaded rotor misalignment criteria and bearing overheat identification conditions with dynamically generated warning rules. Injecting a neural symbolic interaction unit (NSI) creates a triple constraint mechanism: the rotor misalignment rule is bound to the double-power frequency channel of the vibration spectrum component in the fourth-order eigentensor; the bearing overheat rule is associated with the peak monitoring channel of the temperature gradient component; and the eddy current dispersion warning rule is linked to the spectral broadening dimension of the magnetic domain fluctuation component. By dynamically setting confidence thresholds, the activation weight of the rotor misalignment predicate is increased when the vibration channel data exceeds a set increase, and the constraint strength of the warning rule on the tensor component is strengthened when the spectral broadening of the magnetic domain channel exceeds a limit.

[0121] The warning parameters generated by dynamic rules are fed back to the digital twin module, triggering molecular dynamics simulation to increase the mesh resolution in the high-risk quadrant. The criteria for static rule updates are synchronized with the federated decision module to optimize the alarm thresholds for edge-layer neural symbolic reasoning. Constraint signals output by the rule activation engine guide the tensor compression engine to prune irrelevant modes: when the warning rule is inactive, high-frequency noise channels in the magnetic domain fluctuation component are removed; when the bearing overheating rule is triggered, the full-band characteristics of the temperature gradient component are retained. This mechanism ensures that the core tensor always retains feature dimensions strongly correlated with the active fault predicate, enabling the diagnostic feature vector to accurately characterize abnormal thermal and mechanical coupling states.

[0122] Specifically, in the steam turbine fault diagnosis system based on multi-parameter fusion monitoring of the present invention, the thermally induced vibration contribution coefficient is input into the vibration measurement point vertex attribute channel of the hypergraph feature extraction module to quantify the contribution intensity of thermal deformation to vibration;

[0123] The inversion value of the rotor eccentricity output by the quantum noise correction mechanism is input into the hyperedge construction unit to dynamically calculate the thermal conductivity weight of the heat conduction hyperedge;

[0124] Among them, the collaborative correlation between the vibration measurement point attributes and the heat conduction hyperedge weights eliminates the feature distortion caused by the scale mismatch of thermal-vibration data.

[0125] The thermally induced vibration contribution coefficient, a physical indicator that quantifies the impact of thermal deformation on mechanical vibration, is input into the hypergraph feature extraction module's vertex attribute channel for vibration measurement points. This coefficient value is bound to the bearing seat's spatial coordinates corresponding to the measurement point position, imbuing the vibration vertex with thermodynamic attribute features. When the graph convolution engine aggregates the vibration spectrum, the coefficient value in the vertex attribute channel is used as a weighting factor in the calculation, ensuring that the vibration frequency domain features carry a quantitative label for the thermal stress contribution, establishing an explicit association between the vibration signal and the temperature gradient.

[0126] The inverted rotor eccentricity value output by the quantum noise correction mechanism reflects the degree of shaft deformation caused by thermal expansion. This value is input into the hyperedge weight calculation engine of the hyperedge construction unit. The heat conduction hyperedge weights between adjacent axial temperature measurement points are dynamically updated based on the inverted eccentricity value: when the eccentricity increases significantly, the hyperedge weight is increased to strengthen the contribution of the heat conduction path in the feature aggregation; when the eccentricity is stable, the weight is reduced to weaken the interference of non-critical heat transfer paths. This weight adjustment enables coupled modeling of the heat conduction process and mechanical deformation, allowing the hyperedge structure to respond to changes in the thermal stress state.

[0127] The thermal contribution coefficient in the vertex attribute channel of the vibration measurement point and the heat conduction hyperedge weight jointly construct a scale normalization model: the vertex attribute quantifies the impact of thermal deformation to the vibration spectrum dimension (mechanical response scale), and the hyperedge weight maps the rotor deformation to the heat transfer path (thermodynamic scale). When the graph convolution engine performs feature aggregation within the hyperedge, the temperature gradient feature is multiplied by the hyperedge weight to obtain the thermodynamic scale conversion value, and the vibration spectrum feature is superimposed on the vertex attribute value to complete the mechanical scale calibration. The two sets of features are spatially convolved under the same physical dimension, and the rotor eccentricity field in the output fused feature tensor simultaneously represents the thermal expansion amplitude and the vibration response intensity, eliminating the feature mismatch caused by physical dimension differences in traditional methods.

[0128] The vertex attribute channel continuously receives updated thermal vibration contribution coefficients from the causal alignment module, dynamically correcting the thermal influence weights of vibrating vertices. The hyperedge construction unit adjusts the credibility of the heat conduction path in real time based on the eccentricity inversion value fed back by quantum noise correction. The hypergraph structure calibrates the dual paths of thermal-vibration correlations, ensuring that vibration spectrum anomalies during temperature transients are accurately identified as thermal distortion (such as rotor thermal bending), avoiding misdiagnosis as mechanical failure. This scale fusion method, based on physical mechanisms, eliminates the risk of cross-scale distortion in thermal-mechanical coupling analysis at the data representation level.

[0129] Specifically, the steam turbine fault diagnosis system based on multi-parameter fusion monitoring of the present invention, the dynamic strategy generation unit of the federal decision module receives the microcrack initiation probability cloud map output by the micro-twin, generates a load reduction instruction based on the degradation level, and sends it to the steam turbine control system;

[0130] The cloud-based federation aggregation module of the federated decision-making module generates hyperedge optimization parameters based on the gradient of encrypted fault cases of multiple power plants, outputs them to the hyperedge construction unit of the hypergraph feature extraction module, and dynamically updates the weight coefficients of the force and magnetic coupling hyperedges;

[0131] Among them, the load reduction instruction reduces the thermal stress mutation, the hyperedge weight optimization improves the accuracy of thermal and vibration correlation characterization, and synergistically suppresses thermal and mechanical coupling distortion.

[0132] The dynamic strategy generation unit receives the microcrack initiation probability cloud map output by the microtwin and analyzes the crack density distribution values ​​in different quadrants of the bearing working surface. Combined with the degradation level indicator (combining the crack density threshold and the wear accumulation rate calculation), when the crack density in a specific quadrant exceeds the critical threshold and the degradation level enters the acceleration period, a gradient load reduction instruction matrix is ​​generated. This matrix contains the main steam temperature drop rate limit and the power ramp rate upper limit. The instructions are transmitted to the turbine DCS controller via hard wiring. The execution process adopts a segmented smooth adjustment mode to stabilize the cylinder thermal stress change rate within the material tolerance range.

[0133] The cloud-based federated aggregation module collects encrypted gradient data from fault cases across multiple power plants. This gradient is derived from the core tensor feature patterns of the neural symbolic fusion module under historical fault conditions of similar units. Hyperedge optimization parameters are generated through homomorphic encrypted weighted averaging. For the force-magnetic coupling hyperedge structure, the median of the vibration-magnetic domain spectral coherence coefficient from bearing wear cases is extracted as a baseline reference value. These parameters are then transmitted to the hyperedge construction unit via a quantum encrypted channel, replacing the initial weight coefficients of the original heat conduction hyperedge. This update triggers the hypergraph feature extraction module to reconstruct the hyperedge connection strengths, ensuring that the force-magnetic coupling hyperedge weights reflect the common wear characteristics of the unit cluster.

[0134] The load reduction instruction reduces the amplitude of the steam temperature mutation and weakens the thermal shock intensity of the rotor from the heat source input side. This operation reduces the temperature gradient peak by about 40%, slows down the thermal expansion rate of the cylinder, and makes the rotor eccentricity required for the calculation of the heat conduction hyperedge weight more stable. The optimization of the hyperedge weight coefficient improves the accuracy of the vibration-magnetic domain correlation characterization: the updated weight strengthens the main frequency correlation between high-frequency vibration and magnetic domain eddy current, and weakens the secondary frequency interference caused by load fluctuations. The thermodynamic side (temperature stress) and the mechanical side (vibration response) are aligned on a unified time scale. In the fused feature tensor output by the hypergraph convolution, the phase deviation between the thermal expansion eccentricity field and the wear eddy current intensity is reduced to less than 15 milliradians.

[0135] Real-time load reduction records are fed into the causal alignment module of the next monitoring cycle, serving as the basis for dynamic adjustment of the steam temperature rise rate threshold. Hyperedge weight update parameters are synchronously fed back to the discrete element simulation unit of the digital twin module to optimize the calculation model of the oil wear debris transport trajectory. This forms a negative feedback loop of "decision intervention → parameter optimization → model update → data calibration," enabling the system to maintain a coupled balance of thermal and mechanical quantities over the course of a continuous operating cycle. This closed loop compresses the vibration characteristic drift caused by thermal hysteresis in traditional methods within a 5% tolerance band, significantly improving the positioning accuracy of the fault tree analysis model.

[0136] To address the distortion problem of thermal and mechanical coupling characterization in supercritical steam turbine rotor systems, this technical solution builds a complete technical closed loop through a three-level collaborative mechanism:

[0137] The quantum hybrid sensing module deploys a hybrid network of diamond NV color center sensors and fiber quantum gyroscopes, establishing a picosecond time reference through the distribution of entangled photon pairs. A photon entanglement synchronization unit adds verifiable spatiotemporal tags to high-frequency vibration signals, low-frequency temperature updates, and magnetic domain quantum states, encapsulating and generating quantum frames with quantum state verification. This frame serves as a unified carrier for multi-source heterogeneous data, eliminating the phase shift risk associated with traditional linear interpolation and providing a physically verifiable synchronization data foundation for the causal alignment module.

[0138] The causal alignment module analyzes the temperature gradient sequence in the quantum frame, generating thermally induced vibration contribution coefficients that are loaded into the hypergraph's vibration measurement point vertex attribute channels to quantify the contribution of thermal deformation to mechanical vibration. Simultaneously, the inverted rotor eccentricity values ​​are used to dynamically calculate heat conduction hyperedge weights, establishing a physical mapping between thermal expansion and structural deformation. The hypergraph feature extraction module, using a heterogeneous graph convolution engine, aggregates three channel features: temperature gradient (thermodynamic scale), vertex attributes (vibration contribution calibration values), and hyperedge weights (thermal deformation correlation) within a spatiotemporal hypergraph that preserves physical topology, outputting a cross-scale aligned fused feature tensor.

[0139] The federated decision-making module generates load reduction instructions based on microcrack probability cloud maps, reducing the intensity of thermal stress excitation caused by sudden steam temperature changes. The cloud-based federated aggregation module integrates the encrypted gradients of multiple power plant fault cases to generate optimized parameters for the force-magnetic coupling hyperedge weights, which are fed back to the hyperedge construction unit. The former reduces the thermal source distortion amplitude at the feature acquisition end, while the latter continuously optimizes the scale conversion accuracy of the feature fusion module. These two paths work together to compress the deviation between the cross-scale simulated and measured vibration envelope energy of the digital twin module to within the tolerance band, eliminating the coupled drift between thermal hysteresis and mechanical response.

[0140] In summary, from the establishment of quantum-level timing benchmarks, the fusion of feature scales under the hypergraph topology to the coordinated regulation of parameters and loads in the federated closed loop, a complete technology chain covering data collection, feature fusion and decision optimization has been formed, systematically overcoming the cross-scale mismatch problem of thermal and mechanical quantities.

Claims

1. The steam turbine fault diagnosis system based on multi-parameter fusion monitoring is characterized by: include: The quantum hybrid sensing module deploys diamond NV color center sensors and fiber quantum gyroscopes to form a hybrid sensing network that collects and outputs multi-source heterogeneous data of the rotor system; a causal alignment module, receiving the multi-source heterogeneous data, establishing a thermal and mechanical coupling causal graph model, performing a timing alignment operation to generate and output a thermally induced vibration contribution coefficient; A hypergraph feature extraction module inputs the thermally induced vibration contribution coefficient, constructs a spatiotemporal hypergraph structure, and aggregates the heat conduction, force, and magnetic coupling features to obtain a fused feature tensor; a neural symbolic fusion module that receives the fused feature tensor, compresses it into a core tensor that preserves thermal-magnetic-resonance correlation, and injects symbolic rules to constrain feature fusion to generate a diagnostic feature vector; The digital twin diagnosis module inputs the diagnostic feature vector, calls the multi-scale twin to compare the feature vector with the simulation prediction to output the fault location and degradation level; a federated decision module that receives the degradation level, generates a maintenance strategy, and dynamically optimizes load distribution to the turbine control system; Among them, the quantum hybrid sensing module forms a closed-loop data flow to the federal decision-making module.

2. The steam turbine fault diagnosis system based on multi-parameter fusion monitoring according to claim 1 is characterized in that: The quantum hybrid sensing module includes: The quantum sensor array is embedded in the bearing seat surface to collect and output magnetic domain thermal fluctuation signals, and non-contactly monitors the rotor angular velocity quantum noise to generate raw quantum data; A photon entanglement synchronization unit inputs the original quantum data, distributes entangled photon pairs to the sensor receiving end, and adds a picosecond time stamp tag to the data to generate synchronized data; The data encapsulation unit receives the synchronization data, generates a quantum frame carrying a quantum state verification tag, partitions and stores the magnetic domain intensity, quantum noise spectrum and vibration waveform, and outputs the multi-source heterogeneous data.

3. The steam turbine fault diagnosis system based on multi-parameter fusion monitoring according to claim 2 is characterized in that: The causal alignment module performs: Thermodynamic index extraction: Analyzing the steam temperature gradient from the quantum frame as the thermodynamic layer input; Counterfactual alignment operation: when the temperature rise rate of the steam temperature gradient exceeds a threshold, an adversarial spectrum is generated based on the real-time vibration spectrum, and the Wasserstein distance between the actual spectrum and the adversarial spectrum is calculated to output a thermally induced vibration contribution coefficient; Quantum noise correction: When the quantum gyro noise in the quantum frame suddenly increases in the 50-150Hz frequency band, the pure angular velocity signal is reconstructed to correct the calculation results of the cylinder thermal expansion.

4. The steam turbine fault diagnosis system based on multi-parameter fusion monitoring according to claim 3 is characterized in that: In the hypergraph feature extraction module: A vertex attribute injection unit is used to load the thermally induced vibration contribution coefficient into the attribute channel of the vibration measurement point vertex; The super-edge construction unit is based on the physical topological structure: the heat conduction super-edge connects the axially adjacent temperature measurement points, and the force and magnetic coupling super-edge connects the vibration measurement points of the same bearing seat and the magnetic domain unit; The graph convolution engine aggregates the multimodal features of temperature, vibration, and magnetic domains within the hyperedge, calculates and outputs the fused feature tensor, including the coupling coefficient between the rotor eccentricity field induced by thermal expansion and the eddy current intensity of the worn magnetic domain.

5. The steam turbine fault diagnosis system based on multi-parameter fusion monitoring according to claim 4 is characterized in that: In the neural symbolic fusion module: A tensor construction unit receives the fused feature tensor and constructs a fourth-order feature tensor according to time, quantum state, physical quantity and spatial dimension; Neural-Symbolic Interaction Unit, which injects symbolic rules: dynamically associates tensor components with fault predicate confidence, and reversely prunes modes that are not related to real-time faults; A tensor compression engine decomposes the fourth-order eigentensor to generate a core tensor that retains thermal-magnetic-vibration correlation and outputs a diagnostic eigenvector.

6. The steam turbine fault diagnosis system based on multi-parameter fusion monitoring according to claim 5 is characterized in that: The digital twin diagnosis module includes: Input the diagnostic feature vector and call the multi-scale twin cluster: Molecular dynamics twinning analyzes magnetic domain fluctuation signals to simulate dislocation evolution, discrete element twinning analyzes oil wear debris transport trajectory, and finite element twinning calculates friction dynamics based on thermal expansion eccentric fields; Dynamic verification unit: Compares the simulated prediction of the twin output with the measured vibration envelope energy. When the deviation exceeds the threshold, it adaptively reconstructs the fault tree nodes and generates fault location and degradation level.

7. The steam turbine fault diagnosis system based on multi-parameter fusion monitoring according to claim 6 is characterized in that: The federated decision module performs: receiving the degradation level, calling the neural symbolic engine at the edge layer to load the core tensor output by the neural symbolic fusion module, performing real-time reasoning and generating a primary diagnosis report; The cloud-based federated aggregation module aggregates the encrypted fault case gradients of multiple power plants and uses homomorphic encryption weighted averaging to update the diagnostic model parameters; The dynamic strategy generation unit integrates the primary diagnostic report and the updated model parameters, calculates the remaining life distribution according to the degradation level, generates a maintenance strategy and triggers steam parameter adjustment instructions to the turbine control system.

8. The steam turbine fault diagnosis system based on multi-parameter fusion monitoring according to claim 7 is characterized in that: The symbolic rule injection unit in the neural symbolic fusion module performs: The static rule loading unit loads the judgment criteria from the fault knowledge base: when the axial vibration 2× increase is greater than 30%, it is considered rotor misalignment, and when the temperature is greater than 120°C, it is considered bearing overheating; The dynamic rule generation unit receives the crack cloud map output by the digital twin diagnosis module and generates a three-level warning rule when the eddy current dispersion degree is greater than 0.7; The rule activation engine injects the judgment criteria and warning rules into the neural symbolic interaction unit, and constrains the association of tensor components with the confidence of the fault predicate.

9. The steam turbine fault diagnosis system based on multi-parameter fusion monitoring according to claim 8, characterized in that: The thermally induced vibration contribution coefficient is input into the vibration measurement point vertex attribute channel of the hypergraph feature extraction module to quantify the contribution intensity of thermal deformation to vibration; The inversion value of the rotor eccentricity output by the quantum noise correction mechanism is input into the hyperedge construction unit to dynamically calculate the thermal conductivity weight of the heat conduction hyperedge; Among them, the collaborative correlation between the vibration measurement point attributes and the heat conduction hyperedge weights eliminates the feature distortion caused by the scale mismatch of thermal-vibration data.

10. The steam turbine fault diagnosis system based on multi-parameter fusion monitoring according to claim 9, characterized in that: The dynamic strategy generation unit of the federated decision module receives the microcrack initiation probability cloud map output by the micro-twin, generates a load reduction instruction based on the degradation level, and sends the instruction to the turbine control system; The cloud-based federation aggregation module of the federated decision-making module generates hyperedge optimization parameters based on the gradient of encrypted fault cases of multiple power plants, outputs them to the hyperedge construction unit of the hypergraph feature extraction module, and dynamically updates the weight coefficients of the force and magnetic coupling hyperedges; Among them, the load reduction instruction reduces the thermal stress mutation, the hyperedge weight optimization improves the accuracy of thermal and vibration correlation characterization, and synergistically suppresses thermal and mechanical coupling distortion.

Images