Thermal power station fan fault diagnosis method and system and storage medium

Abstract

The invention provides a thermal power station draught fan fault diagnosis method and system and a storage medium. The method comprises the steps that vibration, sound, temperature and pressure data of draught fans are collected and subjected to time-space alignment; constructing a three-dimensional physical model of the fan; carrying out sparse representation on the vibration and sound signals by adopting an improved K-SVD (K-Singular Value Decomposition) algorithm to generate codes for representing bearing fatigue and blade cracks; analyzing temperature and pressure data through heat conduction and CFD flow field simulation, and identifying an overheating area and a pressure abnormal path; establishing a multi-physical field coupling relationship through a tensor fusion algorithm, and constructing a deep belief network diagnosis model; and outputting a fault type and a three-dimensional space positioning result by using the diagnosis model. Through the steps of multi-physics field data acquisition, three-dimensional model construction, improved algorithm sparse representation, heat conduction and flow field simulation analysis, deep belief network model construction, real-time diagnosis and the like, the accuracy and efficiency of fault diagnosis are remarkably improved.

Description

Technical Field

[0001] The present invention relates to the technical field of fan fault diagnosis, and in particular to a method, system and storage medium for diagnosing fan faults in a thermal power station. Background Art

[0002] In thermal power generation, fans serve as core auxiliary equipment for boiler systems and steam turbine units, and their operational stability directly impacts power generation efficiency and unit safety. Traditional fan fault diagnosis technology primarily relies on single-field physical field monitoring and empirically driven analysis methods, often using single sensor data such as vibration monitoring, infrared temperature measurement, or pressure analysis. For example, bearing faults are identified based on vibration time-domain indicators (such as peak value and RMS value) or frequency-domain characteristics (such as spectral peak frequency).

[0003] Existing methods often overlook the multi-physics coupling characteristics of wind turbine faults, limiting the accuracy and reliability of fault diagnosis. Especially under complex operating conditions, monitoring data from a single physical field cannot fully reflect the actual operating status of the wind turbine, leading to frequent misdiagnosis or missed diagnosis. Summary of the Invention

[0004] The present invention aims to at least solve the technical problem in the prior art that fault misdiagnosis or missed diagnosis is prone to occur, and in particular innovatively proposes a method, system and storage medium for diagnosing fan faults in thermal power plants.

[0005] In order to achieve the above-mentioned object of the present invention, the present invention provides a method for diagnosing fan faults in a thermal power station, the method comprising:

[0006] S1. Collect multi-physics field data of the wind turbine, including vibration data, sound data, temperature data, and pressure data, through timestamp synchronization and sensor spatial coordinate calibration, and perform spatiotemporal alignment on the multi-physics field data;

[0007] S2. Construct a three-dimensional physical model of the wind turbine including blade geometric parameters and flow field distribution, map the multi-physics field data into the three-dimensional physical model of the wind turbine, and generate a spatial discretization feature matrix including vibration modal components, sound pressure level distribution, temperature gradient field, and pressure cloud map;

[0008] S3. Perform sparse representation on the vibration time-frequency spectrum and the acoustic resonance demodulation signal to generate a time-frequency ridge code representing bearing fatigue and a blade crack azimuth code;

[0009] S4. Through solving the inverse heat conduction problem and CFD flow field simulation, the temperature data and pressure data are integrated and analyzed to identify the overheating area and the pressure anomaly propagation path;

[0010] S5. Using the time-frequency ridge code and the blade crack azimuth code as constraints of the spatial discretization feature matrix, establishing a multi-physics field coupling relationship based on the overheating area and the pressure anomaly propagation path through a tensor fusion algorithm, and constructing a deep belief network diagnostic model including a fault mode recognition layer and a spatial positioning layer based on the multi-physics field coupling relationship;

[0011] S6. Based on the time-frequency ridge coding and blade crack azimuth coding that characterize bearing fatigue, use the deep belief network diagnostic model to perform real-time diagnosis and output the fault type and three-dimensional spatial positioning results.

[0012] As an optional embodiment of the present invention, optionally, constructing the three-dimensional physical model of the wind turbine in step S2 includes:

[0013] S201, performing a three-dimensional scan on the fan impeller, bearing chamber, and casing to obtain point cloud data and establish an initial fan model, and then modifying the initial fan model;

[0014] S202, calibrate the spatial coordinate system of the wind turbine and establish an assembly parameter database including blade installation angle, shaft coaxiality, and casing deformation;

[0015] S203, performing structural domain meshing and fluid domain meshing on the initial wind turbine model based on the assembly parameter database;

[0016] S204, establishing a transformation matrix between the sensor coordinate system and the initial wind turbine model coordinate system, and integrating blade geometric parameters, flow field distribution characteristics, and assembly parameters into the initial wind turbine model to obtain the wind turbine three-dimensional physical model;

[0017] S205. Use computational fluid dynamics methods to simulate the internal flow field of the fan to obtain flow velocity distribution information, pressure distribution information, and temperature distribution information of the internal flow field of the fan, and verify and optimize the three-dimensional physical model of the fan based on the flow velocity distribution information, pressure distribution information, and temperature distribution information.

[0018] As an optional embodiment of the present invention, optionally, in step S3, an improved K-SVD algorithm is used to perform sparse representation on the vibration time spectrum and the sound resonance demodulation signal. The expression of the improved K-SVD algorithm is:

[0019]

[0020] Where D represents the bimodal dictionary, X represents the sparse coding matrix, N represents the number of samples, || || F represents the Frobenius norm, V :,i represents the vibration time-frequency matrix, D vrepresents the vibration mode dictionary, X v,i represents the sparse coding of the i-th vibration sample, λ represents the acoustic modal weight coefficient, D a represents the acoustic modal dictionary, X a,i represents the sparse coding of the i-th acoustic sample, μ represents the structured sparse penalty factor, || || 1,1 represents the L1,1 mixing norm, Ω represents the fault-sensitive weight matrix, and X represents the joint sparse coding matrix.

[0021] As an optional embodiment of the present invention, optionally, in step S4, the temperature data and pressure data are fused and analyzed by solving an inverse heat conduction problem and simulating a CFD flow field to identify overheating areas and pressure anomaly propagation paths, including:

[0022] S401. Establishing transient heat conduction control equations containing dynamic thermophysical parameters;

[0023] S402. Based on the transient heat conduction control equation, a conjugate gradient method combined with regularization is used to invert the dynamic convection heat transfer coefficient distribution through the measured temperature data to form a dynamic correction mechanism for thermal boundary conditions;

[0024] S403: Based on the dynamic correction mechanism of the thermal boundary conditions, a moving Gaussian heat source model is introduced, and a state transfer equation is constructed in combination with the bearing vibration data. The spatial tracking of the friction heat generation position is achieved through a particle filter algorithm to obtain the heat source position;

[0025] S404: inputting the heat source position into a pre-built turbulence model and transition model, implementing a sliding grid technology to simulate the impeller rotating flow field, and obtaining a flow field simulation result including the distribution of the overheating area;

[0026] S405, using the flow field simulation results, extracting the main modes of pressure pulsation through intrinsic orthogonal decomposition, and combining it with finite-time Lyapunov exponent calculation to obtain a geometric topological representation of the pressure anomaly propagation path;

[0027] S406, performing spatiotemporal alignment on the flow field simulation results of the overheating area distribution and the geometric topological representation of the pressure anomaly propagation path to obtain a fused temperature-pressure characteristic map;

[0028] S407: Based on the temperature-pressure characteristic diagram, identify the overheating area and the pressure anomaly propagation path.

[0029] As an optional embodiment of the present invention, optionally, the transient heat conduction control equation in step S401 is expressed as:

[0030]

[0031] Where ρ represents the material density, c p represents the constant pressure specific heat capacity, T represents the transient temperature field, t represents the time variable, represents the Hamiltonian operator, k(T) represents thermal conductivity, h represents the convective heat transfer coefficient, T ∞ represents the ambient fluid temperature, ∈ represents the surface emissivity, σ represents the Stefan-Boltzmann constant, T sur Indicates the surrounding surface radiation temperature, Q fric represents the frictional heating term.

[0032] As an optional embodiment of the present invention, optionally, constructing a deep belief network diagnostic model including a fault mode recognition layer and a spatial positioning layer in step S5 includes:

[0033] S501, using the time-frequency ridge code and the blade crack azimuth code as constraint dimensions, embedding the constraint dimensions into the spatial discretization feature matrix through tensor outer product, obtaining a three-dimensional feature tensor containing vibration-acoustic-thermal-fluid four-field coupling information, and generating an initial fusion carrier of multi-physics field data;

[0034] S502: constructing a nonlinear tensor fusion kernel function including a residual term of a heat conduction equation and a flow field continuity equation through the initial fusion carrier and based on the coordinate set of the overheating area and the pressure anomaly propagation path;

[0035] S503. Using HOSVD to perform multimodal decomposition on the nonlinear tensor fusion kernel function, extract the core tensor that represents the coupling relationship between multiple physical fields, and establish a correlation probability model between vibration modal frequency, acoustic emission amplitude, temperature gradient, and pressure pulsation based on the core tensor through nuclear norm constraints to form a causal reasoning basis for fault propagation.

[0036] S504. Based on the causal reasoning foundation, a DBN structure is constructed in which the fault pattern recognition layer and the spatial positioning layer share underlying features, wherein the pattern recognition channel uses a time domain convolution module to capture the time-frequency ridge evolution characteristics, and the positioning channel introduces a spatial self-attention mechanism to analyze the spatial correlation between the crack azimuth and the overheating area;

[0037] S505. Based on the historical fault case library, a generative adversarial network is constructed to perform sample enhancement, generating an enhanced data set containing compound fault modes, and based on the DBN structure, a multi-task loss function is used to achieve collaborative optimization of pattern classification accuracy and positioning accuracy to obtain the deep belief network diagnostic model.

[0038] As an optional embodiment of the present invention, optionally, the expression of the nonlinear tensor fusion kernel function in step S502 is:

[0039]

[0040] Where K(Y) represents the nonlinear tensor fusion kernel function of Y, Y represents the three-dimensional feature tensor containing the vibration-acoustic-thermal-fluid four-field coupling information, W h represents the heat conduction kernel weight matrix, W p represents the fluid mechanics kernel weight matrix, W f represents the fracture mechanics kernel weight matrix, λ represents the physical constraint enhancement coefficient, Δ·v represents the flow field divergence, ⊙ represents the Hadamard product, ReLU( ) represents the linear rectification function, T represents the temperature field, t represents time, α represents the thermal diffusivity, represents the temperature field Laplace operator.

[0041] As an optional embodiment of the present invention, optionally, in step S6, the expression for the fault type output by the deep belief network diagnostic model is:

[0042]

[0043] Among them, P(f m |D) indicates that under the condition of given input data D, the mth type of failure mode f m The posterior probability of occurrence, exp( ) represents the natural exponential function, represents the transpose of the weight vector of the mth class of the fault mode recognition layer, z represents the shared hidden layer feature vector, b fm represents the bias scalar of the mth class of the fault pattern recognition layer, M represents the total number of fault types, represents the transpose of the weight vector of the nth class of the fault mode recognition layer, b fn The bias scalar representing the nth class of the fault mode recognition layer;

[0044] The expression of the three-dimensional spatial positioning result output by the diagnostic model is:

[0045]

[0046] Where x, y, and z represent the three-dimensional coordinates of the fault source, and W l Represents the spatial positioning layer regression weight matrix, ReLU( ) represents the linear rectification function, W h represents the spatial positioning channel hidden layer weight matrix, z represents the shared hidden layer feature vector, b h represents the bias vector of the hidden layer of the spatial positioning channel, b l Represents the regression bias vector of the spatial positioning layer.

[0047] In another aspect, the present invention further provides a thermal power plant fan fault diagnosis system, comprising a data acquisition module, a diagnosis processing module and a result display module;

[0048] The data acquisition module is used to collect fan vibration signals, acoustic signals, temperature signals and pressure signals;

[0049] The diagnostic processing module is used to execute the thermal power station fan fault diagnosis method;

[0050] The result display module is used to display the fault diagnosis results, including the fault type and fault location.

[0051] In another aspect, the present invention provides a computer-readable storage medium comprising:

[0052] a memory having a computer program stored thereon;

[0053] A processor is used to execute the program in the memory to implement the thermal power station fan fault diagnosis method.

[0054] Beneficial effects of the present invention: The present invention improves the accuracy of fault diagnosis through the steps of multi-physics field data acquisition and spatiotemporal alignment, three-dimensional physical model construction and feature matrix generation, improved algorithm for sparse representation and encoding, heat conduction and flow field simulation analysis, deep confidence network diagnostic model construction, and real-time diagnosis and output. Specifically, the present invention first ensures the integrity and accuracy of the collected multi-physics field data such as vibration, sound, temperature, and pressure through timestamp synchronization and sensor spatial coordinate calibration; then constructs a three-dimensional physical model of the wind turbine including blade geometric parameters and flow field distribution, maps the multi-physics field data into the model, generates a spatial discretization feature matrix, and realizes three-dimensional visualization and feature extraction of the data; then uses an improved K-SVD algorithm to sparsely represent the vibration and sound signals, generates time-frequency ridge line coding and crack azimuth coding, and improves the accuracy and efficiency of feature extraction; then, through solving the inverse problem of heat conduction and computational fluid dynamics (CFD) flow field simulation, identifies overheating areas and pressure anomaly propagation paths, providing a physical basis for fault diagnosis; finally, constructs a deep belief network diagnostic model based on the multi-physics field coupling relationship to achieve accurate identification and spatial positioning of fault modes, and uses the deep belief network diagnostic model for real-time diagnosis, outputting the fault type and three-dimensional spatial positioning results. The diagnostic method of the present invention can significantly improve the accuracy and efficiency of fault diagnosis of thermal power plant wind turbines.

[0055] Additional aspects and advantages of the present invention will be set forth in part in the description which follows and, in part, will be obvious from the description which follows, or may be learned by practice of the present invention. BRIEF DESCRIPTION OF THE DRAWINGS

[0056] The above and / or additional aspects and advantages of the present invention will become apparent and readily understood from the following description of the embodiments with reference to the accompanying drawings, in which:

[0057] Figure 1 This is a flow chart of a method for diagnosing fan faults in a thermal power station provided by the present invention;

[0058] Figure 2 This is a structural block diagram of a thermal power station fan fault diagnosis system provided by the present invention. DETAILED DESCRIPTION

[0059] The following describes embodiments of the present invention in detail. Examples of the embodiments are shown in the accompanying drawings, wherein the same or similar reference numerals throughout represent the same or similar elements or elements having the same or similar functions. The embodiments described below with reference to the accompanying drawings are exemplary and are intended only to explain the present invention and are not to be construed as limiting the present invention.

[0060] Example 1

[0061] like Figure 1 As shown, a method for diagnosing fan faults in a thermal power station comprises:

[0062] S1. Collect multi-physics field data of the wind turbine, including vibration data, sound data, temperature data, and pressure data, through timestamp synchronization and sensor spatial coordinate calibration, and perform spatiotemporal alignment on the multi-physics field data;

[0063] It's important to note that in step S1, a high-precision clock synchronization system ensures consistent timestamps for data collected by each sensor. Three-dimensional spatial positioning technology is also used to calibrate the sensors' spatial coordinates to ensure data accuracy and consistency. The spatiotemporal alignment process integrates this data, ensuring that different physical field data at the same point in time correspond to each other.

[0064] S2. Construct a three-dimensional physical model of the wind turbine including blade geometric parameters and flow field distribution, map the multi-physics field data into the three-dimensional physical model of the wind turbine, and generate a spatial discretization feature matrix including vibration modal components, sound pressure level distribution, temperature gradient field, and pressure cloud map;

[0065] S3. Use the improved K-SVD algorithm to perform sparse representation on the vibration time-frequency spectrum and acoustic resonance demodulation signal to generate the time-frequency ridge code and blade crack azimuth code representing bearing fatigue;

[0066] It should be noted in step S3 that the improved K-SVD algorithm effectively improves the feature extraction capability of vibration and sound signals by introducing the orthogonal matching pursuit strategy and optimizing the dictionary learning and sparse coding process, so that the time-frequency ridge coding and crack azimuth coding can more accurately reflect the fatigue state of the wind turbine bearing and the location information of the blade crack.

[0067] The vibration time-frequency spectrum is obtained by performing time-frequency analysis on the vibration signal. Time-frequency analysis expands the vibration signal across both time and frequency, revealing the frequency components and their variations over time. This analysis method is particularly effective in identifying fault characteristics in rotating components such as fan bearings, as faults often cause the vibration signal to abnormally increase or decrease at specific frequencies.

[0068] The acoustic resonance demodulation signal is obtained by performing resonance demodulation on the acoustic signal. Resonance demodulation is a signal processing technique that uses the resonance phenomenon in acoustic signals to extract characteristic information related to faults. When cracks or damage occur on wind turbine blades or other components, they resonate at specific frequencies. These resonance signals are captured by the acoustic sensor and extracted through resonance demodulation, thereby being used for fault diagnosis.

[0069] Based on the vibration time-frequency spectrum, the characteristic frequency components caused by bearing fatigue and their temporal variation trends are identified, and then a time-frequency ridge code is generated. The time-frequency ridge code can clearly outline the characteristic trajectory of bearing fatigue failure in the time-frequency domain. The generation of the blade crack azimuth code is based on the spatial distribution characteristics of the acoustic resonance demodulation signal. By analyzing the changes in the intensity, frequency, and phase of the resonance signal with different blade positions, combined with the geometric parameters of the wind turbine blade, the approximate azimuth of the crack is calculated and encoded. This process fully utilizes the sensitivity of acoustic signals to structural damage to achieve accurate identification of the blade crack azimuth.

[0070] S4. Through solving the inverse heat conduction problem and CFD flow field simulation, the temperature data and pressure data are integrated and analyzed to identify the overheating area and the pressure anomaly propagation path;

[0071] S5. Using the time-frequency ridge code and the blade crack azimuth code as constraints of the spatial discretization feature matrix, establishing a multi-physics field coupling relationship based on the overheating area and the pressure anomaly propagation path through a tensor fusion algorithm, and constructing a deep belief network diagnostic model including a fault mode recognition layer and a spatial positioning layer based on the multi-physics field coupling relationship;

[0072] S6. Based on the time-frequency ridge coding and blade crack azimuth coding that characterize bearing fatigue, use the deep belief network diagnostic model to perform real-time diagnosis and output the fault type and three-dimensional spatial positioning results.

[0073] It should be noted that in step S6: the deep confidence network diagnostic model can accurately identify the type of wind turbine fault through its internal fault mode recognition layer. This layer uses a deep learning algorithm to process the input time-frequency ridge code and blade crack azimuth code, and quickly determines the type of fault by comparing it with the preset fault feature library. At the same time, the spatial positioning layer is responsible for outputting the specific location of the fault in the three-dimensional space of the wind turbine. This layer analyzes the vibration modal components, sound pressure level distribution, temperature gradient field and pressure cloud map in the spatial discretization feature matrix, and combines the multi-physical field coupling relationship obtained by the tensor fusion algorithm to accurately calculate the three-dimensional coordinates of the fault source.

[0074] The principle behind this thermal power plant wind turbine fault diagnosis method is as follows: First, timestamp synchronization and sensor spatial coordinate calibration ensure the accuracy and consistency of multi-physics field data, including key information such as vibration, sound, temperature, and pressure. Subsequently, a constructed three-dimensional physical model of the wind turbine is used to map this data to the physical model, generating a spatially discretized feature matrix. This step enables three-dimensional data visualization and feature extraction.

[0075] In the feature extraction stage, the improved K-SVD algorithm is used to perform sparse representation on the vibration and sound signals, generating time-frequency ridge coding and crack azimuth coding. These coding improves the accuracy and efficiency of feature extraction.

[0076] Next, the temperature and pressure data were fused and analyzed by solving an inverse heat conduction problem and performing CFD flow field simulations. This process identified overheating areas and the propagation paths of pressure anomalies.

[0077] When constructing a deep belief network diagnostic model, the time-frequency ridge encoding and blade crack azimuth encoding were used as constraints in the spatially discretized feature matrix. Based on the overheating region and the pressure anomaly propagation path, a multi-physics field coupling relationship was established using a tensor fusion algorithm. This model incorporates not only a fault mode recognition layer but also a spatial localization layer, enabling accurate fault mode identification and spatial localization.

[0078] Finally, the deep belief network diagnostic model is used for real-time diagnosis, quickly outputting the fault type and 3D spatial location results. This step improves the accuracy and efficiency of fault diagnosis.

[0079] As an optional embodiment of the present invention, optionally, constructing the three-dimensional physical model of the wind turbine in step S2 includes:

[0080] S201, performing a three-dimensional scan on the fan impeller, bearing chamber, and casing to obtain point cloud data and establish an initial fan model, and then modifying the initial fan model;

[0081] It should be noted in step S201 that the step of modifying the initial fan model includes adjusting the initial fan model's geometric dimensions, blade shape, and flow path structure based on the fan design drawings and measured data to ensure the model's accuracy and reliability. Furthermore, the fan model's boundary conditions, including inlet velocity, outlet pressure, and wall conditions, must be set to simulate actual fan operating conditions. This modified and configured three-dimensional physical model of the fan more accurately reflects the fan's actual operating conditions.

[0082] S202, calibrate the spatial coordinate system of the wind turbine and establish an assembly parameter database including blade installation angle, shaft coaxiality, and casing deformation;

[0083] It should be noted in step S202 that the step of calibrating the spatial coordinate system of the entire fan is intended to ensure that the spatial positional relationships of the various parts of the fan are accurate. Using high-precision measurement equipment and methods, the spatial coordinates of key components of the fan, such as the impeller, bearing chamber, and casing, are calibrated to establish an accurate spatial coordinate system. In this coordinate system, the position, orientation, and mutual relationship of the various parts of the fan can be accurately described. At the same time, an assembly parameter database containing blade mounting angle, shaft coaxiality, and casing deformation is established to record key parameters of the fan during the assembly process. These parameters are of great significance to the operating performance and fault diagnosis of the fan. The blade mounting angle determines the force applied to the blades and the direction of the airflow, the shaft coaxiality is related to the smoothness and vibration of the fan operation, and the casing deformation may affect the sealing performance and flow field distribution of the fan. By monitoring the changes in these parameters in real time, potential faults of the fan can be discovered in a timely manner.

[0084] S203, performing structural domain meshing and fluid domain meshing on the initial wind turbine model based on the assembly parameter database;

[0085] The purpose of meshing in step S203 is to discretize the fan model into mesh units required for numerical calculations in order to perform subsequent fluid dynamics analysis and heat conduction analysis. Structural domain meshing is mainly aimed at the solid parts of the fan, such as the impeller, bearing chamber and casing, and is divided using structured mesh units such as hexahedrons or tetrahedrons to ensure the accuracy and convergence of the calculation. Fluid domain meshing is aimed at the fluid parts inside the fan, such as airflow and lubricating oil, and is divided using unstructured mesh units to adapt to the complex flow patterns and boundary conditions of the fluid. Through meshing, the fan model can be converted into a numerical calculation model.

[0086] S204, establishing a transformation matrix between the sensor coordinate system and the initial wind turbine model coordinate system, and integrating blade geometric parameters, flow field distribution characteristics, and assembly parameters into the initial wind turbine model to obtain the wind turbine three-dimensional physical model;

[0087] It should be noted that in step S204, the transformation matrix between the sensor coordinate system and the initial fan model coordinate system is established in order to achieve accurate correspondence between the data collected by the sensor and the fan model. Through high-precision coordinate transformation, the vibration, sound, temperature, pressure and other data collected by the sensor can be accurately mapped to the corresponding positions in the fan model. This step is the key to achieving data spatiotemporal alignment and subsequent fault diagnosis. At the same time, integrating the blade geometric parameters, flow field distribution characteristics and assembly parameters into the initial fan model can make the three-dimensional physical model of the fan more realistically reflect the actual operating status of the fan. The integration of these parameters not only improves the accuracy of the model, but also provides richer information for subsequent fault diagnosis.

[0088] S205. Use computational fluid dynamics methods to simulate the internal flow field of the fan to obtain flow velocity distribution information, pressure distribution information, and temperature distribution information of the internal flow field of the fan, and verify and optimize the three-dimensional physical model of the fan based on the flow velocity distribution information, pressure distribution information, and temperature distribution information.

[0089] It should be noted that in step S205, simulation of the internal flow field of the fan using computational fluid dynamics (CFD) methods can provide a deeper understanding of the fluid flow conditions within the fan. Through simulation, key information such as the velocity distribution, pressure distribution, and temperature distribution of the internal flow field of the fan can be obtained.

[0090] First, velocity distribution information reveals the flow speed and direction of the fluid inside the wind turbine, helping to identify potential fluid dynamics issues such as vortices, backflow, or uneven flow rates, which could affect the efficiency and stability of the wind turbine.

[0091] Secondly, pressure distribution information is crucial for evaluating the fan's pressure loss and fluid dynamics. The simulated pressure distribution map can identify high- and low-pressure areas, as well as changes in pressure gradients. This information helps optimize fan design, reduce pressure loss, and improve efficiency.

[0092] Finally, temperature distribution information is crucial for evaluating the thermal performance of wind turbines. The simulated temperature distribution maps can identify areas of overheating and significant temperature gradients. This information helps prevent thermal failures such as bearing overheating, blade deformation, and seal failure.

[0093] After obtaining these simulation results, the wind turbine 3D physical model can be verified and optimized. By comparing the simulation results with measured data, the model's accuracy can be assessed and appropriate adjustments can be made. This includes adjusting blade geometry, flow path structure, boundary conditions, and other factors to more accurately reflect the wind turbine's actual operating conditions. A verified and optimized 3D wind turbine physical model.

[0094] As an optional embodiment of the present invention, optionally, the expression of the improved K-SVD algorithm in step S3 is:

[0095]

[0096] Where D represents the bimodal dictionary, X represents the sparse coding matrix, N represents the number of samples, || || F represents the Frobenius norm, V :,i represents the vibration time-frequency spectrum matrix of the i-th vibration sample, D v represents the vibration mode dictionary, X v,i represents the sparse coding of the i-th vibration sample, λ represents the acoustic modal weight coefficient, D a represents the acoustic modal dictionary, X a,i represents the sparse coding of the i-th acoustic sample, μ represents the structured sparse penalty factor, || || 1,1 represents the L1,1 mixed norm, Ω represents the fault-sensitive weight matrix, X represents the joint sparse coding matrix, T v represents the sparsity of vibration signal, T a Indicates the sparsity of the acoustic signal.

[0097] As an optional embodiment of the present invention, optionally, in step S4, the temperature data and pressure data are fused and analyzed by solving an inverse heat conduction problem and simulating a CFD flow field to identify overheating areas and pressure anomaly propagation paths, including:

[0098] S401. Establishing a transient heat conduction control equation containing dynamic thermophysical parameters;

[0099] S402. Based on the transient heat conduction control equation, a conjugate gradient method combined with regularization is used to invert the dynamic convection heat transfer coefficient distribution through the measured temperature data to form a dynamic correction mechanism for thermal boundary conditions;

[0100] It should be noted in step S402 that the conjugate gradient method combined with regularization effectively addresses the ill-posedness of the inverse heat conduction problem, improving the accuracy and stability of the inversion results. Furthermore, the resulting dynamic correction mechanism for thermal boundary conditions allows for real-time adjustment of the thermal boundary conditions based on measured temperature data, making the simulation results more realistic.

[0101] S403: Based on the dynamic correction mechanism of the thermal boundary conditions, a moving Gaussian heat source model is introduced, and a state transfer equation is constructed in combination with the bearing vibration data of the fan. The spatial tracking of the friction heat generation position is achieved through a particle filter algorithm to obtain the heat source position;

[0102] It should be noted in step S403 that the moving Gaussian heat source model is introduced to simulate the heat distribution generated by friction within the fan. This model can describe the temporal and spatial variations of the heat source, making the simulation results more realistic. Furthermore, by combining bearing vibration data with the state transition equation, the location of frictional heat generation can be spatially tracked. The particle filter algorithm can efficiently process this complex data and accurately locate the heat source.

[0103] The specific process for obtaining the heat source location includes: first, using a particle filter algorithm to iteratively solve the state transition equation and continuously update the probability distribution of the heat source location. Second, based on the updated probability distribution, the most likely value of the heat source location is determined. Finally, by comparing the heat source location with the positions of key components in the three-dimensional physical model of the wind turbine, the possible causes and impacts of the fault can be further analyzed. By comprehensively considering multiple sources of information such as temperature data, pressure data, vibration data, and acoustic data, the method of the present invention can achieve accurate diagnosis and spatial location of wind turbine faults in thermal power plants.

[0104] S404: inputting the heat source position into a pre-built turbulence model and transition model, implementing a sliding grid technology to simulate the impeller rotating flow field, and obtaining a flow field simulation result including the distribution of the overheating area;

[0105] It is important to note that in step S404, the flow field simulation results for the overheated area distribution are compared with the pressure data to identify the propagation path of the pressure anomaly. This process combines heat conduction analysis and fluid dynamics simulation to more comprehensively reveal the thermal and fluid conditions within the wind turbine. By comparing the simulation results with the measured data, the accuracy of the model can be further verified and necessary adjustments and optimizations can be made to ensure that it more accurately reflects the actual operating conditions of the wind turbine.

[0106] S405, using the flow field simulation results, extracting the main modes of pressure pulsation through intrinsic orthogonal decomposition, and combining it with finite-time Lyapunov exponent calculation to obtain a geometric topological representation of the pressure anomaly propagation path;

[0107] It should be noted in step S405 that intrinsic orthogonal decomposition is an effective data dimensionality reduction technique that can extract key features from complex flow field data. In this embodiment, intrinsic orthogonal decomposition is used to extract the main modes of pressure pulsation. These main modes represent the main trends and characteristics of pressure changes in the flow field. Subsequently, combined with the calculation of the finite-time Lyapunov exponent, the stability and evolution laws of these main modes can be further analyzed to obtain a geometric topological representation of the pressure anomaly propagation path. This representation can intuitively display the propagation path and evolution process of the pressure anomaly inside the fan.

[0108] S406, performing spatiotemporal alignment on the flow field simulation results of the overheating area distribution and the geometric topological representation of the pressure anomaly propagation path to obtain a fused temperature-pressure characteristic map;

[0109] It should be noted that in step S406, the fused temperature-pressure characteristic map is used as one of the inputs to the deep belief network diagnostic model to enhance the model's ability to capture fault characteristics. This step fully utilizes the correlation between temperature and pressure data, improving the comprehensiveness and accuracy of fault diagnosis. Through the complex learning and reasoning mechanisms of the deep belief network, it is possible to deeply explore the implicit information and fault characteristics in the data, thereby achieving more accurate fault diagnosis.

[0110] S407: Based on the temperature-pressure characteristic diagram, identify the overheating area and the pressure anomaly propagation path.

[0111] It should be noted in step S407 that the specific process of identifying the overheating area and the pressure anomaly propagation path includes the following key steps: First, the fused temperature-pressure characteristic map is carefully analyzed to identify areas in the characteristic map with abnormally high temperature or abnormal pressure changes. These areas are often the key locations where internal faults of the fan occur. Secondly, advanced image processing technology and algorithms are used to perform boundary extraction and morphological processing on these abnormal areas to more clearly outline the contours of the overheating area and the pressure anomaly propagation path. Finally, the identified overheating area and pressure anomaly propagation path are verified and optimized by combining the three-dimensional physical model of the fan and the measured data. By comparing the simulation results with the measured data, the location and nature of the fault can be further confirmed. This process not only improves the accuracy and efficiency of fault diagnosis, but also provides a strong guarantee for the safe operation of the fan.

[0112] As an optional embodiment of the present invention, optionally, the expression of the transient heat conduction control equation in step S401 is:

[0113]

[0114] Where ρ refers to the density of the materials used in the internal components of the fan, c p represents the constant pressure specific heat capacity, T represents the transient temperature field, t represents the time variable, represents the Hamiltonian operator, k(T) represents thermal conductivity, h represents the convective heat transfer coefficient, T ∞ represents the ambient fluid temperature, ∈ represents the emissivity of the surface materials of the components inside the fan, σ represents the Stefan-Boltzmann constant, T sur Indicates the surrounding surface radiation temperature, Q fric represents the frictional heating term.

[0115] As an optional embodiment of the present invention, optionally, constructing a deep belief network diagnostic model including a fault mode recognition layer and a spatial positioning layer in step S5 includes:

[0116] S501, using the time-frequency ridge code and the blade crack azimuth code as constraint dimensions, embedding the constraint dimensions into the spatial discretization feature matrix through tensor outer product, obtaining a three-dimensional feature tensor containing vibration-acoustic-thermal-fluid four-field coupling information, and generating an initial fusion carrier of multi-physics field data;

[0117] It should be noted that in step S501, through the tensor outer product operation, the two constraint dimensions of time-frequency ridge coding and blade crack azimuth coding can be effectively combined with the spatial discretization feature matrix to form a three-dimensional feature tensor containing rich information. This step realizes the preliminary fusion of the four physical field data of vibration, acoustics, thermal and fluid dynamics. The generated three-dimensional feature tensor not only contains the data features of each physical field, but also realizes the association and constraint between the data through the introduction of the constraint dimension, thereby improving the accuracy and effectiveness of data fusion. This initial fusion carrier will serve as one of the inputs of the deep belief network diagnostic model for subsequent fault mode recognition and spatial positioning.

[0118] S502: constructing a nonlinear tensor fusion kernel function including a residual term of a heat conduction equation and a flow field continuity equation through the initial fusion carrier and based on the coordinate set of the overheating area and the pressure anomaly propagation path;

[0119] It's important to note that in step S502, the nonlinear tensor fusion kernel function fully accounts for the impact of overheating and the propagation path of pressure anomalies on the wind turbine's operating status. By combining the residual term of the heat conduction equation with the flow field continuity equation, a more comprehensive and accurate fault characterization can be constructed. This step enhances data correlation.

[0120] S503. Using HOSVD to perform multimodal decomposition on the nonlinear tensor fusion kernel function, extract the core tensor that represents the coupling relationship between multiple physical fields, and establish a correlation probability model between vibration modal frequency, acoustic emission amplitude, temperature gradient, and pressure pulsation based on the core tensor through nuclear norm constraints to form a causal reasoning basis for fault propagation.

[0121] It should be noted that in step S503, the multimodal decomposition of the nonlinear tensor fusion kernel function using high-order singular value decomposition (HOSVD) can effectively extract the core tensor that characterizes the coupling relationship of multiple physical fields. The core tensor contains the key coupling information between the four physical fields of vibration, acoustics, thermals and fluid dynamics. Subsequently, based on the core tensor, through the nuclear norm constraint, an association probability model can be established, which can describe the complex relationship between vibration modal frequency, acoustic emission amplitude, temperature gradient and pressure pulsation. By in-depth analysis of this model, the path and mechanism of fault propagation inside the wind turbine can be revealed, thereby achieving accurate prediction and location of the fault.

[0122] S504. Based on the causal reasoning foundation, a DBN structure is constructed in which the fault pattern recognition layer and the spatial positioning layer share underlying features, wherein the pattern recognition channel uses a time domain convolution module to capture the time-frequency ridge evolution characteristics, and the positioning channel introduces a spatial self-attention mechanism to analyze the spatial correlation between the crack azimuth and the overheating area;

[0123] It should be noted in step S504 that when constructing the deep belief network (DBN) structure in which the fault pattern recognition layer and the spatial positioning layer share underlying features, this embodiment fully considers the complexity and diversity of wind turbine fault characteristics. The pattern recognition channel uses a time-domain convolution module, which can capture the evolution characteristics of time-frequency ridges, thereby effectively identifying different fault modes. As one of the important features of the vibration signal, the evolution law of the time-frequency ridge can reflect the changes in the operating state of the wind turbine. Therefore, the introduction of the time-domain convolution module can significantly improve the accuracy of fault pattern recognition.

[0124] The localization channel incorporates a spatial self-attention mechanism, which analyzes the spatial correlation between crack azimuth and overheating areas. In wind turbine failures, cracks and overheating are often important fault indicators, and their location and distribution are crucial for fault location and repair. The spatial self-attention mechanism can more accurately capture the spatial distribution of these key features, thereby accurately locating the fault location.

[0125] This embodiment also uses a causal reasoning foundation as guidance in constructing the DBN structure. This foundation can reveal the path and mechanism of fault propagation within the wind turbine, providing more in-depth information on fault characteristics. By incorporating this information into the DBN structure, the comprehensiveness and accuracy of fault diagnosis can be further improved.

[0126] The DBN structure of this embodiment specifically includes: an input layer, a hidden layer, and an output layer. Among them, the input layer is used to receive the fused multi-physics field data, which are pre-processed to form a feature vector containing rich fault information. The hidden layer is composed of a plurality of restricted Boltzmann machines (RBMs) stacked together, and each RBM is responsible for extracting high-level features from the data. In the hidden layer, through layer-by-layer learning and feature extraction, the implicit information and fault characteristics in the data can be gradually mined. The output layer performs fault pattern recognition and spatial positioning based on the features extracted by the hidden layer. Specifically, the output layer includes two branches: one for fault pattern recognition, which outputs the probability distribution of the fault category; the other is used for spatial positioning, which outputs the probability distribution map of the fault location. By training and optimizing the parameters in the DBN structure, accurate diagnosis and spatial positioning of wind turbine faults can be achieved.

[0127] S505. Based on the historical fault case library, a generative adversarial network is constructed to perform sample enhancement, generating an enhanced data set containing compound fault modes, and based on the DBN structure, a multi-task loss function is used to achieve collaborative optimization of pattern classification accuracy and positioning accuracy to obtain the deep belief network diagnostic model.

[0128] It should be noted in step S505 that in order to improve the generalization ability and robustness of the deep belief network diagnostic model, this embodiment uses a generative adversarial network (GAN) for sample enhancement. The generative adversarial network consists of a generator and a discriminator. Through the game process between the two, virtual samples with a high degree of similarity to real samples can be generated. In this step, a generative adversarial network is first constructed based on a historical fault case library. The network can learn the distribution characteristics of historical fault data and generate an enhanced data set containing complex fault modes. This enhanced data set not only enriches the diversity of training samples, but also contains more complex fault characteristics, which helps to improve the performance of the diagnostic model.

[0129] Subsequently, the model was trained and optimized using the generated augmented dataset, combined with the previously constructed Deep Belief Network (DBN) architecture. During training, a multi-task loss function was employed that simultaneously considered the accuracy of fault mode classification and the precision of fault location. By collaboratively optimizing these two objectives, a Deep Belief Network diagnostic model was developed that could both accurately identify fault modes and precisely locate fault locations.

[0130] As an optional embodiment of the present invention, optionally, the expression of the nonlinear tensor fusion kernel function in step S502 is:

[0131]

[0132] Where K(Y) represents the nonlinear tensor fusion kernel function of Y, Y represents the three-dimensional feature tensor containing the vibration-acoustic-thermal-fluid four-field coupling information, W h represents the heat conduction kernel weight matrix, W p represents the fluid mechanics kernel weight matrix, W f represents the fracture mechanics kernel weight matrix, λ represents the physical constraint enhancement coefficient, Δ·v represents the flow field divergence, ⊙ represents the Hadamard product, ReLU( ) represents the linear rectification function, T represents the temperature field, t represents time, α represents the thermal diffusivity, represents the temperature field Laplace operator.

[0133] As an optional embodiment of the present invention, optionally, in step S6, the diagnostic model outputs the expression of the fault type and the three-dimensional spatial positioning result as follows:

[0134]

[0135] Among them, P(f m |D) indicates that under the condition of given input data D, the mth type of failure mode f m The posterior probability of occurrence, exp( ) represents the natural exponential function, represents the transpose of the weight vector of the mth class of the fault mode recognition layer, z represents the shared hidden layer feature vector, b fm represents the bias scalar of the mth class of the fault pattern recognition layer, M represents the total number of fault types, represents the transpose of the weight vector of the nth class of the fault mode recognition layer, b fn The bias scalar representing the nth class of the fault mode recognition layer;

[0136] The expression of the three-dimensional spatial positioning result output by the diagnostic model is:

[0137]

[0138] Where x, y, and z represent the three-dimensional coordinates of the fault source, and W l Represents the spatial positioning layer regression weight matrix, ReLU( ) represents the linear rectification function, W h represents the spatial positioning channel hidden layer weight matrix, z represents the shared hidden layer feature vector, b h represents the bias vector of the hidden layer of the spatial positioning channel, b l Represents the regression bias vector of the spatial positioning layer.

[0139] Example 2

[0140] like Figure 2 As shown, a thermal power plant fan fault diagnosis system is used to implement a thermal power plant fan fault diagnosis method described in Example 1;

[0141] The system includes a data acquisition module, a data preprocessing module, a real-time monitoring module and a result display module; the data acquisition module is used to collect fan vibration signals, acoustic signals, temperature signals and pressure signals; the data preprocessing module is used to preprocess the collected fan vibration signals, acoustic signals, temperature signals and pressure signals; the real-time monitoring module uses the diagnostic model to perform real-time diagnosis on the preprocessed signals, and quickly outputs the fault type and three-dimensional spatial positioning results; the result display module is used to intuitively display the fault diagnosis results, including the fault type and fault location.

[0142] In this embodiment, the principle of the thermal power plant fan fault diagnosis system is as follows: the system first obtains the fan's vibration signal, acoustic signal, temperature signal, and pressure signal through the data acquisition module. These data are the basis for fault diagnosis, and they reflect various status information during the operation of the fan. Subsequently, the data preprocessing module cleans, filters, and extracts features from these raw signals to eliminate noise and interference and extract feature information useful for fault diagnosis. The preprocessed signals are sent to the real-time monitoring module, which uses a trained deep belief network diagnostic model to analyze the signals. The diagnostic model can automatically identify different fault modes and accurately locate the three-dimensional spatial position of the fault inside the fan. This process is based on the fusion and analysis of multi-physics field data, fully considering the coupling relationship between vibration, acoustics, thermal and fluid dynamics.

[0143] Finally, the result display module presents the fault diagnosis results in an intuitive manner, including the fault type and location. This allows operation and maintenance personnel to quickly understand the operating status of the wind turbine and take appropriate repair measures in a timely manner to prevent the fault from further deteriorating and ensure the safe and stable operation of the thermal power plant.

[0144] Example 3

[0145] A computer-readable storage medium comprising:

[0146] a memory having a computer program stored thereon;

[0147] The processor is used to execute the program in the memory to implement a thermal power plant fan fault diagnosis method in Example 1.

[0148] It should be noted that the electronic device according to the embodiment of the present disclosure includes a processor and a memory for storing processor executable instructions, wherein the processor is configured to implement any of the above-mentioned methods for diagnosing wind turbine faults in a thermal power station when executing the executable instructions.

[0149] It should be noted that the number of processors can be one or more. Furthermore, the electronic device in the embodiments of the present disclosure may also include an input device and an output device. The processor, memory, input device, and output device may be connected via a bus or other means, which are not specifically limited here.

[0150] Memory, as a computer-readable storage medium, can be used to store software programs, computer-executable programs, and various modules, such as the program or module corresponding to the method for diagnosing wind turbine faults in a thermal power plant described in the embodiments of the present disclosure. The processor executes the software programs or modules stored in the memory to perform various functional applications and data processing in the electronic device.

[0151] The input device can be used to receive input numbers or signals. The signals can be key signals related to user settings and function control of the device / terminal / server. The output device can include a display device such as a display screen.

[0152] While embodiments of the present invention have been shown and described, it will be appreciated by those skilled in the art that various changes, modifications, substitutions, and alterations may be made to the embodiments without departing from the principles and spirit of the invention, and that the scope of the invention is defined by the claims and their equivalents.

Claims

1. A method for diagnosing fan faults in a thermal power station, characterized in that: The method comprises: S1. Collect multi-physics field data of the wind turbine, including vibration data, sound data, temperature data, and pressure data, through timestamp synchronization and sensor spatial coordinate calibration, and perform spatiotemporal alignment on the multi-physics field data; S2. Construct a three-dimensional physical model of the wind turbine including blade geometric parameters and flow field distribution, map the multi-physics field data into the three-dimensional physical model of the wind turbine, and generate a spatial discretization feature matrix including vibration modal components, sound pressure level distribution, temperature gradient field, and pressure cloud map; S3. Perform sparse representation on the vibration time-frequency spectrum and the acoustic resonance demodulation signal to generate a time-frequency ridge code representing bearing fatigue and a blade crack azimuth code; S4. Through solving the inverse heat conduction problem and CFD flow field simulation, the temperature data and pressure data are integrated and analyzed to identify the overheating area and the pressure anomaly propagation path; S5. Using the time-frequency ridge code and the blade crack azimuth code as constraints of the spatial discretization feature matrix, establishing a multi-physics field coupling relationship based on the overheating area and the pressure anomaly propagation path through a tensor fusion algorithm, and constructing a deep belief network diagnostic model including a fault mode recognition layer and a spatial positioning layer based on the multi-physics field coupling relationship; S6. Based on the time-frequency ridge coding and blade crack azimuth coding that characterize bearing fatigue, use the deep belief network diagnostic model to perform real-time diagnosis and output the fault type and three-dimensional spatial positioning results.

2. The method for diagnosing a fault of a fan in a thermal power plant according to claim 1, wherein: Constructing the three-dimensional physical model of the wind turbine in step S2 includes: S201, performing a three-dimensional scan on the fan impeller, bearing chamber, and casing to obtain point cloud data and establish an initial fan model, and then modifying the initial fan model; S202, calibrate the spatial coordinate system of the wind turbine and establish an assembly parameter database including blade installation angle, shaft coaxiality, and casing deformation; S203, performing structural domain meshing and fluid domain meshing on the initial wind turbine model based on the assembly parameter database; S204, establishing a transformation matrix between the sensor coordinate system and the initial wind turbine model coordinate system, and integrating blade geometric parameters, flow field distribution characteristics, and assembly parameters into the initial wind turbine model to obtain the wind turbine three-dimensional physical model; S205. Use computational fluid dynamics methods to simulate the internal flow field of the fan to obtain flow velocity distribution information, pressure distribution information, and temperature distribution information of the internal flow field of the fan, and verify and optimize the three-dimensional physical model of the fan based on the flow velocity distribution information, pressure distribution information, and temperature distribution information.

3. The method for diagnosing a fault of a fan in a thermal power plant according to claim 2, wherein: In step S3, the vibration time spectrum and the sound resonance demodulation signal are sparsely represented using the improved K-SVD algorithm. The expression of the improved K-SVD algorithm is: Where D represents the bimodal dictionary, X represents the sparse coding matrix, N represents the number of samples, |||| F represents the Frobenius norm, V :,i represents the vibration time-frequency matrix, D v represents the vibration mode dictionary, X v,i represents the sparse coding of the i-th vibration sample, λ represents the acoustic modal weight coefficient, D a represents the acoustic modal dictionary, X a,i represents the sparse coding of the i-th acoustic sample, μ represents the structured sparse penalty factor, |||| 1,1 represents the L1,1 mixing norm, Ω represents the fault-sensitive weight matrix, and X represents the joint sparse coding matrix.

4. The method for diagnosing a fault of a fan in a thermal power plant according to claim 1, wherein: In step S4, the temperature and pressure data are fused and analyzed by solving the inverse heat conduction problem and simulating the CFD flow field to identify overheating areas and pressure anomaly propagation paths, including: S401. Establishing a transient heat conduction control equation containing dynamic thermophysical parameters; S402. Based on the transient heat conduction control equation, a conjugate gradient method combined with regularization is used to invert the dynamic convection heat transfer coefficient distribution through the measured temperature data to form a dynamic correction mechanism for thermal boundary conditions; S403: Based on the dynamic correction mechanism of the thermal boundary conditions, a moving Gaussian heat source model is introduced, and a state transfer equation is constructed in combination with the bearing vibration data. The spatial tracking of the friction heat generation position is achieved through a particle filter algorithm to obtain the heat source position; S404: inputting the heat source position into a pre-built turbulence model and transition model, implementing a sliding grid technology to simulate the impeller rotating flow field, and obtaining a flow field simulation result including the distribution of the overheating area; S405, using the flow field simulation results, extracting the main modes of pressure pulsation through intrinsic orthogonal decomposition, and combining it with finite-time Lyapunov exponent calculation to obtain a geometric topological representation of the pressure anomaly propagation path; S406, performing spatiotemporal alignment on the flow field simulation results of the overheating area distribution and the geometric topological representation of the pressure anomaly propagation path to obtain a fused temperature-pressure characteristic map; S407: Based on the temperature-pressure characteristic diagram, identify the overheating area and the pressure anomaly propagation path.

5. The method for diagnosing a fault of a fan in a thermal power plant according to claim 4, characterized in that: In step S401, the transient heat conduction control equation is expressed as: Where ρ represents the material density, c p represents the constant pressure specific heat capacity, T represents the transient temperature field, t represents the time variable, represents the Hamiltonian operator, k(T) represents thermal conductivity, h represents the convective heat transfer coefficient, T ∞ represents the ambient fluid temperature, ∈ represents the surface emissivity, σ represents the Stefan-Boltzmann constant, T sur Indicates the surrounding surface radiation temperature, Q fric represents the frictional heating term.

6. The method for diagnosing a fault of a fan in a thermal power plant according to claim 1, wherein: In step S5, a deep belief network diagnostic model including a fault mode recognition layer and a spatial positioning layer is constructed, including: S501, using the time-frequency ridge code and the blade crack azimuth code as constraint dimensions, embedding the constraint dimensions into the spatial discretization feature matrix through tensor outer product, obtaining a three-dimensional feature tensor containing vibration-acoustic-thermal-fluid four-field coupling information, and generating an initial fusion carrier of multi-physics field data; S502: constructing a nonlinear tensor fusion kernel function including a residual term of a heat conduction equation and a flow field continuity equation through the initial fusion carrier and based on the coordinate set of the overheating area and the pressure anomaly propagation path; S503. Using HOSVD to perform multimodal decomposition on the nonlinear tensor fusion kernel function, extract the core tensor that represents the coupling relationship between multiple physical fields, and establish a correlation probability model between vibration modal frequency, acoustic emission amplitude, temperature gradient, and pressure pulsation based on the core tensor through nuclear norm constraints to form a causal reasoning basis for fault propagation. S504. Based on the causal reasoning foundation, a DBN structure is constructed in which the fault pattern recognition layer and the spatial positioning layer share underlying features, wherein the pattern recognition channel uses a time domain convolution module to capture the time-frequency ridge evolution characteristics, and the positioning channel introduces a spatial self-attention mechanism to analyze the spatial correlation between the crack azimuth and the overheating area; S505. Based on the historical fault case library, a generative adversarial network is constructed to perform sample enhancement, generating an enhanced data set containing compound fault modes, and based on the DBN structure, a multi-task loss function is used to achieve collaborative optimization of pattern classification accuracy and positioning accuracy to obtain the deep belief network diagnostic model.

7. The method for diagnosing a fault of a fan in a thermal power plant according to claim 6, wherein: In step S502, the expression of the nonlinear tensor fusion kernel function is: Where K(Y) represents the nonlinear tensor fusion kernel function of Y, Y represents the three-dimensional feature tensor containing the vibration-acoustic-thermal-fluid four-field coupling information, W h represents the heat conduction kernel weight matrix, W p represents the fluid mechanics kernel weight matrix, W f represents the fracture mechanics kernel weight matrix, λ represents the physical constraint enhancement coefficient, Δ·v represents the flow field divergence, ⊙ represents the Hadamard product, ReLU() represents the linear rectification function, T represents the temperature field, t represents time, α represents the thermal diffusivity, represents the temperature field Laplace operator.

8. The method for diagnosing a fault of a fan in a thermal power plant according to claim 1, wherein: In step S6, the expression of the fault type output by the deep belief network diagnosis model is: Among them, P(f m |D) indicates that under the condition of given input data D, the mth type of failure mode f m The posterior probability of occurrence, exp() represents the natural exponential function, represents the transpose of the weight vector of the mth class of the fault mode recognition layer, z represents the shared hidden layer feature vector, b fm represents the bias scalar of the mth class of the fault pattern recognition layer, M represents the total number of fault types, represents the transpose of the weight vector of the nth class of the fault mode recognition layer, b fn The bias scalar representing the nth class of the fault mode recognition layer; The expression of the three-dimensional spatial positioning result output by the diagnostic model is: Where x, y, and z represent the three-dimensional coordinates of the fault source, and W l Represents the spatial positioning layer regression weight matrix, ReLU() represents the linear rectification function, W h represents the spatial positioning channel hidden layer weight matrix, z represents the shared hidden layer feature vector, b h represents the bias vector of the hidden layer of the spatial positioning channel, b l Represents the regression bias vector of the spatial positioning layer.

9. A thermal power plant fan fault diagnosis system, characterized in that: It includes data acquisition module, diagnosis processing module and result display module; The data acquisition module is used to collect fan vibration signals, acoustic signals, temperature signals and pressure signals; The diagnostic processing module is used to execute the thermal power plant fan fault diagnosis method according to any one of claims 1 to 8; The result display module is used to display the fault diagnosis results, including the fault type and fault location.

10. A computer-readable storage medium, characterized in that include: a memory having a computer program stored thereon; A processor is used to execute the program in the memory to implement the thermal power plant fan fault diagnosis method according to any one of claims 1 to 8.

Images