A method for detecting faults in a gas turbine lubricating oil supply system.

By collecting and processing physical field data in the gas turbine lubricating oil supply system, and combining the finite volume method and phase field evolution method, the lubricating oil flow trajectory is identified, solving the problem of difficulty in capturing early fault characteristics in existing technologies, and realizing accurate fault location and early warning.

CN123069098APending Publication Date: 2026-11-10NO 703 RES INST OF CHINA SHIPBUILDING IND CORP
View PDF 0 Cites 0 Cited by

Patent Information

Application Number
CN202611221016.4
Authority / Receiving Office
CN · China
Patent Type
Applications(China)
Current Assignee / Owner
Filing Date
2026-08-12
Publication Date
2026-11-10

AI Technical Summary

Technical Problem

Existing technologies are insufficient to effectively capture early fault characteristics in gas turbine lubricating oil supply systems, and are susceptible to interference from load fluctuations and environmental changes, making it impossible to achieve accurate location and early warning.

Method used

Based on the fluid network topology of the gas turbine lubricating oil supply system, physical field data is collected by sensors to establish a three-dimensional data structure. Combining the finite volume method and phase field evolution method, the velocity field distribution and virtual tracer particle flow trajectory are calculated to extract dynamic features. The phase field evolution equation is used to identify fault propagation paths and perform adaptive enhancement and noise suppression.

Benefits of technology

It achieves adaptive separation of fault characteristics and background noise, reduces false alarm rate and false negative rate, accurately identifies fault propagation mechanism and impact range, and meets the early warning requirements of gas turbine lubricating oil system.

✦ Generated by Eureka AI based on patent content.

Smart Images

  • Figure CN123069098A_ABST
    Figure CN123069098A_ABST
Patent Text Reader

Abstract

This invention discloses a method for detecting faults in a gas turbine lubricating oil supply system, relating to the field of lubricating oil supply detection technology. The method includes: integrating physical field data collected by sensors into a spatiotemporally aligned multi-physics coupled feature dataset; obtaining a dynamic feature set characterizing the lubricating oil flow trajectory based on the multi-physics coupled feature dataset; obtaining the enhanced fault feature set and background noise separation results based on adaptive enhancement and noise suppression; and comparing the results with fault mode features in a database using cosine similarity to obtain fault type identification results. This invention, based on the real-time velocity field distribution of the entire system and utilizing the interface growth characteristics of phase field evolution, leverages the cascade evolution of river networks to reveal the fault propagation mechanism and impact range, while effectively suppressing residual noise interference. This achieves adaptive enhancement and accurate separation of fault features, reducing the false alarm rate and false negative rate of fault detection.
Need to check novelty before this filing date? Find Prior Art

Description

Technical Field

[0001] This invention relates to the field of lubricating oil supply detection technology, specifically a method for detecting faults in the lubricating oil supply system of a gas turbine. Background Technology

[0002] Oil system failures are characterized by their high degree of concealment, rapid chain reaction, and high degree of harm. They are a major cause of serious accidents such as bearing burnout, gearbox failure, and rotor seizure in gas turbines, accounting for about one-third of all mechanical failures in gas turbines. Once a failure occurs, it will cause huge economic losses and safety hazards.

[0003] Currently, fault detection technology for gas turbine lubricating oil supply systems mainly relies on traditional detection methods based on single physical quantity thresholds. These methods determine system anomalies by setting upper and lower limits for parameters such as pressure, temperature, and flow rate. While these methods are simple to implement and have low deployment costs, they also utilize data-driven intelligent detection methods, employing machine learning algorithms to extract features and recognize patterns from sensor data.

[0004] However, both single physical quantity thresholds and data-driven intelligent detection methods are susceptible to interference from factors such as gas turbine load fluctuations and ambient temperature changes. Furthermore, they fail to fully integrate the fluid network topology characteristics of the lubricating oil system, ignore the intrinsic relationship between faults and lubricating oil flow dynamics, and struggle to capture fluid field anomalies caused by faults. Consequently, the amplitude of early fault characteristics is extremely low, which is submerged by noise data and fails to intuitively present the fault propagation mechanism and impact range.

[0005] While existing technologies have attempted to introduce computational fluid dynamics methods to analyze lubricating oil flow, most are limited to velocity field calculations under steady-state conditions. They lack the ability to characterize complex coupled faults and cannot meet the engineering requirements for early warning and precise location of gas turbine lubricating oil systems. Summary of the Invention

[0006] To address the shortcomings of existing technologies, this invention provides a method for detecting faults in a gas turbine lubricating oil supply system, comprising the following specific steps: S100, based on the fluid network topology of the gas turbine lubricating oil supply system, physical field data is collected by sensors, and then spatial coordinate mapping is performed after time stamp calibration to establish a three-dimensional data structure containing time, space and physical quantities. After being combined with the time stamp, three-dimensional data points are formed. After outlier removal, the data is integrated into a spatiotemporally aligned multi-physics coupling feature dataset. S200. Based on the multi-physics coupling feature dataset and fluid network topology, the finite volume method is used to calculate the velocity field distribution of lubricating oil in the entire gas turbine lubricating oil supply system at each moment. Based on numerical simulation, several virtual tracer particles are released along the flow direction of lubricating oil. Statistical analysis is performed on the flow trajectory and particle parameters of the virtual tracer particles in the lubricating oil system to obtain a dynamic feature set that characterizes the flow trajectory of lubricating oil. S300: Based on mapping each feature point in the dynamic feature set to the same two-dimensional plane, and obtaining the comprehensive feature value of each trajectory, a phase field variable is introduced, and the comprehensive feature value is used as the driving term of the phase field evolution equation. After dividing the two-dimensional plane into several grid cells of equal size, the phase field evolution equation is solved to obtain the phase field distribution at each time. The phase field distribution is used as the terrain surface, and the value of the phase field variable of each grid cell is used as the height value of the terrain surface. Based on the self-organization principle of the river network, the region composed of all grid cells through which the main fault propagation path passes is used as the initial fault candidate region set. After completing the region replication based on the initial fault candidate region set, adaptive enhancement and noise suppression are performed to obtain the enhanced fault feature set and the background noise separation result. S400. Based on the enhanced fault feature set, multi-dimensional feature extraction is performed, and cosine similarity is compared with fault mode features in the database to obtain fault type identification results.

[0007] Preferably, S100 collects physical parameters, including pressure field, temperature field, flow field, vibration field and oil quality parameters, at the lubricating oil inlet and outlet nodes based on the fluid network topology of the gas turbine lubricating oil supply system. The original signals of the collected physical field data are timestamped. Using the global clock of the gas turbine control system as a reference, the local timestamp of each sampling point in the data collected by different sensors is compared with the timestamp of the global clock to calculate the time difference. Based on the time difference, the timestamp of the sampling point is adjusted to the timestamp of the global clock. The adjusted sampling points are rearranged in chronological order. Each sensor is assigned a unique spatial coordinate identifier. It should be noted that the identifier consists of the device number where the sensor is located and the specific location number on the device. The measurement value of each sampling point is combined with the corresponding spatial coordinate identifier and the adjusted timestamp to form a three-dimensional data point. The S100 process uses the median filtering method to remove outliers from the three-dimensional data points to obtain a spatiotemporally aligned multiphysics coupling feature dataset.

[0008] Preferably, step S200, based on a multi-physics coupling feature dataset and a fluid network topology, uses the finite volume method to solve the flow control equations of the lubricating oil in the system, and calculates the velocity field distribution of the lubricating oil in the entire gas turbine lubricating oil supply system at each moment, specifically: Each pipeline in the gas turbine lubricating oil supply system is divided into several control volumes of equal length. Based on these control volumes, discretized flow control equations are established according to the laws of conservation of mass and momentum. The difference between the mass flow rate flowing into and out of the control volume is calculated based on the law of conservation of mass, and this difference is used as the rate of change of fluid mass within the control volume. The resultant force of all forces acting on the control volume is calculated based on the law of conservation of momentum, and this resultant force is used as the rate of change of fluid momentum within the control volume. The continuous differential equations of the difference and resultant force are integrated over the control volume to obtain discretized algebraic equations, where the time derivative term uses a forward difference scheme and the spatial derivative term uses a central difference scheme. In the specific calculations, the laws of conservation of mass and momentum refer to the mass conservation equation and the momentum conservation equation, respectively. Based on the pressure and flow rates of the physical parameters collected at the lubricating oil inlet and outlet nodes using the multiphysics coupling feature dataset as boundary conditions, the discretized flow control equations are solved using an iterative method after being substituted into the discretized algebraic equations. Based on the flow rate of each control volume obtained from the solution, the velocity value at the center of each control volume is calculated to obtain the velocity field distribution of the gas turbine lubricating oil supply system. It should be noted that the velocity value is equal to the flow rate of the control volume divided by the cross-sectional area of ​​the control volume.

[0009] Preferably, after obtaining the velocity field distribution and flow velocity value, S200 releases several virtual tracer particles in the velocity field starting from the lubricating oil tank outlet and along the flow direction of the lubricating oil. The initial position of each virtual tracer particle is uniformly distributed on the cross-section of the lubricating oil tank outlet. Specifically, the influence of different particle numbers on the statistical characteristics of the flow trajectory is analyzed through numerical simulation experiments. When the number of particles increases to a certain extent, the statistical characteristics no longer change significantly, and the particle number at this point is taken as the final number of virtual tracer particles. The cross-section of the lubricating oil tank outlet is divided into small regions with an area equal to the number of virtual tracer particles. A virtual tracer particle is released at the center of each small region, and the flow trajectory of each virtual tracer particle in the gas turbine lubricating oil supply system is tracked. The physical parameters of each virtual tracer particle passing through each lubricating oil inlet and outlet node are recorded. Specifically: Based on the velocity field distribution at each moment, the velocity of each virtual tracer particle at the current moment is calculated, and this velocity value is used as the flow velocity value at the center of the control volume where the virtual tracer particle is located. Based on the velocity and time step of the virtual tracer particle, the position of the virtual tracer particle at the next moment is calculated. The position coordinates are equal to the position coordinates at the current moment plus the velocity multiplied by the time step. It should be noted that the time step is equal to the length of the control volume divided by the product of the maximum flow velocity in the system and the Courant number. The Courant number, which can guarantee the absolute stability of the particle flow trajectory calculation, is determined through numerical experiments. Substituting this Courant number into the formula yields the final time step. There is a possibility that the virtual tracer particle moves from one control volume to another. When a virtual tracer particle passes through each lubricating oil inlet / outlet node, the corresponding time and physical parameters are updated. This process is repeated until all virtual tracer particles reach the return oil manifold and / or exceed the preset maximum tracking time, thus obtaining the flow trajectory of each virtual tracer particle in the gas turbine lubricating oil supply system. Statistical analysis is performed on the flow trajectories of all virtual tracer particles to extract the dynamic features of each flow trajectory. After uniformly normalizing all extracted dynamic features, features of different dimensions and orders of magnitude are converted into the same numerical range to obtain a set of dynamic features of the lubricating oil flow trajectory composed of dynamic feature points.

[0010] Preferably, step S300 maps each feature point in the normalized dynamic feature set of lubricating oil flow trajectory to the same two-dimensional plane, specifically as follows: The comprehensive feature value of the corresponding lubricating oil flow trajectory is obtained by summing all the normalized dynamic feature values ​​in each lubricating oil flow trajectory. The comprehensive feature value is used as the horizontal axis of the two-dimensional plane, and the time corresponding to the lubricating oil flow trajectory is used as the vertical axis of the two-dimensional plane. Each lubricating oil flow trajectory corresponds to a point on the two-dimensional plane, and the points corresponding to all the lubricating oil flow trajectories together form the feature point distribution on the two-dimensional plane. By introducing phase field variables, a phase field evolution equation consisting of volume free energy and gradient energy terms is constructed. The comprehensive eigenvalues ​​in the normalized dynamic characteristic set of the lubricating oil flow trajectory are used as the driving terms of the phase field evolution equation, specifically: For each point on the two-dimensional plane, the average value of the comprehensive characteristic value of all lubricating oil flow trajectories is calculated and used as the original driving term value, which is added to the right side of the phase field evolution equation as an additional driving term.

[0011] Preferably, the S300 solves the phase field evolution equation to obtain the phase field distribution at each time step, specifically as follows: The two-dimensional plane is divided into several grid cells of equal size. By analyzing the distribution density of feature points, the maximum grid cell side length that can guarantee that each grid cell contains at least one feature point on average is calculated and used as the final grid cell side length. The distribution of phase field variables is initialized by setting the phase field variables of all grid cells to zero. The phase field evolution equation is discretized using the finite difference method to obtain the new values ​​of the phase field variables of each grid cell at each time step. The time derivative term uses a forward difference scheme, the first spatial derivative term uses a central difference scheme, and the second spatial derivative term uses a five-point difference scheme. The time step is calculated based on the numerical stability condition of the explicit finite difference method. This time step is equal to the square of the grid cell side length divided by the ratio of the stability coefficient to the product of the mobility and gradient energy coefficients. The stability coefficient is determined by numerical... The values ​​are determined experimentally to ensure the numerical stability of the explicit finite difference method. The operation is repeated until the preset maximum evolution time and / or the change in the phase field distribution is less than the preset convergence threshold. The maximum evolution time is calculated by analyzing the growth rate of the fault feature region and calculating the time required for the fault feature region to cover all relevant feature points from its initial appearance. The convergence threshold is calculated by analyzing the initial change rate of the phase field variables and calculating the minimum change that can distinguish between the actual evolution and the numerical error. The product of this minimum change and the accuracy coefficient is used as the convergence threshold to obtain the phase field distribution composed of the phase field variables of each grid cell.

[0012] Preferably, in step S300, the value of the phase field variable of each grid cell on the two-dimensional plane is used as the height value of the terrain surface corresponding to the grid cell. All grid cells on the terrain surface with height values ​​greater than a preset threshold are used as the starting point of the feature flow. The preset threshold is calculated by statistically analyzing the distribution of height values ​​of all grid cells, calculating the average value and standard deviation of the height value distribution, and using the sum of the multiples of the average value and standard deviation as the threshold to exclude most background noise points. Based on each starting point of the feature flow, the slope between the corresponding grid cell and eight adjacent grid cells is calculated. The slope is calculated by dividing the height difference between two grid cells by the distance between the two grid cells. The distance between horizontally and vertically adjacent grid cells is one grid cell side length, and the distance between diagonally adjacent grid cells is the square root of two grid cell side lengths. The direction of the feature flow is based on the direction of the steepest slope. If there are multiple directions with the same and steepest slope, the water flow is evenly distributed to these directions. After the feature flow flows from the current grid cell to the adjacent grid cell, it continues to repeat on the adjacent grid cells until the feature flow reaches the lowest point of the terrain surface and / or flows out of the boundary of the two-dimensional plane. The network structure of fault propagation paths formed by characteristic flows on the terrain surface is recorded. Using characteristic flows as fault propagation paths, the process includes: assigning a unique number to each fault propagation path; calculating the length of each path based on its start and end points; using the sum of the side lengths of all grid cells traversed by the path as the length; using the sum of the cumulative intensities of all characteristic flows at each grid cell as the characteristic intensity; and using the total area of ​​the grid cells into which all characteristic flows eventually converge as the influence range of the corresponding fault propagation path. Then, the catchment area in the self-organizing criticality principle of the flow network is considered. For example, if a fault propagation path converges into 20 grid cells, and each grid cell has an area of ​​0.01, the influence range is 0.2. The obtained parameters are integrated to form the fault propagation path network structure.

[0013] Preferably, step S300 calculates the fractal dimension and power-law exponent of the fault propagation path network structure, specifically as follows: Based on box counting, a two-dimensional plane is divided into several square boxes of the same size, and the number of boxes containing fault propagation paths is counted. The size of the boxes is changed and the operation is repeated to obtain several data pairs of box size and corresponding box number. The absolute value of the slope of the obtained regression line is used as the fractal dimension through double log-linear regression. Based on the influence range and characteristic intensity of the obtained fault propagation path, data pairs of influence range and corresponding characteristic intensity are formed, and the slope of the regression line obtained by double log-linear regression is used as the power law exponent. The fault propagation path with the largest feature intensity and influence range in the fault propagation path network is extracted as the main fault propagation path, which corresponds to the main river in the self-organization principle of river network. The phase field region corresponding to the main fault propagation path and all grid cells passed through the path are used as the initial set of fault candidate regions.

[0014] Preferably, S300 extracts the position information and phase field variable values ​​of all grid cells traversed by the main fault propagation path in the initial fault candidate region set, and combines them to form a fault candidate region. The structure of the fault candidate region includes the boundary coordinates of the region, the average value of the phase field variables of all grid cells in the region, and the area of ​​the region. The fault matching degree of the fault candidate region is obtained by summing the average value of the phase field variables of the fault candidate region, the characteristic intensity of the fault propagation path, and the influence range of the fault propagation path. Based on the fault matching degree of each fault candidate region in the initial fault candidate region set, the fault candidate regions are sorted from high to low according to the fault matching degree. The top few fault candidate regions with the highest fault matching degree are selected for region copying operation to obtain copied fault candidate regions. The copied fault candidate regions are exactly the same as the original fault candidate regions, and the number of copied regions is defined by the user. In this application, the top 2 regions are selected for copying. The more regions copied, the better the noise separation result and the enhancement effect, but the heavier the computational burden. Then, the boundary coordinates of the copied fault candidate region are expanded outward and / or shrunken inward, and the corresponding phase field region is translated a certain distance on the two-dimensional plane so that the boundary of the copied fault candidate region does not contact the boundary of the corresponding fault candidate region. Then, the fault matching degree between the fault candidate region and the adjusted copied fault candidate region is calculated. The region with the highest fault matching degree is selected as the new original fault candidate region, and the new original fault candidate region replaces the fault candidate region with the lowest fault matching degree in the initial set of fault candidate regions. After replacing all the duplicated fault candidate regions, the phase field region corresponding to the fault candidate region with the highest fault matching degree in the final set of replaced fault candidate regions is taken as the enhanced fault feature region, and the remaining phase field regions are background noise regions. The physical field data of the enhanced fault feature region is extracted to obtain the fault feature set.

[0015] Preferably, S400 performs multi-dimensional feature extraction on the enhanced fault feature set, extracting feature parameters including the center position, area, average intensity, intensity change rate, and shape parameters of the fault feature region, and uses cosine similarity to calculate the similarity between the extracted multi-dimensional feature parameters and the features of each fault mode in the database, and selects the fault mode with the highest similarity as the recognition result.

[0016] Beneficial effects This invention provides a method for detecting faults in a gas turbine lubricating oil supply system, which has the following advantages: Based on the real-time velocity field distribution of the entire system and utilizing the interface growth characteristics of phase field evolution, the comprehensive characteristic value of the lubricating oil flow trajectory is used as the driving term of phase field evolution, so that the fault feature region naturally aggregates and grows during the evolution process, while the background noise region is gradually suppressed, initially achieving the separation of fault features and background noise. The phase field distribution is transformed into a terrain surface, and the cascade evolution of the river network is used to identify the main fault propagation path with the strongest feature intensity and the widest influence range, revealing the fault propagation mechanism and influence range. Then, based on the initial set of fault candidate regions, regional replication and adaptive optimization are performed to replace low-matching regions, further enhancing the salience of fault features, while effectively suppressing the interference of residual noise, realizing adaptive enhancement and accurate separation of fault features, and reducing the false alarm rate and false negative rate of fault detection. Attached Figure Description

[0017] Figure 1 This is a schematic diagram of the process of the present invention. Detailed Implementation

[0018] The present invention will now be further described with reference to the accompanying drawings.

[0019] Please see Figure 1 This invention provides a method for detecting faults in a gas turbine lubricating oil supply system, comprising the following specific steps: S100, based on the fluid network topology of the gas turbine lubricating oil supply system, physical field data is collected by sensors, and then spatial coordinate mapping is performed after time stamp calibration to establish a three-dimensional data structure containing time, space and physical quantities. After being combined with the time stamp, three-dimensional data points are formed. After outlier removal, the data is integrated into a spatiotemporally aligned multi-physics coupling feature dataset. S100, based on the fluid network topology of the gas turbine lubricating oil supply system, collects physical parameters including pressure field, temperature field, flow field, vibration field, and oil quality parameters at the lubricating oil inlet and outlet nodes. These lubricating oil inlet and outlet nodes refer to key locations such as the lubricating oil tank outlet, main oil pump inlet and outlet, filter inlet and outlet, cooler inlet and outlet, oil supply ports for each bearing, and the return oil manifold. In some other specific embodiments, the pressure field parameter acquisition frequency is determined by analyzing the pressure fluctuation signal under normal operation of the gas turbine lubricating oil supply system using Fast Fourier Transform (FFT). The highest frequency component of the signal is extracted, and the acquisition frequency is set as the product of this highest frequency component and a safety factor. The safety factor is determined by analyzing the harmonic content of the pressure fluctuation signal to ensure that all effective frequency components can be completely captured. For example, the main oil pump inlet and outlet nodes are obtained through FFT analysis. The highest frequency component of the oil pump pressure fluctuation is 20Hz. Harmonic content analysis shows the presence of 2-3 times higher frequency harmonics. Therefore, a safety factor of 2.5 is taken, and the final pressure field acquisition frequency is set to 50Hz. Similarly, the temperature field and flow field parameter acquisition frequencies are analyzed using the same fast Fourier transform method to extract the highest frequency components of the temperature and flow fluctuation signals. The acquisition frequency is set to the product of the corresponding highest frequency component and the safety factor. The vibration field parameter acquisition frequency is obtained by extracting the fundamental frequency and harmonic frequencies of the vibration signal of the rotating components of the gas turbine through spectrum analysis. The acquisition frequency is set to the product of the highest harmonic frequency and the safety factor. The oil quality parameter acquisition frequency is obtained by analyzing historical data of lubricating oil quality changes over time, fitting the time constant of lubricating oil quality changes, and setting the acquisition frequency to the reciprocal of this time constant. The original signals of the acquired physical field data are timestamped. Using the global clock of the gas turbine control system as a reference, the local timestamp of each sampling point in the data collected by different sensors is compared with the timestamp of the global clock to calculate the time difference. Based on the time difference, the timestamp of the sampling point is adjusted to the timestamp of the global clock, and the adjusted sampling points are rearranged in chronological order. Each sensor is assigned a unique spatial coordinate identifier. It should be noted that this identifier consists of the device number on which the sensor is located and its specific location number on the device. The measured value of each sampling point is combined with the corresponding spatial coordinate identifier and the adjusted timestamp to form a three-dimensional data point. For example, the spatial coordinate identifier of the main oil pump outlet pressure sensor is P001-01, where P001 represents the main oil pump and 01 represents the outlet position. If the pressure value collected at 10:00:00.001 is 0.5MPa, then the resulting three-dimensional data point is (10:00:00.001, P001-01, 0.5MPa). S100 uses median filtering to remove outliers from 3D data points, obtaining a spatiotemporally aligned multiphysics coupling feature dataset. In some other specific embodiments, this application provides an outlier removal method, including: for the time series data of 3D data points of each physical quantity, by analyzing the time scale of normal fluctuations of the physical quantity, calculating the window length that can cover the complete fluctuation cycle, and using this window length as the size of the sliding window; starting from the first sampling point of the time series, sliding the window sequentially to the next sampling point; for the sampling points within each window, arranging their measurements in ascending order, and taking the measurement value at the middle position as the filtered value of the center sampling point of the window; when the window slides to both ends of the time series, supplementing the sampling points within the window using a symmetrical expansion method, that is, for the first sampling point at the beginning of the sequence, the sampling points within the window are the sum of the sum of the sampling point and subsequent adjacent sampling points, with the number of repetitions equal to half the size of the sliding window rounded up; for the last sampling point at the end of the sequence, the sampling points within the window are the sum of the sum of the preceding adjacent sampling points and the sampling point, with the number of repetitions equal to half the size of the sliding window rounded up.

[0020] S200. Based on the multi-physics coupling feature dataset and fluid network topology, the finite volume method is used to calculate the velocity field distribution of lubricating oil in the entire gas turbine lubricating oil supply system at each moment. Based on numerical simulation, several virtual tracer particles are released along the flow direction of lubricating oil. Statistical analysis is performed on the flow trajectory and particle parameters of the virtual tracer particles in the lubricating oil system to obtain a dynamic feature set that characterizes the flow trajectory of lubricating oil. Based on a multiphysics coupling feature dataset and fluid network topology, S200 uses the finite volume method to solve the flow control equations of lubricating oil in the system, calculating the velocity field distribution of lubricating oil in the entire gas turbine lubricating oil supply system at each moment. Specifically: Each pipeline in the gas turbine lubricating oil supply system is divided into several control volumes of equal length. Based on these control volumes, discretized flow control equations are established according to the laws of conservation of mass and momentum. The difference between the mass flow rate flowing into and out of the control volume is calculated based on the law of conservation of mass, and this difference is used as the rate of change of fluid mass within the control volume. The resultant force of all forces acting on the control volume is calculated based on the law of conservation of momentum, and this resultant force is used as the rate of change of fluid momentum within the control volume. The continuous differential equations of the difference and resultant force are integrated over the control volume to obtain discretized algebraic equations, where the time derivative term uses a forward difference scheme and the spatial derivative term uses a central difference scheme. In the specific calculations, the laws of conservation of mass and momentum refer to the mass conservation equation and the momentum conservation equation, respectively. Based on the pressure and flow rates of the physical parameters collected at the lubricating oil inlet and outlet nodes using the multiphysics coupling feature dataset as boundary conditions, the discretized flow control equations are solved by an iterative method after being substituted into the discretized algebraic equations. Based on the flow rate of each control volume obtained by the solution, the velocity value at the center of each control volume is calculated to obtain the velocity field distribution of the gas turbine lubricating oil supply system. It should be noted that the velocity value is equal to the flow rate of the control volume divided by the cross-sectional area of ​​the control volume. In the above-mentioned method of solving discretized flow control equations using an iterative method, this application provides a specific solution process: First, an initial velocity field and pressure field distribution are assumed. Then, the velocity field is calculated based on the initially assumed pressure field. Next, the pressure field is updated based on the calculated velocity field. The alternating update process of the velocity field and pressure field is repeated until the difference between the calculation results of two adjacent iterations is less than a preset convergence threshold. This convergence threshold is calculated by analyzing the amplitude of normal fluctuations of the physical quantity and determining the minimum difference that can distinguish between real fluctuations and numerical errors. The product of this minimum difference and the accuracy coefficient is used as the convergence threshold. To uniformly handle all physical quantities, a uniform relative convergence threshold is used. For example, if the normal fluctuation amplitude of pressure is 0.05 MPa, the minimum difference that can distinguish between real fluctuations and numerical errors is 0.0005 MPa. Therefore, the convergence threshold is set to 0.0005 MPa.

[0021] After obtaining the velocity field distribution and flow velocity values, S200 releases several virtual tracer particles in the velocity field starting from the oil tank outlet and along the flow direction of the lubricating oil. The initial positions of each virtual tracer particle are uniformly distributed on the cross-section of the oil tank outlet. Specifically, numerical simulation experiments are used to analyze the influence of different particle numbers on the statistical characteristics of the flow trajectory. When the particle number increases to a certain extent, the statistical characteristics no longer change significantly; this particle number is taken as the final number of virtual tracer particles. The cross-section of the oil tank outlet is divided into small regions with an area equal to the number of virtual tracer particles, and one virtual tracer particle is released at the center of each small region. For example, when the particle number increases from 50 to 100, the change in the average flow velocity statistical value is less than 1%, therefore 100 virtual tracer particles are selected. The area of ​​the oil tank outlet is 100 cm². 2 Therefore, the area of ​​each small region is 1 cm². 2 One particle is released at the center of each small area; Then, the flow trajectory of each virtual tracer particle in the gas turbine lubricating oil supply system is tracked, and the physical parameters of each virtual tracer particle as it passes through each lubricating oil inlet and outlet node are recorded, specifically: Based on the velocity field distribution at each moment, the velocity of each virtual tracer particle at the current moment is calculated, and the velocity value is used as the flow velocity value at the center of the control volume where the virtual tracer particle is located. Based on the velocity and time step of the virtual tracer particle, the position of the virtual tracer particle at the next moment is calculated. The position coordinate is equal to the position coordinate at the current moment plus the velocity multiplied by the time step. It should be noted that the time step is equal to the length of the control volume divided by the product of the maximum flow velocity in the system and the Coulomb number. The Coulomb number that can guarantee the absolute stability of the particle flow trajectory calculation is determined through numerical experiments. The Coulomb number is substituted into the formula to obtain the final time step. When a virtual tracer particle moves from one control volume to another, the control volume number where the virtual tracer particle is located is updated. When a virtual tracer particle passes through each lubricating oil inlet / outlet node, the corresponding time and physical parameters are updated and recorded. The operation is repeated until all virtual tracer particles have reached the return oil manifold and / or exceeded the preset maximum tracking time, and the flow trajectory of each virtual tracer particle in the gas turbine lubricating oil supply system is obtained. It should be noted that the maximum tracking time is determined by statistically analyzing the flow time of all virtual tracer particles from the lubricating oil tank outlet to the return oil manifold under normal operating conditions. A cumulative probability function of the flow time distribution is constructed, and the time value corresponding to the cumulative probability value of the vast majority of virtual tracer particles completing the flow process is taken as the maximum tracking time. For example, if the flow time of 100 virtual tracer particles is statistically analyzed, and the flow time of 95 particles is between 8 and 12 seconds, the time corresponding to the cumulative probability reaching 95% is 12 seconds. Therefore, the maximum tracking time is set to 12 seconds. Statistical analysis is performed on the flow trajectories of all virtual tracer particles to extract the dynamic features of each flow trajectory. After uniformly normalizing all the extracted dynamic features, features of different dimensions and orders of magnitude are converted into the same numerical range to obtain a set of dynamic features of lubricating oil flow trajectory composed of dynamic feature points. The dynamic features extracted for each flow trajectory include the average velocity, maximum velocity, minimum velocity, pressure gradient, temperature gradient, residence time, and velocity, pressure, and temperature differences between adjacent flow trajectories. The average velocity is the sum of the velocity values ​​at all sampling points on the trajectory divided by the number of sampling points. The maximum velocity is the maximum velocity value among all sampling points on the trajectory, and the minimum velocity is the minimum velocity value among all sampling points on the trajectory. The pressure gradient is the difference between the pressure values ​​at the last sampling point and the pressure values ​​at the first sampling point, divided by the total length of the trajectory. The temperature gradient is the difference between the temperature values ​​at the last sampling point and the temperature values ​​at the first sampling point, divided by the total length of the trajectory. The residence time is the difference between the time at the last sampling point and the time at the first sampling point. The velocity, pressure, and temperature differences between adjacent flow trajectories are the absolute values ​​of the subtraction of the corresponding parameter values ​​at the same location on two adjacent trajectories.

[0022] S300: Based on mapping each feature point in the dynamic feature set to the same two-dimensional plane, and obtaining the comprehensive feature value of each trajectory, a phase field variable is introduced, and the comprehensive feature value is used as the driving term of the phase field evolution equation. After dividing the two-dimensional plane into several grid cells of equal size, the phase field evolution equation is solved to obtain the phase field distribution at each time. The phase field distribution is used as the terrain surface, and the value of the phase field variable of each grid cell is used as the height value of the terrain surface. Based on the self-organization principle of the river network, the region composed of all grid cells through which the main fault propagation path passes is used as the initial fault candidate region set. After completing the region replication based on the initial fault candidate region set, adaptive enhancement and noise suppression are performed to obtain the enhanced fault feature set and the background noise separation result. S300 maps each feature point in the normalized dynamic feature set of lubricating oil flow trajectory to the same two-dimensional plane, specifically: The comprehensive feature value of the corresponding lubricating oil flow trajectory is obtained by summing all the normalized dynamic feature values ​​in each lubricating oil flow trajectory. The comprehensive feature value is used as the horizontal axis of the two-dimensional plane, and the time corresponding to the lubricating oil flow trajectory is used as the vertical axis of the two-dimensional plane. Each lubricating oil flow trajectory corresponds to a point on the two-dimensional plane, and the points corresponding to all the lubricating oil flow trajectories together form the feature point distribution on the two-dimensional plane. A phase field variable is introduced, which is a continuous scalar field with values ​​ranging from zero to unit. When the phase field variable is equal to zero, it indicates that the location is in the normal state region; when the phase field variable is equal to unit, it indicates that the location is in the fault characteristic region; when the phase field variable is between zero and unit, it indicates that the location is in the transition region between the normal state and the fault characteristic region. A phase field evolution equation composed of volume free energy and gradient energy terms is constructed. The comprehensive characteristic value in the normalized dynamic characteristic set of lubricating oil flow trajectory is used as the driving term of the phase field evolution equation, specifically: Based on each point on the two-dimensional plane, the average value of the comprehensive characteristic value of all lubricating oil flow trajectories is calculated and used as the original driving term value. For example, if there are 3 trajectories passing through a certain point, and the comprehensive characteristic values ​​are 2.95, 3.0, and 3.05 respectively, then the original driving term value is 3. After normalizing all the original driving term values, the minimum value of all the original driving term values ​​is subtracted from each original driving term value, and then divided by the difference between the maximum and minimum values ​​of all the original driving term values ​​to obtain the normalized driving term value, which ranges from zero to a unit value. It is added to the right side of the phase field evolution equation as an additional driving term. In the above, the driving term added in this application has the following effect: when the comprehensive characteristic value at a certain point is large, the driving term will accelerate the evolution of the phase field variable at that point towards a unit value, thereby promoting the growth of the fault characteristic region. When the comprehensive characteristic value at a certain point is small, the effect of the driving term can be ignored. To further understand the phase field evolution equation proposed in this application, which is composed of volume free energy term and gradient energy term, the specific operation is as follows: Define a free energy functional, which consists of two parts: volume free energy term and gradient energy term; The volume free energy term adopts the form of a double potential well function, which has a minimum value at the point where the phase field variable is equal to zero and the point where the phase field variable is equal to a unit value, respectively corresponding to two stable potential wells in the normal state and the fault characteristic state. The gradient energy term is equal to the gradient energy coefficient multiplied by half the square of the gradient of the phase field variable. The gradient energy coefficient is a positive constant, and it is calculated by taking the square of the transition region width and dividing it by the product of the interface energy coefficient and the logarithmic transformation coefficient. The interface energy coefficient and the logarithmic transformation coefficient are determined through phase field numerical simulation experiments to ensure the stability and clarity of the phase field interface. The transition region width is analyzed through numerical experiments to determine the impact of different transition region widths on the stability of the phase field interface and the clarity of fault feature boundaries, and a transition region width that can simultaneously ensure interface stability and boundary clarity is selected. Therefore, the phase field evolution equation is in the form that the rate of change of the phase field variable with respect to time is equal to the negative mobility multiplied by the variational derivative of the free energy functional with respect to the phase field variable; the mobility is a positive constant, which is calculated by taking the gradient energy coefficient and dividing it by the energy difference between the normal state and the fault characteristic state; the energy difference between the normal state and the fault characteristic state is determined by the phase field numerical simulation experiment.

[0023] S300 solves the phase field evolution equation to obtain the phase field distribution at each time step, specifically: The two-dimensional plane is divided into several grid cells of equal size. By analyzing the distribution density of feature points, the maximum grid cell side length that ensures that each grid cell contains at least one feature point on average is calculated and used as the final grid cell side length. In the specific implementation, the distribution density of feature points in the horizontal axis direction is 10 per unit length and in the vertical axis direction is 5 per unit length. Therefore, the grid cell side length is set to 0.1 to ensure that each grid cell contains an average of 0.5 feature points. The distribution of phase field variables is initialized by setting the phase field variables of all grid cells to zero. The phase field evolution equation is discretized using the finite difference method to obtain the new values ​​of the phase field variables of each grid cell at each time step. The time derivative term uses a forward difference scheme, the first spatial derivative term uses a central difference scheme, and the second spatial derivative term uses a five-point difference scheme. The time step is calculated based on the numerical stability condition of the explicit finite difference method. This time step is equal to the square of the grid cell side length divided by the ratio of the stability coefficient to the product of the mobility and gradient energy coefficients. The stability coefficient is determined by numerical... The values ​​are determined experimentally to ensure the numerical stability of the explicit finite difference method. The operation is repeated until the preset maximum evolution time and / or the change in the phase field distribution is less than the preset convergence threshold. The maximum evolution time is calculated by analyzing the growth rate of the fault feature region and calculating the time required for the fault feature region to cover all relevant feature points from its initial appearance. The convergence threshold is calculated by analyzing the initial change rate of the phase field variables and calculating the minimum change that can distinguish between the actual evolution and the numerical error. The product of this minimum change and the accuracy coefficient is used as the convergence threshold to obtain the phase field distribution composed of the phase field variables of each grid cell.

[0024] S300 uses the phase field variable value of each grid cell on the two-dimensional plane as the height value of the terrain surface corresponding to the grid cell. In this application, the obtained phase field distribution is regarded as a terrain surface in this step. The larger the value of the phase field variable, the higher the height of the corresponding terrain surface. For example, if the phase field variable of a certain grid cell is 0.8, then the terrain height corresponding to that grid cell is 0.8. Thus, the flow process of the characteristic flow is simulated on the terrain surface. The characteristic flow starts from the highest point of the terrain surface and flows downward along the direction of the steepest slope. Here, the characteristic flow corresponds to the water in the self-organizing principle of river networks, with all heights on the terrain surface... Grid cells with height values ​​greater than a preset threshold are used as the starting point of the feature flow. This preset threshold is calculated by statistically analyzing the distribution of height values ​​of all grid cells, calculating the average and standard deviation of the height value distribution, and using the sum of the multiples of the average and standard deviation as the threshold to exclude the vast majority of background noise points. Based on each starting point of the feature flow, the slope between the corresponding grid cell and its eight adjacent grid cells is calculated. The slope is calculated by dividing the height difference between two grid cells by the distance between the two grid cells. The distance between horizontally and vertically adjacent grid cells is one grid cell side length, and the distance between diagonally adjacent grid cells is the square root of two grid cell side lengths. The direction of the feature flow is based on the direction of the steepest slope. If there are multiple directions with the same and steepest slope, the water flow is evenly distributed to these directions. After the feature flow flows from the current grid cell to the adjacent grid cell, it continues to repeat on the adjacent grid cells until the feature flow reaches the lowest point of the terrain surface and / or flows out of the boundary of the two-dimensional plane. In some other specific embodiments, if multiple directions have the same and maximum slope, the water flow is evenly distributed to these directions. Therefore, when a characteristic flow encounters multiple possible flow directions, the intensity of the characteristic flow is distributed according to Horton's Law. This application provides its specific operation, defining the levels of fault propagation paths. The source fault propagation path is the first-level fault propagation path, which corresponds to the source river in the self-organizing principle of river networks. When two fault propagation paths of the same level merge, the level of the formed fault propagation path is one level higher than the original. When fault propagation paths of different levels merge, the level of the formed fault propagation path is the same as the higher-level fault propagation path. Horton's Law states that the average characteristic intensity of the nth-level fault propagation path is approximately the branch ratio multiple of the average characteristic intensity of the (n-1)th-level fault propagation path, which corresponds to the cascading relationship of river flow in the self-organizing principle of river networks. The branch ratio is calculated by statistically analyzing the number of fault propagation paths at each level in the fault propagation path network under normal conditions, calculating the ratio of the number of fault propagation paths at two adjacent levels, and taking the average value as the branch ratio. For example, if statistics show that there are 16 first-level fault propagation paths, 4 second-level paths, and 1 third-level path, then the branch ratio is 16 / 4 = 4. In the process of simulating characteristic flow, when two fault propagation paths merge, the characteristic intensity of the merged fault propagation path is calculated according to their level and Horton's law; for example, if two first-level fault propagation paths merge and each has a characteristic intensity of 0.1, then the characteristic intensity of the second-level fault propagation path formed after merging is 0.1·4=0.4. The network structure of fault propagation paths formed by characteristic flows on the terrain surface is recorded. Using characteristic flows as fault propagation paths, the process includes: assigning a unique number to each fault propagation path; calculating the length of each path based on its start and end points; using the sum of the side lengths of all grid cells traversed by the path as the length; using the sum of the cumulative intensities of all characteristic flows at each grid cell as the characteristic intensity; and using the total area of ​​the grid cells into which all characteristic flows eventually converge as the influence range of the corresponding fault propagation path. Then, the catchment area in the self-organizing criticality principle of the flow network is considered. For example, if a fault propagation path converges into 20 grid cells, and each grid cell has an area of ​​0.01, the influence range is 0.2. The obtained parameters are integrated to form the fault propagation path network structure.

[0025] S300 calculates the fractal dimension and power-law exponent of the fault propagation path network structure, specifically as follows: Based on box counting, a two-dimensional plane is divided into several square boxes of the same size, and the number of boxes containing fault propagation paths is counted. The size of the boxes is changed, and the operation is repeated to obtain several data pairs of box sizes and corresponding box numbers. The absolute value of the slope of the regression line obtained by double log-linear regression is used as the fractal dimension. In some other specific implementations, when the box size s=0.1, the number of boxes containing fault propagation paths N=100; when s=0.2, N=25; the slope of the double log-linear regression is -2, so the fractal dimension is 2. Based on the influence range and characteristic intensity of the obtained fault propagation path, data pairs of influence range and corresponding characteristic intensity are formed, and the slope of the regression line obtained by double log-linear regression is used as the power law exponent. The fractal dimension and power law exponent obtained by this application based on the above series of operations are because, when the system is in a normal state, the fractal dimension and power law exponent of the fault propagation path network remain within a relatively stable range; when the system fails, the growth of the fault feature region will cause changes in the terrain surface, thereby causing the fractal dimension and power law exponent of the fault propagation path network to deviate from the normal range. Therefore, the fault propagation path with the largest feature intensity and influence range in the fault propagation path network is extracted as the main fault propagation path, which corresponds to the main river in the self-organization principle of river network. The phase field region corresponding to the main fault propagation path and all grid cells passed through the path are used as the initial set of fault candidate regions. In some other specific embodiments, this application provides a method for extracting the fault propagation path with the largest feature intensity and influence range in the fault propagation path network as the backbone fault propagation path. All fault propagation paths are sorted according to their feature intensity from largest to smallest. For example, the feature intensities of 10 fault propagation paths are 0.8, 0.6, 0.3, 0.1, 0.05, 0.02, 0.01, 0.01, 0.005, and 0.002, respectively. The cumulative distribution of the feature intensity of all fault propagation paths is calculated, and a cumulative feature intensity distribution curve is plotted. The percentage of cumulative feature intensity corresponding to the inflection point of the curve is taken as the selection criterion for the backbone fault propagation path. The top few fault propagation paths whose cumulative feature intensity reaches this percentage are selected as the backbone fault propagation paths. Correspondingly, the cumulative feature intensity of the first two fault propagation paths is 0.8 + 0.6 = 1.4, accounting for 73.8% of the total intensity of 1.897. The inflection point of the curve occurs here, therefore the first two are selected as the backbone fault propagation paths.

[0026] S300 extracts the location information and phase field variable values ​​of all grid cells traversed by the main fault propagation path in the initial fault candidate region set, and combines them to form a fault candidate region. The structure of the fault candidate region includes the boundary coordinates of the region, the average value of the phase field variables of all grid cells in the region, and the area of ​​the region. The fault matching degree of the fault candidate region is obtained by summing the average value of the phase field variables of the fault candidate region, the characteristic intensity of the fault propagation path, and the influence range of the fault propagation path. Based on the fault matching degree of each fault candidate region in the initial fault candidate region set, the fault candidate regions are sorted from high to low according to the fault matching degree. The top few fault candidate regions with the highest fault matching degree are selected for region copying operation to obtain copied fault candidate regions. The copied fault candidate regions are exactly the same as the original fault candidate regions, and the number of copied regions is defined by the user. In this application, the top 2 regions are selected for copying. The more regions copied, the better the noise separation result and the enhancement effect, but the heavier the computational burden. Then, the boundary coordinates of the copied fault candidate region are expanded outward and / or shrunken inward, and the corresponding phase field region is translated a certain distance on the two-dimensional plane so that the boundary of the copied fault candidate region does not contact the boundary of the corresponding fault candidate region. Then, the fault matching degree between the fault candidate region and the adjusted copied fault candidate region is calculated. The region with the highest fault matching degree is selected as the new original fault candidate region, and the new original fault candidate region replaces the fault candidate region with the lowest fault matching degree in the initial set of fault candidate regions. After replacing all the duplicated fault candidate regions, the phase field region corresponding to the fault candidate region with the highest fault matching degree in the final set of replaced fault candidate regions is taken as the enhanced fault feature region, and the remaining phase field regions are background noise regions. The physical field data of the enhanced fault feature region is extracted to obtain the fault feature set.

[0027] S400. Based on the enhanced fault feature set, multi-dimensional feature extraction is performed, and cosine similarity is compared with fault mode features in the database to obtain fault type identification results.

[0028] The S400 performs multi-dimensional feature extraction on the enhanced fault feature set, extracting feature parameters including the center position, area, average intensity, intensity change rate, and shape parameters of the fault feature region. It uses cosine similarity to calculate the similarity between the extracted multi-dimensional feature parameters and the features of each fault mode in the database, and selects the fault mode with the highest similarity as the recognition result.

[0029] The database includes five common fault modes and their corresponding characteristics: oil pump wear, filter blockage, pipeline leakage, cooler scaling, and pressure regulating valve jamming. In the multi-dimensional feature extraction of the enhanced fault feature set, the center position is calculated by summing the abscissas of all grid cells within the region and dividing by the number of grid cells; the ordinates of all grid cells within the region are summed and divided by the number of grid cells to obtain the center ordinate. The area is equal to the number of grid cells in the region multiplied by the area of ​​each grid cell. The average intensity is equal to the average value of the phase field variables of all grid cells within the region. The rate of change of intensity is equal to the maximum value minus the minimum value of the phase field variables of all grid cells within the region, then divided by the average intensity of the region. Shape parameters include aspect ratio and roundness; the aspect ratio is equal to the length of the bounding rectangle of the region divided by its width, and the roundness is equal to four times pi multiplied by the area of ​​the region divided by the square of the region's perimeter.

Claims

1. A method for detecting faults in a gas turbine lubricating oil supply system, characterized in that: The specific steps include the following: S100, based on the fluid network topology of the gas turbine lubricating oil supply system, physical field data is collected by sensors, and then spatial coordinate mapping is performed after time stamp calibration to establish a three-dimensional data structure containing time, space and physical quantities. After being combined with the time stamp, three-dimensional data points are formed. After outlier removal, the data is integrated into a spatiotemporally aligned multi-physics coupling feature dataset. S200. Based on the multi-physics coupling feature dataset and fluid network topology, the finite volume method is used to calculate the velocity field distribution of lubricating oil in the entire gas turbine lubricating oil supply system at each moment. Based on numerical simulation, several virtual tracer particles are released along the flow direction of lubricating oil. Statistical analysis is performed on the flow trajectory and particle parameters of the virtual tracer particles in the lubricating oil system to obtain a dynamic feature set that characterizes the flow trajectory of lubricating oil. S300: Based on mapping each feature point in the dynamic feature set to the same two-dimensional plane, and obtaining the comprehensive feature value of each trajectory, a phase field variable is introduced, and the comprehensive feature value is used as the driving term of the phase field evolution equation. After dividing the two-dimensional plane into several grid cells of equal size, the phase field evolution equation is solved to obtain the phase field distribution at each time. The phase field distribution is used as the terrain surface, and the value of the phase field variable of each grid cell is used as the height value of the terrain surface. Based on the self-organization principle of the river network, the region composed of all grid cells through which the main fault propagation path passes is used as the initial fault candidate region set. After completing the region replication based on the initial fault candidate region set, adaptive enhancement and noise suppression are performed to obtain the enhanced fault feature set and the background noise separation result. S400. Based on the enhanced fault feature set, multi-dimensional feature extraction is performed, and cosine similarity is compared with fault mode features in the database to obtain fault type identification results.

2. The method for detecting faults in a gas turbine lubricating oil supply system according to claim 1, characterized in that: The S100, based on the fluid network topology of the gas turbine lubricating oil supply system, collects physical parameters, including pressure field, temperature field, flow field, vibration field, and oil quality parameters, at the lubricating oil inlet and outlet nodes. The original signals of the acquired physical field data are timestamped. Using the global clock of the gas turbine control system as a reference, the local timestamp of each sampling point in the data collected by different sensors is compared with the timestamp of the global clock to calculate the time difference. Based on the time difference, the timestamp of the sampling point is adjusted to the timestamp of the global clock, and the adjusted sampling points are rearranged in chronological order. Each sensor is assigned a unique spatial coordinate identifier. The measurement value of each sampling point is combined with the corresponding spatial coordinate identifier and the adjusted timestamp to form a three-dimensional data point. The S100 performs outlier removal on the three-dimensional data points using the median filtering method to obtain a spatiotemporally aligned multiphysics coupling feature dataset.

3. The method for detecting faults in a gas turbine lubricating oil supply system according to claim 1, characterized in that: The S200, based on a multi-physics coupled feature dataset and fluid network topology, uses the finite volume method to solve the flow control equations of lubricating oil in the system, and calculates the velocity field distribution of lubricating oil in the entire gas turbine lubricating oil supply system at each moment. Specifically: Each pipeline in the gas turbine lubricating oil supply system is divided into several control volumes of equal length. Based on the divided control volumes, discretized flow control equations are established according to the laws of conservation of mass and momentum. The difference between the mass flow rate flowing into and out of the control volume is calculated according to the law of conservation of mass, and the obtained difference is used as the rate of change of fluid mass within the control volume. The resultant force of all forces acting on the control volume is calculated based on the law of conservation of momentum, and is used as the rate of change of fluid momentum within the control volume. Integrating the continuous differential equations of the resulting difference and resultant force over the control volume yields a discretized algebraic equation. Based on the pressure and flow rates of the physical parameters collected at the lubricating oil inlet and outlet nodes using the multi-physics coupling feature dataset as boundary conditions, the discretized flow control equations are solved using an iterative method after being substituted into the discretized algebraic equations. Based on the flow rate of each control volume obtained from the solution, the velocity field distribution of the gas turbine lubricating oil supply system is obtained by calculating the flow velocity at the center of each control volume.

4. The method for detecting faults in a gas turbine lubricating oil supply system according to claim 3, characterized in that: After obtaining the velocity field distribution and flow velocity value, S200 releases a number of virtual tracer particles in the velocity field starting from the oil tank outlet and along the flow direction of the lubricating oil. The initial position of each virtual tracer particle is evenly distributed on the cross-section of the oil tank outlet. Then, the flow trajectory of each virtual tracer particle in the gas turbine lubricating oil supply system is tracked, and the physical parameters of each virtual tracer particle as it passes through each lubricating oil inlet and outlet node are recorded, specifically: Based on the velocity field distribution at each moment, the velocity of each virtual tracer particle at the current moment is calculated, and the velocity value is used as the flow velocity value at the center of the control volume where the virtual tracer particle is located. Based on the velocity and time step of the virtual tracer particle, the position of the virtual tracer particle at the next moment is calculated. When a virtual tracer particle moves from one control volume to another, the control volume number where the virtual tracer particle is located is updated. When a virtual tracer particle passes through each lubricating oil inlet / outlet node, the corresponding time and physical parameters are updated and recorded. The operation is repeated until all virtual tracer particles reach the return oil manifold and / or exceed the preset maximum tracking time, and then the flow trajectory of each virtual tracer particle in the gas turbine lubricating oil supply system is obtained. Statistical analysis was performed on the flow trajectories of all virtual tracer particles to extract the dynamic features of each flow trajectory. After uniformly normalizing all the extracted dynamic features, a set of dynamic features of lubricating oil flow trajectory composed of dynamic feature points was obtained.

5. The method for detecting faults in a gas turbine lubricating oil supply system according to claim 1, characterized in that: S300 maps each feature point in the normalized dynamic feature set of lubricating oil flow trajectory to the same two-dimensional plane, specifically: The comprehensive feature value of the corresponding lubricating oil flow trajectory is obtained by summing all the normalized dynamic feature values ​​in each lubricating oil flow trajectory. The comprehensive feature value is used as the horizontal axis of the two-dimensional plane, and the time corresponding to the lubricating oil flow trajectory is used as the vertical axis of the two-dimensional plane. Each lubricating oil flow trajectory corresponds to a point on the two-dimensional plane, and the points corresponding to all the lubricating oil flow trajectories together form the feature point distribution on the two-dimensional plane. By introducing phase field variables, a phase field evolution equation consisting of volume free energy and gradient energy terms is constructed. The comprehensive eigenvalues ​​in the normalized dynamic characteristic set of the lubricating oil flow trajectory are used as the driving terms of the phase field evolution equation, specifically: Based on each point on the two-dimensional plane, the average value of the comprehensive characteristic value of all lubricating oil flow trajectories is calculated and used as the original driving term value. After normalizing all the original driving term values, they are added to the right side of the phase field evolution equation as an additional driving term.

6. The method for detecting faults in a gas turbine lubricating oil supply system according to claim 5, characterized in that: The S300 solves the phase field evolution equation to obtain the phase field distribution at each time step, specifically: The two-dimensional plane is divided into several grid cells of equal size. By analyzing the distribution density of feature points, the maximum grid cell side length that can guarantee that each grid cell contains at least one feature point on average is calculated and used as the final grid cell side length. The distribution of phase field variables is initialized by setting all phase field variables of all grid cells to zero. The phase field evolution equation is discretized using the finite difference method to obtain the new value of the phase field variable of each grid cell at each time step. This process is repeated until the preset maximum evolution time and / or the change in phase field distribution is less than the preset convergence threshold, thus obtaining the phase field distribution composed of the phase field variables of each grid cell.

7. The method for detecting faults in a gas turbine lubricating oil supply system according to claim 6, characterized in that: The S300 takes the value of the phase field variable of each grid cell on the two-dimensional plane as the height value of the terrain surface corresponding to the grid cell, takes all grid cells on the terrain surface with height values ​​greater than a preset threshold as the starting point of the feature flow, and calculates the slope between the corresponding grid cell and eight adjacent grid cells based on each starting point of the feature flow. The slope is calculated by dividing the height difference between two grid cells by the distance between the two grid cells. Based on the direction of the steepest slope as the flow direction of the feature flow, the feature flow is moved from the current grid cell to the adjacent grid cell and then repeated on the adjacent grid cells until the feature flow reaches the lowest point of the terrain surface and / or flows out of the boundary of the two-dimensional plane. The network structure of fault propagation paths formed by feature flows on the terrain surface is recorded. The feature flows are used as fault propagation paths. The process includes: assigning a unique number to each fault propagation path; calculating the length of each fault propagation path based on the starting and ending positions of each fault propagation path; using the sum of the intensities of all feature flows that have flowed through each grid cell as the feature intensity; and using the total area of ​​the grid cell where all feature flows finally converge to the corresponding fault propagation path as the influence range of the corresponding fault propagation path. The obtained parameters are then integrated to form the fault propagation path network structure.

8. The method for detecting faults in a gas turbine lubricating oil supply system according to claim 7, characterized in that: The S300 calculates the fractal dimension and power-law exponent of the fault propagation path network structure, specifically as follows: Based on box counting, a two-dimensional plane is divided into several square boxes of the same size, and the number of boxes containing fault propagation paths is counted. The size of the boxes is changed and the operation is repeated to obtain several data pairs of box size and corresponding box number. The absolute value of the slope of the obtained regression line is used as the fractal dimension through double log-linear regression. Based on the influence range and characteristic intensity of the obtained fault propagation path, data pairs of influence range and corresponding characteristic intensity are formed, and the slope of the regression line obtained by double log-linear regression is used as the power law exponent. The fault propagation path with the largest feature intensity and influence range in the fault propagation path network is extracted as the main fault propagation path. The phase field region corresponding to the main fault propagation path and all grid cells passed through the path are used as the initial set of fault candidate regions.

9. The method for detecting faults in a gas turbine lubricating oil supply system according to claim 1, characterized in that: The S300 extracts the position information and phase field variable values ​​of all grid cells traversed by the main fault propagation path in the initial fault candidate region set, and combines them to form a fault candidate region. The fault matching degree of the fault candidate region is obtained by summing the average value of the phase field variable of the fault candidate region, the characteristic intensity of the fault propagation path, and the influence range of the fault propagation path. Based on the fault matching degree of each fault candidate region in the initial fault candidate region set, the fault candidate regions are sorted from high to low according to the fault matching degree. The top few fault candidate regions with the highest fault matching degree are selected for region copying operation to obtain the fault candidate regions to be copied. Then, the boundary coordinates of the copied fault candidate region are expanded outward and / or shrunken inward, and the corresponding phase field region is translated a certain distance on the two-dimensional plane so that the boundary of the copied fault candidate region does not contact the boundary of the corresponding fault candidate region. Then, the fault matching degree between the fault candidate region and the adjusted copied fault candidate region is calculated. The region with the highest fault matching degree is selected as the new original fault candidate region, and the new original fault candidate region replaces the fault candidate region with the lowest fault matching degree in the initial set of fault candidate regions. After replacing all the duplicated fault candidate regions, the phase field region corresponding to the fault candidate region with the highest fault matching degree in the final set of replaced fault candidate regions is taken as the enhanced fault feature region, and the remaining phase field regions are background noise regions. The physical field data of the enhanced fault feature region is extracted to obtain the fault feature set.

10. A method for detecting faults in a gas turbine lubricating oil supply system according to claim 1, characterized in that: The S400 performs multi-dimensional feature extraction on the enhanced fault feature set, extracting feature parameters including the center position, area, average intensity, intensity change rate, and shape parameters of the fault feature region. The cosine similarity is used to calculate the similarity between the extracted multi-dimensional feature parameters and the features of each fault mode in the database, and the fault mode with the highest similarity is selected as the recognition result.