Power generation equipment state fault diagnosis method and system based on artificial intelligence

Abstract

The invention discloses a power generation equipment state fault diagnosis method and system based on artificial intelligence, and the method comprises the steps: actively injecting a mechanical excitation signal of a preset frequency spectrum into a key part according to a physical topological structure of power generation equipment, and carrying out the fusion to generate a time-space-frequency three-dimensional data volume; inputting the three-dimensional data volume into a physical embedded variational auto-encoder, and outputting an equipment state pure feature tensor; inputting the pure feature tensor into a graph space-time causal reasoning network to generate a fault propagation causal graph with probability weight; performing multi-agent diagnosis on the fault propagation causal atlas, and outputting a fault diagnosis report which has a credibility interval and comprises fault positioning and root cause analysis; and mapping the fault diagnosis report to the digital twin of the equipment in real time, and outputting a self-adaptive maintenance strategy sequence which minimizes the expected value of the whole life cycle operation and maintenance cost. According to the embodiment of the invention, the accuracy and anti-interference capability of fault diagnosis can be improved, and the operation and maintenance cost can be effectively reduced.

Description

Technical Field

[0001] The present invention belongs to the technical field of fault diagnosis, and in particular to a method and system for diagnosing power generation equipment status faults based on artificial intelligence. Background Art

[0002] As power generation equipment evolves toward higher parameters and larger capacities, traditional fault diagnosis methods face challenges such as delayed response, high false alarm rates, and difficulty identifying rare faults. Existing technologies primarily rely on single-sensor data or offline analysis, making it difficult to capture the multi-physics coupling characteristics of complex equipment operating conditions. Furthermore, environmental noise interference can lead to insufficient reliability in diagnostic results. Furthermore, root cause analysis often relies on expert experience, making it difficult to quantify the impact of different maintenance strategies on the equipment's lifecycle costs. Summary of the Invention

[0003] The purpose of the present invention is to provide a method and system for diagnosing power generation equipment status faults based on artificial intelligence to address the deficiencies in the prior art, improve the accuracy and anti-interference ability of fault diagnosis, and effectively reduce operation and maintenance costs.

[0004] One embodiment of the present application provides a method for diagnosing power generation equipment status faults based on artificial intelligence, the method comprising: Based on the physical topology of the power generation equipment, mechanical excitation signals with a preset spectrum are actively injected into key components. The transient responses of vibration sensors, infrared thermal imagers, and current transformers under excitation are simultaneously collected. Combined with environmental parameters and equipment operation logs, a fusion of time-space-frequency three-dimensional data is generated. Inputting the three-dimensional data volume into a physical embedded variational autoencoder, constructing a physical constraint loss function using the device thermal-mechanical coupling equation, separating the intrinsic physical feature vectors representing the health status of the component from the redundant feature vectors interfered by environmental noise, and outputting a pure feature tensor of the device state; The clean feature tensor is input into the graph spatiotemporal causal reasoning network, and a causal prior graph is constructed based on the historical failure cases of the equipment throughout its life cycle. The adversarial sample generation mechanism is used to enhance the causal association mining of rare failures and generate a fault propagation causal graph with probability weights. Perform multi-agent diagnosis on the fault propagation causal graph, deploy competitive agents to simulate the fault discrimination logic of the equipment designer, operator, and component manufacturer, respectively, fuse the dispute results of the three parties through a Nash equilibrium strategy, and output a fault diagnosis report with credibility intervals, including fault location and root cause analysis; The fault diagnosis report is mapped to the equipment digital twin in real time. Combining reinforcement learning with Monte Carlo fault tree simulation, the equipment degradation trajectory under different maintenance strategies is previewed in virtual space, and an adaptive maintenance strategy sequence that minimizes the expected value of the full life cycle operation and maintenance cost is output.

[0005] Optionally, based on the physical topology of the power generation equipment, a mechanical excitation signal with a preset frequency spectrum is actively injected into key components, and the transient responses of the vibration sensor, infrared thermal imager, and current transformer under the excitation are synchronously collected. The environmental parameters and equipment operation logs are combined to generate a three-dimensional time-space-frequency data volume, including: Based on the bearing stiffness distribution matrix in the physical topology of the power generation equipment, the resonance sensitive areas are identified and the coordinate mapping table of key components is generated; Based on the coordinate mapping table of key components, an amplitude-modulated swept-frequency mechanical excitation signal is injected into the specified position to generate a synchronous trigger pulse sequence; Using a synchronized trigger pulse sequence, the time domain waveform of the vibration sensor, the temperature field matrix of the infrared thermal imager, and the harmonic distortion spectrum of the current transformer are collected in parallel to obtain time-aligned multi-source response data packets. Combining environmental parameters with the noise floor in the device operation log, an adaptive filter is used to separate the environmental interference component from the multi-source response data packet and extract the device's intrinsic transient response set. The device's intrinsic transient response set is reorganized into tensors according to the time axis, spatial coordinates, and frequency dimensions, and a time-space-frequency three-dimensional data volume is output.

[0006] Optionally, the three-dimensional data volume is input into a physical embedded variational autoencoder, a physical constraint loss function is constructed using the device thermal-mechanical coupling equation, intrinsic physical feature vectors representing the health status of the component and redundant feature vectors interfered by environmental noise are separated, and a pure feature tensor of the device state is output, including: Based on the device's thermal-mechanical coupling equations, a physical regularization module for the variational autoencoder is constructed; The three-dimensional data volume is input into the variational autoencoder network, and the initial intrinsic feature vector and redundant feature vector are generated through the feature decoupling layer; Use the physical regularization module to perform finite element simulation on the initial eigenvectors to generate the theoretical thermal response field; Calculate the residual matrix between the theoretical thermal response field and the measured infrared temperature field, and construct a loss function for physical constraints by combining KL divergence; The variational autoencoder parameters are reversely optimized through the loss function, and the pure feature tensor of the device state is output.

[0007] Optionally, the clean feature tensor is input into a graph spatiotemporal causal reasoning network, a causal prior graph is constructed based on historical failure cases throughout the equipment life cycle, causal association mining of rare failures is enhanced through an adversarial sample generation mechanism, and a fault propagation causal graph with probability weights is generated, including: The device state graph node network is constructed based on the spatial topological relationship of the pure feature tensor to obtain a spatiotemporal topological graph structure with timestamps. Extract causal rules based on the equipment's historical lifecycle failure case database and generate a constraint matrix for the causal prior graph; The spatiotemporal topological graph structure and the constraint matrix of the causal prior graph are input into the spatiotemporal causal reasoning network, features are extracted through gated temporal convolution, and the spatiotemporal enhanced feature tensor is output; Generate adversarial examples corresponding to specific rare fault directions based on spatiotemporal enhancement feature tensors, and enhance causal correlation features through gradient sign method; The enhanced causal correlation features are used to train a Bayesian graph network, so that the trained Bayesian graph network can output a fault propagation causal graph with probability weights.

[0008] Optionally, the fault propagation causal graph is subjected to multi-agent diagnosis, and competitive agents are deployed to simulate the fault discrimination logic of the equipment designer, the operator, and the component manufacturer, respectively. The dispute results of the three parties are integrated through a Nash equilibrium strategy, and a fault diagnosis report with a credibility interval, including fault location and root cause analysis, is output, including: Load equipment design specifications, operation and maintenance records, and material databases based on node attributes of the fault propagation causal graph to generate a tripartite knowledge vector group; The three-party knowledge vector group is input into the competitive agent group, where the design agent outputs a failure probability vector based on stress simulation, the operation and maintenance agent outputs a failure probability vector based on maintenance history, and the manufacturing agent outputs a failure probability vector based on fatigue model. Aggregate the nodes whose fault probability vector detection differences between the three parties are greater than a preset threshold to generate a set of dispute points; Construct a three-party payoff matrix based on the dispute focus set and solve the Nash equilibrium strategy vector; Bootstrap sampling is performed on the Nash equilibrium strategy vector to calculate the preset percentage confidence interval of the fault location; Based on the confidence interval and causal graph, the root cause path with a probability weight greater than the preset weight is traced back, and a diagnostic report with a credibility interval is output.

[0009] Optionally, the fault diagnosis report is mapped to the equipment digital twin in real time, and reinforcement learning and Monte Carlo fault tree simulation are combined to preview the equipment degradation trajectory under different maintenance strategies in a virtual space, and output an adaptive maintenance strategy sequence that minimizes the expected value of the full life cycle operation and maintenance cost, including: Based on the root cause path sequence in the fault diagnosis report, the equipment degradation model is initialized in the digital twin to obtain the component degradation rate equation; Based on the component degradation rate equation, a three-dimensional discrete action space is defined to generate the maintenance strategy search domain; In the maintenance strategy search domain, a number of paths are explored using a proximal strategy optimization algorithm, and a preview trajectory and cost dataset are output; Perform Monte Carlo fault tree analysis on the cost data set to calculate the expected value of the full life cycle operation and maintenance cost; The optimal strategy is selected according to the expected cost value and the adaptive maintenance strategy sequence is output.

[0010] Another embodiment of the present application provides a power generation equipment status fault diagnosis system based on artificial intelligence, the system comprising: The acquisition module is used to actively inject mechanical excitation signals with a preset frequency spectrum into key components based on the physical topology of the power generation equipment. It simultaneously collects the transient responses of vibration sensors, infrared thermal imagers, and current transformers under excitation, and combines environmental parameters with equipment operation logs to generate a three-dimensional time-space-frequency data volume. a separation module, configured to input the three-dimensional data volume into a physical embedded variational autoencoder, construct a physical constraint loss function using the device thermal-mechanical coupling equation, separate the intrinsic physical feature vectors representing the health status of the component from the redundant feature vectors interfered by environmental noise, and output a pure feature tensor of the device state; A construction module is used to input the clean feature tensor into a graph spatiotemporal causal reasoning network, construct a causal prior graph based on historical failure cases throughout the equipment life cycle, enhance causal association mining of rare failures through an adversarial sample generation mechanism, and generate a fault propagation causal graph with probability weights; A diagnostic module is used to perform multi-agent diagnosis on the fault propagation causal graph, deploy competitive agents to simulate the fault discrimination logic of the equipment designer, operator, and component manufacturer, fuse the dispute results of the three parties through a Nash equilibrium strategy, and output a fault diagnosis report with a credibility interval, including fault location and root cause analysis; The output module is used to map the fault diagnosis report to the equipment digital twin in real time, combine reinforcement learning with Monte Carlo fault tree simulation, preview the equipment degradation trajectory under different maintenance strategies in virtual space, and output an adaptive maintenance strategy sequence that minimizes the expected value of the operation and maintenance cost throughout the life cycle.

[0011] Yet another embodiment of the present application provides a storage medium, wherein the storage medium stores a computer program, wherein the computer program is configured to execute any of the above methods when run.

[0012] Yet another embodiment of the present application provides an electronic device, comprising a memory and a processor, wherein the memory stores a computer program, and the processor is configured to run the computer program to execute any of the above methods.

[0013] Compared with the existing technology, the present invention provides an artificial intelligence-based method for diagnosing power generation equipment status faults. According to the physical topology of the power generation equipment, a mechanical excitation signal with a preset spectrum is actively injected into key components to fuse and generate a time-space-frequency three-dimensional data body; the three-dimensional data body is input into a physical embedded variational autoencoder to output a pure feature tensor of the equipment state; the pure feature tensor is input into a graph space-time causal reasoning network to generate a fault propagation causal graph with probability weights; multi-agent diagnosis is performed on the fault propagation causal graph to output a fault diagnosis report with credibility intervals, including fault location and root cause analysis; the fault diagnosis report is mapped to the equipment digital twin in real time, and an adaptive maintenance strategy sequence is output that minimizes the expected value of the operation and maintenance cost throughout the entire life cycle, thereby improving the accuracy and anti-interference ability of fault diagnosis and effectively reducing the operation and maintenance cost. BRIEF DESCRIPTION OF THE DRAWINGS

[0014] Figure 1 A hardware structure block diagram of a computer terminal for a power generation equipment status fault diagnosis method based on artificial intelligence provided by an embodiment of the present invention; Figure 2 A schematic diagram of a flow chart of a method for diagnosing power generation equipment status faults based on artificial intelligence provided by an embodiment of the present invention; Figure 3 A schematic structural diagram of an artificial intelligence-based power generation equipment status fault diagnosis system provided in an embodiment of the present invention. DETAILED DESCRIPTION

[0015] The embodiments described below with reference to the accompanying drawings are exemplary and are only used to explain the present invention, and are not to be construed as limiting the present invention.

[0016] The embodiment of the present invention first provides a method for diagnosing power generation equipment status faults based on artificial intelligence. The method can be applied to electronic devices such as computer terminals, specifically ordinary computers.

[0017] The following describes it in detail by taking running on a computer terminal as an example. Figure 1 The hardware structure block diagram of a computer terminal for a power generation equipment status fault diagnosis method based on artificial intelligence provided by an embodiment of the present invention. Figure 1 As shown, the computer device includes a processor, a memory, and a network interface connected via a system bus, wherein the memory may include a non-volatile storage medium and an internal memory.

[0018] The non-volatile storage medium can store an operating system and a computer program. The computer program includes program instructions, which, when executed, can cause a processor to execute any one of the power generation equipment state fault diagnosis methods based on artificial intelligence.

[0019] The processor is used to provide computing and control capabilities and support the operation of the entire computer equipment.

[0020] The internal memory provides an environment for the operation of the computer program in the non-volatile storage medium. When the computer program is executed by the processor, the processor can execute any power generation equipment status fault diagnosis method based on artificial intelligence.

[0021] The network interface is used for network communication, such as sending assigned tasks, etc. Those skilled in the art will understand that Figure 1 The structure shown in the figure is only a block diagram of a part of the structure related to the solution of the present application, and does not constitute a limitation on the computer device to which the solution of the present application is applied. The specific computer device may include more or fewer components than shown in the figure, or combine certain components, or have a different component arrangement.

[0022] It should be understood that the processor may be a central processing unit (CPU), other general-purpose processors, digital signal processors (DSP), application-specific integrated circuits (ASIC), field-programmable gate arrays (FPGA), other programmable logic devices, discrete gate or transistor logic devices, discrete hardware components, etc. The general-purpose processor may be a microprocessor or any conventional processor, etc.

[0023] See also Figure 2 , an embodiment of the present invention provides a method for diagnosing power generation equipment status faults based on artificial intelligence, which may include the following steps: S201: Based on the physical topology of the power generation equipment, a mechanical excitation signal with a preset frequency spectrum is actively injected into key components. The transient responses of the vibration sensor, infrared thermal imager, and current transformer under the excitation are simultaneously collected. The transient responses are combined with environmental parameters and equipment operation logs to generate a three-dimensional time-space-frequency data volume. Specifically, based on the bearing stiffness distribution matrix in the physical topology of the power generation equipment, the resonance sensitive areas can be identified and a coordinate mapping table of key components can be generated; ‌Bearing stiffness distribution matrix analysis‌ Analyze the bearing stiffness data in the equipment design drawings and construct a two-dimensional stiffness distribution matrix (i.e., a stiffness distribution matrix). Stiffness values ​​are expressed in Newtons per millimeter (N / mm). For example, the stiffness of a generator rotor bearing at an axial position of 1200 mm drops sharply from 8000 N / mm to 5000 N / mm, marking this point as the stiffness mutation point.

[0024] Combined with modal analysis and experimental verification, the system tested the excitation frequency response using a hammer method. When the excitation frequency reached 85 Hz, the amplitude in that area exceeded the warning value of 0.15 mm, confirming it as a high-risk resonance zone. The stiffness gradient of adjacent components (such as the gearbox) was also tested. If the stiffness difference between adjacent areas exceeded 2000 Newtons per millimeter, it was simultaneously marked as a sensitive area.

[0025] Generate a coordinate mapping table: Use the center of the equipment base as the origin of the three-dimensional coordinate system (0,0,0), with the axial direction as the X-axis (in millimeters), the radial direction as the Y-axis, and the circumferential direction as the Z-axis. Categorize the coordinates of the resonance points by risk level. For example, bearing A's coordinates are labeled (X=1200, Y=50, Z=0, risk level: high), and gearbox B's coordinates are labeled (X=800, Y=200, Z=30, risk level: medium). The output table includes component name, coordinate values, maximum amplitude, and recommended monitoring frequency.

[0026] Based on the coordinate mapping table of key components, an amplitude-modulated swept-frequency mechanical excitation signal is injected into the specified position to generate a synchronous trigger pulse sequence; ‌Excitation signal parameter design‌ Frequency Sweep Range Setting: Based on the frequency characteristics of the resonant sensitive zone, a linearly increasing frequency sweep from 50 Hz to 500 Hz (at a sweep rate of 5 Hz per second) is set. In the high-frequency range (above 300 Hz), an amplitude attenuation mechanism is used, with an initial amplitude of 5 Newtons (N) and a 20% decrease for every 100 Hz increase to prevent equipment overload damage. For example, at coordinate point A on the bearing, an excitation with an initial frequency of 50 Hz and an amplitude of 5 Newtons is injected, with the frequency increasing by 1 Hz every 0.2 seconds for 90 seconds.

[0027] Modulation logic optimization: Superimpose second harmonic excitation (twice the fundamental frequency) in the stiffness mutation area. For example, when the fundamental frequency is 85 Hz, simultaneously inject a 170 Hz component (amplitude accounts for 30%) to enhance the excitation effect on microcracks.

[0028] ‌Synchronous trigger pulse generation‌ Pulse timing arrangement: Using the excitation signal starting point as time reference T0, a pulse sequence with 10-ms intervals is generated. T0 triggers the vibration sensor, T0+5 ms triggers the infrared thermal imager, and T0+8 ms triggers the current transformer, ensuring that the time alignment deviation of multi-source data is less than 1 ms.

[0029] Fault-tolerant retransmission mechanism: If a sensor does not respond, it will resend the pulse three times within 50 milliseconds; if there is no response for three consecutive times, it will automatically switch to the backup sensor node and record the alarm log.

[0030] Using a synchronized trigger pulse sequence, the time domain waveform of the vibration sensor, the temperature field matrix of the infrared thermal imager, and the harmonic distortion spectrum of the current transformer are collected in parallel to obtain time-aligned multi-source response data packets. ‌Multi-source data synchronous acquisition‌ Vibration sensor: Collects time-domain waveforms of axial and radial acceleration at a sampling rate of 10 kHz (10,000 data points per second) and a range of ±15 g. Focus is placed on recording the amplitude envelope within the resonant frequency band (e.g., 85 Hz ± 5 Hz), extracting the peak-to-peak value and root mean square (RMS) value.

[0031] Thermal imager: Captures a 256×192 pixel temperature field matrix (spatial resolution 2 mm per pixel) with a temperature accuracy of ±0.5°C. A frame is captured every 5 milliseconds, recording the hotspot coordinates (e.g., temperature 89.5°C at coordinate (120,95)) and the temperature rise rate (e.g., 0.8°C per second).

[0032] Current transformer: Collects the current harmonic distortion spectrum from 0 to 2000 Hz, decomposing it into 50 harmonic components. Records total harmonic distortion (THD) and characteristic harmonic amplitudes (e.g., 8.3 amps for the 5th harmonic). Sampling rate: 5 kHz.

[0033] Packet Encapsulation and Alignment Timestamp Binding: Based on the pulse sequence, each data source is marked with an absolute time (accurate to the millisecond). For example, the timestamp of data packet number P2023-085 is "2023-05-18 14:30:25.123," which includes vibration data (axial acceleration 0.35 times the acceleration of gravity), temperature field (hotspot 89.5 degrees Celsius), and current harmonics (5th harmonic 8.3 amps).

[0034] Data integrity check: The cyclic redundancy check (CRC) is used to verify that the transmission is lossless. If the data packet loss rate is greater than 1%, the automatic re-collection process is triggered.

[0035] Combining environmental parameters with the noise floor in the device operation log, an adaptive filter is used to separate the environmental interference component from the multi-source response data packet and extract the device's intrinsic transient response set. Environmental noise quantitative modeling and base library construction Noise source parameterization calibration: Based on the typical operating environment of power generation equipment, the three core noise sources are quantified: Wind turbine aerodynamic noise: Measured using A-weighting (a noise evaluation index that simulates the human ear's auditory characteristics), the typical value is 25 decibels. For example, a wind turbine at rated speed measured 3 meters from the nacelle has a background noise level of 24.8 decibels.

[0036] Grid fluctuation interference: Continuously monitor the voltage deviation range and set a threshold of ±0.5%. For example, if the measured voltage fluctuation at a power plant's grid connection point is +0.3% to -0.4%, it should be recorded as the effective noise floor.

[0037] Impact of ambient temperature drift: Based on ±2°C per hour, dynamic correction is performed in conjunction with weather station data. For example, if the ambient temperature at a hydropower station rises from 28°C to 30°C between 8:00 and 9:00 a.m., the temperature field correction parameters must be adjusted accordingly.

[0038] Noise floor library dynamic update Extract historical operating condition data from equipment operation logs and establish a noise-environment mapping relationship: Vibration Noise Floor: This indicates the quantitative relationship between ambient temperature and vibration noise. For example, if you log a value like "Gearbox vibration noise floor 0.02 times the acceleration of gravity (g) at 2:00 PM on May 20, 2025, ambient temperature 30°C," this data will serve as the filtering benchmark for the same temperature conditions on that day.

[0039] Electrical interference floor: This records the harmonic characteristics of power grid anomalies. For example, "During a thunderstorm on June 1, 2025, power grid voltage flicker caused the total harmonic distortion (THD) of the current to temporarily rise to 2.0%." Such events should be marked as special noise samples.

[0040] Multi-source data adaptive filtering and interference separation Least Mean Square (LMS) adaptive filtering algorithm implementation: Vibration signal processing: The system inputs raw vibration data (e.g., 0.35 times the acceleration of gravity at a bearing measurement point) and a noise floor (0.02 times the acceleration of gravity). The LMS algorithm (an adaptive filtering method that iteratively minimizes the power of the error signal) updates the filter coefficients in real time, outputting the intrinsic vibration component of 0.33 times the acceleration of gravity. After filtering, the signal waveform integrity must be verified, for example, to eliminate periodic interference caused by the wind blade pass frequency (1.2 Hz).

[0041] Temperature field correction: Based on the infrared thermal imager's raw data (e.g., an apparent generator winding temperature of 89.5°C), combined with ambient thermal radiation parameters (direct sunlight causing a 3°C false increase), spatial filtering is used to eliminate non-device hot spots. For example, after correction for an abnormally high temperature in a cooling fan area, the actual component temperature is determined to be 86.5°C.

[0042] Current harmonic purification: This system uses harmonic data collected by current transformers (e.g., total harmonic distortion (THD) of 4.5%) to identify interference components (approximately 1.3%) introduced by grid fluctuations and outputs the device's true harmonic characteristics, achieving a THD of 3.2%. The system focuses on monitoring the amplitudes of the fifth (250 Hz) and seventh (350 Hz) harmonics. For example, the fifth harmonic can be corrected from 8.3 amps to 7.8 amps.

[0043] Signal quality hierarchical control The signal-to-noise ratio (SNR) threshold is set at 40 decibels. For example, a filtered vibration signal with an SNR of 42 decibels (corresponding to a noise power of 0.005 times the acceleration of gravity) is considered acceptable. If the SNR is ≤ 40 decibels (e.g., at a remote measurement point with an SNR of 38 decibels due to transmission loss), wavelet threshold secondary filtering (a signal denoising method based on time-frequency analysis) is triggered until the signal meets the standard.

[0044] Data integrity is verified using a cyclic redundancy check (CRC), a coding technique for detecting data transmission errors. Re-collection is initiated when the packet loss rate exceeds 1%. For example, if 95 out of 9,000 time slices are lost during a monitoring session, the system automatically re-sends the acquisition command.

[0045] Intrinsic transient response set generation and verification Structured data output specifications Time dimension: Absolute timestamps are marked in 10-millisecond intervals, such as "2025-06-17 14:30:25.123".

[0046] Spatial dimension: Positioning is based on the topological coordinates of the equipment. For example, the front bearing position of the generator is marked as (X=1200 mm, Y=50 mm, Z=0 mm).

[0047] Physical parameters: Vibration: Preserve axial / radial acceleration time domain waveform (sampling rate 10 kHz) and FFT spectrum (0-2000 Hz, resolution 1 Hz).

[0048] Temperature: Records a 256×192 pixel temperature matrix, where the hotspot coordinate (120,95) corresponds to 86.5 degrees Celsius.

[0049] Current: Stores the 50th harmonic amplitude and THD value, such as 5th harmonic 7.8 amps and THD 3.2%.

[0050] The device's intrinsic transient response set is reorganized into tensors according to the time axis, spatial coordinates, and frequency dimensions, and a time-space-frequency three-dimensional data volume is output.

[0051] ‌Three-dimensional data volume structure construction‌ Time axis: Sliced ​​at 10 millisecond intervals, covering the 0-90 second excitation period, for a total of 9000 time slices; Spatial axis: A grid coordinate system is established based on the 256×192 pixels of the infrared thermal imager, with each pixel mapping the actual position (e.g., pixel (100,50) corresponds to 800 mm axially and 100 mm radially).

[0052] Frequency axis: The vibration signal is decomposed into 512 frequency bands (with a resolution of approximately 1 Hz) using the Fast Fourier Transform (FFT). The current harmonics retain the 50th harmonic component (such as the 50 Hz fundamental wave and the 2500 Hz inner harmonic).

[0053] ‌Tensor Filling Rules and Output‌ Multi-source data fusion rules: vibration spectrum is mapped to physical coordinate points (for example, the coordinate points of bearing A are filled with its 0-2000 Hz spectrum); temperature data is filled into the spatial grid according to pixel position; current harmonics are associated with the coordinates of electrical connection components (for example, the coordinate points of the generator junction box are filled with harmonic data).

[0054] Output specification: Data format: Hierarchical Data Format (HDF5); Dimension Label: Time dimension: relative time (milliseconds); space dimension: three-dimensional coordinates (mm); frequency dimension: frequency value (Hertz).

[0055] Typical data volume size: 9000 (time slices) × 49152 (spatial points) × 562 (frequencies + harmonics); data volume: approximately 25 gigabytes (GB).

[0056] First, by analyzing the physical connection relationship of the power generation equipment, the location of the core components that need to be monitored is determined, and a mechanical excitation signal within a specific frequency range is actively applied. During the period of the excitation signal, the dynamic changes of vibration, temperature, and current signals are synchronously recorded. At the same time, combined with parameters such as temperature and humidity of the equipment's environment and historical operation records, all data are integrated into structured data blocks based on three dimensions: time, spatial position, and frequency components. Through active excitation and synchronous acquisition of multi-source data, the problems of weak signals and high interference in traditional passive monitoring are overcome. The construction of a three-dimensional data body provides a complete information foundation for subsequent analysis, including the spatiotemporal dynamic characteristics of the equipment, significantly improving the comprehensiveness and accuracy of state perception.

[0057] S202: Input the three-dimensional data volume into a physical embedded variational autoencoder, construct a physical constraint loss function using the device thermal-mechanical coupling equation, separate the intrinsic physical feature vectors representing the health status of the component and the redundant feature vectors interfered by environmental noise, and output a pure feature tensor of the device state; Specifically, a physical regularization module of the variational autoencoder can be constructed based on the device's thermal-mechanical coupling equation; Thermal-mechanical physics mechanism modeling and module integration For core components of power generation equipment (such as steam turbine rotors and generator stator windings), a computable set of physical equations is established based on the thermal expansion characteristics and mechanical deformation laws of the materials: The heat conduction equation quantifies temperature diffusion behavior. For example, the thermal conductivity of a gas turbine blade material is set at 42 watts per meter per degree Celsius (a unit of heat transfer efficiency). When the ambient temperature rises by 2 degrees Celsius per hour, the blade surface temperature rise rate is measured to be 0.8 degrees Celsius per second.

[0058] The stress-strain relationship is based on Hooke's law (the basic law of elastic deformation), which defines the material's elastic modulus as 206 GPa (gigaPascal). When the centrifugal force on the rotor increases by 100 kilonewtons, the deformation increases by 0.12 mm.

[0059] The continuous equations are discretized into a set of finite element node equations. For example, the generator stator is divided into 12,000 tetrahedral elements. Each node is associated with three sets of physical quantities: temperature, displacement, and stress, forming a node equation matrix.

[0060] Physical regularization module is deeply embedded Structural Design: A physical computation layer is added to the encoder end of a variational autoencoder (a deep learning model that integrates data generation and feature extraction). This layer contains 500,000 adjustable parameters and is used to map the features extracted by the neural network to the inputs of the physical equations. For example, when the encoder outputs "rotor radial thermal deformation coefficient 0.15," the physical layer automatically converts it into the actual deformation of 0.08 mm.

[0061] Dynamic Parameter Loading: This feature establishes a real-time retrieval mechanism for material properties. If the input feature contains the "Ni-Base Alloy 718" tag, its linear expansion coefficient of 13 million parts per degree Celsius and yield strength of 1100 MPa (million Pascals) are automatically loaded, and the heat conduction equation boundary conditions are injected.

[0062] ‌Loadcase Adaptive Boundary Condition Configuration‌ Dynamically adjust simulation constraints based on device operation logs: When the turbine unit detects the current water head height of 85 meters, it automatically sets the runner water pressure limit to 8.3 MPa; When the speed sensor feedback value reaches 1500 rpm, the centrifugal force load coefficient is corrected to 1.25 times the reference value to ensure that the physical model strictly matches the actual operating status.

[0063] The three-dimensional data volume is input into the variational autoencoder network, and the initial intrinsic feature vector and redundant feature vector are generated through the feature decoupling layer; ‌3D data volume standardization‌ Time Dimension Alignment: All sensor data is resampled at 10 millisecond intervals. Vibration waveforms are sampled at a rate of 10 kHz (10,000 samples per second), infrared temperature fields are refreshed at a rate of 1 Hz (once per second), and current harmonic analysis is performed at the 50th harmonic (fundamental at 50 Hz). Timestamps are strictly synchronized using an interpolation algorithm.

[0064] Spatial Topology Coding: This tool creates a 3D grid coordinate system based on the mechanical structure of the equipment. For example, the generator's front bearing seat is defined as the origin (X=0 mm, Y=0 mm, Z=0 mm), and the coordinates of the adjacent gearbox are (X=400 mm, Y=150 mm, Z=30 mm). The grid spacing is accurate to the millimeter level.

[0065] Spectrum Normalization: The vibration frequency range of 0-2000 Hz is divided into 400 equal bands. The amplitude of each band is divided by 15 times the full-scale gravity acceleration (gravity acceleration unit: g). The temperature value is linearly mapped to the range of 0-1 from 20-150 degrees Celsius.

[0066] ‌Feature Decoupling Layer Operation Mechanism‌ ‌Five-layer convolutional feature extraction network‌: The first layer of 3×3×3 convolution kernels (length × width × height) captures local features. For example, the 85 Hz component in the bearing vibration signal is enhanced to a feature map brightness value of 0.87 (range 0-1); The third layer of dilated convolution identifies large-scale correlations, such as the correlation coefficient between winding temperature rise and cooling wind speed reaching 0.92.

[0067] ‌Dual-channel separation design‌: The intrinsic channel outputs a 128-dimensional vector representing the inherent state of the equipment. Typical values ​​include: [bearing wear index 0.15, insulation aging rate 0.83, shaft misalignment 0.07]; The redundant channel outputs a 64-dimensional vector to record environmental interference. Typical values ​​are: [grid harmonic residual 0.21, ambient temperature drift error 0.04, and solar radiation noise 0.33].

[0068] Separation effect verification Use Fast Fourier Transform (signal spectrum analysis method) to verify feature purity: The retention rate of the 85 Hz component of the equipment characteristic frequency in the intrinsic characteristics is greater than 95%, while the 1.2 Hz component of the fan blade pass frequency is attenuated to less than 3%; The signal-to-noise ratio (signal-to-noise power ratio) of redundant features is strictly controlled below 30 decibels to ensure that interference components are effectively removed.

[0069] Use the physical regularization module to perform finite element simulation on the initial eigenvectors to generate the theoretical thermal response field; Engineering Conversion of Eigenvectors to Physical Parameters ‌Normalized value restoration‌: "Bearing wear index 0.15" corresponds to an actual expansion of radial clearance of 0.08 mm; "winding temperature rise slope 0.92" is converted into an actual heating rate of 1.8 degrees Celsius per second; "shaft alignment deviation 0.07" is mapped to an axis offset angle of 0.15 milliradians (one thousandth of a radian).

[0070] Intelligent material library call: When the rotor material is identified as 30 chromium-molybdenum-vanadium, its creep limit temperature of 625 degrees Celsius and thermal expansion coefficient of 12 millionths per degree Celsius are automatically loaded.

[0071] ‌Multiphysics Coupling Solution Process‌ Temperature Field Calculation: Solve the unsteady heat conduction equation and output a three-dimensional temperature distribution matrix. A typical example: a simulated temperature of 857°C (material temperature limit 900°C) at the leading edge of a first-stage gas turbine rotor blade, with a temperature gradient of 120°C per millimeter.

[0072] Stress Field Derivation: Thermal stresses are calculated based on temperature field results, and mechanical centrifugal force is superimposed to generate a Mises stress (composite stress index) cloud map. The peak stress at the rotor disc groove is 583 MPa, approaching the material yield strength of 600 MPa.

[0073] Data reshaping: The finite element mesh data is converted into a 256×192 infrared pixel matrix using a bicubic interpolation algorithm, which is consistent with the dimensions of the measured thermal imager data.

[0074] Calculate the residual matrix between the theoretical thermal response field and the measured infrared temperature field, and construct a loss function for physical constraints by combining KL divergence; ‌Residual Diagnosis and Anomaly Location‌ Pixel-by-pixel comparison: The simulated temperature at coordinate (205,178) on the tube wall of a boiler superheater is 586 degrees Celsius, while the infrared measurement is 612 degrees Celsius, with an absolute residual value of 26 degrees Celsius.

[0075] Spatial cluster analysis: Areas with 15 consecutive pixels with residuals greater than 20 degrees Celsius (approximately 48 square millimeters) are marked as "physical law conflict areas," indicating potential material defects or carbon deposit failures.

[0076] Physical Constraint Loss Function Construction Probabilistic residuals: Statistical residual distribution histogram, for example, residuals between 10 and 20 degrees Celsius account for 35%, and residuals above 30 degrees Celsius account for 5%.

[0077] KL Divergence (a measure of the difference in probability distribution): Calculate the difference between the ideal residual distribution and the actual distribution, assuming a zero-mean Gaussian distribution with a standard deviation of 5 degrees Celsius. A typical KL divergence is 0.45 (target value < 0.1).

[0078] Multi-objective weighted fusion: Physical constraint loss = mean absolute residual × 0.6 + KL divergence × 0.4.

[0079] Example: Mean residual 18.3°C → 18.3×0.6 = 10.98; KL divergence 0.45 → 0.45×0.4 = 0.18; Total loss 11.16.

[0080] The variational autoencoder parameters are reversely optimized through the loss function, and the pure feature tensor of the device state is output.

[0081] Adaptive Optimization Mechanism Dynamic learning rate control: Initial value is 0.01 (neural network weight adjustment step size). If the loss value decreases by less than 1% for 5 consecutive iterations, the learning rate is reduced to 80% of the original value.

[0082] Regularization weight adjustment: When the loss value is greater than 10.0, the physical constraint weight is increased to 0.7, and when the loss value is less than 2.0, it is reduced to 0.3 to balance data fitting and compliance with physical laws.

[0083] ‌Clean Eigentensor Quality Criteria‌ Physical consistency verification: Vibration characteristic frequency prediction deviation < ±0.3% (e.g., the 85 Hz component error limit is ±0.25 Hz); temperature field distribution spatial correlation coefficient > 0.98 (pixel-level similarity index).

[0084] Data structure specification: The output tensor size is 128×256×256 (feature channels × spatial height × spatial width). The feature values ​​are located according to the device coordinates: [Coordinate X=0, Y=0, Z=0]: Bearing wear factor 0.83, centering deviation 0.12, lubrication status 0.05; [Coordinate X=400, Y=150, Z=30]: Insulation degradation 0.67, contact resistance 0.21, heat dissipation efficiency 0.09.

[0085] Fuse protection mechanism: If the loss value does not decrease after 20 consecutive iterations, the "Physical Constraint Failure" alarm (code F3001) is triggered, and the non-converged features are saved for fault tracing analysis.

[0086] A deep learning model incorporating the physical principles of the equipment is used to process 3D data. The model establishes constraints based on physical laws such as heat conduction equations and mechanical equilibrium equations, enforcing physical plausibility during feature extraction. This process automatically distinguishes between key features reflecting the equipment's true state and irrelevant features caused by environmental interference. The introduction of physical constraints frees the AI ​​model from its "black box" nature, and the extracted features have clear physical meaning. This physics-based feature extraction significantly reduces false alarm rates and provides reliable feature input for subsequent fault diagnosis.

[0087] S203: Input the clean feature tensor into a graph spatiotemporal causal reasoning network, construct a causal prior graph based on historical failure cases throughout the equipment life cycle, enhance causal association mining of rare failures through an adversarial sample generation mechanism, and generate a fault propagation causal graph with probability weights; Specifically, the device state graph node network can be constructed based on the spatial topological relationship of the pure feature tensor to obtain a spatiotemporal topological graph structure with timestamps; Device space topology modeling and node definition Key Component Coordinate Mapping: Based on the 3D model of power generation equipment (such as turbine rotors and generator stators), the spatial coordinates (X, Y, Z) in the 128-dimensional clean feature tensor are bound to their physical locations. For example, the generator front bearing position (X=0 mm, Y=0 mm, Z=0 mm) is defined as node N01, and the rear bearing (X=1200 mm, Y=0 mm, Z=0 mm) is defined as node N02. The spacing between adjacent nodes is ≤1 mm.

[0088] Node attribute loading: Each node is associated with health indicators in the feature tensor. For example, node N01 is associated with [wear factor 0.83, misalignment 0.12, lubrication status 0.05], and node N15 (cooling fan) is associated with [dynamic balance index 0.67, bearing temperature rise 0.21].

[0089] Timestamp embedding: A 64-bit timestamp (e.g., 2023-08-20 14:30:25.123) is appended to each node at 10-millisecond intervals to record the feature collection moment, forming a four-dimensional spatiotemporal coordinate system (X, Y, Z, T).

[0090] Node connection relationship construction Mechanical transmission chain connection: Define edge relationships based on the physical structure of the equipment. For example, the turbine high-pressure cylinder rotor node (N05) and the generator rotor node (N08) are connected by a coupling, and the edge weight is set to a torque transmission efficiency of 0.98 (dimensionless).

[0091] Thermal Influence Domain Connection: Calculates the temperature field gradient. If the temperature difference between two nodes is greater than 50 degrees Celsius and the distance is less than 300 mm, a "heat conduction edge" is automatically created with a weight of 1 / (distance × temperature difference) (unit: per mm per degree Celsius). For example, if the distance between nodes N12 (superheater tube) and N13 (support frame) is 200 mm and the temperature difference is 80 degrees Celsius, the weight is 1 / (200 × 80) = 6.25 × 10 -5 .

[0092] Dynamic Timing Correlation: If the vibration phase difference between two nodes is less than 5 milliseconds, a "synchronous vibration edge" is established with a weight of 1-phase difference / 10 (dimensionless).

[0093] Graph structure optimization and verification Redundant edge pruning: Remove weakly connected edges with weight < 0.01 (such as heat conduction edges with distance > 500 mm and temperature difference < 20 degrees Celsius).

[0094] Topology consistency check: Verifies that the node degree distribution conforms to the physical laws of the equipment. For example, a turbine bearing node should have 3-5 connected edges. If an isolated node (degree = 0) is detected, alarm code G4001 is triggered.

[0095] Extract causal rules based on the equipment's historical lifecycle failure case database and generate a constraint matrix for the causal prior graph; ‌Fault Case Knowledge Extraction‌ Multi-source data correlation: Analyze 2,000 fault cases from 10 years of maintenance records to create a "fault symptom-root cause" mapping table. For example, case F2073 records: "Vibration amplitude suddenly increased to 8 mm / s → Root cause: Coupling misalignment exceeded the limit by 0.2 mm."

[0096] Physical Mechanism Regularization: Convert equipment design specifications into causal constraints. For example, "If the generator winding temperature rise exceeds 80 degrees Celsius for 10 minutes, the insulation aging rate doubles" is recorded as rule R092.

[0097] Material failure model integration: Generate SN curve (stress-life curve) based on fatigue test data. For example, the fatigue life of a bearing steel under a stress amplitude of 350 MPa is 1.2×10 6 Second cycle.

[0098] Constraint Matrix Generation Algorithm Node Impact Weight Calculation: Define causal relationship strength as the number of historical co-occurrences divided by the total number of faults. For example, if nodes N01 (front bearing) and N08 (generator rotor) co-occur 153 times in rotor imbalance faults, and the total number of faults is 2000, then the causal strength is 153 / 2000 = 0.0765.

[0099] Quantification of propagation delay: This method calculates the delay distribution from the cause node anomaly to the effect node anomaly. For example, the average delay from bearing wear (the cause) to gearbox vibration (the effect) is 48 hours, with a standard deviation of ±3 hours.

[0100] Constraint Matrix Construction: Create an N×N matrix (N is the number of nodes). Matrix element A[i,j] stores: causal strength (0–1), mean latency (hours), and standard deviation (hours). Example: A[01,08] = [0.0765, 48, 3].

[0101] ‌Rare Failure Enhancement Mechanism‌ Small Sample Data Enhancement: For rare faults that occur less than 5 times (such as stator winding inter-turn short circuit), synthetic minority class oversampling technology (a data balancing algorithm) is used to generate 10 times more virtual cases.

[0102] Expert rule injection: Manually label implicit causal relationships, such as "Cooling water pH < 6.5 for 30 days → Impeller corrosion rate increases by 300%" as rule R205.

[0103] The spatiotemporal topological graph structure and the constraint matrix of the causal prior graph are input into the spatiotemporal causal reasoning network, features are extracted through gated temporal convolution, and the spatiotemporal enhanced feature tensor is output; Network architecture design Graph Convolutional Layer: Utilizes a third-order neighborhood aggregation algorithm, fusing features from three hops of neighbors at each node. For example, node N01 aggregates features from N02, N08, and N12, with an aggregation weight of 1 / (1 + distance) (unit: per millimeter).

[0104] Gated Temporal Convolution Module: Temporal convolution kernel size = 32 (covering 3.2 seconds of duration), dilation factor = 4 (expanding the receptive field to 12.8 seconds).

[0105] The forget gate filters out non-periodic noise (such as random vibration caused by power grid fluctuations) and retains the device's characteristic frequencies (such as the 85 Hz resonance component).

[0106] ‌Residual connection‌: The output of each layer is superimposed on the original input to prevent gradient disappearance.

[0107] Constraint Matrix Application Mechanism Causal strength weighting: When aggregating neighborhoods, edges with causal strength > 0.05 in the constraint matrix are given a weight 3 times greater than the original weight.

[0108] Delay Alignment Compensation: Time-shift the feature sequence based on the mean delay in the constraint matrix. For example, for a node pair with a causal strength of 0.0765, the feature sequence of the result node is aligned 48 hours earlier.

[0109] Conflict resolution: If the measured delay deviates from the constraint matrix by more than 2 standard deviations, the rule update process is triggered.

[0110] ‌Spatiotemporal feature enhanced output‌ Feature dimension expansion: Input 128-dimensional pure features, process them through a 5-layer network, and output 256-dimensional enhanced features.

[0111] ‌Key Feature Marking‌: Automatically highlight fault-sensitive features, such as the "rotor imbalance index" with a 120% increase in energy in the 85 Hz frequency band.

[0112] Data structure specifications: Output tensor size 256 × N × T (feature channels × number of nodes × time points), time resolution 10 milliseconds.

[0113] Generate adversarial examples corresponding to specific rare fault directions based on spatiotemporal enhancement feature tensors, and enhance causal correlation features through gradient sign method; Rare Fault Directed Generation Fault Mode Focusing: Select a fault type with a historical occurrence rate of less than 1% (e.g., generator rotor axial movement) and define the fault direction vector V_rare in the feature space. For example, V_rare = [0, 0.92, 0, ..., 0.15] (0.92 corresponds to the axial displacement feature).

[0114] Adversarial example generation: Perturb the original features along the V_rare direction with a perturbation step of 0.05 (feature normalization value). Generate an adversarial example sequence of length 100. Example: original features [0.12, 0.07, ..., 0.03] → adversarial examples [0.12, 0.12, ..., 0.18] (axial displacement feature enhancement).

[0115] Gradient Sign Method Enhancement Mechanism Causal feature extraction: Calculate the partial derivative of the fault prediction probability with respect to the input features P / X.

[0116] Feature significance ranking: Select | P / The top 10% features with X|>0.8 are taken as key causal features.

[0117] Adversarial reinforcement: Apply a ±0.1 perturbation to the key features, generating 50 sets of positive and negative perturbation samples. Positive perturbation: +0.1 to the key features; negative perturbation: -0.1 to the key features.

[0118] Adversarial Training Optimization ‌Mixed dataset construction‌: original samples and adversarial samples are mixed in a ratio of 8:2.

[0119] Improved loss function: Added a causal consistency penalty. If the change of non-critical features in the adversarial example is greater than 0.05, an additional loss value of 0.3 is added.

[0120] Improved detection rate of rare faults: After three rounds of adversarial training, the recognition rate of rotor axial movement faults increased from 68% to 92%.

[0121] The enhanced causal correlation features are used to train a Bayesian graph network, so that the trained Bayesian graph network can output a fault propagation causal graph with probability weights.

[0122] ‌Bayesian Graph Network Architecture‌ Node Conditional Probability Table: Defines 256 states for each node (corresponding to the discretization intervals of the enhanced feature). The combination of parent node states determines the conditional probability. Example: P(N08 Fault | N01 Wear > 0.7, N05 Vibration > 4 mm / s) = 0.93.

[0123] Variational Inference Engine: Uses the stochastic gradient variational Bayesian algorithm (a probabilistic model optimization method) and converges after 1,000 iterations.

[0124] Prior distribution setting: By default, the node failure rate follows the Beta distribution (α=1, β=99), corresponding to a 1% prior failure probability.

[0125] Probability Weight Calculation Probabilistic conversion of causal strength: The causal strength in the constraint matrix is ​​converted to conditional probability: P(result node abnormality | cause node abnormality) = causal strength × 0.95 (confidence scaling factor).

[0126] Delay uncertainty modeling: Transmission delay is modeled as a Gaussian distribution, for example, P(N08 fault delay | N01 anomaly) = Gaussian(48,3) (unit: hours).

[0127] Root cause probability backtracking: Using the belief propagation algorithm (a graphical model inference method), the root cause probability is calculated backwards from the leaf nodes: Observing that the temperature of node N12 exceeds the standard → Backtracking to calculate the probability that N01 is the root cause = 0.76; if N08 vibration exceeds the standard at the same time → The root cause probability increases to 0.91.

[0128] ‌Causal Graph Output Specification‌ Graph structure: Directed acyclic graph, where the node diameter is proportional to the failure probability and the edge width is proportional to the causal strength.

[0129] Probability labeling: Each node is labeled with the current failure probability (0-1), and each edge is labeled with the propagation delay (mean ± standard deviation).

[0130] Dynamic update mechanism: Receive new feature tensors every 10 minutes, incrementally update probability values, and use a historical data attenuation factor of 0.95.

[0131] Graph neural networks are used to analyze the spatiotemporal correlations between the status characteristics of various equipment components. Combined with the causal relationships summarized in a historical fault case database, a logical network for fault propagation is constructed. Data augmentation techniques are used to strengthen the learning of causal characteristics for low-probability but high-impact fault types. This causal inference network not only locates the current fault but also predicts potential cascading failure paths. Adversarial training effectively addresses the shortage of rare fault samples, enabling the system to provide early warning of events.

[0132] S204: Perform multi-agent diagnosis on the fault propagation causal graph. Competitive agents are deployed to simulate the fault discrimination logic of the equipment designer, operator, and component manufacturer, respectively. The dispute results of the three parties are integrated through a Nash equilibrium strategy to output a fault diagnosis report with a credibility interval, including fault location and root cause analysis. Specifically, the equipment design specifications, operation and maintenance records, and material database can be loaded according to the node attributes of the fault propagation causal graph to generate a tripartite knowledge vector group; Multi-source knowledge structured extraction Design Specification Parsing: This process reads equipment design documents (e.g., the GB / T 9239 standard for dynamic balancing of steam turbine rotors) and extracts key parameters to form a design knowledge vector. For example, the permissible unbalance limit for a 600 MW generator rotor is 80 g·mm (the product of mass and eccentricity), corresponding to the vector element [dynamic balance limit: 0.08] (normalized to kg·m).

[0133] Maintenance Record Mining: We analyzed a 10-year maintenance database to compile records of similar equipment failure repairs. For example, in bearing replacement logs, lubrication failures accounted for 63% of the time, corresponding to the vector [Lubrication Failure Probability: 0.63]. Winding overheating repairs took an average of 4.5 hours, recorded as [Repair Duration: 4.5].

[0134] Material Database Matching: Retrieve performance parameters based on the node material label. For example, if node N08 (coupling) is labeled "42CrMo alloy steel," then loading its fatigue limit stress of 510 MPa (million Pascals) and yield strength of 930 MPa will generate the vector [Fatigue limit: 510, Yield strength: 930].

[0135] ‌Knowledge Vector Group Construction Rules‌ ‌Dimensionality Unification‌: All three dimensions use a 128-dimensional format, with missing values ​​filled with zeros. Design-oriented dimensions focus on theoretical parameters (e.g., safety factor 1.8, stress concentration factor 2.3), operation-oriented dimensions focus on statistical values ​​(e.g., mean time between failures 8000 hours, spare parts replacement rate 0.12), and manufacturing-oriented dimensions include material properties (e.g., creep rate 3×10 -8 per hour, hardness HRC32).

[0136] Spatiotemporal Correlation: Attach a time-decay weight to each node. For example, maintenance records from three years ago have a weight of 0.7, while current records have a weight of 1.0. Also, consider the design specification version difference coefficient (2015 version has a coefficient of 0.8, 2020 version has a coefficient of 1.0).

[0137] Outlier filtering: Eliminate data that exceeds the physically feasible domain. For example, if a log record shows "bearing temperature rise 300 degrees Celsius" (the actual material melting point is 1500 degrees Celsius), an alarm will be triggered and the industry average will be used instead.

[0138] The three-party knowledge vector group is input into the competitive agent group, where the design agent outputs a failure probability vector based on stress simulation, the operation and maintenance agent outputs a failure probability vector based on maintenance history, and the manufacturing agent outputs a failure probability vector based on fatigue model. ‌Agent Core Algorithm Implementation‌ ‌Designer Agent‌: Stress Field Finite Element Simulation: Generate a 3D mesh model based on the design code vectors and calculate the nodal Mises stress (a composite stress index). If the simulated stress value of node N01 (620 MPa) exceeds the material yield strength of 600 MPa, the failure probability = 1 - exp(-(620 - 600) / 50) = 0.33 (exponential probability mapping function).

[0139] Dynamic operating condition correction: The load is adjusted based on the real-time speed. For example, at 1500 rpm, the centrifugal load factor is 1.2, and the failure probability is simultaneously amplified by 1.2 times.

[0140] ‌Operation and Maintenance Agent‌: Maintenance History Bayesian Inference: Construct a failure probability-maintenance record matrix. For example, node N08 (coupling) has been repaired six times in the past three years, and similar equipment has an average annual failure rate of 0.8. Therefore, the prior probability of failure is 6 / (3 × 0.8) = 2.5 (the portion exceeding the baseline is considered an incremental risk). Using the Sigmoid function to compress the probability to the range [0, 1] yields 0.92.

[0141] Environmental Factor Compensation: If the current ambient temperature is > 35°C (baseline value 25°C), the probability of temperature-related failures increases by (35-25)×0.05=0.5 (5% probability increase per degree Celsius).

[0142] ‌Manufacturer Agent‌: Fatigue life model: Miner linear damage accumulation law (a fatigue failure prediction method) is used. Assume that the current stress amplitude of node N12 (gear) is 300 MPa, and the material SN curve (stress-life curve) shows that the life under this stress is 10 6 Cycles, 8×10 5 times, then the damage degree = 8 / 10 = 0.8, and the failure probability = damage degree² = 0.64 (nonlinear acceleration model).

[0143] Material degradation monitoring: If infrared detection shows that the local temperature of a node is greater than the material recrystallization temperature (e.g. 800 degrees Celsius for nickel-based alloys), the fatigue limit is automatically reduced by 30%.

[0144] ‌Probability Vector Output Specification‌ Dimension Alignment: All three parties output an N-dimensional vector (N is the number of nodes), where the element value is the failure probability in the interval [0, 1].

[0145] ‌Time decay factor‌: The output of the operation and maintenance party needs to be multiplied by the time factor. For example, the weight of the detection data last week is 1.0, and the weight of the data three months ago is 0.6.

[0146] Uncertainty Notation: The manufacturer adds a ±15% confidence band for the dispersion of material properties, and the designer adds a ±8% deviation range for the simulation error.

[0147] Aggregate the nodes whose fault probability vector detection differences between the three parties are greater than a preset threshold to generate a set of dispute points; ‌Difference Detection Algorithm‌ Node-by-node comparison: Calculate the maximum absolute difference between the three probabilities. Formula: Difference = max(|P_design - P_maint|, |P_design - P_manuf|, |P_maint - P_manuf|).

[0148] Example: Node N01 tripartite probability [0.33, 0.92, 0.64] → Difference value = max(|0.33-0.92|,|0.33-0.64|,|0.92-0.64|) = 0.59.

[0149] Threshold judgment: The preset difference threshold is 0.3 (adjustable parameter). If the difference value is greater than 0.3, it is marked as a disputed node.

[0150] Spatial clustering: Dispute nodes are clustered into neighborhoods with a radius of 200 mm to form disputed areas (e.g., the generator front bearing area includes nodes {N01, N03, N05}).

[0151] ‌Controversy Focus Collection Generation‌ ‌Multi-dimensional dispute quantification‌: Conflict of technical routes: The design side is based on theoretical simulation, the operation and maintenance side relies on historical data, and the manufacturing side focuses on material failure.

[0152] Time scale difference: Operations and maintenance focus on short-term maintainability (hours), while designers consider long-term safety (years).

[0153] ‌Focus Prioritization‌: Arrange in descending order by difference value, and nodes with difference value > 0.5 are marked as "key disputes"; The weight of the failure consequences is added. For example, if the downtime loss is greater than 1 million yuan per hour, the priority of the node is increased by 3 levels.

[0154] ‌Collection data structure‌: The output format is a list [node ID, design probability, operation and maintenance probability, manufacturing probability, difference value]. Example: [N01, 0.33, 0.92, 0.64, 0.59] [N08, 0.15, 0.87, 0.21, 0.72].

[0155] Construct a three-party payoff matrix based on the dispute focus set and solve the Nash equilibrium strategy vector; Profit Matrix Modeling Definition of dispute node game: Each dispute node is considered an independent game, and the three parties are game participants.

[0156] ‌Strategy Space Partition‌: Designer strategy: {maintain, not maintain} Operation and maintenance strategy: {immediate repair, delayed monitoring} Manufacturer's strategy: {Replace parts, adjust process} ‌Return function quantification‌: Strategy portfolio design party benefits, operation and maintenance party benefits, manufacturing party benefits (Maintenance, immediate repair, parts replacement) 0.9 0.7 0.6 (Maintenance, immediate repair, process adjustment) 0.8 0.9 0.8 ............ ‌Revenue calculation rules‌: The designer's benefit = safety improvement factor × 0.6 - cost factor × 0.4; the operation and maintenance party's benefit = availability improvement × 0.7 - maintenance cost × 0.3; the manufacturer's benefit = brand reputation × 0.5 + spare parts profit × 0.5.

[0157] Nash equilibrium solution Mixed Strategy Equilibrium Calculation: Use the Lemke-Howson algorithm (a game theory solution method) to iteratively solve the optimal response strategy of the three parties.

[0158] Balanced strategy output: Equilibrium solution for node N01: The designer chooses "maintenance" with a probability of 0.8, the operator chooses "immediate repair" with a probability of 0.7, and the manufacturer chooses "part replacement" with a probability of 0.6; composite strategy vector: [0.8, 0.7, 0.6]; conflict resolution mechanism: If the probability of any party in the equilibrium solution is less than 0.3, the expert arbitration process is triggered, and the minimum guaranteed probability is manually set to 0.4.

[0159] Bootstrap sampling is performed on the Nash equilibrium strategy vector to calculate the preset percentage confidence interval of the fault location; ‌Bootstrap Sampling Process‌ ‌Resampled dataset construction‌: The original tripartite probability vector is used as the population, and 1000 subsample sets are randomly selected with replacement (sample size = 80% of the original data size).

[0160] ‌Equilibrium strategy recalculation‌: Repeat the Nash equilibrium solution for each subsample set to obtain 1000 sets of strategy vectors.

[0161] Sampling example: Original policy vector: [0.8, 0.7, 0.6]; resampling 1: [0.82, 0.68, 0.57]; resampling 2: [0.76, 0.73, 0.62]. The sampling rounds are shown in Table 1.

[0162] Table 1

[0163] Confidence Interval Calculation Quantile statistics method: Sort the probability values ​​of 1000 design strategies, take the 25th value as the 95% confidence lower limit (2.5% quantile), and the 975th value as the upper limit (97.5% quantile).

[0164] Example: After the design probability sequence is sorted, the 25th digit = 0.74, the 975th digit = 0.85 → confidence interval [0.74, 0.85].

[0165] Node-level interval synthesis: The final failure probability of node N01 = balanced strategy mean × node original probability. If the original probability of N01 is 0.33 and the balanced strategy mean is 0.8, then the synthesized probability = 0.33 × 0.8 = 0.264. Similarly, the confidence interval is calculated: original probability interval [0.30, 0.36] × strategy interval [0.74, 0.85] → synthesized interval [0.222, 0.306].

[0166] Based on the confidence interval and causal graph, the root cause path with a probability weight greater than the preset weight is traced back, and a diagnostic report with a credibility interval is output.

[0167] Root Cause Path Backtracking Algorithm‌ Threshold filtering: Select nodes with a node failure probability greater than 0.3 (preset weight) as candidate root causes.

[0168] Causal graph tracing: Traverse the causal edges backward from the leaf node (such as vibration exceeding node N12) and select the upstream node with the highest probability weight.

[0169] Path probability = product of conditional probabilities along the path × node probability. Example: path N01 → N08 → N12, conditional probability P(N08|N01) = 0.93, P(N12|N08) = 0.87, N01 probability 0.33 → path probability = 0.33 × 0.93 × 0.87 = 0.267 Multi-path competition: Paths with probability greater than 0.15 are retained and sorted in descending order of probability.

[0170] ‌Diagnostic report generation specifications‌ Example of structured output: Fault location report Root cause node: N01 (front bearing); Probability: 0.264 [95% confidence interval: 0.222-0.306]; Transmission path: N01→N08→N12 [path probability: 0.267]; N01→N05→N12 [path probability: 0.198]; Dispute Resolution Instructions The designer claims: the probability of maintenance necessity is 0.8 [0.74-0.85]; Operation and maintenance suggestion: Probability of immediate repair 0.7 [0.65-0.75]; Manufacturer's solution: Probability of replacing parts is 0.6 [0.55-0.65].

[0171] Visualization Enhancement: Root cause nodes (probability > 0.3) are marked in red in the 3D device model, and propagation paths are marked in yellow. Confidence intervals are superimposed on the probability histogram as error bars.

[0172] Three intelligent agents with different professional perspectives collaboratively analyze the fault map: the design team focuses on theoretical calculations, the operations team on practical experience, and the manufacturer on material properties. These three parties reach consensus through game theory, and the final diagnosis results are accompanied by probability assessments and confidence intervals for different perspectives. This multi-perspective diagnosis mechanism avoids the limitations of single-point judgment, and the Nash equilibrium strategy ensures the scientific and acceptable conclusions. Reports with confidence intervals provide a risk quantification basis for decision-making, significantly improving the actionability of diagnostic results.

[0173] S205, mapping the fault diagnosis report to the equipment digital twin in real time, combining reinforcement learning with Monte Carlo fault tree simulation, rehearsing the equipment degradation trajectory under different maintenance strategies in a virtual space, and outputting an adaptive maintenance strategy sequence that minimizes the expected value of the full life cycle operation and maintenance cost.

[0174] Specifically, the equipment degradation model can be initialized in the digital twin based on the root cause path sequence of the fault diagnosis report to obtain the component degradation rate equation; Root cause path analysis and physical mechanism mapping Analyze the root cause node sequence in the fault diagnosis report (for example: front bearing wear → coupling misalignment → excessive gearbox vibration) and establish a degradation transmission chain based on the equipment drive chain topology. For the root cause node, the front bearing (identification number N01), load the parameters from the material property library: material type 42 chromium-molybdenum alloy steel, hardness Rockwell hardness 32, wear resistance coefficient 1.5×10 -8 (dimensionless parameter).

[0175] Dynamic load characteristics were extracted through vibration spectrum analysis: the contact stress at a main frequency of 85 Hz was 18 megapascals, which, combined with a rotational speed of 1500 rpm, translated into a sliding velocity of 3.2 meters per second. The Hertzian contact wear model was used to construct the following equation: Wear change rate = (dynamic contact stress × sliding velocity) / (material hardness × wear resistance coefficient) = (18 × 3.2) / (32 × 1.5 × 1010 -8 ) = 0.15 (dimensionless parameter).

[0176] For electrical faults (such as stator winding insulation aging), the Arrhenius thermal aging model is used to calculate the rate multiplier based on a temperature rise slope of 1.8 degrees Celsius per second: aging rate multiplier = e^[(activation energy / gas constant)×(1 / reference temperature - 1 / actual temperature)]; substituting an activation energy of 1.2 electron volts, a gas constant of 8.314 joules per mole per kelvin, a reference temperature of 80 degrees Celsius (353 degrees Kelvin), and an actual temperature of 110 degrees Celsius (383 degrees Kelvin): the multiplier = e^[(1.2 / 8.314)×(1 / 353-1 / 383)] ≈ 4.3 times.

[0177] Multiple failure mode coupling and environmental correction When a component experiences concurrent degradation (e.g., simultaneous wear and fatigue cracking in a bearing), a damage superposition equation is established: Total damage = 0.7 × (Current wear / Wear threshold) + 0.3 × (Actual number of stress cycles / Fatigue life). Weighting factors are determined based on historical fault database statistics (e.g., wear accounts for 70% and fatigue accounts for 30%).

[0178] Dynamic correction of environmental factors: If the report indicates that the ambient humidity is ‌90%‌ (exceeding the baseline value of 80%), the corrosion component will be increased by ‌5%‌ according to the rule that the corrosion rate increases by 5% for every 10% humidity increase.

[0179] Model Validation and Parameter Calibration Input the current measured bearing clearance of 0.52 mm into the digital twin, run the degradation model for 24 hours, and output a predicted value of 0.523 mm.

[0180] Compared with the real-time data of the laser micrometer, which is 0.524 mm, the relative error is 0.19% (less than the threshold of 1%), and the model is verified. If the error exceeds the limit, the wear coefficient will be automatically adjusted (for example, ±0.1×10 -8 ) and iterate again.

[0181] Sample output: Part identification number: N01; Degradation type: Contact wear; Rate equation: 0.15 + 0.05 × humidity correction factor (unit: microns per hour); Applicable conditions: Speed ​​1200-1800 rpm, temperature <150 degrees Celsius.

[0182] Based on the component degradation rate equation, a three-dimensional discrete action space is defined to generate the maintenance strategy search domain; ‌3D Motion Space Quantization Coding‌ Time dimension: Based on the equipment thermal inertia cycle, the maintenance window is divided into 4-hour intervals, covering a 720-hour (30-day) period, generating 180 discrete time points (sequence: 0, 4, 8, ..., 716).

[0183] Operation type dimension: Three basic actions are defined: Maintenance (code B): such as adding lubricating oil, which takes 0.5 hours and costs 800 yuan, with the effect of reducing the wear change rate by 30%; Inspection (code J): such as adjusting the bearing clearance, which takes 2 hours and costs 5,000 yuan, and the effect of restoring the wear to 80% of the initial value; Replacement (code G): such as replacing the entire bearing, which takes 8 hours and costs 30,000 yuan, and the effect of returning the wear to zero.

[0184] Execution intensity dimension: Each type of action is divided into three levels (maintenance intensity level 1: refueling 200 ml, level 2: 400 ml, level 3: 600 ml). The intensity difference causes the cost to fluctuate by ±20% and the effect to change by ±15%.

[0185] Physics Constraints and State Pruning Status feasibility rules: When the wear amount is less than 0.1 mm, replacement is prohibited (to avoid excessive maintenance); when the insulation aging rate is greater than 5 times the baseline value, maintenance or replacement is forced to be triggered.

[0186] Economic constraints: The upper limit for daily maintenance costs is 20,000 yuan (if the cost of a replacement action exceeds 30,000 yuan, it must be eliminated).

[0187] Timing constraint: Repair of the same component is prohibited within 48 hours after replacement.

[0188] Search domain compression: Traverse theoretical combinations (180 time points × 3 operation types × 3 intensity levels = 1620 dimensions). If the current wear change rate is less than 0.1, remove all replacement actions, reducing the search space to 1080 dimensions.

[0189] An example of a search domain matrix includes the information shown in Table 2.

[0190] Table 2

[0191] In the maintenance strategy search domain, a number of paths are explored using a proximal strategy optimization algorithm, and a preview trajectory and cost dataset are output; Reinforcement Learning Agent Design‌ ‌State Representation‌: Encodes the device state as a ‌128-dimensional vector‌, including: Normalized wear (current value 0.52 → mapped to the [0,1] interval to 0.65); temperature exceedance ratio (measured 110°C / threshold 90°C = 1.22); vibration harmonic energy ratio (85 Hz component accounts for 62% of the total).

[0192] Reward Function Design: Multi-objective weighted aggregation formula: Comprehensive reward = equipment availability × 60 - normalized maintenance cost × 30 - failure risk coefficient × 10.

[0193] Among them, availability = (total duration - downtime) / total duration × 100; normalized maintenance cost = actual cost / 50,000 yuan; failure risk factor = min(damage amount / failure threshold, 1.0).

[0194] Proximal Strategy Optimization Process ‌Interaction sampling‌: The agent performs action ACT0308 (replace bearing strength level 3 at the 240th hour); the word twin returns to a new state (insulation aging rate returns to zero) and a reward value ‌75.6‌ (calculation: availability 88%×60 - 30000 / 50000×30- 0.05×10=52.8-18-0.5=34.3→normalized to 75.6).

[0195] ‌Policy update mechanism‌: The probability ratio of the new and old strategies is limited to the range of [0.8, 1.2] (if the new probability is 0.15 and the old probability is 0.10, the adjustment factor takes the minimum value (0.15 / 0.10×0.8, 1.2)=1.2); the upper limit of the gradient clipping step is set to 0.002.

[0196] ‌Preview trajectory and cost data output‌ Explore 50 policy paths, each containing 30 action sequences: Path P09 sequence: Time 0 hours → ACT0101 (Maintenance intensity 1) → Cost +800 yuan → Wear change rate reduced to 0.12; Time 240 hours → ACT0308 (Replacement intensity 3) → Cost +30,000 yuan → Insulation aging returns to zero; Total cost 30,800 yuan, availability 88%, failure risk coefficient 0.05.

[0197] Cost dataset example: Path number: 23; Maintenance cost (yuan): 0000; Downtime loss (yuan): 8000; Total cost (yuan): 8000; Availability: 5%.

[0198] Perform Monte Carlo fault tree analysis on the cost data set to calculate the expected value of the full life cycle operation and maintenance cost; Fault Tree Event Probability Calibration Basic event probability: Based on statistics from the equipment's 10-year operation and maintenance database: Probability of bearing wear failure = 6 failures / 10 years = 0.6 per year; Probability of insulation breakdown = 12 events / 5000 operating days = 0.0024 per day.

[0199] Logic gate rules: OR gate: The parent event is triggered when any sub-event occurs (such as bearing failure or insulation breakdown causing shutdown); AND gate: The parent event is triggered only when all sub-events occur at the same time (such as cooling failure and temperature exceeding the limit causing tripping).

[0200] Monte Carlo Simulation Process ‌Random Sampling‌: Generate ‌10,000 sets‌ of ‌[0,1] uniformly distributed random numbers: Bearing wear sampling: random number 0.43 < 0.6 → mark occurs; insulation breakdown sampling: random number 0.998 > 0.0024 → mark does not occur.

[0201] Calculation of top event probability: The number of unit shutdown events is 1500 → Probability = 1500 / 10000 = 0.15; Economic loss correlation: Single downtime loss = 8 hours of downtime × 30,000 yuan loss per hour = 240,000 yuan; insulation breakdown loss = 1,200,000 yuan cost of replacing the winding.

[0202] Expected Cost Synthesis and Distribution Fitting Single strategy expected cost formula: Expected total cost = maintenance cost + Σ(top event probability × corresponding loss).

[0203] Example path P23: Maintenance cost 48,000 yuan + (0.15×240,000 + 0.002×1,200,000) = 48,000+36,000+2,400=86,400 yuan.

[0204] Probability distribution output: expected value 86,400 yuan, standard deviation ±12,000 yuan; 95% confidence interval [74,400, 98,400] yuan.

[0205] The optimal strategy is selected according to the expected cost value and the adaptive maintenance strategy sequence is output.

[0206] Multi-objective decision-making model Indicator weight distribution: Cost expectation weight 0.65; Equipment availability weight 0.25; Strategy complexity weight 0.10 (evaluated by the number of action types).

[0207] Normalized score calculation: Cost score = (maximum cost 150,000 yuan - current cost 86,400 yuan) / (maximum cost - minimum cost 50,000 yuan) = 0.636; Availability score = 92% / 100% = 0.92; Complexity score = 1 - number of action types (3 categories) / 10 = 0.7; Overall score = 0.65 × 0.636 + 0.25 × 0.92 + 0.10 × 0.7 = 0.713.

[0208] Adaptive Strategy Sequence Generation Dynamic response mechanism: If the real-time monitoring data deviates from the preview trajectory by more than 5% (e.g., the measured wear volume is 0.55 mm > the predicted 0.52 mm), a strategy reassessment is triggered. If the ambient temperature suddenly rises by more than 10 degrees Celsius, the strategy library switches to the high-temperature strategy library (e.g., increasing the intensity of the cooling action).

[0209] ‌Sequence output specification‌: The optimal strategy sequence is as follows: time 0 hours → maintenance action (B) intensity 1 → wear change rate reduced to 0.12; time 48 hours → inspection action (J) intensity 2 → wear reset to 0.01 mm; time 240 hours → maintenance action (B) intensity 3 → wear change rate reduced to 0.08; total expected cost 86,400 yuan, availability 92%, and comprehensive score 0.713.

[0210] ‌Visual Verification and Economic Comparison‌ Digital Twin Mapping: In the 3D model, replacement parts (such as bearings) are marked in red, and maintenance parts (such as gearboxes) are marked in yellow; a dynamic curve shows that the remaining life has been extended from 3 months to 8 months.

[0211] ‌Execution Effectiveness Verification‌: The actual maintenance cost of 85,000 yuan fell within the predicted range of [74,400, 98,400 yuan] → verified effective; this represents a 32.5% reduction compared to the traditional scheduled maintenance strategy (expected cost of 128,000 yuan).

[0212] Based on diagnostic results, the long-term effects of various maintenance options are simulated in the digital twin model. Reinforcement learning is used to optimize decision paths, while probabilistic statistical methods are used to assess the risks and costs of different options. Ultimately, an optimal maintenance plan is generated that considers the economic benefits of the equipment's entire lifecycle. Combining real-time diagnostics with long-term planning achieves closed-loop management from condition monitoring to maintenance decision-making. Virtual simulation avoids the high cost of trial and error in practice, and the optimized maintenance strategy can extend equipment life while reducing operation and maintenance costs.

[0213] It can be seen that according to the physical topology of the power generation equipment, mechanical excitation signals with a preset spectrum are actively injected into key components to generate a fusion of time-space-frequency three-dimensional data volume; the three-dimensional data volume is input into the physical embedded variational autoencoder to output the pure feature tensor of the equipment state; the pure feature tensor is input into the graph space-time causal reasoning network to generate a fault propagation causal graph with probability weights; multi-agent diagnosis is performed on the fault propagation causal graph to output a fault diagnosis report with a credibility interval, including fault location and root cause analysis; the fault diagnosis report is mapped to the equipment digital twin in real time, and an adaptive maintenance strategy sequence is output that minimizes the expected value of the operation and maintenance cost throughout the life cycle, thereby improving the accuracy and anti-interference ability of fault diagnosis and effectively reducing the operation and maintenance cost.

[0214] Another embodiment of the present invention provides a power generation equipment status fault diagnosis system based on artificial intelligence, see Figure 3 , the system may include: Acquisition module 301 is used to actively inject mechanical excitation signals with a preset frequency spectrum into key components based on the physical topology of the power generation equipment. It simultaneously collects the transient responses of vibration sensors, infrared thermal imagers, and current transformers under excitation, and combines environmental parameters with equipment operation logs to generate a three-dimensional time-space-frequency data volume. A separation module 302 is configured to input the three-dimensional data volume into a physical embedded variational autoencoder, construct a physical constraint loss function using the device thermal-mechanical coupling equation, separate the intrinsic physical feature vectors representing the health status of the component from the redundant feature vectors interfered by environmental noise, and output a pure feature tensor of the device state; Construction module 303 is used to input the clean feature tensor into a graph spatiotemporal causal reasoning network, construct a causal prior graph based on historical failure cases throughout the equipment life cycle, enhance causal association mining of rare failures through an adversarial sample generation mechanism, and generate a fault propagation causal graph with probability weights; Diagnosis module 304 is configured to perform multi-agent diagnosis on the fault propagation causal graph, deploy competitive agents to simulate the fault discrimination logic of the equipment designer, operator, and component manufacturer, respectively, fuse the dispute results of the three parties through a Nash equilibrium strategy, and output a fault diagnosis report with credibility intervals, including fault location and root cause analysis; The output module 305 is used to map the fault diagnosis report to the equipment digital twin in real time, combine reinforcement learning with Monte Carlo fault tree simulation, preview the equipment degradation trajectory under different maintenance strategies in virtual space, and output an adaptive maintenance strategy sequence that minimizes the expected value of the full life cycle operation and maintenance cost.

[0215] An embodiment of the present invention further provides a storage medium storing a computer program, wherein the computer program is configured to execute the steps of any one of the above method embodiments when running.

[0216] Specifically, in this embodiment, the above-mentioned storage medium may be configured to store a computer program for performing the following steps: S201: Based on the physical topology of the power generation equipment, a mechanical excitation signal with a preset frequency spectrum is actively injected into key components. The transient responses of the vibration sensor, infrared thermal imager, and current transformer under the excitation are simultaneously collected. The transient responses are combined with environmental parameters and equipment operation logs to generate a three-dimensional time-space-frequency data volume. S202: Input the three-dimensional data volume into a physical embedded variational autoencoder, construct a physical constraint loss function using the device thermal-mechanical coupling equation, separate the intrinsic physical feature vectors representing the health status of the component and the redundant feature vectors interfered by environmental noise, and output a pure feature tensor of the device state; S203: Input the clean feature tensor into a graph spatiotemporal causal reasoning network, construct a causal prior graph based on historical failure cases throughout the equipment life cycle, enhance causal association mining of rare failures through an adversarial sample generation mechanism, and generate a fault propagation causal graph with probability weights; S204: Perform multi-agent diagnosis on the fault propagation causal graph. Competitive agents are deployed to simulate the fault discrimination logic of the equipment designer, operator, and component manufacturer, respectively. The dispute results of the three parties are integrated through a Nash equilibrium strategy to output a fault diagnosis report with a credibility interval, including fault location and root cause analysis. S205, mapping the fault diagnosis report to the equipment digital twin in real time, combining reinforcement learning with Monte Carlo fault tree simulation, rehearsing the equipment degradation trajectory under different maintenance strategies in a virtual space, and outputting an adaptive maintenance strategy sequence that minimizes the expected value of the full life cycle operation and maintenance cost.

[0217] An embodiment of the present invention further provides an electronic device, comprising a memory and a processor, wherein the memory stores a computer program, and the processor is configured to run the computer program to perform the steps in any one of the above method embodiments.

[0218] Specifically, the electronic device may further include a transmission device and an input / output device, wherein the transmission device is connected to the processor, and the input / output device is connected to the processor.

[0219] Specifically, in this embodiment, the processor may be configured to execute the following steps through a computer program: S201: Based on the physical topology of the power generation equipment, a mechanical excitation signal with a preset frequency spectrum is actively injected into key components. The transient responses of the vibration sensor, infrared thermal imager, and current transformer under the excitation are simultaneously collected. The transient responses are combined with environmental parameters and equipment operation logs to generate a three-dimensional time-space-frequency data volume. S202: Input the three-dimensional data volume into a physical embedded variational autoencoder, construct a physical constraint loss function using the device thermal-mechanical coupling equation, separate the intrinsic physical feature vectors representing the health status of the component and the redundant feature vectors interfered by environmental noise, and output a pure feature tensor of the device state; S203: Input the clean feature tensor into a graph spatiotemporal causal reasoning network, construct a causal prior graph based on historical failure cases throughout the equipment life cycle, enhance causal association mining of rare failures through an adversarial sample generation mechanism, and generate a fault propagation causal graph with probability weights; S204: Perform multi-agent diagnosis on the fault propagation causal graph. Competitive agents are deployed to simulate the fault discrimination logic of the equipment designer, operator, and component manufacturer, respectively. The dispute results of the three parties are integrated through a Nash equilibrium strategy to output a fault diagnosis report with a credibility interval, including fault location and root cause analysis. S205, mapping the fault diagnosis report to the equipment digital twin in real time, combining reinforcement learning with Monte Carlo fault tree simulation, rehearsing the equipment degradation trajectory under different maintenance strategies in a virtual space, and outputting an adaptive maintenance strategy sequence that minimizes the expected value of the full life cycle operation and maintenance cost.

[0220] The above describes in detail the structure, features and effects of the present invention based on the embodiments shown in the drawings. The above is only a preferred embodiment of the present invention, but the scope of implementation of the present invention is not limited to what is shown in the drawings. Any changes made in accordance with the concept of the present invention, or modifications to equivalent embodiments with equivalent changes, which do not exceed the spirit covered by the description and drawings, should be within the scope of protection of the present invention.

Claims

1. A method for diagnosing power generation equipment status faults based on artificial intelligence, characterized in that: The method comprises: Based on the physical topology of the power generation equipment, mechanical excitation signals with a preset spectrum are actively injected into key components. The transient responses of vibration sensors, infrared thermal imagers, and current transformers under excitation are simultaneously collected. Combined with environmental parameters and equipment operation logs, a fusion of time-space-frequency three-dimensional data is generated. Inputting the three-dimensional data volume into a physical embedded variational autoencoder, constructing a physical constraint loss function using the device thermal-mechanical coupling equation, separating the intrinsic physical feature vectors representing the health status of the component from the redundant feature vectors interfered by environmental noise, and outputting a pure feature tensor of the device state; The clean feature tensor is input into the graph spatiotemporal causal reasoning network, and a causal prior graph is constructed based on the historical failure cases of the equipment throughout its life cycle. The adversarial sample generation mechanism is used to enhance the causal association mining of rare failures and generate a fault propagation causal graph with probability weights. Perform multi-agent diagnosis on the fault propagation causal graph, deploy competitive agents to simulate the fault discrimination logic of the equipment designer, operator, and component manufacturer, respectively, fuse the dispute results of the three parties through a Nash equilibrium strategy, and output a fault diagnosis report with credibility intervals, including fault location and root cause analysis; The fault diagnosis report is mapped to the equipment digital twin in real time. Combining reinforcement learning with Monte Carlo fault tree simulation, the equipment degradation trajectory under different maintenance strategies is previewed in virtual space, and an adaptive maintenance strategy sequence that minimizes the expected value of the full life cycle operation and maintenance cost is output.

2. The method according to claim 1, characterized in that Based on the physical topology of the power generation equipment, a mechanical excitation signal with a preset spectrum is actively injected into key components. The transient responses of the vibration sensor, infrared thermal imager, and current transformer under excitation are synchronously collected. The environmental parameters and equipment operation logs are combined to generate a three-dimensional time-space-frequency data volume, including: Based on the bearing stiffness distribution matrix in the physical topology of the power generation equipment, the resonance sensitive areas are identified and the coordinate mapping table of key components is generated; Based on the coordinate mapping table of key components, an amplitude-modulated swept-frequency mechanical excitation signal is injected into the specified position to generate a synchronous trigger pulse sequence; Using a synchronized trigger pulse sequence, the time domain waveform of the vibration sensor, the temperature field matrix of the infrared thermal imager, and the harmonic distortion spectrum of the current transformer are collected in parallel to obtain time-aligned multi-source response data packets. Combining environmental parameters with the noise floor in the device operation log, an adaptive filter is used to separate the environmental interference component from the multi-source response data packet and extract the device's intrinsic transient response set. The device's intrinsic transient response set is reorganized into tensors according to the time axis, spatial coordinates, and frequency dimensions, and a time-space-frequency three-dimensional data volume is output.

3. The method according to claim 2, characterized in that The three-dimensional data volume is input into a physical embedded variational autoencoder, and a physical constraint loss function is constructed using the device thermal-mechanical coupling equation to separate the intrinsic physical feature vectors representing the health status of the component and the redundant feature vectors interfered by environmental noise, and output a pure feature tensor of the device state, including: Based on the device's thermal-mechanical coupling equations, a physical regularization module for the variational autoencoder is constructed; The three-dimensional data volume is input into the variational autoencoder network, and the initial intrinsic feature vector and redundant feature vector are generated through the feature decoupling layer; Use the physical regularization module to perform finite element simulation on the initial eigenvectors to generate the theoretical thermal response field; Calculate the residual matrix between the theoretical thermal response field and the measured infrared temperature field, and construct a loss function for physical constraints by combining KL divergence; The variational autoencoder parameters are reversely optimized through the loss function, and the pure feature tensor of the device state is output.

4. The method according to claim 3, characterized in that The clean feature tensor is input into the graph spatiotemporal causal reasoning network, a causal prior graph is constructed based on historical failure cases throughout the equipment life cycle, and the causal association mining of rare failures is enhanced through the adversarial sample generation mechanism to generate a fault propagation causal graph with probability weights, including: The device state graph node network is constructed based on the spatial topological relationship of the pure feature tensor to obtain a spatiotemporal topological graph structure with timestamps. Extract causal rules based on the equipment's historical lifecycle failure case database and generate a constraint matrix for the causal prior graph; The spatiotemporal topological graph structure and the constraint matrix of the causal prior graph are input into the spatiotemporal causal reasoning network, features are extracted through gated temporal convolution, and the spatiotemporal enhanced feature tensor is output; Generate adversarial examples corresponding to specific rare fault directions based on spatiotemporal enhancement feature tensors, and enhance causal correlation features through gradient sign method; The enhanced causal correlation features are used to train a Bayesian graph network, so that the trained Bayesian graph network can output a fault propagation causal graph with probability weights.

5. The method according to claim 4, characterized in that The multi-agent diagnosis of the fault propagation causal graph is performed, and competitive agents are deployed to simulate the fault discrimination logic of the equipment designer, the operation and maintenance party, and the component manufacturer respectively. The dispute results of the three parties are integrated through the Nash equilibrium strategy, and a fault diagnosis report with a credibility interval, including fault location and root cause analysis, is output, including: Load equipment design specifications, operation and maintenance records, and material databases based on node attributes of the fault propagation causal graph to generate a tripartite knowledge vector group; The three-party knowledge vector group is input into the competitive agent group, where the design agent outputs a failure probability vector based on stress simulation, the operation and maintenance agent outputs a failure probability vector based on maintenance history, and the manufacturing agent outputs a failure probability vector based on fatigue model. Aggregate the nodes whose fault probability vector detection differences between the three parties are greater than a preset threshold to generate a set of dispute points; Construct a three-party payoff matrix based on the dispute focus set and solve the Nash equilibrium strategy vector; Bootstrap sampling is performed on the Nash equilibrium strategy vector to calculate the preset percentage confidence interval of the fault location; Based on the confidence interval and causal graph, the root cause path with a probability weight greater than the preset weight is traced back, and a diagnostic report with a credibility interval is output.

6. The method according to claim 5, characterized in that The fault diagnosis report is mapped to the equipment digital twin in real time, and reinforcement learning and Monte Carlo fault tree simulation are combined to preview the equipment degradation trajectory under different maintenance strategies in a virtual space, and output an adaptive maintenance strategy sequence that minimizes the expected value of the full life cycle operation and maintenance cost, including: Based on the root cause path sequence in the fault diagnosis report, the equipment degradation model is initialized in the digital twin to obtain the component degradation rate equation; Based on the component degradation rate equation, a three-dimensional discrete action space is defined to generate the maintenance strategy search domain; In the maintenance strategy search domain, a number of paths are explored using a proximal strategy optimization algorithm, and a preview trajectory and cost dataset are output; Perform Monte Carlo fault tree analysis on the cost data set to calculate the expected value of the full life cycle operation and maintenance cost; The optimal strategy is selected according to the expected cost value and the adaptive maintenance strategy sequence is output.

7. A power generation equipment status fault diagnosis system based on artificial intelligence, characterized in that: The system comprises: The acquisition module is used to actively inject mechanical excitation signals with a preset frequency spectrum into key components based on the physical topology of the power generation equipment. It simultaneously collects the transient responses of vibration sensors, infrared thermal imagers, and current transformers under excitation, and combines environmental parameters with equipment operation logs to generate a three-dimensional time-space-frequency data volume. a separation module, configured to input the three-dimensional data volume into a physical embedded variational autoencoder, construct a physical constraint loss function using the device thermal-mechanical coupling equation, separate the intrinsic physical feature vectors representing the health status of the component from the redundant feature vectors interfered by environmental noise, and output a pure feature tensor of the device state; A construction module is used to input the clean feature tensor into a graph spatiotemporal causal reasoning network, construct a causal prior graph based on historical failure cases throughout the equipment life cycle, enhance causal association mining of rare failures through an adversarial sample generation mechanism, and generate a fault propagation causal graph with probability weights; A diagnostic module is used to perform multi-agent diagnosis on the fault propagation causal graph, deploy competitive agents to simulate the fault discrimination logic of the equipment designer, operator, and component manufacturer, fuse the dispute results of the three parties through a Nash equilibrium strategy, and output a fault diagnosis report with a credibility interval, including fault location and root cause analysis; The output module is used to map the fault diagnosis report to the equipment digital twin in real time, combine reinforcement learning with Monte Carlo fault tree simulation, preview the equipment degradation trajectory under different maintenance strategies in virtual space, and output an adaptive maintenance strategy sequence that minimizes the expected value of the operation and maintenance cost throughout the life cycle.

8. The system according to claim 7, characterized in that The acquisition module is specifically used to: Based on the bearing stiffness distribution matrix in the physical topology of the power generation equipment, the resonance sensitive areas are identified and the coordinate mapping table of key components is generated; Based on the coordinate mapping table of key components, an amplitude-modulated swept-frequency mechanical excitation signal is injected into the specified position to generate a synchronous trigger pulse sequence; Using a synchronized trigger pulse sequence, the time domain waveform of the vibration sensor, the temperature field matrix of the infrared thermal imager, and the harmonic distortion spectrum of the current transformer are collected in parallel to obtain time-aligned multi-source response data packets. Combining environmental parameters with the noise floor in the device operation log, an adaptive filter is used to separate the environmental interference component from the multi-source response data packet and extract the device's intrinsic transient response set. The device's intrinsic transient response set is reorganized into tensors according to the time axis, spatial coordinates, and frequency dimensions, and a time-space-frequency three-dimensional data volume is output.

9. A storage medium, characterized in that: The storage medium stores a computer program, wherein the computer program is configured to execute the method according to any one of claims 1 to 6 when executed.

10. An electronic device comprising a memory and a processor, characterized in that: A computer program is stored in the memory, and the processor is configured to run the computer program to perform the method according to any one of claims 1 to 6.

Images