An artificial intelligence-based power generation equipment state fault diagnosis method and system

Abstract

The application discloses a power generation equipment state fault diagnosis method and system based on artificial intelligence, and the method comprises the following steps: according to the physical topology structure of the power generation equipment, a mechanical excitation signal with a preset frequency spectrum is actively injected into a key component, and a time-space-frequency three-dimensional data body is fused and generated; the three-dimensional data body is input into a physically embedded variational autoencoder, and a device state pure feature tensor is output; the pure feature tensor is input into a graph space-time causal reasoning network, and a fault propagation causal graph with a probability weight is generated; the fault propagation causal graph is subjected to multi-agent diagnosis, and a fault diagnosis report containing fault positioning and root cause analysis with a credibility interval is output; and the fault diagnosis report is mapped to a device digital twin in real time, and an adaptive maintenance strategy sequence that minimizes the life cycle operation and maintenance cost expectation value is output. By using the embodiment of the application, the accuracy and anti-interference ability of fault diagnosis can be improved, and the operation and maintenance cost can be effectively reduced.

Description

TECHNICAL FIELD

[0001] The application belongs to the technical field of fault diagnosis, and particularly relates to a power generation equipment state fault diagnosis method and system based on artificial intelligence. BACKGROUND

[0002] With the development of power generation equipment towards high parameters and large capacity, the traditional fault diagnosis method faces challenges such as response lag, high false alarm rate, and difficulty in identifying rare faults. The existing technology mainly relies on single sensor data or offline analysis, which is difficult to capture the multi-physical field coupling characteristics of the equipment under complex working conditions, and the environmental noise interference leads to insufficient reliability of the diagnosis results. In addition, the fault root cause analysis mainly relies on expert experience, and it is difficult to quantify the impact of different maintenance strategies on the life cycle cost of the equipment. SUMMARY

[0003] The purpose of the present application is to provide a power generation equipment state fault diagnosis method and system based on artificial intelligence to solve the problems in the prior art and improve the accuracy and anti-interference ability of fault diagnosis, and effectively reduce the operation and maintenance cost.

[0004] One embodiment of the present application provides a power generation equipment state fault diagnosis method based on artificial intelligence, which comprises the following steps:

[0005] According to the physical topology structure of the power generation equipment, a mechanical excitation signal of a preset frequency spectrum is actively injected into a key component, the transient response of a vibration sensor, an infrared thermal imager and a current transformer under excitation is synchronously collected, environmental parameters and equipment operation logs are combined, and a time-space-frequency three-dimensional data body is fused and generated;

[0006] The three-dimensional data body is input into a physically embedded variational autoencoder, a physical constraint loss function is constructed by using a device thermal-mechanical coupling equation, intrinsic physical feature vectors representing the health state of the component and redundant feature vectors disturbed by environmental noise are separated, and a pure feature tensor of the equipment state is output;

[0007] The pure feature tensor is input into a graph spatio-temporal causal reasoning network, a causal prior graph is constructed based on historical full life cycle fault cases of the equipment, the causal correlation mining of rare faults is strengthened through an adversarial sample generation mechanism, and a fault propagation causal graph with probability weight is generated;

[0008] The fault propagation causal graph is diagnosed by multiple intelligent agents, competitive intelligent agents are deployed to simulate the fault judgment logic of the equipment design party, the operation and maintenance party and the component manufacturing party respectively, the controversial results of the three parties are fused through a Nash equilibrium strategy, and a fault diagnosis report with a credibility interval and containing fault location and root cause analysis is output;

[0009] The fault diagnosis report is mapped to a device digital twin in real time, combined with reinforcement learning and Monte Carlo fault tree simulation, and the device degradation trajectory under different maintenance strategies is preplayed in a virtual space, and an adaptive maintenance strategy sequence that minimizes the expected value of the whole life cycle operation and maintenance cost is output.

[0010] Optionally, according to the physical topology of the power generation equipment, a preset frequency spectrum mechanical excitation signal is actively injected into the key component, the transient responses of the vibration sensor, the infrared thermal imager and the current transformer under excitation are synchronously collected, the environmental parameters and the equipment operation log are combined, and a time-space-frequency three-dimensional data body is fused to include:

[0011] According to the bearing stiffness distribution matrix in the physical topology of the power generation equipment, a resonance sensitive area is identified, and a key component coordinate mapping table is generated;

[0012] Based on the key component coordinate mapping table, an amplitude-modulated swept-frequency mechanical excitation signal is injected into the specified position, and a synchronous trigger pulse sequence is generated;

[0013] Using the synchronous 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, and time-aligned multi-source response data packets are obtained;

[0014] Combined with the noise floor in the environmental parameters and the equipment operation log, the environmental interference components are separated from the multi-source response data packets through an adaptive filter, and a set of intrinsic transient responses of the equipment is extracted;

[0015] The set of intrinsic transient responses of the equipment is tensor-reorganized according to the time axis, the spatial coordinates and the frequency dimension, and a time-space-frequency three-dimensional data body is output.

[0016] Optionally, the three-dimensional data body is input into a physically embedded variational autoencoder, a physical constraint loss function is constructed using the equipment thermal-mechanical coupling equation, intrinsic physical feature vectors representing the component health state and redundant feature vectors disturbed by environmental noise are separated, and a pure equipment state feature tensor is output, including:

[0017] According to the equipment thermal-mechanical coupling equation, a physical regularization module of the variational autoencoder is constructed;

[0018] The three-dimensional data body is input into the variational autoencoder network, and initial intrinsic feature vectors and redundant feature vectors are generated through a feature decoupling layer;

[0019] The initial intrinsic feature vectors are subjected to finite element simulation using the physical regularization module, and a theoretical thermal-mechanical response field is generated;

[0020] The residual matrix of the theoretical thermal-mechanical response field and the measured infrared temperature field is calculated, and a loss function with physical constraints is constructed in combination with the KL divergence;

[0021] The variational autoencoder parameters are optimized by inversely using a loss function, and the output device-state clean feature tensor is obtained.

[0022] Optionally, the step of inputting the pure feature tensor into the graph spatiotemporal causal inference network, constructing a causal prior graph based on equipment historical full lifecycle failure cases, and strengthening the causal association mining of rare failures through an adversarial sample generation mechanism to generate a failure propagation causal graph with probability weights includes:

[0023] Based on the spatial topological relationships of the pure feature tensor, a device state graph node network is constructed to obtain a spatiotemporal topological graph structure with timestamps;

[0024] Causal rules are extracted from the equipment's historical full lifecycle failure case library to generate a constraint matrix for the causal prior graph;

[0025] The constraint matrix of the spatiotemporal topological graph structure and the causal prior graph is input into the graph spatiotemporal causal inference network. Features are extracted through gated temporal convolution, and the spatiotemporal enhanced feature tensor is output.

[0026] Based on the spatiotemporal enhancement feature tensor, an adversarial sample set corresponding to specific rare fault directions is generated, and the causal correlation features are enhanced by the gradient sign method.

[0027] A Bayesian graph network is trained using enhanced causal association features, and the trained Bayesian graph network is used to output a fault propagation causal graph with probability weights.

[0028] Optionally, the multi-agent diagnosis of the fault propagation causal graph involves deploying competing agents to simulate the fault judgment logic of the equipment designer, maintenance provider, and component manufacturer, respectively. A Nash equilibrium strategy is used to fuse the disputed results from the three parties, outputting a fault diagnosis report with a confidence interval, including fault location and root cause analysis, comprising:

[0029] Based on the node attributes of the fault propagation causal graph, load equipment design specifications, operation and maintenance records, and material databases to generate a third-party knowledge vector group;

[0030] The three knowledge vector groups are input into a competitive group of intelligent agents, 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.

[0031] Aggregate the fault probability vectors of the three parties to detect nodes whose difference value is greater than a preset threshold, and generate a set of points of contention.

[0032] Construct a three-party payoff matrix based on the set of points of contention, and solve for the Nash equilibrium strategy vector;

[0033] Bootstrap sampling is performed on the Nash equilibrium strategy vector to calculate the preset percentage confidence interval for fault location;

[0034] Based on the confidence interval and causal graph, backtrack the root cause paths with probability weights greater than preset weights, and output a diagnostic report with a confidence interval.

[0035] Optionally, the step of mapping the fault diagnosis report to the equipment digital twin in real time, and combining reinforcement learning and Monte Carlo fault tree simulation, to pre-simulate the equipment degradation trajectory under different maintenance strategies in virtual space, and outputting an adaptive maintenance strategy sequence that minimizes the expected value of the total lifecycle maintenance cost, includes:

[0036] Based on the root cause path sequence in the fault diagnosis report, the device degradation model is initialized in the digital twin to obtain the component degradation rate equation.

[0037] A three-dimensional discrete action space is defined based on the component degradation rate equation, and a maintenance strategy search domain is generated.

[0038] In the maintenance strategy search domain, several paths are explored using a proximal strategy optimization algorithm, and the pre-simulated trajectory and cost dataset are output.

[0039] Perform Monte Carlo fault tree analysis on the cost dataset to calculate the expected value of the entire lifecycle operation and maintenance cost;

[0040] Select the optimal strategy based on the expected cost value and output an adaptive maintenance strategy sequence.

[0041] Another embodiment of this application provides an artificial intelligence-based power generation equipment condition fault diagnosis system, the system comprising:

[0042] The acquisition module is used to actively inject mechanical excitation signals of a preset spectrum into key components according to the physical topology of the power generation equipment, and simultaneously acquire the transient response of vibration sensors, infrared thermal imagers, and current transformers under excitation. Combined with environmental parameters and equipment operation logs, it fuses and generates a three-dimensional data volume of time, space, and frequency.

[0043] The separation module is used 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 vector characterizing the health state of the component from the redundant feature vector affected by environmental noise, and output the device state clean feature tensor.

[0044] The module is used to input the pure feature tensor into the spatiotemporal causal reasoning network, construct a causal prior graph based on the equipment's historical full life cycle failure cases, strengthen the causal association mining of rare failures through an adversarial sample generation mechanism, and generate a failure propagation causal graph with probability weights.

[0045] The diagnostic module is used to perform multi-agent diagnosis on the fault propagation causal graph. It deploys competing agents to simulate the fault judgment logic of the equipment designer, the maintenance operator, and the component manufacturer, respectively. It integrates the disputed results of the three parties through the Nash equilibrium strategy and outputs a fault diagnosis report with a confidence interval, which includes fault location and root cause analysis.

[0046] The output module is used to map the fault diagnosis report to the device digital twin in real time. Combining reinforcement learning and Monte Carlo fault tree simulation, it pre-simulates the device degradation trajectory under different maintenance strategies in virtual space and outputs an adaptive maintenance strategy sequence that minimizes the expected value of the total life cycle maintenance cost.

[0047] Another embodiment of this application provides a storage medium storing a computer program, wherein the computer program is configured to execute the method described in any of the preceding claims when running.

[0048] Another embodiment of this application provides an electronic device including 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 method described in any of the preceding claims.

[0049] Compared with existing technologies, this invention provides an artificial intelligence-based method for diagnosing the condition-related faults of power generation equipment. Based on the physical topology of the power generation equipment, it actively injects mechanical excitation signals with a preset spectrum into key components, fusing them to generate a three-dimensional data volume of time, space, and frequency. This three-dimensional data volume is then input into a physically embedded variational autoencoder, outputting a clean feature tensor of the equipment's condition. This clean feature tensor is then input into a graph-temporal-causal inference network to generate a fault propagation causal graph with probability weights. Multi-agent diagnosis is performed on the fault propagation causal graph, outputting a fault diagnosis report with a confidence interval, including fault location and root cause analysis. The fault diagnosis report is mapped in real-time to the equipment's digital twin, outputting an adaptive maintenance strategy sequence that minimizes the expected value of the entire lifecycle operation and maintenance cost. This improves the accuracy and anti-interference capability of fault diagnosis, effectively reducing operation and maintenance costs. Attached Figure Description

[0050] Figure 1 A hardware structure block diagram of a computer terminal for a method for diagnosing the status faults of power generation equipment based on artificial intelligence, provided in an embodiment of the present invention;

[0051] Figure 2 A flowchart illustrating a method for diagnosing the condition faults of power generation equipment based on artificial intelligence, provided in an embodiment of the present invention;

[0052] Figure 3 This is a schematic diagram of a power generation equipment condition fault diagnosis system based on artificial intelligence, provided in an embodiment of the present invention. Detailed Implementation

[0053] The embodiments described below with reference to the accompanying drawings are exemplary and are only used to explain the present invention, and should not be construed as limiting the present invention.

[0054] This invention first provides an artificial intelligence-based method for diagnosing the condition faults of power generation equipment. This method can be applied to electronic devices, such as computer terminals, specifically ordinary computers.

[0055] The following detailed explanation uses a computer terminal as an example. Figure 1 This is a hardware structure block diagram of a computer terminal for a method of diagnosing the condition faults of power generation equipment based on artificial intelligence, provided in an embodiment of the present invention. (See diagram below.) Figure 1 As shown, the computer device includes a processor, memory, and network interface connected via a system bus, wherein the memory may include non-volatile storage media and internal memory.

[0056] Non-volatile storage media can store operating systems and computer programs. These computer programs include program instructions that, when executed, cause the processor to perform any artificial intelligence-based method for diagnosing faults in power generation equipment.

[0057] The processor provides computing and control capabilities, supporting the operation of the entire computer device.

[0058] Internal memory provides an environment for the execution of computer programs in non-volatile storage media. When the computer program is executed by the processor, it enables the processor to perform any artificial intelligence-based method for diagnosing the condition of power generation equipment.

[0059] This network interface is used for network communication, such as sending assigned tasks. Those skilled in the art will understand that... Figure 1 The structure shown is merely a block diagram of a portion of the structure related to the present application and does not constitute a limitation on the computer device to which the present application is applied. Specific computer devices may include more or fewer components than those shown in the figure, or combine certain components, or have different component arrangements.

[0060] It should be understood that the processor can be a Central Processing Unit (CPU), but it can also be other general-purpose processors, digital signal processors (DSPs), application-specific integrated circuits (ASICs), field-programmable gate arrays (FPGAs), or other programmable logic devices, discrete gate or transistor logic devices, discrete hardware components, etc. Among these, a general-purpose processor can be a microprocessor or any conventional processor.

[0061] See Figure 2 The present invention provides an artificial intelligence-based method for diagnosing the condition faults of power generation equipment, which may include the following steps:

[0062] S201, based on the physical topology of the power generation equipment, actively injects mechanical excitation signals with a preset spectrum into key components, and simultaneously collects the transient responses of vibration sensors, infrared thermal imagers, and current transformers under excitation. Combined with environmental parameters and equipment operation logs, it generates a three-dimensional data volume of time, space, and frequency.

[0063] Specifically, based on the bearing stiffness distribution matrix in the physical topology of the power generation equipment, resonance-sensitive areas can be identified, and a coordinate mapping table of key components can be generated.

[0064] Analysis of bearing stiffness distribution matrix

[0065] Analyze the bearing stiffness data in the equipment design drawings to construct a two-dimensional stiffness distribution matrix (i.e., stiffness distribution matrix). The unit of stiffness value is Newtons per millimeter (N / mm). For example, the stiffness value of a generator rotor bearing at an axial position of 1200 mm drops sharply from 8000 N / mm to 5000 N / mm, which is marked as the stiffness abrupt change point.

[0066] Modal analysis experiments were conducted to verify that the vibration frequency response was tested using the hammer impact method. When the vibration frequency reached 85 Hz, the amplitude in this area exceeded the warning value of 0.15 mm, confirming it as a high-risk area for resonance. Simultaneously, the stiffness gradient of adjacent components (such as gearboxes) was detected. If the stiffness difference between adjacent areas exceeded 2000 N / mm, it was simultaneously marked as a sensitive area.

[0067] Generate a coordinate mapping table: The origin of the 3D coordinate system is set at the center of the equipment base (0,0,0), with the axial direction as the X-axis (unit: mm), the radial direction as the Y-axis, and the circumferential direction as the Z-axis. Resonance point coordinates are categorized by risk level; for example, bearing A's coordinates are marked as (X=1200, Y=50, Z=0, risk level: high), and gearbox B's coordinates are marked as (X=800, Y=200, Z=30, risk level: medium). The output table includes component name, coordinate values, maximum amplitude, and recommended monitoring frequency.

[0068] Based on the coordinate mapping table of key components, an amplitude-modulated sweep frequency mechanical excitation signal is injected into a specified position to generate a synchronous trigger pulse sequence;

[0069] Excitation signal parameter design

[0070] Frequency sweep range setting: Based on the frequency characteristics of the resonant sensitive area, a linearly increasing frequency sweep from 50 Hz to 500 Hz is set (sweep speed 5 Hz per second). For the high-frequency band (above 300 Hz), an amplitude attenuation mechanism is used, with an initial amplitude of 5 Newtons, decreasing by 20% for every 100 Hz increase to avoid equipment overload damage. For example, for bearing coordinate point A, an initial excitation of 50 Hz with an amplitude of 5 Newtons is injected, increasing the frequency by 1 Hz every 0.2 seconds for 90 seconds.

[0071] Modulation logic optimization: Superimpose second harmonic excitation (twice the fundamental frequency) in the stiffness abrupt change region, for example, when the fundamental frequency is 85 Hz, simultaneously inject a 170 Hz component (amplitude ratio of 30%) to enhance the excitation effect on microcracks.

[0072] Synchronous trigger pulse generation

[0073] Pulse timing arrangement: Using the excitation signal start point as the time base T0, a pulse sequence with an interval of 10 milliseconds is generated. The vibration sensor is triggered at time T0, the infrared thermal imager is triggered at T0+5 milliseconds, and the current transformer is triggered at T0+8 milliseconds, ensuring that the time alignment deviation of multi-source data is less than 1 millisecond.

[0074] Fault-tolerant retransmission mechanism: If a sensor does not respond, retransmit three pulses within 50 milliseconds; if there is no response for three consecutive times, automatically switch to the backup sensor node and record the alarm log.

[0075] By using a synchronous 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 acquired in parallel to obtain time-aligned multi-source response data packets.

[0076] Multi-source data synchronous acquisition

[0077] Vibration sensor: Acquires time-domain waveforms of axial and radial acceleration at a sampling rate of 10 kHz (i.e., 10,000 data points per second), with a range of ±15 times gravitational acceleration (g). It focuses on recording the amplitude envelope of the resonant frequency band (e.g., 85 Hz ± 5 Hz) and extracting the peak-to-peak value and root mean square (RMS) value.

[0078] Infrared thermal imager: Captures a 256×192 pixel temperature field matrix (spatial resolution 2 mm per pixel), with a temperature accuracy of ±0.5 degrees Celsius. Captures one frame every 5 milliseconds, recording the hotspot coordinates (e.g., coordinates (120,95) with a temperature of 89.5 degrees Celsius) and the temperature rise rate (e.g., 0.8 degrees Celsius per second).

[0079] Current transformer: Acquires the harmonic distortion spectrum of current from 0 to 2000 Hz and decomposes it into 50th harmonic components. Records the total harmonic distortion (THD) and characteristic harmonic amplitudes (e.g., 8.3 amperes for the 5th harmonic) at a sampling rate of 5 kHz.

[0080] Packet encapsulation and alignment

[0081] Timestamp binding: Based on the pulse sequence, each data source is marked with absolute time (accurate to milliseconds). 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 gravitational acceleration), temperature field (hot spot 89.5 degrees Celsius), and current harmonics (5th harmonic 8.3 amperes).

[0082] Data integrity verification: The transmission is verified to be lossless by using Cyclic Redundancy Check (CRC). If the data packet loss rate is greater than 1%, an automatic data re-acquisition process is triggered.

[0083] By combining environmental parameters with the noise floor in the equipment operation log, an adaptive filter is used to separate environmental interference components from multi-source response data packets and extract the intrinsic transient response set of the equipment.

[0084] Environmental noise quantification modeling and baseline library construction

[0085] Noise source parameterization calibration:

[0086] For typical operating environments of power generation equipment, the three core noise sources are quantified:

[0087] Wind turbine aerodynamic noise: Measured using A-weighted (a noise evaluation index that simulates the characteristics of human hearing), with a typical value of 25 dB. For example, a wind turbine at rated speed had a background noise level of 24.8 dB measured 3 meters from the nacelle.

[0088] 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.

[0089] Environmental temperature drift impact: Based on ±2 degrees Celsius per hour, dynamically corrected using meteorological station data. For example, if the ambient temperature at a hydropower station rises from 28°C to 30°C between 8:00 AM and 9:00 AM, the temperature field correction parameters need to be adjusted accordingly.

[0090] Dynamic updates of the noise base library

[0091] Extract historical operating condition data from equipment operation logs and establish a noise-environment mapping relationship:

[0092] Vibration and noise baseline: This indicates the quantitative relationship between ambient temperature and vibration and noise. For example, if the log record is "At 14:00 on May 20, 2025, the vibration and noise baseline of the gearbox is 0.02 times the gravitational acceleration (g) when the ambient temperature is 30℃", this data will be used as the filtering benchmark under the same temperature conditions on that day.

[0093] Electrical interference substrate: Record the harmonic characteristics during power grid anomalies. For example, "During a thunderstorm on June 1, 2025, voltage flicker in the power grid caused the total harmonic distortion (THD) of the current to temporarily rise to 2.0%". Such events need to be marked as special noise samples.

[0094] Multi-source data adaptive filtering and interference separation

[0095] Implementation of the Least Mean Square (LMS) adaptive filtering algorithm:

[0096] Vibration signal processing:

[0097] The input consists of raw vibration data (e.g., axial acceleration of 0.35 times gravitational acceleration at a bearing measuring point) and a noise floor (0.02 times gravitational acceleration). The filter coefficients are updated in real time using the LMS algorithm (an adaptive filtering method that iteratively minimizes the power of the error signal), and the output is the intrinsic vibration component of 0.33 times gravitational acceleration. After filtering, the integrity of the signal waveform needs to be verified, such as eliminating periodic interference caused by the frequency (1.2 Hz) of the wind turbine blades.

[0098] Temperature field correction: Based on raw data from infrared thermal imagers (e.g., the apparent temperature of a generator winding is detected to be 89.5 degrees Celsius), combined with environmental thermal radiation parameters (direct sunlight causing a 3-degree Celsius overestimation), spatial domain filtering is used to eliminate non-equipment heat sources. For example, after correction, the abnormally high temperature in a certain cooling fan area is determined to be 86.5 degrees Celsius.

[0099] Harmonic current purification: Harmonic data collected from current transformers (e.g., Total Harmonic Distortion (THD) 4.5%) is analyzed, and interference components introduced by grid fluctuations (approximately 1.3%) are identified. The true harmonic characteristics of the output device are calculated to have a THD of 3.2%. The amplitudes of the 5th (250 Hz) and 7th (350 Hz) characteristic harmonics are monitored, such as correcting the 5th harmonic from 8.3 amperes to 7.8 amperes.

[0100] Signal quality classification and control

[0101] The signal-to-noise ratio (SNR, the power ratio of signal to noise) threshold is set to 40 dB. For example, if the SNR of a filtered vibration signal is 42 dB (corresponding to a noise power of 0.005 times the gravitational acceleration), it is considered acceptable. If the SNR is ≤ 40 dB (e.g., if the SNR at a remote measuring point is 38 dB due to transmission loss), then wavelet threshold secondary filtering (a signal denoising method based on time-frequency analysis) is triggered until the standard is met.

[0102] Data integrity verification uses Cyclic Redundancy Check (CRC, a coding technique for detecting data transmission errors), and re-collection is initiated when the packet loss rate exceeds 1%. For example, if 95 out of 9000 time slices are lost in a monitoring session, the system will automatically resend the collection command.

[0103] Generation and Verification of Intrinsic Transient Response Sets

[0104] Structured data output specifications

[0105] Time dimension: Absolute timestamps are marked at 10-millisecond intervals, such as "2025-06-17 14:30:25.123".

[0106] Spatial dimension: Positioned according to the topological coordinates of the equipment, such as the position of the front bearing of the generator marked as (X=1200 mm, Y=50 mm, Z=0 mm).

[0107] Physical parameters:

[0108] Vibration: Retain the time-domain waveforms of axial / radial acceleration (sampling rate 10 kHz) and FFT spectrum (0-2000 Hz, resolution 1 Hz).

[0109] Temperature: Record a 256×192 pixel temperature matrix, with hotspot coordinates (120, 95) corresponding to 86.5 degrees Celsius.

[0110] Current: Stores the amplitude and THD value of the 50th harmonic, such as 7.8 amps and 3.2% THD for the 5th harmonic.

[0111] The intrinsic transient response set of the device is recombined into tensors according to the time axis, spatial coordinates, and frequency dimension to output a three-dimensional data volume of time-space-frequency.

[0112] Structured construction of 3D data volumes

[0113] Time axis: Slices are taken at 10-millisecond intervals, covering an excitation cycle of 0-90 seconds, 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 to the actual position (e.g., pixel (100,50) corresponds to 800 mm axially and 100 mm radially).

[0114] Frequency axis: The vibration signal is decomposed into 512 frequency bands (with a resolution of about 1 Hz) by Fast Fourier Transform (FFT); the current harmonics retain the 50th harmonic component (such as the 50 Hz fundamental wave and the inner harmonics of 2500 Hz).

[0115] Tensor Filling Rules and Output

[0116] Multi-source data fusion rules: Vibration spectrum is mapped to physical coordinate points (e.g., the coordinate point of bearing A is 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 (e.g., the coordinate point of generator junction box is filled with harmonic data).

[0117] Output specifications:

[0118] Data format: Hierarchical data format (HDF5);

[0119] Dimensional tags:

[0120] Time dimension: relative time (milliseconds); Spatial dimension: three-dimensional coordinates (millimeters); Frequency dimension: frequency value (hertz).

[0121] Typical data volume size: 9000 (time slices) × 49152 (spatial points) × 562 (frequency + harmonics); data volume: approximately 25 gigabytes (GB).

[0122] First, by analyzing the physical connections of the power generation equipment, the locations of the core components requiring monitoring are determined, and mechanical excitation signals within a specific frequency range are actively applied. During the application of the excitation signal, the dynamic changes of vibration, temperature, and current signals are simultaneously recorded. Combined with parameters such as temperature and humidity of the equipment's environment and historical operating records, all data are integrated into a structured data block according to three dimensions: time, spatial location, and frequency components. This active excitation and simultaneous acquisition of multi-source data overcomes the problems of weak signals and significant interference in traditional passive monitoring. The construction of the three-dimensional data volume provides a complete information foundation encompassing the spatiotemporal dynamic characteristics of the equipment for subsequent analysis, significantly improving the comprehensiveness and accuracy of state perception.

[0123] S202, 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, the intrinsic physical feature vector characterizing the health state of the component and the redundant feature vector affected by environmental noise are separated, and the device state pure feature tensor is output.

[0124] Specifically, a physical regularization module for the variational autoencoder can be constructed based on the device's thermo-mechanical coupling equation.

[0125] Thermo-mechanical physical mechanism modeling and module integration

[0126] For core components of power generation equipment (such as turbine rotors and generator stator windings), a set of calculable physical equations is established based on the thermal expansion characteristics and mechanical deformation laws of materials:

[0127] The heat conduction equation quantifies temperature diffusion behavior. For example, the thermal conductivity of a certain type of gas turbine blade material is set to 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 measured temperature rise rate of the blade surface is 0.8 degrees Celsius per second.

[0128] The stress-strain relationship is based on Hooke's law (the fundamental law of elastic deformation). The elastic modulus of the material is defined as 206 gigapascals. When the centrifugal force on the rotor increases by 100 kilonewtons, the deformation increases by 0.12 millimeters.

[0129] The continuous equations are discretized into a set of finite element nodal equations. For example, the generator stator is divided into 12,000 tetrahedral elements, and each node is associated with three sets of physical quantities: temperature, displacement, and stress, forming a nodal equation matrix.

[0130] Deep embedding of physical regularization module

[0131] Structural Design: A physical computation layer is added to the end of the encoder of the variational autoencoder (a deep learning model that integrates data generation and feature extraction). This layer contains 500,000 adjustable parameters used to map the features extracted by the neural network to the input terms of the physical equation. For example, when the encoder outputs "rotor radial thermal deformation coefficient 0.15", the physical layer automatically converts it into an actual deformation of 0.08 mm.

[0132] Dynamic parameter loading: Establish a real-time material property retrieval mechanism. If the input feature contains the tag "nickel-based alloy 718", its linear expansion coefficient of 13 parts per million per degree Celsius and yield strength of 1100 MPa (megapascals) will be automatically retrieved, and the boundary conditions of the heat conduction equation will be injected.

[0133] Adaptive boundary condition configuration for operating conditions

[0134] Dynamically adjust simulation constraints based on equipment operation logs:

[0135] When the turbine unit detects that the current water head is 85 meters, it automatically sets the runner water pressure boundary to 8.3 MPa.

[0136] 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 conditions.

[0137] 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.

[0138] Standardization of 3D Data Volumes

[0139] Time dimension alignment: All sensor data are resampled at 10-millisecond intervals. Vibration waveform sampling rate is 10 kHz (10,000 sampling points per second), infrared temperature field refresh rate is 1 Hz (once per second), and current harmonic analysis is performed on the 50th harmonic (fundamental 50 Hz). Strict timestamp synchronization is achieved through interpolation algorithms.

[0140] Spatial topology coding: Establishing a three-dimensional grid coordinate system based on the mechanical structure of the equipment. For example, the generator front bearing housing 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), with the grid spacing accurate to the millimeter level.

[0141] Spectrum normalization: The vibration frequency band of 0-2000 Hz is divided into 400 frequency bands. The amplitude of each frequency band is divided by 15 times the full scale of gravitational acceleration (the unit of gravitational acceleration is g). The temperature value is linearly mapped to the 0-1 interval according to the range of 20-150 degrees Celsius.

[0142] Feature decoupling layer operation mechanism

[0143] Five-layer convolutional feature extraction network:

[0144] The first layer of 3×3×3 convolutional kernels (length×width×height) captures local features, such as the 85 Hz component in the bearing vibration signal being enhanced to a feature map brightness value of 0.87 (range 0-1).

[0145] The third layer of void convolution identifies a wide range of correlations, such as the correlation coefficient between winding temperature rise and cooling wind speed reaching 0.92.

[0146] Dual-channel separate design:

[0147] The intrinsic channel outputs a 128-dimensional vector, representing the inherent state of the equipment. Typical values ​​are: [bearing wear index 0.15, insulation aging rate 0.83, shaft alignment deviation 0.07].

[0148] The redundant channel outputs a 64-dimensional vector to record environmental interference. Typical values ​​are: [power grid harmonic residual amount 0.21, environmental temperature drift error 0.04, solar radiation noise 0.33].

[0149] Separation effect verification

[0150] The purity of features was verified using Fast Fourier Transform (a signal spectrum analysis method).

[0151] In terms of intrinsic characteristics, the retention rate of the 85 Hz component of the equipment's characteristic frequency is >95%, while the 1.2 Hz component of the wind turbine blades attenuates to <3%.

[0152] The signal-to-noise ratio (signal-to-noise power ratio) of the redundancy features is strictly controlled below 30 dB to ensure effective removal of interference components.

[0153] The initial intrinsic eigenvectors are simulated using the physical regularization module to generate the theoretical thermodynamic response field.

[0154] Engineering conversion from eigenvectors to physical parameters

[0155] Normalized value restoration:

[0156] "Bearing wear index 0.15" corresponds to an actual radial clearance increase of 0.08 mm; "winding temperature rise slope 0.92" is converted into a real 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).

[0157] Intelligent Material Library Retrieval: When the rotor material is identified as 30CrMoV, its creep limit temperature of 625 degrees Celsius and thermal expansion coefficient of 12 parts per million per degree Celsius are automatically loaded.

[0158] Multiphysics Coupling Solution Process

[0159] Temperature field calculation: Solve the unsteady-state heat conduction equation and output a three-dimensional temperature distribution matrix. Typical case: Simulated temperature of the leading edge of the first-stage moving blade of a gas turbine is 857 degrees Celsius (material temperature limit 900 degrees Celsius), with a temperature gradient of 120 degrees Celsius per millimeter.

[0160] Stress field derivation: Thermal stress was calculated based on temperature field results, and a Mises stress (composite stress index) cloud map was generated by superimposing mechanical centrifugal force. The peak stress at the rotor disc tenon joint is 583 MPa, which is close to the material yield strength of 600 MPa.

[0161] Data remodeling: The finite element mesh data is converted into a 256×192 infrared pixel matrix using a bicubic interpolation algorithm, which is consistent with the dimension of the measured thermal imager data.

[0162] The residual matrix between the theoretical thermodynamic response field and the measured infrared temperature field is calculated, and the loss function of physical constraints is constructed by combining KL divergence.

[0163] Residual diagnosis and anomaly localization

[0164] Pixel-by-pixel comparison: The simulated temperature at coordinates (205, 178) of a boiler superheater tube wall is 586 degrees Celsius, while the actual infrared measurement is 612 degrees Celsius, with an absolute residual of 26 degrees Celsius.

[0165] Spatial clustering analysis: Regions with a residual of more than 20 degrees Celsius for 15 consecutive pixels (approximately 48 square millimeters in area) are marked as "physical law conflict zones", indicating potential material defects or carbon buildup faults.

[0166] Construction of Physical Constraint Loss Function

[0167] Residual probabilistics: Statistical residual distribution histogram, for example, residuals in the 10-20 degree Celsius range account for 35%, and residuals in the >30 degree Celsius range account for 5%.

[0168] KL divergence calculation (a measure of probability distribution difference): Assuming an ideal residual distribution is a zero-mean Gaussian distribution (standard deviation 5 degrees Celsius), calculate the difference between the actual distribution and this distribution. Typical example: KL divergence = 0.45 (target value < 0.1).

[0169] Multi-objective weighted fusion: Physical constraint loss value = mean absolute residual × 0.6 + KL divergence × 0.4.

[0170] Example: Average residual 18.3 degrees Celsius → 18.3 × 0.6 = 10.98; KL divergence 0.45 → 0.45 × 0.4 = 0.18; Total loss value 11.16.

[0171] The variational autoencoder parameters are optimized by inversely using a loss function, and the output device-state clean feature tensor is obtained.

[0172] Adaptive optimization mechanism

[0173] Dynamic learning rate control: Initial value 0.01 (step size for adjusting neural network weights). If the loss value decreases by less than 1% for 5 consecutive iterations, the learning rate is reduced to 80% of the original value.

[0174] Regularization weight adjustment: When the loss value is >10.0, the physical constraint weight is increased to 0.7, and when the loss value is <2.0, it is reduced to 0.3 to balance data fitting and compliance with physical laws.

[0175] Pure Feature Tensor Quality Standard

[0176] Physical consistency verification: Vibration characteristic frequency prediction deviation < ±0.3% (e.g., 85 Hz component error limit ±0.25 Hz); spatial correlation coefficient of temperature field distribution > 0.98 (pixel-level similarity index).

[0177] Data structure specifications: Output tensor size 128×256×256 (feature channel × spatial height × spatial width), feature values ​​located according to device coordinates: [coordinates X=0, Y=0, Z=0]: bearing wear factor 0.83, centering deviation 0.12, lubrication status 0.05; [coordinates X=400, Y=150, Z=30]: insulation degradation 0.67, contact resistance 0.21, heat dissipation efficiency 0.09.

[0178] Circuit breaker protection mechanism: If the loss value does not decrease after 20 consecutive iterations, a "physical constraint failure" alarm (code F3001) is triggered, and the non-convergence feature is saved for fault source analysis.

[0179] A deep learning model incorporating the physical principles of the equipment is used to process 3D data. The model constructs constraints based on physical laws such as heat conduction equations and mechanical equilibrium equations, forcibly maintaining physical plausibility during feature extraction. This process automatically distinguishes between key features reflecting the true state of the equipment and irrelevant features caused by environmental interference. The introduction of physical constraints frees the AI ​​model from its "black box" nature, ensuring that the extracted features have clear physical meaning. This feature purification based on physical mechanisms significantly reduces the false alarm rate, providing reliable feature input for subsequent fault diagnosis.

[0180] S203, input the pure feature tensor into the graph spatiotemporal causal reasoning network, construct a causal prior graph based on the equipment's historical full life cycle failure cases, strengthen the causal association mining of rare failures through an adversarial sample generation mechanism, and generate a failure propagation causal graph with probability weights.

[0181] Specifically, a device state graph node network can be constructed based on the spatial topological relationships of pure feature tensors to obtain a spatiotemporal topological graph structure with timestamps;

[0182] Equipment spatial topology modeling and node definition

[0183] Key component coordinate mapping: Based on the 3D model of the power generation equipment (such as the turbine rotor and generator stator), the spatial coordinates (X, Y, Z) in the 128-dimensional pure feature tensor are bound to the physical locations. For example, the position of the generator front bearing (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, with the distance error between adjacent nodes ≤1 mm.

[0184] Node attribute loading: Health indicators in the associated feature tensor for each node. For example, node N01 is bound to [wear factor 0.83, centering deviation 0.12, lubrication status 0.05], and node N15 (cooling fan) is bound to [dynamic balance index 0.67, bearing temperature rise 0.21].

[0185] Timestamp embedding: At 10-millisecond intervals, a 64-bit timestamp is attached to each node (e.g., 2023-08-20 14:30:25.123) to record the feature acquisition time and form a spatiotemporal four-dimensional coordinate (X,Y,Z,T).

[0186] Building node connection relationships

[0187] Mechanical transmission chain connection: The edge relationship is defined according to the physical structure of the equipment. For example, the rotor node (N05) of the high-pressure cylinder of the steam turbine and the rotor node (N08) of the generator are connected by a coupling, and the edge weight is set to the torque transmission efficiency of 0.98 (dimensionless).

[0188] Thermal Influence Domain Connection: Calculate the temperature field gradient. If the temperature difference between two nodes is greater than 50 degrees Celsius and the distance is less than 300 millimeters, an automatic "heat conduction edge" is established, with a weight of 1 / (distance × temperature difference) (unit: per millimeter per degree Celsius). For example, if nodes N12 (superheater tube) and N13 (support frame) are 200 millimeters apart with a temperature difference of 80 degrees Celsius, the weight is 1 / (200 × 80) = 6.25 × 10⁻⁶. -5 .

[0189] Dynamic temporal correlation: If the phase difference between the vibrations of two nodes is less than 5 milliseconds, a "synchronous vibration edge" is established, with a weight of 1 - phase difference / 10 (dimensionless).

[0190] Graph Structure Optimization and Verification

[0191] Redundant edge pruning: Delete weak connection edges with a weight < 0.01 (such as heat conduction edges with a distance > 500 mm and a temperature difference < 20 degrees Celsius).

[0192] Topology consistency check: Verifies whether the node degree distribution conforms to the physical laws of the equipment. For example, a turbine bearing node should have 3-5 connecting edges. If an isolated node (degree=0) is detected, alarm code G4001 is triggered.

[0193] Causal rules are extracted from the equipment's historical full lifecycle failure case library to generate a constraint matrix for the causal prior graph;

[0194] Fault Case Knowledge Extraction

[0195] Multi-source data association: Analyze 2,000 fault cases from 10 years of operation and maintenance records to establish a "fault phenomenon - 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".

[0196] Physical mechanism regularization: Transforming equipment design specifications into causal constraints. For example, "If the temperature rise of the generator winding is greater than 80 degrees Celsius for 10 minutes, the insulation aging rate will double" is recorded as rule R092.

[0197] Material failure model integration: SN curves (stress-life curves) are generated based on fatigue test data. For example, a certain bearing steel has a fatigue life of 1.2 × 10⁻⁶ under a stress amplitude of 350 MPa. 6 The loop continues.

[0198] Constraint matrix generation algorithm

[0199] Node influence weight calculation: Define causal relationship strength = number of historical co-occurrences / total number of failures. For example, if node N01 (front bearing) and node N08 (generator rotor) co-occur 153 times in rotor imbalance failure, and the total number of failures is 2000, then the causal strength = 153 / 2000 = 0.0765.

[0200] Quantification of propagation delay: Statistical distribution of the time delay from cause node anomaly to effect node anomaly. For example, the average time delay of bearing wear (cause) leading to gearbox vibration (effect) is 48 hours, with a standard deviation of ±3 hours.

[0201] Constraint Matrix Construction: Create an N×N matrix (N is the number of nodes), where matrix elements A[i,j] store: causal strength (0~1); mean delay (hours); standard deviation of delay (hours). Example: A[01,08]=[0.0765, 48, 3].

[0202] Rare Fault Reinforcement Mechanism

[0203] Small sample data augmentation: For rare faults that occur less than 5 times (such as inter-turn short circuits in stator windings), synthetic minority class oversampling technology (a data balancing algorithm) is used to generate 10 times the number of virtual cases.

[0204] Expert rule injection: Manually label implicit causes, such as "cooling water pH value < 6.5 for 30 consecutive days → impeller corrosion rate increases by 300%" is recorded as rule R205.

[0205] The constraint matrix of the spatiotemporal topological graph structure and the causal prior graph is input into the graph spatiotemporal causal inference network. Features are extracted through gated temporal convolution, and the spatiotemporal enhanced feature tensor is output.

[0206] Network architecture design

[0207] Graph convolutional layer: Employs a three-order neighborhood aggregation algorithm, where each node fuses features from its three neighbors. For example, node N01 aggregates features from N02, N08, and N12, with the aggregation weight = 1 / (1 + distance) (unit: per millimeter).

[0208] 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).

[0209] Forget gate filters retain the device's characteristic frequencies (such as the 85 Hz resonant component) unless there is periodic noise (such as random vibrations caused by power grid fluctuations).

[0210] Residual connections: The output of each layer is superimposed on the original input to avoid gradient vanishing.

[0211] Constraint Matrix Application Mechanism

[0212] Causality strength weighting: When clustering neighborhoods, edges with causal strength > 0.05 in the constraint matrix are assigned a weight of 3 times.

[0213] Time Delay Alignment Compensation: Based on the mean time delay in the constraint matrix, the feature sequence is time-shifted. For example, for a node pair with a causal strength of 0.0765, the feature sequence of the result node is aligned 48 hours in advance.

[0214] Conflict resolution: If the deviation between the measured delay and the constraint matrix is ​​greater than twice the standard deviation, the rule update process is triggered.

[0215] Spatiotemporal feature enhancement output

[0216] Feature Dimension Expansion: Input 128-dimensional pure features, process through 5 layers of network, and output 256-dimensional enhanced features.

[0217] Key feature marker: Automatically highlight fault-sensitive features, such as "rotor imbalance index" with a 120% energy increase in the 85 Hz frequency band.

[0218] Data structure specifications: Output tensor size 256×N×T (feature channels×number of nodes×time points), time resolution 10 milliseconds.

[0219] Based on the spatiotemporal enhancement feature tensor, an adversarial sample set corresponding to specific rare fault directions is generated, and the causal correlation features are enhanced by the gradient sign method.

[0220] Rare Fault Targeted Generation

[0221] Fault Mode Focus: Select fault types with a historical occurrence rate of <1% (such as 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).

[0222] Adversarial example generation: Perturb the original features along the V_rare direction: perturbation step size = 0.05 (feature normalization value); generate an adversarial example sequence of length 100; Example: Original features [0.12,0.07,...,0.03] → adversarial example [0.12,0.12,...,0.18] (axial displacement feature enhancement).

[0223] Gradient sign reinforcement mechanism

[0224] Causal correlation feature extraction: Calculate the partial derivative of the fault prediction probability with respect to the input features. P / X.

[0225] Feature saliency ranking: Select | P / The top 10% of features with X | > 0.8 are considered as key causal features.

[0226] Adversarial enhancement: Apply ±0.1 perturbations to key features, generating 50 sets of positive and 50 sets of negative perturbation samples. Positive perturbation: key feature +0.1; negative perturbation: key feature -0.1.

[0227] Optimization of adversarial training

[0228] Hybrid dataset construction: Original samples and adversarial samples are mixed in an 8:2 ratio.

[0229] Loss function improvement: Add a causal consistency penalty term. If the change of non-key features in the adversarial sample is greater than 0.05, add an additional loss value of 0.3.

[0230] Rare fault detection rate improved: After three rounds of adversarial training, the rotor axial movement fault identification rate increased from 68% to 92%.

[0231] A Bayesian graph network is trained using enhanced causal association features, and the trained Bayesian graph network is used to output a fault propagation causal graph with probability weights.

[0232] Bayesian graph network architecture

[0233] Node Conditional Probability Table: 256 states are defined for each node (corresponding to the discretization interval of the enhanced feature), and 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.

[0234] Variational inference engine: Employs stochastic gradient variational Bayesian algorithm (a probabilistic model optimization method), converges after 1000 iterations.

[0235] Prior distribution settings: The node failure rate follows a Beta distribution by default (α=1, β=99), corresponding to a 1% prior failure probability.

[0236] Probability weight calculation

[0237] Causal strength probabilistic: The causal strength in the constraint matrix is ​​converted into conditional probability: P(effect node anomaly | cause node anomaly) = causal strength × 0.95 (confidence scaling factor).

[0238] Delay uncertainty modeling: Transmission delay is modeled as a Gaussian distribution, for example, P(N08 fault delay|N01 anomaly)=Gaussian(48,3) (unit: hours).

[0239] Root cause probability backtracking: The belief propagation algorithm (a graph model reasoning method) is used to calculate the root cause probability in reverse from the leaf nodes: if the temperature of node N12 exceeds the standard, the probability of N01 being the root cause is calculated back to 0.76; if the vibration of N08 exceeds the standard is observed at the same time, the root cause probability increases to 0.91.

[0240] Causal Graph Output Specification

[0241] Graph structure: Directed acyclic graph, node diameter is proportional to failure probability, and edge width is proportional to causal strength.

[0242] Probability labeling: Each node is labeled with its current failure probability (0~1), and each edge is labeled with its propagation delay (mean ± standard deviation).

[0243] Dynamic update mechanism: New feature tensors are received every 10 minutes, and probability values ​​are updated incrementally. The historical data decay factor is 0.95.

[0244] By utilizing graph neural networks to analyze the spatiotemporal correlations of the state characteristics of various equipment components and combining this with causal relationships summarized from a historical failure case database, a logical network for failure propagation is constructed. Specifically for failure types with low probability of occurrence but high severity, data augmentation techniques are used to enhance the learning of their causal features. The causal inference network can not only locate the current failure but also predict potential cascading failure paths. An adversarial training mechanism effectively addresses the problem of insufficient rare failure samples, enabling the system to possess early warning capabilities for events.

[0245] S204, perform multi-agent diagnosis on the fault propagation causal graph, deploy competitive agents to simulate the fault judgment logic of the equipment designer, the maintenance provider, and the component manufacturer respectively, and fuse the disputed results of the three parties through the Nash equilibrium strategy to output a fault diagnosis report with a confidence interval that includes fault location and root cause analysis.

[0246] Specifically, equipment design specifications, operation and maintenance records, and material databases can be loaded based on the node attributes of the fault propagation causal graph to generate a third-party knowledge vector group;

[0247] Multi-source knowledge structured extraction

[0248] Design Specification Analysis: Read the equipment design documents (such as the turbine rotor dynamic balancing standard GB / T 9239) and extract key parameters to form a design knowledge vector. For example, the allowable unbalance limit of 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] (the value after normalization to kilogram-meter units).

[0249] Maintenance record mining: Analyze the 10-year maintenance database and statistically analyze the fault repair records of similar equipment. For example, in the bearing replacement log, lubrication failure accounted for 63%, corresponding to the vector [lubrication failure probability: 0.63]; the average repair time for winding overheating was 4.5 hours, denoted as [repair time: 4.5].

[0250] Material database matching: Retrieve performance parameters based on node material labels. For example, if node N08 (coupling) is labeled "42CrMo alloy steel", then apply its fatigue limit stress of 510 MPa (megapascal) and yield strength of 930 MPa, and generate the vector [fatigue limit: 510, yield strength: 930].

[0251] Knowledge Vector Group Construction Rules

[0252] Dimensional unification: All three dimensions adopt a 128-dimensional format, with missing items padded with zeros. Design-oriented dimensions focus on theoretical parameters (such as a safety factor of 1.8 and a stress concentration factor of 2.3), operation and maintenance-oriented dimensions focus on statistical values ​​(such as a mean time between failures of 8000 hours and a spare parts replacement rate of 0.12), and manufacturing-oriented dimensions include material properties (such as a creep rate of 3×10⁻⁶). -8 (Hourly rate, hardness HRC32).

[0253] Spatiotemporal correlation: Add a time decay weight to each node. For example, the weight of the operation and maintenance record 3 years ago is 0.7, and the weight of the current record is 1.0; design specification version difference coefficient (2015 version coefficient 0.8, 2020 version coefficient 1.0).

[0254] Outlier filtering: Remove data that exceeds the physical feasible range. For example, if a log record says "bearing temperature rise of 300 degrees Celsius" (the actual material melting point is 1500 degrees Celsius), trigger an alarm and replace it with the industry average.

[0255] The three knowledge vector groups are input into a competitive group of intelligent agents, 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.

[0256] Implementation of core algorithms for intelligent agents

[0257] Design agent:

[0258] Stress field finite element simulation: A three-dimensional mesh model is generated based on the design specification vector, and the nodal Mises stress (composite stress index) is calculated. If the simulated stress value of node N01 is 620 MPa > the material yield strength is 600 MPa, then the failure probability = 1 - exp(-(620-600) / 50) = 0.33 (exponential probability mapping function).

[0259] Dynamic operating condition correction: Adjust the load according to the real-time rotation speed. For example, when the centrifugal load coefficient is 1.2 at 1500 rpm, the failure probability is amplified by 1.2 times.

[0260] Operations and maintenance intelligent agent:

[0261] Maintenance history Bayesian inference: Construct a failure probability-maintenance record matrix. For example, node N08 (coupling) has been maintained 6 times in the past 3 years, and the average annual failure rate of similar equipment is 0.8 times. Therefore, the prior probability of failure = 6 / (3×0.8) = 2.5 (the part exceeding the benchmark value is regarded as risk increment), which is compressed to the interval [0,1] by the Sigmoid function to get 0.92.

[0262] Environmental factor compensation: If the current ambient temperature is >35 degrees Celsius (baseline value 25 degrees Celsius), the probability of temperature rise-related failures increases by (35-25)×0.05=0.5 (5% increase in probability per degree Celsius).

[0263] Manufacturing agent:

[0264] Fatigue life model: Miner's linear damage accumulation rule (a fatigue failure prediction method) is adopted. Assuming the current stress amplitude at node N12 (gear) is 300 MPa, the material's SN curve (stress-life curve) shows a life of 10 under this stress. 6 The cycle has run 8×10 5 If the damage is 8 / 10 = 0.8, then the failure probability = damage² = 0.64 (nonlinear acceleration model).

[0265] Material degradation monitoring: If the infrared detector detects that the local temperature at 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%.

[0266] Probability Vector Output Specification

[0267] Dimension Alignment: All three parties output an N-dimensional vector (N is the number of nodes), with element values ​​representing the fault probability in the interval [0,1].

[0268] Time decay factor: The output of the operation and maintenance party needs to be multiplied by the time efficiency coefficient. For example, the data detected last week has a weight of 1.0, and the data from three months ago has a weight of 0.6.

[0269] Uncertainty annotation: The manufacturer adds a ±15% confidence band to account for the dispersion of material properties, and the designer annotates the simulation error with a ±8% deviation range.

[0270] Aggregate the fault probability vectors of the three parties to detect nodes whose difference value is greater than a preset threshold, and generate a set of points of contention.

[0271] Difference detection algorithm

[0272] Node-by-node comparison: Calculate the maximum value of the absolute difference between the three probabilities. Formula: Difference = max(|P_design - P_maint|, |P_design - P_manuf|, |P_maint - P_manuf|).

[0273] Example: The three-way probability of node N01 is [0.33, 0.92, 0.64] → the difference value is max(|0.33-0.92|,|0.33-0.64|,|0.92-0.64|)=0.59.

[0274] Threshold determination: 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.

[0275] Spatial clustering: Cluster the disputed nodes in a neighborhood with a radius of 200 mm to form a disputed region (e.g., the front bearing area of ​​the generator contains nodes {N01, N03, N05}).

[0276] Generate a set of points of contention

[0277] Multi-dimensional Quantification of Controversy:

[0278] Technical route conflict: the designers rely on theoretical simulations, the operators rely on historical data, and the manufacturers focus on material failure.

[0279] Differences in time scales: Operations and maintenance personnel focus on short-term maintainability (hours), while designers consider long-term security (years).

[0280] Focus priority sorting:

[0281] Sort by difference value in descending order; nodes with difference value > 0.5 are marked as "key disputes";

[0282] The weight of the consequences of failures is added, and the priority of nodes that cause downtime losses of more than RMB 1 million per hour is increased by 3 levels.

[0283] Set data structure: Output format is a list [node ID, designer probability, maintenance probability, manufacturer probability, variance value]. Example:

[0284] [N01, 0.33, 0.92, 0.64, 0.59]

[0285] [N08, 0.15, 0.87, 0.21, 0.72].

[0286] Construct a three-party payoff matrix based on the set of points of contention, and solve for the Nash equilibrium strategy vector;

[0287] Payoff matrix modeling

[0288] Definition of dispute node game: Each dispute node is regarded as an independent game, and the three parties are the game participants.

[0289] Strategy space partitioning:

[0290] Design strategy: {Maintain, Don't maintain}

[0291] Operations and maintenance strategy: {Immediate repair, delay monitoring}

[0292] Manufacturer's strategy: {Replace parts, adjust processes}

[0293] Quantization of the payoff function:

[0294] Benefits for strategy design, operation and maintenance, and manufacturing

[0295] (Maintenance, immediate repair, replacement of parts) 0.9 0.7 0.6

[0296] (Maintenance, immediate repair, process adjustment) 0.8 0.9 0.8 ......

[0297] Revenue Calculation Rules:

[0298] Designer's revenue = safety improvement coefficient × 0.6 - cost coefficient × 0.4; Operation and maintenance revenue = availability improvement × 0.7 - maintenance cost × 0.3; Manufacturer's revenue = brand reputation × 0.5 + spare parts profit × 0.5.

[0299] Nash equilibrium solution

[0300] Hybrid strategy equilibrium calculation: The Lemke-Howson algorithm (a game theory solution method) is used to iteratively solve for the optimal response strategy of the three parties.

[0301] Equilibrium Strategy Output:

[0302] Node N01 equilibrium solution: The probability of the designer choosing "maintain" is 0.8, the probability of the operator choosing "immediate repair" is 0.7, and the probability of the manufacturer choosing "replace part" is 0.6; the synthesis strategy vector is [0.8, 0.7, 0.6]; the conflict resolution mechanism is: if the probability of any party in the equilibrium solution is <0.3, the expert arbitration process is triggered, and the minimum guarantee probability of 0.4 is manually set.

[0303] Bootstrap sampling is performed on the Nash equilibrium policy vector to calculate the preset percentage confidence interval for fault location;

[0304] Bootstrap sampling process

[0305] Resampling dataset construction: The original three-way probability vector is used as the population, and 1000 subsets are randomly drawn with replacement (sample size = 80% of the original data size).

[0306] Equilibrium policy recalculation: Repeat the Nash equilibrium solution for each subset of samples to obtain 1000 policy vectors.

[0307] 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.

[0308] Table 1

[0309]

[0310] Confidence interval calculation

[0311] Quantile statistic 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).

[0312] Example: After sorting the probability sequence of the designer, the 25th position = 0.74, the 975th position = 0.85 → confidence interval [0.74, 0.85].

[0313] Node-level interval synthesis: The final failure probability of node N01 = the mean of the equilibrium strategy × the original probability of the node; if the original probability of N01 is 0.33 and the mean of the equilibrium strategy is 0.8, then the synthesized probability = 0.33 × 0.8 = 0.264; Similarly, calculate the confidence interval: original probability interval [0.30, 0.36] × strategy interval [0.74, 0.85] → synthesized interval [0.222, 0.306].

[0314] Based on the confidence interval and causal graph, backtrack the root cause paths with probability weights greater than preset weights, and output a diagnostic report with a confidence interval.

[0315] Root Cause Path Backtracking Algorithm

[0316] Threshold filtering: Select nodes with a failure probability > 0.3 (preset weight) as candidate root causes.

[0317] Cause-effect graph tracing: Traverse the causal edges in reverse from the leaf node (such as the vibration exceeding the standard node N12) and select the upstream node with the highest probability weight.

[0318] Path probability = Product of conditional probabilities along the path × Node's own probability. Example: Path N01→N08→N12, conditional probability P(N08|N01)=0.93, P(N12|N08)=0.87, N01's own probability 0.33 → Path probability = 0.33 × 0.93 × 0.87 = 0.267

[0319] Multi-path competition: Paths with a retention probability > 0.15 are sorted in descending order of probability value.

[0320] Diagnostic Report Generation Standards

[0321] Structured output example:

[0322] Fault location report

[0323] Root cause node: N01 (front bearing);

[0324] Probability: 0.264 [95% confidence interval: 0.222-0.306];

[0325] Transmission path:

[0326] N01→N08→N12 [Path probability: 0.267];

[0327] N01→N05→N12 [Path probability: 0.198];

[0328] Dispute Resolution Instructions

[0329] The designers claim that the probability of maintaining the necessity is 0.8 [0.74-0.85].

[0330] The maintenance team recommends an immediate repair probability of 0.7 [0.65-0.75].

[0331] Manufacturer's solution: Probability of component replacement 0.6 [0.55-0.65].

[0332] Visualization enhancement: Root cause nodes (probability > 0.3) are highlighted in red in the 3D device model, and propagation paths are marked in yellow; confidence intervals are superimposed on the probability histogram in the form of error bars.

[0333] By collaboratively analyzing the fault map using three intelligent agents with different professional perspectives—the designer focusing on theoretical calculations, the operations and maintenance team on practical experience, and the manufacturer on material properties—a consensus is reached among the three parties through game theory. The final diagnostic results include probability assessments and confidence intervals for different viewpoints. This multi-perspective diagnostic mechanism avoids the limitations of single-judgment approaches, while the Nash equilibrium strategy ensures the scientific validity and acceptability of the conclusions. The report with confidence intervals provides a quantitative basis for decision-making, significantly improving the operability of the diagnostic results.

[0334] S205, the fault diagnosis report is mapped to the equipment digital twin in real time. Combining reinforcement learning and Monte Carlo fault tree simulation, the equipment degradation trajectory under different maintenance strategies is pre-simulated in the virtual space, and the adaptive maintenance strategy sequence that minimizes the expected value of the whole life cycle operation and maintenance cost is output.

[0335] Specifically, based on the root cause path sequence in the fault diagnosis report, the equipment degradation model can be initialized in the digital twin to obtain the component degradation rate equation;

[0336] Root cause path analysis and physical mechanism mapping

[0337] Analyze the root cause sequence in the fault diagnosis report (example: front bearing wear → coupling misalignment → excessive gearbox vibration), and establish a degradation transmission chain based on the equipment transmission chain topology. For the root cause node, the front bearing (identification number N01), load the parameters from the material property library: material type is 42CrMo alloy steel, hardness is Rockwell hardness grade 32, and wear resistance coefficient is 1.5×10⁻⁶. -8 (Dimensionless parameter).

[0338] Dynamic load characteristics were extracted through vibration spectrum analysis: the contact stress at the dominant frequency of 85 Hz was 18 megapascals, which, combined with a rotational speed of 1500 rpm, resulted in a sliding speed of 3.2 m / s. The Hertzian contact wear model was applied to construct the equation: Wear change rate = (Dynamic contact stress × Sliding speed) / (Material hardness × Wear resistance coefficient) = (18 × 3.2) / (32 × 1.5 × 10¹⁰) -8 = 0.15 (dimensionless parameter).

[0339] For electrical faults (such as stator winding insulation aging), based on a temperature rise slope of 1.8 degrees Celsius per second, the Arrhenius thermal aging model is used to calculate the rate multiplier: Aging rate multiplier = e^[(activation energy / gas constant)×(1 / reference temperature - 1 / actual temperature)]; Substituting the activation energy of 1.2 electron volts, the gas constant of 8.314 joules per mole per Kelvin, the reference temperature of 80 degrees Celsius (353 Kelvin), and the actual temperature of 110 degrees Celsius (383 Kelvin): Multiplier = e^[(1.2 / 8.314)×(1 / 353-1 / 383)]≈ 4.3 times.

[0340] Multiple Failure Mode Coupling and Environmental Modification

[0341] When components experience concurrent degradation (e.g., bearings simultaneously develop wear and fatigue cracks), a damage superposition equation is established: Total damage = 0.7 × (Current wear amount / Wear threshold) + 0.3 × (Actual stress cycle count / Fatigue life). Weighting coefficients are determined based on historical fault database statistics (e.g., wear accounts for 70%, fatigue accounts for 30%).

[0342] Dynamic correction of environmental factors: If the report indicates an ambient humidity of 90% (exceeding the baseline value of 80%), the corrosion rate will be increased by 5% for every 10% increase in humidity, and the corrosion component will be added by 5%.

[0343] Model validation and parameter calibration

[0344] Input the current measured bearing clearance of 0.52 mm into the digital twin, run the degradation model for 24 hours, and output the predicted value of 0.523 mm.

[0345] The real-time data from the laser micrometer is 0.524 mm, with a relative error of 0.19% (less than the threshold of 1%), indicating successful model validation. If the error exceeds the limit, the wear resistance coefficient will be automatically adjusted (e.g., ±0.1×10). -8 Then recalculate iteratively.

[0346] Output example:

[0347] Part number: N01; Degradation type: contact wear; Rate equation: 0.15 + 0.05 × humidity correction factor (unit: micrometers per hour); Applicable conditions: speed 1200~1800 rpm, temperature <150 degrees Celsius.

[0348] A three-dimensional discrete action space is defined based on the component degradation rate equation, and a maintenance strategy search domain is generated.

[0349] 3D motion space quantization encoding

[0350] Time dimension: Based on the thermal inertia cycle of the equipment, the maintenance window is divided into 4-hour intervals, covering a 720-hour (30-day) cycle, generating 180 discrete time points (sequence: 0, 4, 8, ..., 716).

[0351] Operation type dimension:

[0352] Three basic actions are defined: Maintenance (code B): such as lubricant replenishment, taking 0.5 hours, costing 800 yuan, with the effect of reducing the wear change rate by 30%; Repair (code J): such as bearing clearance adjustment, taking 2 hours, costing 5000 yuan, with the effect of restoring the wear amount to 80% of the initial value; Replacement (code G): such as replacing the entire bearing, taking 8 hours, costing 30000 yuan, with the effect of returning the wear amount to zero.

[0353] Execution intensity dimension: Each type of action is divided into three levels (maintenance intensity level 1: 200 ml oil, level 2: 400 ml, level 3: 600 ml). The difference in intensity will cause the cost to fluctuate by ±20% and the effect to change by ±15%.

[0354] Physical constraints and state pruning

[0355] Condition Feasibility Rules: Replacement is prohibited when wear is <0.1 mm (to avoid over-maintenance); maintenance or replacement is forcibly triggered when insulation aging rate is >5 times the reference value.

[0356] Economic constraints: The daily maintenance cost is capped at 20,000 yuan (if the replacement cost is 30,000 yuan, it must be excluded).

[0357] Timing constraint: Repair of the same component is prohibited within 48 hours after the replacement action.

[0358] Search domain compression: Traverse theoretical combinations (180 time points × 3 operation types × 3 intensity levels = 1620 dimensions); if the current wear change rate is <0.1, delete all replacement actions, and the space is reduced to 1080 dimensions.

[0359] Examples of search domain matrices are shown in Table 2.

[0360] Table 2

[0361]

[0362] In the maintenance strategy search domain, several paths are explored using a proximal strategy optimization algorithm, and the pre-simulated trajectory and cost dataset are output.

[0363] Reinforcement learning agent design

[0364] State representation: Encode the device state as a 128-dimensional vector, including:

[0365] Normalized wear (current value 0.52 → mapped to the [0,1] interval to get 0.65); temperature exceedance ratio (actual 110℃ / threshold 90℃ = 1.22); vibration harmonic energy ratio (85 Hz component accounts for 62% of the total).

[0366] Reward function design: Multi-objective weighted aggregation formula: Comprehensive reward = Equipment availability × 60 - Normalized maintenance cost × 30 - Failure risk coefficient × 10.

[0367] Among them, availability = (total duration - downtime) / total duration × 100; normalized maintenance cost = actual cost / 50,000 yuan; failure risk coefficient = min(damage amount / failure threshold, 1.0).

[0368] Near-end strategy optimization process

[0369] Interactive sampling: The agent executes action ACT0308 (replace bearing strength level 3 at the 240th hour); the twin returns to the new state (insulation aging rate returns to zero) and the reward value is 75.6 (calculation: availability 88%×60 - 30000 / 50000×30 - 0.05×10=52.8-18-0.5=34.3→normalized to 75.6).

[0370] Strategy update mechanism:

[0371] 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 is set to the minimum value (0.15 / 0.10×0.8, 1.2)=1.2); the upper limit of the gradient pruning step size is set to 0.002.

[0372] Pre-simulation trajectory and cost data output

[0373] Explore 50 strategy paths, each containing a sequence of 30 actions:

[0374] Path P09 sequence: Time 0 hours → ACT0101 (Maintenance Intensity 1) → Cost +800 yuan → Wear change rate drops to 0.12; Time 240 hours → ACT0308 (Replacement Intensity 3) → Cost +30000 yuan → Insulation aging reaches zero; Total cost 30800 yuan, availability 88%, failure risk coefficient 0.05.

[0375] Example cost dataset: Path number: 23; Maintenance cost (RMB): 0000; Downtime loss (RMB): 8000; Total cost (RMB): 8000; Availability: 5%.

[0376] Perform Monte Carlo fault tree analysis on the cost dataset to calculate the expected value of the entire lifecycle operation and maintenance cost;

[0377] Fault Tree Event Probability Calibration

[0378] Basic event probability: Based on 10 years of equipment operation and maintenance database statistics: bearing wear failure probability = 6 failures / 10 years = 0.6 per year; insulation breakdown probability = 12 events / 5000 operating days = 0.0024 per day.

[0379] Logic gate rules: OR gate: The parent event is triggered when any sub-event occurs (such as bearing failure or insulation breakdown causing a shutdown); AND gate: The parent event is triggered only when all sub-events occur simultaneously (such as cooling failure and temperature exceeding the limit causing a shutdown).

[0380] Monte Carlo Simulation Process

[0381] Random sampling: Generate 10,000 sets of uniformly distributed random numbers in the [0,1] range.

[0382] Sampling of bearing wear: random number 0.43 < 0.6 → marked as occurred; Sampling of insulation breakdown: random number 0.998 > 0.0024 → marked as not occurred.

[0383] Top event probability calculation: The number of times the unit shutdown event occurred was 1500 → Probability = 1500 / 10000 = 0.15;

[0384] Economic loss correlation: Single downtime loss = 8 hours of downtime × 30,000 yuan per hour = 240,000 yuan; Insulation breakdown loss = 1,200,000 yuan for winding replacement.

[0385] Expected cost synthesis and distribution fitting

[0386] The formula for the expected cost of a single strategy is: Expected total cost = maintenance cost + Σ(probability of the top event × corresponding loss).

[0387] Example path P23: Maintenance cost 48,000 yuan + (0.15×240,000 + 0.002×1200,000) = 48,000 + 36,000 + 2,400 = 86,400 yuan.

[0388] Probability distribution output: Expected value 86,400 yuan, standard deviation ±12,000 yuan; 95% confidence interval [74,400, 98,400] yuan.

[0389] Select the optimal strategy based on the expected cost value and output an adaptive maintenance strategy sequence.

[0390] Multi-objective decision-making model

[0391] Indicator weighting: Expected cost weight 0.65; Equipment availability weight 0.25; Strategy complexity weight 0.10 (evaluated by the number of action types).

[0392] Normalized score calculation: Cost score = (highest cost 150,000 yuan - current cost 86,400 yuan) / (highest cost - lowest cost 50,000 yuan) = 0.636; Availability score = 92% / 100% = 0.92; Complexity score = 1 - number of action types 3 / 10 = 0.7; Overall score = 0.65×0.636 + 0.25×0.92 + 0.10×0.7 = 0.713.

[0393] Adaptive policy sequence generation

[0394] Dynamic response mechanism: Real-time monitoring data deviates from the pre-simulated trajectory by more than 5% (e.g., actual wear amount is 0.55 mm > predicted 0.52 mm), triggering strategy reassessment; when the ambient temperature rises sharply by more than 10 degrees Celsius, switch to the high-temperature strategy library (e.g., increase the intensity of cooling action).

[0395] Sequence output specifications:

[0396] Optimal strategy sequence: 0 hours → maintenance action (B) intensity 1 → wear change rate reduced to 0.12; 48 hours → inspection action (J) intensity 2 → wear amount reset to 0.01 mm; 240 hours → maintenance action (B) intensity 3 → wear change rate reduced to 0.08; total expected cost 86,400 yuan, availability 92%, comprehensive score 0.713.

[0397] Visual verification and economic comparison

[0398] Digital twin mapping: Replacement parts (such as bearings) are marked in red and maintenance parts (such as gearboxes) are marked in yellow in the 3D model; dynamic curves show the remaining lifespan extended from 3 months to 8 months.

[0399] Implementation effect verification: The actual maintenance cost of 85,000 yuan fell within the predicted range of [74,400, 98,400] yuan → verification of effectiveness; a reduction of 32.5% compared to the traditional periodic maintenance strategy (expected cost of 128,000 yuan).

[0400] Based on diagnostic results, the long-term effects of various maintenance plans are simulated in a digital twin model. Reinforcement learning is used to optimize the decision-making path, while probabilistic statistical methods are employed to assess the risks and costs of different plans. Ultimately, an optimal maintenance plan considering the economic benefits throughout the equipment's entire lifecycle is generated. Combining real-time diagnostics with long-term planning achieves closed-loop management from condition monitoring to maintenance decision-making. Virtual simulation avoids the high costs of actual trial and error, and the optimized maintenance strategy can extend equipment life while reducing operating costs.

[0401] As can be seen, based on the physical topology of the power generation equipment, mechanical excitation signals with preset frequencies are actively injected into key components to generate a three-dimensional data volume of time, space, and frequency. The three-dimensional data volume is then input into a physically embedded variational autoencoder, which outputs a pure feature tensor of the equipment state. This pure feature tensor is then input into a graph-temporal causal inference network to generate a fault propagation causal graph with probability weights. Multi-agent diagnosis is performed on the fault propagation causal graph, and a fault diagnosis report with a confidence interval, including fault location and root cause analysis, is output. The fault diagnosis report is mapped to the equipment digital twin in real time, and an adaptive maintenance strategy sequence that minimizes the expected value of the entire life cycle operation and maintenance cost is output. This improves the accuracy and anti-interference capability of fault diagnosis and effectively reduces operation and maintenance costs.

[0402] Another embodiment of the present invention provides a power generation equipment condition fault diagnosis system based on artificial intelligence, see [link to relevant documentation]. Figure 3 The system may include:

[0403] The acquisition module 301 is used to actively inject mechanical excitation signals with a preset spectrum into key components according to the physical topology of the power generation equipment, and simultaneously acquire the transient response of vibration sensors, infrared thermal imagers and current transformers under excitation. Combined with environmental parameters and equipment operation logs, it fuses and generates a three-dimensional data volume of time-space-frequency.

[0404] The separation module 302 is used 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 vector characterizing the health state of the component from the redundant feature vector affected by environmental noise, and output the device state clean feature tensor.

[0405] The construction module 303 is used to input the pure feature tensor into the graph spatiotemporal causal reasoning network, construct a causal prior graph based on the equipment's historical full life cycle failure cases, strengthen the causal association mining of rare failures through an adversarial sample generation mechanism, and generate a failure propagation causal graph with probability weights.

[0406] The diagnostic module 304 is used to perform multi-agent diagnosis on the fault propagation causal graph. It deploys competing agents to simulate the fault judgment logic of the equipment designer, the maintenance operator, and the component manufacturer, respectively. It integrates the disputed results of the three parties through the Nash equilibrium strategy and outputs a fault diagnosis report with a confidence interval, which includes fault location and root cause analysis.

[0407] The output module 305 is used to map the fault diagnosis report to the equipment digital twin in real time, and combine reinforcement learning and Monte Carlo fault tree simulation to pre-enact the equipment degradation trajectory under different maintenance strategies in virtual space, and output the adaptive maintenance strategy sequence that minimizes the expected value of the whole life cycle operation and maintenance cost.

[0408] This invention also provides a storage medium storing a computer program, wherein the computer program is configured to execute the steps in any of the above method embodiments when running.

[0409] Specifically, in this embodiment, the storage medium can be configured to store a computer program for performing the following steps:

[0410] S201, based on the physical topology of the power generation equipment, actively injects mechanical excitation signals with a preset spectrum into key components, and simultaneously collects the transient responses of vibration sensors, infrared thermal imagers, and current transformers under excitation. Combined with environmental parameters and equipment operation logs, it generates a three-dimensional data volume of time, space, and frequency.

[0411] S202, 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, the intrinsic physical feature vector characterizing the health state of the component and the redundant feature vector affected by environmental noise are separated, and the device state pure feature tensor is output.

[0412] S203, input the pure feature tensor into the graph spatiotemporal causal reasoning network, construct a causal prior graph based on the equipment's historical full life cycle failure cases, strengthen the causal association mining of rare failures through an adversarial sample generation mechanism, and generate a failure propagation causal graph with probability weights.

[0413] S204, perform multi-agent diagnosis on the fault propagation causal graph, deploy competitive agents to simulate the fault judgment logic of the equipment designer, the maintenance provider, and the component manufacturer respectively, and fuse the disputed results of the three parties through the Nash equilibrium strategy to output a fault diagnosis report with a confidence interval that includes fault location and root cause analysis.

[0414] S205, the fault diagnosis report is mapped to the equipment digital twin in real time. Combining reinforcement learning and Monte Carlo fault tree simulation, the equipment degradation trajectory under different maintenance strategies is pre-simulated in the virtual space, and the adaptive maintenance strategy sequence that minimizes the expected value of the whole life cycle operation and maintenance cost is output.

[0415] This invention also provides an electronic device, including 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 of the above method embodiments.

[0416] Specifically, the aforementioned electronic device may further include a transmission device and an input / output device, wherein the transmission device is connected to the aforementioned processor, and the input / output device is connected to the aforementioned processor.

[0417] Specifically, in this embodiment, the processor can be configured to perform the following steps via a computer program:

[0418] S201, based on the physical topology of the power generation equipment, actively injects mechanical excitation signals with a preset spectrum into key components, and simultaneously collects the transient responses of vibration sensors, infrared thermal imagers, and current transformers under excitation. Combined with environmental parameters and equipment operation logs, it generates a three-dimensional data volume of time, space, and frequency.

[0419] S202, 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, the intrinsic physical feature vector characterizing the health state of the component and the redundant feature vector affected by environmental noise are separated, and the device state pure feature tensor is output.

[0420] S203, input the pure feature tensor into the graph spatiotemporal causal reasoning network, construct a causal prior graph based on the equipment's historical full life cycle failure cases, strengthen the causal association mining of rare failures through an adversarial sample generation mechanism, and generate a failure propagation causal graph with probability weights.

[0421] S204, perform multi-agent diagnosis on the fault propagation causal graph, deploy competitive agents to simulate the fault judgment logic of the equipment designer, the maintenance provider, and the component manufacturer respectively, and fuse the disputed results of the three parties through the Nash equilibrium strategy to output a fault diagnosis report with a confidence interval that includes fault location and root cause analysis.

[0422] S205, the fault diagnosis report is mapped to the equipment digital twin in real time. Combining reinforcement learning and Monte Carlo fault tree simulation, the equipment degradation trajectory under different maintenance strategies is pre-simulated in the virtual space, and the adaptive maintenance strategy sequence that minimizes the expected value of the whole life cycle operation and maintenance cost is output.

[0423] The above description, based on the embodiments shown in the figures, details the structure, features, and effects of the present invention. The above description is only a preferred embodiment of the present invention, but the present invention is not limited to the scope of implementation shown in the figures. Any changes made in accordance with the concept of the present invention, or equivalent embodiments modified to have equivalent changes, that do not exceed the spirit covered by the specification and figures, should be within the protection scope of the present invention.

Claims

1. A method for diagnosing condition-related faults in power generation equipment based on artificial intelligence, characterized in that, The method includes: Based on the physical topology of the power generation equipment, a pre-defined frequency mechanical excitation signal is actively injected into key components. Simultaneously, the transient responses of vibration sensors, infrared thermal imagers, and current transformers under excitation are acquired. Combined with environmental parameters and equipment operation logs, a three-dimensional time-space-frequency data volume is generated. Specifically, based on the bearing stiffness distribution matrix in the power generation equipment's physical topology, resonance-sensitive areas are identified, and a coordinate mapping table for key components is generated. Based on this table, amplitude-modulated sweep frequency mechanical excitation signals are injected into designated locations to generate a synchronous trigger pulse sequence. Using this sequence, the time-domain waveforms of vibration sensors, the temperature field matrix of infrared thermal imagers, and the harmonic distortion spectrum of current transformers are acquired in parallel to obtain time-aligned multi-source response data packets. Combining environmental parameters and the noise floor in the equipment operation logs, an adaptive filter separates environmental interference components from the multi-source response data packets, extracting the equipment's intrinsic transient response set. The intrinsic transient response set is then tensor-reconstructed according to the time axis, spatial coordinates, and frequency dimensions to output a three-dimensional time-space-frequency data volume. 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. The intrinsic physical feature vector representing the health status of the component and the redundant feature vector affected by environmental noise are separated, and the device state pure feature tensor is output. The pure feature tensor is input into the graph spatiotemporal causal reasoning network. A causal prior graph is constructed based on the equipment's historical full life cycle failure cases. The causal association mining of rare failures is enhanced through an adversarial sample generation mechanism, generating a failure propagation causal graph with probability weights. Multi-agent diagnosis is performed on the fault propagation causal graph. Competitive agents are deployed to simulate the fault judgment logic of the equipment designer, the maintenance operator, and the component manufacturer, respectively. The disputed results of the three parties are fused through the Nash equilibrium strategy, and a fault diagnosis report with a confidence interval, including fault location and root cause analysis, is output. The fault diagnosis report is mapped to the equipment digital twin in real time. By combining reinforcement learning and Monte Carlo fault tree simulation, the equipment degradation trajectory under different maintenance strategies is pre-simulated in virtual space, and the adaptive maintenance strategy sequence that minimizes the expected value of the whole life cycle operation and maintenance cost is output.

2. The method according to claim 1, characterized in that, The process involves inputting the three-dimensional data volume into a physically embedded variational autoencoder, constructing a physical constraint loss function using the device's thermal-mechanical coupling equation, separating the intrinsic physical feature vector characterizing the component's health state from the redundant feature vector affected by environmental noise, and outputting a device state-clean feature tensor, including: Based on the thermo-mechanical coupling equation of the device, 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. The initial intrinsic eigenvectors are simulated using the physical regularization module to generate the theoretical thermodynamic response field. The residual matrix between the theoretical thermodynamic response field and the measured infrared temperature field is calculated, and the loss function of physical constraints is constructed by combining KL divergence. The variational autoencoder parameters are optimized by inversely using a loss function, and the output device-state clean feature tensor is obtained.

3. The method according to claim 2, characterized in that, The process involves inputting the pure feature tensor into a graph-spatiotemporal causal inference network, constructing a causal prior graph based on historical failure cases throughout the device's lifecycle, and enhancing the causal association mining of rare failures through an adversarial sample generation mechanism to generate a failure propagation causal graph with probability weights, including: Based on the spatial topological relationships of the pure feature tensor, a device state graph node network is constructed to obtain a spatiotemporal topological graph structure with timestamps; Causal rules are extracted from the equipment's historical full lifecycle failure case library to generate a constraint matrix for the causal prior graph; The constraint matrix of the spatiotemporal topological graph structure and the causal prior graph is input into the graph spatiotemporal causal inference network. Features are extracted through gated temporal convolution, and the spatiotemporal enhanced feature tensor is output. Based on the spatiotemporal enhancement feature tensor, an adversarial sample set corresponding to specific rare fault directions is generated, and the causal correlation features are enhanced by the gradient sign method. A Bayesian graph network is trained using enhanced causal association features, and the trained Bayesian graph network is used to output a fault propagation causal graph with probability weights.

4. The method according to claim 3, characterized in that, The process involves multi-agent diagnosis of the fault propagation causal graph, deploying competing agents to simulate the fault judgment logic of the equipment designer, maintenance provider, and component manufacturer, respectively. A Nash equilibrium strategy is used to fuse the disputed results from the three parties, outputting a fault diagnosis report with a confidence interval, including fault location and root cause analysis. Based on the node attributes of the fault propagation causal graph, load equipment design specifications, operation and maintenance records, and material databases to generate a third-party knowledge vector group; The three knowledge vector groups are input into a competitive group of intelligent agents, 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 fault probability vectors of the three parties to detect nodes whose difference value is greater than a preset threshold, and generate a set of points of contention. Construct a three-party payoff matrix based on the set of points of contention, and solve for the Nash equilibrium strategy vector; Bootstrap sampling is performed on the Nash equilibrium strategy vector to calculate the preset percentage confidence interval for fault location; Based on the confidence interval and causal graph, backtrack the root cause paths with probability weights greater than preset weights, and output a diagnostic report with a confidence interval.

5. The method according to claim 4, characterized in that, The process involves mapping the fault diagnosis report to the equipment's digital twin in real time, combining reinforcement learning and Monte Carlo fault tree simulation to pre-simulate 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 total lifecycle maintenance cost, including: Based on the root cause path sequence in the fault diagnosis report, the device degradation model is initialized in the digital twin to obtain the component degradation rate equation. A three-dimensional discrete action space is defined based on the component degradation rate equation, and a maintenance strategy search domain is generated. In the maintenance strategy search domain, several paths are explored using a proximal strategy optimization algorithm, and the pre-simulated trajectory and cost dataset are output. Perform Monte Carlo fault tree analysis on the cost dataset to calculate the expected value of the entire lifecycle operation and maintenance cost; Select the optimal strategy based on the expected cost value and output an adaptive maintenance strategy sequence.

6. A power generation equipment condition fault diagnosis system based on artificial intelligence, characterized in that, The system includes: The acquisition module is used to actively inject mechanical excitation signals of a preset spectrum into key components based on the physical topology of the power generation equipment. It simultaneously acquires the transient responses of vibration sensors, infrared thermal imagers, and current transformers under excitation. Combined with environmental parameters and equipment operation logs, it fuses these signals 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, it identifies resonance-sensitive areas and generates a coordinate mapping table for key components. Based on this table, it injects amplitude-modulated swept-frequency mechanical excitation signals into designated locations, generating a synchronous trigger pulse sequence. Using this sequence, it acquires the time-domain waveforms of vibration sensors, the temperature field matrix of infrared thermal imagers, and the harmonic distortion spectrum of current transformers in parallel, obtaining time-aligned multi-source response data packets. Combining environmental parameters and noise floor data from the equipment operation logs, it separates environmental interference components from the multi-source response data packets using an adaptive filter, extracting the equipment's intrinsic transient response set. Finally, it reassembles the intrinsic transient response set using tensors along the time axis, spatial coordinates, and frequency dimensions, outputting a three-dimensional time-space-frequency data volume. The separation module is used 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 vector characterizing the health state of the component from the redundant feature vector affected by environmental noise, and output the device state clean feature tensor. The module is used to input the pure feature tensor into the spatiotemporal causal reasoning network, construct a causal prior graph based on the equipment's historical full life cycle failure cases, strengthen the causal association mining of rare failures through an adversarial sample generation mechanism, and generate a failure propagation causal graph with probability weights. The diagnostic module is used to perform multi-agent diagnosis on the fault propagation causal graph. It deploys competing agents to simulate the fault judgment logic of the equipment designer, the maintenance operator, and the component manufacturer, respectively. It integrates the disputed results of the three parties through the Nash equilibrium strategy and outputs a fault diagnosis report with a confidence interval, which includes fault location and root cause analysis. The output module is used to map the fault diagnosis report to the device digital twin in real time. Combining reinforcement learning and Monte Carlo fault tree simulation, it pre-simulates the device degradation trajectory under different maintenance strategies in virtual space and outputs an adaptive maintenance strategy sequence that minimizes the expected value of the total life cycle maintenance cost.

7. A storage medium, characterized in that, The storage medium stores a computer program, wherein the computer program is configured to execute the method of any one of claims 1-5 when it is run.

8. An electronic device comprising a memory and a processor, characterized in that, The memory stores a computer program, and the processor is configured to run the computer program to perform the method of any one of claims 1-5.

Images