Real-time control system for power equipment based on supercomputer CT three-dimensional imaging

Abstract

The present application relates to the technical field of circuit device or system of power supply or power distribution, and relates to a real-time control system of power equipment based on supercomputer CT three-dimensional imaging.The present application generates a physical state transition probability database by obtaining equipment parameters; triggers a tomographic scan to obtain a raw data stream of ray attenuation, extracts a coherent spectrum feature vector in combination with the database; generates a suppressive power quality disturbance waveform sequence injection power grid through a physical electrical coupling reverse deduction model; captures a digital trigger pulse at an extreme value moment, synchronously calls a scanning array to perform transient relocation focusing, and obtains a high signal-to-noise ratio transient attenuation data stream; and finally compares the differences in microscopic responses before and after interference to modify waveform parameters in a closed loop.The present application establishes a feedback link from microscopic physical observation to macroscopic electrical control, realizes high-precision dynamic observation and active physical intervention on the internal microscopic failure process, and constructs an adaptive closed-loop control system.

Description

Technical Field

[0001] This invention relates to the field of circuit devices or systems for power supply or distribution, and to a real-time control system for power equipment based on supercomputing CT three-dimensional imaging. Background Technology

[0002] Large-scale power equipment, such as power transformers, gas-insulated switchgear, and high-voltage instrument transformers, are the cornerstone of maintaining the safe and stable operation of modern power grids. These devices are subjected to the combined effects of high electric fields, high temperatures, and strong mechanical stresses over long periods, inevitably leading to gradual degradation of the physical properties of their internal insulating media and conductive structures. This degradation often begins at the microscopic level, such as the formation of tiny bubbles in insulating oil, localized carbonization paths within solid insulating media, or the loosening of preload in metal fasteners due to creep. These microscopic defects typically do not exhibit obvious external electrical characteristics in their early stages, but their continuous growth and accumulation are the main root cause of sudden and catastrophic equipment failures. Therefore, the core challenge and pain point currently facing the field of intelligent operation and maintenance of power systems lies in how to penetrate the thick metal casing of equipment, accurately and in real time grasp the dynamic evolution of the microscopic physical state of internal components, and effectively intervene before the fault evolves into an irreversible state.

[0003] To address the aforementioned challenges, the industry's current prevalent technologies primarily revolve around condition-based monitoring and prediction. One mainstream approach is online monitoring and fault diagnosis based on multi-source sensor information. This approach involves deploying various types of sensors externally or at key locations on the equipment, such as chromatographic sensors for analyzing dissolved gases in oil, ultra-high frequency sensors or acoustic sensors for detecting partial discharge signals, and infrared thermal imagers for monitoring surface temperature distribution. The control center comprehensively analyzes the data streams collected by these sensors and, based on a pre-defined fault feature library or expert system rules, indirectly infers potential abnormal states within the equipment. Another common technique is condition prediction based on digital twin technology. This method constructs a high-fidelity digital model of the physical equipment, inputting real-time operating parameters such as load current and ambient temperature. Through multiphysics simulation, it extrapolates the stress distribution and aging rate within the equipment, thereby predicting its health status and remaining lifespan.

[0004] However, the aforementioned traditional solutions still have inherent technical drawbacks when dealing with complex and ever-changing actual operating environments and early microscopic defects. Sensor-based monitoring methods are essentially "symptom" detection, detecting physical or chemical byproducts generated after microscopic defects develop to a certain stage, such as discharge pulses, abnormal gases, or localized overheating. This leads to an unavoidable time lag in fault detection, and it is difficult to accurately locate defects in three-dimensional space, let alone provide direct physical information on the morphology, size, and dynamic evolution of defects. Digital twin-based prediction methods heavily rely on the initial fidelity of the model and the accuracy of physical laws. During long-term operation of equipment, the actual material aging path and the generation of random defects often cause deviations between the physical entity and the digital model. Due to the lack of direct observation data on the actual physical state inside the equipment for model correction, the reliability of the prediction results decreases over time. Most importantly, both solutions remain at the level of "monitoring" and "alarming," lacking the technical means to actively intervene physically to inhibit the continued development of identified early microscopic defects.

[0005] Based on the above problems, the present invention aims to solve the technical problems in the prior art that it is difficult to perform real-time, high-precision imaging observation of the dynamic process of micro-physical failure inside power equipment, and that there is a lack of active intervention methods to inhibit the continuous development of the failure process. Summary of the Invention

[0006] In view of this, in order to solve the problems mentioned in the background technology, a real-time control system for power equipment based on supercomputing CT three-dimensional imaging is proposed.

[0007] The objective of this invention can be achieved through the following technical solution: This invention provides a real-time control system for power equipment based on supercomputing CT three-dimensional imaging, including: a probability database generation module, used to obtain the basic design parameters of the power equipment, perform physical state degradation simulation, and obtain a physical state transition probability database.

[0008] The attenuation raw data stream generation module is used to trigger a tomographic scanning array deployed around the power equipment to perform a transillumination scan of the power equipment and obtain the X-ray attenuation raw data stream.

[0009] The feature vector generation module is used to acquire the raw data stream of ray attenuation, and to perform frequency domain component conversion by calling the parallel computing power of the supercomputing cluster based on the physical state transition probability database to obtain the coherent spectrum feature vector.

[0010] The disturbance waveform sequence generation module is used to obtain the coherence spectrum feature vector and perform causal analysis calculation through the physical-electrical coupling inverse deduction model to obtain the suppressive power quality disturbance waveform sequence.

[0011] The digital trigger pulse generation module is used to inject the suppressive power quality disturbance waveform sequence into the grid-connected node, perform extreme moment capture, and obtain the digital trigger pulse.

[0012] The transient attenuation data stream generation module is used to acquire digital trigger pulses and synchronously call the tomographic scanning array to perform transient repositioning and focusing operations through digital trigger pulses, thereby obtaining a high signal-to-noise ratio transient attenuation data stream.

[0013] The waveform parameter sequence generation module is used to acquire a high signal-to-noise ratio transient decay data stream, compare the microscopic response difference between the high signal-to-noise ratio transient decay data stream and the coherent spectrum feature vector, and obtain the corrected waveform parameter sequence.

[0014] Compared with the prior art, the embodiments of the present invention have at least the following advantages or beneficial effects: (1) The present invention constructs a physical state transition probability database that characterizes the dynamic degradation trend of the microstructure of the component, and combines the ray attenuation data stream of sparse tomography with spatiotemporal frequency domain analysis to extract the coherent spectrum feature vector encoding the growth of micro-physical failures. This can upgrade the traditional tomography technology from static defect “photographing” to “videography” of the dynamic evolution process of micro-defects. The physical state transition probability database establishes a quantitative mapping relationship between the ray attenuation gradient and specific physical failure modes, such as carbonization or peeling, so that the system can accurately identify suspicious regions of interest from massive amounts of raw data. Furthermore, by applying frequency band analysis to the spatiotemporal data segment passing through the region, the system can demodulate the unique high-frequency signal fingerprint caused by the periodic growth or oscillation of micro-defects from complex background noise. Furthermore, this method forms the technical basis for direct, real-time imaging and dynamic tracking of the invisible microscopic physical degradation process inside the equipment, realizing high-precision quantitative characterization of the physical behavior characteristics in the incipient stage of a fault. Its spatial positioning accuracy and dynamic capture capability have better technical performance than traditional monitoring methods based on indirect signals.

[0015] (2) This invention introduces a physical-electrical coupling reverse deduction model to perform causal analysis on the coherent spectrum feature vector and generate a suppressive power quality disturbance waveform sequence for physical field interference, thereby transforming passive state monitoring into active physical intervention. The core mechanism of the physical-electrical coupling reverse deduction model is that it reversely solves the coherent spectrum feature vector characterizing the physical dynamics of defects into specific electric field parameters that can produce destructive interference physical effects; subsequently, through execution units such as local active power filters, these parameters are transformed into precise voltage disturbance waveforms that can be injected into the power grid. The alternating electric field generated by this disturbance waveform in the local area of ​​the equipment can apply a reverse, suppressive electrostrictive force or Coulomb torque to the defect growth process such as oscillating discharge of small air gaps or dielectric polarization wear. Furthermore, this invention establishes a direct feedback link from real-time observation of microscopic physical state to macroscopic electrical control input, realizing active suppression and "physical stabilization" of internal microscopic defect growth, and optimizing the traditional operation and maintenance mode that can only passively wait for faults to occur or perform shutdown maintenance.

[0016] (3) This invention utilizes the digital trigger pulse generated at the extreme moment of the injected waveform to synchronously invoke the sparse tomographic scanning array to perform transient repositioning and focusing operations. By comparing the high signal-to-noise ratio transient attenuation data stream before and after interference, and iteratively correcting the control model in a closed loop, an adaptive closed-loop control system for precise regulation of microscopic physical processes can be constructed. The digital trigger pulse ensures that the moment of verification scanning is strictly synchronized with the moment of strongest physical interference, while the cross-focusing scanning strategy improves the signal-to-noise ratio of the verification signal by several times, thereby enabling the precise capture of the weak physical response changes generated by interference. By quantitatively calculating the characteristic degradation attenuation ratio before and after interference, the system can accurately evaluate the effectiveness of the current control strategy and, based on the deviation residual from the expected stabilization index, reversely correct the weight matrix of the physical-electric coupling reverse deduction model. Furthermore, this constitutes a complete "observation-intervention-verification-optimization" closed loop, making the control strategy no longer a static open-loop execution, but capable of dynamic and continuous self-optimization based on real physical feedback, effectively improving the robustness, accuracy, and long-term effectiveness of the control action. Attached Figure Description

[0017] To more clearly illustrate the technical solutions of the embodiments of the present invention, the accompanying drawings used in the description of the embodiments will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0018] Figure 1 This is a schematic diagram of the system module structure connection of the present invention.

[0019] Figure 2 This invention provides a synchronization timing mechanism for the injection of physical interference waveforms and the triggering of transient extrema.

[0020] Figure 3 The curve is the adaptive optimization iteration curve of the closed-loop residual evaluation feedback and control model of the present invention. Detailed Implementation

[0021] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0022] Please see Figure 1 The present invention provides a real-time control system for power equipment based on supercomputing CT three-dimensional imaging, including: a probability database generation module, an attenuation raw data stream generation module, a feature vector generation module, a disturbance waveform sequence generation module, a digital trigger pulse generation module, a transient attenuation data stream generation module, and a waveform parameter sequence generation module.

[0023] The probability database generation module and the attenuation raw data stream generation module are both connected to the feature vector generation module. The feature vector generation module is connected to the disturbance waveform sequence generation module. The disturbance waveform sequence generation module is connected to the digital trigger pulse generation module. The digital trigger pulse generation module is connected to the transient attenuation data stream generation module. The feature vector generation module and the transient attenuation data stream generation module are both connected to the waveform parameter sequence generation module.

[0024] The probability database generation module is used to obtain the basic design parameters of power equipment, perform physical state degradation simulation, and obtain a physical state transition probability database.

[0025] In a specific embodiment of the present invention, the basic design parameters of the power equipment are obtained, physical state degradation simulation is performed, and a physical state transition probability database is obtained. This includes: obtaining geometric topology information and material factory test parameters from the basic design parameters, performing registration fusion and mesh generation, and obtaining a high-fidelity digital twin benchmark model.

[0026] A high-fidelity digital twin baseline model is obtained, and a preset accelerated aging load sequence is applied to obtain multiple discrete physical degradation stages.

[0027] Obtain the density and elemental composition of the material during the physical degradation stage, calculate the attenuation effect of the virtual high-energy ray beam passing through the component, and obtain the ray attenuation coefficient corresponding to the current physical state.

[0028] Obtain the ray attenuation coefficients of two adjacent physical degradation stages, calculate the difference between the ray attenuation coefficients, and obtain the change in the ray attenuation coefficient.

[0029] Obtain the change in physical state between two adjacent physical degradation stages, and divide the change in ray attenuation coefficient by the change in physical state to obtain the gradient of ray attenuation coefficient.

[0030] The gradient of the ray attenuation coefficient is obtained, and the gradient of the ray attenuation coefficient is structurally correlated with the spatial location of the component, the material type and the degradation mode to obtain a physical state transition probability database.

[0031] Specifically, the implementation process for acquiring the basic design parameters of power equipment, constructing a basic twin architecture, and establishing a physical state transition probability database characterizing the dynamic degradation trend of component microstructures is as follows: The data processing server of the supercomputing cluster first loads the target power equipment, such as a 3D computer-aided design model file of a power transformer with a rated voltage of 500kV. The file format is STEP or IGES. The server decompiles the model file, extracting the precise geometric topology information and spatial coordinate data of all core components inside the equipment, including windings, core, insulating paperboard, and fastening bolts. At the same time, the server retrieves the material factory test parameters that precisely correspond to each component from the material database provided by the manufacturer. These parameters cover material density, Young's modulus, dielectric constant, thermal conductivity, and initial X-ray linear attenuation coefficient. Subsequently, the server invokes the finite element analysis software interface to register and fuse the geometric topology information with the material factory test parameters in three-dimensional space. Through adaptive mesh generation technology, the entire device is discretized into hundreds of millions of tetrahedral or hexahedral mesh elements, each assigned specific physical properties, thus constructing a high-fidelity digital twin benchmark model that is highly consistent with the physical entity in terms of geometric dimensions and material physical properties. Based on this high-fidelity digital twin benchmark model, the supercomputing cluster's data processing server initiates a multiphysics coupling simulation program. For predefined aging-sensitive components in the model, particularly the insulating medium subjected to high electric fields and thermal stress, and the fastening conductive components subjected to mechanical stress, a preset accelerated aging load sequence is applied. The simulation program simulates a series of discrete physical degradation stages in the time domain, from the device's brand-new state to its end-of-life. For example, the carbonization degree of the insulating cardboard increases linearly from 0% to 30%, or the preload loss of copper fasteners due to creep increases from 0% to 15%. In each discrete physical degradation stage, the simulation program calculates the attenuation effect of a virtual high-energy X-ray beam passing through the component using the Beer-Lambert law, based on the material's density and elemental composition at that state, precisely determining the X-ray attenuation coefficient uniquely corresponding to the current physical state. By comparing the X-ray attenuation coefficients of two adjacent physical degradation stages, the X-ray attenuation coefficient gradient characterizing the degradation rate of that stage is calculated. Finally, the supercomputing cluster's data processing server aggregates the X-ray attenuation coefficient gradients of all aging-sensitive components across all physical degradation stages, structurally associating these gradient values ​​with the component's spatial location, material type, and specific degradation mode (e.g., carbonization, moisture absorption, or mechanical delamination) that triggered the change. The server constructs a key-value database structure indexed by the X-ray attenuation coefficient gradient values, where the value portion is a list containing the potential physical failure modes indicated by the gradient value, along with the probability of occurrence of each mode calculated based on extensive simulation statistics or historical failure case data.Once this process is completed, a physical state transition probability database is generated to indicate the carbonization or stripping response of a specific material in subsequent steps.

[0032] In this step, the core calculations involve the physical laws governing ray attenuation and solving for the degradation gradient. The ray attenuation process follows the Beer-Lambert law, whose mathematical expression is:

[0033]

[0034] The gradient of the ray attenuation coefficient characterizing the degradation rate Then it is calculated using the following formula:

[0035]

[0036] in, Indicates the intensity of the radiation after it penetrates matter; The initial radiation intensity before incident; It is the base of the natural logarithm; The radiation attenuation coefficient of a material is typically expressed in units of 1000 ppm. It is usually set to 0.250. ; It is the thickness through which the rays penetrate the material. In the gradient calculation formula, For the first The gradient of the ray attenuation coefficient at each physical degradation stage; and The first The and the first The radiation attenuation coefficient calculated for each physical degradation stage; Representative from the first The first stage to the second The change in physical state at each stage is a dimensionless normalized parameter, for example, a change of five percent in the degree of carbonization.

[0037] In the above formula, the radiation attenuation coefficient The dimension of is the reciprocal of length, for example Penetration thickness The dimension of the exponent is length, for example, cm, therefore the exponent term... Since it is a dimensionless pure number, the physical dimensions of both sides of the formula, namely ray intensity, are guaranteed. and The dimensions are consistent. In the gradient... In the calculation, the molecule The dimensions are denominator Since it is a dimensionless parameter, the gradient The dimensions of are consistent with those of the X-ray attenuation coefficient, both being 1. The entire calculation process is physically self-consistent.

[0038] A high-fidelity digital twin benchmark model refers to a digital simulation model that integrates the precise three-dimensional geometry of the equipment, the distribution of material physical properties, and initial operating state parameters. Its fidelity requirement is to reproduce the key behaviors of the equipment in terms of macroscopic electrical characteristics and microscopic material responses. Physical degradation stages are a series of predefined discrete time points or state points representing the evolution of equipment components from a healthy state to a failure state, used to simulate the gradual aging process in simulation. The ray attenuation coefficient is a physical quantity characterizing a material's ability to absorb rays of specific energy, directly related to the material's density and atomic number. The ray attenuation coefficient gradient is the ratio of the change in the ray attenuation coefficient to the change in physical state between two adjacent physical degradation stages; this gradient directly reflects the rate and mode of material aging. The physical state transition probability database is a storage structure that establishes a probabilistic mapping relationship between observable ray attenuation data features, i.e., the ray attenuation coefficient gradient, and invisible microscopic physical states within the equipment, such as carbonization and delamination failure modes. Regarding the setting of simulation parameters, assuming the application scenario is to monitor the aging of the oil-paper insulation inside the transformer, the simulation upper limit of its carbonization degree is set at 30%. This is based on industry operating procedures. If the insulation paper exceeds this aging degree, its mechanical and electrical properties will experience an irreversible and rapid decline, which is usually regarded as the critical point that requires maintenance or replacement. Therefore, this setting has a clear engineering practice basis.

[0039] For example, suppose the target power equipment is a 220kV high-voltage instrument transformer, whose core component is an epoxy resin insulated tie rod. The supercomputing cluster's data processing server loads the tie rod's STEP format 3D model and material properties, where the initial density of the epoxy resin is 1.2 g / cm³ and the initial X-ray attenuation coefficient is 0.250. The server constructed a high-fidelity digital twin baseline model containing 500,000 hexahedral elements using finite element mesh generation. Subsequently, the simulation program simulated its carbonization process under long-term electrical stress, setting six discrete physical degradation stages corresponding to carbonization degrees of 0%, 2%, 4%, 6%, 8%, and 10%, respectively. The simulation calculated the X-ray attenuation coefficients for each stage. The values ​​are 0.250 in sequence. 0.254 0.259 0.265 0.272 0.280 Therefore, the gradient of the ray attenuation coefficient between adjacent stages can be calculated. For example, from stage 1 (2% carbonization) to stage 2 (4% carbonization), the change in physical state is 2%, and the corresponding gradient is... The server aggregates the calculation results from all stages and generates a series of mapping entries in the physical state transition probability database. For example, it generates a record stating that when the gradient value of the ray attenuation coefficient in a certain region is detected... At that time, the database query result was {Physical state: Primary carbonization of epoxy resin, probability of occurrence: 0.92}. By performing such calculations on all components and all degradation modes, a complete database was ultimately formed.

[0040] The attenuation raw data stream generation module is used to trigger a tomographic scanning array deployed around the power equipment to perform a transillumination scan of the power equipment and obtain the X-ray attenuation raw data stream.

[0041] In a specific embodiment of the present invention, a tomographic scanning array deployed around the power equipment is triggered to perform a transillumination scan on the power equipment and obtain a raw data stream of X-ray attenuation. This includes: acquiring a synchronous digital start command, activating multiple high-energy collimated X-ray sources deployed around the power equipment through the synchronous digital start command, and obtaining a pulse detection field.

[0042] A preset scanning strategy is obtained, and the high-energy collimated X-ray source is controlled to output a high-energy X-ray beam to obtain the transmitted X-ray that penetrates the power equipment.

[0043] A preset time-division multiplexed scanning sequence is acquired. The detector array distributed on the opposite side of the high-energy collimated X-ray source is activated by polling the time-division multiplexed scanning sequence. The photon flux of the transmitted X-ray is collected and converted from analog to digital to obtain the raw data stream of X-ray attenuation.

[0044] Specifically, the process of triggering a sparse tomography array deployed around power equipment to scan the internal physical medium of the target power equipment and acquire raw data streams of X-ray attenuation is as follows: The tomography control system deployed on-site first activates multiple, typically 3 to 8, high-energy collimated X-ray sources surrounding the target power equipment, such as the metal casing of a 220kV high-voltage transformer. This activation is accomplished by sending a synchronous digital start command, accurate to the microsecond level, ensuring that all X-ray sources are in standby mode, thereby forming a potential pulse detection field with its spatial location precisely calibrated around the equipment. Next, according to a preset scanning strategy, the tomography control system outputs a set of precise control parameters to the selected high-energy collimated X-ray source, such as an industrial-grade electron linear accelerator, through a digital-to-analog conversion circuit. These parameters include setting its anode voltage to the range of 6 million to 9 million electron volts, and setting the pulse width of its pulse forming network to 1-4 microseconds and the pulse repetition frequency to 200-400 Hz. Driven by these parameters, the high-energy collimated X-ray source outputs a high-energy, narrow-beam, pulsed high-energy X-ray beam shaped by a tungsten alloy collimator. This beam is precisely projected, its path pre-planned by a high-fidelity digital twin reference model to ensure unimpeded penetration of the target electrical equipment's metal casing, insulating oil, windings, and epoxy resin-insulated tie rods—complex internal media and metal components. Simultaneously with beam emission, the tomographic scanning control system synchronously polls and activates high-sensitivity detector arrays, such as solid-state detector arrays composed of cadmium zinc telluride crystals, distributed directly opposite each high-energy collimated X-ray source according to a preset time-division multiplexed scanning sequence. Only when the corresponding X-ray source emits a pulse is the detector array on that side activated and enter the data acquisition window. The detector array captures and quantifies the photon flux of the high-energy X-ray beam, whose intensity has attenuated after penetrating the equipment, and converts it into an electrical pulse signal through built-in preamplification and shaping circuitry. Finally, the array's back-end data acquisition card converts the electrical pulse signal of each detection unit into a 16-bit or higher bit-depth digital signal using high-speed analog-to-digital conversion, and adds a precise global positioning system timestamp and the physical address code of the detector unit. This results in a structured dataset containing attenuation intensity information at all detection angles. This dataset is the raw data stream of X-ray attenuation, which is transmitted to the subsequent data processing server of the supercomputing cluster via a high-speed fiber optic bus.

[0045] In this step, the acquired raw ray attenuation data stream can be represented as a dataset, with the following mathematical form:

[0046]

[0047] in, This represents the complete raw data stream of ray attenuation; It is in the At the moment of the first scan pulse, by the first... The beam emitted by the high-energy collimated X-ray source, after penetrating the equipment, is detected by the detector array on its opposite side. The radiation intensity value measured by each detector unit; It is the total number of high-energy collimated X-ray sources; It is the total number of detector elements in a single detector array; It represents the total number of pulses in a single scan task.

[0048] In the above definition, all variables are measurable physical quantities or counts. The unit is the photon count or a digital quantization unit output by the detector that is proportional to the energy deposition; it is a dimensionless pure number. All are dimensionless integer indices. This expression clearly defines the structure of the data stream, its physical meaning is explicit, and it does not involve the problem of inconsistent dimensions.

[0049] Sparse tomography arrays refer to arrays consisting of a limited number, typically fewer than 10, high-energy X-ray sources and detector pairs arranged discontinuously around the target, contrasting with dense rings of hundreds of source-detector pairs used in the medical field. High-energy collimated X-ray sources are devices capable of generating energies above megaelectron volts with extremely small beam divergence angles; examples include linear electron accelerators, whose high-energy rays possess the ability to penetrate thick metal structures. Pulsed detection fields are temporally discontinuous detection energy fields formed in space by X-ray sources operating in pulsed mode; their pulsed characteristics facilitate high-speed dynamic scanning. High-sensitivity detector arrays are collections of detectors capable of efficiently detecting individual high-energy photons with good energy resolution, typically made of materials such as cadmium zinc telluride or scintillation crystals. Raw X-ray attenuation data streams are intensity data sets directly acquired and digitized from the detector array without any reconstruction algorithms; they contain attenuation information along each detection path and serve as direct input for subsequent physical state analysis. The selection of X-ray source energy, set in the range of 6 million to 9 million electron volts, is based on engineering trade-offs. This energy range can effectively penetrate typical ultra-high voltage power equipment, such as the steel casing and internal components of a 220kV instrument transformer, while maintaining sufficient resolution contrast for insulating media of different densities. It is a commonly used energy window in this field for non-destructive testing of large industrial components.

[0050] For example, suppose an operator performs a rapid scan on a 220kV high-voltage transformer. The tomographic scan control system first activates five high-energy collimated X-ray sources surrounding the transformer. The system's X-ray sources are activated, and their anode voltages are set to generate a 7 Mvolt pulsed X-ray beam with a pulse width of 2 microseconds. This beam passes through the transformer along a pre-defined path in a high-fidelity digital twin reference model, specifically through the epoxy resin insulating tie rod, which is the object of analysis. On the opposite side, a detector array D1 consisting of 1024 detector units is simultaneously activated. Assume that, without the transformer obstruction, a detector unit can receive a reference photon count of 250,000 in a single pulse. In this penetration scan, detector unit number 512 in the D1 array, located directly opposite the center of the epoxy resin insulating tie rod, actually collects a photon count of 85,432. This value, along with its source-detector pair identifier and the global positioning system timestamp of the acquisition time (2023-10-26T10:30:01.123456Z), is packaged into a single data record. Subsequently, the system sequentially triggers the X-ray sources and simultaneously acquires data from the corresponding detector arrays D2 to D5. All five source-detector combinations complete a full data acquisition, generating thousands of similar data records. The collection of these records constitutes a single frame of raw X-ray attenuation data and is immediately transmitted to the supercomputing cluster's data processing server for subsequent analysis.

[0051] The feature vector generation module is used to acquire the raw data stream of ray attenuation, and to perform frequency domain component conversion by calling the parallel computing power of the supercomputing cluster based on the physical state transition probability database to obtain the coherent spectrum feature vector.

[0052] In a specific embodiment of the present invention, the process of obtaining the original data stream of ray attenuation and performing frequency domain component conversion based on the parallel computing power of the supercomputing cluster using the physical state transition probability database to obtain the coherent spectrum feature vector includes: obtaining the actual measured intensity value in the original data stream of ray attenuation and comparing the actual measured intensity value with the theoretical ray intensity value to obtain an attenuation anomaly distribution map.

[0053] Obtain the decay anomaly distribution map, and use the physical state transition probability database as a pattern matching template to perform a sliding window scan on the decay anomaly distribution map to obtain the suspected region of interest.

[0054] Obtain the ray attenuation data segment matrix for a specific time window that passes through the suspected region of interest. Based on the data segmentation strategy in the spatial or temporal domain, call multiple computing nodes of the supercomputing cluster to perform a two-dimensional discrete Fourier transform on the ray attenuation data segment matrix for the specific time window to obtain a two-dimensional spectrum.

[0055] A two-dimensional spectrum is obtained, and a digital high-pass filter is applied to extract the peak point coordinates and energy amplitude in the high-frequency spectrum region to obtain the coherent spectrum feature vector.

[0056] Specifically, after receiving the raw ray attenuation data stream, the supercomputing cluster's data processing server calls the physical state transition probability database and introduces frequency domain component transformation in both spatial and temporal dimensions to extract the coherent spectrum feature vector encoding the growth of microscopic physical failures. The process is as follows: First, the supercomputing cluster's data processing server retrieves a high-fidelity digital twin benchmark model. Based on the ray path geometry and material properties of each source-detector pair, it calculates the theoretical ray intensity value that each detector unit should receive under ideal health conditions. The server compares the actual measured intensity values ​​in the raw ray attenuation data stream with this theoretical value point-by-point, calculating the attenuation deviation on each detector path. Subsequently, the server maps these deviation values ​​back to three-dimensional space, forming an attenuation anomaly distribution map. It uses the physical state transition probability database as a pattern matching template, performing a sliding window scan on the attenuation anomaly distribution map to identify regions that highly match typical failure modes recorded in the database, such as primary carbonization of epoxy resin, exhibiting specific gradients, morphologies, or aggregation characteristics in space. Once the matching degree exceeds a preset confidence threshold, such as 0.85, the server designates the three-dimensional spatial region as a potentially interesting region of interest. Next, the supercomputing cluster's data processing server precisely extracts the attenuation data corresponding to all ray beams passing through the suspected region of interest within a preset specific time window, such as the past 500 milliseconds, from the cached, continuously acquired raw ray attenuation data stream. This forms a two-dimensional matrix of ray attenuation data segments for a specific time window. The rows of this matrix represent the spatial dimension, i.e., the cell index along the detector array, and the columns represent the temporal dimension, i.e., the pulse sequence index. The server then applies frequency band resolution operations to this data matrix. Specifically, using the supercomputing cluster's Message Passing Interface (MPI) or Unified Computing Device Architecture (CUDA), the two-dimensional data matrix is ​​divided into blocks according to spatial rows or temporal columns, distributed to multiple parallel central processing unit (CPU) or graphics processing unit (GPU) nodes, and a distributed two-dimensional fast Fourier transform is performed to convert the data from the spatial-temporal domain to the spatial wavenumber-temporal frequency domain. Finally, the server applies a digital high-pass filter to the generated two-dimensional spectrum. The cutoff frequency of this digital high-pass filter is set significantly higher than the power frequency and its low harmonics of the power grid system, above 150 Hz, to filter out low-frequency background noise introduced by power frequency interference and macroscopic vibrations. This allows the extraction of high-frequency spatial and temporal periodic fluctuation signal components induced solely by the evolution of microscopic defects, such as the minute extension of carbonization paths or the breathing effect of air gaps. The server searches for and locates several peak points with the most significant energy within the high-frequency spectrum region, extracts the coordinates of these peak points (i.e., spatial wavenumber and temporal frequency) and the corresponding energy amplitude, and concatenates these values ​​in a predetermined order to form a high-dimensional physical state fingerprint that uniquely characterizes the current microscopic failure dynamics, i.e., a coherent spectrum feature vector.

[0057] The parallel computing power of a supercomputing cluster refers to the use of a distributed computing architecture to decompose a massive two-dimensional data segment matrix into multiple independent subtasks, which are then computed simultaneously by multiple processing nodes to meet the ultra-low latency and real-time requirements of microsecond-level transient capture. A digital high-pass filter is an algorithm component characterized by allowing only high-frequency microscopic vibration signals with frequencies higher than a preset threshold to pass through, thereby shielding macroscopic equipment from conventional mechanical vibration noise.

[0058] In this step, the core frequency band analysis operation is a two-dimensional discrete Fourier transform, the mathematical expression of which is:

[0059]

[0060] in, A matrix representing discrete ray attenuation data segments within a specific time window. For spatial sampling point index, Index of time sampling points; and These represent the total number of sampling points in the spatial and temporal dimensions, respectively. It is the transformed two-dimensional spectrum, where It is a spatial wavenumber or spatial frequency index. It is a time-frequency index; It is the imaginary unit.

[0061] Regarding the meaning of the symbols, It is a dimensionless digital intensity value. All are dimensionless integer indices. This transformation converts a real matrix... Convert to a complex matrix The transformation itself is mathematically rigorous and does not involve dimensionless problems. The entire calculation process, in a physical sense, decomposes the signal from the spatiotemporal domain into a superposition of sine waves of different frequencies, which is self-consistent.

[0062] The suspected region of interest (ROI) refers to a local three-dimensional space that, through comparison with a physical state transition probability database, is preliminarily determined to be highly likely to be experiencing microscopic physical failure. The specific time window ray attenuation data segment is a two-dimensional data subset extracted from massive raw data streams that precisely corresponds spatiotemporally to the suspected ROI; it serves as the direct input for subsequent spectral analysis. Frequency band analysis is a mathematical processing method that converts signals from the time or spatial domain to the frequency domain. In this embodiment, it is specifically implemented through a two-dimensional fast Fourier transform to reveal hidden periodic features in the data. The coherence spectrum feature vector is a numerical vector composed of the coordinates and amplitudes of key high-frequency peaks extracted from the frequency band analysis results. Like a fingerprint, it identifies the dynamic characteristics of a specific microscopic failure process. Regarding the confidence threshold setting, 0.85 is a typical value based on engineering experience, aiming to balance detection sensitivity and false alarm rate, ensuring that subsequent detailed analysis is triggered only when the data pattern highly matches a known failure mode.

[0063] For example, the data processing server of the supercomputing cluster received a ray intensity record from a detector unit, with a value of 85,432. Simultaneously, the server calculated from a high-fidelity digital twin benchmark model that the theoretical received intensity for this ray path under healthy conditions should be 95,000. The server found that the actual value was lower than the theoretical value, indicating abnormal attenuation. Calculations showed that the attenuation coefficient of this path matched the attenuation coefficient gradient characteristics of the "epoxy resin primary carbonization" mode recorded in the physical state transition probability database with a similar degree of agreement (0.91), exceeding the threshold of 0.85. The server jointly analyzed data from 20 surrounding detector units and found that they also exhibited similar anomalies. Therefore, the region with a volume of approximately 5 mm³, located inside the epoxy resin insulated tie rod and traversed by these 21 ray paths, was designated as a region of suspected interest. The server then extracted the ray intensity data from 21 spatial sampling points corresponding to the past 512 scan pulses (512 time sampling points) of this region, forming a 21-row, 512-column matrix of ray attenuation data segments for a specific time window. The server performs a two-dimensional Fast Fourier Transform on the matrix, obtaining a 21x512 complex spectrum matrix. After applying a high-pass filter, the algorithm finds a significant energy peak in the high-frequency region of the spectrum matrix, located at spatial wavenumber index k=4 and time frequency index k=512. At this point, the normalized amplitude is 0.62. The server extracts and encodes this information to generate a coherent spectral feature vector. For example, a segment of the vector is [4,78,0.62], and this vector is immediately forwarded to the next processing module.

[0064] See Figure 2The disturbance waveform sequence generation module is used to obtain the coherent spectrum feature vector and perform causal analysis calculation through the physical-electrical coupling inverse deduction model to obtain the suppressive power quality disturbance waveform sequence.

[0065] In a specific embodiment of the present invention, a coherent spectrum feature vector is obtained, and a causal analysis operation is performed through a physical-electrical coupling inverse deduction model to obtain a suppressive power quality disturbance waveform sequence, including: obtaining the time frequency component and amplitude component in the coherent spectrum feature vector, and performing analysis through a physical-electrical coupling inverse deduction model to obtain the vibration mode frequency and energy intensity of micro-defect growth.

[0066] The vibration mode frequency and energy intensity are obtained, and the reverse compensation electric field torque parameters are calculated based on the principles of electrostriction and Coulomb force.

[0067] Obtain the reverse compensation electric field torque parameters, and construct the time series data containing the amplitude, frequency, phase and duration of a specific high-frequency voltage waveform to obtain the suppressed power quality disturbance waveform sequence.

[0068] Specifically, the process of introducing a physical-electrical coupling inverse deduction model, performing causal analysis on the coherent spectrum feature vector, and solving for the output suppressed power quality disturbance waveform sequence is as follows: The data processing server of the supercomputing cluster first loads a pre-built physical-electrical coupling inverse deduction model from its pre-installed algorithm library. This model is a deep neural network trained on a large amount of multi-physics finite element simulation data of the target power equipment, using a backpropagation multilayer perceptron or support vector regression algorithm model. Internally, it contains a nonlinear function or high-dimensional lookup table that can accurately map the conversion relationship between microscopic physical deformation characteristics and macroscopic electrical stress response. The server then loads the coherent spectrum feature vector generated in the previous step as the sole input into the designated input interface of the deduction model. Subsequently, the server drives the physical-electrical coupling inverse deduction model to perform deep analysis on the coherent spectrum feature vector. The model directly identifies the time frequency component in the vector as the core vibration mode frequency of microscopic defect growth, and converts the amplitude component in the vector into the energy intensity of the vibration mode through a preset energy conversion function, thereby quantitatively evaluating the dielectric loss thermal effect properties during defect growth. Based on the analyzed vibration modes and energy intensity, the core solver of the inverse model initiates computation, aiming to calculate a reverse compensation field capable of neutralizing or suppressing the micro-vibration. For physical failures caused by micro-gap discharges or dielectric polarization oscillations, the model, based on the principles of electrostriction and Coulomb force, accurately calculates the external compensation electric field torque parameters required to generate a reverse mechanical stress field with opposite phase and equivalent torque. These parameters typically include the frequency, amplitude, and phase information of the compensation electric field. This external compensation electric field superimposes with the built-in electric field of the internal defect, thereby achieving a destructive interference physical effect. Finally, based on the calculated reverse compensation electric field torque parameters, the model constructs an electrical signal—a suppressive power quality disturbance waveform sequence—to generate this physical field interference effect in actual grid operation. This sequence is defined as a set of digitized time-series data, detailing the amplitude, frequency, phase, and duration of a specific high-frequency voltage waveform that needs to be superimposed on the fundamental power supply of the equipment.

[0069] The core calculation in this step lies in the reverse derivation model of physical-electrical coupling, which can be simplified to solving a set of control parameters. The process, which is carried out by a reverse mapping function. describe:

[0070]

[0071] in, The coherent spectral eigenvector representing the input can be expressed as: It includes spatial wavenumber, temporal frequency, and amplitude. It is the set of output control parameters, i.e. , representing the amplitude, frequency, phase, and duration of the compensation voltage, respectively. This function The goal is to make by The compensating torque generated by the generated compensating electric field with by Characterized defect vibration torque They are equal in size but opposite in direction. That is, they satisfy:

[0072]

[0073] in ,and with by Defined compensation voltage It is related through the electromechanical coupling coefficient matrix of the material.

[0074] In the above function expression, Each component, as defined in the previous step, is a dimensionless value or a value with specific physical units. The output... middle, The unit is volt. The unit is Hertz. The unit is radians. The unit is seconds. Inverse mapping function. Its internal physical model and numerical algorithms ensure the correct conversion of the physical meaning from input to output and the matching of dimensions. For example, it converts the dimensionless amplitude of the input... Converted to voltage amplitude via a gain coefficient with dimensionless V. .

[0075] Among them, the physical-electric coupling reverse inference model typically employs a deep neural network with multiple hidden layers. It uses the feature vectors obtained from finite element simulation as input samples and the corresponding suppression electric field parameters as labels for supervised training. The destructive interference physical effect refers to the phenomenon where an alternating electric field of a specific frequency and phase injected externally induces an electrostrictive force or Coulomb force at the microscopic defects of the material, which is completely opposite in direction to the stress experienced by the defect itself during growth. The two forces cancel each other out, thereby achieving "physical stabilization".

[0076] For example, the supercomputing cluster's data processing server loads a physical-electrical coupling inverse model and receives an example-generated coherent spectrum feature vector, [4, 78, 0.62]. The model immediately parses this vector and identifies a suspected region of interest within the epoxy resin insulated tie rod, containing a microscopic vibration mode with a dominant frequency of 78 Hz and a normalized energy amplitude of 0.62. The model then activates its internal inverse solver to calculate compensation parameters that suppress this 78 Hz vibration. Based on the model's pre-stored electromechanical coupling coefficients for the epoxy resin material, the solver calculates that a compensation electric field with the same frequency of 78 Hz, a phase opposite to the defect vibration (i.e., a phase difference of π radians), and a voltage amplitude of 31 volts needs to be applied. This voltage amplitude is obtained by multiplying the defect energy amplitude of 0.62 by a pre-set control gain coefficient of 50V within the model. Simultaneously, the model sets the duration of the interference waveform to 500 milliseconds. Finally, the supercomputing cluster's data processing server integrates these calculated parameters—amplitude 31V, frequency 78Hz, phase π, and duration 500ms—to generate a suppressive power quality disturbance waveform sequence. This sequence, with a precision of 256 samples per cycle, is represented as a digital waveform data file containing 39 complete cycles and is ready to be sent to the control execution unit on the grid side.

[0077] The digital trigger pulse generation module is used to inject the suppressive power quality disturbance waveform sequence into the grid-connected node, perform extreme moment capture, and obtain the digital trigger pulse.

[0078] In a specific embodiment of the present invention, the suppressive power quality disturbance waveform sequence is injected into the grid-connected node, extreme moment capture is performed, and a digital trigger pulse is obtained, including: acquiring the suppressive power quality disturbance waveform sequence, executing a digital-to-pulse width modulation conversion algorithm, and obtaining a control command set.

[0079] The control command set is acquired, and the local active power filter is driven by the control command set to synthesize a simulated voltage waveform. The simulated voltage waveform is then superimposed on the fundamental frequency of the power grid system to obtain the local interference environment.

[0080] The actual voltage waveform at the output port of the local active power filter is obtained. The actual voltage waveform is compared with the preset extreme voltage threshold. When the instantaneous value first reaches the extreme voltage threshold, the monostable trigger is triggered to obtain a digital trigger pulse.

[0081] Specifically, the process of injecting a suppressive power quality disturbance waveform sequence into the grid-connected node, superimposing it to generate a local interference environment, and simultaneously generating digital trigger pulses physically locked to the timing logic is as follows: First, a waveform injection controller deployed near the power equipment receives and parses the suppressive power quality disturbance waveform sequence generated in the previous step and stored in digital timing data format. The controller's internal digital signal processor executes a digital-to-pulse-width modulation (PWM) conversion algorithm in real time, mapping each discrete amplitude point of the waveform sequence to a set of pulse-width modulation (PWM) signals with a specific duty cycle. This series of PWM signals constitutes a control command set that can directly drive the switching action of power electronic devices. Subsequently, the waveform injection controller sends the control command set to a local active power filter that is electrically closely associated with the target power equipment via high-speed optical fiber. This filter is typically connected to the third winding or the corresponding compensation branch of the equipment. This command drives the insulated-gate bipolar transistor (IGBT) power module of the local active power filter to perform high-frequency switching action, accurately synthesizing an analog voltage waveform at its AC output that is completely consistent with the suppressive power quality disturbance waveform sequence. The synthesized voltage is forcibly superimposed onto the 50Hz fundamental voltage of the power grid system during normal operation, thereby completing the precise grid injection operation of the suppressed power quality disturbance waveform sequence in the electromagnetic environment around the equipment. Simultaneously, the waveform injection controller uses a high-precision voltage sensor to capture the actual voltage waveform at the output port of the local active power filter in real time. The controller's internal hardware comparator continuously compares this real-time feedback waveform with a preset extreme voltage threshold. When the instantaneous value of the feedback voltage first reaches or exceeds this threshold, for example, when it reaches a positive or negative peak, it indicates that the extreme moment of the injection discharge has arrived. At this moment, the hardware comparator immediately outputs a high-level toggle signal. This signal triggers a monostable multivibrator, which instantaneously generates a transistor-to-transistor logic level signal with a pulse width precisely set to 1 microsecond, a voltage amplitude of 5 volts, and an extremely steep rising edge. This signal is the digital trigger pulse used to force the downstream equipment clock cycle to lock.

[0082] In one specific embodiment of the present invention, the extreme voltage threshold used by the hardware comparator is not a fixed static value, but is dynamically correlated with the compensation voltage amplitude output by the inverse deduction model, typically ±31V. This setting is based on maximizing physical observation efficiency: by precisely locking the trigger threshold at the voltage peak point of the injected interference waveform, the system can ensure that the CT transient scanning action is executed synchronously within the microsecond-level window period when the electrostrictive force or Coulomb torque interference is most intense and the physical suppression response is most significant. This mechanism is clearly demonstrated in the specification through a complete calculation derivation mapping characteristic amplitudes to voltage parameters. Its numerical source is reliable, its physical meaning is clear, and its timing logic is rigorous, meeting the legal requirements for full disclosure.

[0083] The implementation of this step mainly relies on digital signal processing and control logic, and does not involve complex mathematical operations that require special explanation, so the formula explanation is omitted.

[0084] The control instruction set refers to a series of coded, low-level drive signals used to control the power semiconductor devices in the local active power filter, such as insulated-gate bipolar transistors, to perform on / off operations. Its most common form is a pulse-width modulation (PWM) signal sequence. A local active power filter is an advanced power electronic device based on a voltage source inverter. It can actively and rapidly generate and inject compensation current or voltage of arbitrary waveforms into the power grid according to external commands. The system operating fundamental frequency refers to the main power frequency sine wave in the power system that carries the task of power transmission; in China, its frequency is 50 Hz. A digital trigger pulse is a standardized digital logic signal with high time accuracy and stable level. Its key characteristic is that its generation time is strictly synchronized with a certain physical event, namely the extreme point of the suppressive power quality disturbance waveform sequence. Its function is to provide an absolute time reference or a "preemptive" interrupt signal for other systems or equipment that require precise timing coordination. The pulse width of the digital trigger pulse is set to 1 microsecond. This is based on the consideration that the signal can be reliably captured by modern digital control systems, while being narrow enough to provide high time resolution. It is a typical parameter configuration for achieving high-precision synchronization in digital systems.

[0085] For example, the waveform injection controller receives a digital file describing a sequence of suppressive power quality disturbance waveforms with an amplitude of 31 volts, a frequency of 78 Hz, a phase of π, and a duration of 500 milliseconds. The controller immediately converts this waveform data into a series of pulse-width modulated signals with dynamically varying duty cycles and sends them to a local active power filter connected near a 220kV transformer. This filter is directed to superimpose a 78Hz sinusoidal voltage signal with a peak value of 31 volts onto the normal power frequency voltage on the transformer's supply line. After the superposition begins, the controller's voltage sensor continuously monitors the output. Since the injected waveform starts with a phase of π, its first extreme point will be a negative peak. When the sensor detects that the instantaneous value of the superimposed waveform first reaches -31 volts, for example, at 3.205 milliseconds after the injection begins, a hardware comparator within the controller immediately triggers, generating a first digital trigger pulse with a voltage of 5 volts and a pulse width of 1 microsecond. Subsequently, whenever the injected waveform reaches its positive peak, i.e., +31 volts, or its negative peak, i.e., -31 volts, the controller repeats this process, generating a total of 78 digital trigger pulse sequences with precise timing over a total duration of 500 milliseconds, for precise synchronization control of subsequent steps.

[0086] The transient attenuation data stream generation module is used to acquire digital trigger pulses and synchronously call the tomographic scanning array to perform transient repositioning and focusing operations through digital trigger pulses, thereby obtaining a high signal-to-noise ratio transient attenuation data stream.

[0087] In a specific embodiment of the present invention, acquiring a digital trigger pulse and synchronously calling the tomographic scanning array to perform a transient repositioning and focusing operation through the digital trigger pulse to obtain a high signal-to-noise ratio transient attenuation data stream includes: acquiring the digital trigger pulse and parsing the digital trigger pulse into a low-level interrupt control signal.

[0088] The underlying interrupt control signal is obtained, the three-dimensional spatial coordinates of the pre-determined suspected region of interest are retrieved, and the angle adjustment of each high-energy collimated X-ray source in the tomographic array is calculated.

[0089] The angle adjustment amount is obtained, and the high-energy collimated X-ray source is driven to adjust the projection angle so that the detection X-ray beam is cross-focused onto the suspected region of interest.

[0090] After focusing, the suspected region of interest is acquired, and synchronous scanning pulses are triggered to collect the attenuated X-ray photon data, resulting in a high signal-to-noise ratio transient attenuated data stream.

[0091] Specifically, the implementation process for using digital trigger pulses to lock and synchronously invoke the sparse tomography array, perform transient repositioning and focusing operations, and generate a high signal-to-noise ratio transient attenuation data stream for comparison and verification is as follows: First, the digital trigger pulse generated in the previous step, with nanosecond-level synchronization accuracy, is directly transmitted to the underlying control system of the sparse tomography array via a dedicated, low-latency coaxial cable or fiber optic link. Upon receiving the pulse, the system resolves it as a high-priority hardware-level interrupt control signal. This operation bypasses the system's conventional software-based task scheduling unit, ensuring extremely low latency and rapid response in subsequent actions. Upon receiving the underlying interrupt control signal, the underlying control system immediately retrieves the pre-loaded, determined three-dimensional spatial coordinates of the suspected region of interest from its internal cache. The system's built-in motion control algorithms, such as inverse kinematics solvers, calculate the required angle adjustment for the servo positioning motors or piezoelectric actuators of each X-ray source based on the target coordinates and the current positions of each high-energy collimated X-ray source. The system then sends parallel drive commands to the positioning mechanisms of all high-energy collimated X-ray sources, rapidly adjusting their projection angles to achieve cross-focusing of all probe beam paths, precisely converging and passing through the suspected region of interest. This repositioning process is completed within milliseconds, ensuring that the scanning array is in a focused, ready state within the effective window period of physical interference caused by the suppressive power quality disturbance waveform sequence. Immediately after confirming that all X-ray source angles are adjusted correctly, the underlying control system triggers a high-intensity synchronous scanning pulse. The X-ray flux of this scanning pulse is set to 5 to 10 times that of a conventional scan, but the duration is compressed to tens to hundreds of microseconds, aiming to instantaneously acquire maximum signal strength. Within this brief, strong pulse window, the detector arrays located opposite each X-ray source synchronously perform target-area penetration, specifically collecting and quantifying the attenuated X-ray photon data generated after passing through the focused suspected region of interest. This results in a structured dataset with an effectively improved signal-to-noise ratio and strict synchronization with external physical interference effects; this dataset is the high signal-to-noise ratio transient attenuation data stream.

[0092] In this step, improving the signal-to-noise ratio is the key technical effect, and its improvement rate is significant. The following formula can be used for quantitative evaluation:

[0093]

[0094] in, The signal-to-noise ratio measured after performing cross-firing operation; The signal-to-noise ratio was measured during the initial sparse scan. This represents the total number of high-energy collimated X-ray sources participating in the focusing operation. This formula is based on the fundamental assumption that the signal-to-noise ratio (SNR) is primarily determined by Poisson-distributed shot noise, in which case the SNR is related to the number of signal photons. It is directly proportional to the square root, that is .when When multiple X-ray sources are focused on the same region, the total number of photons received in that region is approximately the same as the original number. Therefore, the new signal-to-noise ratio is approximately equal to that of the original. times.

[0095] In this formula, and All are dimensionless pure numbers, representing the ratio of signal strength to noise strength. This is the number of X-ray sources, and it is also a dimensionless integer. Therefore, the entire formula is dimensionally consistent, clearly revealing the physical principle of improving the signal-to-noise ratio by increasing the number of focusing sources.

[0096] Among them, the low-level interrupt control signal is a highest-priority instruction that can bypass the operating system software scheduling and be processed directly by the hardware level to achieve near-zero latency event response. Cross-focusing is a scanning strategy that precisely controls the projection angles of multiple X-ray sources, causing their beams to converge within a specific small volume in space, thereby greatly increasing the X-ray flux density in that area. The effective window period refers to the duration during which the suppressive power quality disturbance waveform sequence produces measurable physical effects within the device. The high signal-to-noise ratio transient attenuation data stream is a set of data acquired within the effective window period using the cross-focusing strategy. Its characteristics include a signal strength far exceeding the background noise, and the data itself accurately records the instantaneous physical response of the device when subjected to external electrical disturbances. Setting the scanning pulse intensity to 5 to 10 times that of conventional scanning is to obtain sufficient signal photon counts within an extremely short exposure time to ensure the signal-to-noise ratio; this is a typical trade-off choice in high-speed dynamic imaging technology.

[0097] For example, the low-level control system of the sparse tomography array receives the digital trigger pulse generated in the previous step at 8.01 milliseconds. The system immediately recognizes this hardware interrupt and reads the coordinates of the suspected region of interest (ROI), a 5 mm³ area located inside the epoxy-insulated tie rod, from the cache. The system instructs the piezoelectric ceramic actuators of the five high-energy collimating X-ray sources to adjust their angles; the entire focusing repositioning process takes 2.1 milliseconds. At 10.11 milliseconds, the system confirms that focusing is complete and then triggers a high-intensity scan pulse lasting 100 microseconds. During this pulse, detector unit 512 in detector array D1, which had previously recorded anomalies, now receives photons from the backscattered and transmitted photons from all five X-ray sources converging at the suspected ROI. Its photon count increases dramatically from 85,432 in the initial sparse scan to 420,000. Its signal-to-noise ratio increases from... Proportional to, rising to Proportional to this, the signal-to-noise ratio improvement is approximately 2.22, which is consistent with the theoretical value. The match was excellent. Similar high-intensity signals were also acquired by the detector units located in the other four detector arrays, which were positioned directly over the suspected region of interest. These five sets of high-intensity, high signal-to-noise ratio attenuation data records, acquired within 100 microseconds, together constituted a high signal-to-noise ratio transient attenuation data stream, which was immediately packaged and used for the final comparative analysis.

[0098] See Figure 3 The waveform parameter sequence generation module is used to acquire high signal-to-noise ratio transient decay data stream, compare the microscopic response difference between the high signal-to-noise ratio transient decay data stream and the coherent spectrum feature vector, and obtain the corrected waveform parameter sequence.

[0099] In a specific embodiment of the present invention, a high signal-to-noise ratio transient attenuation data stream is acquired, and the microscopic response difference between the high signal-to-noise ratio transient attenuation data stream and the coherent spectrum feature vector is compared to obtain a corrected waveform parameter sequence. This includes: acquiring the high signal-to-noise ratio transient attenuation data stream, performing frequency domain feature extraction operations, and obtaining a secondary coherent spectrum feature vector.

[0100] Obtain the eigenvectors of the secondary coherence spectrum and the eigenvectors of the coherence spectrum, calculate the relative changes between the energy amplitudes corresponding to the core vibration mode frequencies, and obtain the characteristic degradation attenuation ratio.

[0101] Obtain the feature degradation decay ratio, compare the feature degradation decay ratio with the expected stabilization index, and obtain the deviation residual margin.

[0102] Obtain the deviation residual margin, calculate the partial derivative of the deviation residual margin with respect to the coupling weight matrix before correction, and obtain the gradient value.

[0103] Obtain the gradient value and the preset learning rate, and multiply the gradient value and the learning rate to obtain the weight adjustment amount.

[0104] Obtain the original coupling weight matrix, and subtract the weight adjustment amount from the original coupling weight matrix to obtain the modified coupling weight matrix.

[0105] Obtain the corrected coupling weight matrix and coherence spectrum eigenvector, input the physical-electric coupling reverse deduction model to generate the next cycle micro-control waveform sequence, and obtain the corrected waveform parameter sequence.

[0106] Specifically, the implementation process for comparing the microscopic response differences in the feedback of high signal-to-noise ratio transient attenuation data streams, deduce the microscopic medium physical disturbance rejection effectiveness before and after grid control, and output the corrected waveform parameter sequence in a closed-loop iterative manner is as follows: First, the data processing server of the supercomputing cluster receives and processes the high signal-to-noise ratio transient attenuation data stream generated in the previous step. The server applies frequency domain feature extraction operations to it, that is, for this new data synchronized with physical interference, it also extracts the data segments that pass through the suspected region of interest, performs a two-dimensional fast Fourier transform, and applies the same high-pass filter to calculate and extract the secondary coherence spectrum feature vector for comparison and verification. Next, the server reads from its memory and compares the initially recorded coherence spectrum feature vector with the newly generated secondary coherence spectrum feature vector. The core of the comparison is to quantify the changes of the two vectors in the key frequency domain components, especially the energy amplitude corresponding to the core vibration mode frequency representing defect growth initially identified. By calculating the relative change between these two amplitudes, the server quantifies the feature degradation attenuation ratio caused by the interference of the suppressive power quality disturbance waveform sequence with the microscopic physical failure process. Subsequently, the server compares the calculated characteristic degradation attenuation ratio with a preset expected stabilization index, such as 80%, representing the ideal control effect, to evaluate the deviation residual margin between the two. If the residual margin is greater than a preset tolerance threshold, such as 5%, it indicates that the current intervention effect has not met the expected stabilization index, and the control strategy needs to be adjusted. To this end, the server reverse-engineers the internal parameters of the physical-electrical coupling inverse derivation model, especially the key coupling weight matrix responsible for mapping defect energy to compensation voltage amplitude. The correction is based on finding a new weight value through gradient descent or other optimization algorithms, so that the model can output a corrected waveform parameter that is expected to minimize the deviation residual margin when receiving the same initial coherence spectrum eigenvector. Finally, the server uses the corrected physical-electrical coupling inverse derivation model and the initial coherence spectrum eigenvector to perform a complete forward calculation again, generating a micro-control waveform sequence for the next cycle of continuous closed-loop execution of intervention scheduling, and prepares to inject it into the power grid in the next control cycle, thereby completing a complete closed-loop iterative control.

[0107] In this step, the key quantitative evaluation and parameter correction can be mathematically expressed as follows:

[0108] Feature degradation attenuation ratio Calculation:

[0109]

[0110] Deviation residual margin Calculation:

[0111]

[0112] Coupling weight matrix Update:

[0113]

[0114] in, It is the energy amplitude of the vibration mode of the defect core in the eigenvector of the coherent spectrum; It is the energy amplitude of the corresponding mode in the eigenvector of the secondary coherence spectrum; It is a pre-set, expected calming indicator; and These are the coupling weight matrices before and after the correction, or their key coefficients; It is the learning rate of the gradient descent algorithm.

[0115] In the above formula, and It is a dimensionless amplitude with the same dimension, therefore It is a dimensionless percentage. and These are all dimensionless percentages. The weight update formula is a standard gradient descent iterative form. To ensure consistency of physical dimensions, the learning rate... It implicitly contains dimension conversion coefficients, and its specific physical dimensions are set as follows: This makes the parameter adjustment term on the right side of the equation... offset as Ensure that it is coupled with the weight matrix Maintain the same voltage dimensions. The entire calculation process is mathematically and physically consistent.

[0116] The secondary coherence spectrum eigenvector is extracted from the high signal-to-noise ratio transient attenuation data stream through frequency domain analysis. Its format is identical to the initial coherence spectrum eigenvector, specifically designed to reflect the microscopic dynamic changes of the system after applying electrical interference. The characteristic degradation attenuation ratio is a quantitative indicator used to measure the extent to which the injected suppressive waveform successfully suppresses the activity intensity of microscopic defects within the equipment. The expected stabilization index is a preset engineering target representing the minimum defect activity suppression level expected by this control system. The deviation residual margin is the gap between the actual control effect and the expected target, serving as the core error signal driving the iterative optimization of the control strategy. The coupling weight matrix is ​​a set of core adjustable parameters in the physical-electrical coupling inverse derivation model. It determines the conversion gain from the identified defect intensity to the generated compensation voltage intensity and is the object directly optimized by the closed-loop control algorithm.

[0117] For example, the data processing server of the supercomputing cluster receives a generated high signal-to-noise ratio transient decay data stream. After frequency domain analysis, the server extracts the secondary coherence spectrum feature vector. The analysis reveals that at the core frequency of 78 Hz, the new energy amplitude has significantly decreased from the initial 0.62 to 0.15. Based on this, the server calculates the feature degradation attenuation ratio, obtaining... The server compares this result with the preset expected stabilization metric, i.e., 80%, and calculates the deviation residual margin. Since the residual was less than the 5% tolerance threshold, the system determined that the intervention was basically successful. However, for further optimization, the system still initiated a weight correction procedure. Through gradient descent algorithm analysis, the system found that under the current operating conditions, the control gain coefficient, originally 50V, in the physical-electrical coupling reverse derivation model should be fine-tuned. According to calculations, the new gain coefficient was corrected to 52.5V. Finally, the server used this updated model, along with the initial coherence spectrum eigenvector [4,78,0.62], to recalculate and generate the next cycle's micro-control waveform sequence. The amplitude of the new waveform sequence became... The frequency and phase remain unchanged. This new waveform sequence will be used in the next 500-millisecond control cycle in order to further increase the feature degradation attenuation ratio to a level closer to 80%.

[0118] The above content is merely an example and illustration of the concept of the present invention. Those skilled in the art can make various modifications or additions to the specific embodiments described, or use similar methods to replace them, as long as they do not deviate from the concept of the invention or exceed the scope defined by the present invention, and all such modifications and additions should fall within the protection scope of the present invention.

Claims

1. A real-time control system for power equipment based on supercomputing CT three-dimensional imaging, characterized in that, include: The probability database generation module is used to obtain the basic design parameters of power equipment, perform physical state degradation simulation, and obtain a physical state transition probability database. The attenuation raw data stream generation module is used to trigger a tomographic scanning array deployed around the power equipment to perform a transillumination scan of the power equipment and obtain the X-ray attenuation raw data stream. The feature vector generation module is used to acquire the raw data stream of ray attenuation, and to perform frequency domain component conversion based on the parallel computing power of the supercomputing cluster using the physical state transition probability database to obtain the coherent spectrum feature vector. The disturbance waveform sequence generation module is used to obtain the coherent spectrum feature vector and perform causal analysis calculation through the physical-electrical coupling inverse deduction model to obtain the suppressive power quality disturbance waveform sequence. The digital trigger pulse generation module is used to inject the suppressive power quality disturbance waveform sequence into the grid connection node, perform extreme moment capture, and obtain the digital trigger pulse; The transient attenuation data stream generation module is used to acquire digital trigger pulses and synchronously call the tomographic scanning array to perform transient repositioning and focusing operations through digital trigger pulses to obtain a high signal-to-noise ratio transient attenuation data stream; The waveform parameter sequence generation module is used to acquire a high signal-to-noise ratio transient decay data stream, compare the microscopic response difference between the high signal-to-noise ratio transient decay data stream and the coherent spectrum feature vector, and obtain the corrected waveform parameter sequence.

2. The real-time control system for power equipment based on supercomputing CT three-dimensional imaging according to claim 1, characterized in that: The process of obtaining the basic design parameters of the power equipment, performing physical state degradation simulation, and obtaining a physical state transition probability database includes: Obtain the geometric topology information and material factory test parameters from the basic design parameters, perform registration and fusion and mesh generation to obtain a high-fidelity digital twin reference model; A high-fidelity digital twin baseline model is obtained, and a preset accelerated aging load sequence is applied to obtain multiple discrete physical degradation stages. The physical degradation stage is obtained, and the ratio of the change in the ray attenuation coefficient to the change in the physical state between two adjacent physical degradation stages is calculated based on the physical law of ray attenuation to obtain the ray attenuation coefficient gradient. The gradient of the ray attenuation coefficient is obtained, and the gradient of the ray attenuation coefficient is structurally correlated with the spatial location of the component, the material type and the degradation mode to obtain a physical state transition probability database.

3. The real-time control system for power equipment based on supercomputing CT three-dimensional imaging according to claim 1, characterized in that: The triggering mechanism deploys a tomographic scanning array around the power equipment to perform a transillumination scan of the power equipment, obtaining a raw data stream of X-ray attenuation, including: Obtain the synchronous digital start command, and activate multiple high-energy collimated X-ray sources deployed around the power equipment to obtain the pulse detection field; A preset scanning strategy is obtained, and the high-energy collimated X-ray source is controlled to output a high-energy X-ray beam to obtain the transmitted X-ray that penetrates the power equipment. A preset time-division multiplexed scanning sequence is acquired. The detector array distributed on the opposite side of the high-energy collimated X-ray source is activated by polling the time-division multiplexed scanning sequence. The photon flux of the transmitted X-ray is collected and converted from analog to digital to obtain the raw data stream of X-ray attenuation.

4. The real-time control system for power equipment based on supercomputing CT three-dimensional imaging according to claim 1, characterized in that: The process of acquiring the raw ray attenuation data stream involves using the parallel computing power of the supercomputing cluster, based on the physical state transition probability database, to perform frequency domain component transformation, resulting in a coherent spectrum feature vector, including: Obtain the actual measured intensity value from the raw data stream of ray attenuation, compare the actual measured intensity value with the theoretical ray intensity value, and obtain the attenuation anomaly distribution map; Obtain the attenuation anomaly distribution map, and use the physical state transition probability database as a pattern matching template to perform a sliding window scan on the attenuation anomaly distribution map to obtain the suspected region of interest. Obtain the ray attenuation data segment matrix for a specific time window that passes through the suspected region of interest. Based on the data segmentation strategy in the spatial or temporal domain, call multiple computing nodes of the supercomputing cluster to perform a two-dimensional discrete Fourier transform on the ray attenuation data segment matrix for the specific time window to obtain a two-dimensional spectrum. A two-dimensional spectrum is obtained, and a digital high-pass filter is applied to extract the peak point coordinates and energy amplitude in the high-frequency spectrum region to obtain the coherent spectrum feature vector.

5. The real-time control system for power equipment based on supercomputing CT three-dimensional imaging according to claim 1, characterized in that: The process of obtaining the coherent spectrum feature vector and performing causal analysis through a physical-electrical coupling inverse deduction model yields a suppressive power quality disturbance waveform sequence, including: The time frequency component and amplitude component in the coherent spectrum feature vector are obtained, and the vibration mode frequency and energy intensity of micro-defect growth are obtained by analyzing the physical-electric coupling inverse deduction model. The vibration mode frequency and energy intensity are obtained, and the reverse compensation electric field torque parameters are calculated based on the principles of electrostriction and Coulomb force. Obtain the reverse compensation electric field torque parameters, and construct the time series data containing the amplitude, frequency, phase and duration of a specific high-frequency voltage waveform to obtain the suppressed power quality disturbance waveform sequence.

6. The real-time control system for power equipment based on supercomputing CT three-dimensional imaging according to claim 1, characterized in that: The process of injecting the suppressive power quality disturbance waveform sequence into the grid-connected node, capturing extreme moments, and obtaining digital trigger pulses includes: The waveform sequence of suppressive power quality disturbance is obtained, and a digital-to-pulse-width modulation conversion algorithm is executed to obtain the control command set; The control command set is acquired, and the local active power filter is driven by the control command set to synthesize a simulated voltage waveform. The simulated voltage waveform is then superimposed on the fundamental frequency of the power grid system to obtain the local interference environment. The actual voltage waveform at the output port of the local active power filter is obtained. The actual voltage waveform is compared with the preset extreme voltage threshold. When the instantaneous value first reaches the extreme voltage threshold, the monostable trigger is triggered to obtain a digital trigger pulse.

7. The real-time control system for power equipment based on supercomputing CT three-dimensional imaging according to claim 1, characterized in that: The process of acquiring a digital trigger pulse, synchronously calling the tomographic scanning array to perform a transient repositioning and focusing operation via the digital trigger pulse, and obtaining a high signal-to-noise ratio transient attenuation data stream includes: Acquire digital trigger pulses and parse them into underlying interrupt control signals; The underlying interrupt control signal is obtained, the three-dimensional spatial coordinates of the pre-determined suspected region of interest are retrieved, and the angle adjustment of each high-energy collimating X-ray source in the tomographic array is calculated. The angle adjustment amount is obtained, and the high-energy collimated X-ray source is driven to adjust the projection angle so that the detection X-ray beam is cross-focused onto the suspected region of interest; After focusing, the suspected region of interest is acquired, and synchronous scanning pulses are triggered to collect the attenuated X-ray photon data, resulting in a high signal-to-noise ratio transient attenuated data stream.

8. The real-time control system for power equipment based on supercomputing CT three-dimensional imaging according to claim 1, characterized in that: The process of acquiring a high signal-to-noise ratio transient attenuation data stream, comparing the microscopic response differences between the high signal-to-noise ratio transient attenuation data stream and the coherent spectrum eigenvector, and obtaining a corrected waveform parameter sequence includes: Acquire a high signal-to-noise ratio transient decay data stream, perform frequency domain feature extraction operations, and obtain the secondary coherence spectrum feature vector; Obtain the eigenvectors of the secondary coherence spectrum and the eigenvectors of the coherence spectrum, calculate the relative change between the energy amplitudes corresponding to the core vibration mode frequencies, and obtain the characteristic degradation attenuation ratio. Obtain the feature degradation decay ratio, compare the feature degradation decay ratio with the expected stabilization index, and obtain the deviation residual margin; Obtain the deviation residual margin, and use the gradient descent algorithm to reverse correct the coupling weight matrix in the physical-electrical coupling reverse derivation model. Then, recalculate in the forward direction to obtain the corrected waveform parameter sequence.

9. The real-time control system for power equipment based on supercomputing CT three-dimensional imaging according to claim 2, characterized in that: The process of obtaining the physical degradation stage involves calculating the ratio of the change in the ray attenuation coefficient between two adjacent physical degradation stages to the change in the physical state, based on the physical law of ray attenuation, to obtain the ray attenuation coefficient gradient, including: Obtain the density and elemental composition of the material during the physical degradation stage, calculate the attenuation effect of the virtual high-energy ray beam passing through the component, and obtain the ray attenuation coefficient corresponding to the current physical state; Obtain the ray attenuation coefficients of two adjacent physical degradation stages, calculate the difference between the ray attenuation coefficients, and obtain the change in the ray attenuation coefficient. Obtain the change in physical state between two adjacent physical degradation stages, and divide the change in ray attenuation coefficient by the change in physical state to obtain the gradient of ray attenuation coefficient.

10. The real-time control system for power equipment based on supercomputing CT three-dimensional imaging according to claim 8, characterized in that: The process of obtaining the deviation residual margin involves reversing the coupling weight matrix in the physical-electrical coupling reverse derivation model based on the gradient descent algorithm, recalculating in the forward direction, and obtaining the corrected waveform parameter sequence, including: Obtain the deviation residual margin, calculate the partial derivative of the deviation residual margin with respect to the coupling weight matrix before correction, and obtain the gradient value; Obtain the gradient value and the preset learning rate, and multiply the gradient value and the learning rate to obtain the weight adjustment amount; Obtain the coupling weight matrix before correction, and subtract the weight adjustment amount from the coupling weight matrix before correction to obtain the coupling weight matrix after correction; Obtain the corrected coupling weight matrix and coherence spectrum eigenvector, input the physical-electric coupling reverse deduction model to generate the next cycle micro-control waveform sequence, and obtain the corrected waveform parameter sequence.

Images