Intelligent fault diagnosis method integrating state monitoring and multi-mode large model

Abstract

The invention discloses an intelligent fault diagnosis method fusing state monitoring and a multi-modal large model, and the method specifically comprises the steps: synchronously collecting time sequence data and a space image through a heterogeneous sensor group and monitoring equipment disposed in power grid equipment, and forming original data; based on the original data, a physical constraint feature vector is generated in combination with an equipment thermodynamic equation and a material deformation rule; performing health index prediction through the lightweight LSTM network based on the physical constraint feature vector; when detecting that the health indexes continuously decrease, clustering an HI time sequence curve by adopting a Gaussian mixture model, judging a degradation stage according to a clustering center distance, and obtaining a stage recognition result; and based on finite element simulation parameters, introducing a reinforcement learning model, optimizing the simulation parameters by taking maintenance cost minimization as a target, and outputting a predictive maintenance work order. According to the invention, intelligent fault diagnosis and accurate maintenance of the power grid equipment are realized, the fault processing efficiency and accuracy are improved, and the power failure loss is reduced.

Description

Technical Field

[0001] This invention relates to the field of power grid fault detection technology, and in particular to a fault intelligent diagnosis method that integrates condition monitoring and multimodal large model. Background Technology

[0002] In power grid systems, core equipment such as transformers and transmission lines undertake the critical tasks of power transmission and conversion. The stability of their operation directly affects the safety and reliable power supply of the entire power grid. Once these core equipment malfunctions, it is highly likely to cause large-scale power outages, resulting in huge economic losses and inconvenience to social production and daily life.

[0003] Traditionally, fault monitoring of power grid equipment has relied primarily on single-parameter-based monitoring methods, such as setting temperature thresholds to monitor whether equipment is overheating. However, as power grid equipment becomes increasingly complex and operating conditions become more variable, this single-parameter monitoring approach is no longer sufficient to meet actual needs and cannot effectively address complex fault modes, such as winding insulation aging and partial discharge.

[0004] Currently, the field of power grid equipment monitoring and fault diagnosis faces many challenges: 1. The problem of isolated multi-source data: Power grid systems contain multiple data sources, such as time-series data collected by SCADA systems, image data acquired by infrared thermal imaging equipment, and equipment operation logs. These data are independent of each other and lack effective correlation and fusion mechanisms, resulting in an inability to comprehensively and accurately reflect the actual operating status of the equipment. For example, it is difficult to intuitively determine the location of localized overheating in equipment using only time-series data. While infrared thermal imaging images can visually display temperature distribution, the lack of integration with historical equipment operating data makes it difficult to accurately assess fault development trends.

[0005] 2. Complex Fault Modes: Faults in equipment such as transformers often manifest as coordinated anomalies in multiple parameters. Taking transformer faults as an example, not only will the temperature rise abnormally, but gas composition and vibration parameters will also change accordingly. Traditional fault diagnosis models usually only consider changes in a single parameter or a few parameters, making it difficult to capture the complex correlations between multimodal data. Consequently, they cannot accurately identify the fault type and severity, leading to frequent misdiagnosis or missed diagnosis.

[0006] 3. Delayed Emergency Response: In the existing fault diagnosis process, there is a lack of automatic linkage mechanism with the contingency plan database when equipment failure is detected. Maintenance personnel need to manually consult relevant contingency plans and formulate maintenance plans based on experience. This process is not only time-consuming but also easily affected by human factors, leading to extended repair time and further increasing power outage losses. Summary of the Invention

[0007] The purpose of this invention is to provide a fault intelligent diagnosis method that integrates condition monitoring and multimodal large model, thereby realizing intelligent fault diagnosis and precise maintenance of power grid equipment, improving fault handling efficiency and accuracy, and reducing power outage losses, thus solving at least one of the aforementioned problems in the prior art.

[0008] In a first aspect, the present invention provides a fault intelligent diagnosis method that integrates state monitoring and multimodal large model, the method specifically comprising: By deploying heterogeneous sensor groups and monitoring equipment on power grid equipment, time-series data and spatial images are collected synchronously to form raw data; Based on the original data, and combining the equipment's thermodynamic equations with the material deformation rules, a physical constraint feature vector is generated. Based on physically constrained feature vectors, a lightweight LSTM network is used to perform health indicator prediction. When a continuous decline in health indicators is detected, the HI time series curve is clustered using a Gaussian mixture model. The degradation stage is determined based on the distance between cluster centers to obtain the stage identification results. The finite element simulation parameters are then dynamically switched based on the stage identification results. Based on finite element simulation parameters, a reinforcement learning model is introduced to optimize the simulation parameters with the goal of minimizing maintenance costs, and predictive maintenance work orders are output. Perform on-site operations based on predictive maintenance work orders and collect post-maintenance status data using AR devices.

[0009] Secondly, the present invention provides a fault intelligent diagnosis system that integrates condition monitoring and multimodal large model, the system specifically comprising: The first diagnostic module is used to synchronously collect time-series data and spatial images through heterogeneous sensor groups and monitoring equipment deployed on power grid equipment to form raw data. The second diagnostic module is used to generate physical constraint feature vectors based on the original data, combined with the equipment's thermodynamic equations and material deformation rules; The third diagnostic module is used to perform health indicator prediction based on physical constraint feature vectors using a lightweight LSTM network. The fourth diagnostic module is used to cluster the HI time series curve using a Gaussian mixture model when a continuous decline in health indicators is detected. It determines the degradation stage based on the distance between cluster centers, obtains the stage identification result, and dynamically switches the finite element simulation parameters based on the stage identification result. The fifth diagnostic module is used to optimize simulation parameters based on finite element simulation parameters by introducing a reinforcement learning model with the goal of minimizing maintenance costs, and output predictive maintenance work orders. The sixth diagnostic module is used to perform on-site operations based on predictive maintenance work orders and collect post-maintenance status data through AR devices.

[0010] Thirdly, the present invention provides a computer device, including: a memory and a processor, and a computer program stored in the memory, wherein when the computer program is executed on the processor, it implements the fusion state monitoring and multimodal large model fault intelligent diagnosis method as described in any of the above methods.

[0011] Fourthly, the present invention provides a computer-readable storage medium having a computer program stored thereon, wherein the computer program, when executed by a processor, implements the fusion state monitoring and multimodal large model fault intelligent diagnosis method as described in any of the above methods.

[0012] Compared with the prior art, the present invention has at least one of the following technical effects: 1. This invention enables intelligent diagnosis and precise maintenance of power grid equipment faults, improving fault handling efficiency and accuracy, and reducing power outage losses.

[0013] 2. This invention combines multidisciplinary knowledge to process raw data and generates a physical constraint feature vector that comprehensively reflects the mechanical-thermodynamic state of the equipment, thereby improving the comprehensiveness and accuracy of fault feature extraction.

[0014] 3. This invention accurately segments infrared thermal images and calculates the rate of temperature change, providing detailed data support for analyzing the heat dissipation status of equipment and helping to detect heat dissipation anomalies in a timely manner.

[0015] 4. Based on professional theories and signal processing methods, this invention accurately analyzes the resonant frequency offset from the vibration spectrum, providing a key indicator for detecting winding mechanical deformation.

[0016] 5. This invention combines the temperature change rate and the resonant frequency offset to generate a joint feature identifier, which can more effectively identify abnormal coupling areas of equipment and assist in fault diagnosis.

[0017] 6. This invention uses a lightweight LSTM network to predict health indicators, and combines a re-prediction mechanism with maintenance strategy matching to achieve real-time assessment and early warning of equipment health status, and promptly push maintenance work orders.

[0018] 7. This invention uses Gaussian mixture model clustering to determine the degradation stage and dynamically switches finite element simulation parameters, making the simulation more closely match the actual state of the equipment and improving the pertinence of maintenance strategies.

[0019] 8. This invention introduces a reinforcement learning model to optimize finite element simulation parameters, and generates predictive maintenance work orders through multi-objective optimization, thereby minimizing maintenance costs and ensuring the scientific nature of maintenance decisions.

[0020] 9. This invention utilizes AR equipment to collect and analyze post-maintenance status data, promptly marking abnormal maintenance items and generating supplementary instructions to ensure maintenance quality and guarantee the healthy operation of the equipment. Attached Figure Description

[0021] To more clearly illustrate the technical solutions in the embodiments of this application, the drawings used in the embodiments will be briefly introduced below. Obviously, the drawings described below are only some embodiments of this application. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0022] Figure 1 This is a flowchart illustrating a fault intelligent diagnosis method that integrates state monitoring and multimodal large model according to an embodiment of the present invention. Figure 2 This is a schematic diagram of the structure of a fault intelligent diagnosis system that integrates state monitoring and multimodal large model according to an embodiment of the present invention; Figure 3 This is a schematic diagram of the structure of a computer device provided in an embodiment of the present invention. Detailed Implementation

[0023] In the following description, specific details such as particular system architectures and techniques are set forth for illustrative purposes and not for limitation, in order to provide a thorough understanding of the embodiments of this application. However, those skilled in the art will understand that this application may also be implemented in other embodiments without these specific details. In other instances, detailed descriptions of well-known systems, apparatuses, circuits, and methods have been omitted so as not to obscure the description of this application with unnecessary detail.

[0024] It should be understood that, when used in this application specification and the appended claims, the term "comprising" indicates the presence of the described features, integrals, steps, operations, elements and / or components, but does not exclude the presence or addition of one or more other features, integrals, steps, operations, elements, components and / or a collection thereof.

[0025] It should also be understood that the term “and / or” as used in this application specification and the appended claims means any combination of one or more of the associated listed items and all possible combinations, and includes such combinations.

[0026] As used in this application specification and the appended claims, the term "if" may be interpreted, depending on the context, as "when," "once," "in response to determination," or "in response to detection." Similarly, the phrase "if determined" or "if detected [the described condition or event]" may be interpreted, depending on the context, as meaning "once determined," "in response to determination," "once detected [the described condition or event]," or "in response to detection [the described condition or event]."

[0027] Furthermore, in the description of this application and the appended claims, the terms "first," "second," "third," etc., are used only to distinguish descriptions and should not be construed as indicating or implying relative importance.

[0028] References to "one embodiment" or "some embodiments" as described in this specification mean that one or more embodiments of this application include a specific feature, structure, or characteristic described in connection with that embodiment. Therefore, the phrases "in one embodiment," "in some embodiments," "in other embodiments," "in still other embodiments," etc., appearing in different parts of this specification do not necessarily refer to the same embodiment, but rather mean "one or more, but not all, embodiments," unless otherwise specifically emphasized. The terms "comprising," "including," "having," and variations thereof mean "including but not limited to," unless otherwise specifically emphasized.

[0029] In this application embodiment, the entity executing the process includes a terminal device. This terminal device includes, but is not limited to, devices capable of executing the methods disclosed in this application, such as servers, computers, smartphones, and tablets. Figure 1 A flowchart illustrating an embodiment of the intelligent fault diagnosis method integrating state monitoring and multimodal large model disclosed in this invention is shown below in detail: S101, through heterogeneous sensor groups and monitoring equipment deployed in power grid equipment, synchronously collects time-series data and spatial images to form raw data.

[0030] In this embodiment, a distributed fiber Bragg grating temperature sensor array is deployed on the surface of the transformer body using a grid-like layout, focusing on key areas such as the oil tank, bushings, tap changers, and radiators. During sensor installation, it is crucial to ensure a tight fit with the equipment surface to avoid measurement errors due to poor contact. The longitudinal spacing of the sensor array is set to 30cm, and the lateral spacing is adjusted to 20-50cm based on the equipment size to ensure that the temperature monitoring resolution meets the requirements for locating localized overheating. Sensor data is transmitted to an edge computing terminal via optical fiber, with a sampling frequency of 1 time per minute, synchronously recording timestamps and device numbers to form a raw data stream containing temperature values, location coordinates, and timing information.

[0031] A triaxial piezoelectric vibration sensor is installed on the transformer winding base (at the connection between the upper and lower yokes) using bolts to monitor vibration signals in the axial (X-axis), radial (Y-axis), and vertical (Z-axis) directions of the winding. Before installation, the sensor must be horizontally calibrated to ensure alignment with the equipment's mechanical structure. The vibration signal is transmitted to the data acquisition module via shielded twisted-pair cable, with a sampling frequency set to 20kHz to capture high-frequency vibration components. The data acquisition module incorporates an anti-aliasing filter to effectively filter out high-frequency noise and adjusts the gain of the raw signal to ensure the dynamic range covers vibration amplitudes under both normal and fault conditions. The acquired vibration data includes triaxial acceleration values, frequency components, and timing information, which are stored synchronously with temperature data on the edge computing terminal.

[0032] A binocular monitoring bracket is erected around the transformer equipment (3-5 meters from the main body), and an infrared thermal imager and a visible light camera are installed simultaneously. The infrared thermal imager uses an uncooled focal plane detector with a wavelength range covering 8-14μm, a temperature resolution ≤0.05℃, and a frame rate of 5 frames / second to ensure dynamic monitoring capabilities. The visible light camera uses a high-definition industrial camera with a resolution of no less than 1080P, and its frame rate is synchronized with the infrared thermal imager. The optical axes of the two devices are parallel and their fields of view overlap, and the timing of image acquisition is aligned through a time synchronization module. The acquired image data includes an infrared thermal image (pixel values ​​corresponding to surface temperature), a visible light image (equipment appearance status), and timestamp information, which is transmitted to an edge computing terminal via gigabit Ethernet and stored in association with temperature and vibration time-series data.

[0033] The edge computing terminal incorporates a multi-source data fusion module, using a timestamp alignment algorithm to synchronize and correlate temperature, vibration time-series data, and infrared / visible light images. Specifically, based on the sampling time point of the temperature sensor, interpolation processing is performed on the vibration data and image data to ensure that all data correspond to the same time segment. The synchronized dataset includes temperature values, vibration characteristic quantities (such as peak and RMS values), infrared thermal image temperature distribution, and visible light image appearance status at each monitoring point of the device, forming a structured raw dataset. This dataset is uploaded to a cloud database via an encrypted communication protocol, providing a foundation for subsequent physical constraint feature extraction and health indicator prediction.

[0034] This embodiment achieves full-dimensional perception of the operating status of power grid equipment through the collaborative deployment of heterogeneous sensor groups and multimodal monitoring equipment, effectively solving the data isolation problem of traditional single-parameter monitoring and providing data support for the accurate diagnosis of complex fault modes.

[0035] S102, based on the original data, combines the equipment's thermodynamic equations and material deformation rules to generate a physical constraint feature vector.

[0036] In this embodiment, a moving average filter is applied to the temperature sensor data to eliminate environmental noise interference; a bandpass filter is applied to the vibration signal to retain frequency components related to equipment mechanical faults; non-uniformity correction is performed on the infrared thermal image to eliminate temperature measurement deviations caused by differences in detector response; and histogram equalization is performed on the visible light image to enhance the discernibility of equipment appearance details. Subsequently, a timestamp synchronization algorithm is used to align the multimodal data to a unified time base, ensuring that temperature, vibration, and image data correspond to the same equipment operating time, forming a structured synchronized dataset.

[0037] Taking the surface temperature distribution of a transformer tank as an example, a three-dimensional heat conduction model is established using two-dimensional temperature field data provided by infrared thermal images, combined with equipment structural parameters (such as tank wall thickness and material thermal conductivity). The heat conduction equation is discretized using the finite difference method, and the internal temperature gradient distribution of the tank is iteratively calculated, with a focus on analyzing the temperature rise rate and heat conduction path in hotspot areas (such as the bushing root). Based on the calculated temperature gradient and the material's thermal expansion coefficient, the thermal stress distribution of the tank wall is calculated using linear thermal stress theory, extracting the maximum thermal stress value and its location as thermodynamic characteristic parameters. Simultaneously, by comparing the real-time calculated temperature rise rate with the normal threshold in the equipment's historical operating data, a thermal anomaly index (0-100% quantization value) is generated to reflect the degree of overheating risk of the equipment.

[0038] For triaxial vibration signals collected from the winding base, time-domain features (such as peak value and RMS value) and frequency-domain features (such as dominant frequency and band energy) are first extracted. Combining the material properties (elastic modulus and Poisson's ratio) of the transformer core and windings, a vibration-deformation conversion model is established: the displacement-time history curve is obtained based on the vibration acceleration integral, and then the dynamic deformation of the winding under vibration excitation is calculated using the modal superposition method. The distribution of deformation in the axial and radial directions of the winding is analyzed in detail, identifying the deformation amplitude and phase characteristics of deformation concentration areas (such as the winding ends). The deformation characteristics are compared with the normal deformation range in the equipment's factory test data to generate a deformation anomaly index (0-100% quantization value) to quantify the aging degree of the winding's mechanical structure.

[0039] The thermal stress characteristic parameters and thermal anomaly index obtained from thermodynamic analysis are fused with the deformation amplitude and deformation anomaly index obtained from deformation analysis. Simultaneously, visible light image analysis results are incorporated: target detection algorithms are used to identify equipment appearance defects (such as sleeve cracks and radiator leaks), generating appearance defect type codes and severity classifications (mild / moderate / severe). All the above characteristic parameters are categorized and combined according to their physical meaning to form a three-level feature structure containing thermodynamic feature sub-vectors (thermal stress, thermal anomaly index), mechanical deformation feature sub-vectors (deformation amplitude, deformation anomaly index), and appearance state feature sub-vectors (defect type, severity). Finally, through normalization, each sub-vector parameter is mapped to the [0,1] interval, and concatenated to generate a 12-dimensional physical constraint feature vector. This vector simultaneously contains multi-physics field coupling information of the equipment's thermal, mechanical, and appearance aspects, providing highly interpretable input parameters for subsequent health indicator prediction.

[0040] This embodiment combines the device's thermodynamic equations with material deformation rules to transform raw sensor data into feature vectors with clear physical meaning, effectively solving the problem of traditional feature extraction methods lacking physical constraints and significantly improving the reliability and interpretability of fault diagnosis results.

[0041] S103, based on physical constraint feature vectors, performs health indicator prediction through a lightweight LSTM network.

[0042] In this embodiment, the input physical constraint feature vector is preprocessed. For thermodynamic features (such as thermal stress and thermal anomaly index), mechanical deformation features (such as deformation amplitude and deformation anomaly index), and appearance features (such as defect type encoding and severity grading) contained in the vector, normalization is performed by grouping them according to feature type. For numerical features (such as thermal stress and deformation amplitude), a maximum-minimum normalization method is used to map them to the [0,1] interval, eliminating the influence of dimensional differences on model training. For classification encoding features (such as defect type), one-hot encoding is used to convert them into binary vectors to ensure consistent distances between different categories. For severity grading features, values ​​of 0.3, 0.6, and 1.0 are assigned for mild, moderate, and severe severity, respectively, converting them into continuous numerical values ​​to retain severity difference information. Finally, the processed sub-vectors are concatenated in their original order to generate a standardized physical constraint feature vector sequence, which serves as the input data for the LSTM network.

[0043] A lightweight LSTM network was designed to reduce model complexity and computational overhead. The network structure consists of a single-layer LSTM unit and a fully connected output layer. The hidden layer dimension of the LSTM unit is set to 32, reducing the parameter size by more than 50% compared to traditional LSTM networks. The weight matrices of the input gate, forget gate, and output gate adopt a shared parameter strategy to reduce model storage requirements through parameter reuse. At the same time, a gated linear unit (GLU) is introduced to replace the traditional tanh activation function, accelerating training convergence while maintaining nonlinear expressiveness. When initializing the network parameters, the weight matrix of the LSTM unit is initialized with a Xavier uniform distribution, and the bias term is initialized to 0; the weight matrix of the fully connected layer is initialized with a He normal distribution to adapt to the output characteristics of the ReLU activation function.

[0044] To capture the temporal evolution of equipment health status, a sliding window method is used to construct training samples. Using the current time *t* as a baseline, a fixed-length sequence of standardized physical constraint feature vectors is truncated forward to form the input window. The true values ​​of health indicators after time *t* (such as the percentage of remaining equipment lifespan) are used as the output labels. The step size of the sliding window is set to 1 hour, meaning a new sample is generated for every additional hour of data, ensuring temporal continuity between samples. Simultaneously, to enhance the model's adaptability to different equipment operating stages, subsets are divided in the dataset according to equipment operating time (e.g., initial break-in, mid-term stabilization, late-term aging), mixed proportionally, and then randomly shuffled to avoid overfitting caused by biased training data distribution.

[0045] The network was trained using the Adaptive Moment Estimator (Adam) optimizer with an initial learning rate of 0.001, dynamically adjusted during training using a cosine annealing strategy to balance the need for rapid convergence in the early stages with the need for fine-grained optimization in the later stages. The batch size was set to 16 to balance memory usage and gradient estimation stability. The training epochs were set to 100, with the validation set loss calculated after each epoch. Training was terminated early if the loss decreased by less than 0.1% for five consecutive epochs to prevent overfitting. The loss function was a weighted combination of mean squared error (MSE) and smooth L1 loss, where MSE penalizes large error samples and Smooth L1 loss enhances the model's robustness to outliers. After training, the parameters of the model that performed best on the validation set were saved as the final prediction model.

[0046] A trained lightweight LSTM network is used to predict health indicators. The model is input with a sequence of standardized physical constraint feature vectors collected in real time, and outputs the predicted health indicator value (e.g., percentage of remaining lifespan) for the current moment. To reduce prediction error, a moving average filter is introduced to smooth the prediction results over 10 consecutive moments, eliminating short-term fluctuations. Furthermore, the predicted value is dynamically corrected by combining baseline health indicator values ​​from historical equipment operating data (e.g., the average lifespan of the same model of equipment under similar operating conditions): if the current predicted value is lower than 80% of the baseline value, the result is adjusted by a correction factor of 0.8 to avoid misjudgments due to individual differences; if the predicted value is higher than 120% of the baseline value, the original value is maintained to prevent overly optimistic estimations. Finally, the corrected predicted health indicator value is output, providing a basis for subsequent degradation stage identification and maintenance decisions.

[0047] This embodiment achieves significant reduction in model computational resource requirements while maintaining prediction accuracy through deep fusion of a lightweight LSTM network and physical constraint feature vectors. It solves the problem of difficult deployment of traditional LSTM networks in real-time monitoring of power grid equipment, and provides an efficient and reliable prediction tool for equipment health management.

[0048] S104. When a continuous decline in health indicators is detected, the Gaussian mixture model is used to cluster the HI time series curve. The degradation stage is determined based on the distance between the cluster centers, and the stage identification results are obtained. The finite element simulation parameters are then dynamically switched based on the stage identification results.

[0049] In this embodiment, the real-time monitored health indicator (HI) time-series data is preprocessed: Statistical characteristics of the HI data within a fixed time window are calculated, including mean, variance, maximum, minimum, and trend slope. The trend slope is obtained by fitting the relationship between HI values ​​and time within the window using linear regression, and is used to quantify the rate of HI decline. Simultaneously, to eliminate short-term fluctuations, a moving average filter is used to smooth the original HI sequence, generating a smoothed HI time-series curve. Further morphological features of the curve are extracted, including the curvature peak (reflecting the severity of the HI decline), fluctuation frequency (reflecting data stability), and inflection point position (reflecting points of change in the degradation trend), forming a comprehensive feature vector containing both statistical and morphological features, which serves as input data for GMM clustering.

[0050] Based on comprehensive feature vectors, a Gaussian mixture model (GMM) is constructed to fit the degradation pattern distribution of the HI time-series curve. It is assumed that the equipment degradation process includes multiple potential stages (e.g., initial stabilization, intermediate degradation, and final critical stage), each corresponding to a Gaussian distribution. The GMM describes the overall degradation pattern by weighted combination of multiple Gaussian distributions. Model parameters include the mean vector, covariance matrix, and mixture weights of each Gaussian distribution. The Expectation-Maximization (EM) algorithm is used for training. During initialization, K-means clustering is used to divide the feature vectors into K classes, with the center of each class serving as the initial mean of the Gaussian distribution. The covariance matrix is ​​initialized as the identity matrix, and the mixture weights are initialized as a uniform distribution. During the EM iteration, the E-step calculates the posterior probability of each sample belonging to each Gaussian distribution, and the M-step updates the mean, covariance, and weight parameters based on the posterior probability until the model's log-likelihood converges or the maximum number of iterations is reached. After training, the parameters of each Gaussian distribution are saved as a baseline model for degradation stage identification.

[0051] The real-time collected comprehensive feature vectors are input into the trained Gaussian Mixture Model (GMM) to calculate the posterior probability of each Gaussian distribution. Samples are then classified into the stage corresponding to the distribution with the highest posterior probability. To improve the stability of stage identification, a sliding window mechanism is used. A window is formed by extracting feature vectors from 10 consecutive time steps backward from the current time step. Clustering is performed on all samples within the window, and the frequency of each stage is counted. If the frequency of a certain stage exceeds 70%, the current window is determined to belong to that stage; otherwise, it is marked as a transition stage. Furthermore, to quantify the confidence of stage identification, the Mahalanobis distance from the sample within the window to the center of the Gaussian distribution of its respective stage is calculated. If the mean distance exceeds a threshold, the confidence of stage identification is reduced, triggering a manual review mechanism. Finally, the degradation stage identification result at the current time step (e.g., initial stability, mid-term degradation, final critical stage) and the corresponding confidence level are output.

[0052] Based on the degradation stage identification results, the input parameters of the finite element simulation model are dynamically adjusted to match the current actual degradation state of the equipment. For different degradation stages, corresponding parameter mapping rules are preset—for example, in the initial stable stage, material parameters under the rated operating conditions of the equipment are used (such as setting the dielectric constant of the insulating material to the nominal value of 2.5, and the conductivity of the conductive core to 5.8 × 10⁻⁶). 7 S / m); During the intermediate degradation stage, parameters are adjusted by linear interpolation based on the stage confidence level (e.g., at a confidence level of 80%, the dielectric constant is adjusted to 2.3 and the conductivity to 5.5 × 10⁻⁶). 7 S / m), reflecting the performance degradation caused by insulation aging; in the final critical stage, extreme operating parameters are directly used (such as dielectric constant decreasing to 2.0, conductivity increasing to 6.0×10). 7The simulation (S / m) simulates the nonlinear characteristics of the equipment when it approaches a fault. Simultaneously, to enhance the simulation's sensitivity to local anomalies, infrared thermal imaging data is used to locate the high-temperature region of the equipment. The mesh in this region is then refined in the simulation model (mesh size reduced to 50% of the original size), and the material parameter adjustment weights for this region are increased to ensure that the simulation results accurately reflect the impact of local degradation on the overall state of the equipment.

[0053] To ensure the effectiveness of simulation parameters after dynamic switching, a parameter verification and feedback optimization mechanism is established. The simulated equipment status indicators (such as temperature distribution and stress concentration areas) are compared with real-time monitoring data (such as temperature measurement points in the SCADA system and spectral characteristics of vibration sensors), and the root mean square error (RMSE) between the two is calculated. If the RMSE exceeds a preset threshold, the parameter optimization process is triggered—using a Bayesian optimization algorithm to minimize the RMSE within a preset parameter range (such as dielectric constant [1.8, 2.8], conductivity [5.0 × 10⁻⁶]). 7 6.5×10 7 The system searches for the optimal parameter combination (S / m) and updates the optimized parameters to the parameter mapping rule base. Simultaneously, it records the time of each parameter switch, the stage identification results, and the RMSE changes before and after optimization, forming a parameter adjustment log to provide data support for subsequent model improvements. This mechanism enables dynamic matching between finite element simulation parameters and the actual degradation state of the equipment, improving the accuracy of predictive maintenance work orders.

[0054] This embodiment achieves accurate identification of equipment degradation stages through GMM clustering and dynamically adjusts finite element simulation parameters based on stage characteristics, solving the problem of state assessment deviation caused by fixed parameters in traditional simulation, and providing more reliable technical support for power grid equipment fault prediction and maintenance decision-making.

[0055] S105, based on finite element simulation parameters, introduces a reinforcement learning model to optimize simulation parameters with the goal of minimizing maintenance costs, and outputs predictive maintenance work orders.

[0056] In this embodiment, a high-precision finite element simulation model is constructed based on the physical structure of the power grid equipment (such as transformer winding layout and transmission line tower structure) and material properties (such as dielectric constant of insulating materials and conductivity of metallic conductors). The model needs to cover the key components and potential fault areas of the equipment (such as transformer winding hotspots and transmission line connection points), and be divided into fine mesh cells (such as mesh size ≤ 5mm in the winding area) to ensure simulation accuracy. The initial simulation parameters are set according to the rated operating conditions of the equipment. For example, the dielectric constant of the transformer insulation material is set to the nominal value of 2.5, and the conductivity of the conductive core is set to 5.8 × 10⁻⁶. 7The simulation output parameters are set to S / m, ambient temperature to 25℃, and load rate to 80%. Simultaneously, the correlation between simulation output parameters and equipment health status is defined. For example, a maintenance warning is triggered when the simulated maximum winding temperature exceeds 120℃, and insulation aging is determined to be accelerated when the CO concentration in the gas exceeds 300ppm.

[0057] The finite element simulation process is abstracted as a reinforcement learning environment, where the goal of the agent (reinforcement learning model) is to minimize maintenance costs by adjusting simulation parameters while meeting safety constraints (such as temperature and stress not exceeding thresholds). The state space is defined to contain three types of information: the real-time status of the equipment, the current values ​​of simulation parameters, and historical maintenance records.

[0058] The action space is designed as a strategy for adjusting simulation parameters, with each action corresponding to a set of parameter modification instructions. For example, action 1 is "reduce dielectric constant by 0.1 and conductivity by 5%", and action 2 is "increase ambient temperature by 2℃ and decrease load rate by 10%", etc. To avoid simulation distortion caused by excessive parameter adjustment, a parameter variation range for each action is set (e.g., dielectric constant adjustment range ≤ ±0.2, load rate adjustment range ≤ ±15%), and continuous parameter adjustments are converted into a finite number of selectable actions through discretization (e.g., the dielectric constant can take 5 discrete values: 2.3, 2.4, 2.5, 2.6, and 2.7).

[0059] A reward function is constructed using maintenance cost penalties, equipment safety rewards, and maintenance effectiveness rewards. A deep Q-network (DQN) is used as the reinforcement learning model, with the state vector as input and the Q-value of each action as output (representing the expected cumulative reward for performing that action). The trained reinforcement learning model is deployed to the power grid equipment monitoring system to receive equipment status data in real time and generate optimal simulation parameter adjustment strategies.

[0060] This embodiment uses a reinforcement learning model to dynamically optimize finite element simulation parameters, which significantly reduces maintenance costs while ensuring the safe operation of equipment, and generates accurate predictive maintenance work orders, effectively solving the problems of fixed parameters, high costs, and poor results in traditional maintenance methods.

[0061] S106 executes on-site operations based on predictive maintenance work orders and collects post-maintenance status data through AR devices.

[0062] In this embodiment, maintenance personnel receive predictive maintenance work orders pushed by the system via mobile terminals (such as tablets or AR glasses). The work order content includes equipment identification (such as "110kV transformer #3"), maintenance type (such as "winding insulation detection"), key detection indicators (such as "winding temperature, insulation resistance, partial discharge amplitude"), and expected state range (such as "winding temperature ≤90℃, insulation resistance ≥500MΩ, partial discharge amplitude ≤10pC"). The work order parsing module built into the AR device performs structured processing on the work order content, extracts key information, and generates a list of detection tasks. For example, "winding insulation detection" is broken down into three sub-tasks: "temperature measurement", "insulation resistance test", and "partial discharge detection". Each sub-task is configured with corresponding detection tools (such as infrared thermal imagers, megohmmeters, and UHF sensors) and data acquisition templates (such as recording the measurement point location, time, and value for temperature data). Meanwhile, the AR device matches the spatial coordinate system and key component location information of the corresponding equipment in the local database based on the equipment's 3D model (pre-imported via laser scanning or CAD drawings) and the equipment identifier in the work order (e.g., the relative coordinates of the transformer winding in the equipment are (X=0.5m, Y=1.2m, Z=0.3m)).

[0063] After maintenance personnel arrive at the equipment site wearing AR devices, they initiate the inspection task. The AR device identifies the physical equipment (such as a QR code or feature mark on the transformer casing) through a camera, calls a pre-registered 3D model of the equipment for spatial positioning and attitude estimation, and overlays virtual inspection guidance information (such as red arrows and highlighted areas) onto the real-world image of the equipment to guide maintenance personnel to quickly locate the inspection point. For example, when performing the "winding temperature measurement" task, the AR device displays a red highlighted box at the corresponding location on the winding and labels it "Please measure the temperature at this location"; when the maintenance personnel point the infrared thermal imager at the highlighted area, the AR device confirms the correct inspection point through image recognition algorithms (such as YOLO-based target detection), automatically triggers temperature data acquisition, and encapsulates the acquisition time, measurement point coordinates (such as "upper part of winding A phase (0.5,1.2,0.3)"), temperature value (such as "85℃"), and equipment identification into a structured data packet and stores it in the local cache. For data that requires multi-angle collection (such as infrared thermal imaging which requires shooting from both the front and side of the device), the AR device guides maintenance personnel to complete multi-view collection through voice prompts (such as "Please move to the right side of the device and keep a distance of 1 meter") and virtual path guidance (displaying a blue arrow path on the ground), ensuring data integrity.

[0064] The data fusion module built into the AR device performs correlation processing on the collected multimodal data. On one hand, it aligns time-series data (such as temperature curves over time) with spatial image data (such as infrared thermal images) using timestamps and spatial coordinates. For example, it matches the temperature measurement time "14:30:00" with the infrared thermal image acquisition time and maps the spatial coordinates (0.5, 1.2, 0.3) of the temperature measurement point to the corresponding pixel position in the infrared thermal image, generating "temperature-thermal image" correlated data. On the other hand, it combines the device's historical operating data (such as the average temperature of the winding over the past 30 days being 75℃) with the expected state range in the work order (winding temperature ≤90℃) to perform real-time verification of the collected data. If abnormal data is detected (e.g., the current temperature is 85℃, which is within the limit, but is 10℃ higher than the historical average and continues to rise), the AR device will alert the maintenance personnel through voice alarms (e.g., "Temperature is abnormally high, please recheck") and highlight the interface (displaying the temperature value in red), and automatically trigger the retest process (e.g., requiring the temperature at this point to be measured 3 times and the average value is taken); if the retest result is still abnormal, a data anomaly mark will be generated and uploaded to the system backend along with the original data for further analysis by the fault diagnosis model.

[0065] After completing all testing tasks, maintenance personnel click the "Data Upload" button on the AR device interface. The device then transmits its locally cached structured data package (containing test data such as temperature, insulation resistance, and partial discharge, as well as metadata such as device identification, acquisition time, and spatial coordinates) to the power grid equipment monitoring system via a 5G / WiFi network with encryption. Upon receiving the data, the system backend first performs an integrity check (e.g., checking if all sub-task data is included). If data is missing, a supplementary acquisition command is sent to the AR device (e.g., "Partial discharge data is missing, please retest"). If the data is complete, the system compares and analyzes the post-maintenance status data with the device's historical data and benchmark data from similar devices, generating a maintenance effectiveness evaluation report (e.g., "After maintenance, the winding temperature decreased by 10℃, insulation resistance increased by 20%, and the partial discharge amplitude decreased to 5pC, demonstrating significant maintenance effectiveness"), and pushes the report to the maintenance personnel's mobile terminal. Simultaneously, the system updates the device's health indicators based on the post-maintenance status data (e.g., adjusting the health indicator from "mid-stage degradation" to "healthy state"), providing more accurate input for subsequent fault prediction.

[0066] In some embodiments, step S102 above, which involves generating a physical constraint feature vector based on the original data and combining the device's thermodynamic equations with material deformation rules, specifically includes: The raw data was separated into temperature sequences, vibration spectra, and infrared thermal images. Based on the equipment's heat dissipation structure topology diagram, spatial region segmentation is performed on the infrared thermal image, and the temperature change rate of each heat sink unit is calculated by combining the temperature sequence. Based on the theory of winding mechanical deformation, the axial and radial resonant frequency offsets are analyzed from the vibration spectrum. By associating the temperature change rate and resonant frequency offset of each heat sink unit, a combined mechanical-thermodynamic feature identifier is output, forming a physical constraint feature vector.

[0067] In this embodiment, a data stream classification algorithm (such as rule matching based on timestamps and sensor identifiers) is used to split the raw data into three sub-data sets: temperature sequence, vibration spectrum, and infrared thermal image.

[0068] To analyze the local thermal state of the equipment's heat dissipation system, it is necessary to segment the infrared thermal image into spatial regions in conjunction with the equipment's heat dissipation structure topology diagram. The heat dissipation structure topology diagram is generated from the equipment's CAD model or laser scanning point cloud and includes the geometric boundaries and spatial coordinate information of key components such as heat sinks, air ducts, and oil ducts. For example, the transformer heat sink topology diagram defines the vertex coordinates of each heat sink (e.g., the four vertex coordinates of heat sink 1 are (x1,y1,z1), (x2,y2,z2), (x3,y3,z3), (x4,y4,z4)) and the spacing between adjacent heat sinks.

[0069] Based on the topology map, a polygon region filling algorithm (such as scan line filling) is used to segment the infrared thermal image into multiple heat sink unit regions, each corresponding to an independent heat sink. For example, pixels in the infrared thermal image whose coordinates satisfy the area enclosed by the vertices of heat sink 1 are marked as "heat sink 1 region". Subsequently, the temperature change rate of each heat sink unit is calculated by combining the temperature sequence: the average temperature of all pixels in each heat sink region is taken to generate the time-temperature curve of that heat sink (e.g., "heat sink 1 temperature curve: 10:00, 65℃; 10:05, 66℃; ...; 10:30, 70℃"). Then, the slope of the curve is fitted by linear regression to obtain the temperature change rate (e.g., the temperature change rate of heat sink 1 is 0.2℃ / min). This process is repeated to complete the calculation of the temperature change rate of all heat sink units (e.g., a transformer with 12 heat sinks).

[0070] Under mechanical stress (such as short-circuit impact or transportation vibration), the winding will deform, causing a shift in its natural resonant frequency. To quantify the degree of deformation, the axial and radial resonant frequency shifts need to be analyzed from the vibration spectrum. First, a theoretical resonant frequency reference value is determined based on the equipment design parameters (e.g., the design value of the winding's axial natural frequency is 100Hz, and the radial frequency is 150Hz). Then, frequency components close to the reference value are searched in the vibration spectrum (e.g., 102Hz in the axial spectrum and 148Hz in the radial spectrum), and the actual resonant frequency is confirmed using a peak detection algorithm (e.g., local maximum search). Finally, the shift between the actual frequency and the reference value is calculated (e.g., axial shift = 102Hz - 100Hz = +2Hz, radial shift = 148Hz - 150Hz = -2Hz). If no obvious resonant peak is detected in the spectrum (e.g., the amplitude is below the threshold of 0.02g), it is marked as "unidentified," and the shift is recorded as 0Hz. Finally, the axial and radial resonant frequency offsets of the winding can be obtained (e.g., "axial offset +2Hz, radial offset -2Hz"), reflecting the current mechanical deformation state of the winding.

[0071] To construct features that reflect the mechanical-thermodynamic coupling state of the equipment, it is necessary to associate the temperature change rate of the heat sink with the resonant frequency offset of the winding. Based on the coupling mechanism of the thermodynamic and mechanical deformation of the equipment (such as the change in contact resistance caused by the deformation of the winding, which leads to local overheating; the blockage of the heat sink causes the winding temperature to rise, which intensifies the mechanical stress), cross-modal feature association rules are designed: (1) Spatial association: According to the topology of the equipment structure (such as the winding being located inside the heat sink, 200mm away from the center of heat sink 1), the resonant frequency offset of the winding is spatially matched with the temperature change rate of the adjacent heat sink (such as heat sink 1-3), for example, "the axial offset of the winding +2Hz is associated with the temperature change rate of heat sink 1 of 0.2℃ / min and the temperature change rate of heat sink 2 of 0.15℃ / min". (2) Trend correlation: Analyze the consistency of the time trend of temperature change rate and resonant frequency offset. If both increase synchronously (e.g., temperature change rate increases from 0.1℃ / min to 0.2℃ / min, axial offset increases from +1Hz to +2Hz), it is marked as "positive correlation"; if the trend is opposite (e.g., temperature increases but offset decreases), it is marked as "negative correlation"; if there is no obvious trend, it is marked as "irrelevant". (3) Weight allocation: According to historical data statistics (e.g., the influence of heat sink 1 on winding temperature is 0.4, and that of heat sink 2 is 0.3), assign weights to each correlation feature. For example, "temperature change rate of heat sink 1 weight 0.4, weight of heat sink 2 weight 0.3, axial offset weight 0.3". Based on the above correlation rules, generate mechanical-thermodynamic joint feature identifiers, such as "[temperature change rate of heat sink 1 0.2℃ / min (weight 0.4), temperature change rate of heat sink 2 0.15℃ / min (weight 0.3), axial offset +2Hz (weight 0.3), correlation trend: positive correlation]". All joint feature identifiers are classified according to equipment components (such as windings, heat sinks) or physical processes (such as heat conduction, mechanical vibration), and combined into physical constraint feature vectors (such as "vector dimension 1: winding axial offset +2Hz; dimension 2: heat sink 1 temperature change rate 0.2℃ / min; dimension 3: correlation trend positive correlation;..."), which serve as inputs for subsequent health indicator prediction.

[0072] This embodiment achieves accurate characterization of the mechanical-thermodynamic coupling state of the equipment through multi-source data separation, spatial modeling of heat dissipation structure, vibration spectrum analysis, and cross-modal feature correlation.

[0073] Furthermore, the step of performing spatial region segmentation on the infrared thermal image based on the device's heat dissipation structure topology diagram and calculating the temperature change rate of each heat sink unit specifically includes: Load the heat dissipation structure topology diagram of the power grid equipment, and combine it with the physical coordinate data of the heat sink to generate a standardized heat dissipation unit spatial grid; For the real-time acquired infrared thermal image, based on the spatial grid of the heat dissipation unit, the key points of the heat sink outline are located by the corner detection algorithm. The actual image and the topological map are then registered by affine transformation to obtain the feature registration result. Based on the feature registration results, the infrared thermal image is segmented into independent regions corresponding to the heat dissipation unit, and the pixel temperature matrix of each independent region is extracted. For each independent region's pixel temperature matrix, the temperature matrix of adjacent frames is compared according to the time series to calculate the rate of temperature change per unit time.

[0074] In this embodiment, a heat dissipation structure topology diagram of the power grid equipment is loaded from an equipment design database or a 3D CAD model. This topology diagram needs to include key information such as the geometric dimensions, installation location, and adjacency relationships of the heat sinks. For example, the transformer heat sink topology diagram is stored in the form of a two-dimensional vector diagram, defining the coordinates of the four vertices of each heat sink, the spacing between heat sinks, and their arrangement direction.

[0075] Based on the physical coordinate data in the topology diagram, a standardized spatial mesh for heat dissipation units is generated. Using the center of the heat sink as the mesh node, rectangular mesh cells are divided according to the actual size of the heat sink, and each mesh cell is assigned a unique identifier. Simultaneously, the mapping relationship between the mesh cells and the actual physical location of the device is recorded to ensure consistency between the mesh space and the device's physical space. The final standardized spatial mesh is stored in JSON or XML format, containing the coordinates, dimensions, and identifier information of the mesh cells.

[0076] For real-time acquired infrared thermal images, they need to be registered with a standardized spatial grid to eliminate spatial deviations caused by factors such as image shooting angle and equipment installation tilt. First, a corner detection algorithm (such as Harris corner detection) is used to extract key feature points of the heat sink outline from the infrared thermal image, such as detecting inflection points of the heat sink edge (e.g., the boundary between the heat sink and the background, gaps between heat sinks), generating a set of actual image feature points. Then, theoretical feature points corresponding to the heat sink outline are extracted from the standardized spatial grid, and an affine transformation algorithm (such as parameter estimation based on least squares) is used to calculate the transformation matrix (including rotation, translation, and scaling parameters) between the actual image and the theoretical topology. For example, if the heat sink in the actual image is shifted 20 pixels to the right and rotated 5 degrees clockwise, the transformation matrix needs to adjust the image coordinates to match the topological grid. Finally, the transformation matrix is ​​applied to the entire infrared thermal image to complete the affine transformation registration between the image and the topological grid, generating feature registration results to ensure that subsequent segmentation is based on the accurate physical structure of the device.

[0077] Based on the feature registration results, the infrared thermal image is segmented into independent regions corresponding to the heat dissipation units. Each grid cell in the standardized spatial grid is traversed, and the pixel region covered by that grid cell in the registered image is extracted. For example, if "Heat Disk 1 - Region A" is defined in the topology map as having coordinates from (x1-100, y1-25) to (x1+100, y1+25), then pixel blocks with the same coordinate range are extracted from the registered image as independent regions. For each independent region, its pixel temperature matrix is ​​extracted, and all pixel values ​​within the region are arranged in row and column order into a two-dimensional matrix. The physical dimensions of the matrix and the acquisition timestamp are recorded. Through this step, the infrared thermal image is finely segmented from the overall to the local level, generating a temperature dataset corresponding one-to-one with the heat dissipation units of the device.

[0078] For the temperature matrix of each independent region, the matrix data of adjacent frames are compared according to the time series to calculate the temperature change rate per unit time. Continuously acquired infrared thermal imaging sequences are obtained from the equipment monitoring system. The feature registration and region segmentation extraction steps described above are repeated for each frame to obtain the temperature matrix of each heat sink unit at each time point. Then, the temperature matrices of adjacent frames are compared pixel-by-pixel, and the temperature difference of each pixel in adjacent frames is calculated. The average difference of all pixels within a region is then taken to obtain the average temperature change of the heat sink unit within adjacent time intervals. Finally, the temperature change rate is calculated based on the time interval, and the temperature change is divided by the time interval to obtain the temperature change rate per unit time. This process is repeated to complete the calculation of the temperature change rate of all heat sink units across all time periods.

[0079] This embodiment achieves accurate calculation of the temperature change rate at the heat sink unit level through topological graph mesh modeling, image affine registration, independent region segmentation, and dynamic temperature analysis. This significantly improves the spatial resolution and temporal accuracy of equipment thermal status monitoring, providing reliable data support for early warning of overheating faults in power grid equipment.

[0080] Furthermore, the step of analyzing the axial and radial resonant frequency shifts from the vibration spectrum based on the winding mechanical deformation theory specifically includes: Based on the transformer winding design drawings, establish the reference resonant frequency values ​​for the axial and radial directions, and simultaneously associate them with the no-load operating conditions of the equipment. For the raw time-domain signal acquired by the triaxial vibration sensor, bandpass filtering is performed to retain the target frequency band, and windowing is applied to suppress spectral leakage to obtain the preprocessed time-domain signal; Based on the preprocessed time-domain signal, a power spectral density curve is generated through fast spectrum transformation to accurately locate the main resonance peak frequency. By comparing the main resonance peak frequency and the reference resonance frequency value, the percentage offsets in the axial and radial directions are calculated respectively to obtain the axial and radial resonance frequency offsets.

[0081] In this embodiment, firstly, detailed design drawings of the winding are obtained from the transformer design department, including the winding axial height, radial thickness, number of coil turns, insulation material properties, and support structure parameters. Based on these parameters, and combined with modal analysis models in the winding mechanical deformation theory (such as simplifying the winding as a multi-degree-of-freedom spring-mass system), the axial and radial reference resonant frequencies of the winding under healthy conditions are calculated using finite element simulation software (such as ANSYS or COMSOL).

[0082] The calculated reference resonant frequency is correlated with the equipment's no-load operating condition: During the transformer no-load test, vibration signals are collected using a triaxial vibration sensor (installed on the top of the transformer tank, capable of simultaneously acquiring axial, radial, and tangential vibrations) to verify the accuracy of the reference resonant frequency. If the deviation between the main resonance peak frequency of the measured vibration spectrum under no-load conditions and the calculated value is less than 5%, the reference value is confirmed to be valid; if the deviation exceeds the threshold, the design parameters need to be adjusted (such as correcting the elastic modulus of the insulation material) and recalculated until the reference value matches the measured no-load value, ensuring that the reference model can accurately reflect the vibration characteristics of the winding under healthy conditions.

[0083] For the raw time-domain signals (such as voltage signals with a sampling frequency of 10kHz and a duration of 10 seconds, with an amplitude range of -5V to +5V) collected by triaxial vibration sensors during transformer operation under load, preprocessing is required to suppress noise and retain the target frequency band. First, based on the winding vibration characteristics analysis, the target frequency band range is determined. The main energy of axial vibration is concentrated in the 600-1000Hz range, and the main energy of radial vibration is concentrated in the 1000-1500Hz range. Therefore, a bandpass filter is designed, setting the passband of the axial channel to 600-1000Hz and the passband of the radial channel to 1000-1500Hz. The raw signal is filtered to remove low-frequency interference (such as 50Hz power frequency noise) and high-frequency noise (such as sensor electronic noise), retaining the target frequency band signal containing winding resonance information. Subsequently, the Hanning window is selected as the window function, characterized by smooth attenuation at both ends of the time domain, which can effectively reduce the "sidelobe effect" in spectrum analysis. The Hanning window is applied to the filtered signal to generate the windowed time-domain signal. This step ensures the accuracy of subsequent spectrum transformation and avoids frequency component aliasing caused by signal truncation.

[0084] Based on the preprocessed time-domain signal, it is converted into a frequency-domain signal using a fast spectral transform (such as Fast Fourier Transform, FFT), generating a power spectral density curve (PSD). Specifically, an FFT is performed on each windowed signal segment to obtain a complex frequency-domain representation, and then the square of its amplitude (i.e., the power spectral density) is calculated, with units of V. 2 / Hz. Averaging the power spectral density of all signal segments (e.g., averaging the PSD of 10 signal segments) generates a smooth power spectral density curve, which clearly reflects the distribution of vibrational energy at different frequencies. Then, traversing all frequency points within the target frequency band, the frequency corresponding to the maximum power spectral density is found; this is the main resonance peak frequency. For example, if the power spectral density curve of the axial channel reaches its maximum at 865Hz, then the axial main resonance peak frequency is confirmed to be 865Hz; if the radial channel reaches its maximum at 1220Hz, then the radial main resonance peak frequency is confirmed to be 1220Hz. To improve positioning accuracy, a quadratic interpolation method (such as parabolic interpolation) can be used to refine the analysis of frequency points near the main resonance peak, further correcting the frequency values.

[0085] For the axial direction, substituting the measured main resonance peak frequency (e.g., 865.3Hz) and the reference resonance frequency (e.g., 850Hz) into the offset calculation formula "Offset = (Measured value - Reference value) / Reference value × 100%", we get the axial offset as (865.3-850) / 850×100%≈1.8%; for the radial direction, substituting the measured main resonance peak frequency (e.g., 1220Hz) and the reference resonance frequency (e.g., 1200Hz) into the formula, we get the radial offset as (1220-1200) / 1200×100%≈1.67%.

[0086] Finally, the axial and radial resonant frequency offsets are obtained (e.g., 1.8% axially and 1.67% radially). These offsets directly reflect the degree of mechanical deformation of the winding in the axial and radial directions. Based on engineering experience, when the axial or radial offset exceeds 3%, it can be determined that there is significant deformation of the winding (e.g., loose insulation pads or broken support bars), requiring further inspection and repair. If the offset is between 1% and 3%, it indicates that the winding is in the early deformation stage, requiring enhanced monitoring.

[0087] This embodiment provides an objective and quantifiable indicator for assessing the mechanical condition of transformer windings by quantifying the resonant frequency offset, effectively improving the accuracy and timeliness of fault diagnosis.

[0088] Furthermore, the temperature change rate and resonant frequency offset of each associated heat sink unit are used to output a combined mechanical-thermodynamic feature identifier, forming a physical constraint feature vector, specifically including: The temperature change rate of each heat sink unit is bound to the resonant frequency offset in the axial and radial directions of the same region, respectively, to form thermo-mechanical axial coupling parameters and thermo-mechanical radial coupling parameters. Based on the thermo-mechanical axial coupling parameters and thermo-mechanical radial coupling parameters, the thermal expansion coefficient of the equipment material is introduced to calculate the theoretical deformation threshold range; When the actual coupling parameters exceed the theoretical deformation threshold range, they are marked as abnormal associated regions, and a mechanical-thermodynamic joint feature identifier is generated. The mechanical-thermodynamic joint feature identifier includes a list of normal coupling region numbers and the location and over-limit level of the abnormal coupling region.

[0089] In this embodiment, based on the transformer heat sink layout and winding zoning design, the heat sink is divided into multiple units (e.g., 8 heat sink units are evenly divided along the circumference of the oil tank, with each unit covering 1 / 8 of the oil tank surface). Temperature data for each unit is collected in real time using distributed temperature sensors (e.g., PT100 platinum resistance thermometers, installed in the middle of each heat sink unit) (e.g., sampling frequency of 1 time / minute, continuous collection for 24 hours). The temperature change rate of each unit is calculated: using a 10-minute time window, the temperature difference between the current moment and the moment 10 minutes ago is calculated, and then divided by the time interval (10 minutes) to obtain the temperature change rate (unit: ℃ / min). For example, if the temperature of heat sink unit 3 rises from 45℃ to 47℃ in a certain time window, its temperature change rate is (47-45) / 10 = 0.2℃ / min.

[0090] Simultaneously, based on the axial and radial resonant frequency offsets of the windings (e.g., axial offset 1.8%, radial offset 1.67%), regions are bound according to the correspondence between heat sink units and winding regions (e.g., heat sink unit 1 corresponds to the axial region A of the high-voltage winding, and heat sink unit 2 corresponds to the radial region B of the low-voltage winding). The temperature change rate of heat sink unit 1 (e.g., 0.15℃ / min) is bound to the axial resonant frequency offset of region A (e.g., 1.8%) to form thermo-mechanical axial coupling parameters (e.g., "Unit 1 - Axial: Temperature change rate 0.15℃ / min, offset 1.8%)); the temperature change rate of heat sink unit 2 (e.g., 0.22℃ / min) is bound to the radial resonant frequency offset of region B (e.g., 1.67%) to form thermo-mechanical radial coupling parameters (e.g., "Unit 2 - Radial: Temperature change rate 0.22℃ / min, offset 1.67%)".

[0091] Based on thermo-mechanical axial coupling parameters and thermo-mechanical radial coupling parameters, the thermal expansion coefficient of the equipment material is introduced (e.g., the linear expansion coefficient of copper conductors commonly used in transformer windings is 16.5 × 10⁻⁻⁻⁴). 6 / ℃, the coefficient of linear expansion of the insulating paperboard is 5×10⁻ 6The theoretical deformation threshold range is calculated based on the temperature (°C) and mechanical deformation constraints (such as axial constraint of the winding by the end pressure plate and radial constraint by the support bars). In specific operations, for thermo-mechanical axial coupling parameters (such as temperature change rate of 0.15°C / min and offset of 1.8%), combined with the axial thermal expansion coefficient of the copper conductor, the axial thermal expansion caused by temperature increase is estimated (e.g., when the temperature increases by 10°C, the axial expansion = thermal expansion coefficient × temperature change × axial length = 16.5 × 10⁻⁻⁻⁶). 6 ×10×1.2m=0.198mm); At the same time, according to the winding mechanical deformation model (such as simplifying the winding into a spring-mass system), the resonant frequency offset of 1.8% is converted into equivalent mechanical deformation (such as by modal analysis, the 1.8% offset corresponds to 0.15mm of axial deformation).

[0092] Considering both thermal expansion and mechanical deformation, and taking into account material creep and fatigue accumulation effects (such as a 10% decrease in the material's elastic modulus after long-term operation), a theoretical deformation threshold range is set. The upper limit of the axial deformation threshold is 1.2 times the sum of thermal expansion and mechanical deformation, such as (0.198 + 0.15) × 1.2 ≈ 0.42 mm; the lower limit is 0.8 times the difference between the two, such as (0.198 - 0.15) × 0.8 ≈ 0.04 mm. The radial deformation threshold range is calculated using a similar method, such as radial thermal expansion = 5 × 10⁻⁻⁴ mm. 6 ×10×0.3m=0.015mm, mechanical deformation 0.12mm, threshold range is (0.015+0.12)×1.2≈0.162mm to (0.015-0.12)×0.8≈-0.084mm, the absolute value is 0-0.162mm.

[0093] The actual coupling parameters (such as the measured deformation of 0.5 mm corresponding to the axial thermo-mechanical coupling parameters of heat sink unit 3) are compared with the theoretical deformation threshold range (such as 0.04-0.42 mm). If the measured deformation exceeds the upper limit of the threshold (0.5 mm > 0.42 mm), the winding area corresponding to the heat sink unit (such as area A) is marked as an abnormal correlation area; if the measured deformation is within the threshold range (such as 0.2 mm ∈ [0.04, 0.42]), it is marked as a normal coupling area. The winding regions corresponding to all heat sink units are traversed to compile a list of normal coupling zone numbers (e.g., "Normal Zone: Region A2, Region B1, Region C3"), and the locations of abnormal coupling zones (e.g., "Abnormal Zone: Region A1 - Axial") and their exceedance levels are recorded (based on the extent of exceedance, e.g., exceedance within 10% is Level 1, 10%-30% is Level 2, and above 30% is Level 3; if the measured deformation of Region A1 is 0.5mm, the upper limit of the threshold is 0.42mm, and the exceedance is (0.5-0.42) / 0.42≈19%, then the exceedance level is Level 2). Finally, the list of normal coupling zones and the information of abnormal coupling zones are combined into a mechanical-thermodynamic joint feature identifier (e.g., "Normal Zone: A2, B1, C3; Abnormal Zone: A1 - Axial - Level 2"), which can intuitively reflect the distribution of the thermo-mechanical state of the equipment.

[0094] Key information from the combined mechanical-thermodynamic feature identifier (such as the temperature change rate, resonant frequency offset, deformation, and over-limit level of each region) is extracted into a physical constraint feature vector (e.g., vector dimensions are [temperature change rate, axial offset, radial offset, axial deformation, radial deformation, over-limit level], corresponding to the data for region A1 as [0.15, 1.8%, 1.2%, 0.5mm, 0.3mm, Level 2]). This feature vector integrates multi-dimensional information from thermodynamics and mechanical deformation and can be used as input parameters for transformer condition assessment. When the over-limit level in the feature vector is Level 2 or 3, an early warning signal is triggered (e.g., pushed to the maintenance terminal via the SCADA system), prompting a check of the heat sink's heat dissipation efficiency (e.g., cleaning accumulated dust) or the winding mechanical structure (e.g., tightening the support bars); when all regions are normally coupled, the thermo-mechanical state of the equipment is confirmed to be stable, and the regular monitoring cycle is maintained.

[0095] This embodiment achieves collaborative analysis of the multi-physical field state of the transformer by quantifying the thermo-mechanical coupling relationship, which significantly improves the comprehensiveness and accuracy of fault diagnosis.

[0096] In some embodiments, step S103 above, which involves performing health indicator prediction using a lightweight LSTM network based on physical constraint feature vectors, specifically includes: Based on the physical constraint feature vector, the normalization module transforms each feature value to a preset numerical range to obtain standardized features. Standardized features are input into a pre-trained lightweight LSTM network to perform rolling time window prediction and output health indicator values ​​and their confidence levels within a preset future time period. When the confidence level is lower than the preset confidence threshold, the entire LSTM model is re-predicted, and a health risk level label is generated. Based on the health risk level identifier, a preset maintenance strategy library is matched, and when the risk level is determined to be high, a maintenance work order is pushed to the operation and maintenance terminal in real time.

[0097] In this embodiment, the normalization module converts each feature value to a preset numerical range (e.g., [0,1]). Specifically, for numerical features (e.g., temperature change rate 0.15℃ / min, deformation 0.5mm), the maximum-minimum normalization method is used: the maximum value (e.g., maximum temperature change rate 0.3℃ / min, maximum deformation 1.0mm) and minimum value (e.g., minimum temperature change rate 0.05℃ / min, minimum deformation 0.1mm) of the feature in historical data are traversed, and the normalized value is calculated (e.g., normalized value of temperature change rate = (0.15-0.05) / (0.3-0.05) = 0.4, normalized value of deformation = (0.5-0.1) / (1.0-0.1) = 0.44); for categorical features (e.g., the level of exceeding the limit "Level 2"), one-hot encoding is used to convert them into binary vectors (e.g., "Level 2" is encoded as [0,1,0], assuming the level is divided into Level 1, Level 2, and Level 3). All normalized numerical features are concatenated with the encoded categorical features to form a standardized feature vector (e.g., the concatenated vector: [0.4,0.6,0.4,0.44,0.33,0,1,0], where 0.6 is the 1.8% normalized value of the resonance frequency offset and 0.33 is the 0.3mm normalized value of the radial deformation). This step eliminates the dimensional differences between different features, improving model training efficiency and prediction stability.

[0098] The standardized feature vectors are input into a pre-trained lightweight LSTM network (trained on historical data, with a 2-layer LSTM structure, 32 units per layer, and an output dimension of 1, reducing the total number of parameters by 60% compared to a full LSTM). Using the current time as a baseline, the standardized feature vectors from the previous N time steps (e.g., N=10, representing monitoring data from the past 10 minutes) are selected as the input sequence to predict health indicator values ​​for the next M time steps (e.g., M=5, predicting health indicators for the next 5 minutes). For example, if the current time is T, the input sequence is the standardized feature vectors from time steps [T-10, T-9, ..., T-1], and the output is the predicted health indicator values ​​from time step T+1 to T+5 (e.g., the predicted values ​​for winding hotspot temperatures are 85℃, 86℃, 87℃, 88℃, and 89℃, respectively). Simultaneously, the network outputs the confidence score for each predicted value (e.g., estimated using Monte Carlo dropout, the confidence score for the predicted value of 85℃ at time step T+1 is 92%, indicating the reliability of the prediction). Through lightweight design, the model takes less than 50ms for a single prediction, meeting real-time requirements.

[0099] The confidence level of the health indicators output by the lightweight LSTM network is compared with a preset confidence threshold (e.g., 90%). If the confidence level of the predicted value at a certain time is lower than the threshold (e.g., the confidence level of the predicted value of 87℃ at time T+3 is 88% < 90%), then that time is marked as a low-confidence time, and the full LSTM model is re-predicted. The full LSTM model has a more complex structure (e.g., 4 layers of LSTM units, 128 units per layer), more parameters but higher prediction accuracy. Its input data is the same as the lightweight model, but it captures long-term dependencies through a deeper network structure. After re-prediction, if the difference between the output of the full model and the lightweight model exceeds a preset threshold (e.g., the temperature difference exceeds 2℃), then the result of the full model is taken as the standard, and the predicted sequence of health indicators is updated (e.g., the predicted values ​​at times T+3 to T+5 are corrected to 89℃, 90℃, and 91℃); if the difference is within the threshold, then the result of the lightweight model is retained. Meanwhile, based on the predicted health indicator values ​​and their changing trends (such as the hotspot temperature continuing to rise and exceeding the warning value of 90℃ in the next 5 minutes), combined with confidence information (such as the proportion of low confidence moments exceeding 30%), risk levels are classified as low risk, medium risk, and high risk. If the predicted future temperature values ​​are all below 90℃ and the confidence level is above 90%, it is considered low risk. If there are times when the temperature exceeds 90℃ or there are low confidence moments, it is considered medium risk. If the temperature exceeds 95℃ or the proportion of low confidence moments exceeds 50%, it is considered high risk.

[0100] Based on the health risk level identifier, the system matches corresponding maintenance actions from a preset maintenance strategy library. For example, the maintenance strategy library includes "Low risk: Strengthen monitoring, extend the cycle to 2 hours / time; Medium risk: On-site inspection, check the heat dissipation system and load status; High risk: Immediate shutdown for maintenance." When a high-risk level is determined (e.g., the health risk level identifier is "High Risk"), the system automatically generates a maintenance work order. The work order includes information such as equipment number, risk level, predicted failure time, and recommended maintenance actions, such as "Equipment T-001, High risk, hot spot temperature is expected to reach 95℃ within 10 minutes, immediate shutdown is required to check winding insulation," and pushes it to maintenance personnel in real time through the SCADA system or mobile maintenance terminal (push methods include SMS, APP notification, audible and visual alarms, etc.).

[0101] After receiving a work order, maintenance personnel perform repairs according to the recommended actions, such as checking for blocked heat sinks, overload, and aging winding insulation. Upon completion of the repairs, they provide feedback on the results, such as "Heat sink cleaning completed, load adjusted to rated value, equipment restored to operation." The system records this feedback information in the historical database for subsequent model optimization and strategy updates. This process achieves closed-loop management of equipment health status, significantly improving maintenance efficiency and equipment reliability.

[0102] To adapt to changes in equipment operating conditions (such as seasonal load fluctuations and material aging), the lightweight LSTM network and the full LSTM model are dynamically updated periodically (e.g., monthly). The latest monitoring data (such as physical constraint feature vectors and measured health indicator values ​​from the past 30 days) is collected and divided into training and validation sets. Transfer learning methods are used to fine-tune model parameters (e.g., fixing the weights of the bottom LSTM units and adjusting only the parameters of the top output layer) to adapt the model to the data distribution under new operating conditions. Simultaneously, the judgment rules for health risk level identification are revised based on maintenance feedback (such as the results of maintenance work orders). For example, if a temperature prediction value is found to be high under a certain operating condition but no actual fault occurs, the warning value and risk level threshold for that condition are adjusted to improve the long-term prediction accuracy of the model. Through this dynamic update mechanism, the model is ensured to maintain optimal predictive performance at all times, meeting the health management needs of the equipment throughout its entire lifecycle.

[0103] In some embodiments, step S104 above, when a continuous decline in health indicators is detected, uses a Gaussian mixture model to cluster the HI time series curve, determines the degradation stage based on the cluster center distance, obtains the stage identification result, and dynamically switches the finite element simulation parameters based on the stage identification result, specifically including: Real-time monitoring of health indicator time-series curves; when a monotonically decreasing indicator value is detected for K consecutive sampling periods, a degradation stage analysis command is triggered. In response to the degradation stage analysis command, historical health index data of the equipment is retrieved, and a probability distribution cluster is generated through a Gaussian mixture model. The cluster centers of the probability distribution cluster correspond to the initial, middle or final degradation stage. Calculate the Mahalanobis distance between the current health indicator trajectory and each cluster center, determine the current degradation stage based on the minimum distance principle, and obtain the stage identification result; Based on the stage identification results, the corresponding finite element simulation parameters are automatically loaded. If it is determined to be the initial degradation stage, the standard material parameter library is enabled. If it is determined to be the final degradation stage, the historical damage case parameter set is loaded.

[0104] In this embodiment, a sensor network (such as vibration acceleration sensors, temperature sensors, and current transformers) deployed on key components of the equipment (such as gearbox bearings and motor stator windings) is used to collect multi-dimensional data reflecting the health status of the equipment in real time (such as vibration spectrum energy, winding temperature, and current harmonic distortion rate). Based on a predefined health index calculation model (such as weighted summation of vibration spectrum energy and normalized mapping of winding temperature and load), a health index time series curve (HI(t)) is generated.

[0105] The health indicator time-series curve is continuously updated with a fixed sampling period (e.g., data is collected once per minute). When a monotonically decreasing indicator value is detected for K consecutive sampling periods (e.g., K=5, i.e., 5 consecutive minutes) (e.g., HI(t) decreases from 85 to 82, 80, 78, 76, 75), and the decrease exceeds a preset threshold (e.g., a single decrease ≥1% or a cumulative decrease ≥5%), it is determined that the equipment health indicator shows a significant degradation trend, triggering a degradation stage analysis command. This command is pushed to the edge computing node (e.g., embedded industrial control computer) through an industrial IoT platform (e.g., an OPC UA server) to initiate the subsequent degradation stage identification process.

[0106] In response to the degradation stage analysis command, the edge computing node retrieves historical health indicator data (such as daily HI values ​​collected over the past 6 months, forming a time series containing 2000 data points) of similar components of the same equipment (such as gearbox bearings of the same model) from the equipment's historical database (such as SQL Server or InfluxDB time series database). A sliding window averaging method (with a window length of 10 sampling points) is used to smooth the time series data, eliminating short-term noise interference (such as HI value fluctuations caused by instantaneous vibration and impact). Secondly, the smoothed data is normalized (mapped to the [0,1] interval) to eliminate dimensional differences between different equipment or operating conditions. The preprocessed historical health indicator data is then input into a Gaussian Mixture Model (GMM) for probability distribution clustering: GMM assumes that the data is generated by a mixture of multiple Gaussian distributions, and iteratively optimizes the mean (μ), covariance (Σ), and mixing coefficient (π) of each Gaussian distribution using the Expectation-Maximization (EM) algorithm, ultimately generating M clusters (M=3 in this embodiment, corresponding to the initial, middle, and final degradation stages, respectively). The center of each cluster (μ1, μ2, μ3) represents the typical health indicator characteristics of that degradation stage. For example, in the initial stage, μ1=0.9, it means that the equipment is in a healthy state and the HI value is close to 1; in the middle stage, μ2=0.6, it means that the equipment begins to show slight degradation; and in the final stage, μ3=0.3, it means that the equipment is close to a failure state.

[0107] Extract health indicator data from the previous N sampling periods (e.g., N=20, i.e., the past 20 minutes) to form the current health indicator trajectory vector (e.g., [HI(t-19), HI(t-18), ..., HI(t)]). Calculate the Mahalanobis distance between the current trajectory vector and each cluster center of the GMM. The Mahalanobis distance considers the covariance structure of the data and can more accurately measure the similarity between the vector and the cluster centers. Specifically, the current trajectory vector is first standardized based on the covariance matrices Σ1, Σ2, Σ3 (corresponding to the initial, middle, and final clusters, respectively) obtained from GMM training (e.g., Z=(X-μ) / √Σ). Then, the Euclidean distance between the standardized vector and each cluster center is calculated (e.g., D1=√((Z-μ1)ᵀΣ1⁻). 1 (Z-μ1)).

[0108] Compare the Mahalanobis distances (D1, D2, D3) between the current trajectory vector and the three cluster centers, and select the cluster with the smallest distance as the current degradation stage (e.g., if D2 is the smallest, then the device is determined to be in the intermediate degradation stage). Generate the stage identification result (e.g., "Intermediate degradation, distance from intermediate cluster center 0.2, distance from initial cluster center 0.5, distance from final cluster center 0.8").

[0109] If it is determined to be in the initial degradation stage (e.g., D1 minimum), then the standard material parameter library is activated (e.g., the elastic modulus of gear steel E=210GPa, Poisson's ratio ν=0.3, density ρ=7850kg / m³). 3 This parameter library is based on equipment design specifications and material manuals, and reflects the physical characteristics of the equipment in a healthy state; If it is determined to be in the intermediate degradation stage (such as D2 minimum), then a set of minor damage parameters is applied (such as elastic modulus decreases by 5% to E=200GPa, surface roughness increases to Ra=1.6μm). This set of parameters is obtained through disassembly inspection and material testing of historical intermediate degradation equipment, reflecting the performance changes of the equipment in the minor degradation state. If it is determined to be in the final stage of degradation (such as D3 minimum), then the historical damage case parameter set is loaded (such as parameters extracted from the final failure cases of the same type of equipment, including elastic modulus reduction of 20% to E=168GPa, surface crack depth of 0.5mm, and contact fatigue damage coefficient increased to 1.5). This parameter set is generated by fault tree analysis (FTA) and finite element inversion calculation, reflecting the nonlinear characteristics of the equipment under severe degradation or failure conditions.

[0110] After the parameters are loaded, the edge computing node will transfer the parameters to the local finite element simulation software (such as the API interface of ANSYS APDL or COMSOL Multiphysics) to start the equipment structure simulation for the current degradation stage (such as contact stress analysis of gearbox bearings and thermo-mechanical coupling analysis of motor stator), generate simulation results (such as maximum stress value and temperature distribution cloud map), and provide maintenance personnel with a visual assessment of the degree of equipment degradation and maintenance suggestions (such as "mid-term degradation, it is recommended to replace the lubricating oil within 3 months; end-stage degradation, it is recommended to stop the machine for maintenance immediately").

[0111] Collect the latest monitoring data (such as time-series curves of newly collected health indicators and equipment maintenance records from the past 3 months) and divide them into training and validation sets. Use incremental learning algorithms (such as the online EM algorithm) to update the GMM parameters (μ, Σ, π) so that the cluster centers can reflect the latest degradation characteristics of the equipment (e.g., if the HI value range of the mid-term degradation stage is found to expand from 0.5-0.7 to 0.4-0.8 under new operating conditions, adjust μ2 and Σ2 of the mid-term cluster centers). Based on the results of equipment disassembly inspection and fault analysis (such as the material parameter change patterns in newly discovered end-stage degradation cases), correct the finite element simulation parameter set (e.g., add new damage mode parameters, such as fretting wear coefficient and corrosion damage factor) to improve the accuracy of the model's long-term predictions.

[0112] In some embodiments, step S105 above, which involves introducing a reinforcement learning model based on finite element simulation parameters to optimize the simulation parameters with the goal of minimizing maintenance costs and outputting predictive maintenance work orders, specifically includes: Load the finite element simulation parameters corresponding to the current degradation stage and construct a digital twin simulation environment that includes the physical constraints of the equipment; In a digital twin simulation environment, a dual-delay deep deterministic policy gradient algorithm is configured, using real-time health indicators and working condition data as the state space and adjustable simulation parameters as the action space to form a reinforcement learning model. With maintenance cost, remaining equipment lifespan, and power outage losses as joint optimization objectives, a reinforcement learning model iteratively executes a parameter adjustment-simulation-evaluation loop. When the cumulative reward value reaches the convergence threshold, the optimal combination of simulation parameters is extracted to generate a predictive maintenance work order that includes maintenance timing, method, and resources.

[0113] In this embodiment, based on the current degradation stage (early, middle, or late) of the equipment determined by Gaussian mixture model clustering in the previous steps, a parameter set corresponding to the stage is loaded from a preset finite element simulation parameter library. For example, if it is determined to be in the middle degradation stage, a parameter set containing an elastic modulus reduction of 10% (E=189GPa), surface roughness Ra=1.2μm, and contact fatigue damage coefficient of 1.2 is loaded. This parameter set is generated based on the disassembly test and material performance analysis of historical middle-degraded equipment.

[0114] Based on the loaded parameter set, a digital twin model of the equipment is constructed in finite element simulation software (such as ANSYS Mechanical or ABAQUS). First, the three-dimensional geometric model of the equipment (such as gears, bearings, and housing structure of a gearbox) is imported and a high-precision mesh is generated (e.g., a 0.1 mm hexahedral mesh is used for the gear contact area). Second, material properties (such as elastic modulus, Poisson's ratio, and density), boundary conditions (such as bearing fixing constraints and gear meshing contact), and load conditions (such as rated torque of 1000 N·m and rotational speed of 1500 r / min) are defined. Finally, the simulation analysis type (such as static analysis and transient dynamic analysis) and solver parameters (such as 1000 iteration steps and convergence tolerance of 1e-6) are set to form a digital twin simulation environment that includes the physical constraints of the equipment (such as stress not exceeding yield strength and deformation less than allowable clearance).

[0115] The Dual Delayed Deep Deterministic Policy Gradient (TD3) algorithm was chosen as the reinforcement learning framework. This algorithm effectively solves the overestimation bias problem in the traditional Deep Deterministic Policy Gradient (DDPG) algorithm by introducing a dual critic network and a delayed policy update mechanism, thereby improving the stability of parameter optimization.

[0116] Define the state space of the reinforcement learning model. Health indicators (HI) and operating condition data (such as load rate, speed, and ambient temperature) are collected from real-time equipment monitoring systems (e.g., SCADA systems, vibration monitors) to form a state vector. For example, the state vector may include dimensions such as the current HI value (0.65, corresponding to mid-term degradation), the HI change rate over the past 5 minutes (-0.02 / min), the current load rate of 80%, the speed of 1450 r / min, and the ambient temperature of 35℃, comprehensively reflecting the current operating status and degradation trend of the equipment.

[0117] Define the action space. Adjustable key parameters from the finite element simulation are used as action outputs. For example, the action vector can include an elastic modulus adjustment coefficient (range 0.9-1.1, allowing fluctuations of ±10% from the baseline value), a surface roughness adjustment coefficient (range 0.8-1.5, reflecting changes in wear level), and a contact fatigue damage coefficient adjustment coefficient (range 1.0-2.0, simulating cumulative fatigue effects). By adjusting these parameters, the changes in the physical properties of the equipment under different degrees of degradation can be simulated.

[0118] Define the reward function. Design a multi-objective weighted reward function with maintenance cost, remaining equipment lifespan, and power outage loss as joint optimization objectives. For example, the reward function can be expressed as: R = -α·C_maintenance - β·(1 / RUL) -γ·C_outage, where C_maintenance is the maintenance cost (e.g., bearing replacement cost of 5000 yuan, labor cost of 2000 yuan), RUL is the remaining equipment lifespan (failure time predicted through simulation, in hours), C_outage is the power outage loss (e.g., a single power outage causing production line downtime loss of 10000 yuan / hour), and α, β, and γ are weighting coefficients (e.g., α=0.4, β=0.3, γ=0.3, determined through historical data regression analysis). A larger reward value indicates a better parameter adjustment effect.

[0119] Initial parameters (such as weight matrix and bias vector) for the Actor Network and Critics Network are randomly generated, and the Replay Buffer is set to a capacity of 10,000 records to store historical state-action-reward-next state samples. In each training step, actions (such as elastic modulus adjustment coefficient 1.02 and surface roughness adjustment coefficient 1.1) are first generated through the Actor Network based on the current state (e.g., HI=0.65, load rate 80%). The actions are then mapped to finite element simulation parameters and input into the digital twin model to perform simulation analysis (e.g., calculating gear contact stress distribution). Secondly, based on simulation results (e.g., maximum contact stress 280MPa, not exceeding the material yield strength 500MPa) and historical equipment failure data (e.g., similar gears have an average remaining life of 500 hours at a contact stress of 300MPa), the remaining life of the equipment (RUL = 800 hours) is predicted. This is then combined with maintenance costs (e.g., a minor repair is recommended after 3 months under current conditions, costing 3000 yuan) and power outage losses (e.g., a 4-hour downtime for minor repairs, resulting in a loss of 40000 yuan) to calculate the reward value (R = -0.4×3000 -0.3×(1 / 800) -0.3×40000≈-12000 -0.000375 -12000≈-24000, with negative values ​​indicating a cost minimization objective). Finally, the current state, action, reward, and next state are combined to form a sample and stored in the experience replay buffer. Randomly sample small batches of samples (e.g., batch_size=64) from the experience replay buffer, calculate the value (Q-value) of the current action through the commentator network, and take the minimum value through the dual commentator network to reduce overestimation bias; update the action network parameters based on the Q-value (maximize the Q-value through gradient ascent), and simultaneously delay the update of the target action network and target commentator network parameters (synchronize once every 100 steps) to improve training stability.

[0120] Repeat the above iterative process until the cumulative reward value fluctuates less than a preset threshold (e.g., 5%) within 100 consecutive training steps. The model is then considered to have converged. At this point, the optimal combination of simulation parameters output by the action network is extracted (e.g., elastic modulus adjustment coefficient 1.05, surface roughness adjustment coefficient 1.2, contact fatigue damage coefficient adjustment coefficient 1.5). This parameter combination can optimize the reward function, i.e. minimize maintenance costs, maximize remaining lifespan, and minimize power outage losses.

[0121] The optimal combination of simulation parameters output from the reinforcement learning model is input into the digital twin model to perform final simulation analysis (e.g., calculating the stress-life curve of the equipment under optimal parameters). This is combined with historical equipment operating data (e.g., the HI trend over the past 6 months) and maintenance records (e.g., last maintenance time, replaced parts model) to predict the equipment's future degradation trajectory (e.g., the HI value will decrease to 0.4 in 3 months, entering the final degradation stage) and remaining lifespan (RUL = 600 hours). Predictive maintenance work orders are generated based on the prediction results, including maintenance timing, method, and resources. These predictive maintenance work orders are pushed to the equipment operation and maintenance management system (e.g., EAM system), notifying maintenance personnel via SMS, email, or WeChat, and simultaneously updating the equipment's digital profile (e.g., recording the predicted remaining lifespan and optimal parameter combination) to provide data support for subsequent maintenance strategy optimization.

[0122] Collect the latest equipment monitoring data (such as HI time-series curves, maintenance records, and failure cases from the past 3 months) and simulation results (such as stress distribution and remaining life prediction under new operating conditions) to construct an incremental training dataset; use transfer learning technology to quickly adapt the pre-trained model (trained based on historical data) to new operating conditions, reducing the computational cost of retraining; compare the reward values ​​of the model before and after the update through A / B testing to verify the update effect and ensure that the model can always output the optimal combination of simulation parameters.

[0123] In some embodiments, step S106 above, which involves performing on-site operations based on predictive maintenance work orders and collecting post-maintenance status data via AR devices, specifically includes: Based on predictive maintenance work orders, key operation items are extracted using natural language processing technology to generate AR-visualized instruction sequences; Based on AR visualization command sequences, the 3D animation of maintenance steps is overlaid onto the corresponding position of the real equipment through AR devices, and the operator's actions are monitored in real time. When deviation from the steps is detected, a correction prompt is triggered. After completing each maintenance operation, activate the AR device's infrared camera, visible light camera, and near-field sensor to simultaneously collect data on the device's surface temperature distribution, appearance, and electromagnetic field strength. Based on the equipment surface temperature distribution, appearance condition and electromagnetic field strength data, the digital twin model is retrieved to predict the status, and the difference degree of key parameters is calculated. When the difference degree of key parameters exceeds the preset difference threshold, abnormal maintenance items are marked. For abnormal maintenance items, supplementary maintenance instructions are automatically generated and fed back to the digital twin system to reassess the health status of the equipment.

[0124] In this embodiment, predictive maintenance work orders are obtained from an equipment operation and maintenance management system (such as an EAM system), and these work orders are stored in structured text format. A natural language processing algorithm based on a BERT pre-trained model is used to segment the work order text, perform part-of-speech tagging, and entity recognition, extracting key operation items, operation parameters, and operation sequences. A unique identifier is generated for each operation item and associated with a corresponding 3D model, operation tool, and safety prompt; simultaneously, the temporal relationship of the instruction sequence is constructed according to the operation sequence to form a complete AR visualization instruction set.

[0125] Maintenance personnel wearing AR glasses (such as Microsoft HoloLens 2 or Magic Leap 2) enter the maintenance site. The AR device scans the real equipment using SLAM (Simultaneous Localization and Mapping) technology, identifies the 3D position and orientation of key components of the equipment, and precisely aligns the pre-stored 3D model with the real equipment.

[0126] Based on the AR visualization command sequence, a 3D animation of maintenance steps is dynamically rendered at the corresponding location on the real device. Simultaneously, operating parameters are displayed at the edge of the maintenance personnel's field of vision. The AR device's RGB camera and depth sensor collect the maintenance personnel's movements in real time and compare them with a preset standard movement model; when a deviation is detected, the AR device immediately triggers a correction prompt.

[0127] After each maintenance operation is completed, the AR device automatically activates its integrated sensor array to collect device status data: an infrared camera scans the surface temperature distribution of the device to generate a thermal image; a visible light camera captures the device's appearance to generate a high-resolution image; and a near-field sensor measures the device's electromagnetic field strength to generate an electromagnetic field distribution curve. The AR device associates and stores the collected multimodal data (thermal image, appearance image, electromagnetic field curve) with the operation item identifier and uploads it in real time to the edge computing node via a 5G / Wi-Fi 6 network. The edge node performs preliminary data processing to reduce the subsequent transmission and computational burden.

[0128] The processed state data is obtained from the edge computing node, and the predicted state data after the corresponding maintenance operation is retrieved from the equipment digital twin model: the digital twin model predicts the surface temperature distribution, appearance and electromagnetic field strength of the equipment after maintenance through finite element simulation based on the operation parameters in the maintenance work order and the historical operation data of the equipment.

[0129] Calculate the difference in key parameters between measured and predicted state data. For temperature data, calculate the root mean square error (RMSE) between the measured and predicted temperature fields; for appearance data, compare the measured and predicted images using an image similarity algorithm (such as the structural similarity index SSIM); for electromagnetic field data, calculate the correlation coefficient between the measured and predicted curves. When the difference in key parameters exceeds a preset difference threshold (e.g., temperature RMSE > 3℃, appearance SSIM < 0.9, electromagnetic field correlation coefficient < 0.95), the corresponding maintenance item is marked as abnormal.

[0130] For maintenance items marked as abnormal, the system automatically generates supplementary maintenance instructions. Based on the analysis of the cause of the abnormality, supplementary instructions are issued; simultaneously, these instructions are pushed to the maintenance personnel's AR device, guiding the corrective operation in the form of 3D animation. After the maintenance personnel complete the supplementary maintenance, the AR device collects equipment status data again and uploads it to the digital twin system; the digital twin model re-predicts the equipment status based on the corrected maintenance parameters and compares it with the measured data. If the difference drops to within the threshold, the abnormality is confirmed to have been eliminated; otherwise, supplementary instructions continue to be generated until the equipment status meets the standard. Finally, the digital twin system comprehensively evaluates the equipment health status based on the execution results of all maintenance items (including normal and abnormal corrections): if all key parameters meet the safety thresholds, the equipment maintenance is deemed qualified, and the equipment health record is updated (such as recording the maintenance time, operation items, and status data); if unresolved abnormalities still exist, a higher-level maintenance process is triggered (such as requesting remote expert support or arranging return to the factory for repair) to ensure the long-term stable operation of the equipment.

[0131] Reference Figure 2 An embodiment of the present invention provides a fault intelligent diagnosis system 2 that integrates state monitoring and multimodal large model, the system 2 specifically comprising: The first diagnostic module 201 is used to synchronously collect time-series data and spatial images through heterogeneous sensor groups and monitoring equipment deployed on power grid equipment to form raw data; The second diagnostic module 202 is used to generate a physical constraint feature vector based on the original data, combined with the equipment thermodynamic equation and material deformation rules; The third diagnostic module 203 is used to perform health indicator prediction based on physical constraint feature vectors through a lightweight LSTM network. The fourth diagnostic module 204 is used to cluster the HI time series curve using a Gaussian mixture model when a continuous decline in health indicators is detected, determine the degradation stage based on the distance between cluster centers, obtain the stage identification result, and dynamically switch the finite element simulation parameters based on the stage identification result. The fifth diagnostic module 205 is used to optimize simulation parameters based on finite element simulation parameters by introducing a reinforcement learning model with the goal of minimizing maintenance costs, and output predictive maintenance work orders. The sixth diagnostic module 206 is used to perform on-site operations based on predictive maintenance work orders and collect post-maintenance status data through AR devices.

[0132] It is understandable that, such as Figure 1 The content of the embodiments of the fusion state monitoring and multimodal large model fault intelligent diagnosis method shown is applicable to the embodiments of the fusion state monitoring and multimodal large model fault intelligent diagnosis system. The specific functions implemented by the embodiments of the fusion state monitoring and multimodal large model fault intelligent diagnosis system are as follows: Figure 1 The embodiment of the fusion state monitoring and multimodal large model fault intelligent diagnosis method shown is the same, and the beneficial effects achieved are the same as those described above. Figure 1 The beneficial effects achieved by the embodiment of the fusion state monitoring and multimodal large model fault intelligent diagnosis method shown are also the same.

[0133] It should be noted that the information interaction and execution process between the above systems are based on the same concept as the method embodiments of the present invention. For details on their specific functions and technical effects, please refer to the method embodiments section, which will not be repeated here.

[0134] Those skilled in the art will clearly understand that, for the sake of convenience and brevity, the above-described division of functional units and modules is merely an example. In practical applications, the above functions can be assigned to different functional units and modules as needed, that is, the internal structure of the system can be divided into different functional units or modules to complete all or part of the functions described above. The functional units and modules in the embodiments can be integrated into one processing unit, or each unit can exist physically separately, or two or more units can be integrated into one unit. The integrated unit can be implemented in hardware or as a software functional unit. Furthermore, the specific names of the functional units and modules are only for easy differentiation and are not intended to limit the scope of protection of this application. The specific working process of the units and modules in the above system can be referred to the corresponding process in the foregoing method embodiments, and will not be repeated here.

[0135] Reference Figure 3 The present invention also provides a computer device 3, including: a memory 302 and a processor 301, and a computer program 303 stored in the memory 302. When the computer program 303 is executed on the processor 301, it implements the fusion state monitoring and multimodal large model fault intelligent diagnosis method as described in any of the above methods.

[0136] The computer device 3 may be a desktop computer, laptop, handheld computer, or cloud server, etc. The computer device 3 may include, but is not limited to, a processor 301 and a memory 302. Those skilled in the art will understand that... Figure 3 The computer device 3 is merely an example and does not constitute a limitation on the computer device 3. It may include more or fewer components than shown in the figure, or combine certain components, or different components, such as input / output devices, network access devices, etc.

[0137] The processor 301 may be a Central Processing Unit (CPU), or it may be other general-purpose processors, digital signal processors (DSPs), application-specific integrated circuits (ASICs), field-programmable gate arrays (FPGAs), or other programmable logic devices, discrete gate or transistor logic devices, discrete hardware components, etc. A general-purpose processor may be a microprocessor or any conventional processor.

[0138] In some embodiments, the memory 302 may be an internal storage unit of the computer device 3, such as a hard disk or memory of the computer device 3. In other embodiments, the memory 302 may be an external storage device of the computer device 3, such as a plug-in hard disk, smart media card (SMC), secure digital (SD) card, flash card, etc., equipped on the computer device 3. Furthermore, the memory 302 may include both internal and external storage units of the computer device 3. The memory 302 is used to store the operating system, applications, boot loader, data, and other programs, such as the program code of the computer program. The memory 302 can also be used to temporarily store data that has been output or will be output.

[0139] This invention also provides a computer-readable storage medium storing a computer program thereon. When the computer program is run by a processor, it implements the fusion state monitoring and multimodal large model fault intelligent diagnosis method as described in any of the above methods.

[0140] In this embodiment, if the integrated unit is implemented as a software functional unit and sold or used as an independent product, it can be stored in a computer-readable storage medium. Based on this understanding, all or part of the processes in the methods of the above embodiments can be implemented by a computer program instructing related hardware. The computer program can be stored in a computer-readable storage medium, and when executed by a processor, it can implement the steps of the various method embodiments described above. The computer program includes computer program code, which can be in the form of source code, object code, executable files, or certain intermediate forms. The computer-readable medium can include at least: any entity or device capable of carrying computer program code to a photographing device / terminal device, a recording medium, a computer memory, a read-only memory (ROM), a random access memory (RAM), an electrical carrier signal, a telecommunication signal, and a software distribution medium. Examples include USB flash drives, portable hard drives, magnetic disks, or optical disks. In some jurisdictions, according to legislation and patent practice, computer-readable media cannot be electrical carrier signals or telecommunication signals.

[0141] In the above embodiments, the descriptions of each embodiment have different focuses. For parts that are not described in detail or recorded in a certain embodiment, please refer to the relevant descriptions of other embodiments.

[0142] Those skilled in the art will recognize that the units and algorithm steps of the various examples described in conjunction with the embodiments disclosed herein can be implemented in electronic hardware, or a combination of computer software and electronic hardware. Whether these functions are implemented in hardware or software depends on the specific application and design constraints of the technical solution. Those skilled in the art can use different methods to implement the described functions for each specific application, but such implementation should not be considered beyond the scope of this application.

[0143] In the embodiments disclosed in this application, it should be understood that the disclosed devices / terminal equipment and methods can be implemented in other ways. For example, the device / terminal equipment embodiments described above are merely illustrative. For instance, the division of modules or units is only a logical functional division, and in actual implementation, there may be other division methods. For example, multiple units or components may be combined or integrated into another system, or some features may be ignored or not executed. Furthermore, the displayed or discussed mutual coupling or direct coupling or communication connection may be through some interfaces; the indirect coupling or communication connection between devices or units may be electrical, mechanical, or other forms.

[0144] The units described as separate components may or may not be physically separate. The components shown as units may or may not be physical units; that is, they may be located in one place or distributed across multiple network units. Some or all of the units can be selected to achieve the purpose of this embodiment according to actual needs.

Claims

1. A fault intelligent diagnosis method integrating condition monitoring and multimodal large model, characterized in that, The method specifically includes: By deploying heterogeneous sensor groups and monitoring equipment on power grid equipment, time-series data and spatial images are collected synchronously to form raw data; Based on the original data, and combining the equipment's thermodynamic equations with the material deformation rules, a physical constraint feature vector is generated. Based on physically constrained feature vectors, a lightweight LSTM network is used to perform health indicator prediction. When a continuous decline in health indicators is detected, the HI time series curve is clustered using a Gaussian mixture model. The degradation stage is determined based on the distance between cluster centers to obtain the stage identification results. The finite element simulation parameters are then dynamically switched based on the stage identification results. Based on finite element simulation parameters, a reinforcement learning model is introduced to optimize the simulation parameters with the goal of minimizing maintenance costs, and predictive maintenance work orders are output. Perform on-site operations based on predictive maintenance work orders and collect post-maintenance status data using AR devices.

2. The method according to claim 1, characterized in that, The process of generating a physical constraint feature vector based on the original data, combined with the equipment's thermodynamic equations and material deformation rules, specifically includes: The raw data was separated into temperature sequences, vibration spectra, and infrared thermal images. Based on the equipment's heat dissipation structure topology diagram, spatial region segmentation is performed on the infrared thermal image, and the temperature change rate of each heat sink unit is calculated by combining the temperature sequence. Based on the theory of winding mechanical deformation, the axial and radial resonant frequency offsets are analyzed from the vibration spectrum. By associating the temperature change rate and resonant frequency offset of each heat sink unit, a combined mechanical-thermodynamic feature identifier is output, forming a physical constraint feature vector.

3. The method according to claim 2, characterized in that, The step of performing spatial region segmentation on the infrared thermal image based on the device's heat dissipation structure topology diagram and calculating the temperature change rate of each heat sink unit specifically includes: Load the heat dissipation structure topology diagram of the power grid equipment, and combine it with the physical coordinate data of the heat sink to generate a standardized heat dissipation unit spatial grid; For the real-time acquired infrared thermal image, based on the spatial grid of the heat dissipation unit, the key points of the heat sink outline are located by the corner detection algorithm. The actual image and the topological map are then registered by affine transformation to obtain the feature registration result. Based on the feature registration results, the infrared thermal image is segmented into independent regions corresponding to the heat dissipation unit, and the pixel temperature matrix of each independent region is extracted. For each independent region's pixel temperature matrix, the temperature matrix of adjacent frames is compared according to the time series to calculate the rate of temperature change per unit time.

4. The method according to claim 2, characterized in that, The method of analyzing the axial and radial resonant frequency shifts from the vibration spectrum based on the winding mechanical deformation theory specifically includes: Based on the transformer winding design drawings, establish the reference resonant frequency values ​​for the axial and radial directions, and simultaneously associate them with the no-load operating conditions of the equipment. For the raw time-domain signal acquired by the triaxial vibration sensor, bandpass filtering is performed to retain the target frequency band, and windowing is applied to suppress spectral leakage to obtain the preprocessed time-domain signal; Based on the preprocessed time-domain signal, a power spectral density curve is generated through fast spectrum transformation to accurately locate the main resonance peak frequency. By comparing the main resonance peak frequency and the reference resonance frequency value, the percentage offsets in the axial and radial directions are calculated respectively to obtain the axial and radial resonance frequency offsets.

5. The method according to claim 2, characterized in that, The temperature change rate and resonant frequency offset of each associated heat sink unit are used to output a combined mechanical-thermodynamic feature identifier, forming a physical constraint feature vector, specifically including: The temperature change rate of each heat sink unit is bound to the resonant frequency offset in the axial and radial directions of the same region, respectively, to form thermo-mechanical axial coupling parameters and thermo-mechanical radial coupling parameters. Based on the thermo-mechanical axial coupling parameters and thermo-mechanical radial coupling parameters, the thermal expansion coefficient of the equipment material is introduced to calculate the theoretical deformation threshold range; When the actual coupling parameters exceed the theoretical deformation threshold range, they are marked as abnormal associated regions, and a mechanical-thermodynamic joint feature identifier is generated. The mechanical-thermodynamic joint feature identifier includes a list of normal coupling region numbers and the location and over-limit level of the abnormal coupling region.

6. The method according to claim 1, characterized in that, The method of predicting health indicators using a lightweight LSTM network based on physically constrained feature vectors specifically includes: Based on the physical constraint feature vector, the normalization module transforms each feature value to a preset numerical range to obtain standardized features. Standardized features are input into a pre-trained lightweight LSTM network to perform rolling time window prediction and output health indicator values ​​and their confidence levels within a preset future time period. When the confidence level is lower than the preset confidence threshold, the entire LSTM model is re-predicted, and a health risk level label is generated. Based on the health risk level identifier, a preset maintenance strategy library is matched, and when the risk level is determined to be high, a maintenance work order is pushed to the operation and maintenance terminal in real time.

7. The method according to claim 1, characterized in that, When a continuous decline in health indicators is detected, a Gaussian mixture model is used to cluster the HI time series curve. The degradation stage is determined based on the cluster center distance to obtain the stage identification result. The finite element simulation parameters are then dynamically switched based on the stage identification result. Specifically, this includes: Real-time monitoring of health indicator time-series curves; when a monotonically decreasing indicator value is detected for K consecutive sampling periods, a degradation stage analysis command is triggered. In response to the degradation stage analysis command, historical health index data of the equipment is retrieved, and a probability distribution cluster is generated through a Gaussian mixture model. The cluster centers of the probability distribution cluster correspond to the initial, middle or final degradation stage. Calculate the Mahalanobis distance between the current health indicator trajectory and each cluster center, determine the current degradation stage based on the minimum distance principle, and obtain the stage identification result; Based on the stage identification results, the corresponding finite element simulation parameters are automatically loaded. If it is determined to be the initial degradation stage, the standard material parameter library is enabled. If it is determined to be the final degradation stage, the historical damage case parameter set is loaded.

8. The method according to claim 7, characterized in that, The method, based on finite element simulation parameters, introduces a reinforcement learning model to optimize the simulation parameters with the goal of minimizing maintenance costs, and outputs predictive maintenance work orders, specifically including: Load the finite element simulation parameters corresponding to the current degradation stage and construct a digital twin simulation environment that includes the physical constraints of the equipment; In a digital twin simulation environment, a dual-delay deep deterministic policy gradient algorithm is configured, using real-time health indicators and working condition data as the state space and adjustable simulation parameters as the action space to form a reinforcement learning model. With maintenance cost, remaining equipment lifespan, and power outage losses as joint optimization objectives, a reinforcement learning model iteratively executes a parameter adjustment-simulation-evaluation loop. When the cumulative reward value reaches the convergence threshold, the optimal combination of simulation parameters is extracted to generate a predictive maintenance work order that includes maintenance timing, method, and resources.

9. The method according to claim 1, characterized in that, The process of performing on-site operations based on predictive maintenance work orders and collecting post-maintenance status data using AR devices specifically includes: Based on predictive maintenance work orders, key operation items are extracted using natural language processing technology to generate AR-visualized instruction sequences; Based on AR visualization command sequences, the 3D animation of maintenance steps is overlaid onto the corresponding position of the real equipment through AR devices, and the operator's actions are monitored in real time. When deviation from the steps is detected, a correction prompt is triggered. After completing each maintenance operation, activate the AR device's infrared camera, visible light camera, and near-field sensor to simultaneously collect data on the device's surface temperature distribution, appearance, and electromagnetic field strength. Based on the equipment surface temperature distribution, appearance condition and electromagnetic field strength data, the digital twin model is retrieved to predict the status, and the difference degree of key parameters is calculated. When the difference degree of key parameters exceeds the preset difference threshold, abnormal maintenance items are marked. For abnormal maintenance items, supplementary maintenance instructions are automatically generated and fed back to the digital twin system to reassess the health status of the equipment.

10. A fault intelligent diagnosis system integrating condition monitoring and multimodal large model, characterized in that, The system specifically includes: The first diagnostic module is used to synchronously collect time-series data and spatial images through heterogeneous sensor groups and monitoring equipment deployed on power grid equipment to form raw data. The second diagnostic module is used to generate physical constraint feature vectors based on the original data, combined with the equipment's thermodynamic equations and material deformation rules; The third diagnostic module is used to perform health indicator prediction based on physical constraint feature vectors using a lightweight LSTM network. The fourth diagnostic module is used to cluster the HI time series curve using a Gaussian mixture model when a continuous decline in health indicators is detected. It determines the degradation stage based on the distance between cluster centers, obtains the stage identification result, and dynamically switches the finite element simulation parameters based on the stage identification result. The fifth diagnostic module is used to optimize simulation parameters based on finite element simulation parameters by introducing a reinforcement learning model with the goal of minimizing maintenance costs, and output predictive maintenance work orders. The sixth diagnostic module is used to perform on-site operations based on predictive maintenance work orders and collect post-maintenance status data through AR devices.

Images