Power equipment implicit state predictive maintenance method

Abstract

The invention discloses a predictive maintenance method for an implicit state of power equipment, and belongs to the technical field of intelligent operation and maintenance of the power equipment. According to the method, multi-mode information such as SCADA data, acoustic vibration data, infrared thermal image data and partial discharge data is collected, space-time alignment and attention mechanism fusion are carried out, and a hidden state code representing the internal health state of equipment is extracted; a dynamic state deduction model combining physical constraint and data driving is constructed, and prediction of the future state evolution trajectory of the equipment is achieved; and performing a virtual maintenance experiment in the digital twin based on a prediction result, and generating an optimal maintenance strategy through multi-objective optimization. According to the method, the problems that the hidden state of the equipment cannot be sensed and the prediction capability is lacked in the prior art are solved, the conversion from passive maintenance to predictive maintenance is realized, and the accuracy and foresight of operation and maintenance of the power equipment are improved.

Description

Technical Field

[0001] This invention relates to the field of intelligent operation and maintenance technology for power equipment, and in particular to a method for predictive maintenance of the latent condition of power equipment. Background Technology

[0002] As a critical infrastructure supporting rail transit operations, the reliability of railway power systems directly affects the safety and efficiency of the entire transportation network. Currently, operation and maintenance management in this field mainly relies on the combined application of three types of technology systems.

[0003] Supervisory Control and Data Acquisition (SCADA) systems form the core of operation and maintenance management. Through sensor networks deployed on various power equipment, they collect real-time operating parameters such as current, voltage, temperature, and switch status, displaying them centrally on a monitoring interface and triggering alarms for any anomalies. Simultaneously, equipment history database systems have completely replaced traditional paper records, comprehensively documenting basic information, historical operation and maintenance records, fault data, and maintenance reports from equipment commissioning to the present, providing crucial references for analyzing long-term equipment operating characteristics. In recent years, with the popularization of 3D visualization technology, solutions have gradually been adopted that establish 3D geometric models of power equipment and display them synchronously with the SCADA monitoring interface and history information in embedded windows or split-screen formats, aiming to enhance the intuitiveness of the operation and maintenance process.

[0004] However, in-depth analysis reveals the following limitations in the existing technological system that urgently need to be overcome:

[0005] First, the current SCADA system suffers from limited state perception and insufficient ability to detect latent defects. The operating parameters monitored by existing SCADA systems only reflect the surface-level, outcome-based operating status of equipment, failing to effectively detect critical, latent changes within the equipment, such as the aging of insulation materials, fatigue accumulation in mechanical structures, and contact wear—the intrinsic physicochemical processes that determine equipment lifespan. This lack of perception makes it difficult for the system to provide sufficiently early warning signals for predictive maintenance.

[0006] Secondly, the level of data integration is low, and information silos are significant. There is a lack of deep logical connections and intelligent integration mechanisms between SCADA real-time data, historical data, and 3D models. Data is fragmented between systems, requiring maintenance personnel to frequently switch between different interfaces and perform manual comparisons. This makes it difficult to quickly establish the correlation between data anomalies and specific equipment components or historical faults, seriously affecting the efficiency and accuracy of fault analysis.

[0007] Third, 3D models are static and rigid, lacking dynamic extrapolation capabilities. Most existing 3D models are static geometric displays, unable to dynamically update their visual state based on real-time data (such as the precise location of switches or internal heat distribution), and even less capable of predicting future trends based on the current state. This limits the value of these models primarily to basic training and initial demonstrations, failing to support deeper situational analysis and decision optimization.

[0008] In summary, existing technologies have significant shortcomings in terms of the depth of state perception, the intelligence of data fusion, and the dynamism of model application, which hinders the transformation and upgrading of operation and maintenance models from reactive maintenance to predictive maintenance. Therefore, there is an urgent need in this field for a new type of intelligent operation and maintenance solution that can deeply integrate multi-source information, accurately perceive the implicit state of equipment, and possess dynamic extrapolation capabilities. Summary of the Invention

[0009] The purpose of this invention is to provide a predictive maintenance method for the latent condition of power equipment, in order to solve the aforementioned problems existing in the prior art.

[0010] The technical solution, a predictive maintenance method for the latent condition of power equipment, includes the following steps:

[0011] Collect multimodal operating data of power equipment, including real-time operating data of SCADA system, acoustic vibration data, infrared thermal imaging data, and partial discharge data;

[0012] Spatiotemporal alignment and feature fusion are performed on multimodal operational data to extract implicit state codes that characterize the internal health status of the equipment;

[0013] A dynamic state inference model is constructed based on a physical constraint model and a deep learning network. The implicit state code is iteratively predicted in multiple steps to generate the evolution trajectory of the device's future state.

[0014] Based on the aforementioned evolutionary trajectory, a virtual maintenance experiment is conducted in a digital twin, and the optimal maintenance strategy is generated through multi-objective optimization.

[0015] According to a further improvement of the present invention, the acquisition of multimodal operation data of power equipment includes:

[0016] The three-phase current data, three-phase voltage data, equipment temperature data and switch status data in the SCADA system are collected through the data acquisition interface. The data is then cleaned and normalized to obtain a standardized set of electrical operating parameters.

[0017] Mechanical vibration waveform data of the equipment is collected by a vibration acceleration sensor, and fast Fourier transform and spectral feature extraction are performed to obtain a mechanical vibration feature vector containing characteristic frequency amplitude, harmonic components and spectral entropy.

[0018] The infrared image sequence of the equipment surface is acquired by an infrared thermal imager, and non-uniformity correction and radiometric calibration are performed to reconstruct the two-dimensional temperature field and calculate the temperature gradient distribution to obtain the temperature distribution state matrix.

[0019] Partial discharge pulse sequences were collected using a high-frequency current transformer, and pulse feature extraction and cluster analysis were performed to obtain an index of discharge severity.

[0020] According to a further improvement of the present invention, spatiotemporal alignment and feature fusion of multimodal operational data are performed, including:

[0021] Using the timestamp of the SCADA system as a reference, an interpolation method is used to synchronize the sampling times of acoustic vibration data, infrared thermal imaging data, and partial discharge data to obtain a time-aligned multimodal dataset.

[0022] Establish a spatial mapping relationship between the equipment's 3D model and infrared thermal image data, and achieve spatial registration through perspective projection transformation;

[0023] An attention mechanism fusion network is constructed, and the weight coefficients of each modality in the multimodal dataset are calculated using the device operating condition as the query vector.

[0024] Multimodal features are weighted and fused based on weight coefficients, and feature dimensionality is reduced by using a deep belief network to extract low-dimensional latent state codes.

[0025] According to a further improvement of the present invention, a dynamic state deduction model is constructed, including:

[0026] Based on the material properties and physical principles of the equipment, differential equations and inequality constraints for the evolution of state variables are established, and a physical knowledge base is constructed.

[0027] Design a physical information neural network architecture and embed the constraints in the physical knowledge base as regularization terms into the loss function;

[0028] A physical information neural network is trained using historical state encoding sequences to obtain a deductive model of physical constraints;

[0029] The Monte Carlo Dropout technique is used to sample the prediction process multiple times and calculate the confidence interval of the prediction results.

[0030] According to a further improvement of the present invention, the evolution trajectory of the future state of the device is generated, including:

[0031] The hidden state code at the current moment is input into the trained physical information neural network to predict the state code at the next moment.

[0032] The predicted state code is used as new input, and multi-step prediction is performed iteratively to generate a state evolution sequence within a specified time range;

[0033] Based on the results of multiple Monte Carlo samplings, the mean and variance of the state prediction values ​​are calculated to form an evolution trajectory with confidence intervals.

[0034] The evolutionary trajectory is dynamically displayed in a 3D visualization interface using color gradients and transparency changes.

[0035] According to a further improvement of the present invention, virtual maintenance experiments are conducted in a digital twin, including:

[0036] Different load conditions, environmental factors, and maintenance operations are set in the digital twin model of the equipment to generate a set of virtual experimental scenarios;

[0037] For each virtual scenario, the dynamic state inference model is driven to simulate long-term state evolution, using the current device state as the initial condition.

[0038] Record the evolution curves of key performance indicators in each virtual scenario, including health index, failure probability, and maintenance cost;

[0039] A database of maintenance strategy effectiveness is established based on simulation results to provide data support for multi-objective optimization.

[0040] According to a further improvement of the present invention, an optimal maintenance strategy is generated through multi-objective optimization, including:

[0041] Establish a multi-objective optimization function with the goals of minimizing maintenance costs, minimizing equipment risks, and maximizing operational efficiency;

[0042] Constraints include resource availability, time window, and technical feasibility.

[0043] A non-dominated sorting genetic algorithm with an elitist strategy is used to search the solution space to obtain the Pareto optimal solution set;

[0044] The final maintenance scheme is selected from the Pareto optimal solution set based on the decision-maker's preferences, and a structured maintenance work order is generated.

[0045] According to a further improvement of the present invention, the method further includes an augmented reality-assisted execution step:

[0046] The key information of the optimal maintenance strategy is converted into augmented reality visualization elements;

[0047] The visual elements are overlaid onto the real device scene using AR devices to provide operation guidance and parameter prompts;

[0048] Based on gesture recognition and voice interaction, feedback from on-site personnel is received, and maintenance plans are dynamically adjusted.

[0049] Record data from the maintenance process to update equipment history and optimize simulation models.

[0050] According to a further improvement of the present invention, the specific implementation of the attention mechanism fusion network includes:

[0051] The feature vectors of each modality are mapped to a unified latent space through independent embedding layers;

[0052] Using equipment load rate and ambient temperature as query vectors, calculate the similarity score with each modal key vector;

[0053] The similarity score is converted into attention weights using the Softmax function;

[0054] The attention weights are used to sum the modality value vectors in a weighted manner to obtain context-aware fusion features.

[0055] According to a further improvement of the present invention, the physical knowledge base includes the following constraint model:

[0056] Thermal aging model of insulating materials based on the Arrhenius equation; contact electrical wear model based on cumulative breaking current; mechanical fatigue accumulation model based on stress-cycle number; temperature field evolution model based on heat conduction theory; and physical boundary constraints for each state variable.

[0057] Beneficial effects: This invention enables precise perception of the latent state of equipment and accurate prediction of its future state, transforming maintenance decisions from reactive to proactive; by constructing a predictive maintenance closed loop driven by digital twins, it improves equipment reliability and operational economy. Attached Figure Description

[0058] Figure 1 This is a flowchart illustrating the overall steps of an embodiment of the present invention.

[0059] Figure 2 This is a flowchart of multi-source heterogeneous data acquisition and preprocessing according to an embodiment of the present invention.

[0060] Figure 3 This is a flowchart of multimodal data fusion and latent state extraction based on attention mechanism according to an embodiment of the present invention.

[0061] Figure 4 This is a flowchart of the process of constructing a dynamic state deduction model of physical constraints and performing trajectory prediction according to an embodiment of the present invention.

[0062] Figure 5 This is a flowchart of interactive maintenance decision-making and execution based on digital twins, according to an embodiment of the present invention. Detailed Implementation

[0063] To enable those skilled in the art to better understand the present invention, the technical solutions of the present invention will be clearly and completely described below with reference to the accompanying drawings of the embodiments of the present invention. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort should fall within the scope of protection of the present invention.

[0064] like Figures 1 to 5 As shown, this embodiment provides an overall framework for a predictive maintenance method for the latent condition of power equipment. Specifically, the method of this embodiment includes the following steps:

[0065] S1: Multi-source heterogeneous data acquisition and preprocessing

[0066] S1.1: Collect real-time operating data from the SCADA system, perform data cleaning and outlier detection, and obtain a standardized dataset of electrical operating parameters.

[0067] Specifically, this embodiment acquires the most basic equipment operating parameters from existing data acquisition and monitoring control systems. In actual operation, raw data such as three-phase current, three-phase voltage, equipment temperature, switch opening and closing status, and protection device signals from key monitoring points of the power equipment are periodically collected through a configured industrial communication protocol interface. This data is aggregated from the field monitoring unit to the data center at fixed time intervals (e.g., once per second).

[0068] Preferably, during the data cleaning stage, a multi-verification mechanism is employed to ensure data quality. First, a data integrity check is performed to identify and mark data loss segments caused by network transmission interruptions. Further, an outlier detection algorithm based on statistical principles is applied, such as combining box plots and the 3σ principle, to establish a normal fluctuation range for each monitored parameter. Values ​​significantly exceeding the physically possible range (such as current momentarily dropping to zero and then returning to normal) are identified as anomalous data. According to a further improvement in this embodiment, identified anomalous data is not simply discarded, but rather reasonably repaired using linear interpolation or spline interpolation methods based on the data trends before and after it, thereby maintaining the continuity of the data time series.

[0069] In one optional implementation, the cleaned data also needs to be normalized. This is because the dimensions and numerical ranges of different physical quantities vary greatly; for example, current values ​​may be in the thousands of amperes, while temperature values ​​are only tens of degrees Celsius. By using a min-max normalization method, each parameter is mapped to the [0,1] interval, eliminating the influence of dimensions and laying the foundation for subsequent multimodal data fusion. Ultimately, an electrical operating parameter dataset containing timestamps, device identifiers, and a series of standardized parameter values ​​is obtained.

[0070] S1.2: Collect acoustic vibration data of the equipment, perform spectral transformation and feature extraction to obtain mechanical vibration feature vectors.

[0071] In this embodiment, unlike the monitoring scope of traditional SCADA systems, a deeper perception of the mechanical state of the equipment is introduced. Specifically, high-precision vibration acceleration sensors and acoustic sensors are installed in key parts of the power equipment (such as the transformer tank wall and the circuit breaker operating mechanism housing) to collect structural vibration and noise signals generated during equipment operation at a sampling frequency higher than 20kHz, thereby obtaining raw time-domain waveform data.

[0072] For example, the acquired raw waveform data is preprocessed, including DC component removal and windowing, to reduce spectral leakage. Further, a Fast Fourier Transform (FFT) is performed to convert the vibration signal in the time domain into a spectrum in the frequency domain. This transform reveals the energy distribution of the signal at different frequency components, which is crucial for identifying specific mechanical fault modes (such as bearing wear or component loosening). In a preferred embodiment, not only is the classical spectrum calculated, but envelope spectrum analysis is further performed to better extract the impact fault features hidden in the vibration signal.

[0073] According to a further improvement of this embodiment, a set of characteristic parameters that can comprehensively characterize the mechanical state are extracted from the spectrum. These parameters include, but are not limited to: amplitude at characteristic frequencies (such as rotational frequency, meshing frequency) to monitor the vibration intensity of specific components; spectral centroid, reflecting the concentrated area of ​​vibration energy; spectral kurtosis, which is very sensitive to impact signals; and the energy proportion of multiple frequency bands. Finally, these characteristic parameters are combined in a predetermined order to form a high-dimensional mechanical vibration characteristic vector, which serves as a digital representation of the mechanical health status of the equipment.

[0074] S1.3: Collect infrared thermal image data from the equipment, reconstruct the temperature field and perform gradient analysis to obtain the temperature distribution state matrix.

[0075] In one alternative implementation, to achieve comprehensive visual monitoring of the equipment's thermal condition, an infrared thermal imager is used to periodically or continuously scan the power equipment. The infrared thermal imager outputs a sequence of grayscale images representing the intensity of infrared radiation on the object's surface. First, these raw infrared images need to undergo non-uniformity correction to eliminate fixed pattern noise caused by differences in the detector's own pixel response.

[0076] Specifically, the next key step is radiometric calibration. By pointing the thermal imager at a blackbody source of known temperature, a mapping model between image grayscale values ​​and the true absolute temperature of the object is established. Applying this model, the grayscale value of each pixel in each frame of the infrared image is converted into a precise temperature value, thereby obtaining the temperature distribution map of the device. According to a further improvement in this embodiment, to obtain a more accurate temperature field, bilinear interpolation or cubic convolution interpolation algorithms are used to upsample the original thermal imager resolution, reconstructing a smoother and more detailed two-dimensional temperature field.

[0077] Preferably, to further quantify the characteristics of heat distribution, gradient calculations are performed on the temperature field. Using gradient operators such as Sobel or Prewitt, the spatial rate of change of the temperature field in the X and Y directions is calculated, thus obtaining a temperature gradient distribution map. High gradient regions typically correspond to thermal defects or connection failure points. Finally, the calibrated and reconstructed absolute temperature field is combined with the calculated temperature gradient field to form a temperature distribution state matrix. This matrix not only contains the temperature of each "pixel" but also implies its spatial variation trend, providing a rich data foundation for thermal state assessment.

[0078] S1.4: Collect partial discharge data from the equipment, perform pattern recognition and pulse clustering to obtain discharge characteristics and severity indicators.

[0079] In a further improvement to this embodiment, a partial discharge monitoring system consisting of a high-frequency current transformer and a high-bandwidth oscilloscope is deployed to detect early defects inside the equipment, especially in the insulation system. This system captures transient current pulse waveforms, ranging from nanoseconds to microseconds, flowing through the equipment's grounding wire; these are the partial discharge pulse sequences.

[0080] In the data processing stage, the captured pulse waveforms are first denoised using wavelet thresholding to suppress environmental electromagnetic interference. Further, for each valid discharge pulse, three key features are extracted: discharge quantity (estimated by the pulse's integral area), discharge phase (the phase angle of the corresponding power grid voltage at the time of pulse occurrence), and pulse waveform characteristics (such as rise time and pulse width). In a preferred embodiment, tens of thousands of discharge pulses within power frequency cycles are collected to construct a discharge spectrum with sufficient statistical significance.

[0081] For example, unsupervised machine learning algorithms, such as K-means clustering or DBSCAN clustering, are then used to automatically classify all discharge pulses based on their discharge magnitude and phase characteristics. This process separates the discharge signals into different clusters, each potentially corresponding to a specific discharge type (e.g., internal discharge, surface discharge, corona discharge). Finally, the average discharge magnitude of each cluster center, the discharge frequency of that cluster, and its phase distribution characteristics are calculated. Based on these statistical results, combined with industry standards (such as IEC 60270), a comprehensive discharge severity index can be calculated, quantifying the threat level of current partial discharge activity to equipment insulation.

[0082] S2: Multimodal Data Fusion and Latent State Extraction Based on Attention Mechanism

[0083] S2.1: Construct a spatiotemporal alignment framework to perform time synchronization and spatial registration of multi-source data, resulting in a spatiotemporally aligned multimodal dataset.

[0084] Specifically, since the multimodal data collected in step S1 comes from different independent systems with different timestamp bases and spatial reference systems, direct fusion would lead to "information misalignment." Therefore, this step is a prerequisite for effective fusion. Regarding time synchronization, the high-precision network time protocol clock of the SCADA system is used as a unified reference. For data streams such as acoustic vibration and infrared thermal imaging, the timestamp information within the data packets is used, and Lagrange interpolation or cubic spline interpolation algorithms are employed to resample them to a time series completely consistent with the SCADA data.

[0085] In a preferred embodiment, spatial registration primarily targets infrared thermal imaging data. A three-dimensional geometric model of the device is pre-established, and the spatial position and orientation (i.e., extrinsic parameter matrix) of the infrared thermal imager relative to this three-dimensional model are determined using camera calibration techniques. Perspective projection or orthographic projection transformation is used to map each pixel in the two-dimensional infrared image to a specific surface location on the three-dimensional model. Conversely, the grid points on the surface of the three-dimensional model are projected onto the two-dimensional infrared image, thus establishing a one-to-one mapping relationship between the three-dimensional model and the two-dimensional thermal image. Ultimately, a spatiotemporally aligned multimodal dataset is obtained, in which observations of multiple states (electrical, mechanical, thermal, etc.) at the same time and physical location can be acquired at any given point in time.

[0086] S2.2: Apply an attention fusion network to calculate the weights of each modality's contribution to the latent state, and generate weighted multimodal fusion features.

[0087] In this embodiment, an attention-based neural network model is designed to adaptively learn the importance of different data modalities under different operating conditions. The input to the network is the feature vectors of each modality (electrical parameters, vibration features, temperature matrix features, and discharge features) aligned using S2.1. The network first maps the features with different dimensions and distributions to a unified latent space through independent embedding layers to obtain their respective feature representations.

[0088] For example, the core of the attention mechanism is a learnable query-key-value computation process. Specifically, the current overall operating conditions of the device (such as load current and ambient temperature) are taken as the "query." The feature representations of each modality's data are taken as the "keys" and "values." By calculating the similarity between the query and each key (usually using dot product or additive attention) and normalizing it through the Softmax function, a set of attention weights is obtained. This set of weights intuitively reflects the relative importance of each data modality in judging the overall health status of the device under the current specific operating conditions. For example, under heavy load conditions, the weight of thermal imaging data may be higher; while when specific harmonics are present, the weight of vibration data may increase.

[0089] In a further improvement to this embodiment, the network then uses the calculated attention weights to perform a weighted summation of the respective "values" (i.e., feature representations of each modality), ultimately generating a fused feature vector that integrates information from all modalities and is dynamically weighted according to the context. This process simulates the thought process of human experts who emphasize different clues depending on the situation during diagnosis.

[0090] S2.3: High-level abstraction of weighted fusion features is performed through deep belief networks to extract low-dimensional device implicit state codes.

[0091] In one alternative implementation, a deep belief network (DBN) is used as a feature extractor to extract the most essential and compact health status information from the high-dimensional weighted fusion features. The DBN is composed of multiple layers of restricted Boltzmann machines and possesses powerful unsupervised feature learning capabilities. The weighted fusion feature vector output from S2.2 is input into the DBN.

[0092] Specifically, each layer of the DBN's Restricted Model (RBM) is pre-trained using a contrastive divergence algorithm to learn how to reconstruct its input data with fewer neurons, thereby capturing statistical regularities at different levels. Lower-level RBMs learn simple local features, while higher-level RBMs combine these features to form more global and abstract feature representations as the layers deepen. Preferably, after pre-training, the entire DBN undergoes supervised fine-tuning using backpropagation to optimize its performance for the final prediction task.

[0093] The final layer of the network is a bottleneck layer, where the number of neurons is intentionally set much smaller than that of the input layer. This bottleneck layer forces the network to learn how to retain the most critical parts of the input data with as little information as possible. Its output is a low-dimensional, dense vector called "latent state encoding." This encoding strips away noise and redundant information from the original data, retaining the core intrinsic factors driving the evolution of the device's state, and is the essential representation of the device's true health state.

[0094] S3: Construct a dynamic state deduction model of physical constraints and perform trajectory prediction.

[0095] S3.1: Establish a physical constraint model for the equipment, define the evolution law and boundary conditions of the state variables, and construct a physical knowledge base.

[0096] Further improvements to this embodiment, to ensure the prediction model results conform to basic physical laws and engineering common sense, firstly, a physical knowledge base is constructed. This knowledge base is defined in the form of mathematical equations or inequality constraints, based on the design principles of the equipment, materials science, and failure physics. For example, for transformer insulation paper, an Arrhenius model of its aging is introduced, linking insulation life to operating temperature; for circuit breaker contact electrical wear, a wear model based on cumulative breaking current is introduced; and for mechanical components, a fatigue accumulation model based on stress-cycle count is introduced.

[0097] In one alternative implementation, these models are typically represented as differential or algebraic equations. These equations are discretized to facilitate representation and processing in a computer. For example, continuous-time differential equations are transformed into difference equations. Furthermore, feasible domain boundaries for state variables are defined, such as the temperature not falling below ambient temperature, insulation resistance not being negative, and wear not exceeding 100%. All these discretized equations and inequalities together constitute a set of physical constraints, representing deterministic prior knowledge of the device's behavior.

[0098] S3.2: Construct a physical information neural network, embed physical constraints as a loss function into the deep learning model, and generate a model for inferring physical laws.

[0099] Specifically, a feedforward neural network is constructed, whose input is the hidden state encoding at the current moment and a future time interval, and whose output is the predicted state at the future moment. Unlike traditional neural networks that only minimize the error between the predicted value and the true value, the innovation of the physical information neural network lies in the design of its loss function.

[0100] For example, PINN's loss function consists of two parts: a conventional "data loss," which is the mean squared error between the network's prediction and a small amount of available real historical data; and a crucial "physical loss." The physical loss is calculated by substituting the network's prediction into the physical constraint equations (such as differential equations) established in S3.1, and calculating the residuals on both sides of the equations. A perfect prediction should result in zero residuals for these physical equations. Therefore, the training objective of PINN is to minimize the sum of the data loss and the physical loss.

[0101] In this way, physical laws are embedded into the training process of the neural network in the form of "soft constraints." The network not only learns from historical data but is also forced to learn from physical laws. Even in situations where historical data is lacking, physical constraints can guide the network to make reasonable predictions, thereby enhancing the model's generalization ability and the reliability of extrapolated predictions. After training, a future state prediction model that understands both data and physics is obtained and can be used for extrapolation.

[0102] S3.3: Based on the current implicit state encoding, perform multi-step iterative prediction to generate the evolution trajectory and uncertainty range of the device's future state.

[0103] In a preferred embodiment, an iterative prediction method is used when long-term prediction is required. First, the hidden state code S_t at the current time t and the next time step Δt are input into a trained PINN to obtain the predicted state value S_{t+1} at time t+Δt. Then, S_{t+1} is used as a new input to the same PINN to predict the state S_{t+2} at time t+2Δt. This process is repeated until the required time range is predicted, thereby generating a complete state evolution trajectory from the present to the future.

[0104] A further improvement to this embodiment is the use of Monte Carlo Dropout technology in the prediction phase to quantify the uncertainty of the prediction. Specifically, a certain proportion of neurons in the neural network are randomly shut down during each forward propagation calculation. Multiple (e.g., 100) predictions with Dropout are performed on the same initial state S_t, resulting in 100 slightly different prediction trajectories. These trajectories form a "prediction band." The width of this prediction band reflects the model's uncertainty regarding future predictions: concentrated trajectories indicate high confidence, while dispersed trajectories indicate low confidence. Finally, the mean of this prediction trajectory (as the most probable evolutionary path) and its upper and lower boundaries (as confidence intervals) are output, providing more comprehensive information for risk assessment.

[0105] S4: Interactive Maintenance Decision-Making and Execution Based on Digital Twins

[0106] S4.1: Set up a virtual experimental environment in the digital twin, load different operating parameters and maintenance schemes, and generate a set of virtual scenes.

[0107] Specifically, a high-fidelity 3D model of the equipment is combined with a pre-trained dynamic simulation model using S3 to create an immersive virtual experimental environment. Maintenance personnel can define a series of "what if..." scenarios within this environment. For example, scenario one: the equipment runs continuously under peak summer load without any maintenance; scenario two: a preventative cleaning is performed after three months; scenario three: a suspected aging component is replaced after six months; scenario four: the equipment's load distribution strategy is adjusted.

[0108] In one alternative implementation, these scenarios are configured via a graphical interface. Users only need to set key parameters (such as future load curves, planned maintenance times, and maintenance types), and the system automatically translates these operational instructions into input parameters that the simulation model can recognize. Each scenario corresponds to a complete set of initial conditions and external intervention sequences. All these scenarios constitute a set of virtual scenarios to be evaluated, preparing for subsequent batch simulations and scheme comparisons.

[0109] S4.2: Drive the simulation model to perform batch simulations of each virtual scenario, evaluate the long-term effects of each scheme, and obtain a simulation result database.

[0110] According to a further improvement in this embodiment, the system automatically schedules computing resources to perform long-term simulations of each scenario in the virtual scene set. For each scenario, starting from the current real state, the system drives the physical information neural network simulation model in S3 according to the load conditions and maintenance plan defined for that scenario, simulating the state evolution of the equipment over the next year or even longer.

[0111] For example, during the simulation, the system records time-series data of a series of key performance indicators, such as changes in health indices, cumulative failure probabilities, total power consumption, and maintenance costs. After the simulation, the output results of each scenario are summarized and analyzed to calculate the final comprehensive benefits, such as total maintenance costs, equipment availability, and risk losses due to failures. Finally, the configuration information of all scenarios and their corresponding simulation results (including complete historical trajectories and summary indicators) are systematically stored in a simulation results database, which clearly reveals the different long-term consequences that different decisions may lead to.

[0112] S4.3: Based on a multi-objective optimization algorithm, it weighs maintenance costs, equipment risks, and operational efficiency to output the recommended optimal maintenance strategy.

[0113] In a preferred embodiment, since this is a typical multi-objective optimization problem, the objectives of "lowest cost," "lowest risk," and "highest efficiency" are often conflicting. An optimization model is established, with decision variables being parameters of the maintenance strategy (such as maintenance timing and type), and its objective function, based on the simulation results of S4.2, aims to minimize total cost, minimize failure risk, and maximize equipment availability.

[0114] Specifically, an advanced multi-objective evolutionary algorithm, such as NSGA-II (Non-dominated sorting genetic algorithm with elitist strategy), is employed to solve this optimization problem. This algorithm searches a vast space of possible solutions by simulating biological evolution processes such as selection, crossover, and mutation. Instead of outputting a single "optimal solution," it finds a set of "Pareto optimal solutions." In this set, improvement in any objective must come at the expense of at least one other objective deteriorating. Finally, the system presents this Pareto optimal solution set to the decision-maker, with each solution corresponding to a feasible maintenance strategy and its expected multi-objective outcomes. The decision-maker can then select a solution from this set of non-dominated solutions based on current management preferences (e.g., prioritizing cost control or power supply reliability), thus achieving scientific, transparent, and optimal maintenance decisions.

[0115] The preferred embodiments of the present invention have been described in detail above. However, the present invention is not limited to the specific details in the above embodiments. Within the scope of the technical concept of the present invention, various equivalent transformations can be made to the technical solutions of the present invention, and these equivalent transformations all fall within the protection scope of the present invention.

Claims

1. A power equipment implicit state predictive maintenance method characterized by, The method comprises the following steps: Collecting multi-modal operation data of power equipment, including real-time operation data of SCADA system, acoustic vibration data, infrared thermal image data and partial discharge data; Temporal and spatial alignment and feature fusion are performed on the multi-modal operation data to extract implicit state codes representing the internal health status of the equipment; A dynamic state deduction model is constructed based on physical constraint model and deep learning network to iteratively predict the implicit state codes for multiple steps and generate the evolution trajectory of the future state of the equipment; Based on the evolution trajectory, virtual maintenance experiments are carried out in the digital twin, and the optimal maintenance strategy is generated through multi-objective optimization.

2. The method of claim 1, wherein, Collecting multi-modal operation data of power equipment, including: Collecting three-phase current data, three-phase voltage data, equipment temperature data and switch state data in the SCADA system through a data acquisition interface, performing data cleaning and normalization processing to obtain a standardized electrical operation parameter data set; Collecting mechanical vibration waveform data of the equipment through a vibration acceleration sensor, performing fast Fourier transform and frequency spectrum feature extraction to obtain a mechanical vibration feature vector containing characteristic frequency amplitude, harmonic component and frequency spectrum entropy; Collecting a sequence of infrared images of the equipment surface through an infrared thermal imager, performing non-uniformity correction and radiation calibration, reconstructing a two-dimensional temperature field and calculating a temperature gradient distribution to obtain a temperature distribution state matrix; Collecting a sequence of partial discharge pulses through a high-frequency current transformer, performing pulse feature extraction and clustering analysis to obtain a discharge severity index.

3. The method of claim 1, wherein, Temporal and spatial alignment and feature fusion are performed on the multi-modal operation data, including: Taking the timestamp of the SCADA system as the reference, the sampling time of the acoustic vibration data, the infrared thermal image data and the partial discharge data is synchronized by interpolation method to obtain a time-aligned multi-modal data set; Establishing a spatial mapping relationship between the three-dimensional model of the equipment and the infrared thermal image data, and realizing spatial registration through perspective projection transformation; Constructing an attention mechanism fusion network to calculate the weight coefficients of each modality data in the multi-modal data set with the equipment operating condition as the query vector; Based on the weight coefficients, the multi-modal features are weighted and fused, and the low-dimensional implicit state codes are extracted through a deep belief network.

4. The method of claim 1, wherein, Constructing a dynamic state deduction model, including: Based on the material properties and physical principles of the equipment, differential equations and inequality constraints of state variable evolution law are established to construct a physical knowledge base; Designing a physical information neural network architecture, embedding the constraint conditions in the physical knowledge base into the loss function as a regularization term; Training the physical information neural network using historical state code sequences to obtain a physical constraint deduction model; Using Monte Carlo Dropout technology to sample the prediction process multiple times to calculate the confidence interval of the prediction result.

5. The method of claim 1, wherein, Generating the evolution trajectory of the future state of the equipment, including: Inputting the current implicit state code into the trained physical information neural network to predict the state code at the next time; Iteratively performing multi-step prediction using the predicted state code as the new input to generate a state evolution sequence within a specified time range; Based on the results of multiple Monte Carlo sampling, the mean and variance of the state prediction value are calculated to form an evolution trajectory with a confidence interval; In the three-dimensional visualization interface, the evolution trajectory is dynamically displayed using color gradient and transparency changes.

6. The method of claim 1, wherein, Virtual maintenance experiments are conducted in the digital twin, including: Different load conditions, environmental factors, and maintenance operations are set in the digital twin model of the device to generate a set of virtual experiment scenarios; For each virtual scenario, the dynamic state inference model is driven to simulate long-term state evolution with the current device state as the initial condition; The evolution curves of key performance indicators, including health index, failure probability, and maintenance cost, are recorded under each virtual scenario; Based on the simulation results, a maintenance strategy effectiveness database is established to provide data support for multi-objective optimization.

7. The method of claim 1, wherein, The optimal maintenance strategy is generated through multi-objective optimization, including: A multi-objective optimization function is established with the objectives of minimizing maintenance cost, minimizing device risk, and maximizing operational efficiency; Resource availability, time window, and technical feasibility are used as constraint conditions; The non-dominated sorting genetic algorithm with elitism is used to search in the solution space to obtain a set of Pareto optimal solutions; Based on the decision maker's preference, the final maintenance plan is selected from the set of Pareto optimal solutions, and a structured maintenance work order is generated.

8. The method of claim 1, wherein, It also includes the steps of augmented reality assisted execution: Key information of the optimal maintenance strategy is converted into augmented reality visualization elements; The visualization elements are superimposed on the real device scene through AR devices to provide operation guidance and parameter prompts; Based on gesture recognition and voice interaction, feedback from on-site personnel is received to dynamically adjust the maintenance plan; The data of the maintenance execution process is recorded for updating the device history and optimizing the inference model.

9. The method of claim 3, wherein, The specific implementation of the attention mechanism fusion network includes: The feature vectors of each modality data are mapped to a unified hidden space through independent embedding layers; Using device load rate and environmental temperature as query vectors, the similarity scores with each modality key vector are calculated; The similarity scores are converted to attention weights through the Softmax function; The weighted sum of each modality value vector is obtained using the attention weights to get the context-aware fusion features.

10. The method of claim 4, wherein, The physical knowledge base includes the following constraint models: Insulating material thermal aging model based on Arrhenius equation; contact electrical wear model based on cumulative breaking current; Mechanical fatigue accumulation model based on stress-cycle number; temperature field evolution model based on heat conduction theory; And the physical boundary constraint conditions of each state variable.

Images