Station building degradation track prediction method fusing physical prior and spatio-temporal knowledge embedding

Abstract

The invention belongs to the technical field of power distribution station operation state monitoring, and discloses a station building degradation track prediction method fusing physical prior and spatio-temporal knowledge embedding. Comprising the steps of collecting multi-source heterogeneous data; constructing a physical constraint equation of key physical quantities in the station building equipment; constructing a space-time knowledge graph, adopting a graph attention network as a graph embedding model, embedding attention weights of nodes to neighbor nodes, and obtaining a node embedding vector set; performing feature extraction on the multi-source heterogeneous data to obtain mixed features; a degradation track prediction result is obtained based on physical information neural network analysis, health degree evaluation is carried out on the degradation track prediction result, a health degree score is obtained, and corresponding early warning is carried out according to the health degree score; according to the method, high-precision prediction of the degradation track of the station building equipment in a fault-free data scene is realized, and meanwhile, the physical rationality and operation and maintenance guidance of a prediction result are guaranteed.

Description

Technical Field

[0001] This invention relates to the field of power distribution station operation status monitoring technology, and in particular to a method for predicting station degradation trajectories by integrating physical priors and spatiotemporal knowledge. Background Technology

[0002] As a core infrastructure for the safe and stable operation of power systems, the degradation trajectory prediction of substation equipment is crucial for identifying equipment deterioration trends in advance and formulating precise operation and maintenance strategies. However, existing degradation prediction technologies still have significant limitations and cannot meet actual needs: On the one hand, although purely data-driven prediction methods can capture statistical patterns in data, they are highly dependent on a large amount of historical fault data and complete degradation samples. However, actual substation equipment generally has long lifespan and low failure rate characteristics, and fault data, especially early degradation data, is scarce. This leads to insufficient model training and poor generalization ability, making it difficult to achieve accurate extrapolation prediction in scenarios lacking fault data. It is also prone to deviating from the true degradation trend due to overfitting normal data. On the other hand, although purely physical mechanism-driven prediction methods are based on equipment degradation mechanisms... While constructing equation models does not rely on fault data, the complex structure and multi-physics coupling of station equipment make it difficult to establish comprehensive and accurate mechanistic equations. Furthermore, it ignores the dynamic correlation information of multi-source heterogeneous monitoring data, resulting in limited prediction accuracy and an inability to effectively characterize the subtle degradation process of equipment. In addition, existing technologies generally fail to solve the problem of effective fusion of multi-source heterogeneous data, resulting in data silos. Moreover, they do not consider the spatiotemporal correlation between station equipment, leading to a lack of comprehensive feature representation. This further restricts the accurate perception of the overall degradation state of the equipment by the prediction model. Ultimately, existing technologies are unable to achieve accurate identification of early equipment degradation and reliable prediction of degradation trajectories, failing to meet the needs of full life-cycle health management of station equipment and safe operation and maintenance of power systems.

[0003] The patent application CN118801579B discloses a health assessment system for substations based on multi-source information, which includes: a sensor unit for collecting operational data from the substation and transmitting it to a data processing unit and a data analysis unit; a data processing unit for preprocessing the operational data collected by the sensor unit and transmitting the preprocessing results to the data analysis unit; a data analysis unit for scoring and analyzing the operational data collected and processed by the sensor unit and the data processing unit, and transmitting the scoring and analysis results to a data display unit; and a data display unit for presenting the results obtained by the data analysis unit to relevant personnel in an intuitive manner.

[0004] The above methods cannot achieve accurate extrapolation predictions in the absence of historical fault data.

[0005] To address the aforementioned issues, this invention proposes a method for predicting station degradation trajectories that integrates physical priors with spatiotemporal knowledge embedding. Summary of the Invention

[0006] The purpose of this invention is to propose a method for predicting station degradation trajectories that integrates physical priors and spatiotemporal knowledge embedding to solve the problems mentioned in the background art. Existing technologies cannot achieve accurate extrapolation predictions in the absence of historical failure data.

[0007] To achieve the above objectives, the present invention adopts the following technical solution: A method for predicting station degradation trajectories by integrating prior physical knowledge with spatiotemporal knowledge includes the following steps: Collect and preprocess station equipment operation data to obtain multi-source heterogeneous data; Construct physical constraint equations for the degradation process of key physical quantities in station equipment, as well as corresponding boundary constraint equations, and obtain the physical constraint equations. A spatiotemporal knowledge graph containing a set of nodes, a set of relationships, and a set of attributes is constructed. A graph attention network is used as the graph embedding model to embed the attention weights of nodes to their neighbors, thereby obtaining a set of node embedding vectors. Feature extraction is performed on multi-source heterogeneous data to obtain data-driven features. Physical constraint equations are analyzed to obtain physical derived features, which are 3-dimensional feature vectors obtained by splicing wear rate, stiffness decay rate and aging acceleration factor. The nodes are standardized by Z-score by embedding the vector set, data-driven features, and physical derived features. The standardized nodes are then concatenated with the embedding vector set, data-driven features, and physical derived features to obtain hybrid features. The hybrid features are used as input to the physical information neural network to obtain the degradation trajectory prediction results. The health of the degradation trajectory prediction results is evaluated to obtain a health score, and corresponding warnings are issued based on the health score.

[0008] Preferably, the station equipment operation data includes electrical data, mechanical data, environmental data, visual images, and inspection data.

[0009] Preferably, the key physical quantities include switchgear contacts, circuit breaker springs, and transformer windings.

[0010] Preferably, the method for obtaining the physical constraint equations includes: The cumulative number of opening and closing cycles and contact resistance are obtained. Combined with the electrical wear coefficient, mechanical wear coefficient and the current at the corresponding moment, an equation for the amount of contact wear is constructed. The initial stiffness and stress cycle number of the spring are obtained, and the spring stiffness decay equation is constructed by combining the attenuation coefficient and the fatigue life at the corresponding stress level. Obtain the dielectric loss angle and winding temperature, and combine the temperature acceleration coefficient and reference temperature to construct the winding insulation aging index equation. Construct boundary constraint equations, including: wear amount is non-negative; spring stiffness is not lower than the lower stiffness threshold; winding insulation aging index is not higher than the upper aging index threshold. The physical constraint equation is obtained by combining the contact wear equation, the spring stiffness decay equation, the winding insulation aging index equation, and the boundary constraint equation.

[0011] Preferably, the method for constructing a spatiotemporal knowledge graph that includes a set of nodes, a set of relationships, and a set of attributes includes: Obtain the station's coverage area, including spatial scope, equipment scope, and time scope; Entities are categorized into equipment unit entities, component entities, sensor entities, spatial region entities, and event / physical quantity entities; Define the relationships between entities, and analyze the domain, range, and type of the relationships; Define the attributes of an entity, including static attributes and dynamic attributes; A pre-defined hierarchical structure of station buildings, floors, bays, and installation sites is adopted to associate each spatial area entity with the superior area; Bind all entities within the station building to absolute coordinates based on the station building's coordinate system and relative coordinates based on the coordinate system of their respective spatial regions; Acquire station equipment testing data and automatically extract and convert it based on preset conversion tools; obtain data entities and map the data entities into a graph format according to the correspondence between entities and attributes to obtain a spatiotemporal knowledge graph.

[0012] Preferably, the method for obtaining the node embedding vector set includes: Randomly assign an initial vector to each node; For nodes neighboring nodes The attention weights are calculated by combining the attention parameters, the linear transformation matrix, and the node embedding vector. Nodes based on attention weights The embedding vector is updated; The updated embedding vectors of each node are concatenated to obtain the node embedding vector set.

[0013] Preferably, the data-driven features include data features and image features; Methods for obtaining data features include: Statistical analysis is performed on time-series data including electrical data, mechanical data, environmental data, and inspection data to obtain the mean and variance. The absolute mean of the time-series data is obtained, and the peak factor is calculated based on the absolute mean. Perform Fourier transform on the mechanical data to obtain the Fourier transform signal, extract the Fourier transform signal to obtain the main frequency, spectral energy and spectral entropy; concatenate the main frequency, spectral energy and spectral entropy to obtain the frequency domain features; Wavelet transform is performed on the mechanical data to calculate the wavelet entropy; the wavelet entropy is then used as a time-frequency domain feature. Data features are obtained by concatenating time-domain features, frequency-domain features, and time-frequency-domain features.

[0014] Preferably, the image features include visual features and infrared features; Methods for obtaining image features include: The device appearance image is used as input to the convolutional network to obtain convolutional features; The image of the equipment's appearance is converted to grayscale and subjected to Gaussian denoising to obtain a preprocessed image; The preprocessed image is used as input to the localization model to obtain the coordinates of the bounding box of the damaged area; the localization model is the YOLOv8n model. The preprocessed image is used as input to the U-Net model to obtain the probability of wear regions. Image regions with wear region probabilities higher than the probability threshold are segmented as ROI regions. The number of pixels in the damaged area and the number of pixels in the ROI area included in the bounding box coordinates of the damaged area are counted, and the ratio of the number of pixels in the damaged area to the number of pixels in the ROI area is calculated to obtain the pixel area ratio of the worn area. Visual features are obtained by concatenating convolutional features and the pixel area ratio of the worn area. The infrared image is converted based on the radiation calibration model of the infrared thermal imager to obtain the pixel temperature; The region is identified where the temperature of consecutive pixels is higher than a temperature threshold and the number of pixels with a temperature higher than the temperature threshold is greater than G. The number of pixels with a temperature higher than the temperature threshold and the total number of pixels in the infrared image are counted. The ratio of the number of pixels to the total number of pixels is calculated to obtain the proportion of hot areas. The highest temperature value among pixels whose temperature exceeds the temperature threshold is counted. Calculate the average temperature of pixels whose temperature is below the temperature threshold, and calculate the difference between the highest temperature value and the average temperature value to obtain the temperature gradient. Infrared features are obtained by stitching together the proportion of hot zones and temperature gradients.

[0015] Preferably, the method for obtaining the screening results includes: The mutual information entropy between data features, image features, and physical derived features and degradation indicators is calculated separately, and features greater than the degradation threshold are selected as the screening results.

[0016] Preferably, the physical information neural network is constructed from a 2-layer LSTM, a Transformer encoder with 4 attention heads, and a spatiotemporal attention layer; A data fitting loss is established based on the error between the actual degradation index and the predicted value, and a physical constraint loss is established based on equation constraints and boundary constraints. The total loss function is calculated based on the data fitting loss and the physical constraint loss. The physical information neural network with the minimum total loss function is obtained by optimization based on the Adam optimizer and used as the training output physical information neural network. Methods to obtain a health score include: The degradation trajectory prediction results are quantified based on a piecewise function constructed according to the normal threshold and the fault threshold of the degradation index to obtain a health score. When the health score is below the score threshold, a health warning is triggered; when the rate of change of the health score is above the rate of change threshold, a trend warning is triggered.

[0017] Compared with existing technologies, this invention provides a method for predicting station degradation trajectories by integrating physical priors and spatiotemporal knowledge embedding, which has the following beneficial effects: This invention collects and preprocesses multi-source heterogeneous data from station equipment to construct a comprehensive equipment status representation to supplement information gaps in historical fault data. It constructs physical constraint equations and boundary constraint equations for switchgear contact wear, circuit breaker spring stiffness decay, and transformer winding insulation aging index, injecting physical mechanisms of equipment degradation into predictions and avoiding predictions that violate physical laws when there is no fault data. Furthermore, it constructs a spatiotemporal knowledge graph containing entities such as equipment units, components, and sensors, along with spatiotemporal, physical, and functional relationships. Combined with a GAT model to generate node embedding vector sets, it transforms equipment-related information into structured features to enhance the model's generalization ability in small-sample scenarios. Finally, it extracts data-driven features and physically derived features, and uses mutual information entropy for filtering and strong correlation with degradation indicators. The system employs a combined approach, removing false features to improve feature quality. By constructing a physical information neural network and optimizing the model using a total loss function that integrates data fitting loss and physical constraint loss, it can accurately capture subtle degradation trends of equipment without relying on historical fault data, ensuring the rationality of extrapolation predictions. Finally, by calculating health scores through piecewise functions and triggering health and trend warnings, it effectively solves the core problem of existing technologies being unable to achieve accurate extrapolation predictions when historical fault data is lacking. This enables high-precision prediction of station equipment degradation trajectories in fault-free data scenarios, while ensuring the physical rationality and operational guidance of the prediction results. It significantly improves the reliability and generalization ability of equipment degradation prediction, providing strong technical support for the full lifecycle health management and operational decision-making of station equipment. Attached Figure Description

[0018] Figure 1This is a schematic diagram of the method flow mentioned in Embodiment 1 of the present invention; Figure 2 This is a schematic diagram of the data flow mentioned in Embodiment 1 of the present invention. Detailed Implementation

[0019] 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.

[0020] Example 1: Please see Figure 1-2 The present invention provides a method for predicting station degradation trajectories that integrates physical priors and spatiotemporal knowledge, comprising the following steps: The system collects and preprocesses station equipment operation data to obtain multi-source heterogeneous data. This data includes electrical, mechanical, environmental, visual images, and inspection data. Electrical data includes corresponding current, voltage, contact resistance, partial discharge, and the number of electrical data samples. Mechanical data includes vibration acceleration, opening and closing times, and operating torque. Environmental data includes temperature, humidity, and dust concentration. Visual images include equipment appearance images and infrared images. Inspection data includes bolt tightness status, etc. By collecting electrical, mechanical, environmental, visual images, and inspection data from station equipment and preprocessing this multi-source heterogeneous data, a state characterization system covering all dimensions of equipment electrical, mechanical, environmental, appearance, and key components can be constructed. This system overcomes the limitations of existing technologies that rely on historical fault data to mine degradation patterns. Even in scenarios lacking fault samples, it can provide rich prior information on equipment operation and degradation through the subtle degradation signals, multi-dimensional state correlation patterns, and external influencing factors provided by complementary multi-source data. This information is then used to inject rich prior information on equipment operation and degradation into subsequent prediction models that integrate physical priors and spatiotemporal knowledge. This supports the model in accurately capturing the inherent logic of equipment degradation trends, thereby achieving accurate extrapolation prediction of the future degradation trajectory of the equipment. This effectively solves the technical problem of insufficient extrapolation prediction accuracy caused by the scarcity of fault data in existing technologies.

[0021] The degradation process physical constraint equations of key physical quantities in the station equipment are constructed, as well as the corresponding boundary constraint equations. The degradation process physical constraint equations and the boundary constraint equations are integrated to construct the physical constraint equations. The physical constraint equations mentioned in other steps refer to the comprehensive physical constraint equations constructed in this step. Key physical quantities include switch cabinet contacts, circuit breaker springs and transformer windings. Methods for obtaining physical constraint equations include: By obtaining the cumulative number of opening and closing cycles and the contact resistance, and combining the electrical wear coefficient, mechanical wear coefficient, and current at the corresponding time, an equation for the contact wear amount is constructed; for example, the contact wear amount at time t. ,in, The coefficient of electrical wear; The mechanical wear coefficient; This is the initial time. for Current at any moment; for Constantly in contact with the resistor; The cumulative number of opening and closing operations at time t; By obtaining the initial stiffness and stress cycle count of the spring, and combining the attenuation coefficient and fatigue life at the corresponding stress level, a spring stiffness attenuation equation is constructed; for example, the spring stiffness attenuation equation at time t. ,in, This is the initial stiffness; This is the fatigue attenuation coefficient; For the first Number of stress cycles per operation; N fl This refers to the fatigue life corresponding to the stress level. Obtain the dielectric loss angle and winding temperature, and combine them with the temperature acceleration factor and reference temperature to construct the winding insulation aging index equation. For example, the winding insulation aging index is: , in, for Constant time interval loss angle; The acceleration factor is the temperature coefficient. for Constant winding temperature; T ref The reference temperature is typically 25°C. Construct boundary constraint equations, including: wear amount is non-negative; spring stiffness is not lower than the lower stiffness threshold; winding insulation aging index is not higher than the upper aging index threshold. By integrating the equations for contact wear, spring stiffness decay, winding insulation aging index, and boundary constraints, a set of equations is obtained, which provides both the mechanism and boundary constraints for the degradation of key equipment components. The essence of this assembly of physical constraint equations is the synergistic constraint of the degradation mechanisms of multiple components: contacts, springs, and windings are the core failure components of the station equipment; their degradation equations respectively characterize the degradation patterns of the electrical, mechanical, and insulation systems. The boundary equations ensure that the degradation of each component does not exceed the safe operating limits. Together, the set of equations constitutes the physical anchor point of the prediction model, avoiding physical inaccuracies in model predictions when there is no fault data.

[0022] A spatiotemporal knowledge graph containing a set of nodes, a set of relationships, and a set of attributes is constructed. A graph attention network, namely GAT, is used as the graph embedding model to embed the attention weights of nodes to their neighboring nodes, thereby obtaining a set of node embedding vectors. Methods for constructing spatiotemporal knowledge graphs that include sets of nodes, relationships, and attributes include: Obtain the station's coverage area, including spatial scope, equipment scope, and time scope; Entities are categorized into equipment unit entities, component entities, sensor entities, spatial region entities, and event / physical quantity entities; Define the relationships between entities, and analyze the domain, range, and type of the relationships; Define the attributes of an entity, including static attributes and dynamic attributes; A pre-defined hierarchical structure of station buildings, floors, bays, and installation sites is adopted to associate each spatial area entity with the superior area; Bind all entities within the station building to absolute coordinates based on the station building's coordinate system and relative coordinates based on the coordinate system of their respective spatial regions; Acquire station equipment testing data and automatically extract and transform it using preset conversion tools, such as ETL tools; obtain data entities and map the data entities into a graph format according to the correspondence between entities and attributes to obtain a spatiotemporal knowledge graph.

[0023] Methods for obtaining node embedding vector sets include: Randomly assign an initial vector to each node; For nodes neighboring nodes The attention weights are calculated by combining the attention parameters, the linear transformation matrix, and the node embedding vector; for example, for nodes... neighboring nodes Attention weights:

[0024] in, For attention parameters; It is a linear transformation matrix; For nodes The set of neighbors; This indicates vector concatenation; The slope is negative, with a value of 0.2. h v For nodes The embedding vector; h u For nodes The embedding vector; The embedding vector; Nodes based on attention weights Update the embedding vectors, such as those of nodes. The updated embedding vector is:

[0025] in, For activation functions; The updated embedding vectors of each node are concatenated to obtain the node embedding vector set.

[0026] By constructing physical constraint equations and boundary constraint equations for the degradation process of key physical quantities of station equipment, and combining this with the construction of a spatiotemporal knowledge graph and the acquisition of the graph embedding vector set, core support is provided for solving the problem that existing technologies cannot achieve accurate extrapolation prediction when historical fault data is lacking. On the one hand, for switchgear contacts, based on the acquired cumulative number of opening and closing operations, contact resistance, and current at the corresponding moment, a contact wear equation is constructed by combining the electrical wear coefficient and the mechanical wear coefficient to quantify the degradation process of contacts caused by electrical erosion and mechanical friction. On the other hand, for circuit breaker springs, the initial stiffness and stress cycle number are used, combined with the attenuation coefficient and the fatigue coefficient of the corresponding stress level. The lifespan equation for spring stiffness decay characterizes the stiffness change pattern during spring fatigue aging. For transformer windings, an insulation aging index equation is constructed based on the dielectric loss angle, winding temperature, temperature acceleration coefficient, and reference temperature. This equation reflects the aging trend of winding insulation with temperature and dielectric loss variations. Furthermore, boundary constraint equations are added to ensure non-negative wear, spring stiffness not falling below a lower threshold, and winding temperature not exceeding an upper threshold. These equations, built upon physical mechanisms, do not rely on historical fault data and provide underlying physical constraints for equipment degradation in the prediction model. This prevents purely data-driven models from deviating from their intended path due to overfitting to normal data when fault samples are scarce. The study examines the actual degradation trend. On the other hand, by clearly defining the spatial scope, equipment scope, and time scope of the station, entities such as equipment units, components, sensors, spatial regions, and events / physical quantities are categorized. Spatial, physical, functional, and temporal relationships and static and dynamic attributes between entities are defined. A hierarchical structure of station buildings, floors, intervals, and installation sites is used to associate spatial region entities and bind their absolute and relative coordinates. ETL tools are used to extract transformation detection data and generate a spatiotemporal knowledge graph. Then, using a graph attention network as the embedding model, an initial vector is assigned to each node. Finally, attention parameters, linear transformation matrices, and node embedding vectors are combined to calculate the node's relationship with its neighbors. The attention weights of nodes are assigned, and the node embedding vectors are updated based on these weights to form a vector set. This process can transform multi-source heterogeneous data into structured features rich in spatiotemporal correlations and semantic relationships between devices. Even when fault data is lacking, the gap in score information can be supplemented by the correlation information between devices. Finally, the mechanism guidance provided by the physical constraint equation and the correlation features provided by the spatiotemporal knowledge graph embedding vector set work together to enable the prediction model to not only follow physical laws to ensure the rationality of the extrapolation direction, but also to capture subtle degradation signals with the help of multi-dimensional correlation information. This effectively breaks through the limitations of existing technologies that rely on historical fault data and achieves accurate extrapolation prediction of equipment degradation trajectories.

[0027] Feature extraction is performed on multi-source heterogeneous data to obtain data-driven features, and physical constraint equations are analyzed to obtain physical-derived features. The node embedding vector set, data-driven features, and physical-derived features are standardized and concatenated to obtain hybrid features. For example, assuming the node embedding vector set contains 10 nodes, each with an embedding dimension of 16, it will be a 1×160 vector after flattening; the data-driven features include 3D time domain, 3D frequency domain, and 10D image features, totaling 1×16 vectors; the physical-derived features are 1×3 vectors; after standardization, they are concatenated to form a hybrid feature with 1×(160+16+3)=1×179 dimensions, ensuring the uniformity of feature dimensions in the input model.

[0028] Data-driven features include data features and image features; Methods for obtaining data features include: Statistical analysis was performed on time-series data including electrical, mechanical, environmental, and inspection data to obtain the mean and variance. The absolute mean of the time-series data was then calculated, and the peak value was determined based on the absolute mean. For example, the absolute mean... Peak factor The mean, variance, and peak factor are concatenated to obtain the time-domain features; This represents the number of time-series data points. For the first One time series data; Perform a Fourier transform on the mechanical data to obtain the Fourier transform signal. Extract the Fourier transform signal to obtain the dominant frequency, spectral energy, and spectral entropy. Concatenate the dominant frequency, spectral energy, and spectral entropy to obtain the frequency domain characteristics; such as the Fourier transform signal. main frequency Spectral energy Spectral entropy ,in, j The imaginary unit; f Frequency is used to describe a signal. Its variation characteristics in the frequency domain; Time is used to describe the signal. Its characteristics of change in the time domain; It is the vibration acceleration; f max This represents the point with the maximum frequency range. For the first Energy percentage of each frequency band; Wavelet transform is performed on the mechanical data to calculate the wavelet entropy; this wavelet entropy is then used as a time-frequency domain feature. (e.g., wavelet entropy) ,in, J The number of decomposition levels; q jl For the first jl Energy proportion of layer wavelet coefficients; Data features are obtained by concatenating time-domain features, frequency-domain features, and time-frequency-domain features.

[0029] By extracting multi-dimensional features from multi-source time-series data of station equipment to construct data-driven features, key data support can be provided to solve the problem that existing technologies struggle to achieve accurate extrapolation and prediction when historical fault data is lacking. For time-series data containing electrical, mechanical, environmental, and inspection data, the mean, variance, and peak factor calculated based on the absolute mean are obtained through statistical analysis. The spliced ​​time-domain features can characterize the basic state of equipment during normal operation from a data statistical perspective. Even without fault data, early potential degradation can be identified through subtle changes such as mean drift and variance increase. Fourier transform is performed on mechanical data to obtain Fourier transform signals. The main frequency, spectral energy, and spectral entropy calculated based on the energy ratio of each frequency band are further extracted. The spliced ​​frequency-domain features can uncover hidden frequency-domain degradation signals in mechanical data, capturing early anomalies of the mechanical system without relying on fault data. Wavelet transform is performed on mechanical data. The wavelet entropy calculated based on the number of decomposition layers and the energy ratio of wavelet coefficients in each layer is used as a time-frequency domain feature, which can accurately capture instantaneous changes in the signal in the time and frequency dimensions, further uncovering subtle degradation information that is easily overlooked in normal operation data. The complete data features formed by splicing time-domain, frequency-domain, and time-frequency-domain features cover the full-dimensional state characterization of multi-source time-series data, from statistical regularities and frequency distributions to transient changes. It constructs a rich benchmark feature library for normal equipment states. Even in the absence of historical fault data, the model can identify characteristic deviations in the equipment operation process based on this feature library, thereby accurately capturing the deterioration trend and providing reliable data-level support for extrapolation prediction. This effectively breaks through the limitations of existing technologies that rely on fault data to mine deterioration patterns and ensures the accuracy of extrapolation prediction.

[0030] Image features include visual features and infrared features; Methods for obtaining image features include: The device appearance image is used as input to the convolutional network to obtain convolutional features; The image of the equipment's appearance is converted to grayscale and subjected to Gaussian denoising to obtain a preprocessed image; The preprocessed image is used as input to the localization model to obtain the coordinates of the bounding box of the damaged area; the localization model is the YOLOv8n model. The preprocessed image is used as input to the U-Net model to obtain the probability of wear regions. Image regions with wear region probabilities higher than the probability threshold are segmented as ROI regions. The number of pixels in the damaged area and the number of pixels in the ROI area included in the bounding box coordinates of the damaged area are counted, and the ratio of the number of pixels in the damaged area to the number of pixels in the ROI area is calculated to obtain the pixel area ratio of the worn area. Visual features are obtained by concatenating convolutional features and the pixel area ratio of the worn area. Infrared images are converted based on the radiation calibration model of infrared thermal imagers to obtain pixel temperatures, such as pixel temperatures. temperature ,in, and These are the calibration coefficients for the infrared thermal imager; The digital quantization value of a pixel; The region is identified where the temperature of consecutive pixels is higher than a temperature threshold and the number of pixels with a temperature higher than the temperature threshold is greater than G. The number of pixels with a temperature higher than the temperature threshold and the total number of pixels in the infrared image are counted. The ratio of the number of pixels to the total number of pixels is calculated to obtain the proportion of hot areas. The highest temperature value among pixels whose temperature exceeds the temperature threshold is counted. Calculate the average temperature of pixels whose temperature is below the temperature threshold, and calculate the difference between the highest temperature value and the average temperature value to obtain the temperature gradient. Infrared features are obtained by stitching together the proportion of hot zones and the temperature gradient. If there is more than one hot zone, the largest proportion of hot zone and the largest temperature gradient are taken for stitching.

[0031] By extracting multi-dimensional features from visual and infrared images of station equipment, image features can be constructed, providing key visual and thermal degradation signal support to address the problem that existing technologies struggle to achieve accurate extrapolation and prediction when historical fault data is lacking. For visual images, the equipment appearance image is first input into a convolutional network to obtain convolutional features containing appearance details. After grayscale conversion and Gaussian denoising preprocessing, the YOLOv8n localization model accurately obtains the bounding box coordinates of the damaged area, and the U-Net model segments the ROI regions with a wear area probability higher than a threshold. By statistically analyzing the ratio of the number of pixels in the damaged area to the number of pixels in the ROI region, a quantified percentage of the wear area pixel area is obtained. The visual features formed by concatenating the convolutional features with this percentage can transform the qualitative degradation of the equipment appearance into a qualitative feature. Quantitative and traceable characteristic indicators can identify early appearance degradation even without historical fault data, by observing small increases in the wear area ratio from 0.5% to 2%, avoiding reliance on fault cases to determine the degree of degradation. For infrared images, the infrared thermal imager radiation calibration model converts the infrared image into temperature data, further filtering out hot areas with more than G consecutive high-temperature pixels. The ratio of hot area pixels to the total number of pixels in the infrared image is used to obtain the hot area ratio. At the same time, the difference between the highest temperature of the hot area and the average temperature of the non-hot area is calculated to obtain the temperature gradient. The infrared feature formed by stitching the two can capture the equipment's hidden thermal degradation. This type of degradation already has temperature anomalies before the fault occurs, and early thermal degradation can be identified by signals such as the expansion of the hot area ratio and the increase in the temperature gradient without the need for historical fault data. Visual and infrared features cover visible and latent thermal degradation of the equipment, respectively, transforming intuitive information from images into quantitative and comparable feature indicators. This constructs an image feature benchmark for normal equipment operation. When early degradation occurs, these features deviate from the benchmark. The prediction model can extrapolate the subsequent degradation trajectory based on the feature deviation trend, without relying on historical fault data to learn fault modes. This effectively overcomes the limitations of existing technologies that cannot capture early degradation or accurately extrapolate due to a lack of fault data.

[0032] Methods for obtaining physically derived features include: By calculating the dynamic change rate of the degradation process physical constraint equations in the physical constraint equations, the dynamic characteristics of degradation are extracted, namely wear rate, stiffness attenuation rate, and aging acceleration factor; the specific methods for obtaining these characteristics include: The wear rate is calculated based on the contact wear equation and wear time:

[0033] in, This represents the change in wear over 24 hours. Wear time; The stiffness attenuation rate is calculated based on the spring stiffness attenuation equation:

[0034] The aging acceleration factor was calculated based on the winding insulation aging index equation.

[0035] The wear rate, stiffness decay rate, and aging acceleration factor are sequentially concatenated into a 3D feature vector to obtain the physical derived features.

[0036] By deriving physically derived characteristics based on the degradation process equations of key physical quantities, this approach provides core physical mechanism-level dynamic support for addressing the difficulty of accurate extrapolation and prediction in the absence of historical fault data in existing technologies. For switchgear contacts, based on the established contact wear equation, the wear rate is obtained by calculating the ratio of wear amount change to wear time over 24 hours. This rate can directly quantify the rate of contact degradation caused by electrical and mechanical wear. Even without fault data, changes in the wear rate can accurately capture early signals of accelerated contact degradation, avoiding the inability to distinguish between normal fluctuations and degradation trends by relying solely on statistical characteristics. Limitations of potential: For circuit breaker springs, the stiffness attenuation rate is calculated based on the spring stiffness attenuation equation. This index can reflect the degree of stiffness loss caused by spring fatigue aging in real time. Without referring to historical fault cases, the change in stiffness attenuation rate can be used to determine whether the spring has entered the accelerated deterioration stage, ensuring that the extrapolation prediction conforms to the physical law of spring fatigue aging. For transformer windings, the aging acceleration factor is derived using the winding insulation aging index equation. This factor can quantify the accelerating effect of temperature on winding insulation aging. Even in the absence of insulation breakdown fault data, the future trend of insulation aging can be predicted by the change of the aging acceleration factor with increasing temperature. The physical derivative features formed by combining wear rate, stiffness decay rate, and aging acceleration factor essentially transform the degradation mechanism of core equipment components into quantifiable and traceable dynamic indicators. These indicators can reflect the true process and speed of degradation without relying on historical fault data, providing physical-level degradation trend anchors for prediction models. This avoids pure data-driven models deviating from the true degradation direction due to overfitting normal data when fault samples are scarce. At the same time, it ensures that the model's extrapolation prediction always follows the physical laws of equipment degradation, effectively filling the gap in degradation trend judgment when fault data is lacking. It provides indispensable physical mechanism support for accurately extrapolating the future degradation trajectory of equipment.

[0037] The hybrid features are used as input to the physical information neural network to obtain the degradation trajectory prediction results. The health of the degradation trajectory prediction results is evaluated to obtain a health score, and corresponding warnings are issued based on the health score.

[0038] The physical information neural network is constructed by a 2-layer LSTM, a Transformer encoder with 4 attention heads, and a spatiotemporal attention layer; A data fitting loss is established based on the error between the actual degradation index and the predicted value. A physical constraint loss is established based on equation constraints and boundary constraints. The total loss function is calculated based on the data fitting loss and the physical constraint loss. The Adam optimizer is used to optimize the physical information neural network, and the one with the minimum total loss function is used as the training output physical information neural network. For example, the data fitting loss is as follows: , The physical constraint loss is established based on the equation constraint loss and the boundary constraint loss, where the equation constraint loss is:

[0039] The boundary constraint loss is:

[0040] In summary, the physical constraint loss is:

[0041] The total loss function is:

[0042] in, To predict the step size; This represents the number of training samples; This is a true indicator of degradation. The results are for predicting the degradation trajectory; For the first training samples Predicted value of contact wear at any given moment; For the first training samples Predicted current value at any given time; For the first training samples Predicted contact resistance value at any time; For the first Predicted cumulative number of opening and closing times for each training sample at time t'; The balance coefficient can be obtained through historical data statistics.

[0043] Methods to obtain a health score include: The degradation trajectory prediction results are quantified using a piecewise function constructed based on the normal threshold and fault threshold of degradation indicators to obtain a health score; for example... Health score at any time ,in, The normal threshold for degradation indicators; The failure threshold is the degradation index. for The predicted degradation trajectory at any given time; When the health score is below the score threshold, a health warning is triggered; when the rate of change of the health score is above the rate of change threshold, a trend warning is triggered.

[0044] By inputting hybrid features into a physical information neural network constructed from a 2-layer LSTM, a Transformer encoder with four attention heads, and a spatiotemporal attention layer, and combining this with a specific loss function to optimize the model and output degradation trajectory prediction results, and then quantifying the health score and triggering corresponding warnings through a piecewise function, this approach effectively solves the problem of existing technologies struggling to achieve accurate extrapolation predictions when historical fault data is lacking. On one hand, this neural network architecture does not rely on historical fault data to learn fault feature patterns; it can accurately uncover potential patterns related to degradation in the data using only existing normal operation data or a small amount of slightly deteriorated data, leveraging the temporal processing capabilities and spatiotemporal attention focusing capabilities of LSTM and Transformer. On the other hand, in the loss function design, the data fitting loss ensures that the prediction results fit the existing monitoring data, while the physical constraint loss, constructed based on equation constraints and boundary constraints, forces the prediction results to conform to the equipment's specifications. The physical mechanism of degradation is addressed to avoid extrapolation biases that violate physical laws due to overfitting normal data when there is no fault data. The introduction of a balance coefficient further optimizes the fit between data fitting and physical laws through historical data statistics. At the same time, the health score is calculated by a piecewise function based on the normal threshold and fault threshold of the degradation index, which can achieve quantitative assessment without relying on health data from historical fault cases. Furthermore, health warnings triggered by health scores below the threshold and trend warnings triggered by score change rates above the threshold can also capture the dynamic trend of equipment degradation in a timely manner even in the absence of fault history. This ensures that the extrapolation prediction is not only accurate, but can also be transformed into health status assessment and warning signals with practical operation and maintenance guidance significance. Overall, it breaks through the limitations of existing technologies that cannot guarantee the rationality of extrapolation prediction and cannot effectively assess health status due to the lack of historical fault data, and realizes accurate prediction and operation and maintenance warning of equipment degradation trajectory in fault-free data scenarios.

[0045] The above description is only a preferred embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any equivalent substitutions or modifications made by those skilled in the art within the scope of the technology disclosed in the present invention, based on the technical solution and inventive concept of the present invention, should be covered within the scope of protection of the present invention.

Claims

1. A method for predicting station degradation trajectories by integrating prior physical knowledge with spatiotemporal knowledge, characterized in that, Includes the following steps: Collect and preprocess station equipment operation data to obtain multi-source heterogeneous data; Physical constraint equations for the degradation process of key physical quantities in station equipment are constructed, along with corresponding boundary constraint equations. The degradation process physical constraint equations and boundary constraint equations are then integrated to construct the physical constraint equations. A spatiotemporal knowledge graph containing a set of nodes, a set of relationships, and a set of attributes is constructed. A graph attention network is used as the graph embedding model to embed the attention weights of nodes to their neighbors, thereby obtaining a set of node embedding vectors. Feature extraction is performed on multi-source heterogeneous data to obtain data-driven features. Physical constraint equations are analyzed to obtain physical derived features, which are 3-dimensional feature vectors obtained by splicing wear rate, stiffness decay rate and aging acceleration factor. The nodes are standardized by Z-score by embedding the vector set, data-driven features, and physical derived features. The standardized nodes are then concatenated with the embedding vector set, data-driven features, and physical derived features to obtain hybrid features. The hybrid features are used as input to the physical information neural network to obtain the degradation trajectory prediction results. The health of the degradation trajectory prediction results is evaluated to obtain a health score, and corresponding warnings are issued based on the health score.

2. The method for predicting station degradation trajectories by integrating physical priors and spatiotemporal knowledge embedding as described in claim 1, characterized in that, The station equipment operation data includes electrical data, mechanical data, environmental data, visual images, and inspection data.

3. The method for predicting station degradation trajectories by integrating physical priors and spatiotemporal knowledge embedding as described in claim 1, characterized in that, The key physical quantities include switchgear contacts, circuit breaker springs, and transformer windings.

4. The method for predicting station degradation trajectories by integrating physical priors and spatiotemporal knowledge embedding as described in claim 3, is characterized in that... Methods for obtaining physical constraint equations include: By obtaining the cumulative number of opening and closing cycles and the contact resistance, and combining the electrical wear coefficient, mechanical wear coefficient, and current at the corresponding moment, an equation for contact wear is constructed: The wear of the contact at time t is: in, k w The coefficient of electrical wear; k m The mechanical wear coefficient; t 0 represents the initial time. for Current at any moment; for Constantly in contact with the resistor; The cumulative number of opening and closing operations at time t; By obtaining the initial stiffness and stress cycle number of the spring, and combining the attenuation coefficient and the fatigue life at the corresponding stress level, a spring stiffness attenuation equation is constructed: The equation for the decrease in spring stiffness at time t is: in, k sp0 This is the initial stiffness; This is the fatigue attenuation coefficient; n l For the first l Number of stress cycles per operation; N fl This refers to the fatigue life corresponding to the stress level. By obtaining the dielectric loss angle and winding temperature, and combining the temperature acceleration factor and reference temperature, an equation for the winding insulation aging index is constructed: The winding insulation aging index is: in, for Constant time interval loss angle; The acceleration factor is the temperature coefficient. for Constant winding temperature; Reference temperature; Construct boundary constraint equations, including: wear amount is non-negative; spring stiffness is not lower than the lower stiffness threshold; winding insulation aging index is not higher than the upper aging index threshold. The equations for contact wear, spring stiffness decay, winding insulation aging index, and boundary constraints are integrated into a set of equations to obtain the physical constraint equations.

5. The method for predicting station degradation trajectories by integrating physical priors and spatiotemporal knowledge embedding as described in claim 1, characterized in that, Methods for constructing spatiotemporal knowledge graphs that include sets of nodes, relationships, and attributes include: Obtain the station's coverage area, including spatial scope, equipment scope, and time scope; Entities are categorized into equipment unit entities, component entities, sensor entities, spatial region entities, and event / physical quantity entities; Define the relationships between entities, and analyze the domain, range, and type of the relationships; Define the attributes of an entity, including static attributes and dynamic attributes; A pre-defined hierarchical structure of station buildings, floors, bays, and installation sites is adopted to associate each spatial area entity with the superior area; Bind all entities within the station building to absolute coordinates based on the station building's coordinate system and relative coordinates based on the coordinate system of their respective spatial regions; Acquire station equipment testing data and automatically extract and convert it based on preset conversion tools; obtain data entities and map the data entities into a graph format according to the correspondence between entities and attributes to obtain a spatiotemporal knowledge graph.

6. The method for predicting station degradation trajectories by integrating physical priors and spatiotemporal knowledge embedding as described in claim 5, is characterized in that, Methods for obtaining node embedding vector sets include: Randomly assign an initial vector to each node; For nodes neighboring nodes By combining attention parameters, linear transformation matrices, and node embedding vectors, the node is calculated. neighboring nodes Attention weights: in, a For attention parameters; E It is a linear transformation matrix; N ( v ) is a node v The set of neighbors; This indicates vector concatenation; The slope is negative, with a value of 0.

2. h v For nodes The embedding vector; h u For nodes The embedding vector; For nodes The embedding vector; Nodes based on attention weights Update the embedding vector; obtain the node The updated embedding vector is: in, For activation functions; The updated embedding vectors of each node are concatenated to obtain the node embedding vector set.

7. The method for predicting station degradation trajectories by integrating physical priors and spatiotemporal knowledge embedding as described in claim 1, characterized in that, The data-driven features include data features and image features; Methods for obtaining data features include: Statistical analysis was performed on time-series data including electrical, mechanical, environmental, and inspection data to obtain the mean and variance. The absolute mean of the time-series data was then calculated, and the peak value was determined based on the absolute mean. Peak factor The mean, variance, and peak factor are concatenated to obtain the time-domain features; This represents the number of time-series data points. For the first One time series data; Perform a Fourier transform on the mechanical data to obtain the Fourier transform signal. Extract the dominant frequency, spectral energy, and spectral entropy from the Fourier transform signal. Concatenate the dominant frequency, spectral energy, and spectral entropy to obtain the frequency domain characteristics. The Fourier transform signal... main frequency Spectral energy Spectral entropy ,in, j The imaginary unit; f For frequency; t For time; It is the vibration acceleration; f max This represents the point with the maximum frequency range. For the first Energy percentage of each frequency band; Wavelet transform is performed on the mechanical data to calculate the wavelet entropy; the wavelet entropy is then used as a time-frequency domain feature; wavelet entropy ,in, The number of decomposition levels; q jl For the first jl Energy proportion of layer wavelet coefficients; Data features are obtained by concatenating time-domain features, frequency-domain features, and time-frequency-domain features.

8. The method for predicting station degradation trajectories by integrating physical priors and spatiotemporal knowledge embedding as described in claim 7, is characterized in that... The image features include visual features and infrared features; Methods for obtaining image features include: The device appearance image is used as input to the convolutional network to obtain convolutional features; The image of the equipment's appearance is converted to grayscale and subjected to Gaussian denoising to obtain a preprocessed image; The preprocessed image is used as input to the localization model to obtain the coordinates of the bounding box of the damaged area; the localization model is the YOLOv8n model. The preprocessed image is used as input to the U-Net model to obtain the probability of wear regions. Image regions with wear region probabilities higher than the probability threshold are segmented as ROI regions. The number of pixels in the damaged area and the number of pixels in the ROI area included in the bounding box coordinates of the damaged area are counted, and the ratio of the number of pixels in the damaged area to the number of pixels in the ROI area is calculated to obtain the pixel area ratio of the worn area. Visual features are obtained by concatenating convolutional features and the pixel area ratio of the worn area. The infrared image is converted based on the radiation calibration model of the infrared thermal imager to obtain the pixel temperature; The region is identified where the temperature of consecutive pixels is higher than a temperature threshold and the number of pixels with a temperature higher than the temperature threshold is greater than G. The number of pixels with a temperature higher than the temperature threshold and the total number of pixels in the infrared image are counted. The ratio of the number of pixels to the total number of pixels is calculated to obtain the proportion of hot areas. The highest temperature value among pixels whose temperature exceeds the temperature threshold is counted. Calculate the average temperature of pixels whose temperature is below the temperature threshold, and calculate the difference between the highest temperature value and the average temperature value to obtain the temperature gradient. Infrared features are obtained by stitching together the proportion of hot zones and the temperature gradient. If there is more than one hot zone, the largest proportion of hot zone and the largest temperature gradient are taken for stitching.

9. The method for predicting station degradation trajectories by integrating physical priors and spatiotemporal knowledge embedding as described in claim 1, characterized in that, The physical information neural network is constructed from a 2-layer LSTM, a Transformer encoder with 4 attention heads, and a spatiotemporal attention layer. A data fitting loss is established based on the error between the actual degradation index and the predicted value: The physical constraint loss is established based on the equation constraint loss and the boundary constraint loss, where the equation constraint loss is: The boundary constraint loss is: In summary, the physical constraint loss is: The total loss function is calculated based on the data fitting loss and the physical constraint loss: in, To predict the step size; This represents the number of training samples; This is a true indicator of degradation. The results are for predicting the degradation trajectory; For the first training samples Predicted value of contact wear at any given moment; For the first training samples Predicted current value at any given time; For the first training samples Predicted value of contact resistance at any time; For the first Predicted cumulative number of opening and closing times for each training sample at time t'; The balancing coefficient is used; the physical information neural network with the minimum total loss function is obtained by optimizing based on the Adam optimizer and used as the training output physical information neural network. Methods to obtain a health score include: The degradation trajectory prediction results are quantified based on a piecewise function constructed according to the normal threshold and the failure threshold of the degradation index to obtain a health score: The health score at time t'' is: ， in, y norm The normal threshold for degradation indicators; y fault The failure threshold is the degradation indicator. for Degradation trajectory prediction results at any given time; When the health score is below the score threshold, a health warning is triggered; when the rate of change of the health score is above the rate of change threshold, a trend warning is triggered.

Images