Intelligent switchgear health assessment method and system for failure prediction

Abstract

The application discloses a fault prediction-oriented intelligent complete switch device health assessment method and system, relates to the technical field of power equipment state monitoring, and deploys multi-modal sensors on key nodes of high and low voltage complete switch devices to synchronously collect full-dimensional operation parameters; a multi-source heterogeneous health feature extraction model is constructed to automatically extract deep health features representing structural looseness, insulation deterioration, poor contact and component aging; a double kernel t distribution random neighborhood embedding algorithm is used for robust dimension reduction and generation of a visual distribution map; low-dimensional features are input into a large language model optimized by power equipment operation and maintenance knowledge to generate a structured health assessment report and probabilistically predict the remaining life of key components; deep features, visual results and large language model analysis conclusions are fused to generate a dynamic health index bound to physical equipment and output graded warning information. The application realizes accurate health assessment and fault prediction of complete switch devices.

Description

Technical Field

[0001] This invention relates to the field of power equipment condition monitoring technology, and more specifically, to a health assessment method and system for intelligent switchgear assemblies oriented towards fault prediction. Background Technology

[0002] High and low voltage switchgear assemblies are core equipment in power systems, responsible for power distribution, control, and protection. Their safe and stable operation directly affects the reliability of the entire power supply system. Switchgear assemblies typically consist of a cabinet structure, primary circuits, secondary circuits, and various components (such as circuit breakers, contactors, and transformers). Their complex structure and close interrelationships mean that any abnormality in any component can trigger a cascading failure, leading to power outages or even equipment damage. Currently, condition monitoring and health assessment of switchgear assemblies suffer from the following shortcomings: First, traditional periodic preventative maintenance relies on fixed-cycle inspections and tests, resulting in significant lag and difficulty in capturing gradual changes in equipment condition. It is also prone to over-maintenance or under-maintenance, leading to high maintenance costs and limited benefits. Second, existing monitoring systems are often limited to monitoring single physical quantities, such as electrical parameters, temperature, or vibration. This fails to comprehensively reflect the complexity of equipment operating conditions, and is particularly difficult to accurately capture early fault characteristics caused by the coupling of multiple factors. Third, traditional data analysis methods are insufficient for processing high-dimensional, multi-source, and heterogeneous equipment operation data. They struggle to automatically uncover deep-seated fault characteristics, have limited ability to analyze the correlation and evolution trends between faults, and cannot achieve accurate quantification and prediction of health status. Fourth, existing assessment systems primarily output numerical indicators or alarm signals, lacking intuitive and interpretable results. This makes it difficult for maintenance personnel to quickly understand equipment status, risk locations, and fault causes, reducing decision-making efficiency.

[0003] With the increasing demands for power supply reliability in power systems and the deepening of smart grid construction, traditional maintenance models and monitoring methods are no longer sufficient to meet actual needs. How to comprehensively utilize multimodal sensing data, combined with advanced deep learning and artificial intelligence technologies, to achieve accurate health assessment and fault prediction of the overall architecture, primary / secondary lines, and internal components of complete switchgear has become a pressing technical problem to be solved in this field. Summary of the Invention

[0004] To address the aforementioned technical problems, this invention proposes a health assessment method and system for intelligent switchgear assemblies oriented towards fault prediction. This effectively solves the problems of delayed response, single assessment dimensions, inaccurate fault diagnosis, and lack of interpretability of assessment results in traditional maintenance modes, providing a complete technical solution for intelligent operation and maintenance and predictive maintenance of high and low voltage switchgear assemblies.

[0005] The first aspect of this invention provides a health assessment method for intelligent switchgear assemblies oriented towards fault prediction, comprising the following steps: Multimodal sensors are deployed at key nodes of high and low voltage switchgear to synchronously collect full-dimensional parameters reflecting the operating status of the equipment. These full-dimensional parameters include mechanical status parameters of the cabinet structure, electrical and thermal status parameters of the primary main circuit, status parameters of the secondary control circuit, and operating condition parameters of internal key components. Based on the full-dimensional parameters, a multi-source heterogeneous health feature extraction model based on deep learning is constructed. An adaptive neural network structure is adopted for different data types to automatically extract deep health features that can characterize structural loosening, insulation degradation, poor contact and component aging. For the deep health features, a robust dimensionality reduction process is performed using a dual-kernel t-distribution random neighborhood embedding algorithm, and the features are mapped to a low-dimensional space to generate a visual distribution map. The low-dimensional features and visualization results are input into a large language model optimized by power equipment operation and maintenance knowledge. The reduced-dimensional feature map and the original multimodal data are parsed to generate a structured natural language health assessment report for a specific cabinet, circuit or component, and to make probabilistic predictions of the remaining effective life of key components. By integrating the deep health features, visualization results, and large language model analysis conclusions, a dynamic health index bound to specific physical devices is generated, and graded early warning information is output according to preset thresholds.

[0006] In this solution, the full-dimensional parameter acquisition specifically includes: Based on the physical structure of high and low voltage switchgear, a multi-layer topology diagram is constructed, including the cabinet architecture layer, primary circuit layer, secondary circuit layer and component layer. The multi-layer topological graph is pre-trained using a graph neural network to learn the state propagation rules and fault impact weights of each node during the evolution of equipment health status, and outputs the importance score of each node. By combining the node importance score output by the graph neural network with the information entropy maximization criterion, a heuristic search algorithm is used to select the sensor deployment node combination that maximizes the overall information gain and determine the location of key nodes. Based on the selected key node locations, deploy corresponding types of multimodal sensors, synchronize the time of each type of sensor through a synchronization triggering device, and synchronously collect parameters in all dimensions. The collected multimodal data is then cleaned, timestamp aligned and standardized in sequence to generate a full-dimensional parameter set.

[0007] This solution extracts deep health characteristics that characterize structural loosening, insulation degradation, poor contact, and component aging, specifically including: The collected full-dimensional parameters are divided into three heterogeneous data types: time-series signal data, image and spectral data, and state parameter data. Based on the heterogeneous data types, a multi-source heterogeneous health feature extraction model with three parallel feature extraction branches is constructed. In the multi-source heterogeneous health feature extraction model, the first branch is a one-dimensional time-series feature extraction network, which adopts a hybrid architecture of a one-dimensional convolutional neural network with residual connections and a bidirectional long short-term memory network to extract local waveform features and sequential dependencies in the time-series signal and generate deep time-series features. The second branch is a two-dimensional image feature extraction network. It adopts an improved residual network architecture and extracts temperature distribution patterns, hot spot features and thermal gradient changes in infrared thermal images, as well as phase distribution features, discharge amplitude patterns and spectrum shape features in partial discharge spectra through multi-layer convolution modules and attention mechanism modules, to generate deep image features. The third branch is a fully connected feature extraction network, which uses a multi-layer fully connected neural network to perform nonlinear transformation and feature combination on the state parameter data, learn the coupling relationship between ambient temperature, load rate, runtime, and number of switching actions, and generate deep state features. The deep temporal features, deep image features, and deep state features are input into the feature fusion module, and a fused multimodal feature vector is generated through adaptive weighting. The fused multimodal feature vector is input into a fully connected layer for dimensional transformation and nonlinear mapping, and the deep health features representing the health status of the device are output.

[0008] In this scheme, a dual-core t-distribution random neighborhood embedding algorithm is used for robust dimensionality reduction and mapping to a low-dimensional space, specifically including: For the aforementioned deep health features, a dual-kernel similarity matrix is ​​constructed that integrates a local linear kernel and a global Gaussian kernel. The local linear kernel calculates the reconstruction weight coefficient based on the linear reconstruction relationship within the neighborhood of the sample point to measure the local geometric structure on the high-dimensional manifold. The global Gaussian kernel calculates the Euclidean distance similarity between sample points using the traditional Gaussian kernel form to measure the overall distribution characteristics of the data. The local linear kernel and the global Gaussian kernel are then weighted and fused using an adaptive fusion coefficient to construct the dual-kernel similarity matrix. Based on the dual-kernel similarity matrix, the similarity relationship in the high-dimensional space is converted into a conditional probability distribution, and a joint probability distribution in the high-dimensional space is generated by symmetry. At the same time, the t-distribution is used as the similarity metric kernel function in the low-dimensional embedding space to define the joint probability distribution in the low-dimensional space. Using KL divergence as the objective function, the difference between the joint probability distribution in high-dimensional space and the joint probability distribution in low-dimensional space is measured. The objective function is optimized and solved to obtain the low-dimensional embedding coordinates corresponding to each high-dimensional deep health feature sample.

[0009] In this solution, generating a visual distribution map specifically includes: The low-dimensional embedded coordinates are plotted in three-dimensional space to generate a visual distribution map. In the visual distribution map, samples with the same health status are clustered to form clusters, samples with different health statuses are separated in space, and the same fault type shows a continuous evolution trajectory as the degree of deterioration deepens. The effectiveness of the dimensionality reduction results is verified by K-nearest neighbor classification accuracy, silhouette coefficient, and distance correlation index. When the index meets the preset threshold, the low-dimensional features and visualization results are output. If the threshold is not met, the adaptive fusion coefficient, neighborhood size, and global bandwidth parameters are adjusted, and the dimensionality reduction is recalculated.

[0010] In this solution, the generation of the structured natural language health assessment report specifically includes: Based on low-dimensional features and visualization results, original full-dimensional parameters, and deep health features, a comprehensive input context for a large language model is constructed. The thinking chain reasoning mechanism guides the large language model to reason step by step according to the logical chain of state recognition, risk positioning, and causal analysis; In the state identification phase, based on the input multimodal feature description and visualization results, the current health status category of the equipment is identified and judgment criteria are generated. In the risk localization phase, combining the equipment topology and fault propagation patterns, the specific location and impact range of the fault are inferred, and the assessment conclusions are associated with specific cabinets, circuits, or components. In the causal analysis phase, fault mechanism knowledge from the domain knowledge base is used to analyze the possible causes and evolution mechanisms of the fault. Based on the reasoning analysis results, the large language model generates a structured natural language health assessment report, which includes basic equipment information, comprehensive health index, sub-item status assessment, risk location, fault cause analysis, and maintenance recommendations.

[0011] In this scheme, the probabilistic prediction of the remaining effective lifespan of key components specifically includes: We acquire degradation-sensitive features, spatiotemporal evolution features of low-dimensional embedded trajectories, and key component degradation models based on physical mechanisms from deep health features, and fuse them to construct a multi-source lifetime feature set. The multi-source lifetime feature set, equipment historical operating information, current operating conditions and key parameters, and degradation theoretical values ​​calculated based on the physical model are input into the large language model for conditional information encoding to generate conditional embedding vectors. The conditional embedding vector is used as the input diffusion probability model. The diffusion probability model adopts a noise prediction network based on the Transformer architecture. It dynamically focuses on the conditional information most relevant to the current prediction task through a cross-attention mechanism. Starting from standard Gaussian noise, multiple remaining lifetime samples are generated through an iterative denoising process to form the posterior distribution of the remaining lifetime. During the training of the diffusion probability model, a physical information neural network constraint mechanism is introduced to construct a physical consistency loss function to measure the difference between the remaining lifetime predicted by the diffusion probability model and the prediction results of the physical information neural network. Through weighted joint optimization, the prediction results are made to conform to physical laws. Perform statistical analysis on the generated remaining lifespan samples and output probabilistic prediction results.

[0012] This solution integrates the aforementioned deep health features, visualization results, and large language model analysis conclusions to generate a dynamic health index bound to a specific physical device, specifically including: Obtain a multi-layer topology graph containing a cabinet architecture layer, a primary loop layer, a secondary loop layer, and a component layer. Based on the multi-layer topology graph, construct a spatiotemporal graph neural network as a health state propagation model. Key nodes are used to determine the state propagation rules and fault impact weights learned by the pre-training of the central graph neural network, which are used as the initialization parameters of the spatiotemporal graph neural network. In the spatiotemporal graph neural network, the state information of neighboring nodes is aggregated through graph convolution operations to simulate the propagation process of fault or degradation state inside the device. A gated recurrent unit network is used to model the state evolution of each node on a continuous time cross-section. The sequence of deep health features of each node at multiple time points is input into the spatiotemporal graph neural network. The node state prediction loss and the overall health index prediction loss are optimized simultaneously through a multi-task learning strategy, and the predicted health status values ​​of each node at the current time and future time are output. A spatiotemporal attention mechanism is introduced to adaptively weight the node health status output by the spatiotemporal graph neural network, and calculate the overall health index of the device.

[0013] In this scheme, tiered early warning information is output based on preset thresholds, specifically including... The Monte Carlo dropout method is used to predict the future evolution trajectory of the health index and quantify the uncertainty. Short-term, medium-term and long-term predictions of the health index change trends and confidence intervals are generated, and the probability distribution of the health index falling to each warning threshold is calculated. Based on the calculated health index and its evolution trend, combined with risk location information and remaining life expectancy prediction results, a graded early warning information is generated. By integrating health indices, tiered early warning information, and health assessment reports, a complete equipment health assessment file is generated and archived, serving as feedback data for continuous learning mechanisms in model iteration and optimization.

[0014] The second aspect of this invention proposes a project resource recommendation system based on the fusion of knowledge graph subgraph partitioning and representation. The system includes: a multimodal data acquisition module, a deep health feature extraction module, a high-dimensional feature dimensionality reduction and visualization module, an intelligent analysis and evaluation module, and a health index calculation and early warning module. The multimodal data acquisition module is responsible for deploying multimodal sensors at key nodes of high and low voltage switchgear to synchronously collect full-dimensional parameters reflecting the operating status of the equipment. The deep health feature extraction module is responsible for constructing a multi-source heterogeneous health feature extraction model based on deep learning. It adopts an appropriate neural network structure for different data types and automatically extracts deep health features that can characterize structural loosening, insulation degradation, poor contact, and component aging. The high-dimensional feature dimensionality reduction and visualization module is responsible for using a dual-core t-distribution random neighborhood embedding algorithm to robustly reduce the dimensionality of high-dimensional deep health features, mapping them to a low-dimensional space and generating a visualization distribution map. The intelligent analysis and evaluation module is responsible for integrating a large language model optimized with power equipment operation and maintenance knowledge, parsing the dimensionality-reduced feature map and the original multimodal data, generating a structured natural language health assessment report, and making probabilistic predictions on the remaining effective lifespan of key components. The health index calculation and early warning module integrates deep health features, visualization results, and large language model analysis conclusions to generate a dynamic health index that is bound to a specific physical device, and outputs graded early warning information according to preset thresholds.

[0015] Compared with the prior art, the beneficial effects of the present invention are as follows: This invention utilizes multimodal sensors to synchronously collect full-dimensional parameters of equipment and combines this with a multi-branch deep neural network to automatically extract deep health features, achieving accurate characterization of potential defects such as structural loosening, insulation degradation, and poor contact. A dual-core t-distribution random neighborhood embedding algorithm is employed for dimensionality reduction visualization, intuitively presenting equipment state clustering and evolution trajectories. A domain-knowledge-optimized large language model is introduced to generate a structured natural language health assessment report, and a diffusion probability model guided by the large language model is integrated with a physical information neural network to achieve probabilistic prediction of remaining lifespan. A dynamic health index is constructed based on a spatiotemporal graph neural network and an attention mechanism, outputting tiered early warning information. This invention effectively solves the problems of delayed response, single assessment dimensions, and lack of interpretability in traditional maintenance methods, providing complete technical support for predictive maintenance of complete switchgear. Attached Figure Description

[0016] To more clearly illustrate the technical solutions in the embodiments or examples of the present invention, the drawings used in the embodiments or examples will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained according to these drawings without creative effort.

[0017] Figure 1 A flowchart of a health assessment method for intelligent switchgear assemblies oriented towards fault prediction is shown. Figure 2 The flowchart for extracting deep health features is shown; Figure 3 A flowchart is shown for probabilistically predicting the remaining effective life of critical components; Figure 4 A block diagram of a health assessment system for intelligent switchgear assemblies designed for fault prediction is shown. Detailed Implementation

[0018] To better understand the above-mentioned objectives, features, and advantages of the present invention, the present invention will be further described in detail below with reference to the accompanying drawings and specific embodiments. It should be noted that, unless otherwise specified, the embodiments and features described in these embodiments can be combined with each other.

[0019] Many specific details are set forth in the following description in order to provide a full understanding of the invention. However, the invention may also be practiced in other ways different from those described herein, and therefore the scope of protection of the invention is not limited to the specific embodiments disclosed below.

[0020] like Figure 1 As shown, this embodiment provides a health assessment method for intelligent switchgear assemblies oriented towards fault prediction, including... Multimodal sensors are deployed at key nodes of high and low voltage switchgear to synchronously collect full-dimensional parameters reflecting the operating status of the equipment. These full-dimensional parameters include mechanical status parameters of the cabinet structure, electrical and thermal status parameters of the primary main circuit, status parameters of the secondary control circuit, and operating condition parameters of internal key components. Based on the full-dimensional parameters, a multi-source heterogeneous health feature extraction model based on deep learning is constructed. An adaptive neural network structure is adopted for different data types to automatically extract deep health features that can characterize structural loosening, insulation degradation, poor contact and component aging. For the deep health features, a robust dimensionality reduction process is performed using a dual-kernel t-distribution random neighborhood embedding algorithm, and the features are mapped to a low-dimensional space to generate a visual distribution map. The low-dimensional features and visualization results are input into a large language model optimized by power equipment operation and maintenance knowledge. The reduced-dimensional feature map and the original multimodal data are parsed to generate a structured natural language health assessment report for a specific cabinet, circuit or component, and to make probabilistic predictions of the remaining effective life of key components. By integrating the deep health features, visualization results, and large language model analysis conclusions, a dynamic health index bound to specific physical devices is generated, and graded early warning information is output according to preset thresholds.

[0021] It should be noted that, based on the physical structure of high and low voltage switchgear, a multi-layer topology diagram is constructed, including a cabinet architecture layer, a primary circuit layer, a secondary circuit layer, and a component layer. The cabinet architecture layer reflects the structural support and mechanical connection relationships of the equipment; the primary circuit layer represents the energy transmission path and key connection points of the main circuit; the secondary circuit layer reflects the logical connections of control, protection, and signal circuits; and the component layer records the functional attributes and failure relationships of each core component (circuit breaker, contactor, transformer, etc.). Historical operating data and fault case data are collected to construct a sample set of equipment state evolution. A graph neural network is used to pre-train the multi-layer topology diagram, learning the state propagation laws and fault impact weights of each node during the equipment health state evolution process, and outputting the importance score of each node. Combining the node importance score output by the graph neural network with the information entropy maximization criterion, using the node importance score as the basic weight, and under the constraint of a given number of sensors, a heuristic search algorithm is used to select the sensor deployment node combination that maximizes the overall information entropy, ensuring that the selected key nodes can reflect the overall health state of the equipment to the greatest extent, and determining the critical nodes. Key node locations; ultimately, the key nodes for the primary circuit determined by the key node selection method include: the incoming connection point, the connection between the busbar and the branch busbar, the incoming and outgoing terminals of each feeder circuit breaker, and the installation location of the current transformer; the key nodes for the secondary circuit determined include: the power input terminal of the control circuit, the signal acquisition terminal of the protection device, the communication interface of the intelligent control unit, and the location of the key status feedback contact; the key nodes for the cabinet structure determined include: the connection point of the cabinet support structure, the location of the door hinge, and the installation point of the insulating support component; the key nodes for the components determined include: the circuit breaker operating mechanism, the contactor contact assembly, the capacitor terminal block, and the surge arrester grounding terminal.

[0022] Based on the selected key node locations, deploy corresponding types of multimodal sensors, including vibration sensors, temperature sensors, current / voltage sensors, partial discharge sensors, infrared thermal imaging sensors, sound sensors, etc. A synchronization triggering device is used to synchronize the time of each type of sensor, enabling synchronous acquisition of all-dimensional parameters. The acquired multimodal data is then sequentially cleaned, timestamp aligned, and standardized to generate a full-dimensional parameter set.

[0023] It should be noted that, as Figure 2 As shown, the collected full-dimensional parameters are divided into three heterogeneous data types: time-series signal data (three-phase current / voltage waveforms, vibration time-domain signals, partial discharge pulse sequences, sound signals, etc.), image and spectrum data (infrared thermal images, partial discharge phase decomposition maps, visible light images, etc.), and state parameter data (temperature values, humidity values, switch status, control signal levels, equipment running time, etc.). Based on the heterogeneous data types, a multi-source heterogeneous health feature extraction model with three parallel feature extraction branches is constructed.

[0024] In the multi-source heterogeneous health feature extraction model, the first branch is a one-dimensional time-series feature extraction network, which adopts a hybrid architecture of a one-dimensional convolutional neural network with residual connections and a bidirectional long short-term memory network to extract local waveform features and dependencies in time-series signals, generating deep time-series features characterizing mechanical wear, electrical aging, and insulation degradation. The second branch is a two-dimensional image feature extraction network, which adopts an improved residual network architecture and extracts temperature distribution patterns, hot spot features, and thermal gradient changes in infrared thermal images, as well as phase distribution features, discharge amplitude patterns, and spectral shape features in partial discharge spectra through multi-layer convolutional modules and attention mechanism modules, generating deep image features characterizing poor contact, insulation defects, and local overheating. The third branch is a fully connected feature extraction network, which adopts a multi-layer fully connected neural network to perform nonlinear transformations and feature combinations on state parameter data, learn the coupling relationship between ambient temperature, load rate, running time, and number of switching actions, and generate deep state features characterizing equipment operating conditions and aging degree.

[0025] The deep temporal features, deep image features, and deep state features are input into a feature fusion module. This module employs a multi-head attention mechanism, adaptively allocating the contribution of different modal features in the fusion process by calculating the correlation weights between features in each branch, generating a fused multimodal feature vector. This fused multimodal feature vector is then input into a fully connected layer for dimensionality transformation and nonlinear mapping, outputting deep health features representing the equipment's health status. The deep health feature vector implicitly contains coupled information about multiple fault modes, such as structural loosening, insulation degradation, poor contact, and component aging. The multi-source heterogeneous health feature extraction model is trained using a combination of unsupervised pre-training and supervised fine-tuning. Unsupervised pre-training of each branch network is performed using unlabeled multimodal historical operating data through an autoencoder structure. Supervised fine-tuning is then performed based on health-labeled sample data, using a combination of cross-entropy loss function and mean squared error loss function. A contrastive learning strategy is introduced to enhance feature discrimination capabilities.

[0026] It should be noted that deep health feature vectors are typically high-dimensional, with complex nonlinear coupling relationships between their dimensions. While high-dimensional features contain rich equipment status information, directly using them for state clustering, fault mode recognition, and visualization analysis presents challenges such as difficulty in effectively distinguishing different health states, hindering maintenance personnel's understanding of equipment status distribution and evolution patterns, and failing to reveal the hidden nonlinear manifold structure within the high-dimensional features. To address this, a dual-kernel t-distribution random neighborhood embedding algorithm is introduced. This algorithm preserves both the local nearest neighbor relationships and the global topological structure of high-dimensional features during dimensionality reduction, enabling low-dimensional visualization of complex high-dimensional health features. For the deep health features, a dual-kernel similarity matrix fusing local linear kernels and a global Gaussian kernel is constructed. The local linear kernel calculates reconstruction weight coefficients based on the linear reconstruction relationships within the neighborhood of sample points, measuring the local geometric structure on the high-dimensional manifold. For each sample point, K nearest neighbor sample points are found within the neighborhood, and the reconstruction weight coefficients are calculated by minimizing the linear reconstruction error. ,in Represents sample points The set of K nearest neighbors, For sample points From its neighborhood samples Contribution weights during reconstruction. Reconstruction weight coefficients reflect the local linear relationship between a sample point and its neighboring samples, effectively characterizing the local geometric structure on a high-dimensional manifold. Local linear kernel similarity is defined as a symmetric similarity based on reconstruction weights. A high similarity value is achieved when two sample points are neighbors and have large reconstruction weights, denoted as... The global Gaussian kernel The sample points are calculated using the traditional Gaussian kernel method. The Euclidean distance similarity between sample points is calculated by integrating the global bandwidth parameter. Scaling measures the overall distribution characteristics of the data, expressed as: The local linear kernel and the global Gaussian kernel are weighted and fused using adaptive fusion coefficients to construct a dual-kernel similarity matrix. ; , For adaptive fusion coefficients.

[0027] Based on the aforementioned dual-core similarity matrix, the similarity relationships in the high-dimensional space are transformed into conditional probability distributions, and then symmetrically generated to produce a joint probability distribution in the high-dimensional space. ,in, Let be the conditional probability, representing the probability of a sample point in a high-dimensional space. From the perspective The probability of being selected as a nearest neighbor for a sample point in a high-dimensional space. From the perspective The probability of being selected as a nearest neighbor is used, and a t-distribution is employed as the similarity metric kernel function in the low-dimensional embedding space to define the joint probability distribution in the low-dimensional space. , ,in For the first in the low-dimensional embedding space The sample point, the first The sample point, the first The sample point, the first One sample point, , , , Let $\mathbf{a}$ be the sample index variable. The KL divergence is used as the objective function to measure the difference between the joint probability distribution in the high-dimensional space and the joint probability distribution in the low-dimensional space, ensuring that similar sample points in the low-dimensional embedding space are as close as possible and dissimilar sample points are as far apart as possible. This achieves a structure-preserving mapping from the high-dimensional feature space to the low-dimensional embedding space. The KL objective function is expressed as: ,in KL divergence is used to measure high-dimensional distributions. With low-dimensional distribution The asymmetric differences between them. The gradient descent method is used to optimize the objective function. The gradient of the objective function with respect to the low-dimensional embedded coordinates is calculated iteratively, and the coordinates are updated along the gradient descent direction. The iteration is repeated until the objective function converges or the preset number of iterations is reached. After optimization convergence, the low-dimensional embedded coordinates corresponding to each high-dimensional deep health feature sample are obtained.

[0028] The low-dimensional embedded coordinates are plotted in two-dimensional or three-dimensional space to generate a visual distribution map. In this map, samples with the same health status cluster together, while samples with different health statuses (such as loose structures, insulation degradation, poor contact, and component aging) are separated in space. The same fault type exhibits a continuous evolution trajectory as the degree of degradation deepens. To ensure the dimensionality reduction effect of the dual-core t-distribution random neighborhood embedding algorithm, quantitative evaluation metrics are introduced to verify the dimensionality reduction results. The K-nearest neighbor classification accuracy is used to evaluate the ability of the dimensionality-reduced features to retain the original category information. By inputting the low-dimensional embedded features into the classifier, the classification accuracy for known health status labels is calculated. The silhouette coefficient is used to evaluate the compactness and separation of the cluster structure. The clustering effect is quantified by calculating the ratio of the average distance between a sample and samples of the same class to the average distance between a sample and samples of different classes. The distance correlation is used to measure the correlation of the distance ranks between samples before and after dimensionality reduction, evaluating the degree of preservation of the global topology. When the metrics meet the preset threshold, the low-dimensional features and visualization results are output. If the threshold is not met, the adaptive fusion coefficient, neighborhood size, and global bandwidth parameters are adjusted, and the dimensionality reduction calculation is performed again.

[0029] It should be noted that a remaining effective life prediction method based on a diffusion probability model guided by a large language model and a physical information neural network is introduced to achieve a probabilistic and quantitative assessment of the remaining life of key components. A knowledge base for the operation and maintenance of power equipment is constructed. This knowledge base covers multi-source heterogeneous text materials, including technical specifications, operating procedures, fault case libraries, maintenance records, expert experience documents, and equipment manufacturer technical manuals for high and low voltage switchgear. Through domain entity recognition and relation extraction techniques, structured knowledge units such as equipment structure knowledge, fault mode knowledge, diagnostic rule knowledge, and maintenance strategy knowledge are extracted from the above materials to form a domain knowledge graph. Secondly, a domain-adaptive pre-training strategy is used to incrementally train the large language model. Based on pre-training with a general corpus, the constructed domain knowledge base is used for continuous pre-training of the model. Instruction fine-tuning technology is introduced to improve the model's task execution capability. Dedicated instruction templates for health assessment tasks are designed, including status parsing instructions, fault diagnosis instructions, risk location instructions, and life prediction instructions.

[0030] Based on low-dimensional features and visualization results, original full-dimensional parameters, and deep health features, a comprehensive input context for a large language model is constructed. Among them, the visualized distribution map is converted into structured descriptive text through image analysis algorithms, the deep health features are converted into key indicator descriptions through feature importance analysis, and the original multimodal data is used to extract abnormal events exceeding the normal range from time series signals and image data through threshold detection and pattern recognition to generate time series descriptions.

[0031] A thought chain reasoning mechanism guides the large language model to reason step by step according to the logical chain of state identification, risk location, and causal analysis. In the state identification stage, based on the multimodal feature description and visualization results of the input, the current health status category of the equipment is identified and judgment criteria are generated. In the risk location stage, combined with the equipment topology and fault propagation law, the specific location and scope of the fault are inferred, and the assessment conclusions are associated with specific cabinets, circuits, or components. In the causal analysis stage, the fault mechanism knowledge in the domain knowledge base is used to analyze the possible causes and evolution mechanisms of the fault. Based on the reasoning and analysis results, the large language model generates a structured natural language health assessment report, which includes basic equipment information, comprehensive health index, sub-item state assessment, risk location, fault cause analysis, and maintenance recommendations. The sub-item status assessment describes the status and risk level of the cabinet structure's mechanical status, primary circuit electrical and thermal status, secondary circuit control status, and key component operating conditions. Risk location uses a combination of text descriptions and cabinet structure diagrams to visually present the specific location of the fault point in the equipment. Maintenance recommendations are based on the fault type and severity, providing graded maintenance recommendations and specific maintenance content, including immediate shutdown for repair, scheduled maintenance in the near future, and continued monitoring and observation.

[0032] The system acquires degradation-sensitive features from deep health characteristics. Through feature evolution trajectory analysis, it identifies feature dimensions that exhibit monotonically increasing or decreasing trends with equipment operating time, such as contact resistance change rate, vibration amplitude trend, and cumulative partial discharge intensity. It also acquires spatiotemporal evolution features of low-dimensional embedded trajectories, connecting the low-dimensional embedded coordinates of various time segments during the equipment's historical operation to form state evolution trajectories. From these trajectories, it extracts spatiotemporal features such as evolution velocity, acceleration, trajectory curvature, and outlier degree, reflecting the dynamic migration patterns of the equipment in the health state space. Furthermore, it acquires degradation models for key components based on physical mechanisms. For different types of key components (circuit breaker operating mechanisms, contactor contacts, insulating supports, capacitors, etc.), it constructs degradation models based on physical mechanisms. For example, for mechanical components, a stress-life model is established based on fatigue cumulative damage theory; for electrical components, a thermal aging model is established based on the Arrhenius equation; and for insulating components, a partial discharge cumulative damage model is established based on electrochemical theory. Finally, it fuses the acquired features to construct a multi-source lifetime feature set.

[0033] The diffusion probability model achieves its ability to generate samples from a noisy distribution by progressively adding noise to the data until it is completely randomized, and then learning an inverse denoising process. In the remaining lifetime prediction task, the model learns the conditional generation process from current health state features to future lifetime distribution. In the forward pass of the diffusion model, Gaussian noise is progressively added to the true remaining lifetime labels of key components to construct intermediate states with increasing noise. In the inverse pass of the diffusion model, a noise prediction network based on the Transformer architecture is trained to learn the process of progressively recovering the true remaining lifetime from the noisy state.

[0034] like Figure 3As shown, the multi-source lifetime feature set, historical equipment operating information, current operating conditions and key parameters, and degradation theoretical values ​​calculated based on the physical model are input into a large language model for conditional information encoding to generate conditional embedding vectors. These conditional embedding vectors are then used as input to a diffusion probability model. This diffusion probability model employs a noise prediction network based on a Transformer architecture, dynamically focusing on the conditional information most relevant to the current prediction task through a cross-attention mechanism. Starting from standard Gaussian noise, it iterative denoising processes generate multiple remaining lifetime samples, constituting the posterior distribution of the remaining lifetime. During the training of the diffusion probability model, a physical information neural network constraint mechanism is introduced. The degradation theoretical values ​​calculated by the physical model are used as soft constraints to construct a physical consistency loss function. For each key component type, a lightweight physical information neural network is constructed to approximate the degradation differential equation of that component. The Physical Information Neural Network (PIN) takes runtime, number of operations, load history, and environmental parameters as inputs and outputs a degradation state prediction based on physical mechanisms. It measures the difference between the remaining lifetime predicted by the diffusion probability model and the prediction result of the PIN. Through weighted joint optimization, the prediction result is made to conform to physical laws, effectively solving the problem of unstable prediction in small sample scenarios.

[0035] During the prediction phase, the large language model transforms the original multi-source feature data into semantically rich conditional embeddings, enabling the diffusion model to understand the meaning of the input information. Probability distribution samples generated by the diffusion model are input into the large language model, which performs statistical analysis and converts them into natural language descriptions. The large language model, combined with its learned knowledge of the power equipment domain, intelligently corrects the initial prediction results of the diffusion model. Statistical analysis is performed on the generated remaining lifetime samples, outputting probabilistic prediction results. These probabilistic prediction results include probability density distribution curves, median remaining lifetime and quantile indicators, and prediction interpretability descriptions. The prediction interpretability descriptions are generated by the large language model, explaining the key factors affecting the prediction results and their mechanisms of action. When actual faults occur or maintenance verification results are fed back, new sample data is added to the training set, and the diffusion probability model and physical information neural network are periodically incrementally updated. Feedback learning optimizes the conditional encoding and prediction correction capabilities of the large language model.

[0036] It should be noted that a multi-layered topology diagram is obtained, including the cabinet architecture layer, primary circuit layer, secondary circuit layer, and component layer. The nodes within each layer and the inter-layer connections together constitute a complete graph structure reflecting the internal coupling relationships of the equipment. Among them, the cabinet architecture layer nodes include mechanical structural elements such as cabinet support structure, door hinges, and insulating support components; the primary circuit layer nodes include electrical main path elements such as incoming terminals, busbars and branch busbars, circuit breaker incoming and outgoing terminals, transformers, and cable joints; the secondary circuit layer nodes include control and signal elements such as control power supply, protection devices, intelligent control units, and status feedback contacts; and the component layer nodes include core component elements such as circuit breaker operating mechanisms, contactor contacts, capacitors, and surge arresters. The primary circuit layer and the component layer are connected by an electrical connection edge, reflecting the influence of components on the electrical characteristics of the main circuit; the cabinet structure layer and the component layer are connected by a mechanical coupling edge, reflecting the transmission relationship between mechanical vibration and structural deformation; the primary circuit layer and the cabinet structure layer are connected by a heat conduction edge, reflecting the heat transfer path between the heat source and the heat dissipation structure; the secondary circuit layer and the primary circuit layer are connected by a control association edge, reflecting the driving relationship of control signals on the operation of the main circuit.

[0037] Based on the multi-layer topological association graph, a spatiotemporal graph neural network is constructed as a health state propagation model. Key nodes are used to determine the state propagation rules and fault impact weights learned during the pre-training of the graph neural network, serving as the initialization parameters for the spatiotemporal graph neural network to achieve dynamic evolution modeling of the device's health state along the topological structure. The spatiotemporal graph neural network consists of a cascaded spatial graph convolution module and a time-series modeling module. The spatial graph convolution module uses the constructed multi-layer topological association graph as its convolutional skeleton, aggregating the state information of neighboring nodes through graph convolution operations to simulate the propagation process of faults or degradation states within the device. Graph convolution operations employ spectral domain graph convolution and spatial domain graph convolution to capture global topological structure features while preserving local neighborhood information. The time-series modeling module uses a gated recurrent unit network to model the state evolution of each node across continuous time segments, learning the temporal dynamic rules of the health state. By using the output of the spatial graph convolution module as the input of the time-series modeling module, joint extraction of spatiotemporal features is achieved.

[0038] The spatiotemporal graph neural network takes as input a sequence of deep health features of each node across multiple time segments and outputs predicted health status values ​​for each node at the current and future times. The network training employs a multi-task learning strategy, simultaneously optimizing the node state prediction loss and the overall health index prediction loss. This ensures the model accurately characterizes node-level states and precisely assesses the overall health level of each node's deep health features. A spatiotemporal attention mechanism is introduced to adaptively weight the node health status output by the spatiotemporal graph neural network. This mechanism includes spatial and temporal attention dimensions. The spatial attention dimension learns the contribution of different nodes to the overall health index, using the node's health status value and its centrality in the network topology as input, and generates normalized spatial attention weights through an attention network. The temporal attention dimension learns the influence of states at different historical times on the current health index, constructing temporal attention weights by calculating the correlation between state features at each historical time segment and the current health index. The spatial and temporal attention weights are then fused to form a spatiotemporal attention weight matrix. Based on this matrix, the node health status is weighted and fused to calculate the overall device health index.

[0039] The Monte Carlo dropout method is used to predict the future evolution trajectory of the health index and quantify its uncertainty. During the inference process of the spatiotemporal graph neural network, some neurons are randomly dropped multiple times to generate multiple health index prediction samples. Short-term, medium-term, and long-term predictions of health index change trends and confidence intervals are generated, and the probability distribution of the health index falling to each warning threshold is calculated. Based on the calculated health index and evolution trend, combined with risk location information and remaining life prediction results, graded warning information is generated. The graded warning information includes four levels: red, orange, yellow, and blue. Each level is associated with the health index threshold, the health index decline rate threshold, and the remaining life prediction threshold. The health index, graded warning information, and health assessment report are integrated to generate a complete equipment health assessment file, which is then archived and used as feedback data for model iterative optimization and continuous learning mechanism.

[0040] Preferably, by calculating the gradient contribution of node features in the graph neural network to the output, the key nodes with the greatest impact on the current health index are identified. By analyzing the information propagation path in the graph neural network, the propagation path of the abnormal status within the equipment is identified, and the propagation path of the abnormal status is output. For example: poor contact at the A-phase connection point - propagation along the bus to adjacent connection points - causing current imbalance in the B-phase - increasing the risk of malfunction of the circuit breaker operating mechanism. The output propagation path of the abnormal status presents the fault diffusion mechanism, helping maintenance personnel understand the equipment status and locate the root cause of the fault.

[0041] like Figure 4As shown, the second embodiment of the present invention provides a health assessment system for intelligent complete switchgear for fault prediction. The system includes: a multimodal data acquisition module, a deep health feature extraction module, a high-dimensional feature dimensionality reduction and visualization module, an intelligent analysis and assessment module, and a health index calculation and early warning module. The multimodal data acquisition module is responsible for deploying multimodal sensors at key nodes of high and low voltage switchgear to synchronously collect full-dimensional parameters reflecting the operating status of the equipment. The deep health feature extraction module is responsible for constructing a multi-source heterogeneous health feature extraction model based on deep learning. It adopts an appropriate neural network structure for different data types and automatically extracts deep health features that can characterize structural loosening, insulation degradation, poor contact, and component aging. The high-dimensional feature dimensionality reduction and visualization module is responsible for using a dual-core t-distribution random neighborhood embedding algorithm to robustly reduce the dimensionality of high-dimensional deep health features, mapping them to a low-dimensional space and generating a visualization distribution map. The intelligent analysis and evaluation module is responsible for integrating a large language model optimized with power equipment operation and maintenance knowledge, parsing the dimensionality-reduced feature map and the original multimodal data, generating a structured natural language health assessment report, and making probabilistic predictions on the remaining effective lifespan of key components. The health index calculation and early warning module integrates deep health features, visualization results, and large language model analysis conclusions to generate a dynamic health index that is bound to a specific physical device, and outputs graded early warning information according to preset thresholds.

[0042] The preferred system also includes: a data management and interaction module, which is responsible for the unified management, storage and interaction of system data, provides a human-computer interaction interface, and supports information interaction and decision-making collaboration between operation and maintenance personnel and the system.

[0043] The multimodal data acquisition module collects equipment operation data in real time, and transmits it to the deep health feature extraction module after preprocessing. The deep health feature extraction module performs deep feature extraction on the multimodal data to generate deep health feature vectors. The high-dimensional feature reduction and visualization module performs dimensionality reduction processing on the deep health features to generate low-dimensional embedded coordinates and a visualized distribution map. The intelligent analysis and evaluation module integrates deep features, visualization results, and raw data to generate a health assessment report and remaining life prediction results through a large language model. The health index calculation and early warning module integrates all the above information, calculates the dynamic health index, and outputs graded early warnings and maintenance suggestions. The data management and interaction module is responsible for the storage, display, and interaction of data throughout the entire process, forming a complete closed loop from data perception to decision-making services.

[0044] The third embodiment of the present invention provides a computer-readable storage medium, which includes a program for a health assessment method of intelligent switchgear for fault prediction. When the program for a health assessment method of intelligent switchgear for fault prediction is executed by a processor, it implements the steps of a health assessment method of intelligent switchgear for fault prediction.

[0045] In the several embodiments provided in this application, it should be understood that the disclosed methods and systems can be implemented in other ways. The system embodiments described above are merely illustrative. For example, the division of units is only a logical functional division, and in actual implementation, there may be other division methods, such as: multiple units or components can be combined, or integrated into another system, or some features can be ignored or not executed. In addition, the coupling, direct coupling, or communication connection between the various components shown or discussed can be through some interfaces, indirect coupling or communication connection of devices or units, and can be electrical, mechanical, or other forms. Furthermore, in the various embodiments of the present invention, all functional units can be integrated into one processing unit, or each unit can be a separate unit, or two or more units can be integrated into one unit; the integrated unit can be implemented in hardware or in the form of hardware plus software functional units.

[0046] Those skilled in the art will understand that all or part of the steps of the above method embodiments can be implemented by hardware related to program instructions. The aforementioned program can be stored in a computer-readable storage medium. When the program is executed, it performs the steps of the above method embodiments. The aforementioned storage medium includes various media capable of storing program code, such as mobile storage devices, read-only memory (ROM), random access memory (RAM), magnetic disks, or optical disks.

[0047] The above description is merely a specific embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any changes or substitutions that can be easily conceived by those skilled in the art within the scope of the technology disclosed in the present invention should be included within the scope of protection of the present invention.

Claims

1. A health assessment method for intelligent switchgear assemblies oriented towards fault prediction, characterized in that, Includes the following steps: Multimodal sensors are deployed at key nodes of high and low voltage switchgear to synchronously collect full-dimensional parameters reflecting the operating status of the equipment. These full-dimensional parameters include mechanical status parameters of the cabinet structure, electrical and thermal status parameters of the primary main circuit, status parameters of the secondary control circuit, and operating condition parameters of internal key components. Based on the full-dimensional parameters, a multi-source heterogeneous health feature extraction model based on deep learning is constructed. An adaptive neural network structure is adopted for different data types to automatically extract deep health features that can characterize structural loosening, insulation degradation, poor contact and component aging. For the deep health features, a robust dimensionality reduction process is performed using a dual-kernel t-distribution random neighborhood embedding algorithm, and the features are mapped to a low-dimensional space to generate a visual distribution map. The low-dimensional features and visualization results are input into a large language model optimized by power equipment operation and maintenance knowledge. The reduced-dimensional feature map and the original multimodal data are parsed to generate a structured natural language health assessment report for a specific cabinet, circuit or component, and to make probabilistic predictions of the remaining effective life of key components. By integrating the deep health features, visualization results, and large language model analysis conclusions, a dynamic health index bound to specific physical devices is generated, and graded early warning information is output according to preset thresholds.

2. The health assessment method for intelligent switchgear assemblies oriented towards fault prediction according to claim 1, characterized in that, The full-dimensional parameter collection specifically includes: Based on the physical structure of high and low voltage switchgear, a multi-layer topology diagram is constructed, including the cabinet architecture layer, primary circuit layer, secondary circuit layer and component layer. The multi-layer topological graph is pre-trained using a graph neural network to learn the state propagation rules and fault impact weights of each node during the evolution of equipment health status, and outputs the importance score of each node. By combining the node importance score output by the graph neural network with the information entropy maximization criterion, a heuristic search algorithm is used to select the sensor deployment node combination that maximizes the overall information gain and determine the location of key nodes. Based on the selected key node locations, deploy corresponding types of multimodal sensors, synchronize the time of each type of sensor through a synchronization triggering device, and synchronously collect parameters in all dimensions. The collected multimodal data is then cleaned, timestamp aligned and standardized in sequence to generate a full-dimensional parameter set.

3. The health assessment method for intelligent switchgear assemblies oriented towards fault prediction according to claim 1, characterized in that, Deep health characteristics that can characterize structural loosening, insulation degradation, poor contact, and component aging are extracted, specifically including: The collected full-dimensional parameters are divided into three heterogeneous data types: time-series signal data, image and spectral data, and state parameter data. Based on the heterogeneous data types, a multi-source heterogeneous health feature extraction model with three parallel feature extraction branches is constructed. In the multi-source heterogeneous health feature extraction model, the first branch is a one-dimensional time-series feature extraction network, which adopts a hybrid architecture of a one-dimensional convolutional neural network with residual connections and a bidirectional long short-term memory network to extract local waveform features and sequential dependencies in the time-series signal and generate deep time-series features. The second branch is a two-dimensional image feature extraction network. It adopts an improved residual network architecture and extracts temperature distribution patterns, hot spot features and thermal gradient changes in infrared thermal images, as well as phase distribution features, discharge amplitude patterns and spectrum shape features in partial discharge spectra through multi-layer convolution modules and attention mechanism modules, to generate deep image features. The third branch is a fully connected feature extraction network, which uses a multi-layer fully connected neural network to perform nonlinear transformation and feature combination on the state parameter data, learn the coupling relationship between ambient temperature, load rate, runtime, and number of switching actions, and generate deep state features. The deep temporal features, deep image features, and deep state features are input into the feature fusion module, and a fused multimodal feature vector is generated through adaptive weighting. The fused multimodal feature vector is input into a fully connected layer for dimensional transformation and nonlinear mapping, and the deep health features representing the health status of the device are output.

4. The health assessment method for intelligent switchgear assemblies oriented towards fault prediction according to claim 1, characterized in that, A robust dimensionality reduction process is performed using a dual-kernel t-distribution random neighborhood embedding algorithm, mapping the data to a lower-dimensional space. Specifically, this includes: For the aforementioned deep health features, a dual-kernel similarity matrix is ​​constructed that integrates a local linear kernel and a global Gaussian kernel. The local linear kernel calculates the reconstruction weight coefficient based on the linear reconstruction relationship within the neighborhood of the sample point to measure the local geometric structure on the high-dimensional manifold. The global Gaussian kernel calculates the Euclidean distance similarity between sample points using the traditional Gaussian kernel form to measure the overall distribution characteristics of the data. The local linear kernel and the global Gaussian kernel are then weighted and fused using an adaptive fusion coefficient to construct the dual-kernel similarity matrix. Based on the dual-kernel similarity matrix, the similarity relationship in the high-dimensional space is converted into a conditional probability distribution, and a joint probability distribution in the high-dimensional space is generated by symmetry. At the same time, the t-distribution is used as the similarity metric kernel function in the low-dimensional embedding space to define the joint probability distribution in the low-dimensional space. Using KL divergence as the objective function, the difference between the joint probability distribution in high-dimensional space and the joint probability distribution in low-dimensional space is measured. The objective function is optimized and solved to obtain the low-dimensional embedding coordinates corresponding to each high-dimensional deep health feature sample.

5. The health assessment method for intelligent switchgear assemblies oriented towards fault prediction according to claim 4, characterized in that, Generate a visual distribution map, specifically including: The low-dimensional embedded coordinates are plotted in three-dimensional space to generate a visual distribution map. In the visual distribution map, samples with the same health status are clustered together to form clusters, while samples with different health statuses are separated in space. The same fault type shows a continuous evolution trajectory as the degree of deterioration deepens. The effectiveness of the dimensionality reduction results is verified by K-nearest neighbor classification accuracy, silhouette coefficient, and distance correlation index. When the index meets the preset threshold, the low-dimensional features and visualization results are output. If the threshold is not met, the adaptive fusion coefficient, neighborhood size, and global bandwidth parameters are adjusted, and the dimensionality reduction is recalculated.

6. The health assessment method for intelligent switchgear assemblies oriented towards fault prediction according to claim 1, characterized in that, The generation of the structured natural language health assessment report specifically includes: Based on low-dimensional features and visualization results, original full-dimensional parameters, and deep health features, a comprehensive input context for a large language model is constructed. The thinking chain reasoning mechanism guides the large language model to reason step by step according to the logical chain of state recognition, risk positioning, and causal analysis; In the state identification phase, based on the input multimodal feature description and visualization results, the current health status category of the equipment is identified and judgment criteria are generated. In the risk location phase, combining the equipment topology and fault propagation patterns, the specific location and impact range of the fault are inferred, and the assessment conclusions are associated with specific cabinets, circuits, or components. In the causal analysis phase, fault mechanism knowledge from the domain knowledge base is used to analyze the possible causes and evolution mechanisms of the fault. Based on the reasoning analysis results, the large language model generates a structured natural language health assessment report, which includes basic equipment information, comprehensive health index, sub-item status assessment, risk location, fault cause analysis, and maintenance recommendations.

7. The health assessment method for intelligent switchgear assemblies oriented towards fault prediction according to claim 1, characterized in that, Probabilistic prediction of the remaining effective life of key components specifically includes: We acquire degradation-sensitive features, spatiotemporal evolution features of low-dimensional embedded trajectories, and key component degradation models based on physical mechanisms from deep health features, and fuse them to construct a multi-source lifetime feature set. The multi-source lifetime feature set, equipment historical operating information, current operating conditions and key parameters, and degradation theoretical values ​​calculated based on the physical model are input into the large language model for conditional information encoding to generate conditional embedding vectors. The conditional embedding vector is used as the input diffusion probability model. The diffusion probability model adopts a noise prediction network based on the Transformer architecture. It dynamically focuses on the conditional information most relevant to the current prediction task through a cross-attention mechanism. Starting from standard Gaussian noise, multiple remaining lifetime samples are generated through an iterative denoising process to form the posterior distribution of the remaining lifetime. During the training of the diffusion probability model, a physical information neural network constraint mechanism is introduced to construct a physical consistency loss function to measure the difference between the remaining lifetime predicted by the diffusion probability model and the prediction results of the physical information neural network. Through weighted joint optimization, the prediction results are made to conform to physical laws. Perform statistical analysis on the generated remaining lifespan samples and output probabilistic prediction results.

8. The health assessment method for intelligent switchgear assemblies oriented towards fault prediction according to claim 1, characterized in that, By integrating the aforementioned deep health features, visualization results, and large language model analysis conclusions, a dynamic health index bound to a specific physical device is generated, specifically including: Obtain a multi-layer topology graph containing a cabinet architecture layer, a primary loop layer, a secondary loop layer, and a component layer. Based on the multi-layer topology graph, construct a spatiotemporal graph neural network as a health state propagation model. Key nodes are used to determine the state propagation rules and fault impact weights learned by the pre-training of the central graph neural network, which are used as the initialization parameters of the spatiotemporal graph neural network. In the spatiotemporal graph neural network, the state information of neighboring nodes is aggregated through graph convolution operations to simulate the propagation process of fault or degradation state inside the device. A gated recurrent unit network is used to model the state evolution of each node on a continuous time cross-section. The sequence of deep health features of each node at multiple time points is input into the spatiotemporal graph neural network. The node state prediction loss and the overall health index prediction loss are optimized simultaneously through a multi-task learning strategy, and the predicted health status values ​​of each node at the current time and future time are output. A spatiotemporal attention mechanism is introduced to adaptively weight the node health status output by the spatiotemporal graph neural network, and calculate the overall health index of the device.

9. The health assessment method for intelligent switchgear assemblies oriented towards fault prediction according to claim 8, characterized in that, Based on preset thresholds, tiered early warning information is output, specifically including... The Monte Carlo dropout method is used to predict the future evolution trajectory of the health index and quantify the uncertainty. Short-term, medium-term and long-term predictions of the health index change trends and confidence intervals are generated, and the probability distribution of the health index falling to each warning threshold is calculated. Based on the calculated health index and its evolution trend, combined with risk location information and remaining life expectancy prediction results, a graded early warning information is generated. By integrating health indices, tiered early warning information, and health assessment reports, a complete equipment health assessment file is generated and archived, serving as feedback data for continuous learning mechanisms in model iteration and optimization.

10. A health assessment system for intelligent switchgear assemblies oriented towards fault prediction, used to implement the health assessment method for intelligent switchgear assemblies oriented towards fault prediction as described in any one of claims 1-9, characterized in that, The system includes: a multimodal data acquisition module, a deep health feature extraction module, a high-dimensional feature dimensionality reduction and visualization module, an intelligent analysis and evaluation module, and a health index calculation and early warning module; The multimodal data acquisition module is responsible for deploying multimodal sensors at key nodes of high and low voltage switchgear to synchronously collect full-dimensional parameters reflecting the operating status of the equipment. The deep health feature extraction module is responsible for constructing a multi-source heterogeneous health feature extraction model based on deep learning. It adopts an appropriate neural network structure for different data types and automatically extracts deep health features that can characterize structural loosening, insulation degradation, poor contact, and component aging. The high-dimensional feature dimensionality reduction and visualization module is responsible for using a dual-core t-distribution random neighborhood embedding algorithm to robustly reduce the dimensionality of high-dimensional deep health features, mapping them to a low-dimensional space and generating a visualization distribution map. The intelligent analysis and evaluation module is responsible for integrating a large language model optimized with power equipment operation and maintenance knowledge, parsing the dimensionality-reduced feature map and the original multimodal data, generating a structured natural language health assessment report, and making probabilistic predictions on the remaining effective lifespan of key components. The health index calculation and early warning module integrates deep health features, visualization results, and large language model analysis conclusions to generate a dynamic health index that is bound to a specific physical device, and outputs graded early warning information according to preset thresholds.

Images