AI-based power equipment health state prediction system and method thereof

Abstract

The invention discloses an AI-based power equipment health state prediction system and method, and belongs to the field of intelligent operation and maintenance of power equipment, and the system comprises a data collection and preprocessing module, a multi-modal deep hierarchical attention fusion module, a health state evaluation and prediction module, and a decision support and linkage module. According to the invention, a multi-modal deep layered attention fusion architecture is innovatively provided, the sequential characteristics of the oil chromatography dissolved gas concentration are analyzed through a bottom-layer sequential characteristic extraction network, and a middle-layer spatial-temporal pattern recognition network is associated with spatial-temporal patterns of an infrared thermogram and an ultrasonic spectrum; the top equipment family relationship network constructs an equipment family defect propagation model to realize full-dimensional analysis and equipment group collaborative awareness of multi-modal data; and meanwhile, an information flow path between different levels of networks is established through a cross-level information fusion mechanism, so that the comprehensiveness and accuracy of feature extraction are improved.

Description

Technical Field

[0001] This invention relates to the field of intelligent operation and maintenance of power equipment, specifically to an artificial intelligence-based power equipment health status prediction system and method, used to monitor, evaluate and predict the health status of key power equipment such as transformers and circuit breakers, and to realize intelligent linkage between maintenance plans and power grid dispatch. Background Technology

[0002] As a crucial infrastructure for the safe and stable operation of the power grid, the health of power equipment directly affects the reliability and security of the entire power system. Traditional power equipment maintenance strategies mainly include periodic maintenance and post-fault repair. This approach often suffers from over-maintenance or under-maintenance, wasting resources and failing to effectively prevent sudden failures, leading to a decline in power grid reliability.

[0003] With the development of sensing and communication technologies, power equipment condition monitoring systems have been widely applied in the power industry. Current power equipment health status monitoring methods mainly rely on single types of monitoring data, such as oil chromatography analysis, infrared thermography, or ultrasonic testing, lacking the ability to comprehensively analyze multi-source heterogeneous data. Furthermore, most existing technologies employ simple threshold-based judgment methods or single-model analysis methods, failing to fully extract deeper information from multi-source data and resulting in limited prediction accuracy.

[0004] Furthermore, existing power equipment health status monitoring systems typically analyze individual devices as independent entities, neglecting the interrelationships and influences between devices, thus failing to achieve collaborative analysis and risk assessment of equipment groups. Power equipment maintenance plans and grid dispatch systems are also often disconnected, making it difficult to optimize the allocation of maintenance resources and ensure the stability of grid operation.

[0005] Therefore, there is an urgent need to develop a power equipment health status prediction system that can integrate multimodal data, achieve full-dimensional analysis, support equipment group collaboration and grid dispatch linkage, so as to improve prediction accuracy, reduce maintenance costs and ensure the safe and stable operation of the power grid. Summary of the Invention

[0006] The purpose of this invention is to provide an AI-based power equipment health status prediction system and method. Through multimodal data fusion and deep hierarchical attention mechanism, it can achieve accurate assessment of the health status of power equipment and fault prediction. Furthermore, through intelligent linkage with the power grid dispatching system, it can optimize the allocation of maintenance resources and the operation mode of the power grid.

[0007] This invention proposes an AI-based power equipment health status prediction system, comprising:

[0008] The data acquisition and preprocessing module is used to acquire oil chromatography, infrared thermography and ultrasonic partial discharge multimodal sensing data of key power equipment such as transformers and circuit breakers, and to clean and standardize the multimodal sensing data.

[0009] A multimodal deep hierarchical attention fusion module, electrically connected to the data acquisition and preprocessing module, is used to receive the multimodal sensing data and process it through a multi-level feature extraction network, wherein the multi-level feature extraction network includes:

[0010] A low-level temporal feature extraction network is used to analyze the temporal features of dissolved gas concentrations in oil chromatography.

[0011] A mid-level spatiotemporal pattern recognition network is used to correlate the spatiotemporal patterns of infrared thermograms and ultrasound spectra.

[0012] A top-level device family relationship network is used to construct a device family defect propagation model; and

[0013] A cross-layer information fusion mechanism is used to establish information flow paths between the bottom, middle and top layers of the network.

[0014] The health status assessment and prediction module is electrically connected to the multimodal deep hierarchical attention fusion module, and is used to assess the health status of power equipment and predict potential faults based on the fusion features output by the multi-level feature extraction network.

[0015] The decision support and linkage module is electrically connected to the health status assessment and prediction module. When the insulation aging rate is detected to exceed the standard, it is used to automatically generate a maintenance plan and link it with the power grid dispatching system to optimize the power flow distribution.

[0016] Preferably, the data acquisition and preprocessing module includes:

[0017] The multimodal sensor acquisition unit is used to acquire oil chromatography data, infrared thermography data, and ultrasonic partial discharge data during the operation of power equipment.

[0018] The data cleaning unit is used to perform missing value processing, outlier detection, and data smoothing on the multimodal sensing data.

[0019] The feature enhancement unit is used to enhance the trend features of oil chromatography data through differential transformation, enhance the temperature anomaly area of ​​infrared thermal imaging through histogram equalization, and enhance the specific frequency band features of ultrasonic signals through spectrum analysis.

[0020] The data standardization unit is used to convert the multimodal sensing data into a unified format and perform standardization processing.

[0021] Preferably, the underlying temporal feature extraction network includes:

[0022] Multi-scale time coding unit, used to extract features from oil chromatographic time series data using different time windows;

[0023] The nonlinear time-series mapping module is used to capture the nonlinear characteristics of gas concentration changes through composite nonlinear transformation;

[0024] A three-level memory gating structure is used to process short-term, medium-term, and long-term changing characteristics respectively;

[0025] Gas interaction modeling unit, used to construct gas interaction matrix, characterizing the strength of the correlation between different gases.

[0026] Preferably, the mid-level spatiotemporal pattern recognition network includes:

[0027] The spatial feature extraction module is used to extract spatial features from infrared thermal images using a multi-scale convolutional structure.

[0028] The spectrum analysis module is used to perform time-frequency domain transformation on ultrasonic partial discharge signals and extract spectral features;

[0029] Cross-modal attention units are used to establish the mapping relationship between thermal imaging and ultrasound data, and to achieve multimodal information fusion;

[0030] A spatiotemporal pattern recognizer is used to identify typical fault mode features based on fused spatiotemporal features.

[0031] Preferably, the top-level device family relationship network includes:

[0032] The graph structure generation module is used to construct multi-relationship graphs based on the physical connections, electrical relationships, and spatial layout of devices.

[0033] A node feature fusion unit is used to integrate features extracted from the bottom and middle layers as feature representations of graph nodes.

[0034] Edge feature calculator, used to calculate edge features based on the relationships between devices, characterizing the probability of defect propagation;

[0035] The graph dynamics analyzer is used to analyze the propagation patterns of defects in graph structures and identify critical propagation paths and vulnerabilities.

[0036] Preferably, the cross-level information fusion mechanism includes:

[0037] The feature fusion path unit is used to establish feature transfer channels from the bottom layer to the middle layer, from the middle layer to the top layer, and feedback connections from the top layer to the bottom layer and the middle layer.

[0038] The feature importance assessment unit is used to dynamically assess the importance of features at different levels based on the information gain principle.

[0039] The adaptive fusion ratio unit is used to automatically adjust the fusion ratio of different levels of features based on device type, operating status, and historical data.

[0040] Spatiotemporal consistency constraint unit is used to ensure the consistency of features at different levels in time and space.

[0041] Preferably, the health status assessment and prediction module includes:

[0042] A multi-layer feature fusion unit is used to integrate features from different levels through early fusion, mid-term fusion, and late-term fusion strategies;

[0043] A health index calculation unit is used to calculate the device's health index based on fusion features;

[0044] A status classification unit is used to map health indices to discrete health status levels;

[0045] The fault prediction unit is used to predict future trends in health index changes and the probability of failure based on historical data.

[0046] The risk assessment unit is used to comprehensively consider the probability of failure, the severity of consequences, and the system impact to assess the risk level of the equipment.

[0047] Preferably, the decision support and linkage module includes:

[0048] The maintenance plan generation unit is used to generate maintenance plans based on the risk assessment results, with priority ranking.

[0049] The resource optimization unit is used to optimize the allocation of maintenance tasks based on maintenance resource constraints.

[0050] The dispatching and coordination unit is used to submit maintenance plans to the power grid dispatching system and receive feedback.

[0051] The power flow optimization unit is used to optimize the power flow distribution based on maintenance plans and grid operating conditions.

[0052] Preferably, the system adopts a distributed deployment architecture, including:

[0053] The edge layer, deployed at the substation site, includes edge computing servers, sensor access gateways, and communication modules, and is used to perform data acquisition and preprocessing as well as lightweight feature extraction.

[0054] The Fog layer, deployed in the regional power dispatch center, includes a regional server cluster, a regional data center, and a regional communication center, and is used to perform mid-level feature extraction and fusion as well as regional equipment health status assessment.

[0055] The cloud layer, deployed in the power grid control center, includes a cloud computing platform, a network-wide data center, and a network control center, and is used to perform top-level network analysis and network-wide health status assessment and prediction.

[0056] An AI-based method for predicting the health status of power equipment includes the following steps:

[0057] Collect multimodal sensing data of key power equipment such as transformers and circuit breakers, including oil chromatography, infrared thermography, and ultrasonic partial discharge, and clean and standardize the multimodal sensing data.

[0058] The multimodal sensing data is processed using a multimodal deep hierarchical attention fusion architecture, which includes:

[0059] The temporal features of dissolved gas concentration in oil chromatography are analyzed using a low-level temporal feature extraction network.

[0060] The spatiotemporal patterns of infrared thermograms and ultrasound spectra are associated through a mid-level spatiotemporal pattern recognition network.

[0061] A defect propagation model for device families is constructed using a top-level device family relationship network; and

[0062] An information flow path is established between the bottom, middle and top layers of the network through a cross-layer information fusion mechanism;

[0063] Based on the fusion features output by the multimodal deep hierarchical attention fusion architecture, the health status of power equipment is evaluated and potential faults are predicted.

[0064] When the insulation aging rate is detected to exceed the standard, a maintenance plan is automatically generated and linked with the power grid dispatch system to optimize power flow distribution.

[0065] The present invention has the following beneficial effects:

[0066] 1. Multimodal data fusion: By fusing multimodal data such as oil chromatography, infrared thermography, and ultrasonic partial discharge, it provides comprehensive equipment status perception capabilities, overcomes the limitations of a single data source, and improves the comprehensiveness and accuracy of health status assessment.

[0067] 2. Hierarchical feature extraction: The hierarchical architecture of bottom-level temporal feature extraction, mid-level spatiotemporal pattern recognition and top-level device family relationship modeling is adopted to realize full-dimensional analysis from micro to macro and capture multi-scale features of device state changes.

[0068] 3. Equipment Group Collaborative Analysis: Through the top-level equipment family relationship network, an inter-equipment correlation model was established, enabling the analysis and prediction of defect propagation patterns and improving the early warning capability for cascading failures.

[0069] 4. Maintenance and Dispatch Linkage: It realizes intelligent linkage between maintenance plans and power grid dispatching system, which optimizes power grid operation mode while ensuring equipment safety, and improves the overall reliability and economy of the system.

[0070] 5. Distributed deployment architecture: Adopting a three-tier deployment architecture of edge-fog-cloud, it achieves reasonable allocation of computing resources and improves system scalability, adapting to the application needs of power grids of different scales.

[0071] Through the above technical solutions, the present invention can improve the accuracy of power equipment health status prediction by 25-35%, advance the fault warning time by 30-90 days, improve the maintenance resource utilization rate by more than 30%, improve the power grid reliability index by 10-15%, significantly reduce operation and maintenance costs and fault risks, and has important economic value and social benefits. Attached Figure Description

[0072] Figure 1 This is a diagram illustrating the overall architecture of the AI-based power equipment health status prediction system of this invention.

[0073] Figure 2 This is a structural diagram of the data acquisition and preprocessing module of the present invention;

[0074] Figure 3 This is a structural diagram of the multimodal deep hierarchical attention fusion module of the present invention;

[0075] Figure 4 This is a structural diagram of the underlying temporal feature extraction network of this invention;

[0076] Figure 5 This is a structural diagram of the mid-layer spatiotemporal pattern recognition network of the present invention;

[0077] Figure 6 This is a structural diagram of the top-level device family relationship network of the present invention;

[0078] Figure 7 This is a structural diagram of the cross-level information fusion mechanism of the present invention;

[0079] Figure 8 This is a structural diagram of the health status assessment and prediction module of the present invention;

[0080] Figure 9 This is a structural diagram of the decision support and linkage module of the present invention;

[0081] Figure 10 This is a flowchart of the AI-based power equipment health status prediction method of the present invention. Detailed Implementation

[0082] Please refer to Figures 1-10 The present invention will now be described in further detail with reference to the accompanying drawings and specific embodiments.

[0083] Reference Figure 1 The AI-based power equipment health status prediction system provided by the present invention includes: a data acquisition and preprocessing module 1, a multimodal deep hierarchical attention fusion module 2, a health status assessment and prediction module 3, and a decision support and linkage module 4.

[0084] The data acquisition and preprocessing module 1 is used to acquire multimodal sensing data from key power equipment such as transformers and circuit breakers, including oil chromatography, infrared thermography, and ultrasonic partial discharge data, and to clean and standardize the multimodal sensing data. In one embodiment of the present invention, the acquisition frequency of the multimodal sensing data is as follows: oil chromatography data 4 times / day, infrared thermography data 24 times / day, and ultrasonic partial discharge data continuously acquired at a sampling rate of 1MHz.

[0085] The multimodal deep hierarchical attention fusion module 2 is electrically connected to the data acquisition and preprocessing module 1, and is used to receive the multimodal sensing data and process it through a multi-level feature extraction network. This module is the core innovation of this invention, realizing full-dimensional perception and analysis from microscopic temporal features to macroscopic topological features through a three-level hierarchical network structure. The multi-level feature extraction network includes: a bottom-level temporal feature extraction network 21, a middle-level spatiotemporal pattern recognition network 22, a top-level device family relationship network 23, and a cross-level information fusion mechanism 24.

[0086] The health status assessment and prediction module 3 is electrically connected to the multimodal deep hierarchical attention fusion module 2. It is used to assess the health status of power equipment and predict potential faults based on the fused features output by the multi-level feature extraction network. This module maps the fused features to a health index and a fault probability, providing a basis for decision-making.

[0087] The decision support and linkage module 4 is electrically connected to the health status assessment and prediction module 3. It is used to automatically generate a maintenance plan and link it with the power grid dispatch system to optimize power flow distribution when an excessive insulation aging rate is detected. Preferably, the system will trigger the maintenance plan generation process when the insulation aging rate exceeds a monthly threshold of 1.5%.

[0088] Reference Figure 2 The data acquisition and preprocessing module 1 of the present invention includes: a multimodal sensor acquisition unit 11, a data cleaning unit 12, a feature enhancement unit 13, and a data standardization unit 14.

[0089] The multimodal sensor acquisition unit 11 is used to acquire oil chromatography data, infrared thermal imaging data, and ultrasonic partial discharge data during the operation of power equipment. In a preferred embodiment of the present invention, an online oil chromatography monitoring device is installed on each key transformer to monitor the concentration of eight key gases, including H2, CH4, C2H2, C2H4, and C2H6; a fixed infrared thermal imager with a spatial resolution of 320×240 pixels and a temperature accuracy of ±0.5°C is installed; and an ultrasonic partial discharge detector with a frequency range of 50kHz-500kHz and a sensitivity of 5pC is installed.

[0090] The data cleaning unit 12 is used to process the multimodal sensing data for missing values, outlier detection, and data smoothing. For missing values, this invention employs a hybrid interpolation method based on temporal correlation and device similarity. This method first attempts to fill short-term missing values ​​using time-series interpolation; when the missing time is long, it uses data from similar devices from the same period as a reference for filling. Outlier detection combines statistical methods with expert rules, calculating the z-score of the data; when |z-score|>3, it is determined to be an outlier. Data smoothing uses wavelet transform, performing a 5-level decomposition of the oil chromatography data using db4 wavelet, removing high-frequency noise, and then reconstructing the data.

[0091] Feature enhancement unit 13 is used to enhance the trend features of oil chromatography data through differential transformation, enhance the temperature anomaly region of infrared thermography through histogram equalization, and enhance the specific frequency band features of ultrasonic signals through spectral analysis. For oil chromatography data, the first and second differences of gas concentration are calculated to highlight the changing trend; for infrared thermography, an adaptive histogram equalization algorithm is used to enhance image contrast; for ultrasonic signals, short-time Fourier transform is used to extract the energy features of the 150-300kHz frequency band, which is highly correlated with insulation degradation.

[0092] The data standardization unit 14 is used to convert the multimodal sensing data into a unified format and perform standardization processing. Standardization uses the Z-score method, and the calculation formula is as follows:

[0093] ,

[0094] in, The standardized value. The original data values, The mean of the data. This represents the standard deviation of the data. Standardization ensures that data of different dimensions and scales can be fused and analyzed within the same framework.

[0095] Reference Figure 3 and Figure 4The underlying temporal feature extraction network 21 of the present invention includes: a multi-scale temporal coding unit 211, a nonlinear temporal mapping module 212, a three-level memory gating structure 213, and a gas interaction modeling unit 214.

[0096] The multi-scale time coding unit 211 is used to extract features from oil chromatographic time-series data using different time windows. This invention employs three different time windows of 1 day, 7 days, and 30 days to capture short-term, medium-term, and long-term variation features, respectively. For the data in each time window, statistical features (mean, variance, kurtosis, skewness) and trend features (slope, curvature) are calculated.

[0097] The nonlinear time-series mapping module 212 is used to capture the nonlinear characteristics of gas concentration changes through composite nonlinear transformation. This module implements a mapping from the original gas concentration value to a representation of the equipment state, considering the nonlinear relationship between gas concentration changes and equipment state. The mapping function is defined as:

[0098] ,

[0099] in, gas In time The mapping results gas In time concentration value, , This is the weight matrix. , For bias vectors, and These are the hyperbolic tangent function and the sigmoid function, respectively.

[0100] The three-level memory gating structure 213 is used to process short-term, medium-term, and long-term change features separately. This structure is an improvement on the traditional LSTM, introducing a multi-timescale memory mechanism that can simultaneously capture change patterns at different timescales. The gating function is defined as:

[0101] ,

[0102] ,

[0103] ,

[0104] ,

[0105] ,

[0106] ,

[0107] Among them, superscript These represent short-term, medium-term, and long-term memory units, respectively. , , These are the forget gate, input gate, and output gate, respectively. In cellular state, In hidden state, and For weights and bias parameters, This represents the Hadamard product (element-wise multiplication).

[0108] Gas interaction modeling unit 214 is used to construct the gas interaction matrix, characterizing the strength of the correlation between different gases. Interaction matrix elements Represents gas With gas The correlation is calculated using the following formula:

[0109] ,

[0110] in, gas and covariance, and Gases and The standard deviation of the interaction matrix is ​​calculated. The interaction matrix is ​​processed using a graph convolutional network to extract the interaction features between gases.

[0111] Reference Figure 3 and Figure 5 The mid-level spatiotemporal pattern recognition network 22 of the present invention includes: a spatial feature extraction module 221, a spectrum analysis module 222, a cross-modal attention unit 223, and a spatiotemporal pattern recognizer 224.

[0112] The spatial feature extraction module 221 is used to extract spatial features from infrared thermal images using a multi-scale convolutional structure. This module employs a three-layer convolutional neural network, with each layer using convolutional kernels of different sizes (3×3, 5×5, 7×7) to capture spatial features at different scales. The mathematical expression for the convolutional layer is:

[0113] ,

[0114] in, For the first Feature map of the layer For convolution kernel weights, For bias, * indicates convolution operation. This is the activation function (the ReLU function is used in this embodiment).

[0115] The spectrum analysis module 222 is used to perform time-frequency domain transformation on the ultrasonic partial discharge signal and extract spectral features. This module uses the Short Time Fourier Transform (STFT) method to convert the time-domain signal into a time-spectrum graph. The calculation formula is as follows:

[0116] ,

[0117] in, For time-domain signals, For window functions (Hanning window is used in this embodiment). For time offset, The angular frequency is used. In this invention, the STFT uses a window length of 10ms, an overlap rate of 50%, and a frequency resolution of 1kHz.

[0118] The cross-modal attention unit 223 is used to establish the mapping relationship between thermal imaging and ultrasound data, realizing multimodal information fusion. This unit adopts a bidirectional attention mechanism, calculating the attention weights of thermal imaging features on ultrasound features and the attention weights of ultrasound features on thermal imaging features, respectively. The attention calculation formula is as follows:

[0119] ,

[0120] ,

[0121] in, For attention weights, To score the similarity, and These are the feature vectors for mode A (thermal imaging) and mode B (ultrasound), respectively. and The transformation matrix for the query and key values. , where is the feature dimension. The fused feature calculation is as follows:

[0122] ,

[0123] in, The adaptive fusion coefficients are dynamically adjusted based on the quality and correlation of the two modalities. In a preferred embodiment of the invention, The initial value is set to 0.5, and it is automatically optimized through the backpropagation process.

[0124] The spatiotemporal pattern recognizer 224 is used to identify typical fault mode features based on fused spatiotemporal features. This recognizer employs a combined structure of gated recurrent units (GRU) and convolutional neural networks (CNN), enabling it to simultaneously process time-series and spatial distribution features. Model training utilizes supervised learning methods, using labeled fault samples. The recognizer outputs the probability distribution of fault types, as well as the temporal and spatial location information of the fault features.

[0125] Reference Figure 3 and Figure 6 The top-level device family relationship network 23 of the present invention includes: a graph structure generation module 231, a node feature fusion unit 232, an edge feature calculator 233, and a graph dynamics analyzer 234.

[0126] Graph structure generation module 231 is used to construct a multi-relationship graph based on the physical connections, electrical relationships, and spatial layout of devices. This module represents electrical equipment as nodes in the graph and the relationships between devices as edges. The graph structure is defined as G=(V,E,W), where V is the set of nodes representing electrical equipment; E is the set of edges representing the relationships between devices; and W is the set of edge weights representing the strength of the relationships. In this invention, three types of relationships are considered: physical connection relationships (such as electrical connections), functional correlations (such as the main transformer and its auxiliary equipment), and spatial proximity (such as physical distance).

[0127] The node feature fusion unit 232 integrates features extracted from the lower and middle layers as feature representations of graph nodes. Node features are represented as vectors $x_i$, containing device type, operating parameters, historical status, and features extracted from the lower and middle layers. Feature fusion employs an attention mechanism, calculated using the following formula:

[0128] ,

[0129] in, Let i be the fusion feature of node i. For features from different levels, These are attention weights, representing the importance of different features. The attention weights are learned automatically during training, with initial values ​​set to [value missing]. , where n is the number of feature sources.

[0130] The edge feature calculator 233 is used to calculate edge features based on the relationships between devices, characterizing the probability of defect propagation. Edge features include the physical connection strength, electrical coupling degree, and spatial distance between devices. Edge weights. The calculation formula is:

[0131] ,

[0132] in, Let be the edge weight from node i to node j. and The features of node i and node j are respectively. The relationship features between node i and node j and For weights and bias parameters, The Sigmoid function normalizes the edge weights to the [0,1] interval.

[0133] The graph dynamics analyzer 234 is used to analyze the propagation patterns of defects in graph structures, identifying critical propagation paths and vulnerabilities. This analyzer employs a graph neural network (GNN) model, simulating the propagation process of defects in a device network through a message passing mechanism. The node state update formula is:

[0134] ,

[0135] in, Let i be the state of node i at layer l. Let i be the set of neighbors of node i. This is an aggregation function (weighted summation is used in this example). and Here are the weights and bias parameters for the l-th layer. This is the activation function.

[0136] Through multi-layer graph convolution operations, the model can capture high-order relationships between devices and defect propagation patterns. The analysis results include device vulnerability scores (indicating the likelihood of a device being affected) and criticality scores (indicating the degree of impact of a device on the system), providing a basis for fault propagation early warning and critical device identification.

[0137] Reference Figure 3 and Figure 7 The cross-level information fusion mechanism 24 of the present invention includes: a feature fusion path unit 241, a feature importance evaluation unit 242, an adaptive fusion ratio unit 243, and a spatiotemporal consistency constraint unit 244.

[0138] Feature fusion path unit 241 is used to establish feature transfer channels from the bottom layer to the middle layer, from the middle layer to the top layer, and feedback connections from the top layer to the bottom and middle layers. This unit realizes the mutual transfer and integration of features at different levels, forming a complete information flow network. During forward propagation, bottom-layer temporal features provide prior knowledge in the time dimension, middle-layer spatiotemporal features provide environmental information in the spatial dimension, and top-layer topological features provide a system-level global perspective. During backward propagation, the global information from the top layer influences the feature extraction processes of the bottom and middle layers through feedback connections, forming a closed-loop optimization mechanism.

[0139] The feature importance evaluation unit 242 is used to dynamically evaluate the importance of features at different levels based on the information gain principle. This unit uses information entropy and mutual information to calculate the importance weights of different features, using the following formula:

[0140] ,

[0141] ,

[0142] in, Features For target variable Information gain The entropy of the target variable, For a given feature Conditional entropy of the target variable under given conditions Features Importance weights.

[0143] The adaptive fusion ratio unit 243 automatically adjusts the fusion ratio of features at different levels based on device type, operating status, and historical data. This unit employs a dynamic weight allocation strategy, adjusting the contribution ratio of features at different levels according to data quality and task requirements. The fusion feature calculation formula is as follows:

[0144] ,

[0145] in, For the features after fusion, For different levels of features, The fusion weights are determined by the following formula:

[0146] ,

[0147] in, Features The quality score (reflecting data reliability) Features The relevance score (reflecting the degree of relevance to the task). The quality score and relevance score are dynamically updated by monitoring model performance and data characteristics.

[0148] The spatiotemporal consistency constraint unit 244 is used to ensure the consistency of features at different levels in time and space. This unit introduces a spatiotemporal consistency loss function, requiring features at different levels to maintain coordination and consistency in the time and space dimensions. The spatiotemporal consistency loss function is defined as follows:

[0149] ,

[0150] ,

[0151] ,

[0152] in, For the total consistency loss, and These represent temporal and spatial consistency losses, and Features Representation in the time and space dimensions, and This refers to the weighting coefficient. In a preferred embodiment of the invention, and Both are set to 0.5, indicating that temporal and spatial consistency are equally important.

[0153] Reference Figure 8 The health status assessment and prediction module 3 of the present invention includes: a multi-layer feature fusion unit 31, a health index calculation unit 32, a status classification unit 33, a fault prediction unit 34, and a risk assessment unit 35.

[0154] The multi-layer feature fusion unit 31 is used to integrate features from different levels through early fusion, mid-term fusion, and late-term fusion strategies. Early fusion is performed during the feature extraction stage, mid-term fusion during the feature representation stage, and late-term fusion during the decision-making stage. The choice of fusion strategy depends on the nature of the features and the task requirements. This invention adopts a hybrid fusion strategy, using early fusion for highly correlated features and late fusion for highly complementary features, to maximize the utilization of information from multiple layers of features.

[0155] The health index calculation unit 32 is used to calculate the device health index based on fusion features. The health index represents a quantitative indicator of the overall health status of the device, ranging from 0 to 100, with higher values ​​indicating better device condition. The health index calculation formula is as follows:

[0156] ,

[0157] in, For health index, The first fusion feature One portion, For the corresponding weights, This is the Sigmoid function, used to normalize the result to the range of 0-100. Weights Automatic optimization through model training reflects the degree to which different features contribute to health status.

[0158] The status classification unit 33 is used to map the health index to discrete health status levels. This invention defines four health status levels: Normal (health index 80-100), Attention (health index 60-80), Warning (health index 40-60), and Danger (health index 0-40). Status classification employs a threshold-based judgment method, considering both the absolute value and trend of the health index. For example, when the health index is in the 60-80 range but shows a continuous downward trend, the system may escalate the status to the "Warning" level.

[0159] The fault prediction unit 34 is used to predict future health index trends and fault probabilities based on historical data. This unit employs a time series prediction model, combined with multi-layer features, to predict the changing trends of equipment health status over a future period (typically 7 days, 30 days, and 90 days). The fault probability calculation formula is as follows:

[0160] ,

[0161] in, From the current time To the future The probability of an internal failure. For time The failure rate function is related to the predicted health index. In this invention, the failure rate function is defined as:

[0162] ,

[0163] in, As the baseline failure rate, The influence coefficient of the health index. For time Predicted health index. Parameters and In this embodiment, the data is obtained by fitting historical fault data. times / day .

[0164] Risk assessment unit 35 is used to comprehensively consider failure probability, severity of consequences, and system impact to assess the equipment risk level. The risk score calculation formula is:

[0165] ,

[0166] in, To score risk, For failure probability, For severity of consequences (range 1-10), The system impact factor (range 1-10) is used. The severity of the consequences is determined based on the equipment type, capacity, and importance, while the system impact factor is determined based on the equipment's topological location in the power grid and load conditions. A risk score greater than 50 is considered high-risk and requires priority handling.

[0167] Reference Figure 9 The decision support and linkage module 4 of the present invention includes: a maintenance plan generation unit 41, a resource optimization unit 42, a scheduling linkage unit 43, and a power flow optimization unit 44.

[0168] The maintenance plan generation unit 41 generates a priority-ranked maintenance plan based on the risk assessment results. The plan generation considers equipment risk level, urgency, and maintenance resource constraints, arranging maintenance tasks in descending order of risk. The maintenance plan includes equipment information, maintenance content, suggested time windows, and resource requirements.

[0169] Resource optimization unit 42 is used to optimize maintenance task allocation based on maintenance resource constraints. This unit employs an integer programming method to maximize the overall risk reduction effect while satisfying resource constraints. The optimization objective function is:

[0170] ,

[0171] ,

[0172] ,

[0173] in, The amount of risk reduction brought about by maintenance task i For decision variables (1 indicates performing the task, 0 indicates not performing it), Let i be the resource requirement of task i for j. Let be the total available amount of resource j.

[0174] The dispatching linkage unit 43 is used to submit the maintenance plan to the power grid dispatching system and receive feedback. This unit realizes two-way communication between the maintenance plan and the power grid dispatching system, converts maintenance requirements into a format recognizable by the dispatching system, and receives feedback from the dispatching system. Feedback includes suggestions for adjusting the maintenance time window, power flow adjustment measures, and system stability assessments. Based on the feedback information, the system may adjust the maintenance plan to adapt to the power grid's operational needs.

[0175] Power flow optimization unit 44 is used to optimize power flow distribution based on maintenance plans and grid operating conditions. This unit analyzes the impact of maintenance plans on grid power flow through power flow calculations and optimization algorithms, and generates the optimal scheduling scheme to ensure system stability and economy. The optimization objective is to minimize line losses and load imbalance while satisfying safety constraints. The mathematical expression of the power flow optimization model is:

[0176] ,

[0177] ,

[0178] ,

[0179] in, Let be the objective function, representing the total line loss. For power flow equation constraints, Safety constraints (such as line capacity limits, node voltage limits, etc.) are taken into account. Optimization results include measures such as generator output adjustment, load transfer, and network topology adjustment.

[0180] Reference Figure 1 The system of this invention adopts a distributed deployment architecture, which includes three layers: edge layer, Fog layer and cloud layer.

[0181] The edge layer is deployed at the substation site and includes an edge computing server, sensor access gateway, and communication module. It performs data acquisition and preprocessing, as well as lightweight feature extraction. The edge layer hardware configuration includes an 8-core CPU, 16GB RAM, 2TB storage, and a GPU accelerator card, sufficient to support local data processing and preliminary analysis. The edge layer software includes data acquisition, preprocessing, lightweight feature extraction, and local early warning modules, capable of independently performing basic status monitoring and early warning functions.

[0182] The Fog layer is deployed in the regional power dispatch center, comprising a regional server cluster, a regional data center, and a regional communication center. It performs mid-level feature extraction and fusion, as well as regional equipment health status assessment. The Fog layer hardware configuration includes a 24-core CPU, 64GB RAM, 10TB storage, and multiple GPU computing nodes, capable of processing data from multiple substations within the region. The Fog layer software includes modules for regional data aggregation, mid-level feature extraction, regional health status assessment, and maintenance plan coordination, enabling collaborative analysis and management of equipment within the region.

[0183] The cloud layer, deployed in the power grid control center, includes a cloud computing platform, a network-wide data center, and a network control center. It performs top-level network analysis and network-wide health status assessment and prediction. The cloud layer utilizes a high-performance computing cluster with petabyte-level storage capacity and powerful computing capabilities, supporting network-wide data analysis and decision support. The cloud layer software includes modules for network-wide data analysis, top-level network analysis, network-wide health status assessment and prediction, and maintenance planning and scheduling linkage, enabling intelligent operation and maintenance management across the entire network.

[0184] The three layers exchange data and collaborate through standard interfaces and secure communication channels, forming a complete distributed intelligent system. This layered deployment architecture offers excellent scalability and flexibility, adapting to the application needs of power grids of different sizes and maintaining basic functionality even in the event of communication interruptions.

[0185] Reference Figure 10 The AI-based power equipment health status prediction method of the present invention includes the following steps:

[0186] Step S1: Collect oil chromatography, infrared thermography and ultrasonic partial discharge multimodal sensing data of key power equipment such as transformers and circuit breakers, and clean and standardize the multimodal sensing data.

[0187] In this step, the system acquires oil chromatography, infrared thermography, and ultrasonic partial discharge data through a multimodal sensor network deployed on power equipment. The acquisition frequency is: oil chromatography data 4 times / day, infrared thermography data 24 times / day, and ultrasonic partial discharge data continuously acquired at a sampling rate of 1MHz. The acquired data undergoes cleaning, including missing value handling, outlier detection, and data smoothing. Then, feature enhancement is performed, using methods such as differential transformation, histogram equalization, and spectral analysis to enhance key features in the original data. Finally, data standardization is performed, converting data from different sources and scales into a unified format and unit to provide a foundation for subsequent analysis.

[0188] Step S2: Process the multimodal sensing data using a multimodal deep hierarchical attention fusion architecture.

[0189] In this step, the system uses a multimodal deep hierarchical attention fusion architecture to process the standardized data. This architecture consists of a three-layer network structure:

[0190] Step S2.1: Analyze the temporal features of dissolved gas concentration in oil chromatography using a low-level temporal feature extraction network.

[0191] The underlying network focuses on temporal feature extraction, employing techniques such as multi-scale time coding, nonlinear temporal mapping, three-level memory gating structure, and gas interaction modeling to capture short-term, medium-term, and long-term variation features in oil chromatography data, as well as the interrelationships between gases.

[0192] Step S2.2: Use a mid-level spatiotemporal pattern recognition network to associate the spatiotemporal patterns of infrared thermograms and ultrasound spectra.

[0193] The mid-layer network focuses on spatiotemporal feature extraction and modality fusion. It employs techniques such as spatial feature extraction, spectrum analysis, and cross-modal attention to achieve collaborative analysis of infrared thermal imaging and ultrasonic spectrum data, capturing the spatial distribution and spectral features of device status.

[0194] Step S2.3: Construct a device family defect propagation model through the top-level device family relationship network.

[0195] The top-level network focuses on modeling the relationships between device groups. It uses techniques such as graph structure generation, node feature fusion, edge feature calculation, and graph dynamics analysis to construct device relationship graphs, analyze defect propagation patterns, and identify critical devices and vulnerabilities.

[0196] Step S2.4: Establish information flow paths between the bottom, middle and top layers of the network through a cross-level information fusion mechanism.

[0197] The cross-level information fusion mechanism enables the mutual transmission and integration of features at different levels. It includes components such as feature fusion path, feature importance assessment, adaptive fusion ratio, and spatiotemporal consistency constraints, ensuring that the system has a comprehensive perception and understanding of the device status.

[0198] Step S3: Based on the fusion features output by the multimodal deep hierarchical attention fusion architecture, assess the health status of the power equipment and predict potential faults.

[0199] In this step, the system calculates the equipment health index based on fused features, maps the health index to discrete health status levels, predicts future trends in the health index and the probability of failure, and comprehensively assesses the equipment risk level. The health status assessment results and failure prediction results provide a basis for subsequent decision-making.

[0200] Step S4: When the insulation aging rate is detected to exceed the standard, an overhaul plan is automatically generated and linked with the power grid dispatch system to optimize power flow distribution.

[0201] In this step, the system generates a priority maintenance plan based on health status assessment and fault prediction results, optimizes maintenance resource allocation, submits the maintenance plan to the power grid dispatch system and receives feedback, and optimizes power flow distribution based on the maintenance plan and power grid operating status. When the insulation aging rate exceeds the monthly threshold of 1.5%, the system triggers the maintenance plan generation process and links with the power grid dispatch system to ensure stable power grid operation while protecting equipment safety.

[0202] Through the above steps, this invention achieves comprehensive perception, accurate assessment and prediction of the health status of power equipment, and through intelligent linkage with the power grid dispatching system, optimizes the allocation of maintenance resources and the operation mode of the power grid, thereby improving the safety, reliability and economy of the power system.

[0203] The above description is merely a preferred embodiment of the present invention and is not intended to limit the invention. Various modifications and variations can be made to the present invention by those skilled in the art. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principle of the present invention should be included within the scope of protection of the present invention.

Claims

1. An AI-based power equipment health status prediction system, characterized in that, include: The data acquisition and preprocessing module is used to acquire oil chromatography, infrared thermography and ultrasonic partial discharge multimodal sensing data of key power equipment such as transformers and circuit breakers, and to clean and standardize the multimodal sensing data. A multimodal deep hierarchical attention fusion module, electrically connected to the data acquisition and preprocessing module, is used to receive the multimodal sensing data and process it through a multi-level feature extraction network, wherein the multi-level feature extraction network includes: A low-level temporal feature extraction network is used to analyze the temporal features of dissolved gas concentrations in oil chromatography. A mid-level spatiotemporal pattern recognition network is used to correlate the spatiotemporal patterns of infrared thermograms and ultrasound spectra. A top-level device family relationship network is used to construct a device family defect propagation model; and A cross-layer information fusion mechanism is used to establish information flow paths between the bottom, middle and top layers of the network. The health status assessment and prediction module is electrically connected to the multimodal deep hierarchical attention fusion module, and is used to assess the health status of power equipment and predict potential faults based on the fusion features output by the multi-level feature extraction network. The decision support and linkage module is electrically connected to the health status assessment and prediction module. When the insulation aging rate is detected to exceed the standard, it is used to automatically generate a maintenance plan and link it with the power grid dispatching system to optimize the power flow distribution.

2. The AI-based power equipment health status prediction system according to claim 1, characterized in that, The data acquisition and preprocessing module includes: The multimodal sensor acquisition unit is used to acquire oil chromatography data, infrared thermography data, and ultrasonic partial discharge data during the operation of power equipment. The data cleaning unit is used to perform missing value processing, outlier detection, and data smoothing on the multimodal sensing data. The feature enhancement unit is used to enhance the trend features of oil chromatography data through differential transformation, enhance the temperature anomaly area of ​​infrared thermal imaging through histogram equalization, and enhance the specific frequency band features of ultrasonic signals through spectrum analysis. The data standardization unit is used to convert the multimodal sensing data into a unified format and perform standardization processing.

3. The AI-based power equipment health status prediction system according to claim 1, characterized in that, The underlying temporal feature extraction network includes: Multi-scale time coding unit, used to extract features from oil chromatographic time series data using different time windows; The nonlinear time-series mapping module is used to capture the nonlinear characteristics of gas concentration changes through composite nonlinear transformation; A three-level memory gating structure is used to process short-term, medium-term, and long-term changing characteristics respectively; Gas interaction modeling unit, used to construct gas interaction matrix, characterizing the strength of the correlation between different gases.

4. The AI-based power equipment health status prediction system according to claim 1, characterized in that, The mid-level spatiotemporal pattern recognition network includes: The spatial feature extraction module is used to extract spatial features from infrared thermal images using a multi-scale convolutional structure. The spectrum analysis module is used to perform time-frequency domain transformation on ultrasonic partial discharge signals and extract spectral features; Cross-modal attention units are used to establish the mapping relationship between thermal imaging and ultrasound data, and to achieve multimodal information fusion; A spatiotemporal pattern recognizer is used to identify typical fault mode features based on fused spatiotemporal features.

5. The AI-based power equipment health status prediction system according to claim 1, characterized in that, The top-level device family relationship network includes: The graph structure generation module is used to construct multi-relationship graphs based on the physical connections, electrical relationships, and spatial layout of devices. A node feature fusion unit is used to integrate features extracted from the bottom and middle layers as feature representations of graph nodes. Edge feature calculator, used to calculate edge features based on the relationships between devices, characterizing the probability of defect propagation; The graph dynamics analyzer is used to analyze the propagation patterns of defects in graph structures and identify critical propagation paths and vulnerabilities.

6. The AI-based power equipment health status prediction system according to claim 1, characterized in that, The cross-level information fusion mechanism includes: The feature fusion path unit is used to establish feature transfer channels from the bottom layer to the middle layer, from the middle layer to the top layer, and feedback connections from the top layer to the bottom layer and the middle layer. The feature importance assessment unit is used to dynamically assess the importance of features at different levels based on the information gain principle. The adaptive fusion ratio unit is used to automatically adjust the fusion ratio of different levels of features based on device type, operating status, and historical data. Spatiotemporal consistency constraint unit is used to ensure the consistency of features at different levels in time and space.

7. The AI-based power equipment health status prediction system according to claim 1, characterized in that, The health status assessment and prediction module includes: A multi-layer feature fusion unit is used to integrate features from different levels through early fusion, mid-term fusion, and late-term fusion strategies; A health index calculation unit is used to calculate the device's health index based on fusion features; A status classification unit is used to map health indices to discrete health status levels; The fault prediction unit is used to predict future trends in health index changes and the probability of failure based on historical data. The risk assessment unit is used to comprehensively consider the probability of failure, the severity of consequences, and the system impact to assess the risk level of the equipment.

8. The AI-based power equipment health status prediction system according to claim 1, characterized in that, The decision support and linkage module includes: The maintenance plan generation unit is used to generate maintenance plans based on the risk assessment results, with priority ranking. The resource optimization unit is used to optimize the allocation of maintenance tasks based on maintenance resource constraints. The dispatching and coordination unit is used to submit maintenance plans to the power grid dispatching system and receive feedback. The power flow optimization unit is used to optimize the power flow distribution based on maintenance plans and grid operating conditions.

9. The AI-based power equipment health status prediction system according to claim 1, characterized in that, The system adopts a distributed deployment architecture, including: The edge layer, deployed at the substation site, includes edge computing servers, sensor access gateways, and communication modules, and is used to perform data acquisition and preprocessing as well as lightweight feature extraction. The Fog layer, deployed in the regional power dispatch center, includes a regional server cluster, a regional data center, and a regional communication center, and is used to perform mid-level feature extraction and fusion as well as regional equipment health status assessment. The cloud layer, deployed in the power grid control center, includes a cloud computing platform, a network-wide data center, and a network control center, and is used to perform top-level network analysis and network-wide health status assessment and prediction.

10. An AI-based method for predicting the health status of power equipment, employing the AI-based power equipment health status prediction system as described in any one of claims 1-9, characterized in that, Includes the following steps: Collect multimodal sensing data of key power equipment such as transformers and circuit breakers, including oil chromatography, infrared thermography, and ultrasonic partial discharge, and clean and standardize the multimodal sensing data. The multimodal sensing data is processed using a multimodal deep hierarchical attention fusion architecture, which includes: The temporal features of dissolved gas concentration in oil chromatography are analyzed using a low-level temporal feature extraction network. The spatiotemporal patterns of infrared thermograms and ultrasound spectra are associated through a mid-level spatiotemporal pattern recognition network. A defect propagation model for device families is constructed using a top-level device family relationship network; and An information flow path is established between the bottom, middle and top layers of the network through a cross-layer information fusion mechanism; Based on the fusion features output by the multimodal deep hierarchical attention fusion architecture, the health status of power equipment is evaluated and potential faults are predicted. When the insulation aging rate is detected to exceed the standard, a maintenance plan is automatically generated and linked with the power grid dispatch system to optimize power flow distribution.

Images