Substation equipment intelligent monitoring method based on artificial intelligence

Abstract

This invention relates to the field of intelligent monitoring technology for power equipment, specifically an artificial intelligence-based intelligent monitoring method for substation equipment. The method includes: collecting multi-source heterogeneous state data of substation equipment through a sensor network, such as infrared thermal images, partial discharge ultrasonic signals, mechanical vibration waveforms, and oil chromatographic component content; preprocessing and time-series alignment to form a structured time-series data stream; extracting equipment fusion feature tensors using an improved deep residual graph convolution algorithm with optimized adjacency matrices; and simultaneously completing anomaly identification, fault classification, and remaining life prediction through a pre-trained multi-task joint learning model; and generating dynamic maintenance decision schemes based on output indicators. This method can accurately characterize the overall health status of equipment, achieve collaborative analysis of multiple monitoring tasks, make operation and maintenance decisions more aligned with the actual operating status of equipment, and improve the precision of substation equipment monitoring and operation and maintenance.

Description

Technical Field

[0001] This invention relates to the field of intelligent monitoring technology for power equipment, and in particular to an intelligent monitoring method for substation equipment based on artificial intelligence. Background Technology

[0002] In existing substation equipment monitoring, infrared thermal images, partial discharge ultrasonic waves, mechanical vibration waveforms, and oil chromatographic data are collected by dispersed sensors. These various status data are analyzed separately after only routine preprocessing, and feature extraction relies on conventional convolutional or neural network algorithms without incorporating the actual physical connections of the substation equipment into the analytical model. In the fault monitoring and lifespan assessment stages, independent models are used to perform anomaly identification, fault classification, or remaining lifespan prediction. The monitoring results are then manually integrated to form maintenance-related instructions.

[0003] The temporal matching degree of status data from multi-source heterogeneous equipment is low. Conventional algorithms cannot combine the physical connection topology of equipment with signal propagation delay to construct correlation features, and the extracted features cannot fully reflect the overall operating status of the equipment. The single-task independent analysis mode will cause information fragmentation between different monitoring tasks, and the results of anomaly judgment, fault classification and life estimation lack correlation. The output monitoring data cannot directly support the formulation of refined maintenance decisions. It is necessary to optimize the feature extraction logic of graph convolution algorithm, adapt it to the physical correlation and signal transmission characteristics of substation equipment, and realize the synchronous execution of multiple monitoring tasks, so as to form corresponding equipment maintenance decision content based on integrated analysis results. Summary of the Invention

[0004] The purpose of this invention is to address the shortcomings of existing technologies by proposing an intelligent monitoring method for substation equipment based on artificial intelligence.

[0005] To achieve the above objectives, the present invention adopts the following technical solution: an intelligent monitoring method for substation equipment based on artificial intelligence, comprising: periodically collecting multi-source heterogeneous equipment status data from a sensor network deployed at monitoring points of various equipment in the substation, wherein the multi-source heterogeneous equipment status data includes equipment infrared thermal images, partial discharge ultrasonic signals, equipment mechanical vibration waveforms, and oil chromatographic component content; preprocessing and time-series aligning the multi-source heterogeneous equipment status data to generate a structured equipment status time-series data stream; and calling an improved deep residual graph convolution algorithm to extract equipment status fusion features from the structured equipment status time-series data stream. A fusion feature tensor representing the overall health of the equipment is extracted. The improved deep residual graph convolution algorithm optimizes the construction of the adjacency matrix based on the equipment's physical connection topology and signal propagation delay. The fusion feature tensor is input into a pre-trained multi-task joint learning model, which simultaneously performs abnormal state identification, fault type classification, and remaining life prediction tasks, and outputs abnormal alarms, fault categories, and predicted life indicators. Based on the abnormal alarms, fault categories, and predicted life indicators, a dynamic maintenance decision scheme is generated, which includes maintenance priority ranking, maintenance resource allocation suggestions, and equipment operation control instructions.

[0006] As a further aspect of the present invention, the multi-source heterogeneous device status data is preprocessed and time-series aligned to generate a structured device status time-series data stream, including: performing hot spot region segmentation and temperature matrix extraction on the infrared thermal image of the device to generate a time-series sequence of the device surface temperature distribution matrix; performing time-frequency transformation and feature wave extraction on the partial discharge ultrasonic signal to generate a time-series sequence of discharge pulse intensity and frequency band energy; performing wavelet packet decomposition on the mechanical vibration waveform of the device to extract the energy features of each frequency band and generate a vibration energy spectrum time-series sequence; performing normalization and moving average filtering on the oil chromatographic component content data to generate a smooth time-series sequence of the concentration ratio of each gas component; establishing a unified timestamp alignment benchmark, and interpolating and synchronizing the device surface temperature distribution matrix time-series sequence, the discharge pulse intensity and frequency band energy time-series sequence, the vibration energy spectrum time-series sequence, and the smooth time-series sequence of the concentration ratio of each gas component according to the unified timestamp alignment benchmark, and encapsulating them into the structured device status time-series data stream.

[0007] As a further aspect of the present invention, the step of calling the improved deep residual graph convolution algorithm to extract equipment state fusion features from the structured equipment state time-series data stream, and extracting a fusion feature tensor characterizing the overall health of the equipment, includes: constructing an equipment connection topology graph with power equipment as nodes and electrical connections and physical proximity between equipment as edges, based on the substation's primary wiring diagram and equipment physical layout diagram; calculating the delay time required for different types of state data to be transmitted on the edges of the equipment connection topology graph according to the signal propagation characteristics and sampling periods of the equipment infrared thermal image, partial discharge ultrasonic signal, equipment mechanical vibration waveform, and oil chromatographic component content; and performing edge processing on each edge of the equipment connection topology graph based on the delay time. The weights are dynamically adjusted to generate a time-aware device state association graph. The adjacency matrix element values ​​of the time-aware device state association graph are jointly determined by the inter-device connection relationship and data transmission delay. The structured device state time-series data stream is used as the input node features and fed into an improved deep residual graph convolution algorithm with the time-aware device state association graph as the graph structure for iterative convolution operations. After each convolution operation, residual connections are introduced to skip the sum of the output of the current layer and the output of the previous layer to alleviate the gradient vanishing problem in deep network training. After multiple layers of graph convolution and residual connections, the features of all nodes are aggregated in the last layer to form the fusion feature tensor representing the overall health of the device.

[0008] As a further aspect of the present invention, the fused feature tensor is input into a pre-trained multi-task joint learning model. The multi-task joint learning model simultaneously performs abnormal state identification, fault type classification, and remaining lifetime prediction tasks, including: the multi-task joint learning model comprises a shared feature encoder and three independent task decoders; the shared feature encoder performs high-order abstract encoding on the input fused feature tensor; the output of the shared feature encoder is simultaneously input to the abnormal state identification decoder, the fault type classification decoder, and the remaining lifetime prediction decoder; the abnormal state identification decoder calculates the probability of the device being in an abnormal state through a fully connected layer and a sigmoid activation function, and compares the probability with a preset threshold to generate the abnormal alarm; the fault type classification decoder outputs the fault type of the device and its corresponding confidence level through a classification layer containing a softmax function to generate the fault category; the remaining lifetime prediction decoder outputs the predicted time for the device to operate from its current state to a predetermined failure threshold through a regression layer to generate the predicted lifetime index.

[0009] As a further aspect of the present invention, the pre-training process of the multi-task joint learning model includes: collecting historical structured equipment state time-series data streams and their corresponding historical anomaly labels, historical fault type labels, and historical remaining lifetime labels to form a training sample set; inputting the training sample set into the multi-task joint learning model, wherein the shared feature encoder extracts features from the input data; the anomaly state recognition decoder calculates the anomaly state probability based on the extracted features, the fault type classification decoder calculates the fault type distribution, and the remaining lifetime prediction decoder calculates the predicted lifetime; calculating the cross-entropy loss between the output of the anomaly state recognition decoder and the historical anomaly labels, calculating the cross-entropy loss between the output of the fault type classification decoder and the historical fault type labels, and calculating the mean squared error loss between the output of the remaining lifetime prediction decoder and the historical remaining lifetime labels; weighted summing of the three losses to obtain the multi-task joint loss; and using a backpropagation algorithm to minimize the multi-task joint loss while simultaneously updating the network parameters of the shared feature encoder and the three task decoders until the model converges.

[0010] As a further aspect of the present invention, based on the abnormal alarm, fault category, and predicted lifespan index, a dynamic maintenance decision-making scheme is generated, including maintenance priority ranking, maintenance resource allocation suggestions, and equipment operation control instructions. This includes: constructing a decision knowledge base, which stores standard maintenance strategies, required spare parts lists, estimated maintenance man-hours, and operation control logic under different fault categories and predicted lifespan indices; when the abnormal alarm is received, a decision-making process is triggered, querying the decision knowledge base based on the fault category and predicted lifespan index to match a baseline maintenance strategy and resource requirements; combining the substation's current maintenance resource inventory and maintenance team scheduling plan, the feasibility of the baseline maintenance strategy and resource requirements is verified and conflict detected; based on the verification results, time windows are adjusted or resources are replaced for maintenance tasks with resource conflicts, generating a feasible maintenance task execution plan; based on the predicted lifespan index and the current operating load of the equipment, the recommended operating power limit or recommended downtime of the equipment is dynamically calculated, forming the equipment operation control instructions; and the feasible maintenance task execution plan, the adjusted resource requirements, and the equipment operation control instructions are integrated to output the final dynamic maintenance decision-making scheme.

[0011] As a further aspect of the present invention, the establishment of a unified timestamp alignment benchmark, which involves interpolating and synchronizing the time series sequences of the equipment surface temperature distribution matrix, the discharge pulse intensity and frequency band energy, the vibration energy spectrum, and the smoothed time series sequences of the concentration ratios of various gas components according to the unified timestamp alignment benchmark, includes: selecting the master clock of the substation monitoring system as the unified timestamp alignment benchmark; establishing a mapping relationship between the original acquisition timestamp and the unified timestamp alignment benchmark for each time series; defining a target sampling frequency under the unified timestamp alignment benchmark; resampling the data points of each sequence to the target sampling frequency using a linear interpolation method for the equipment surface temperature distribution matrix, the discharge pulse intensity and frequency band energy, and the vibration energy spectrum; resampling the smoothed time series sequences of the concentration ratios of various gas components to the target sampling frequency using a nearest neighbor interpolation method due to their slow changes; and combining all sequence data resampled to the same time axis and the same sampling frequency according to the equipment identifier to form the structured equipment status time series data stream.

[0012] As a further aspect of the present invention, the weights of each edge in the device connection topology graph are dynamically adjusted based on the delay time to generate a time-aware device state association graph, including: for each connection edge in the device connection topology graph, obtaining the delay time determined by its corresponding signal propagation characteristics; setting a reference time window and calculating the delay ratio coefficient of the delay time relative to the reference time window; multiplying the initial weight value of each edge in the device connection topology graph by the reciprocal of the corresponding delay ratio coefficient to obtain the time-adjusted edge weight value; introducing a distance attenuation factor, which is calculated based on the physical distance between devices, and further multiplying the time-adjusted edge weight value by the distance attenuation factor; normalizing the final weight values ​​of all edges to ensure that the sum of all weight values ​​is a fixed constant; and constructing a weighted adjacency matrix of the time-aware device state association graph using devices as nodes and the normalized final weight values ​​as edge weights.

[0013] As a further aspect of the present invention, the abnormal state recognition decoder calculates the probability that the device is in an abnormal state through a fully connected layer and a sigmoid activation function, and compares the probability with a preset threshold to generate the abnormal alarm. This includes: the abnormal state recognition decoder receiving a feature vector output from the shared feature encoder; the fully connected layer inside the abnormal state recognition decoder performing a linear transformation on the feature vector, mapping the high-dimensional features to a one-dimensional value; applying a sigmoid activation function to the one-dimensional value after the linear transformation by the fully connected layer for non-linear mapping, constraining the mapping result to the interval between zero and one, the mapping result being the probability that the device is in an abnormal state; comparing the calculated probability of the device being in an abnormal state with a preset alarm threshold in real time; and generating an abnormal alarm message containing a device identifier, an abnormal probability value, and a timestamp when the probability of the device being in an abnormal state continuously exceeds the preset alarm threshold for more than a preset time window.

[0014] As a further aspect of the present invention, calculating the cross-entropy loss between the output of the abnormal state identification decoder and the historical abnormal labels, calculating the cross-entropy loss between the output of the fault type classification decoder and the historical fault type labels, and calculating the mean squared error loss between the output of the remaining lifetime prediction decoder and the historical remaining lifetime labels, includes: during the model training phase, acquiring a batch of data from the training sample set, including the structured equipment state time-series data stream, the corresponding historical abnormal labels, historical fault type labels, and historical remaining lifetime labels; inputting the structured equipment state time-series data stream into the multi-task joint learning model to obtain the abnormal state probability output by the abnormal state identification decoder, the fault type probability distribution output by the fault type classification decoder, and the remaining lifetime prediction solution. The coder outputs the predicted lifetime value; for the output of the abnormal state identification decoder, a binary cross-entropy loss function is used to compare the output abnormal state probability with the historical abnormal labels, and the cross-entropy loss between the output of the abnormal state identification decoder and the historical abnormal labels is calculated; for the output of the fault type classification decoder, a multivariate cross-entropy loss function is used to compare the output fault type probability distribution with the historical fault type labels, and the cross-entropy loss between the output of the fault type classification decoder and the historical fault type labels is calculated; for the output of the remaining lifetime prediction decoder, a mean squared error loss function is used to compare the output predicted lifetime value with the historical remaining lifetime labels, and the mean squared error loss between the output of the remaining lifetime prediction decoder and the historical remaining lifetime labels is calculated.

[0015] Compared with existing technologies, the advantages and positive effects of this invention are as follows: Based on the adjacency matrix construction method of optimizing the deep residual graph convolution algorithm using the physical connection topology of equipment and signal propagation delay, this invention extracts equipment state fusion features from structured equipment state time-series data streams. This method closely matches the physical layout relationships and signal transmission timing characteristics of substation equipment, transforming multi-source heterogeneous equipment state data into a unified fusion feature tensor. It fully correlates the intrinsic relationships between parameters such as infrared thermography, partial discharge ultrasonic signals, mechanical vibration waveforms, and oil chromatographic component content, accurately depicting the overall health status of the equipment. The optimized adjacency matrix construction logic can adapt to the time delay characteristics in signal transmission, ensuring that the feature extraction process aligns with the actual signal propagation patterns of the equipment. The fusion feature tensor can cover multi-dimensional representation information of the equipment's operating status, eliminating the one-sidedness of state representation caused by single-state data feature analysis, improving the matching accuracy between feature data and the actual operating status of the equipment, and making the extracted feature information more closely match the actual operating characteristics of substation equipment.

[0016] By integrating feature tensors into a pre-trained multi-task joint learning model, and simultaneously performing tasks such as abnormal state identification, fault type classification, and remaining life prediction, the model enables the interactive reuse of feature information across different monitoring tasks. This allows for a synergistic correlation between the outputs of abnormal alarms, fault category determinations, and predicted life indicators, avoiding the information fragmentation caused by independent analysis of single tasks. The simultaneous execution of multiple tasks allows various monitoring results to corroborate each other, improving the consistency between state determinations and life estimations. Based on the synchronously output multi-dimensional indicators, maintenance priority ranking, maintenance resource allocation suggestions, and equipment operation control commands can be directly generated. The generation of dynamic maintenance decision-making schemes closely relies on real-time equipment monitoring results, resulting in a higher degree of matching between decision content and equipment fault status and remaining life. This ensures that maintenance decisions are more aligned with the real-time operating status of substation equipment, guaranteeing the consistency between monitoring analysis results and operation and maintenance decisions. Attached Figure Description

[0017] Figure 1 This is a flowchart of the intelligent monitoring method for substation equipment based on artificial intelligence as described in this invention; Figure 2 A flowchart for generating a structured device state timing data stream for preprocessing and timing alignment; Figure 3 A flowchart for improving the deep residual graph convolution algorithm to extract fused feature tensors. Detailed Implementation

[0018] To make the objectives, technical solutions, and advantages of this invention clearer, the invention will be further described in detail below with reference to the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are merely illustrative and not intended to limit the invention.

[0019] In the description of this invention, it should be understood that the terms "length," "width," "upper," "lower," "front," "rear," "left," "right," "vertical," "horizontal," "top," "bottom," "inner," and "outer," etc., indicating orientation or positional relationships, are based on the orientation or positional relationships shown in the accompanying drawings and are only for the convenience of describing the invention and simplifying the description, and do not indicate or imply that the device or element referred to must have a specific orientation, or be constructed and operated in a specific orientation, and therefore should not be construed as a limitation of the invention. Furthermore, in the description of this invention, "a plurality of" means two or more, unless otherwise explicitly specified.

[0020] See Figure 1 This invention provides an intelligent monitoring method for substation equipment based on artificial intelligence. The method includes: periodically collecting multi-source heterogeneous equipment status data through a sensor network deployed at monitoring points of various equipment in the substation. This multi-source heterogeneous equipment status data includes equipment infrared thermal images, partial discharge ultrasonic signals, equipment mechanical vibration waveforms, and oil chromatographic component content. The multi-source heterogeneous equipment status data is preprocessed and time-series aligned to generate a structured equipment status time-series data stream. An improved deep residual graph convolution algorithm is used to extract equipment status fusion features from the structured equipment status time-series data stream, extracting table... The improved deep residual graph convolution algorithm optimizes the construction of the adjacency matrix based on the physical connection topology of the equipment and the signal propagation delay to obtain a fusion feature tensor of the overall health of the equipment. This fusion feature tensor is input into a pre-trained multi-task joint learning model, which simultaneously performs abnormal state identification, fault type classification, and remaining life prediction tasks, and outputs abnormal alarms, fault categories, and predicted life indicators. Based on the abnormal alarms, fault categories, and predicted life indicators, a dynamic maintenance decision scheme is generated, which includes maintenance priority ranking, maintenance resource allocation suggestions, and equipment operation control instructions.

[0021] In one embodiment of the present invention, see [reference] Figure 2The system performs hot spot region segmentation and temperature matrix extraction on the infrared thermal image of the equipment to generate a time series sequence of the surface temperature distribution matrix of the equipment; it performs time-frequency transformation and feature wave extraction on the partial discharge ultrasonic signal to generate a time series sequence of discharge pulse intensity and frequency band energy; it performs wavelet packet decomposition on the mechanical vibration waveform of the equipment to extract the energy features of each frequency band and generate a time series sequence of vibration energy spectrum; and it performs normalization and moving average filtering on the oil chromatographic component content data to generate a smooth time series sequence of the concentration ratio of each gas component. The master clock of the substation monitoring system is selected as the unified timestamp alignment reference. A mapping relationship between the original acquisition timestamp and the unified timestamp alignment reference is established for each time series. Under the unified timestamp alignment reference, a target sampling frequency is defined. For the equipment surface temperature distribution matrix time series, discharge pulse intensity and frequency band energy time series, and vibration energy spectrum time series, a linear interpolation method is used to resample the data points of each series to the target sampling frequency. For the smoothed time series of each gas component concentration ratio, a nearest neighbor interpolation method is used to resample it to the target sampling frequency. All sequence data resampled to the same time axis and the same sampling frequency are combined according to the equipment identifier and encapsulated into a structured equipment status time series data stream.

[0022] In practical implementation, for infrared thermal images of equipment, the temperature matrix of the equipment surface is extracted by segmenting the hot spot region of the image, forming a time series sequence of the equipment surface temperature distribution matrix; for partial discharge ultrasonic signals, time-frequency transformation is performed using short-time Fourier transform to extract the amplitude and phase information of characteristic waves, generating a time series sequence of discharge pulse intensity and frequency band energy containing discharge pulse intensity and specific frequency band energy; for mechanical vibration waveforms of equipment, wavelet packet decomposition algorithm is used to decompose them into different frequency bands, and the energy characteristics of each frequency band are calculated to construct a vibration energy spectrum time series sequence; for oil chromatographic component content data, the maximum and minimum values ​​of each gas concentration are first normalized, and then a moving average filter with a window length of 5 is applied to obtain a smooth time series sequence of the concentration ratio of each gas component.

[0023] In some embodiments, the key to time alignment lies in establishing a unified time reference and performing data resampling. The GPS master clock with the highest accuracy in the substation monitoring system is selected as the unified timestamp alignment reference. Mapping relationships are established between the original acquisition timestamps and the unified timestamp alignment reference for the time series sequences of equipment surface temperature distribution matrix, discharge pulse intensity and frequency band energy, vibration energy spectrum, and smoothed time series sequences of various gas component concentration ratios. A target sampling frequency is defined. (Unit: Hz), where This represents the time interval after resampling. For time series sequences of high-frequency changing equipment surface temperature distribution matrices, discharge pulse intensity and frequency band energy, and vibration energy spectrum, linear interpolation is used to resample them to the target sampling frequency. For smooth time series sequences of low-frequency, slowly varying gas component concentration ratios, the nearest neighbor interpolation method, which has lower computational complexity, is used for resampling. All sequence data that are on the same time axis after interpolation synchronization are combined and encapsulated according to the equipment's unique identifier to generate a structured equipment status time series data stream.

[0024] It is understandable that the accuracy of data preprocessing directly affects the accuracy of subsequent analysis. In practice, processing infrared thermographic images of equipment can effectively distinguish between normal temperature rises and abnormal hot spots. Compared to the unprocessed raw thermographic image, the temperature matrix can more accurately reflect the details of surface temperature field changes. Time-frequency transformation processing of partial discharge ultrasonic signals, compared to simple time-domain analysis, can capture the abrupt changes in signal characteristics in specific frequency bands. For example, in a certain measurement, the frequency band energy sequence after time-frequency transformation showed significant fluctuations several hours before the fault occurred, while the original time-domain waveform did not show any obvious abnormalities. After wavelet packet decomposition, the time series sequences of energy in each frequency band can clearly separate the vibration modes caused by the loosening of different mechanical parts, providing richer diagnostic information compared to a single overall vibration level trend. After normalization and moving average filtering, the oil chromatographic component content data effectively suppresses random noise caused by ambient temperature fluctuations or sampling errors, making the long-term trend of gas concentration ratio smoother and more stable, which is beneficial for capturing subtle early fault signs.

[0025] In one embodiment of the present invention, see [reference] Figure 3Based on the substation's primary wiring diagram and equipment physical layout diagram, an equipment connection topology is constructed, with power equipment as nodes and electrical connections and physical proximity between equipment as edges. According to the signal propagation characteristics and sampling periods of equipment infrared thermal images, partial discharge ultrasonic signals, equipment mechanical vibration waveforms, and oil chromatographic component content, the delay time required for different types of state data to be transmitted along the edges of this equipment connection topology is calculated. For each connection edge in the equipment connection topology, the delay time determined by its corresponding signal propagation characteristics is obtained. A reference time window is set, and the delay ratio coefficient relative to the reference time window is calculated. The initial weight value of each edge in the equipment connection topology is multiplied by the reciprocal of the corresponding delay ratio coefficient to obtain the time-adjusted edge weight value. A distance attenuation factor is introduced, calculated based on the physical distance between equipment, and the time-adjusted edge weight value is further multiplied by this distance attenuation factor. The final weight values ​​of all edges are normalized to ensure that the sum of all weight values ​​is a fixed constant. Using equipment as nodes and the normalized final weight values ​​as edge weights, a weighted adjacency matrix of the time-aware equipment state association graph is constructed. The structured device state time-series data stream is used as the input node features and fed into an improved deep residual graph convolution algorithm with the time-aware device state association graph as the graph structure for iterative convolution operations. After each convolution operation, residual connections are introduced to add the output of the current layer to the output of the previous layer in a skip manner. After multiple layers of graph convolution and residual connections, the features of all nodes are aggregated in the last layer to form a fusion feature tensor that represents the overall health of the device.

[0026] In practical implementation, based on the substation's primary wiring diagram and equipment physical layout diagram, power equipment such as transformers, circuit breakers, and disconnectors are abstracted as graph nodes, and the direct electrical connections between equipment (such as busbar connections and bushing leads) and physical spatial proximity relationships (such as equipment installed on the same bracket) are abstracted as edges, constructing an undirected equipment connection topology graph. For four types of data—equipment infrared thermal images, partial discharge ultrasonic signals, equipment mechanical vibration waveforms, and oil chromatographic component content—the differences in their respective signal propagation media and paths are analyzed, and the required transmission delay time along the edges of the equipment connection topology graph is calculated: equipment infrared thermal images rely on electromagnetic wave propagation, with a minimal delay that can be approximated as zero; partial discharge ultrasonic signals propagate at approximately 340 m / s in air, requiring millisecond-level delay calculations based on equipment spacing; equipment mechanical vibration waveforms propagate along metal structures at approximately 5000 m / s, resulting in microsecond-level delays; oil chromatographic component content reflects the diffusion process of dissolved gases in oil, a slowly varying parameter, requiring a fixed sampling period lag.

[0027] In some embodiments, the adjacency matrix construction of the time-aware device state association graph relies on dynamic adjustment of edge weights. For any edge in the device connection topology graph... Obtain the delay time determined by its corresponding signal propagation characteristics. Set the reference time window (For example, 1 second), calculate the delay ratio. Connect the devices to the edges in the topology diagram. initial weights Multiply by the reciprocal of the delay scaling factor to obtain the time-adjusted edge weights. Introducing a method based on the physical Euclidean distance between devices. Calculated distance attenuation factor (in (where the distance scale is constant), the time-adjusted edge weights are further modified to The corrected weights of all edges are normalized so that the sum of all weights equals the total number of nodes, ultimately yielding the weighted adjacency matrix of the time-aware device state association graph.

[0028] It is understandable that the feature extraction performance of the improved deep residual graph convolution algorithm is closely related to the graph structure. In specific implementation, a structured equipment state time-series data stream is used as the input node feature matrix, which is input into an improved deep residual graph convolution network with a time-aware equipment state association graph as the graph structure. Each layer of the network performs graph convolution operations, aggregating the state information of adjacent nodes, and introducing residual connections after each convolution, adding the output features of the current layer to the output features of the previous layer element-wise. Through multi-layer stacking, the network gradually fuses local and global topological association information, and gathers the high-dimensional features of all nodes at the output layer to form a fusion feature tensor representing the overall health of the equipment. Compared with traditional graph convolution that ignores signal propagation delay, the time-aware graph structure constructed in this embodiment is more in line with the physical signal transmission law of substations and can more accurately characterize the spatiotemporal correlation of equipment states.

[0029] In one embodiment of the present invention, the multi-task joint learning model includes a shared feature encoder and three independent task decoders. The shared feature encoder performs high-order abstract encoding on the input fused feature tensor. The output of the shared feature encoder is simultaneously input to the abnormal state identification decoder, the fault type classification decoder, and the remaining lifetime prediction decoder. The abnormal state identification decoder receives the feature vector output from the shared feature encoder. The fully connected layer inside the abnormal state identification decoder performs a linear transformation on the feature vector, mapping the high-dimensional features to a one-dimensional value. The one-dimensional value after the linear transformation by the fully connected layer is non-linearly mapped using the Sigmoid activation function, constraining the mapping result to the interval between zero and one. This mapping result is the probability that the device is in an abnormal state. The calculated probability that the device is in an abnormal state is compared with a preset alarm threshold in real time. When the probability that the device is in an abnormal state continues to be greater than the preset alarm threshold for more than a preset time window, an abnormal alarm message containing a device identifier, an abnormal probability value, and a timestamp is generated. The fault type classification decoder generates a fault category by outputting the fault type and its corresponding confidence level through a classification layer containing a Softmax function; the remaining lifetime prediction decoder generates a predicted lifetime index by outputting the predicted time for the device to operate from its current state to a predetermined failure threshold through a regression layer.

[0030] In practice, the multi-task joint learning model consists of a shared feature encoder, an abnormal state recognition decoder, a fault type classification decoder, and a remaining life prediction decoder. The shared feature encoder employs a multilayer perceptron structure to perform high-order nonlinear transformations and abstract encoding on the input fused feature tensor, extracting a discriminative and general feature representation. The output feature vector of the shared feature encoder is simultaneously fed into the three task decoders to achieve collaborative analysis and inference of the equipment state.

[0031] In some embodiments, the anomaly detection decoder is responsible for generating device anomaly probabilities and alarm information. The anomaly detection decoder receives a 256-dimensional feature vector from the shared feature encoder output and performs a linear transformation through a fully connected layer containing 128 neurons, compressing the high-dimensional features into one-dimensional real values. A sigmoid activation function is applied to the one-dimensional value output by the fully connected layer, constraining the mapping result to the (0,1) interval to obtain the probability value of the device being in an abnormal state. The calculation formula is: .

[0032] in: For the input feature vector, This is the weight matrix. For bias terms, Use the Sigmoid function. Set the alarm threshold. The probability is 0.85 when the equipment is in an abnormal state. More than 10 consecutive sampling periods greater than When the device is in an abnormal state, an abnormal alarm message containing the device ID, current probability value, and timestamp is generated.

[0033] Table 1: Output Table of Fault Type Classification Decoder

[0034] It is understandable that the fault type classification decoder and the remaining life prediction decoder work in parallel to complete the comprehensive assessment. The fault type classification decoder receives the output of the shared feature encoder and calculates the probability distribution of the device belonging to each fault type through a classification layer with a softmax function. Referring to Table 1, it outputs the fault category and its corresponding confidence score. The remaining life prediction decoder maps abstract features to predicted time values ​​from the current moment of device operation to the performance degradation failure threshold through a linear regression layer, generating a predicted life index. The outputs of the three decoders together constitute a comprehensive description of the device's health status, providing multi-dimensional judgment criteria for subsequent maintenance decisions.

[0035] In one embodiment of the present invention, a training sample set is constructed by collecting historical structured equipment state time-series data streams and their corresponding historical anomaly labels, historical fault type labels, and historical remaining lifetime labels. This training sample set is input into the multi-task joint learning model, where the shared feature encoder extracts features from the input data. The anomaly state identification decoder calculates the anomaly state probability based on the extracted features, the fault type classification decoder calculates the fault type distribution, and the remaining lifetime prediction decoder calculates the predicted lifetime. During the model training phase, a batch of data is obtained from the training sample set, including the structured equipment state time-series data stream, the corresponding historical anomaly labels, historical fault type labels, and historical remaining lifetime labels. This structured equipment state time-series data stream is input into the multi-task joint learning model to obtain the anomaly state probability output by the anomaly state identification decoder, the fault type probability distribution output by the fault type classification decoder, and the predicted lifetime value output by the remaining lifetime prediction decoder. For the output of the anomaly state recognition decoder, a binary cross-entropy loss function is used to compare the output anomaly state probability with the historical anomaly label, calculating the cross-entropy loss between the anomaly state recognition decoder output and the historical anomaly label. For the output of the fault type classification decoder, a multivariate cross-entropy loss function is used to compare the output fault type probability distribution with the historical fault type label, calculating the cross-entropy loss between the fault type classification decoder output and the historical fault type label. For the output of the remaining lifetime prediction decoder, a mean squared error loss function is used to compare the output predicted lifetime value with the historical remaining lifetime label, calculating the mean squared error loss between the remaining lifetime prediction decoder output and the historical remaining lifetime label. These three losses are weighted and summed to obtain the multi-task joint loss. A backpropagation algorithm is used to minimize this multi-task joint loss, while simultaneously updating the network parameters of the shared feature encoder and the three task decoders until the model converges.

[0036] In practice, structured equipment status time-series data streams stored in the substation over the past three years are collected and associated with labeled historical anomaly tags (0 for normal, 1 for abnormal), historical fault type tags (such as winding overheating, insulation dampness, etc.), and historical remaining life tags (in hours), forming a training sample set containing 10,000 samples. The training sample set is divided into a training set and a validation set in an 8:2 ratio. The training set is used to update network parameters, and the validation set is used to monitor whether overfitting occurs during the training process.

[0037] In some embodiments, the training process of the multi-task joint learning model involves the independent calculation and joint optimization of three loss functions. Thirty-two data records are read in batches from the training sample set. Each record includes structured multidimensional time-series data, corresponding historical anomaly labels, historical fault type labels, and historical remaining lifetime labels. The structured multidimensional time-series data is input into the shared feature encoder of the multi-task joint learning model for feature extraction. The output feature vector of the shared feature encoder is simultaneously fed into the anomaly state identification decoder, the fault type classification decoder, and the remaining lifetime prediction decoder. The anomaly state identification decoder outputs a single anomaly state probability value, the fault type classification decoder outputs a discrete distribution vector containing the probabilities of various fault types, and the remaining lifetime prediction decoder outputs the predicted remaining lifetime value (see Table 2).

[0038] Table 2: Single-batch training data table for multi-task joint learning model

[0039] For the output of the anomaly state recognition decoder, a binary cross-entropy loss function is used to calculate the anomaly state recognition loss, and the anomaly state probability output by the decoder is compared with historical anomaly labels. For the output of the fault type classification decoder, a multivariate cross-entropy loss function is used to calculate the fault type classification loss, and the output fault type probability distribution is compared with historical fault type labels. For the output of the remaining lifetime prediction decoder, a mean squared error loss function is used to calculate the remaining lifetime prediction loss, and the output predicted lifetime value is compared with historical remaining lifetime labels. The three calculated losses are then weighted and summed to obtain the total multi-task joint loss. : .

[0040] in: The binary cross-entropy loss represents anomaly detection. Multivariate cross-entropy loss representing fault type classification, The mean squared error loss represents the remaining lifetime prediction. , , These are the balancing weight coefficients for each type of loss.

[0041] It is understandable that updating the model parameters depends on the gradient calculation of the entire network structure using the backpropagation algorithm. To minimize the joint loss of the multi-task system, the Adam optimization algorithm is used to simultaneously update the weights and biases of the shared feature encoder, anomaly detection decoder, fault type classification decoder, and remaining lifetime prediction decoder. The training process is set to a maximum of 200 iterations. When the validation set loss no longer decreases after 10 consecutive training epochs, training is terminated early, and the optimal model parameters are saved to ensure that the multi-task joint learning model possesses accurate anomaly detection, fault classification, and lifetime prediction capabilities.

[0042] In one embodiment of the present invention, a decision knowledge base is constructed, which stores standard maintenance strategies, required spare parts lists, estimated maintenance man-hours, and operation control logic under different fault categories and different predicted life indicators. When an abnormal alarm is received, a decision process is triggered, and the decision knowledge base is queried according to the fault category and predicted life indicator to obtain a baseline maintenance strategy and resource requirements. Combined with the current maintenance resource inventory of the substation and the maintenance team's shift schedule, the feasibility of the baseline maintenance strategy and resource requirements is checked and conflict detection is performed. Based on the check results, the time window is adjusted or resources are replaced for maintenance tasks with resource conflicts to generate a feasible maintenance task execution plan. Based on the predicted life indicator and the current operating load of the equipment, the recommended operating power limit or recommended downtime of the equipment is dynamically calculated to form the equipment operation control command. The feasible maintenance task execution plan, the adjusted resource requirements, and the equipment operation control command are integrated to output the final dynamic maintenance decision scheme.

[0043] In practical implementation, a decision knowledge base is constructed and stored in a relational database. This base includes standard maintenance strategy entries, required spare parts lists, estimated maintenance man-hours, and operational control logic rules for different fault categories and predicted life indicators. When the monitoring system receives an anomaly alarm, it immediately triggers a decision-making process. Based on the fault category identifier and predicted life indicator value carried in the anomaly alarm, the decision knowledge base is queried to obtain the corresponding baseline maintenance strategy description and resource requirement list. Combining the current substation maintenance resource inventory and maintenance team scheduling, the feasibility of the baseline maintenance strategy and resource requirements is verified, and resource conflict detection is performed to identify any spare parts shortages or manpower deficiencies.

[0044] In some embodiments, the generation of maintenance task scheduling and operation control instructions depends on specific calculation and adjustment logic. Based on the feasibility verification results, maintenance tasks with resource conflicts are shifted backward in time windows or alternative resource solutions with equivalent functions are sought to generate a feasible maintenance task execution plan. Based on the predicted lifespan index value and the current operating load level of the equipment, the recommended upper limit of operating power or recommended downtime is dynamically calculated. For high-risk equipment with a short predicted lifespan index, the formula for calculating the recommended upper limit of operating power is as follows: .

[0045] in: The representative suggested an upper limit for operating power. Represents the rated power of the equipment. This represents the cumulative operating time of the device in its current state. This represents a value indicating a life expectancy. This represents the safety adjustment index. It integrates feasible maintenance task execution plans, adjusted resource requirements, and equipment operation control command parameters, encapsulating them into a complete dynamic maintenance decision-making solution document and outputting it to the substation operation and maintenance management system.

[0046] Understandably, the design of the decision knowledge base directly impacts the accuracy and response speed of maintenance decisions. Each rule in the decision knowledge base is summarized and categorized by domain experts based on a historical fault case database, covering handling strategies for various typical fault modes such as winding overheating, insulation dampness, and partial discharge. Compared to traditional static maintenance schedules, dynamic maintenance decision-making schemes can adaptively adjust based on real-time equipment health status scores and resource availability. When predicted life indicators show that the equipment's remaining lifespan is less than 72 hours and the fault category is high-risk, the decision-making scheme will prioritize allocating emergency repair channels and recommend temporary measures to reduce load operation, ensuring maximum grid safety before maintenance resources arrive.

[0047] The above are merely preferred embodiments of the present invention and are not intended to limit the present invention in any other way. Any person skilled in the art may make changes or modifications to the above-disclosed technical content to create equivalent embodiments that can be applied to other fields. However, any simple modifications, equivalent changes, and modifications made to the above embodiments based on the technical essence of the present invention without departing from the scope of the present invention shall still fall within the protection scope of the present invention.

Claims

1. An artificial intelligence-based intelligent monitoring method for substation equipment, characterized in that, The method includes: periodically collecting multi-source heterogeneous equipment status data from a sensor network deployed at monitoring points of various equipment in a substation; the multi-source heterogeneous equipment status data includes equipment infrared thermal images, partial discharge ultrasonic signals, equipment mechanical vibration waveforms, and oil chromatographic component content; preprocessing and temporally aligning the multi-source heterogeneous equipment status data to generate a structured equipment status time-series data stream; using an improved deep residual graph convolution algorithm to extract equipment status fusion features from the structured equipment status time-series data stream, extracting a fusion feature tensor characterizing the overall health of the equipment; the improved deep residual graph convolution algorithm optimizes the construction of the adjacency matrix based on the equipment physical connection topology and signal propagation delay; inputting the fusion feature tensor into a pre-trained multi-task joint learning model; the multi-task joint learning model simultaneously performs abnormal state identification, fault type classification, and remaining life prediction tasks, and outputs abnormal alarms, fault categories, and predicted life indicators; and generating a dynamic maintenance decision scheme including maintenance priority ranking, maintenance resource allocation suggestions, and equipment operation control commands based on the abnormal alarms, fault categories, and predicted life indicators. 2.The substation equipment intelligent monitoring method based on artificial intelligence according to claim 1, characterized in that, The preprocessing and time-series alignment of the multi-source heterogeneous device status data generates a structured device status time-series data stream, including: segmenting hot spot regions and extracting temperature matrices from the infrared thermal images of the devices to generate a time-series sequence of the device surface temperature distribution matrix; performing time-frequency transformation and feature wave extraction on the partial discharge ultrasonic signals to generate a time-series sequence of discharge pulse intensity and frequency band energy; performing wavelet packet decomposition on the mechanical vibration waveforms of the devices to extract energy features of each frequency band and generate a vibration energy spectrum time-series sequence; performing normalization and moving average filtering on the oil chromatographic component content data to generate a smooth time-series sequence of the concentration ratios of each gas component; establishing a unified timestamp alignment benchmark, interpolating and synchronizing the device surface temperature distribution matrix time-series sequence, the discharge pulse intensity and frequency band energy time-series sequence, the vibration energy spectrum time-series sequence, and the smooth time-series sequence of the concentration ratios of each gas component according to the unified timestamp alignment benchmark, and encapsulating them into the structured device status time-series data stream. 3.The substation equipment intelligent monitoring method based on artificial intelligence according to claim 1, characterized in that, The improved deep residual graph convolution algorithm is invoked to extract equipment state fusion features from the structured equipment state time-series data stream, extracting a fusion feature tensor characterizing the overall health of the equipment. This includes: constructing an equipment connection topology graph with power equipment as nodes and electrical connections and physical proximity between equipment as edges, based on the substation's primary wiring diagram and equipment physical layout diagram; calculating the delay time required for different types of state data to be transmitted along the edges of the equipment connection topology graph based on the signal propagation characteristics and sampling periods of the equipment's infrared thermal image, partial discharge ultrasonic signal, equipment mechanical vibration waveform, and oil chromatographic component content; and dynamically adjusting the weights of each edge in the equipment connection topology graph based on the delay time. The system adjusts the state to generate a time-aware device state association graph. The adjacency matrix element values ​​of the time-aware device state association graph are jointly determined by the inter-device connection relationship and data transmission delay. The structured device state time-series data stream is used as the input node features and fed into an improved deep residual graph convolution algorithm with the time-aware device state association graph as the graph structure for iterative convolution operations. After each convolution operation, residual connections are introduced to skip the summation of the output of the current layer and the output of the previous layer to alleviate the gradient vanishing problem in deep network training. After multiple layers of graph convolution and residual connections, the features of all nodes are aggregated in the last layer to form the fusion feature tensor representing the overall health of the device. 4.The method of claim 1, wherein, The fused feature tensor is input into a pre-trained multi-task joint learning model, which simultaneously performs abnormal state identification, fault type classification, and remaining lifetime prediction tasks. The multi-task joint learning model includes a shared feature encoder and three independent task decoders. The shared feature encoder performs high-order abstract encoding on the input fused feature tensor. The output of the shared feature encoder is simultaneously input to the abnormal state identification decoder, the fault type classification decoder, and the remaining lifetime prediction decoder. The abnormal state identification decoder calculates the probability of the device being in an abnormal state using a fully connected layer and a sigmoid activation function, and compares the probability with a preset threshold to generate the abnormal alarm. The fault type classification decoder outputs the fault type and its corresponding confidence level using a classification layer containing a softmax function to generate the fault category. The remaining lifetime prediction decoder outputs the predicted time for the device to operate from its current state to a predetermined failure threshold using a regression layer to generate the predicted lifetime index.

5. The method of claim 4, wherein the method further comprises: The pre-training process of the multi-task joint learning model includes: collecting historical structured equipment state time-series data streams and their corresponding historical anomaly labels, historical fault type labels, and historical remaining lifetime labels to form a training sample set; inputting the training sample set into the multi-task joint learning model, whereby the shared feature encoder extracts features from the input data; calculating the anomaly state recognition decoder based on the extracted features, calculating the fault type classification decoder for fault type distribution, and calculating the remaining lifetime prediction decoder for predicted lifetime; calculating the cross-entropy loss between the output of the anomaly state recognition decoder and the historical anomaly labels, calculating the cross-entropy loss between the output of the fault type classification decoder and the historical fault type labels, and calculating the mean squared error loss between the output of the remaining lifetime prediction decoder and the historical remaining lifetime labels; weighted summing of the three losses to obtain the multi-task joint loss; and using a backpropagation algorithm to minimize the multi-task joint loss while simultaneously updating the network parameters of the shared feature encoder and the three task decoders until the model converges.

6. The intelligent monitoring method for substation equipment based on artificial intelligence according to claim 4, characterized in that, Based on the aforementioned abnormal alarms, fault categories, and predicted life indicators, a dynamic maintenance decision-making scheme is generated, including maintenance priority ranking, maintenance resource allocation suggestions, and equipment operation control instructions. This includes: constructing a decision knowledge base that stores standard maintenance strategies, required spare parts lists, estimated maintenance man-hours, and operation control logic under different fault categories and predicted life indicators; when an abnormal alarm is received, a decision-making process is triggered, querying the decision knowledge base based on the fault category and predicted life indicator to match a baseline maintenance strategy and resource requirements; combining the substation's current maintenance resource inventory and maintenance team scheduling plan, the feasibility of the baseline maintenance strategy and resource requirements is verified and conflict detected; based on the verification results, time windows are adjusted or resources are replaced for maintenance tasks with resource conflicts, generating a feasible maintenance task execution plan; based on the predicted life indicator and the equipment's current operating load, the recommended operating power limit or recommended downtime of the equipment is dynamically calculated, forming the equipment operation control instructions; and the feasible maintenance task execution plan, adjusted resource requirements, and equipment operation control instructions are integrated to output the final dynamic maintenance decision-making scheme.

7. The intelligent monitoring method for substation equipment based on artificial intelligence according to claim 2, characterized in that, The establishment of a unified timestamp alignment benchmark, which involves interpolating and synchronizing the time series sequences of the equipment surface temperature distribution matrix, the discharge pulse intensity and frequency band energy, the vibration energy spectrum, and the smoothed time series sequences of the concentration ratios of various gas components according to the unified timestamp alignment benchmark, includes: selecting the master clock of the substation monitoring system as the unified timestamp alignment benchmark; establishing a mapping relationship between the original acquisition timestamp and the unified timestamp alignment benchmark for each time series; defining a target sampling frequency under the unified timestamp alignment benchmark; resampling the data points of each sequence to the target sampling frequency using a linear interpolation method for the equipment surface temperature distribution matrix, the discharge pulse intensity and frequency band energy, and the vibration energy spectrum; resampling the smoothed time series sequences of the concentration ratios of various gas components to the target sampling frequency using a nearest neighbor interpolation method due to their slow changes; and combining all sequence data resampled to the same time axis and the same sampling frequency according to the equipment identifier to form the structured equipment status time series data stream.

8. The intelligent monitoring method for substation equipment based on artificial intelligence according to claim 3, characterized in that, The weights of each edge in the device connection topology graph are dynamically adjusted based on the delay time to generate a time-aware device state association graph. This includes: for each connection edge in the device connection topology graph, obtaining the delay time determined by its corresponding signal propagation characteristics; setting a reference time window and calculating the delay ratio coefficient relative to the reference time window; multiplying the initial weight value of each edge in the device connection topology graph by the reciprocal of the corresponding delay ratio coefficient to obtain the time-adjusted edge weight value; introducing a distance attenuation factor, calculated based on the physical distance between devices, and further multiplying the time-adjusted edge weight value by the distance attenuation factor; normalizing the final weight values ​​of all edges to ensure that the sum of all weight values ​​is a fixed constant; and constructing a weighted adjacency matrix of the time-aware device state association graph using devices as nodes and the normalized final weight values ​​as edge weights.

9. The intelligent monitoring method for substation equipment based on artificial intelligence according to claim 4, characterized in that, The abnormal state recognition decoder calculates the probability that the device is in an abnormal state through a fully connected layer and a sigmoid activation function, and compares the probability with a preset threshold to generate the abnormal alarm. The process includes: the abnormal state recognition decoder receiving a feature vector output from the shared feature encoder; the fully connected layer inside the abnormal state recognition decoder performing a linear transformation on the feature vector, mapping the high-dimensional features to a one-dimensional value; applying a sigmoid activation function to the one-dimensional value after the linear transformation, performing a non-linear mapping, and constraining the mapping result to the interval between zero and one, the mapping result being the probability that the device is in an abnormal state; comparing the calculated probability of the device being in an abnormal state with a preset alarm threshold in real time; and generating an abnormal alarm message containing a device identifier, an abnormal probability value, and a timestamp when the probability of the device being in an abnormal state continuously exceeds the preset alarm threshold for more than a preset time window.

10. The intelligent monitoring method for substation equipment based on artificial intelligence according to claim 5, characterized in that, The calculation of the cross-entropy loss between the output of the abnormal state identification decoder and historical abnormal labels, the calculation of the cross-entropy loss between the output of the fault type classification decoder and historical fault type labels, and the calculation of the mean squared error loss between the output of the remaining lifetime prediction decoder and historical remaining lifetime labels includes: during the model training phase, acquiring a batch of data from the training sample set, including the structured equipment state time-series data stream, corresponding historical abnormal labels, historical fault type labels, and historical remaining lifetime labels; inputting the structured equipment state time-series data stream into the multi-task joint learning model to obtain the abnormal state probability output by the abnormal state identification decoder, the fault type probability distribution output by the fault type classification decoder, and the mean squared error loss output by the remaining lifetime prediction decoder. The following steps are performed: Measure lifetime value; For the output of the abnormal state identification decoder, use a binary cross-entropy loss function to compare the output abnormal state probability with the historical abnormal labels, and calculate the cross-entropy loss between the output of the abnormal state identification decoder and the historical abnormal labels; For the output of the fault type classification decoder, use a multivariate cross-entropy loss function to compare the output fault type probability distribution with the historical fault type labels, and calculate the cross-entropy loss between the output of the fault type classification decoder and the historical fault type labels; For the output of the remaining lifetime prediction decoder, use a mean squared error loss function to compare the output predicted lifetime value with the historical remaining lifetime labels, and calculate the mean squared error loss between the output of the remaining lifetime prediction decoder and the historical remaining lifetime labels.

Images