Substation equipment temperature intelligent sensing method and system

Abstract

The invention discloses a substation equipment temperature intelligent sensing method and system. The content comprises data acquisition, adaptive preprocessing, graph pre-training, temperature sensing model construction and temperature intelligent sensing. The invention relates to the technical field of substation equipment management, in particular to a substation equipment temperature intelligent sensing method and system.According to the scheme, firstly, a multi-channel covariance is estimated in real time through an adaptive sliding window, a scaling factor is introduced, noise and anomaly isolation is conducted in combination with low-rank-sparse decomposition and microfilterable filtering; high-fidelity interpolation is carried out on the data with the exception removed based on a covariance structure; then, a time-varying sensor graph is constructed, structural priori is pre-trained through a graph self-encoding network, sensor deviation is dynamically compensated through residual regression in a fusion feature space, finally, a high-precision temperature sensing value and a corresponding confidence coefficient score are output, and reliable data support is provided for operation and maintenance decision making and abnormal alarm.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of substation equipment management, and particularly relates to a substation equipment temperature intelligent sensing method and system. BACKGROUND

[0002] With the continuous expansion of the power system scale and the increasing complexity of the operating environment, the temperature state of the substation equipment is directly related to the safe operation of the equipment and the stability of the power grid; in the prior art, a fixed-point temperature sensor is combined with a centralized monitoring system to periodically sample and simply threshold alarm the surface or internal temperature of the equipment; or the whole temperature field of the equipment is scanned through infrared imaging, but there are the following problems: limited measurement points, incomplete coverage, poor real-time performance, susceptible to environmental interference, and inability to quantitatively measure the confidence, etc.

[0003] In addition, the traditional temperature monitoring system often has difficulty in quickly and accurately identifying and compensating for sensor faults, sudden environmental changes or local heat dissipation abnormalities, resulting in low timeliness and accuracy of temperature anomaly early warning; in recent years, although some research attempts to introduce machine learning or simple data fusion methods to improve monitoring accuracy, but generally lacks the ability to cooperatively denoise and spatially interpolate multi-modal high-frequency data, and to deeply model the complex spatio-temporal dependence structure between sensor channels; therefore, there is an urgent need for an innovative method and system capable of realizing adaptive preprocessing, structured feature extraction and intelligent fusion estimation of multi-source sensing data of substation equipment, and providing high-precision temperature sensing and reliability evaluation in real-time online scenarios, to meet the needs of modern power grids for safe, reliable and intelligent operation and maintenance. SUMMARY

[0004] In view of the above, in order to overcome the defects of the prior art, the present application provides a substation equipment temperature intelligent sensing method and system. In view of the problems that the traditional multi-sensor temperature data preprocessing method is difficult to effectively eliminate noise, difficult to find abnormal readings in time, and difficult to accurately reconstruct missing data, the present scheme estimates the multi-channel covariance in real time through an adaptive sliding window, introduces a window stretching factor, combines low-rank-sparse decomposition to separate principal components and abnormal items, uses a differentiable filter to smooth state changes, and assists in self-supervised time and space consistency anomaly detection. For abnormal data segments, multi-modal interpolation reconstruction is performed based on the covariance structure, which realizes accurate isolation of noise and outliers and high-fidelity completion of missing data. In view of the problems that the traditional sensor fusion model is difficult to capture complex space-time dependence between channels and lacks effective structure prior, the present scheme constructs a time-varying dynamic graph with a multi-channel observation vector, calculates the correlation coefficient between channels using a fixed window and normalizes it to generate an adjacency matrix, and further uses a graph auto-encoding network for self-supervised pre-training and mask enhancement learning, which realizes deep mining and representation of the dependence structure between sensor channels. In view of the problem that the traditional temperature estimation method usually relies on a single sensor or shallow fusion, resulting in limited precision and lack of confidence evaluation, the present scheme inputs the preprocessed multi-channel features into a feature mapping network based on graph structure prior, extracts high-dimensional fusion features, combines a residual regression network to dynamically compensate for sensor system bias, and adaptively maps the residual size to output confidence scores, which realizes high-precision real-time sensing and reliable quantification of the temperature of key parts of the equipment.

[0005] The technical scheme adopted by the present application is as follows: a substation equipment temperature intelligent sensing method, which comprises the following steps:

[0006] Step S1: data acquisition, acquiring substation equipment data;

[0007] Step S2: adaptive preprocessing, using a sliding window of adaptive length to estimate the covariance of the multi-channel observation vector in real time, and decomposing it into a low-rank principal component and a sparse abnormal item; then, based on the low-rank part, a differentiable Kalman-like filter is constructed to smooth the state, and then abnormality is identified by scoring in time and space through contrastive learning; for abnormal data segments, multi-channel interpolation reconstruction is completed based on the covariance structure, and the observation signal is output;

[0008] Step S3: Graph pre-training, first, based on the multi-channel observation vector output in step S2, the time-varying Pearson correlation coefficient between each sensor channel is calculated in a fixed window and normalized to generate a dynamic adjacency matrix, and a graph is constructed, in which the nodes represent the channels and the edge weights reflect the dependence between the channels; then, a graph auto-encoder network is built, the encoder maps the normalized adjacency and channel features to node embeddings, and the decoder reconstructs the adjacency relationship through the inner product of the embeddings and linearly maps the aggregated channel features to reconstruct the channel features; finally, the system jointly optimizes the encoder and decoder parameters to learn the pre-trained model parameters that can capture the spatio-temporal dependence structure between sensors by taking the weighted sum of the adjacency reconstruction error and the feature reconstruction error as the pre-training loss;

[0009] Step S4: Temperature perception model construction, the multi-channel observation vector after adaptive preprocessing is input into the feature mapping network constructed by the graph neural network encoder and fully connected layer pre-trained in step S3; then, the residual regression network is called to predict the temperature compensation value at the current time, and the compensation value is added to the selected sensor original measurement value to generate the final perception result of the temperature of the key part of the device; finally, the system generates a confidence score through adaptive mapping according to the compensation value;

[0010] Step S5: Temperature intelligent perception, real-time acquisition of the multi-channel observation vector after adaptive preprocessing and graph fusion and input into the temperature perception model, output of the current temperature estimation and the corresponding confidence, and triggering of an abnormal alarm once the confidence is lower than a preset threshold.

[0011] Further, in step S1, the data acquisition, the point temperature values on the surface of the device are collected by a resistance temperature detector; the point temperature values inside the device are collected by a thermocouple; the air temperature, air humidity, wind speed and wind direction data are collected by an environmental sensor; the vibration acceleration values of the transformer and the switch cabinet are collected by a three-axis accelerometer; all collected data are attached with a UTC timestamp accurate to milliseconds and a unique device number.

[0012] Further, in step S2, the adaptive preprocessing, specifically including the following steps:

[0013] Step S21: Covariance fast estimation, first, based on the data collected in step S1, a multi-channel observation vector is constructed, and then within a time sliding window with a length of , the sample covariance matrix of each channel is estimated in real time, and a window stretching factor is introduced to adaptively adjust the window center;

[0014] Step S22: Sparse decomposition, the covariance matrix is decomposed into principal component noise and abnormal outliers;

[0015] Step S23: Differentiable Kalman-like filter can be constructed based on the low-rank principal components obtained by decomposition;

[0016] Step S24: Self-supervised anomaly detection, simultaneously scoring time consistency and spatial consistency anomalies under the contrast learning framework, constructing feature points, and self-supervised learning normal mode;

[0017] Step S25: Interpolation, considering the spatiotemporal covariance and adaptive window length, reconstructing the abnormal data segment using the multi-channel inter-covariance structure.

[0018] Further, in step S3, this step takes the multi-channel observation vector output in step S2 as the basis, learns the spatiotemporal dependence structure between channels through dynamic graph construction, graph neural network pre-training and self-supervised mask enhancement, and generates pre-training parameters to provide structural prior for subsequent regression. Specifically, it includes the following steps:

[0019] Step S31: Node and edge structure construction, specifically including the following steps:

[0020] Step S311: Node feature definition, let the total number of channels be P, and the channel number of each channel be , define the node set ; for the multi-channel observation vector output in step S2, at any time t, the preprocessed observation value of the pth channel is denoted as , and the observation values of all channels are combined to form a feature vector;

[0021] Step S312: Time-varying correlation weight calculation, select a fixed window length of , for any pair of channels , calculate the Pearson correlation coefficient on the window ; all correlation coefficients are combined to form an adjacency matrix;

[0022] Step S313: Adjacency matrix normalization, first define a full 1 column vector, and then normalize the adjacency matrix;

[0023] Step S32: Graph auto-encoding network pre-training, specifically including the following steps:

[0024] Step S321: Encoding layer calculation, define the encoder parameter matrix, and calculate the hidden representation matrix;

[0025] Step S322: Edge and feature reconstruction, the correlation strength between channels is reconstructed by the inner product of node embedding, and the observation features of each channel are reconstructed by weighted aggregation of embedding using normalized adjacency and linear mapping;

[0026] Step S323: pre-training loss definition, first define the pre-training loss weight; construct the joint loss of the adjacency reconstruction error and the feature reconstruction error, and obtain the parameter set by jointly minimizing all training time.

[0027] Further, in step S4, the temperature perception model is constructed, based on the parameters obtained by pre-training in step S3 and the multi-channel observation vector after pre-processing in step S2, the graph structure prior is fused and deep regression is evaluated to estimate the temperature of the key part of the device and evaluate the confidence, which specifically includes the following steps:

[0028] Step S41: multi-channel feature mapping fusion, using the structure prior parameters pre-trained by the graph neural network, the pre-processed observation values of each sensor channel are mapped to a high-dimensional feature space to capture the spatial and temporal dependence between channels;

[0029] Step S42: residual adaptive regression, based on the fused features, a residual regression network is introduced to output the temperature residual estimation;

[0030] Step S43: temperature evaluation and confidence evaluation, add the residual estimation to the reference temperature directly measured by the sensor to obtain the final temperature prediction; at the same time, the confidence score is adaptively generated according to the absolute value of the residual.

[0031] Further, in step S5, the temperature intelligent perception, real-time acquisition of the data of the substation equipment, after adaptive pre-processing and graph structure fusion, a multi-channel observation vector is obtained, and the multi-channel observation vector is input into the temperature perception model to output the temperature perception value in real time, and generate the confidence score; at the same time, a confidence score threshold is set, when the confidence score is lower than the preset confidence score threshold, the system immediately triggers the temperature abnormality alarm.

[0032] The application provides a substation equipment temperature intelligent perception system, which comprises a data acquisition module, an adaptive pre-processing module, a graph pre-training module, a temperature perception model construction module and a temperature intelligent perception module.

[0033] The data acquisition module acquires the surface point temperature value of the substation equipment, the point temperature value inside the equipment, the air temperature, the air humidity, the wind speed, the wind direction, the vibration acceleration value of the transformer and the switch cabinet, and sends the data to the adaptive pre-processing module.

[0034] The adaptive preprocessing module receives data sent by the data acquisition module, estimates the covariance of the multi-channel observation vector in real time with a sliding window of adaptive length, and decomposes it into a low-rank principal component and a sparse anomaly term; then, based on the low-rank part, a differentiable Kalman-like filter is constructed for state smoothing, and then the anomaly is identified by scoring in time and space through contrastive learning; for the data segment of the anomaly, the multi-channel interpolation reconstruction is completed by combining the covariance structure, the observation signal is output, and the data is sent to the graph pre-training module, the temperature perception model construction module and the temperature intelligent perception module;

[0035] The graph pre-training module receives data sent by the adaptive preprocessing module, calculates the time-varying Pearson correlation coefficient between each sensor channel in pairs and normalizes it to generate a dynamic adjacency matrix, and builds a graph with nodes representing channels and edge weights reflecting channel dependence; then, a graph auto-encoding network is built, the encoder maps the normalized adjacency and channel features to node embeddings, and the decoder reconstructs the adjacency relationship through the inner product of the embeddings and linearly maps the aggregated channel features reconstructed using the normalized adjacency to reconstruct the channel features; finally, the system takes the weighted sum of the adjacency reconstruction error and the feature reconstruction error as the pre-training loss, jointly optimizes the encoder and decoder parameters, learns the pre-training model parameters that can capture the spatio-temporal dependence structure between sensors, and sends the data to the temperature perception model construction module and the temperature intelligent perception module;

[0036] The temperature perception model construction module receives data sent by the adaptive preprocessing module and the graph pre-training module, and inputs the multi-channel observation vector after adaptive preprocessing into the feature mapping network; then, a residual regression network is called to predict the temperature compensation value at the current time, and the compensation value is added to the selected sensor original measurement value to generate the final perception result of the temperature of the key part of the device; finally, the system generates a confidence score through adaptive mapping according to the compensation value, and sends the data to the temperature intelligent perception module;

[0037] The temperature intelligent perception module receives data sent by the adaptive preprocessing module, the graph pre-training module and the temperature perception model construction module, obtains the multi-channel observation vector after adaptive preprocessing and graph fusion in real time and inputs it into the temperature perception model, outputs the current temperature estimate and the corresponding confidence, and triggers an abnormal alarm as soon as the confidence is lower than the preset threshold.

[0038] The above scheme has the following beneficial effects:

[0039] (1) In view of the problems that noise is difficult to effectively eliminate, abnormal readings are difficult to find in time, and missing data is difficult to accurately reconstruct in the traditional multi-sensor temperature data preprocessing method, the scheme realizes accurate isolation of noise and outliers and high-fidelity completion of missing data by using adaptive sliding window to estimate multi-channel covariance in real time, introducing window stretching factor, combining low-rank-sparse decomposition to separate principal components and abnormal items, using differentiable filter to smooth state changes, and supplementing self-supervised time and space consistency anomaly detection. Based on the covariance structure, the multi-modal interpolation reconstruction is realized for the abnormal data segment.

[0040] (2) In view of the problems that the traditional sensor fusion model is difficult to capture the complex space-time dependence between channels and lacks effective structure prior, the scheme realizes deep mining and representation of the dependence structure between sensor channels by constructing a time-varying dynamic graph with a multi-channel observation vector, calculating the correlation coefficient between channels with a fixed window and normalizing to generate an adjacency matrix, and further using a graph auto-encoding network for self-supervised pre-training and mask enhancement learning, thereby providing structure prior parameters rich in space-time association for subsequent regression models.

[0041] (3) In view of the problems that the traditional temperature estimation method usually relies on a single sensor or shallow fusion, resulting in limited precision and lack of confidence evaluation, the scheme realizes high-precision real-time perception and credibility quantification of the temperature of the key parts of the device by inputting the preprocessed multi-channel features into a feature mapping network based on graph structure prior, extracting high-dimensional fusion features, combining a residual regression network to dynamically compensate for sensor system bias, and adaptively mapping the residual size to output confidence scores, thereby providing solid data support for operation and maintenance decision and abnormal alarm. BRIEF DESCRIPTION OF DRAWINGS

[0042] Figure 1 A schematic diagram of a substation equipment temperature intelligent perception method provided by the present application;

[0043] Figure 2 A schematic diagram of a substation equipment temperature intelligent perception system provided by the present application;

[0044] Figure 3 A schematic diagram of step S2;

[0045] Figure 4 A schematic diagram of step S3;

[0046] Figure 5 A schematic diagram of step S4.

[0047] The accompanying drawings are used to provide a further understanding of the present application, and constitute a part of the specification, together with the embodiments of the present application, to explain the present application, and do not constitute a limitation of the present application. DETAILED DESCRIPTION

[0048] The technical solutions in the embodiments of the present application will be clearly and completely described below with reference to the drawings in the embodiments of the present application. Obviously, the described embodiments are only part of the embodiments of the present application, rather than all the embodiments of the present application; based on the embodiments in the present application, all other embodiments obtained by a person of ordinary skill in the art without creative work fall within the protection scope of the present application.

[0049] In the description of the present application, it should be understood that the terms "upper", "lower", "front", "back", "left", "right", "top", "bottom", "inner", "outer" and the like indicate the orientation or positional relationship based on the orientation or positional relationship shown in the drawings, and are only for the convenience of describing the present application and simplifying the description, and do not indicate or imply that the device or element referred to must have a particular orientation, be constructed and operated in a particular orientation, and therefore cannot be understood as a limitation on the present application.

[0050] Embodiment one, refer to Figure 1 The present application provides a kind of substation equipment temperature intelligent perception method, the method comprises the following steps:

[0051] Step S1: data acquisition, the surface point temperature value of substation equipment, the point temperature value in equipment interior, air temperature, air humidity, wind speed, wind direction, transformer and switchgear vibration acceleration value are collected;

[0052] Step S2: adaptive preprocessing, real-time estimate covariance with adaptive length sliding window to multi-channel observation vector, and decompose it into low rank principal component and sparse abnormal item;Then, based on the low rank part, a differentiable Kalman-like filter is constructed to smooth the state, and then the abnormality is identified by scoring in time and space through contrast learning;For the data segment of abnormality, multi-channel interpolation reconstruction is completed by combining the covariance structure, and the observation signal is output;

[0053] Step S3: graph pre-training, first, based on the multi-channel observation vector output in step S2, the time-varying Pearson correlation coefficient between each sensor channel is calculated and normalized, and a dynamic adjacency matrix is generated, and a graph is constructed, in which the nodes represent the channels and the edge weights reflect the dependence between the channels;Then, a graph auto-encoder network is built, the encoder maps the normalized adjacency and channel features to node embeddings, and the decoder reconstructs the adjacency relationship through the inner product of the embeddings and linearly maps the aggregated channel features after weighting with the normalized adjacency;Finally, the system takes the weighted sum of the adjacency reconstruction error and the feature reconstruction error as the pre-training loss, and jointly optimizes the encoder and decoder parameters, so as to learn the pre-trained model parameters that can capture the temporal and spatial dependence structure between sensors.

[0054] Step S4: Constructing a temperature-aware model, inputting the multi-channel observation vector after adaptive preprocessing into the feature mapping network constructed by the graph neural network encoder and subsequent fully connected layers pre-trained in step S3; then, calling the residual regression network to predict the temperature compensation value at the current time, and adding the compensation value to the selected sensor raw measurement value to generate the final perception result of the temperature of the key part of the device; finally, the system generates a confidence score through adaptive mapping according to the compensation value size;

[0055] Step S5: Temperature intelligent perception, real-time acquisition of the multi-channel observation vector after adaptive preprocessing and graph fusion and input into the temperature-aware model, output of the current temperature estimation and corresponding confidence, and triggering of an abnormal alarm once the confidence is lower than a preset threshold.

[0056] Embodiment Two, refer to Figure 1 This embodiment is based on the above embodiment, in step S1, the data acquisition, the point temperature value of the device surface is collected by a resistance temperature detector , the sampling frequency is 1 Hz, and the measurement accuracy is ±0.1°C; the point temperature value inside the device is collected by a thermocouple , the sampling frequency is 10 Hz, and the measurement accuracy is ±0.5°C; the air temperature , air humidity , wind speed and wind direction data are collected by an environmental sensor, the sampling frequency is 1 Hz, and the measurement accuracy is ±0.2°C, ±2% and ±0.1 m / s respectively; the vibration acceleration value of the transformer and the switch cabinet is collected by a three-axis accelerometer , the sampling frequency is 100 Hz, and the resolution is 0.01 g; all collected data are attached with UTC time stamp accurate to milliseconds and device unique number.

[0057] Embodiment Three, refer to Figure 1 and Figure 3 This embodiment is based on the above embodiment, in step S2, the adaptive preprocessing, specifically including the following steps:

[0058] Step S21: Fast Covariance Estimation, first constructing a multi-channel observation vector based on the data collected in step S1 , then estimating the sample covariance matrix of each channel in real time within a time sliding window with a length of , and introducing a window stretching factor to adaptively adjust the window center, which is represented as follows:

[0059] ;

[0060] Wherein, represents the multi-channel observation vector at time t, denotes the observation mean within the window at time t-1, denotes the observation mean within the window at time t, denotes the window scaling factor, denotes the initial window length, which is set to 100 seconds, denotes the jump sensitivity coefficient, and denote the upper and lower bound of the window, respectively, denotes the floor function, denotes the truncation function, which limits the value of to the interval and takes the nearest boundary value if it exceeds the upper or lower bound, i denotes the index of the time point, denotes the multi-channel observation vector at time i, denotes the sliding window covariance matrix, denotes the transpose symbol;

[0061] Step S22: Sparse decomposition, decompose the covariance matrix into the principal component noise and outliers, denoted as follows:

[0062]

[0063] where, denotes the low-rank matrix used to characterize the covariance of the principal component of the signal, denotes the sparse matrix that captures outliers, denotes the matrix nuclear norm, denotes the 1-norm, denotes the balance parameter;

[0064] Step S23: Differentiable filtering, based on the low-rank principal component obtained by decomposition, a differentiable Kalman-like filter is constructed, and the filter gain and state transition matrix are simultaneously learned from , denoted as follows:

[0065]

[0066] where, denotes the filter hidden state, denotes the time-varying state transition matrix obtained by mapping from the low-rank principal component, denotes the filter gain, denotes the identity matrix, denotes the sigmoid function, denotes the differentiable mapping operator;

[0067] ​​Step S24: self-supervised anomaly detection, time consistency and spatial consistency anomaly scoring are simultaneously performed in a contrast learning framework, environmental wind speed, wind direction and vibration acceleration are used to construct feature points, and normal mode is learned by self-supervised learning, which is represented as follows:

[0068]

[0069] wherein, represents the environmental feature vector at time t, represents a differentiable feature mapping network, represents the jth prototype center, represents a temperature adjustment parameter, represents the similarity score of the jth prototype, represents an exponential function with a natural constant as the base; if , time t is abnormal, represents an anomaly decision threshold;

[0070] Step S25: interpolation, considering the spatiotemporal covariance and adaptive window length, the abnormal data segment is reconstructed by the multi-channel inter-covariance structure, which is represented as follows:

[0071]

[0072] wherein, and respectively represent the index set of the observed and missing components, represents the sub-block of the covariance matrix obtained in step S21 on the row, column, represents the estimated value of the missing component obtained by interpolation.

[0073] By performing the above operation, for the problems of traditional multi-sensor temperature data preprocessing method that noise is difficult to effectively eliminate, abnormal readings are difficult to find in time, and missing data is difficult to accurately reconstruct, the present scheme estimates the multi-channel covariance by an adaptive sliding window and introduces a window stretching factor in real time, separates the principal component and the abnormal item by combining low-rank-sparse decomposition, smoothes the state change by using a differentiable filter, and assists the self-supervised time and spatial consistency anomaly detection. For the abnormal data segment, the multi-modal interpolation reconstruction is performed based on the covariance structure, so as to realize the accurate isolation of noise and outliers and the high-fidelity completion of missing data.

[0074] Embodiment four, refer to Figure 1 and Figure 4 ​​The embodiment is based on the above embodiment, in step S3, the step is based on the multi-channel observation vector output in step S2, the spatio-temporal dependence structure between the channels of each sensor is learned by dynamic graph construction, graph neural network pre-training and self-supervised mask enhancement, and pre-training parameters are generated To provide structural prior for subsequent regression, specifically including the following steps:

[0075] Step S31: node and edge structure construction, specifically including the following steps:

[0076] Step S311: node feature definition, let the total number of channels be P, and the channel number of each channel be , define the node set ; for the multi-channel observation vector output in step S2, at any time t, the pre-processed observation value of the pth channel is denoted as , the observation values of all channels are combined into a feature vector, and are denoted as follows:

[0077] ;

[0078] wherein and represent the observation vectors of the 1st and 2nd channels output in step S2, respectively, represents a feature vector composed of observation values of all channels;

[0079] Step S312: time-varying correlation weight calculation, a fixed window length of is selected, specifically 200 seconds, for any pair of channels , the Pearson correlation coefficient on the window is calculated, and is denoted as follows:

[0080] ;

[0081] wherein represents the sample mean of channel p in the window, represents the sample mean of channel q in the window, represents the correlation coefficient between channel p and channel q, which is the weight of the edge in the graph; all correlation coefficients are combined into an adjacency matrix ;

[0082] Step S313: adjacency matrix normalization, first define a full 1 column vector , take the degree matrix , then normalize the adjacency matrix, and is denoted as follows:

[0083] ;

[0084] wherein denotes the normalized adjacency matrix;

[0085] Step S32: Pre-training of the graph auto-encoder network, specifically comprising the following steps:

[0086] Step S321: Encoding layer calculation, defining the encoder parameter matrix , the hidden representation matrix is calculated at the corresponding time t ;

[0087] Step S322: Edge and feature reconstruction, reconstructing the inter-channel correlation strength through the inner product of node embedding, and using the normalized adjacency pair embedding to aggregate and linearly map to reconstruct the observed features of each channel, denoted as follows:

[0088] ;

[0089] wherein, denotes the node embedding matrix output by the encoder, each row is a d-dimensional embedding vector of node , and d represents the hidden dimension; denotes the Gram matrix of node embedding, the th element is ; denotes the reconstructed inter-channel correlation strength, and the th reconstruction value is denoted as ; denotes the trainable decoder parameters, denotes all the reconstructed channel features, and the th component is denoted as

[0090] Step S323: Pre-training loss definition, first defining the pre-training loss weight ; constructing the joint loss of adjacency reconstruction error and feature reconstruction error, and obtaining the parameter set by jointly minimizing all training time t denotes the joint loss.

[0091] ;

[0092] wherein, denotes the joint loss.

[0093] By performing the above operation, in view of the problems that the conventional sensor fusion model is difficult to capture the complex space-time dependence between channels and lacks effective structure priori, the scheme constructs a time-varying dynamic graph by using a multi-channel observation vector, calculates the correlation coefficient between channels by using a fixed window and generates an adjacency matrix by normalization, further performs self-supervised pre-training and mask enhancement learning on the graph by using a graph auto-encoding network, realizes deep mining and representation of the dependence structure between sensor channels, and thus provides structure priori parameters rich in space-time correlation for a subsequent regression model.

[0094] Embodiment five, referring to Figure 1 and Figure 5 , this embodiment is based on the above-mentioned embodiment, in step S4, the temperature perception model is constructed based on the parameters obtained by pre-training in step S3 and the multi-channel observation vector after pre-processing in step S2, the graph structure priori and deep regression are fused to estimate the temperature of the key part of the device and evaluate the confidence, specifically including the following steps:

[0095] Step S41: multi-channel feature mapping fusion, using the structure priori parameters pre-trained by the graph neural network , the pre-processed observation values of each sensor channel are mapped to a high-dimensional feature space to capture the space-time dependence between channels, which is represented as follows:

[0096] ;

[0097] Wherein, represents the fusion feature vector at time t, represents the feature mapping network based on the pre-training parameters in step S32, the internal structure of which is specifically the encoder part of the graph auto-encoding network and adds a layer of full connection mapping;

[0098] Step S42: residual adaptive regression, based on the fusion feature , a residual regression network is introduced to output temperature residual estimation, which is represented as follows:

[0099] ;

[0100] Wherein, represents the temperature residual estimation value at time t, and respectively represent the learnable residual regression weight matrix and the residual regression bias term;

[0101] Step S43: temperature evaluation and confidence evaluation, the residual estimation is added to the reference temperature directly measured by the sensor to obtain the final temperature prediction; at the same time, the confidence score is adaptively generated according to the absolute value of the residual, which is represented as follows:

[0102] ;

[0103] wherein, represents a temperature perception value at time t, represents a selected reference temperature, and the value includes a device surface point temperature value and a device internal temperature value, represents a temperature prediction confidence score, and represents a trainable parameter of the confidence mapping network.

[0104] By performing the above operation, the conventional temperature estimation method generally relies on a single sensor or shallow fusion, which results in limited precision and lack of confidence evaluation. The present scheme inputs the preprocessed multi-channel features into the feature mapping network based on the graph structure prior, extracts high-dimensional fusion features, dynamically compensates for the sensor system deviation by combining the residual regression network, and adaptively maps the residual size to output the confidence score, thereby realizing high-precision real-time perception and credible quantification of the temperature of the key parts of the device, and providing solid data support for operation and maintenance decision and abnormal alarm.

[0105] Embodiment six, referring to Figure 1 , the embodiment is based on the above-mentioned embodiment, in step S5, the temperature intelligent perception, real-time collection of substation equipment data, after adaptive preprocessing and graph structure fusion, a multi-channel observation vector is obtained, and the multi-channel observation vector is input into a temperature perception model to output a temperature perception value in real time, and a confidence score is generated; at the same time, a confidence score threshold is set, when the confidence score is lower than the preset confidence score threshold, the system immediately triggers temperature abnormal alarm.

[0106] Embodiment seven, referring to Figure 1 and Figure 2 , the embodiment is based on the above-mentioned embodiment, and the present application provides a substation equipment temperature intelligent perception system, which comprises a data acquisition module, an adaptive preprocessing module, a graph pre-training module, a temperature perception model building module and a temperature intelligent perception module;

[0107] The data acquisition module acquires the surface point temperature value of the substation equipment, the point temperature value inside the equipment, the air temperature, the air humidity, the wind speed, the wind direction, the vibration acceleration value of the transformer and the switch cabinet, and sends the data to the adaptive preprocessing module;

[0108] The adaptive preprocessing module receives data sent by the data acquisition module, uses a sliding window of adaptive length to estimate the covariance of the multi-channel observation vector in real time, and decomposes it into a low-rank principal component and a sparse anomaly term; then, based on the low-rank part, a differentiable Kalman-like filter is constructed for state smoothing, and then the anomaly is identified by scoring in time and space through contrastive learning; for the data segment of the anomaly, the multi-channel interpolation reconstruction is completed by combining the covariance structure, the observation signal is output, and the data is sent to the graph pre-training module, the temperature perception model construction module and the temperature intelligent perception module;

[0109] The graph pre-training module receives data sent by the adaptive preprocessing module, calculates the time-varying Pearson correlation coefficient between each sensor channel in pairs and normalizes it to generate a dynamic adjacency matrix, and builds a graph with nodes representing channels and edge weights reflecting channel dependence; then a graph auto-encoder network is built, the encoder maps the normalized adjacency and channel features to node embeddings, and the decoder reconstructs the adjacency relationship through embedding inner product and reconstructs the channel features through linear mapping after weighted aggregation of the normalized adjacency; finally, the system takes the weighted sum of adjacency reconstruction error and feature reconstruction error as the pre-training loss, and jointly optimizes the encoder and decoder parameters, so as to learn the pre-training model parameters that can capture the spatio-temporal dependence structure between sensors, and sends the data to the temperature perception model construction module and the temperature intelligent perception module;

[0110] The temperature perception model construction module receives data sent by the adaptive preprocessing module and the graph pre-training module, and inputs the multi-channel observation vector after adaptive preprocessing into the feature mapping network; then, a residual regression network is called to predict the temperature compensation value at the current time, and the compensation value is added to the selected sensor original measurement value to generate the final perception result of the temperature of the key part of the device; finally, the system generates a confidence score through adaptive mapping according to the compensation value, and sends the data to the temperature intelligent perception module;

[0111] The temperature intelligent perception module receives data sent by the adaptive preprocessing module, the graph pre-training module and the temperature perception model construction module, obtains the multi-channel observation vector after adaptive preprocessing and graph fusion in real time and inputs it into the temperature perception model, outputs the current temperature estimate and the corresponding confidence, and triggers an abnormal alarm as soon as the confidence is lower than the preset threshold.

[0112] It is to be understood that the terminology used herein is for the purpose of describing particular embodiments only and is not intended to be limiting; it is not intended to exclude myriad other embodiments of the present application that other presenters can develop. It is also possible, however, that only a single element can be present. Terms such as "a", "an", and "the" are not intended to refer to only a singular entity but include the general class of which a single element is only one species, unless otherwise indicated. Furthermore, the use of the terms "primary" and "secondary", "first" and "second", etc., designate different Stages in the process, and are not intended to otherwise limit the number of stages which can be employed. The terminology includes the words specifically noted above, derivatives thereof, and words of similar import. The designation of a component as "optional" indicates that the component is "optional" and can or can not be present or used in the practice of the application, but that when it is present or used, it can be used in varying embodiments of the present application.

[0113] While the embodiments of the application have been shown and described, it is to be understood that various modifications, substitutions, combinations, staging, and alterations can be made by those skilled in the art without departing from the principles and spirits of the application.

[0114] The above description of the application and its embodiments is not limiting, and the embodiments shown in the drawings are only one of the embodiments of the application, and the actual structure is not limited thereto. In general, if a person skilled in the art is inspired by it, without departing from the purpose of the application, without creative design, similar structure and embodiments of the technical solution can be developed, which should belong to the protection scope of the application.

Claims

1. A method for intelligent temperature sensing of substation equipment, characterized in that, The method includes the following steps: Step S1: Data acquisition, collecting data from substation equipment; Step S2: Adaptive preprocessing, using an adaptive length sliding window to estimate the covariance of the multi-channel observation vector in real time, and decomposing it into low-rank principal components and sparse outliers; then, based on the low-rank part, a differentiable Kalman-like filter is constructed for state smoothing, and then anomalies are identified by scoring in time and space through contrastive learning; for the anomalous data segments, multi-channel interpolation reconstruction is completed in combination with the covariance structure, and the observation signal is output. Step S3: Graph pre-training. First, based on the multi-channel observation vectors output in Step S2, the time-varying Pearson correlation coefficients between each pair of sensor channels are calculated and normalized according to a fixed window to generate a dynamic adjacency matrix. This matrix is ​​used to construct a graph where nodes represent each channel and edge weights reflect channel dependencies. Then, a graph autoencoder network is built. The encoder maps the normalized adjacency and channel features to node embeddings, while the decoder reconstructs adjacency relationships through embedding inner products and reconstructs channel features using a linear mapping after weighted aggregation of normalized adjacency. Finally, the system uses the weighted sum of adjacency reconstruction error and feature reconstruction error as the pre-training loss to jointly optimize the encoder and decoder parameters. Step S4: Construct a temperature sensing model by inputting the adaptively preprocessed multi-channel observation vector into the feature mapping network constructed using the graph neural network encoder and fully connected layer obtained in step S3; then, call the residual regression network to predict the temperature compensation value at the current moment, and add the compensation value to the original measurement value of the selected sensor to generate the final sensing result of the temperature of the key parts of the equipment; finally, the system generates a confidence score based on the compensation value through adaptive mapping. Step S5: Intelligent temperature sensing.

2. The intelligent temperature sensing method for substation equipment according to claim 1, characterized in that: In step S2, the adaptive preprocessing specifically includes the following steps: Step S21: Fast covariance estimation. First, a multi-channel observation vector is constructed based on the data collected in step S1, and then a vector of length is calculated. Within a time sliding window, the sample covariance matrix of each channel is estimated in real time, while a window scaling factor is introduced to adaptively adjust the window center. Step S22: Sparse decomposition, decomposing the covariance matrix into principal component noise and outlier components; Step S23: Differentiable filtering: Based on the low-rank principal components obtained from the decomposition, construct a differentiable Kalman-like filter; Step S24: Self-supervised anomaly detection, simultaneously scoring temporal consistency and spatial consistency anomalies under the contrastive learning framework, constructing feature points, and self-supervised learning of normal modes; Step S25: Interpolation, taking into account both spatiotemporal covariance and adaptive window length, and reconstructing abnormal data segments using a multi-channel covariance structure.

3. The intelligent temperature sensing method for substation equipment according to claim 1, characterized in that: In step S3, based on the multi-channel observation vector output in step S2, the spatiotemporal dependency structure between sensor channels is learned through dynamic graph construction, graph neural network pre-training, and self-supervised mask enhancement, generating pre-training parameters to provide structural priors for subsequent regression. Specifically, this includes the following steps: Step S31: Node and edge structure construction, specifically including the following steps: Step S311: Define node features. Let the total number of channels be P, and the channel number be... Define a set of nodes For the multi-channel observation vector output in step S2, at any time t, the preprocessed observation value of the p-th channel is denoted as... The observations from all channels are combined into a feature vector; Step S312: Calculate time-varying correlation weights, selecting a fixed window length of... For any pair of channels Calculation window The Pearson correlation coefficients are used to construct an adjacency matrix from all the correlation coefficients. Step S313: Adjacency matrix normalization. First, define a column vector of all 1s, and then normalize the adjacency matrix. Step S32: Pre-training of the graph autoencoder network, specifically including the following steps: Step S321: Calculate the coding layer, define the encoder parameter matrix, and calculate the hidden representation matrix; Step S322: Edge and feature reconstruction. The correlation strength between channels is reconstructed by the inner product of node embeddings, and the observed features of each channel are reconstructed by weighted aggregation and linear mapping of the embeddings using normalized adjacency. Step S323: Pre-training loss definition. First, define the pre-training loss weights; construct the joint loss of adjacency reconstruction error and feature reconstruction error, and minimize the joint loss for all training time steps to obtain the parameter set.

4. The intelligent temperature sensing method for substation equipment according to claim 1, characterized in that: In step S4, the construction of the temperature sensing model, based on the parameters obtained from pre-training in step S3 and the multi-channel observation vectors preprocessed in step S2, fuses graph structure priors and depth regression to estimate the temperature of key parts of the equipment and assess the confidence level, specifically includes the following steps: Step S41: Multi-channel feature mapping fusion. Using the pre-trained structural prior parameters of the graph neural network, the preprocessed observations of each sensor channel are mapped to a high-dimensional feature space to capture the spatiotemporal dependence between channels. Step S42: Residual adaptive regression. Based on the fused features, a residual regression network is introduced to output the temperature residual estimate. Step S43: Temperature assessment and confidence assessment. The residual estimate is added to the reference temperature directly measured by the sensor to obtain the final temperature prediction; at the same time, a confidence score is adaptively generated based on the absolute value of the residual.

5. The intelligent temperature sensing method for substation equipment according to claim 1, characterized in that: In step S1, the data acquisition involves collecting temperature values ​​at points on the surface of the device using a resistance temperature detector. The equipment collects point temperature values ​​inside the device using thermocouples; it collects air temperature, humidity, wind speed, and wind direction data using environmental sensors; and it collects vibration acceleration values ​​of the transformer and switchgear using a triaxial accelerometer. All collected data is accompanied by a UTC timestamp accurate to milliseconds and a unique device number.

6. The intelligent temperature sensing method for substation equipment according to claim 1, characterized in that: In step S5, the intelligent temperature sensing collects data from substation equipment in real time, obtains multi-channel observation vectors after adaptive preprocessing and graph structure fusion, and inputs the multi-channel observation vectors into the temperature sensing model to output temperature sensing values ​​in real time and generate a confidence score. At the same time, a confidence score threshold is set. When the confidence score is lower than the preset confidence score threshold, the system immediately triggers a temperature anomaly alarm.

7. A substation equipment temperature intelligent sensing system, used to implement the substation equipment temperature intelligent sensing method as described in any one of claims 1-6, characterized in that: It includes a data acquisition module, an adaptive preprocessing module, a graph pre-training module, a temperature sensing model building module, and a temperature intelligent sensing module.

8. The intelligent temperature sensing system for substation equipment according to claim 7, characterized in that: The data acquisition module collects surface temperature values, internal temperature values, air temperature, air humidity, wind speed, wind direction, and vibration acceleration values ​​of transformers and switchgear in the substation equipment, and sends the data to the adaptive preprocessing module. The adaptive preprocessing module receives data sent by the data acquisition module, uses an adaptive-length sliding window to estimate the covariance of the multi-channel observation vector in real time, and decomposes it into low-rank principal components and sparse outliers. Then, it constructs a differentiable Kalman-like filter based on the low-rank part to perform state smoothing, and then identifies anomalies by scoring in time and space through contrastive learning. For the anomalous data segments, it completes multi-channel interpolation reconstruction by combining the covariance structure, outputs the observation signal, and sends the data to the graph pre-training module, the temperature sensing model construction module, and the temperature intelligent sensing module. The graph pre-training module receives data sent by the adaptive preprocessing module, calculates and normalizes the time-varying Pearson correlation coefficients between each pair of sensor channels according to a fixed window, and generates a dynamic adjacency matrix to construct a graph in which nodes represent each channel and edge weights reflect channel dependencies. Subsequently, a graph autoencoder network is built. The encoder maps normalized adjacency and channel features to node embeddings, while the decoder reconstructs adjacency relationships through embedding inner product and reconstructs channel features by linear mapping after weighted aggregation of normalized adjacency. Finally, the system uses the weighted sum of adjacency reconstruction error and feature reconstruction error as the pre-training loss to jointly optimize the encoder and decoder parameters, and sends the data to the temperature sensing model module and the temperature intelligent sensing module. The temperature sensing model construction module receives data sent by the adaptive preprocessing module and the graph pretraining module, and inputs the multi-channel observation vector after adaptive preprocessing into the feature mapping network; Subsequently, the residual regression network is invoked to predict the temperature compensation value at the current moment, and this compensation value is added to the original measurement value of the selected sensor to generate the final sensing result of the temperature of the key parts of the equipment; finally, the system generates a confidence score through adaptive mapping based on the compensation value and sends the data to the temperature intelligent sensing module. The temperature intelligent sensing module receives data sent by the adaptive preprocessing module, the graph pretraining module, and the temperature sensing model construction module. It acquires the multi-channel observation vector after adaptive preprocessing and graph fusion in real time and sends it into the temperature sensing model. It outputs the current temperature estimate and the corresponding confidence level. Once the confidence level is lower than the preset threshold, an abnormal alarm is triggered immediately.

Images