An on-line monitoring method for gas tank of an air charging cabinet

Abstract

This invention discloses an online monitoring method for gas-insulated switchgear (GES) gas boxes, relating to the field of GES gas box technology. The method includes the following steps: dividing the GES gas box into key monitoring areas to obtain key monitoring regions; collecting thermal imaging images of the key monitoring areas under normal conditions to obtain normal thermal imaging sequences; clustering the normal thermal imaging sequences using an improved Gaussian mixture clustering algorithm to obtain normal thermal imaging clusters; collecting real-time thermal imaging sequences, load current, and ambient temperature of the key monitoring areas of the GES gas box, and inputting them into a Transformer model based on a spatiotemporal attention mechanism to output a GES gas box status label; calculating the membership probability of the real-time thermal imaging sequence to the normal thermal imaging cluster based on the GES gas box status label to obtain the GES gas box health level; and classifying the GES gas box risk level and issuing early warnings based on the GES gas box health level, thereby achieving online monitoring of the GES gas box.

Description

Technical Field

[0001] This invention relates to the field of gas-filled cabinet gas box technology, specifically to an online monitoring method for gas-filled cabinet gas boxes. Background Technology

[0002] As a core piece of equipment in the power distribution network, gas-insulated switchgear is susceptible to abnormal temperature rises in critical components such as the contact areas of moving and stationary contacts and conductive connection nodes encapsulated inside the gas tank during long-term switching operations and load operation due to poor contact and insulation aging. This directly threatens the safe and stable operation of the equipment. Therefore, capturing the thermal response characteristics of key areas through thermal condition monitoring, establishing a normal operation baseline, and promptly identifying anomalies have become core requirements for ensuring the reliable operation and maintenance of gas-insulated switchgear.

[0003] In traditional gas-insulated switchgear thermal status monitoring, a discrete point acquisition method is mainly used to establish a normal temperature baseline: several temperature sensors are arranged in key heat-generating parts inside the gas box of the gas-insulated switchgear to collect the temperature time sequence data of each point during the closing and opening operations under normal and healthy conditions of the gas-insulated switchgear. These discrete temperature data are used as the basis for the normal state to construct a temperature baseline for subsequent anomaly judgment.

[0004] However, discrete points can only obtain temperature values ​​at local single points, and cannot fully restore the overall thermal distribution characteristics of key monitoring areas and the dynamic thermal response patterns during operation. As a result, the established temperature benchmark lacks spatial integrity and temporal continuity, affecting the accuracy of subsequent temperature anomaly identification. Summary of the Invention

[0005] To address the shortcomings of existing technologies, this invention provides an online monitoring method for gas-filled cabinet gas boxes, thereby resolving the problems existing in the background technology.

[0006] To achieve the above objectives, the present invention provides the following technical solution: an online monitoring method for gas-filled gas holders, comprising the following steps: Step S1: Divide the gas chamber of the gas filling cabinet into key monitoring areas to obtain the key monitoring areas of the gas filling cabinet; Step S2: Collect thermal imaging images of the key monitoring area of ​​the gas-filled cabinet under normal conditions to obtain a normal thermal imaging sequence. Then, cluster the normal thermal imaging sequence using an improved Gaussian mixture clustering algorithm to obtain normal thermal imaging clusters. Step S3: Collect real-time thermal imaging sequences, load current, and ambient temperature of key monitoring areas of the gas-insulated switchgear, and input them into the Transformer model based on the spatiotemporal attention mechanism to output the gas-insulated switchgear gas box status label; based on the gas-insulated switchgear gas box status label, calculate the membership probability of the real-time thermal imaging sequence to the normal thermal imaging cluster to obtain the health status of the gas-insulated switchgear. Step S4: Based on the health status of the gas-filled cabinet, classify the risk level of the gas-filled cabinet and issue an early warning, thereby realizing online monitoring of the gas-filled cabinet.

[0007] Preferably, the process of collecting thermal imaging images of the key monitoring area of ​​the gas-filled cabinet under normal conditions to obtain a normal thermal imaging sequence includes the following specific steps: Acquire the t-th frame of thermal imaging image under normal and healthy operating conditions. , Represents the pixel coordinates at time t. The original temperature measurement at the location; the first The fixed set of pixels in the preset reference heatmark region in the frame is , ,in , , , For the pixel boundary values ​​of a fixed set of pixels, and These are the horizontal and vertical indexes, respectively. By applying drift correction to the thermal imaging image using a fixed set of pixels, a drift-corrected thermal imaging image is obtained, ultimately resulting in a normal thermal imaging sequence under normal and healthy operating conditions.

[0008] Preferably, the step of performing drift correction on the thermal imaging image using a fixed pixel set to obtain a drift-corrected thermal imaging image includes the following steps: Template matching algorithm in thermal imaging images The actual pixel position of the upper positioning reference heat-marked area ; Calculate the actual pixel position respectively centroid coordinates and fixed pixel set Determine the centroid coordinates and solve for the actual region. centroid coordinates and reference region Translation offset of the centroid coordinates ; thermal imaging images Perform a translation transformation to obtain a spatially aligned thermal image. ; Spatially aligned thermal images In the middle, calculate the reference area Average temperature inside Based on the benchmark area Average temperature inside Temperature drift correction is performed to obtain the final drift-corrected thermal imaging image. .

[0009] Preferably, the step of clustering normal thermal imaging sequences using an improved Gaussian mixture clustering algorithm to obtain normal thermal imaging clusters includes the following steps: Feature extraction is performed on the normal thermal imaging sequence to obtain the normal thermal imaging vector; initial Gaussian component number , Calculate the initial covariance parameter:

[0010] in, This represents the initial covariance scalar. Scaling factor Let D be the s-th normal thermal imaging vector, and let D be the feature dimension of the normal thermal imaging vector. Let S represent the trace of the matrix, where S is the total number of samples in the normal thermal imaging vector. This represents the mean of a normal thermal imaging vector. T is the transpose symbol; A point is randomly selected from the unassigned normal thermal imaging vectors and used as the initial cluster center of the k-th cluster. Calculate each normal thermal imaging vector To the initial center Mahalanobis distance:

[0011] in, express arrive Mahalanobis distance, Let T be a randomly selected feature point, and let T be the transpose sign. This represents the global covariance matrix used to calculate the initial distance; Mahalanobis distance Normal thermal imaging vectors smaller than a preset distance threshold are grouped into clusters. , This represents the normal thermal imaging cluster corresponding to the k-th Gaussian component. Calculate this normal thermal imaging cluster. vector mean sum vector covariance ; The cluster centers of the clusters are iteratively updated to obtain the initial Gaussian mixture model parameters. After obtaining the initial Gaussian mixture model parameters, the expectation-maximization algorithm is used for optimization. The following expectation and maximization steps are executed iteratively until convergence, and the parameters of the optimal Gaussian mixture model are obtained, thus obtaining the normal thermal imaging cluster.

[0012] Preferably, the iterative update of the cluster centers to obtain the initial Gaussian mixture model parameters includes the following steps: Clustering normal thermal imaging vector mean As the new cluster center, repeat the Mahalanobis distance calculation, redistribute clusters, and recalculate the vector mean. sum vector covariance The process continues until the cluster center converges, ultimately yielding the first... The vector mean, vector covariance, and weights of each Gaussian component are obtained, and then the normal thermal imaging vector is processed serially to obtain the initial Gaussian mixture model parameters.

[0013] Preferably, the process of collecting real-time thermal imaging sequences, load current, and ambient temperature of key monitoring areas of the gas-insulated switchgear and inputting them into a Transformer model based on a spatiotemporal attention mechanism to output a gas-insulated switchgear status label includes the following steps: Construct a Transformer model based on a spatiotemporal attention mechanism; train the Transformer model; Real-time thermal imaging sequences, load current, and ambient temperature of key monitoring areas of the gas-insulated switchgear are collected; the real-time thermal imaging sequences, load current, and ambient temperature are input into a Transformer model based on a spatiotemporal attention mechanism to finally obtain the gas-insulated switchgear's gas box status label.

[0014] Preferably, the construction of the Transformer model based on the spatiotemporal attention mechanism includes the following specific steps: The Transformer model based on the spatiotemporal attention mechanism consists of three parts: an embedding layer, an encoder, and a classification head. The embedding layer receives real-time thermal imaging sequences, load current, and ambient temperature to obtain the embedded input feature sequence. The encoder receives the input feature sequence and contains L encoder layers with the same structure. Each layer consists of a multi-head spatiotemporal attention module, a contextual temporal attention mechanism, and a feedforward network. It uses residual connections and layer normalization to output deep features. The classification head performs global average pooling on deep features, aggregates them into sequence-level feature vectors, and then converts them into probability distributions of gas box status labels through a fully connected layer and a Softmax function.

[0015] Preferably, training the Transformer model includes the following specific steps: The loss function for the Transformer model is:

[0016] in, Represents the total loss function. Cross-entropy loss function for classifying operation types. The cross-entropy loss function for load level classification, Cross-entropy loss function for classifying ambient temperature ranges; The Adam optimizer is used to update the model parameters, and the loss function is minimized through gradient descent to obtain a trained Transformer model.

[0017] Preferably, the step of calculating the membership probability of the real-time thermal imaging sequence to the normal thermal imaging cluster based on the gas chamber status tag of the gas cabinet to obtain the health status of the gas cabinet is as follows:

[0018] in, For the health status of the gas-filled cabinet, The number of effective Gaussian components, This is the feature vector of a real-time thermal imaging sequence. for Gaussian multivariate probability density, The weight of the k-th Gaussian component. Let be the mean vector of the k-th Gaussian component. Let be the covariance matrix of the k-th Gaussian component.

[0019] Preferably, the specific steps for classifying the risk level of the gas-filled cabinet based on its health status are as follows: Low risk: When the gas-filled cabinet is in good health A high health threshold Th2 is preset to determine that thermal behavior patterns are normal. Medium risk: When the preset low health threshold Th1 ≤ the health status of the gas-filled cabinet When the preset high health threshold Th2 is reached, a warning is triggered if the thermal pattern deviates significantly from the baseline. High risk: When the health status of the gas-filled cabinet When the preset low health threshold Th1 is reached, the thermal pattern is determined to deviate significantly from the health benchmark, triggering the highest level alarm.

[0020] This invention provides an online monitoring method for gas-filled cabinet gas boxes, involving machine learning and deep learning technologies, which has the following beneficial effects: (1) The improved Gaussian mixture clustering algorithm adopted solves the defects of traditional Gaussian mixture models, which use random initialization, depend on initial values, and are prone to getting trapped in local optima, through a serial learning initialization mechanism. Traditional Gaussian mixture models directly use random initial parameters to execute the expectation-maximization algorithm, which is very easy to converge to local optima due to improper initial values ​​and cannot accurately fit the multimodal distribution of normal thermal features. However, before the expectation-maximization iteration, this improved algorithm first divides the data clusters serially through the Mahalanobis distance threshold, and calculates the initial mean, covariance and weight of each Gaussian component based on the real data distribution, ensuring that the initial parameters completely fit the real distribution structure of the data. This avoids the local optima problem caused by random initialization from the root, greatly improves the accuracy and stability of the establishment of normal thermal imaging baseline, helps to enhance the reliability of anomaly detection and reduce the false alarm rate.

[0021] (2) The Transformer model utilizes a self-attention mechanism to globally capture long-range dependencies in thermal imaging sequences, effectively integrating the temporal information of the sequence. Through the encoder-decoder structure and multi-head self-attention computation, this model enhances feature extraction capabilities, making multi-dimensional state classification more accurate. Its advantage lies in maintaining high computational efficiency when processing high-dimensional sequence data, and enhancing the interpretability and confidence of state recognition through the output probability distribution via Softmax, providing strong classification support for online monitoring.

[0022] (3) The Transformer model based on the spatiotemporal attention mechanism further integrates spatial and temporal attention mechanisms and is specifically designed for the spatiotemporal dynamic characteristics of thermal imaging sequences. It simultaneously captures the local thermal response details and global long-term time-series evolution of thermal imaging sequences, achieving high-precision identification of closing / opening operation states. At the same time, it can be extended to adapt to multi-dimensional operating condition label identification of load levels and ambient temperature ranges, laying the foundation for accurate matching of health benchmarks and fundamentally solving the problem of false alarms caused by mismatch of operating conditions in traditional methods. Attached Figure Description

[0023] To more clearly illustrate the technical solutions of the embodiments of the present invention, the accompanying drawings used in the description of the embodiments will be briefly introduced below. Obviously, the accompanying drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0024] Figure 1 This is a flowchart illustrating the steps of an online monitoring method for a gas-filled cabinet gas box proposed in this invention. Figure 2 This is a model structure diagram of the Transformer model based on the spatiotemporal attention mechanism in the online monitoring method for gas boxes of gas holders proposed in this invention; Figure 3 This is a hierarchical diagram of the steps involved in obtaining the health status of a gas-filled cabinet in an online monitoring method for gas-filled cabinets proposed in this invention. Detailed Implementation

[0025] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0026] Please see Figures 1-3 The present invention provides a technical solution: an online monitoring method for gas tanks in gas-filled cabinets.

[0027] Step S1: Divide the gas chamber of the gas filling cabinet into key monitoring areas to obtain the key monitoring areas of the gas filling cabinet.

[0028] Delineating the critical monitoring areas of the gas-insulated switchgear's gas box is a prerequisite for all subsequent status identification and early warning analysis. In practice, based on the gas-insulated switchgear's electrical schematic diagram and structural design drawings, combined with thermal simulation analysis results under rated load, short-time withstand current, and maximum operating load conditions, key heat-generating components and main heat transfer paths are identified, determining at least one critical monitoring area. This critical monitoring area must completely cover key heat-generating components such as the contact areas of moving and stationary contacts, main conductive circuit connection nodes, and the contact areas of moving / stationary contacts of disconnecting switches and grounding switches. The core function of this step is to lay the foundation for effectively capturing the dynamic thermal response during the switching process of the conductive circuit and identifying abnormal temperature rises during operation by accurately delineating the critical monitoring areas and deploying infrared thermal imaging units.

[0029] Once the key monitoring area is determined, the infrared thermal imaging unit is fixed to the inner wall of the gas chamber or an insulating support component using a rigid mounting bracket inside the gas chamber. Its orientation and focus are finely adjusted to ensure stable coverage and long-term alignment of its field of view with the key monitoring area. The mounting bracket is made of the same metal material as the gas chamber shell and is secured with anti-loosening bolts to prevent field of view shift caused by vibrations during transportation and operation. To adapt to the sealed environment and strong electric field conditions inside the gas chamber filled with insulating gas, the infrared thermal imaging unit is encapsulated in a pressure-resistant shielded shell made of aluminum alloy or stainless steel, reliably grounded, to shield against transient electromagnetic interference and partial discharge pulse interference generated by the switching operations inside the gas chamber. The shielded shell has a cross-sectional area of ​​not less than [missing information]. The multi-strand copper core wire is reliably connected to the gas box grounding electrode, and the grounding circuit impedance is not greater than [value missing]. To ensure effective shielding and grounding safety, the outer contour adopts a smooth-transition electric field passivation structure, with all edge curvature radii not less than 5mm; for gas-filled switchgear with voltage levels of 24kV and above, all edge curvature radii are not less than 8mm to avoid local electric field concentration in strong electric field environments and to prevent the formation of sharp field strength concentration points. The optical window of the infrared thermal imaging unit must be made of a substrate compatible with insulating gases and with good infrared transmittance, preferably using calcium fluoride or zinc sulfide crystals. The surface of the substrate is coated with an infrared anti-reflection film that matches the response band of the infrared thermal imaging unit to improve infrared light transmittance. The optical window and the metal shell are hermetically sealed using a compatible sealant or metallization sintering process to ensure no gas leakage and stable window transmittance performance during long-term operation.

[0030] Within the fixed field of view of the infrared thermal imaging unit, at least one reference thermal marker unit is simultaneously installed to form a reference thermal marker area. This reference thermal marker area is constructed of a material with an emissivity of not less than 0.90 and a long-term stability better than ±1%, and is prepared using blackbody paint or alumina ceramic sheets with calibrated thermal capacity. The reference thermal marker unit is completely isolated from the main conductive circuit by an insulating support and is fixedly installed on a structural component inside the gas chamber without localized heating, ensuring that its temperature changes synchronously only with the ambient temperature inside the gas chamber. After the infrared thermal imaging unit and the reference thermal marker unit are installed, on-site benchmark calibration is performed: the infrared thermal imaging unit is calibrated for full-range temperature measurement accuracy using a standard blackbody source, the field of view benchmark calibration is completed, and the benchmark pixel coordinates of the key monitoring area and the reference thermal marker area in the thermal imaging image are recorded. A calibration file is generated and encrypted for storage, providing a unified benchmark for subsequent processing of thermal imaging sequences.

[0031] It should be noted that an information output channel permanently compatible with airtight requirements is installed on the gas-filled switchgear's gas box shell. This channel achieves electrical penetration using a through-chamber structure sealed with metal-glass insulators or laser welding. Its connection to the gas box body is verified by helium mass spectrometry leak detection to ensure sealing reliability. This allows for the stable transmission of digital signals, including status identification results, risk classification warning information, and suspected abnormal location indications generated by the infrared thermal imaging unit in subsequent steps, along with compressed thermal image feature data, to an external terminal without introducing additional leakage risks. To resist strong electromagnetic fields and partial discharge interference inside the switchgear, the transmission link uses double-shielded cables and implements differential signal transmission. All output data packets are appended with a cyclic redundancy check code before transmission, and the receiving end verifies the integrity of the check code to ensure that information is not tampered with or lost during transmission.

[0032] Step S2: Collect thermal imaging images of the key monitoring area of ​​the gas-filled cabinet under normal conditions to obtain a normal thermal imaging sequence. Then, cluster the normal thermal imaging sequence using an improved Gaussian mixture clustering algorithm to obtain normal thermal imaging clusters.

[0033] In a specific implementation of step S2 of the present invention, a thermal imaging standard baseline for the gas chamber of the gas holder under normal operating conditions is established, which serves as a comparison benchmark for subsequent status identification and anomaly warning.

[0034] Based on the reference thermal marker area that has been installed and calibrated in step S1, subsequent spatial alignment and temperature drift correction operations are performed. In specific implementation, after the gas-filled cabinet is installed and all its functions are confirmed to be normal, during a specific calibration period in the initial stage of operation, thermal imaging sequence data generated when the equipment performs standard operations under known normal and healthy conditions are collected based on the infrared monitoring foundation established in step S1. The calibration period is set as the factory commissioning phase before the gas-insulated switchgear is put into operation, lasting no less than 72 hours. The operation during the calibration period covers five typical load conditions: no-load, 25% rated load, 50% rated load, 75% rated load, and 100% rated load. The ambient temperature covers the rated operating temperature range of the equipment from -10°C to +40°C. Under each condition, no less than 30 closing operations and 30 opening operations are performed. The steady-state waiting time before each operation is no less than 30 minutes, and the stable observation time after the operation is no less than 60 minutes, to ensure that complete thermal response data including the reference steady state before operation, the transient state during operation, the temperature rise process, the post-steady state, and the cooling and recovery process are collected.

[0035] Specifically, a time index t is defined, and the thermal image corresponding to time t is: , This represents the original temperature measurement at pixel coordinates (x, y) at time t. The location of the reference thermal marker region in the baseline frame (typically selected as frame t=1) is defined as a fixed pixel set. This region is defined by a range of pixel coordinates, expressed as: ,in , , , Pixel boundary values ​​for a fixed set of pixels.

[0036] For each frame t, a template matching algorithm is used in the thermal imaging image. The actual pixel position of the reference heat-marked area in the middle. ; Calculate the actual pixel position Compared with the reference area Spatial transformation parameters, wherein the spatial transformation parameters include at least translation offset. Based on the spatial transformation parameters obtained from the solution, the original thermal image matrix is... Perform spatial transformation and calculate pixel values ​​using bilinear interpolation to obtain a spatially aligned thermal image. This operation ensures that the pixel positions of the reference hot-marked regions in all frames remain consistent with those of the base frame. The translation offset is... For actual area centroid coordinates and reference region The difference in centroid coordinates; when spatial alignment is performed using only translation transformation, the pixel value mapping relationship is: .

[0037] In a preferred embodiment, the spatial alignment primarily corrects translational offsets caused by equipment vibration and stress relaxation, achieving full-sequence spatial alignment solely through the aforementioned translational transformation. This method, while meeting the monitoring accuracy requirements under normal operating conditions of the gas-insulated switchgear, significantly reduces algorithm computation overhead, adapts to the computing power limitations of edge-embedded devices, and ensures the real-time performance and reliability of online monitoring. After the spatial transformation parameters are solved, anomaly checks are performed simultaneously on the bounding rectangle of the reference hot-marked region: when the actual region Compared with the reference area When the aspect ratio deviation of the circumscribed rectangle exceeds 2% and the area deviation exceeds 5%, it is determined to be an abnormal rotation or scaling structural deformation, directly triggering a hardware abnormality alarm, suspending the health status assessment of this monitoring sequence, and prompting maintenance personnel to conduct on-site inspection and calibration.

[0038] Spatially aligned thermal image matrix In the middle, calculate the reference area Average temperature inside ,based on Temperature drift correction is performed, and the thermal imaging image after drift correction is obtained. Each pixel value is the difference between the spatially aligned pixel temperature and the average temperature of the reference area, expressed as: It should be noted that, The average temperature of the reference thermally marked region at the same time t is determined by the fact that the reference region is completely thermally isolated from the main conductive circuit, has no local heat sources, and is made of a highly stable material with calibrated heat capacity. It changes synchronously with the internal temperature of the air chamber, and also includes the zero-point drift component of the infrared sensor itself, serving as a comprehensive quantitative benchmark for common interference terms shared by all pixels across the entire field of view; in the above formula, the original temperature measurement value of the target area pixel... Average temperature at the same time as the reference area Subtraction completely cancels out the two common interference components across the entire field of view: ambient temperature and sensor drift, resulting in the final product. The relative temperature difference between the target pixel and the reference area is related only to the local heating characteristics of the corresponding part of the pixel (x,y), thus realizing the drift correction of temperature measurement and the effective enhancement of local heating characteristics.

[0039] Drift-corrected thermal imaging images Characterizing the temperature difference of each pixel relative to the reference thermal marker area can effectively suppress common interference components caused by changes in ambient temperature and highlight the local heating characteristics of the measured object. Through the above spatial alignment and temperature drift correction steps, the infrared thermal image sequence will be spatially unified to the coordinate system of the reference frame and temperature-wise converted into a relative temperature difference representation, providing stable and reliable input data for subsequent equipment status identification and fault early warning analysis.

[0040] For each operation, record the moment the operation command was issued. and collect from arrive A sequence of thermal images within a time window, where The baseline observation time before operation should be no less than 30 seconds to ensure that a steady-state temperature baseline is acquired before operation. The observation period for post-operation stabilization should be no less than 120 seconds, for gas-insulated switchgear with voltage levels of 24kV and above. The time interval should be no less than 300 seconds to ensure complete capture of the temperature rise process and steady-state characteristics after the operation. The thermal imaging acquisition frequency is f, with a value no less than 10 Hz to ensure no loss of transient thermal response characteristics during the operation. Each thermal imaging sequence contains... Relative temperature difference image of frames .

[0041] The final normal thermal imaging sequence is obtained, which includes normal thermal imaging sequences of closing and opening operations. All normal thermal imaging sequences are marked with their corresponding operation types, and the corresponding load level and ambient temperature range are marked simultaneously. The ambient temperature range is -10°C to +40°C, with each 5°C range being a separate range.

[0042] For each operating condition label (a three-dimensional operating condition label of "operation type + load level + ambient temperature range"), the autoencoder model is trained independently for each operating condition label. All operating condition samples under each label participate in the model training, ensuring that the model can learn common health characteristics under different operating conditions. The normal thermal imaging sequence samples under each type are input into the autoencoder model trained for the corresponding operating condition label, and the output is the normal thermal imaging vector under that operating condition label. Subsequently, for all normal thermal imaging vectors, an improved Gaussian mixture clustering algorithm is applied for probabilistic clustering according to the operating condition label. A dedicated normal thermal imaging cluster and Gaussian mixture model are established for each operating condition label to accurately characterize the multimodal distribution of normal thermal features under different operating conditions.

[0043] It should be noted that the autoencoder model described is a typical encoder-decoder structure. The encoder compresses the input high-dimensional thermal image sequence into a low-dimensional, dense feature vector, aiming to capture the essential information most relevant to health status within the sequence; the decoder then attempts to reconstruct the original sequence from this feature vector. The model undergoes unsupervised training by minimizing the reconstruction error, thereby learning a nonlinear mapping from the original data space to the low-dimensional feature space. In this process, only the encoder part is used to map each normal thermal image sequence sample to a fixed-length normal thermal imaging vector.

[0044] Subsequently, an improved Gaussian mixture clustering algorithm was applied to all normal thermal imaging vectors for probabilistic clustering. The core of the improved Gaussian mixture clustering algorithm is sequential learning initialization, the steps of which include: setting the initial number of Gaussian components k, And calculate the initial covariance parameter:

[0045] in, This represents the initial covariance scalar. This is the scaling factor, which defaults to 10. Let D be the s-th normal thermal imaging vector, and let D be the feature dimension of the normal thermal imaging vector. Let S represent the trace of the matrix, where S is the total number of samples in the normal thermal imaging vector. This represents the mean of a normal thermal imaging vector. , where T is the transpose symbol.

[0046] Randomly initialize cluster centers: Randomly select a point from the currently unassigned set of normal thermal imaging vectors as the initial cluster center of the k-th cluster, denoted as . Then, calculate each normal thermal imaging vector. arrive Mahalanobis distance:

[0047] in, express arrive Mahalanobis distance, Let T be a randomly selected feature point, and let T be the transpose sign. This represents the global covariance matrix used to calculate the initial distance. It is a fixed value and consists of the product of the initial covariance scalar and the identity matrix.

[0048] Mahalanobis distance Normal thermal imaging vectors with a distance less than a preset distance threshold (the distance threshold is usually less than 3) are grouped into clusters. , This represents the normal thermal imaging cluster corresponding to the k-th Gaussian component. Calculate this cluster. vector mean sum vector covariance .

[0049] Then proceed with iterative updates: clusters vector mean As the new cluster center Repeat the above steps to calculate the Mahalanobis distance, redistribute clusters, and recalculate the vector mean. sum vector covariance The process continues until the cluster centers converge (for example, the change in the Euclidean norm between the new vector mean and the old vector mean is less than...). Finally, the number was obtained. Vector mean of Gaussian components Vector covariance and weight , .

[0050] When the number of remaining unassigned samples is less than 1% of the total number of samples, or when a new cluster that meets the sample number requirement cannot be formed, the serial learning process is terminated, and the remaining samples are removed as outliers and not included in the initial Gaussian mixture model. The initial covariance of each new cluster is constructed by multiplying the globally calculated initial covariance scalar with the identity matrix. After the samples within the cluster are assigned, the specific mean and covariance matrix are calculated based on the samples within the cluster. After the serial learning is completed, the number of normal thermal imaging vectors is ignored. The number of remaining valid clusters is the number of effective Gaussian components. The parameters of the remaining clusters are used to form the initial Gaussian mixture model parameters. .

[0051] Obtain the initial Gaussian mixture model parameters Then, the expectation-maximization algorithm is used for optimization, which involves iteratively executing the following expectation step and maximization step until convergence. The expectation step calculates each normal thermal imaging vector. The posterior probability of belonging to the k-th Gaussian component:

[0052] in, Let represent the posterior probability that the s-th normal thermal imaging vector belongs to the k-th Gaussian component. Let k be the multivariate Gaussian probability density function, where k is the index of the Gaussian component. The number of effective Gaussian components.

[0053] The mathematical expression for the multivariate Gaussian probability density is:

[0054] Where D is the dimension of the normal thermal imaging feature vector. Let C be the determinant of the covariance matrix. It is the inverse of the covariance matrix C; The maximization step updates the weights, mean vector, and covariance matrix parameters of each Gaussian component using the maximum likelihood estimation method based on the posterior probability calculated in the expectation step. The update formula is as follows: Weight update: Mean vector update: Covariance matrix update: Where S is the total number of samples in the normal thermal imaging feature vector. The posterior probability is calculated for the expectation step. The weight coefficients of each Gaussian component satisfy the normalization constraint: This ensures the validity of the probability distribution in the Gaussian mixture model.

[0055] The maximization step updates the weights, mean vector, and covariance matrix parameters of each Gaussian component using the standard maximum likelihood estimation method based on the posterior probabilities calculated in the expectation step. The expectation and maximization steps are repeated until the model parameters converge (the change in the log-likelihood function between adjacent iterations is less than a preset threshold). Finally, the parameters of the optimal Gaussian mixture model are obtained. .

[0056] Obtaining optimal parameters Then, the final clustering assignment is performed for each normal thermal imaging vector. Based on the optimal parameters after convergence Calculate its final posterior probability relative to its k Gaussian components. By comparing these posterior probabilities, each feature vector is... Divide into its posterior probability The Gaussian component corresponding to the maximum value (i.e., the k-th cluster). All normal thermal imaging vectors assigned to the same Gaussian component k. The set of elements constitutes a specific cluster, which ultimately yields the normal thermal imaging cluster.

[0057] After obtaining the normal thermal imaging clusters, the effectiveness of the clustering results is verified using the silhouette coefficient and the Bayesian Information Criterion (BIC): the silhouette coefficient must be greater than 0.5 to ensure high similarity of samples within a cluster and significant differentiation between samples between clusters; the Bayesian Information Criterion takes the effective Gaussian component number corresponding to the minimum value to ensure a balance between model fitting accuracy and generalization ability; if the verification fails, the Mahalanobis distance threshold, initial scaling factor and other parameters need to be adjusted and the clustering is re-executed until the verification passes. The final determined Gaussian mixture model is the normal thermal state baseline model for the corresponding operation type.

[0058] It should be noted that the improved Gaussian mixture clustering algorithm used in step S2 addresses the shortcomings of traditional Gaussian mixture models, which rely on random initialization, depend on initial values, and are prone to getting trapped in local optima, through a serial learning initialization mechanism. Traditional Gaussian mixture models directly use random initial parameters to execute the expectation-maximization algorithm, which is highly susceptible to converging to local optima due to improper initial values, failing to accurately fit the multimodal distribution of normal thermal features. In contrast, this improved algorithm, before the expectation-maximization iteration, serially divides the data clusters using a Mahalanobis distance threshold and calculates the initial mean, covariance, and weight of each Gaussian component based on the real data distribution. This ensures that the initial parameters perfectly match the true distribution structure of the data, fundamentally avoiding the local optima problem caused by random initialization. This significantly improves the accuracy and stability of establishing normal thermal imaging baselines, helps enhance the reliability of anomaly detection, and reduces the false alarm rate.

[0059] Meanwhile, the essence of using the Gaussian mixture clustering algorithm in step S2 lies in the fact that during normal operation of the gas-insulated switchgear, its thermal behavior will form multiple differentiated normal thermal modes due to different operating states such as closing and opening, as well as changing load currents. This results in the thermal image characteristics of the health state exhibiting a multimodal distribution in the feature space. Traditional methods based on a single threshold or unimodal distribution cannot accurately describe this complex health benchmark composed of multiple operating conditions. The Gaussian mixture model fits the data through a linear combination of multiple Gaussian components, with each Gaussian component adaptively characterizing the thermal feature distribution under a specific operating condition. This method no longer attempts to define all states with a unified "normal range," but rather acknowledges that the health benchmark itself is a probabilistic mixture composed of several subspaces, thus providing a fundamental guarantee for the subsequent accurate calculation of the membership probability of real-time data relative to this composite benchmark.

[0060] Step S3: Collect real-time thermal imaging sequences, load current, and ambient temperature of key monitoring areas of the gas-insulated switchgear, and input them into the Transformer model based on the spatiotemporal attention mechanism to output the gas-insulated switchgear gas box status label; based on the gas-insulated switchgear gas box status label, calculate the membership probability of the real-time thermal imaging sequence to the normal thermal imaging cluster to obtain the health status of the gas-insulated switchgear.

[0061] Based on the infrared monitoring foundation established in step S1, real-time thermal imaging sequences of key monitoring areas of the gas-filled cabinet are continuously collected. The collection rules for these sequences are completely consistent with those in step S2, based on the time the operation command is issued. Based on the baseline, data was collected from... arrive The continuous frame data within a fixed time window, with the same window length, acquisition frequency f, and total number of frames as the normal thermal imaging sequence in step S2, fully covers the baseline steady state before operation, the transient state during operation, and the dynamic thermal response and stabilization process after operation.

[0062] Real-time thermal imaging sequences of key monitoring areas of the gas-insulated switchgear are collected and input into a Transformer model based on a spatiotemporal attention mechanism. This Transformer model is an improved spatiotemporal attention Transformer model based on a contextual temporal attention (CTAD) mechanism, yielding gas-insulated switchgear status labels (including closing and opening operations). The input data for the spatiotemporal attention-based Transformer model is defined as real-time thermal imaging sequences preprocessed with spatial alignment and temperature drift correction identical to step S2. The spatial alignment reference frame, reference thermal marker region, translation transformation rules, and temperature drift correction formula of the real-time thermal imaging sequence are all completely consistent with the preprocessing specifications of the normal thermal imaging sequence in step S2, ensuring that the real-time data is consistent with the quantization benchmark of the health benchmark. Dimensions H and W represent the height and width of the image in pixels, respectively. This represents the relative temperature difference at pixel coordinates (x, y) at time t.

[0063] Embedding layers of the Transformer model: real-time thermal imaging sequences Construct a three-dimensional input tensor X. Where L is the sequence length. Real-time data of the current load current and the internal ambient temperature of the gas box are synchronously collected and encoded into a working condition feature vector, serving as a supplementary dimension for the model input. Tensor X first passes through an embedding layer, merging each image frame... Divide the image into N non-overlapping blocks, each block being of size N. ,in Given the total number of image patches in a single frame, each image patch is flattened into a dimension of... The vector is then mapped to a linear projection layer (fully connected layer) as... The embedding vector of dimension is used to obtain an intermediate tensor obtained by combining all image patches. To preserve spatiotemporal and operational information, learnable spatiotemporal locations are encoded. The working condition feature vector is element-wise added to the intermediate tensor. This encoding collectively represents the information of each image patch in both the temporal dimension (frame index) and the spatial dimension (the patch's position within the frame). Finally, the embedded input feature sequence is obtained. , which serves as the input to the Transformer encoder.

[0064] The embedded input feature sequence is fed into a Transformer encoder with L layers based on a context attention mechanism. Each layer l (where l = 1, 2, ..., L) of this encoder takes the output of the previous layer as its input. Each encoder layer contains a multi-head self-attention module and a feedforward network (FFN), and applies layer normalization (LayerNorm) and residual connections: First, the multi-head self-attention module passes a learnable weight matrix through a learnable weight matrix. , , Will Projected as Query, Key, and Value respectively, where , , Then, a contextual temporal attention (CTAD) mechanism is introduced to calculate the self-attention weights: ,in It is the dimension of the key vector Key. express Activation function. Multi-head attention module query. ,key ,value In feature dimension The upper part is divided into h heads. Attention is calculated independently on each head, and the results are concatenated. Finally, the attention output is passed through a linear projection layer to collectively focus on information from different subspaces. In each encoder layer, the attention output of the multi-head self-attention module is first compared with the input of that module. Residual connections are made to alleviate the vanishing gradient problem in deep networks and preserve the original information. The results of the connections are then fed into the layer normalization layer for processing to stabilize the distribution of activation values ​​and accelerate training convergence. This step finally yields the intermediate output of the layer.

[0065] Subsequently, the intermediate output undergoes a nonlinear transformation through a feedforward network. This network sequentially performs three steps: a first linear transformation, GELU nonlinear activation, and a second linear transformation. Then, residual connections and layer normalization are applied again; that is, the output of the feedforward network is added to the intermediate output and then fed into the layer normalization layer to obtain the final output of the encoder layer. Repeat this process until the last layer, and the encoder's final output is the deep feature of the last layer. .

[0066] It should be noted that the improvement of CTAD lies in the introduction of contextual attention, which distinguishes between neighboring attention and sequence attention to accurately capture the spatial-temporal dependence in thermal imaging sequences: neighboring attention focuses on the local temporal neighborhood and uses Gaussian kernels to calculate weights to capture detailed dynamics; sequence attention, on the other hand, globally models the correlation of the entire sequence to capture long-range evolutionary trends.

[0067] Deep features of encoder output First, global average pooling is performed in both spatial and temporal dimensions for aggregation: that is, features of each frame are aggregated in the spatial dimension. The features of N image patches are averaged to obtain the global representation of each frame. Then, the global representations of L frames are averaged again over time, and finally aggregated into a sequence-level feature vector representing the entire sequence. , The sequence-level feature vector z is input into the multi-label classification head, and mapped into a multi-dimensional output vector through a fully connected layer. This multi-dimensional output vector consists of three parts: the first part is a 2-dimensional operation type vector, used to calculate the state class probabilities of closing and opening; the second part is... Dimensional load level vector, The number of load levels covered in step S2 is used to calculate the category probability of the current load level; the third part is... A dimensional ambient temperature range vector, The number of ambient temperature ranges covered in step S2 is used to calculate the category probability of the current ambient temperature range.

[0068] The two-dimensional vector o is converted into a probability distribution of closing and opening categories using the Softmax function. The calculation formula is as follows:

[0069]

[0070] in and To predict the state label, Indicates that the circuit is closed. Indicates that the circuit breaker is open. , These are the elements of the two-dimensional output vector.

[0071] Synchronous Pass The function will assign load levels The dimensional vector is transformed into a category probability distribution for each load level, and the load level corresponding to the maximum probability is taken as the final predicted load level label; the ambient temperature range is then processed using the Softmax function. The dimensional vector is converted into a category probability distribution for each temperature range, and the temperature range corresponding to the maximum probability is taken as the final predicted environmental temperature label.

[0072] For the three categories of labels—operation type, load level, and ambient temperature range—preset confidence thresholds are set for each. A recommended confidence threshold of 0.95 is used for the operation type label, and 0.90 is used for the load level and ambient temperature range labels. For each label category, the category corresponding to the maximum value in its category probability distribution is taken as the candidate identification result for that label. The three-dimensional label combination of "operation type + load level + ambient temperature range" is considered valid only if the maximum category probability of all three labels is greater than or equal to the corresponding confidence threshold. This results in the gas holder / airbox status label (including operation type, load level, and ambient temperature range), which serves as the basis for subsequent health benchmark matching. If the maximum category probability of any label is lower than the corresponding confidence threshold, the label identification is deemed unreliable, indicating potential sensor malfunction or signal interference. The health status assessment is then paused until the operation is completed and the sequence is intact, at which point the status identification is re-performed to avoid health benchmark matching errors.

[0073] It should be noted that the directly collected load current and ambient temperature values ​​are continuous and real-time, and may contain fluctuations and noise. The benchmark library established in step S2 is based on discretized "load levels" and "ambient temperature ranges." The Transformer model (using thermal image sequence features and raw operating condition signals as input) can infer the most likely discretized labels, "classifying" continuous real-time signals into preset discrete ranges. This fundamentally avoids benchmark mismatch and health assessment failure caused by misjudgment of operating conditions, and is key to the high reliability of the entire method under varying operating conditions.

[0074] The model is trained using supervised learning. The load level labels of the training samples are obtained by dividing the load current collected in step S2 into preset ranges: based on the rated current of the gas-insulated switchgear. Based on the load current Corresponding no-load load level, Corresponding to 25% of the rated load level, Corresponding to 50% of the rated load level, Corresponding to 75% of the rated load level, Corresponding to 100% rated load level; the real-time input load current is mapped to the model's prediction target according to the above load level classification rules, ensuring that the load level identification during online monitoring is completely consistent with the labeling system during the training phase. The load current input during model training serves as the essential quantitative basis for load level classification, and is fused with the thermal response features of the real-time thermal imaging sequence for learning, ensuring that the model can accurately output the corresponding load level label through the correlation between thermal response features and current information.

[0075] The training samples are normal thermal imaging sequences collected in step S2, which have been labeled with multiple dimensions of "operation type + load level + ambient temperature range". All samples undergo the same preprocessing procedure as in step S2. The core training objective of the model is accurate multi-label classification of operation type, load level, and ambient temperature range. All learnable parameters of the model are optimized using a multi-label cross-entropy loss function. The loss function formula is as follows:

[0076] in, Represents the total loss function. The cross-entropy loss function for operation type classification is as follows: M is the number of training samples. This is the true label for the i-th training sample (1 for closed state and 0 for open state). Let be the predicted probability of the model for the i-th training sample, where i is the index of the training sample. The cross-entropy loss function for load level classification, The cross-entropy loss function for classifying ambient temperature ranges. , This is the weighting coefficient, with a default value of 0.5, used to balance the loss weights of each classification task.

[0077] Cross-entropy loss function for load level classification ,in This represents the total number of load level categories, corresponding to "no-load, 25% rated load, 50% rated load, 75% rated load, and 100% rated load," i.e. , Let the true label of the load level of the i-th training sample be (multi-class one-hot encoding). If the true load level of the sample is the i-th... Class, then The values ​​for the other categories are 0. For the model to determine if the i-th sample belongs to the i-th sample, ... The predicted probability of load class meets the following conditions. Cross-entropy loss function for environmental temperature range classification , This represents the total number of categories within the ambient temperature range, corresponding to "" in step S2 ,Every The rule for dividing an interval, namely , For the true label of the environmental temperature range of the i-th training sample (multi-class one-heat encoding), if the true temperature of the sample belongs to the i-th... If there are intervals, then The values ​​for the other categories are 0. For the model to determine if the i-th sample belongs to the i-th sample, ... The predicted probabilities of temperature ranges satisfy the normalization constraint. The Adam optimizer is used to update the model parameters, with the initial learning rate set to... The actual training effect can be adjusted accordingly. Adjust within the range, weight decay coefficient set to The training iterations are no less than 100 rounds. The training set and the validation set are randomly divided in a ratio of 8:2. The training is terminated early when the total loss function of the validation set does not decrease for 10 consecutive rounds. Finally, the converged model parameters are obtained.

[0078] It should be noted that this model learns the spatiotemporal structural features of the gas-insulated switchgear opening and closing operation, the relative patterns of thermal response corresponding to the load level, and the full-field temperature change features corresponding to the ambient temperature through the spatiotemporal attention mechanism, rather than the absolute level of temperature amplitude. Therefore, it still has a stable multi-label classification capability for thermal imaging sequences with local abnormal temperature rise.

[0079] During the online monitoring phase, after obtaining the multi-dimensional status labels of the gas chamber in the gas holder, the real-time thermal imaging sequence, which has undergone preprocessing identical to that in step S2, is input into the encoder part of the autoencoder of the same operation type that has been trained in step S2. The output is a real-time thermal imaging feature vector with dimensions identical to those of the normal thermal imaging vector in step S2. The real-time thermal imaging feature vector In step S2, the normal thermal imaging clusters and Gaussian mixture models that perfectly correspond to the "operation type + load level + ambient temperature range" are matched. The closed state corresponds to the model established by the closed operation sample, and the open state corresponds to the model established by the open operation sample, ensuring the benchmark matching of the health status calculation. If the real-time identified load level and ambient temperature range exceed the operating condition range covered by the calibration in step S2, an operating condition out-of-bounds warning is triggered, the current health status assessment is paused, and only real-time thermal imaging data and operation status labels are output to avoid misjudgment due to benchmark mismatch. If the real-time identified load level and ambient temperature range are within the operating condition range covered by the calibration in step S2, the membership probability relative to the selected benchmark cluster (i.e., its probability density value under the Gaussian mixture model) is calculated. This probability value intuitively represents the degree of consistency between the current thermal behavior pattern and the health benchmark.

[0080] in, The health status of the gas-insulated cabinet is determined by its numerical value and the real-time thermal imaging feature vector. The probability of membership is completely equivalent to that of clusters in normal thermal imaging, and it is a core quantitative indicator characterizing the real-time operational health status of the gas-insulated switchgear. The effective Gaussian component count is the number of effective Gaussian components retained after removing invalid small clusters in step S2, where the normal thermal imaging sequence is clustered using the improved Gaussian mixture clustering algorithm. It corresponds to the total number of effective thermal imaging clusters under normal operating conditions and is a core structural parameter of the normal operating condition Gaussian mixture model. for Gaussian multivariate probability density, quantized The single-cluster membership probability belonging to the k-th normal thermal imaging cluster is the core computational unit of the Gaussian mixture model under normal operating conditions. The feature vector of the real-time thermal imaging sequence is the only real-time input data in the calculation process of this formula. It is extracted from the real-time acquired thermal imaging sequence through the encoder part of the autoencoder trained in step S2. It has the same dimension as the normal thermal imaging vector in step S2 and is in the same feature space. The weight coefficient of the k-th Gaussian component is determined by the serial learning initialization and expectation-maximization algorithm iteration optimization in step S2. Its physical meaning is the proportion of the effective k-th normal thermal imaging cluster to the total number of normal samples, satisfying the normalization constraint. , The mean vector of the k-th Gaussian component is calculated from the mean of all normal thermal imaging vectors within the k-th normal thermal imaging cluster in step S2. After iterative optimization using the expectation-maximization algorithm, it eventually converges, defining the center position of the k-th type of thermal feature distribution under normal operating conditions. The covariance matrix of the k-th Gaussian component is calculated from the covariance of all normal thermal imaging vectors within the k-th normal thermal imaging cluster in step S2. After iterative optimization by the expectation-maximization algorithm, it finally converges, defining the shape, dispersion, and correlation between the k-th thermal feature distribution under normal operating conditions and each feature dimension.

[0081] It should be noted that, in the complete calculation of the affiliation probability and health level, apart from the real-time thermal imaging feature vector... Apart from the input data generated in real time, all other core parameters, calculation benchmarks, model structures, and matching rules are entirely derived from the normal working condition Gaussian mixture model trained based on normal samples in step S2. Without this normal model, the calculation of health status and subordination probability has no physical meaning or engineering value.

[0082] Health status of gas cabinet The higher the value, the stronger the feature vector of the real-time thermal imaging sequence. The more likely it is to originate from the normal operating state of a healthy system; conversely, a significant decrease in the probability density value means that the current state deviates from the normal baseline distribution, indicating the possibility of abnormality or damage.

[0083] This step, as the core of the real-time execution of the online monitoring method for the gas holder's gas box in this invention, builds upon the hardware acquisition foundation of step S1 and the offline health benchmark establishment results of step S2, realizing the implementation of the core process from real-time thermal imaging data acquisition to quantitative assessment of equipment health status. This step corely includes two key technical aspects: first, a spatiotemporal Transformer model based on a fused contextual attention (CTAD) mechanism to achieve high-precision identification of the gas holder's opening and closing operation status; second, the construction of a health level quantification system deeply coupled with the normal operating condition Gaussian mixture model established offline in step S2, completing an accurate quantitative assessment of the equipment's health status and providing a core basis for risk classification and early warning in the subsequent step S4.

[0084] Compared to existing gas-insulated switchgear thermal status monitoring technologies, the core novelty and creativity of this approach are reflected in the following two points: First, addressing the spatiotemporal dynamic characteristics of the thermal response during gas-insulated switchgear opening and closing operations, it innovatively employs a spatiotemporal Transformer model that integrates a contextual temporal attention (CTAD) mechanism, overcoming the limitations of traditional monitoring methods that can only perform simple state classification based on discrete temperature data. Through a dual-branch design of adjacent attention and sequential attention, it simultaneously captures the local thermal response details and global long-term temporal evolution patterns of the thermal imaging sequence, achieving high-precision identification of closing / opening operation states. Simultaneously, it can be extended to adapt to multi-dimensional operating condition label identification based on load levels and ambient temperature ranges, laying the foundation for accurate matching of health benchmarks and fundamentally solving the problem of false alarms caused by mismatched operating conditions in traditional methods. Second, a health quantification system deeply coupled with the Gaussian mixture model of normal operating conditions established offline in step S2 was constructed. This innovative approach directly defines the health level of the gas-insulated cabinet as the membership probability of the real-time thermal imaging feature vector relative to the normal thermal imaging cluster (i.e., the joint probability density value under the Gaussian mixture model), overcoming the inherent defect of traditional monitoring methods that rely on fixed temperature thresholds for anomaly judgment. This quantification method fully aligns with the multimodal thermal characteristic distribution of the gas-insulated cabinet under normal operating conditions, accurately distinguishing between "normal temperature rise caused by changes in operating conditions" and "abnormal heating caused by equipment degradation," significantly improving the accuracy and reliability of health status assessment, and achieving a leap from qualitative judgment of "whether there is an anomaly" to precise assessment of "health quantification."

[0085] Step S4: Based on the health status of the gas-filled cabinet, classify the risk level of the gas-filled cabinet and issue an early warning, thereby realizing online monitoring of the gas-filled cabinet.

[0086] This invention obtains the feature vector of a real-time thermal imaging sequence. Health status of the gas-filled cabinet Then, it is compared with the threshold determined based on the statistical analysis of the normal thermal imaging sequence in step S2 to classify the risk level: Low risk (healthy level): If If the thermal behavior pattern is determined to be normal, only the monitoring data will be recorded and no warning will be triggered.

[0087] Medium risk (warning level): If If the thermal pattern deviates significantly from the baseline, the system will trigger an early warning.

[0088] High risk (alert level): If If the thermal pattern is determined to be severely deviated from the healthy baseline, the system will trigger the highest level alarm.

[0089] Th2 is the high health threshold, and Th1 is the low health threshold. Both thresholds are determined based on the health status distribution calculated from all normal thermal imaging sequences in step S2, and are fitted using the Gumbel distribution in extreme value theory. Th2 is taken as the 5th percentile of the health status distribution of healthy samples, representing that 95% of healthy samples are healthier than this threshold. Th1 is taken as the 1st percentile of the health status distribution of healthy samples, representing that 99% of healthy samples are healthier than this threshold, ensuring a low false alarm rate in threshold determination.

[0090] When the alarm is classified as medium or high risk, the autoencoder decoder matching the real-time operation type from step S2 is used to complete the health state reconstruction and anomaly localization of the abnormal sequence, and the real-time thermal imaging feature vector extracted in step S3 is used. The input is fed into the corresponding decoder to obtain the feature vector. Reconstructed thermal imaging sequence under ideal health conditions Subsequently, the original real-time thermal imaging sequence was calculated. With reconstructed thermal imaging sequences Pixel-by-pixel reconstruction error:

[0091] in, Let be the reconstruction error at pixel coordinates (x, y) at time t. This is the original real-time thermal imaging sequence that has undergone preprocessing identical to step S2. This is the reconstructed thermal imaging sequence output by the decoder.

[0092] Next, the pixel reconstruction error sequence is analyzed in the time dimension. The maximum value is taken to generate a two-dimensional "anomaly heatmap," while pixels corresponding to the reference heatmark area defined in step S1 are removed to eliminate interference from non-monitored areas. Areas in the heatmap with significantly higher reconstruction errors than surrounding areas are suspected deterioration sites. Ultimately, the system-generated warning information includes the anomaly level, its status (closed / open), and this anomaly heatmap, thus accurately pinpointing the risk location and achieving a leap from "presence" to "where, what, and what trend," ultimately supporting proactive maintenance decision-making. All monitoring time-series data, warning records, and anomaly location heatmaps are encrypted and stored in the time-series database. The model parameters trained in steps S2 and S3 are encrypted and stored in the device's local configuration file, while providing a standard data interface to the upper-level management system. This completes the closed loop from status awareness and diagnostic decision-making to maintenance support.

[0093] It should be noted that, in this document, relational terms such as "first" and "second" are used merely to distinguish one entity or operation from another, and do not necessarily require or imply any such actual relationship or order between these entities or operations. Furthermore, the terms "comprising," "including," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or apparatus that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such a process, method, article, or apparatus. Without further limitations, the phrase "comprising an element defined as..." does not exclude the presence of other identical elements in the process, method, article, or apparatus that includes said element.

[0094] Although embodiments of the invention have been shown and described, it will be understood by those skilled in the art that various changes, modifications, substitutions and alterations can be made to these embodiments without departing from the principles and spirit of the invention, the scope of which is defined by the appended claims and their equivalents.

Claims

1. A method for online monitoring of the gas chamber of a gas-filled switchgear, characterized in that: Includes the following steps: Step S1: Divide the gas chamber of the gas filling cabinet into key monitoring areas to obtain the key monitoring areas of the gas filling cabinet; Step S2: Collect thermal imaging images of the key monitoring area of ​​the gas-filled cabinet under normal conditions to obtain a normal thermal imaging sequence. Then, cluster the normal thermal imaging sequence using an improved Gaussian mixture clustering algorithm to obtain normal thermal imaging clusters. Step S2 involves clustering normal thermal imaging sequences using an improved Gaussian mixture clustering algorithm to obtain normal thermal imaging clusters, including the following steps: Feature extraction is performed on the normal thermal imaging sequence to obtain the normal thermal imaging vector; the initial number of Gaussian components k, Calculate the initial covariance parameter: ； in, This represents the initial covariance scalar. Scaling factor Let D be the s-th normal thermal imaging vector, and let D be the feature dimension of the normal thermal imaging vector. () denotes the trace of the matrix, and S is the total number of samples in the normal thermal imaging vector. This represents the mean of a normal thermal imaging vector. T is the transpose symbol; A point is randomly selected from the unassigned normal thermal imaging vectors and used as the initial cluster center of the k-th cluster. Calculate each normal thermal imaging vector To the initial center Mahalanobis distance: ； in, express arrive Mahalanobis distance, Let T be a randomly selected feature point, and let T be the transpose sign. This represents the global covariance matrix used to calculate the initial distance; Mahalanobis distance Normal thermal imaging vectors smaller than a preset distance threshold are grouped into clusters. , This represents the normal thermal imaging cluster corresponding to the k-th Gaussian component. Calculate this normal thermal imaging cluster. vector mean sum vector covariance ; The cluster centers of the clusters are iteratively updated to obtain the initial Gaussian mixture model parameters. After obtaining the initial Gaussian mixture model parameters, the expectation-maximization algorithm is used for optimization. The expectation step and the maximization step are executed iteratively until convergence, and the parameters of the optimal Gaussian mixture model are obtained, thus obtaining the normal thermal imaging clusters. Step S3: Collect real-time thermal imaging sequences, load current, and ambient temperature of key monitoring areas of the gas-insulated switchgear, and input them into a Transformer model based on a spatiotemporal attention mechanism. The output is the gas-insulated switchgear's gas box status label. The construction of the Transformer model based on a spatiotemporal attention mechanism includes the following specific steps: The Transformer model based on the spatiotemporal attention mechanism consists of three parts: an embedding layer, an encoder, and a classification head. The embedding layer receives real-time thermal imaging sequences, load current, and ambient temperature to obtain the embedded input feature sequence. The encoder receives the input feature sequence and contains L encoder layers with the same structure. Each layer consists of a multi-head spatiotemporal attention module, a contextual temporal attention mechanism, and a feedforward network. It uses residual connections and layer normalization to output deep features. The classification head performs global average pooling on deep features, aggregates them into sequence-level feature vectors, and then converts them into probability distributions of gas box status labels through a fully connected layer and a Softmax function. Based on the gas box status tag of the gas cabinet, the probability of belonging to the real-time thermal imaging sequence to the normal thermal imaging cluster is calculated to obtain the health status of the gas cabinet. Step S4: Based on the health status of the gas-filled cabinet, classify the risk level of the gas-filled cabinet and issue an early warning, thereby realizing online monitoring of the gas-filled cabinet.

2. The online monitoring method for the gas chamber of a gas-filled cabinet according to claim 1, characterized in that: Step S2 involves collecting thermal imaging images of the key monitoring area of ​​the gas-filled cabinet under normal conditions to obtain a normal thermal imaging sequence, which includes the following specific steps: Acquire the t-th frame of thermal imaging image under normal and healthy operating conditions. (x,y), (x, y) represents the original temperature measurement value at pixel coordinates (x, y) at time t; the fixed pixel set of the preset reference thermal marker region in frame t=1 is... , ={(x,y)| ≤x≤ , ≤y≤ },in , , , The pixel boundary values ​​are for a fixed set of pixels, where x and y are the horizontal and vertical coordinate indices, respectively. By applying drift correction to the thermal imaging image using a fixed set of pixels, a drift-corrected thermal imaging image is obtained, ultimately resulting in a normal thermal imaging sequence under normal and healthy operating conditions.

3. The method for online monitoring of the gas chamber of a gas-filled cabinet according to claim 2, characterized in that: In step S2, the thermal imaging image is shift-corrected using a fixed pixel set to obtain the shift-corrected thermal imaging image, which includes the following steps: Template matching algorithm in thermal imaging images The actual pixel position of the reference heat-marked region is located on (x,y). ; Calculate the actual pixel position respectively centroid coordinates and fixed pixel set Determine the centroid coordinates and solve for the actual region. centroid coordinates and reference region Translational offset of the centroid coordinates ( , ); for thermal imaging images Perform a translation transformation on (x,y) to obtain a spatially aligned thermal imaging image. (x,y); Spatially aligned thermal imaging images In (x,y), calculate the reference region. Average temperature inside (t), based on the reference region Average temperature inside (t) Temperature drift correction is performed to obtain the drift-corrected thermal imaging image. .

4. The online monitoring method for the gas chamber of a gas-filled cabinet according to claim 3, characterized in that: Step S2 iteratively updates the cluster centers of the clusters to obtain the initial Gaussian mixture model parameters, including the following steps: Clustering normal thermal imaging vector mean As the new cluster center, repeat the Mahalanobis distance calculation, redistribute clusters, and recalculate the vector mean. sum vector covariance The process continues until the cluster center converges, ultimately yielding the first... The vector mean, vector covariance, and weights of each Gaussian component are obtained, and then the normal thermal imaging vector is processed serially to obtain the initial Gaussian mixture model parameters.

5. The online monitoring method for the gas tank of a gas-filled cabinet according to claim 4, characterized in that: Step S3 involves collecting real-time thermal imaging sequences, load current, and ambient temperature of the key monitoring areas of the gas-insulated switchgear, and inputting them into a Transformer model based on a spatiotemporal attention mechanism. The output is a status label for the gas-insulated switchgear, which includes the following steps: Construct a Transformer model based on a spatiotemporal attention mechanism; train the Transformer model; Real-time thermal imaging sequences, load current, and ambient temperature of key monitoring areas of the gas-insulated switchgear are collected; the real-time thermal imaging sequences, load current, and ambient temperature are input into a Transformer model based on a spatiotemporal attention mechanism to finally obtain the gas-insulated switchgear's gas box status label.

6. The method for online monitoring of the gas chamber of a gas-filled cabinet according to claim 5, characterized in that: Step S3, which involves training the Transformer model, includes the following specific steps: The loss function for the Transformer model is: ； in, Represents the total loss function. Cross-entropy loss function for classifying operation types. The cross-entropy loss function for load level classification, Cross-entropy loss function for classifying ambient temperature ranges; The Adam optimizer is used to update the model parameters, and the loss function is minimized through gradient descent to obtain a trained Transformer model.

7. The online monitoring method for the gas chamber of a gas-filled cabinet according to claim 6, characterized in that: In step S3, based on the gas chamber status tag of the gas cabinet, the probability of the real-time thermal imaging sequence belonging to the normal thermal imaging cluster is calculated to obtain the health status of the gas cabinet, specifically: ； in, For the health status of the gas-filled cabinet, The number of effective Gaussian components, This is the feature vector of a real-time thermal imaging sequence. for Gaussian multivariate probability density, The weight of the k-th Gaussian component. Let be the mean vector of the k-th Gaussian component. Let be the covariance matrix of the k-th Gaussian component.

8. The online monitoring method for the gas box of a gas-filled cabinet according to claim 7, characterized in that: In step S4, the risk level of the gas-filled cabinet is classified according to its health status. The specific steps are as follows: Low risk: When the health status p( If the temperature is greater than or equal to the preset high health threshold Th2, the thermal behavior pattern is considered normal. Medium risk: When the preset low health threshold Th1 ≤ the health status p of the gas-filled cabinet. When the temperature exceeds the preset high health threshold Th2, the thermal pattern is determined to deviate significantly from the baseline, triggering an alert. High risk: When the health status p( When the preset low health threshold Th1 is reached, the thermal pattern is determined to be seriously deviating from the health benchmark, triggering the highest level alarm.

Images