An insulation aging early warning method based on odor sensing graph neural network

Abstract

This application relates to the field of electrical equipment condition monitoring and fault diagnosis, and discloses an insulation aging early warning method based on an odor-sensing graph neural network. This method compares the convection dominance index of the electrical equipment with a critical threshold to determine the internal gas transport pattern. Under the thermal convection dominance pattern, it calculates the physical time delay of characteristic gas transport to each node and maps it as an addressing index to extract historical odor concentration data for spatiotemporal alignment. A directed aerodynamic adjacency matrix is ​​constructed based on the relative spatial height of nodes. The historical odor data and dynamic temperature rise features are concatenated and then fed into a directed graph convolutional neural network to output the diagnostic results. Under the natural diffusion dominance pattern, real-time odor concentration data is extracted, and a static undirected adjacency matrix is ​​constructed based on spatial Euclidean distance and fed into an undirected graph convolutional neural network to output the results. This invention effectively solves the problem of spatiotemporal data mismatch, truly characterizes the asymmetric migration law of fluids, and improves the accuracy of insulation aging diagnosis of electrical equipment.

Description

Technical Field

[0001] This invention relates to the field of electrical equipment condition monitoring and fault diagnosis technology, specifically to an insulation aging early warning method based on an odor sensing graph neural network. Background Technology

[0002] During long-term operation, the internal insulation materials of electrical equipment age under stress from electricity, heat, and other factors, releasing specific characteristic gases. Currently, deploying odor sensors or gas sensor arrays inside the equipment to collect the concentration of these characteristic gases and then combining this data with deep learning algorithms for analysis has become a standard method for early warning of insulation aging in electrical equipment.

[0003] However, in actual industrial operation scenarios, the gas transport process inside electrical equipment is affected by both the dynamic heating of the equipment and its internal structure. The diffusion of characteristic gases exhibits a complex physical process involving alternating thermal convection and natural diffusion. Existing insulation aging monitoring methods typically read data from all sensor nodes simultaneously at the same fixed sampling time and use it directly as model input, without considering the time difference in the transport of characteristic gases from the release source to sensors at different spatial locations. This direct reading mechanism ignores the physical time delay of gas transport, causing the data input into the model to fail to correspond to the gas state released from the same insulation aging source at the same time, resulting in a severe mismatch in the spatiotemporal dimensions of the underlying characteristic data.

[0004] Meanwhile, existing diagnostic methods based on graph neural networks typically construct a symmetrical undirected graph topology based solely on the physical Euclidean distance between sensor nodes when processing sensor array data. Under conditions where significant temperature rises in equipment induce thermal convection, characteristic gases exhibit bottom-up directional migration driven by thermal buoyancy. The transport resistance and concentration distribution of the gas along the forward and reverse paths show significant asymmetry. Traditional static undirected adjacency matrices cannot accurately represent this aerodynamic topological relationship driven by temperature stress. The dual defects of data temporal misalignment and spatial topological distortion make it difficult for conventional algorithms to accurately extract the true spatiotemporal distribution of odor concentration, ultimately reducing the accuracy and adaptability of insulation aging diagnosis. Summary of the Invention

[0005] The technical problem solved by this invention is that existing insulation aging monitoring methods do not fully consider the dynamic alternation of internal thermal convection and natural diffusion conditions in electrical equipment, resulting in mismatches in the spatiotemporal dimensions of characteristic gas concentration data. Furthermore, traditional static graph topology cannot accurately represent the asymmetric gas transport path driven by temperature stress, leading to a decrease in the accuracy of insulation diagnosis.

[0006] To address the above problems, the present invention provides the following technical solution:

[0007] The first aspect of this invention provides an insulation aging early warning method based on an odor sensing graph neural network, comprising the following steps:

[0008] By comparing the current convection dominance index of the electrical equipment with the critical threshold, the gas transport mode inside the current electrical equipment is determined to be either thermal convection dominance or natural diffusion dominance. When the gas transport mode is thermal convection dominance, the physical transport delay of the characteristic gas to each odor sensor node is calculated, and the physical transport delay is mapped to an addressing index to extract historical odor concentration data from the circular buffer. A directed aerodynamic adjacency matrix is ​​constructed based on the spatial relative height of each odor sensor node.

[0009] Historical odor concentration data is spliced ​​with the dynamic temperature rise characteristics of electrical equipment to construct a multimodal node feature matrix; the multimodal node feature matrix and the directed aerodynamic adjacency matrix are input into a pre-deployed directed graph convolutional neural network model to output insulation aging diagnosis results;

[0010] When the gas transport mode is dominated by natural diffusion, the addressing index of each odor sensor node is reset to zero, and the real-time odor concentration data at the current sampling time is extracted.

[0011] A static undirected adjacency matrix is ​​constructed based on the spatial Euclidean distance of each odor sensor node; the node feature matrix composed of real-time odor concentration data and the static undirected adjacency matrix are input into a pre-deployed undirected graph convolutional neural network model to output the insulation aging diagnosis results.

[0012] As a further explanation of the invention, when determining the gas transport mode, a lower limit phase transition threshold and an upper limit phase transition threshold with tolerance ranges are configured. The system performs a hysteresis comparison by combining the gas transport mode at the previous sampling time with the currently calculated convection dominance index to suppress frequent switching of airflow states near the critical point.

[0013] In the thermal convection-dominated mode, the geometric center of the primary conductive circuit on the electrical equipment is extracted as the physical location of the proxy source for the diffusion of characteristic gases. Combining the effective temperature difference inside the equipment with aerodynamic parameters, an empirical reduction factor for convection velocity is introduced to calculate the dynamic convection velocity in the vertical direction. The system employs a spatial dimensionality reduction and decoupling strategy, independently calculating and superimposing the time for the characteristic gas to travel vertically via dynamic convection velocity with the time for natural diffusion in the horizontal plane via concentration gradient, thus obtaining the physical transmission time delay of the characteristic gas to each odor sensor node.

[0014] Preferably, the physical transmission delay of each odor sensor node is divided by a fixed sampling period and rounded to the nearest integer. After exceeding the maximum buffer depth limit protection, a corresponding addressing index is generated. During the memory read phase, the current write pointer is subtracted from the addressing index, and a modulo operation is performed using the maximum buffer depth to locate the time-compensated and aligned register slot to extract historical odor concentration data. This operation compensates for the spatial transmission delay caused by thermal convection, achieving spatiotemporal alignment of data samples at the model input.

[0015] In constructing the directed aerodynamic adjacency matrix, the height difference between the target odor sensor node and the source odor sensor node in the vertical direction is calculated. When the height difference is greater than zero, it is determined to be a forward path, and the forward adjacency weight is calculated by introducing the vertical aerodynamic compressibility coefficient and a small constant; when the height difference is less than or equal to zero, it is determined to be a reverse or horizontal transmission path, and the reverse flow penalty factor and the three-dimensional Euclidean distance between the nodes are introduced to calculate the reverse adjacency weight. This asymmetric weight allocation mechanism characterizes the directional migration characteristics driven by fluid thermal buoyancy.

[0016] In the process of constructing the multimodal node feature matrix, the dynamic temperature rise feature is dimensionally expanded to generate a temperature rise feature column vector. A standardization algorithm is then applied to normalize the historical odor concentration data and the temperature rise feature column vector to zero mean and unit variance, respectively, to unify the physical dimensions. Subsequently, a splicing operation is performed on the feature dimensions to generate the multimodal node feature matrix, enabling the model to simultaneously process local concentration and thermal stress features.

[0017] During forward inference in the directed graph convolutional network, self-loop compensation is introduced into the directed aerodynamic adjacency matrix. Out-degree normalization is performed on the node's out-degree matrix to generate the directed transition matrix, preventing local node feature values ​​from exceeding limits due to asymmetric topology. The directed transition matrix and the multimodal node feature matrix are input into the hidden layer of the directed graph convolutional neural network model. Asymmetric spatial aggregation and nonlinear mapping are performed along the airflow channel to obtain a node-level high-order representation matrix. Global average pooling is then used to aggregate and reduce the dimensionality of this matrix along the spatial dimension to generate a graph-level feature vector. This vector is then mapped through a fully connected layer and a normalized exponential function to a probability distribution vector reflecting the insulation state. The physical state label corresponding to the highest probability is extracted as the final warning result.

[0018] Under the natural diffusion-dominated mode, the system halts the time-delay addressing process due to the lack of directional kinetic energy in gas transport. For the extraction of real-time odor concentration data, the filtered output is initialized to the original odor concentration data during cold start.

[0019] During non-cold start times, a first-order inertial low-pass filter algorithm is invoked to perform smooth recursion by combining the filtered data from the previous time step. When constructing the static undirected adjacency matrix, the reciprocal of the physical Euclidean distance between nodes is calculated and a constant compensation is added as a matrix element. The constructed node feature matrix and the Laplacian operator with self-loop compensation are substituted into an undirected graph convolutional neural network to perform symmetric aggregation, outputting the insulation aging diagnosis results.

[0020] Before determining the gas transport pattern, the process includes offline training of the model. This involves acquiring historical operating samples containing odor concentration and temperature rise characteristics, as well as actual aging state labels represented by a unique encoding format. The cross-entropy loss is calculated using the predicted probability distribution output from the forward propagation and the actual aging state labels. An adaptive moment estimation optimizer is then used to iteratively update the weights and biases of the network layers within the model using the backpropagation algorithm, until the loss fluctuations converge to a preset threshold range, thus completing parameter solidification.

[0021] A second aspect of the present invention provides an insulation aging early warning system based on an odor sensing graph neural network, including a memory and a processor, wherein the memory is used to store a computer program;

[0022] When the processor is used to execute a computer program, it implements the insulation aging early warning method based on odor sensing graph neural network as described in the first aspect.

[0023] This invention provides an insulation aging early warning method based on an odor sensing graph neural network. It has the following beneficial effects:

[0024] 1. This invention dynamically switches between a thermal convection-dominated mode and a natural diffusion-dominated mode by comparing the convection dominance index with a critical threshold. This dual-branch processing logic based on physical transport mechanisms overcomes the shortcomings of a single fixed model that cannot adapt to the complex airflow environment inside electrical equipment. Simultaneously, in conjunction with a hysteresis comparison mechanism with a tolerance range, it suppresses frequent state switching of the system under critical airflow energy conditions, improving the overall operational stability of the insulation aging monitoring system under different load and temperature conditions.

[0025] 2. Under the thermal convection-dominated mode, this invention extracts historical odor concentration data by calculating the physical transport delay of the characteristic gas and mapping it to an addressing index, combined with a circular buffer mechanism. This method directly compensates for the time difference caused by the transport of the characteristic gas at different spatial locations within the device, achieving spatiotemporal alignment of multi-node sampling samples from the underlying data structure. It eliminates data mismatch errors caused by differences in the physical installation positions of sensors, providing high-quality feature input synchronized on the physical time axis for subsequent neural network inference.

[0026] 3. This invention constructs a directed aerodynamic adjacency matrix based on the spatial relative height difference of odor sensor nodes, and concatenates odor concentration data with dynamic temperature rise features to construct a multimodal node feature matrix, which is then substituted into a directed graph convolutional neural network for inference. By distinguishing between forward thermal convection paths and reverse transport paths and assigning asymmetric weights, the physical laws of directional fluid migration driven by temperature stress are realistically characterized. Multimodal feature fusion enables the model to simultaneously perceive odor concentration gradients and global thermal stress states, overcoming the limitations of traditional undirected graph static topologies and improving the reliability of insulation aging early warning. Attached Figure Description

[0027] Figure 1 This is a hardware architecture diagram of an insulation aging early warning system based on an odor sensing graph neural network according to an embodiment of the present invention;

[0028] Figure 2 This is a flowchart illustrating the process of the insulation aging early warning method based on an odor sensing graph neural network according to an embodiment of the present invention.

[0029] Figure 3 This is a flowchart illustrating the system topology initialization process according to an embodiment of the present invention.

[0030] Figure 4 This is a flowchart of the multi-source heterogeneous parameter synchronous sensing process according to an embodiment of the present invention;

[0031] Figure 5 This is a schematic diagram of the asynchronous data buffer principle for odor time series according to an embodiment of the present invention;

[0032] Figure 6 This is a schematic diagram illustrating the calculation principle of a first-order equivalent thermal resistivity-capacitance network according to an embodiment of the present invention.

[0033] Figure 7 This is a schematic diagram illustrating the quantitative evaluation principle of the convection dominance index in an embodiment of the present invention.

[0034] Figure 8 This is a logic diagram for comparing critical thresholds of flow phase transitions in an embodiment of the present invention.

[0035] Figure 9 This is a schematic diagram illustrating the principle of proxy source anchoring and time-delay calculation under thermal convection conditions in an embodiment of the present invention.

[0036] Figure 10 This is a schematic diagram of the fault-tolerant bypass mechanism in the natural diffusion state according to an embodiment of the present invention;

[0037] Figure 11 This is a schematic diagram illustrating the principle of pointer offset spatiotemporal alignment and aerodynamic adjacency matrix generation in an embodiment of the present invention.

[0038] Figure 12 This is a schematic diagram illustrating the principle of static feature extraction and spatial topology degradation in an embodiment of the present invention.

[0039] Figure 13 This is a schematic diagram illustrating the forward inference and state mapping principle of a cross-modal graph neural network according to an embodiment of the present invention.

[0040] Figure 14 This is a visual comparison curve of cross-modal feature splicing and graph-level representation extraction in an embodiment of the present invention;

[0041] Figure 15 This is a bar chart comparing the accuracy and false alarm rate of insulation state diagnosis for different models in this invention. Detailed Implementation

[0042] 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.

[0043] Example:

[0044] See attached document Figure 1 , Figure 1 This is a hardware architecture diagram of an insulation aging early warning system based on an odor sensing graph neural network according to an embodiment of the present invention. The present invention provides an insulation aging early warning system based on an odor sensing graph neural network, which may include: a load monitoring unit, a temperature sensing unit, an odor sensing array, and an edge gateway.

[0045] The load monitoring unit is installed on the primary conductor of the electrical equipment to collect the power frequency load current of the electrical equipment.

[0046] Temperature sensing units are placed in areas outside or inside electrical equipment that are not exposed to heat-generating components, and are used to collect ambient baseline reference temperatures.

[0047] The odor sensing array consists of N odor sensor nodes. N is an integer greater than or equal to 3. The N odor sensor nodes are distributed at three-dimensional spatial coordinate points inside the electrical equipment to synchronously collect the concentration of characteristic gases.

[0048] The edge gateway connects to the load monitoring unit, temperature sensing unit, and odor sensing array via a communication bus. The edge gateway integrates a microprocessor and memory. The memory stores a graph neural network inference model and control execution program.

[0049] See attached document Figure 2 , Figure 2This is a flowchart illustrating the workflow of an insulation aging early warning method based on an odor sensing graph neural network according to an embodiment of the present invention. The present invention provides an insulation aging early warning method based on an odor sensing graph neural network, comprising the following steps:

[0050] S10, during the startup phase, the edge gateway reads the physical space impedance matrix and Euclidean distance matrix inside the electrical equipment;

[0051] S20, the edge gateway synchronously collects power frequency load current, ambient basic reference temperature and concentration data output by each odor sensor node according to the set clock cycle;

[0052] S30, the edge gateway configures an independent first-in-first-out circular buffer for each odor sensor node in the memory, and pushes the collected concentration data into the corresponding first-in-first-out circular buffer according to the time series;

[0053] S40, the edge gateway extracts the power frequency load current and substitutes it into the first-order equivalent thermal resistance-capacitance network model to calculate the estimated dynamic temperature rise of the internal insulation surface of the electrical equipment.

[0054] S50: The edge gateway calculates the real-time temperature difference based on the dynamic temperature rise estimate and the environmental baseline reference temperature, and calculates the convection dominance index based on the real-time temperature difference.

[0055] S60, the edge gateway compares the calculated convection dominance index with the preset convection critical threshold;

[0056] S70, when the convection dominance index is greater than or equal to the preset convection critical threshold, the edge gateway determines that the system is in a thermal convection dominance state, executes the dynamic agent source anchoring algorithm to obtain the spatial reference origin, and calculates the time delay of each odor sensor node by combining the physical space impedance matrix and the Euclidean distance matrix.

[0057] S80, the edge gateway converts the offset pointer according to the calculated time delay, extracts the concentration data of the corresponding time window from each first-in-first-out circular buffer to form an equivalent synchronous emission feature vector, and generates a dynamic aerodynamic adjacency matrix based on the time delay gradient.

[0058] S90, when the convection dominance index is less than the preset convection critical threshold, the edge gateway determines that the system is in a natural diffusion dominance state and forces the time delay of each odor sensor node to be zero.

[0059] S100: The edge gateway reads the latest concentration data from the top of each first-in-first-out circular buffer stack to form a node feature vector, and generates a static graph adjacency matrix based on the Euclidean distance matrix.

[0060] S110, the edge gateway takes the equivalent synchronous emission feature vector or node feature vector as the feature tensor input of the graph node, and the dynamic aerodynamic adjacency matrix or static graph adjacency matrix as the topology input, and feeds them into the graph neural network inference model for forward calculation, and outputs the insulation aging classification label of the electrical equipment.

[0061] See attached document Figure 3 , Figure 3 This is a system topology initialization flowchart according to an embodiment of the present invention. In the insulation aging early warning method provided by the present invention, step S10, in which the edge gateway reads the physical space impedance matrix and Euclidean distance matrix inside the electrical equipment during the startup phase, may specifically include the following steps:

[0062] S101, the edge gateway reads the three-dimensional spatial coordinates of each odor sensor node inside the electrical equipment and calculates and generates an Euclidean distance matrix;

[0063] In this embodiment, the edge gateway acquires the position data of N odor sensor nodes in a preset three-dimensional Cartesian coordinate system. As a preferred implementation, this three-dimensional Cartesian coordinate system can be established with preset reference points such as the vertex of the bottom surface inside the electrical equipment cabinet as the origin. The position data of any odor sensor node is represented as three-dimensional coordinates. ,in For node index, and .

[0064] To evaluate the spatial geometric connectivity between sensor network nodes in subsequent calculations, the edge gateway computes the odor sensor nodes. With odor sensor node Linear spatial distance between The corresponding spatial distance calculation formula is as follows:

[0065] ;

[0066] It should be noted that when At that time, the linear spatial distance of the node itself The value is 0. After the edge gateway has traversed and calculated the linear spatial distances between all odor sensor nodes, it integrates the results to generate a dimension of 0. Euclidean distance matrix This Euclidean distance matrix is ​​used to characterize the ideal geometric topology connectivity of the sensor network inside an electrical device; it is a symmetric matrix with zero elements on one diagonal.

[0067] S102, the edge gateway reads the structural feature data inside the electrical equipment and establishes a physical space impedance matrix characterizing the gas transmission obstruction effect based on the structural feature data;

[0068] In real-world substation operation scenarios, electrical equipment contains physical components such as insulating partitions, busbars, and circuit breakers. When characteristic gases generated by the aging of insulating materials diffuse outwards, their propagation path is not an ideal straight line. These physical components constitute physical obstacles in the gas diffusion path, making the actual gas propagation path often longer than the ideal straight-line spatial distance. Relying solely on simple Euclidean distance to assess the time delay of gas diffusion introduces significant physical errors. Therefore, To quantify the aforementioned transmission blocking effect. Physical space impedance matrix. The dimension is also Its matrix elements Defined as an odor sensor node To the odor sensor node The gas transmission resistance coefficient, whose value range is set to be... .

[0069] Edge gateway gas transmission resistance coefficient The specific control logic for initialization assignment is as follows:

[0070] When the odor sensor node With odor sensor node For the same node (i.e.) If there are no physical components obstructing the straight line connecting the two nodes, the edge gateway determines that the gas transmission channel between the two nodes is free space and sets the gas transmission resistance coefficient. 1.

[0071] When the odor sensor node With odor sensor node When a physical component obstructs the straight line connecting the two nodes, the edge gateway determines that the space between the two nodes is a restricted deflection space. The edge gateway reads preset 3D structural model data to obtain the length of the effective obstacle avoidance path between the two nodes that bypasses the physical component. Specifically, the effective obstacle avoidance path length Characterizing gas molecules from nodes Start by circling around the edge of the solid component to reach the node. The shortest aerodynamic channel distance. The edge gateway calculates the ratio of the effective obstacle avoidance path length to the straight-line spatial distance, and assigns this ratio to the corresponding gas transmission drag coefficient. The logical expression of this formula is as follows:

[0072] ;

[0073] Since the effective obstacle avoidance path length is usually greater than the straight-line spatial distance, the gas transmission drag coefficient is set accordingly. As an anomaly protection mechanism, when there is a completely sealed and isolated chamber inside the electrical equipment, resulting in no effective obstacle avoidance path connecting the two nodes, the edge gateway will treat the obstacle avoidance path length as infinite and trigger the interception mechanism. It will directly set the corresponding gas transmission resistance coefficient to the system's preset maximum resistance penalty threshold to prevent numerical overflow in subsequent network weight calculations.

[0074] For the specific algorithm for obtaining the internal three-dimensional structural feature data of electrical equipment and calculating the length of the effective obstacle avoidance path around the solid components, those skilled in the art can use conventional computer three-dimensional solid modeling analysis and ray obstacle avoidance pathfinding algorithm. The underlying calculation logic is a well-known technology in this field and will not be elaborated here.

[0075] After completing the above calculations, the edge gateway will convert the Euclidean distance matrix... and physical space impedance matrix The data is persistently written to the internal non-volatile memory to complete the system topology initialization process, which will then be called by the subsequent dynamic feature reconfiguration process.

[0076] See attached document Figure 4 , Figure 4 This is a flowchart of multi-source heterogeneous parameter synchronous sensing according to an embodiment of the present invention. In the insulation aging early warning method provided by the present invention, step S20, the edge gateway 40 synchronously collects the power frequency load current, the ambient basic reference temperature, and the concentration data output by each odor sensor node 301 according to a set clock cycle, which may specifically include the following steps:

[0077] S201, Edge Gateway 40 is configured with a global synchronous sampling clock to establish a unified discrete time coordinate system;

[0078] In this embodiment, the electrical, thermal, and chemical parameters of the electrical equipment during operation have inherently different dynamic response bandwidths. To eliminate the temporal misalignment of multi-source heterogeneous data in subsequent graph neural network inference, the edge gateway initializes a hardware timer to generate a fixed system clock tick. The edge gateway defines the time interval between two adjacent system clock tick triggers as a fixed sampling period. Considering the relatively slow macroscopic physical diffusion rate of gas molecules within a confined space, this sampling period is a preferred approach. The value range can be set from 1 second to 60 seconds, thus balancing the computational load of the edge gateway while preserving sufficient dynamic gas change characteristics. The edge gateway uses the instant when the current clock interrupt is triggered as the unified sampling time. At this moment, data freeze and read commands are sent in parallel to all sensing nodes on the communication bus.

[0079] S202, the edge gateway obtains the power frequency load current at the current sampling time t through the load monitoring unit;

[0080] The load monitoring unit senses the power frequency alternating magnetic field of the primary conductor of the electrical equipment and outputs a corresponding current measurement signal to the edge gateway. The edge gateway receives this signal and performs analog-to-digital conversion. To accurately measure the actual electrical load causing heating of the internal conductors, the edge gateway calculates the root mean square value of the AC signal based on a power frequency cycle integration algorithm and calibrates it as the current sampling time. power frequency load current This parameter characterizes the current-carrying state of the main circuit of the electrical equipment and is the fundamental physical input for subsequent calculations of the internal Joule heating power.

[0081] S203, the edge gateway reads the ambient baseline reference temperature at the current sampling time t through the temperature sensing unit;

[0082] The temperature sensing unit is deployed in a reference area that is not directly exposed to radiation from localized heat sources within the electrical equipment. The edge gateway reads the real-time temperature signal it senses and records it as the environmental baseline reference temperature. As a preferred approach, obtaining this environmental baseline reference temperature aims to provide thermodynamic cold-end boundary conditions for subsequent thermal resistance-capacitance network calculations. This helps filter out common-mode interference from external seasonal climate changes that cause overall equipment temperature rise, allowing the system to primarily respond to abnormal temperature rises caused by insulation degradation.

[0083] S204, the edge gateway acquires the characteristic gas concentration data output by each odor sensor node through the odor sensor array;

[0084] The edge gateway sequentially parses digital messages uploaded by N odor sensor nodes in the odor sensor array via the communication bus. From each node's message, the edge gateway extracts concentration measurements characterizing the content of microscopic decomposition or oxidation products of the insulating material. At the current sampling time... , by the The characteristic gas concentration data collected and analyzed by each odor sensor node is defined as... , where index This concentration data characterizes the macroscopic distribution of gas molecules at specific physical coordinates after diffusion through complex internal spaces.

[0085] In real-world industrial fieldbus environments, data uploads from individual odor sensor nodes may experience minor delays or occasional packet loss due to electromagnetic interference. To ensure the spatial integrity of the data structure input into the graph neural network, when an odor sensor node is at the current sampling time... When a communication timeout occurs, the edge gateway triggers the data retention mechanism, calling the node's data retention mechanism from the previous sampling time. Historical concentration data Fill in the gaps to maintain the constant feature dimensions.

[0086] For anti-aliasing filtering, RMS calculation, and encapsulation and parsing of industrial bus communication messages for analog signals, those skilled in the art can employ conventional digital signal processing algorithms and standard serial communication protocols. The underlying execution logic is well-known in the field and will not be elaborated upon here. The edge gateway will use the same sampling time... The following is obtained , and sets Timestamp alignment and structured packaging are performed in memory for subsequent data buffering and cross-modal feature reassembly.

[0087] See attached document Figure 5 , Figure 5 This is a schematic diagram of an asynchronous data buffer principle for odor time series data according to an embodiment of the present invention. In the insulation aging early warning method based on an odor sensing graph neural network provided by the present invention, step S30 involves the edge gateway configuring an independent first-in-first-out (FIFO) circular buffer for each odor sensor node in its memory, and pushing the collected concentration data into the corresponding FIFO circular buffer according to the time series. Specifically, this may include the following steps:

[0088] S301, The edge gateway evaluates the maximum physical delay time of gas transmission within electrical equipment and allocates the data depth of the first-in-first-out circular buffer to each odor sensor node based on the maximum physical delay time;

[0089] In this embodiment, due to the different physical spatial distances between the various odor sensor nodes inside the electrical equipment and the potential sources of insulation aging release, the arrival times of characteristic gas molecules at each node are not consistent. To solve the aforementioned spatiotemporal misalignment problem and bridge this gas transmission time difference caused by spatial distance in the subsequent feature reconstruction stage, thereby achieving data alignment on the logical time section, the edge gateway typically needs to cache the historical concentration data of each node in memory.

[0090] The edge gateway calculates the maximum physical delay time required for gas to traverse the entire interior space based on the maximum internal geometric diagonal dimension of the electrical equipment and the empirical minimum flow velocity of gas under natural diffusion or weak convection conditions. As a preferred method, this empirical minimum flow rate can be obtained through preliminary simulated fluid dynamics analysis or measured statistical data from a field hot-wire anemometer. This is combined with a set fixed sampling period. The edge gateway calculates the minimum data depth of the first-in-first-out circular buffer. The corresponding calculation formula is as follows:

[0091] ;

[0092] in, The symbol represents the rounding up sign. A data depth redundancy constant is preset for the system. By introducing this redundancy constant, additional buffer space can be provided for network fluctuations or extremely low flow rate conditions. As a preferred approach, considering airflow stagnation under extreme conditions, the maximum physical delay time... The value range is typically set to 60 seconds to 300 seconds, and the data depth redundancy constant is... It can be set to 5 to 10 to prevent the pointer from going out of bounds during extreme delay drift.

[0093] The edge gateway is in its internal static random access memory, for Each odor sensor node has a length of [length missing]. A one-dimensional array, used to construct Each buffer is a separate first-in-first-out circular buffer. Each buffer maintains an independent head pointer and a tail pointer to mark the current write position of the latest data and the overwrite position of the oldest data.

[0094] S302, the edge gateway updates the latest concentration data of each odor sensor node at each sampling time and dynamically maintains a sliding window of historical data with a fixed time span;

[0095] After completing the memory allocation described above, the system enters the periodic data pushing phase. At any sampling time... The edge gateway will collect the latest concentration data from each odor sensor node uploaded by the odor sensor array. Each element is pushed into the corresponding first-in-first-out circular buffer of its respective node.

[0096] When the amount of data stored in the buffer has not reached the set data depth During this process, the head pointer increments sequentially as data is written, indicating that the system is in the initial data accumulation phase. To ensure the integrity of the algorithm logic and avoid subsequent extraction errors due to insufficient historical data in the early stages of startup, the edge gateway fills the free memory blocks that have not yet been written with real data with the initial environmental baseline concentration value of the corresponding node before the buffer is full. This initial environmental baseline concentration value is obtained by the edge gateway through zero-point calibration of the odor sensor nodes against a clean air background during the power-on initialization phase.

[0097] Once all FIFO circular buffers are full, the system enters a steady-state sliding window operation mode. With the influx of data at each new sampling time, the edge gateway writes the latest concentration data to the current tail pointer position, directly overwriting the oldest historical data. The head and tail pointers are then synchronously updated via circular shifting. Through this alternating data overwriting operation, each FIFO circular buffer maintains a time span of [missing information]. A dynamic historical data sliding window.

[0098] In the context of continuous data updates, to prevent timing conflicts between underlying data write operations and upper-layer feature extraction operations, the edge gateway configures a read-write mutex mechanism for each first-in-first-out circular buffer. When the system pushes sampled data onto the stack, it requests a mutex to suspend the feature extraction thread. The lock is released only after the write is completed and the pointer is updated, thus ensuring the data consistency of the final output feature sequence.

[0099] For the specific logic implementation of the underlying memory addressing, pointer out-of-bounds reset, and loop toggling of the first-in-first-out circular buffer, those skilled in the art can use conventional embedded C language array pointer manipulation techniques. The underlying mechanism is well-known in the field and will not be elaborated upon here. After asynchronous data buffering, the system transforms the originally continuous physical time stream slices into two-dimensional feature matrices that can be used for structured calls by graph neural networks.

[0100] See attached document Figure 6 , Figure 6 This is a schematic diagram illustrating the calculation principle of a first-order equivalent thermal resistance-capacitance network according to an embodiment of the present invention. In the insulation aging early warning method based on an odor sensing graph neural network provided by the present invention, step S40, where the edge gateway 40 extracts the power frequency load current and substitutes it into the first-order equivalent thermal resistance-capacitance network model, calculates the estimated dynamic temperature rise of the internal insulation surface of the electrical equipment, may specifically include the following steps:

[0101] S401, The edge gateway reads and initializes the pre-configured first-order equivalent thermal resistivity-capacitance network model parameters;

[0102] In this embodiment, in practical engineering applications, high-voltage electrical equipment is enclosed and subject to strong electromagnetic fields. Directly placing contact temperature sensors not only poses a risk of high-voltage insulation breakdown but also makes it difficult to cover all potential heat points. Therefore, the edge gateway employs an electro-thermal analogy principle for non-invasive indirect temperature measurement. The edge gateway reads and initializes pre-configured first-order equivalent thermal resistance-capacitance network model parameters from its memory. The physical process of heat generation and dissipation of conductors inside electrical equipment can be analogized to the resistance-capacitance charging and discharging process in electricity. This model consists of three core physical constants characterizing heat generation and conduction properties, specifically including: equivalent conductor resistance... Equivalent thermal resistance and equivalent heat capacity .

[0103] Equivalent conductor resistance Equivalent thermal resistance characterizes the overall electrical contact and conductor ohmic loss characteristics of the primary circuit of electrical equipment. Its unit is ohms, and it is typically taken as the AC equivalent resistance value under normal operating conditions. Equivalent heat capacity is the barrier effect that characterizes the transfer of heat from a heating conductor through an insulating medium to the surrounding air, and its unit is Kelvin per watt. The physical inertia characterizes the absorption and storage of heat by the conductors and insulation materials inside the equipment, and its unit is joules per Kelvin. As a preferred method, the above three constant parameters can be calculated by extracting the steady-state temperature rise test report data from the electrical equipment at the time of manufacture, or they can be calibrated a priori through finite element thermal field simulation and persistently burned and stored in the edge gateway.

[0104] S402, the edge gateway calculates the transient Joule heating power based on the power frequency load current at the current sampling time;

[0105] The internal heating conditions of electrical equipment directly affect the aging rate of insulation materials and the spatial diffusion evolution of characteristic gases. The edge gateway invokes the current sampling time obtained in step S20. power frequency load current This is used as the system's heat source input. The edge gateway calculates the transient internal Joule heating power based on Joule's law. The specific conversion formula is as follows:

[0106] ;

[0107] After obtaining the transient heating power, according to the laws of energy conservation and heat conduction, part of the heating power inside the system is used to increase the temperature of the heating element itself, and the other part is dissipated to the outside through thermal resistance. The corresponding continuous-time differential control equation is expressed as:

[0108] ;

[0109] S403, the edge gateway uses the first-order forward Euler difference algorithm for discrete iteration to solve and output the dynamic temperature rise estimate;

[0110] To adapt to the microprocessor of the edge gateway relying on a fixed sampling period The hardware mechanism for discrete operation involves the edge gateway employing a first-order forward Euler difference algorithm to transform the aforementioned continuous equations into discrete recursive formulas suitable for iterative execution of software programs.

[0111] ;

[0112] in, The current sampling time The calculated estimated value of dynamic temperature rise on the insulation surface; For the edge gateway in the previous clock cycle, i.e. time The historical temperature rise estimates are calculated and stored. To ensure the numerical convergence of the forward difference algorithm and reduce the risk of iteration divergence, a preset sampling period is used. The value of should be less than the system's equivalent thermal time constant, i.e., satisfy . As a preferred method, It can be set between 1 second and 5 seconds to achieve a balance between calculation accuracy and the computing power overhead of the microprocessor.

[0113] To prevent computational anomalies in the difference equations due to missing initial data or negative fluctuations, the edge gateway employs a dual mechanism: cold start initial value estimation and non-negative limiting protection. When the edge gateway executes the startup procedure (i.e.... To avoid computational anomalies such as memory addressing errors due to a lack of historical data, the system automatically executes a steady-state initial value estimation mechanism. Specifically, if the edge gateway and device are powered on synchronously during cold boot, the detected initial current... Approaching zero, directly assign the initial temperature rise estimate. If the edge gateway connects to monitoring during the operation of the device under load, it is reasonable to assume that the device has reached thermal equilibrium at this moment. The differential term in the control equation is set to zero, and the initial temperature rise is directly derived from the initial power frequency load current, i.e., the value is assigned. This provides a reasonable initial iteration reference point for subsequent smooth iterations. Furthermore, after each difference iteration calculation, if the calculation result... The system performs non-negative limiting processing, setting it to 0 to ensure the non-negative autonomy of the temperature physical parameter.

[0114] For the code refactoring, numerical overflow protection, and integral saturation filtering of first-order difference discrete equations in embedded systems, those skilled in the art can use conventional industrial-grade digital controller software design specifications. The underlying coding and security mechanisms are well-known technologies in the field and will not be elaborated upon here. Through the above calculations, the edge gateway transforms one-dimensional current load data into a dynamic temperature rise estimate reflecting the internal insulation thermal stress state, providing data support for subsequent fluid phase transition assessments.

[0115] See attached document Figure 7 , Figure 7 This is a schematic diagram illustrating the principle of quantitative evaluation of the convection dominance index according to an embodiment of the present invention. In the insulation aging early warning method provided by the present invention, step S50 involves the edge gateway performing thermodynamic calculations based on temperature and physical space parameters, outputting a dimensionless convection dominance index to characterize the gas flow state characteristics inside electrical equipment. Specifically, this may include the following steps:

[0116] S501, the edge gateway combines the ambient base reference temperature with the dynamic temperature rise estimate of the insulation surface to calculate the effective temperature difference and absolute ambient temperature used to drive air thermal convection.

[0117] In this embodiment, the air surrounding the heating element expands due to heat, resulting in a local decrease in density and generating an upward thermodynamic buoyancy in the gravitational field. This buoyancy is the main physical driving force that disrupts the natural diffusion equilibrium of the gas and drives the macroscopic convection transport of the gases characteristic of insulation aging. This is based on the dynamic temperature rise estimate calculated in step S40. In a physical sense, this represents the temperature rise of the insulating heating surface relative to the surrounding ambient gas. The edge gateway extracts this value and sets it as the current sampling time. The effective temperature difference, i.e. .

[0118] To prevent negative temperature rises due to numerical truncation errors in the underlying difference equations under zero-input or extremely low-load conditions, the edge gateway performs non-negative limiting processing after extracting the temperature difference. Specifically, when determining... At that time, the program forces the assignment. This ensures the self-consistency of the physical logic. Furthermore, to meet the consistency requirements of the unit system in subsequent fluid dynamics formulas and to avoid anomalies when calculating the expansion coefficient, the edge gateway calls the obtained environmental baseline reference temperature. The thermodynamic absolute temperature is converted from Celsius to Kelvin using the following formula: .

[0119] S502, the edge gateway introduces aerodynamic parameters to calculate the dimensionless convection dominance index, which characterizes the internal flow pattern of the device;

[0120] The actual flow pattern of a gas depends not only on the temperature difference but also on the combined effects of the physical geometry of the equipment and the viscous barrier effect of the air itself. Combining fluid dynamics and heat transfer theory, the edge gateway employs a dimensionless thermodynamic criterion number to quantify the antagonistic relationship between buoyancy and viscosity. Substituting the effective temperature difference and absolute ambient temperature, the edge gateway calculates the dimensionless convection dominance index. The specific calculation formula is as follows:

[0121] ;

[0122] In the above formula, This is the constant of gravitational acceleration, usually taken as a value of This represents the volumetric expansion coefficient of air at the current sampling time. Based on the ideal gas assumption, the volumetric expansion coefficient can be equivalently replaced by the reciprocal of the absolute ambient temperature, meaning it can be directly assigned a value during program execution. . The characteristic length inside the electrical equipment characterizes the spatial scale of hot airflow development. As a preferred approach, this parameter can be determined during system initialization by measuring the effective vertical exhaust distance from the heating conductor to the top of the cabinet. Kinematic viscosity is the kinematic viscosity of air inside electrical equipment, used to measure the attenuation effect of internal air friction on convective motion. In a typical substation operating environment (e.g., -10°C to 50°C), kinematic viscosity serves as a simplified engineering implementation method. A constant can be approximated. If high accuracy is required for the on-site assessment, the edge gateway can also use a pre-stored standard air thermal property parameter table to... The parameter is obtained by performing dynamic interpolation addressing.

[0123] Through this formula, the edge gateway physically integrates multi-source heterogeneous temperature variables with solid-state spatial structure variables. The calculated convection dominance index... The larger the value, the more dominant thermal buoyancy is compared to viscous resistance, indicating that the characteristic gas tends to be transported over long distances via rapid thermal convection within the enclosed equipment. Through the above evaluation steps, the system completes the logical mapping from the thermal temperature rise state to the aerodynamic flow characteristics, providing a quantitative basis for the subsequent conditional branch comparison of the flow phase transition critical threshold.

[0124] See attached document Figure 8 , Figure 8 This is a logic diagram for comparing critical thresholds of flow phase transition according to an embodiment of the present invention. In the insulation aging early warning method provided by the present invention, step S60 involves the edge gateway extracting the calculated dimensionless convection dominance index and performing threshold comparison in conjunction with a preset hysteresis comparison mechanism to determine the macroscopic transport mode of the gas inside the current electrical equipment. Specifically, this may include the following steps:

[0125] S601, the edge gateway is configured with a tolerant dual critical threshold in the memory to delineate the flow evolution interval;

[0126] In this embodiment, during the propagation of the characteristic gas of insulation aging, the change in the internal airflow state is not instantaneous, but rather involves an evolutionary range from laminar diffusion to turbulent convection. If a single hard threshold is used for logical branch judgment, frequent switching of the system's judgment state can easily occur when the convection dominance index fluctuates slightly near the critical point. To ensure the stability of the flow pattern output and eliminate high-frequency jitter dead zones in algorithm execution, the edge gateway is configured with a tolerant dual critical threshold in its memory, specifically including a lower limit phase transition threshold. Phase transition threshold with upper limit .

[0127] As a preferred approach, the aforementioned lower limit phase transition threshold Phase transition threshold with upper limit The value of is usually empirically mapped based on the critical range of the Grashof number in classical fluid mechanics. For specific models of electrical equipment, technicians can calibrate the parameters through prior computational fluid dynamics simulations or closed-container heating tests. Generally, the upper limit of the phase transition threshold corresponds to the physical critical point of the transition from natural diffusion to thermal convection. The magnitude can be set to to Between; to construct a reasonable anti-interference hysteresis range, the lower limit phase transition threshold. It is usually set at 70% to 85% of the upper limit.

[0128] S602, the edge gateway performs state transition judgment based on the preset hysteresis comparison algorithm and dynamically updates the state flag bit;

[0129] The edge gateway maintains a status flag in memory that represents the gas transport mode. To ensure the integrity of the algorithm logic during system cold start, the edge gateway resets this status flag to its default value during the initialization phase. This characterizes the natural diffusion state during the initial heating phase. After entering a steady-state cycle, when the gas transport mode at the previous sampling time is dominated by natural diffusion, i.e. When the currently calculated convection dominance index is... The edge gateway determines that thermal buoyancy has overcome viscous drag and flips the status flag to the thermal convection-dominated mode, i.e., assigns a value. If the conditions are not met, the original state is maintained. Conversely, if the previous sampling time was already in a state dominated by thermal convection (i.e., When the index falls and satisfies the following conditions, only when the index falls and satisfies the following conditions. When the edge gateway determines that the thermal driving force has decayed, it resets the status flag to natural diffusion dominance, i.e., assigns a value... If it is within the hysteresis range, the convection state will continue.

[0130] S603, the edge gateway performs physical reconstruction of the spatial adjacency topology of the graph neural network based on the status flag bits;

[0131] The aforementioned status flag Mt characterizes the current physical distribution state inside the electrical equipment. When Mt=0, it indicates that the characteristic gas in the enclosed space mainly relies on molecular thermal motion to slowly diffuse isotropically outwards. When Mt=1, it indicates that the rising airflow generated by the heat source has changed the original concentration distribution field, and the gas exhibits obvious bottom-up anisotropic transport characteristics. For these two distinctly different physical distribution states, the edge gateway uses this status flag as a control parameter for the graph neural network preprocessing module, dynamically configuring the spatial adjacency matrix of the graph nodes. Specifically, when Mt=0, the graph neural network calls the static undirected adjacency matrix, whose edge weights are determined by the reciprocal of the three-dimensional Euclidean distance between odor sensor nodes to characterize isotropic concentration gradient diffusion. When Mt=1, the graph neural network switches to a dynamic directed adjacency matrix. The system assigns higher propagation weights to the bottom-up directed edges based on the direction of gravity and the effective temperature difference, thereby simulating the directional transport mechanism of the hot airflow. This method of dynamically switching the network topology based on the actual flow pattern allows the data flow at the bottom layer of the model to better match the actual physical space dynamics inside the device, which helps to reduce feature extraction bias caused by errors in the bottom layer topology.

[0132] See attached document Figure 9 , Figure 9 This is a schematic diagram illustrating the principle of proxy source anchoring and time-delay calculation under thermal convection conditions according to an embodiment of the present invention. In the insulation aging early warning method provided by the present invention, in step S70, when the edge gateway determines, based on the previous comparison step, that the current electrical equipment is in a thermal convection-dominated mode (i.e., the status flag bit...),... When this occurs, the proxy source anchoring and time delay calculation mechanism is triggered to achieve logical time compensation and alignment of spatially distributed heterogeneous sensor data. Specifically, this may include the following steps:

[0133] S701, the edge gateway extracts the geometric center of the primary conductive circuit of the electrical equipment and constructs the proxy source anchoring reference in the internal three-dimensional coordinate system;

[0134] In real industrial operating environments, the specific physical coordinates of microscopic degradation of insulation materials exhibit strong randomness, making it difficult for monitoring systems to identify the true location of characteristic gas release sources in the early warning stage. Without a clear spatial reference origin, subsequent gas transmission time estimation will fall into a calculation dead zone due to missing variables. To address this lack of a reference point, the edge gateway constructs a proxy source anchoring mechanism within its internal three-dimensional coordinate system. In this embodiment, considering that insulation degradation often accompanies or occurs adjacent to major electrical heating nodes, the edge gateway retrieves the factory-preset internal structural three-dimensional parameters or engineering model of the equipment to extract the geometric center of the primary conductive circuit of the electrical equipment, defining its coordinates as the physical location of the proxy source. This proxy source serves as an equivalent virtual starting point for the diffusion of the characteristic gas, providing a definite spatial physical reference for subsequent time-delay calculations.

[0135] S702, the edge gateway combines the effective temperature difference and aerodynamic parameters to calculate the dynamic convection velocity of the characteristic gas in the vertical direction. After establishing the proxy source, the system needs to quantify the gas transport dynamic parameters under thermal convection. Driven by thermal buoyancy, the characteristic gas exhibits a bottom-up directional flow trend. The edge gateway, combining the effective temperature difference obtained in the previous steps, calculates the dynamic convection velocity in the vertical direction. The specific conversion formula is as follows;

[0136] ;

[0137] Where g is the gravitational acceleration constant, The coefficient of volume expansion of air at the current sampling time; For effective temperature difference, The characteristic length inside the electrical equipment. This is a preset empirical reduction factor for convection velocity, used to compensate for the obstruction and turbulence effects caused by internal structural components such as baffles and supporting insulators on the upward flow of air. As a preferred method, this reduction factor... The value range is typically set between 0.3 and 0.7, and technicians can obtain specific values ​​through calibration tests of the wind speed sensor array before shipment. To prevent underflow anomalies in calculation results caused by temperature differences approaching zero, the edge gateway sets a lower limit protection threshold for dynamic convection velocities. For example, in the program, velocity values ​​below 0.01 m / s are forcibly clamped to 0.01 m / s.

[0138] The S703 edge gateway employs a spatial dimensionality reduction and decoupling strategy to calculate the theoretical physical time delay of characteristic gas transmission to each odor sensor node.

[0139] Based on the obtained dynamic convection velocity, the edge gateway independently calculates the gas transport delay for each sensor node in the odor sensing array. Since the sensor nodes are distributed at different locations in three-dimensional space, the time required for each to receive the characteristic gas released from the proxy source varies significantly. For coordinates... The For each odor sensor node, the edge gateway uses a spatial dimensionality reduction and decoupling strategy to calculate its corresponding transmission delay. The corresponding calculation formula is as follows:

[0140] ;

[0141] In the above formula, the first term on the right-hand side of the equation represents the time required for gas to be transported vertically by high-speed thermal convection; the second term represents the time required for gas to diffuse naturally on a horizontal surface by concentration gradient. This is the background natural diffusion rate constant at room temperature, typically determined based on the dynamic characteristics of insulating characteristic gas molecules. The magnitude is significant. Through the two-dimensional decoupled calculation of the superposition of vertical convection and horizontal diffusion described above, the edge gateway obtains the theoretical transmission delay time from the proxy source to each sensor node.

[0142] S704, the edge gateway maps the physical transmission delay of each node to the offset index of the underlying memory, and performs synchronous alignment extraction of asynchronous data.

[0143] After acquiring the continuous physical time delays, the system maps them to the discrete data buffer space of the underlying hardware. The edge gateway then maps the transmission time delays of each node... Divide by a fixed sampling period Then, round to the nearest integer to obtain the corresponding buffer offset index. Considering the limited memory allocation in embedded systems, and to prevent array out-of-bounds errors in the calculated addressing index due to excessive physical distance, the edge gateway introduces an addressing limiting protection mechanism. Specifically, the program internally calculates the index... Maximum buffer depth allocated by the system If a comparison is made, Then force assignment Subsequently, the edge gateway uses this index value to trace back to the first-in-first-out circular buffer corresponding to the odor sensor node and extracts the corresponding historical concentration data.

[0144] This asynchronous data extraction mechanism based on physical time delay can effectively compensate for the physical time difference in the transmission of characteristic gases in complex spaces, recombining physically correlated data collected by different nodes at different times into synchronized feature vectors on the same logical cross-section. For the specific execution logic of matrix translation transformations in the three-dimensional spatial coordinate system and floating-point rounding operations, those skilled in the art can use conventional embedded linear algebra libraries and basic mathematical functions for processing; the underlying execution instructions are well-known technologies in the field and will not be elaborated upon here. After the above calculations, the edge gateway provides a spatiotemporally compensated structured data source for the subsequent graph neural network input.

[0145] See attached document Figure 10 , Figure 10 This is a schematic diagram of a fault-tolerant bypass mechanism under the natural diffusion state according to an embodiment of the present invention. In the insulation aging early warning method provided by the present invention, in step S80, when the edge gateway determines that the current electrical equipment is in the natural diffusion dominant mode (i.e., the status flag bit...),... When this occurs, the system triggers a fault-tolerant bypass mechanism to reduce the risk of computational divergence under low wind speed conditions and ensure the smooth extraction of feature data. Specifically, this may include the following steps:

[0146] S801, the edge gateway stops the time-delay addressing process based on the status flag bit and forces the logical addressing index of each odor sensor node to be zero;

[0147] In this embodiment, under physical conditions dominated by natural diffusion, the effective temperature difference inside the electrical equipment is small, and gas molecules mainly rely on Brownian motion to slowly propagate in all directions. If the dynamic time-delay calculation logic under thermal convection is continued, substituting the lower wind speed into the denominator will cause the calculated transmission time to tend towards infinity, leading to frequent triggering of the buffer addressing limit protection and the extraction of redundant historical data that is out of sync with the current physical state. To prevent the aforementioned computational divergence and logic dead zones, the edge gateway detects... At this time, the spatiotemporal decoupling and offset calculation stages are skipped. Internally, the buffer offset indices of all odor sensor nodes are uniformly assigned to zero, i.e. Under this bypass mechanism, the edge gateway extracts the real-time raw concentration data of each node at the current sampling time, and downgrades the asynchronous time delay compensation to synchronous cross-sectional sampling.

[0148] The S802 edge gateway introduces a low-pass filtering algorithm to smooth the raw concentration data of synchronous sampling in order to suppress background noise interference under low concentration gradients.

[0149] Because the macroscopic concentration changes during natural diffusion are subtle, the effective signal output by the sensor is easily interfered with by thermal noise from the hardware loop or ambient white noise. To improve the signal-to-noise ratio, the edge gateway uses a first-order inertial low-pass filter algorithm for data preprocessing after extracting the real-time raw concentration data. For the... Each odor sensor node outputs a filtered concentration value. The recursive calculation formula is as follows:

[0151] ;

[0152] In this formula, Characterizes the real-time raw odor concentration data extracted at the current sampling moment; Characterizes the filtered concentration data stored by the edge gateway at the previous sampling time. These are the preset filter weight coefficients. As a preferred method, considering the large time constant of the natural diffusion process and the slow accumulation of characteristic gases, The value is typically set between 0.05 and 0.15. For cold starts when the system is first powered on, since there is no historical data stored in memory, the edge gateway... The filter output is initialized directly to the current sampled value at all times, i.e., the value is assigned. This ensures the logical completeness of the filtering algorithm in the time domain recursion.

[0153] S803, the edge gateway constructs a static undirected adjacency matrix based on spatial Euclidean distance, assembles node feature tensors, and completes the reduced-order organization of feature data in the peer structure;

[0154] After data smoothing extraction, the system needs to provide normalized tensor topology input to the graph neural network. In the natural diffusion state, the gas lacks definite directional transport kinetic energy, and its isotropic diffusion characteristics make the original proxy source anchoring unsuitable for the current fluid dynamics. Therefore, the edge gateway uses a static undirected adjacency matrix to characterize the spatial correlation between nodes. The elements in this matrix... Used to quantify the The sensor and the first The weighting of data flow between individual sensors is calculated using the following formula:

[0155] ;

[0156] In the formula, Indicates the first The node and the first The physical Euclidean distance between nodes in a three-dimensional coordinate system, in meters, is acquired during initial device installation through manual ranging or 3D modeling and stored in the edge gateway's non-volatile memory. To prevent the same node from calculating itself (i.e.) The small constant compensation amount that triggers an error when the denominator is zero is usually set to 0. .

[0157] After completing the adjacency matrix After element-wise computation, in order to form a standard data input stream that can be recognized by a graph neural network, the edge gateway further extracts all... Real-time concentration data from odor sensor nodes after filtering Stack them according to the system's preset node numbering order, and splice them to generate a dimension of Node feature matrix Subsequently, the edge gateway will use an adjacency matrix representing spatial topological relationships. With the node feature matrix representing the physical attributes of the business Package and combine them to generate structured graph data objects. .

[0158] This processing method helps avoid algorithm distortion under weak thermal drive and provides a well-structured, dimensionally adapted data format for stable inference of subsequent graph neural networks under low load conditions. For the storage structure of matrix basic elements and the traversal and assignment operations of two-dimensional arrays, those skilled in the art can use conventional memory addressing algorithms; the underlying execution details are well-known in the field and will not be elaborated here.

[0159] See attached document Figure 11 , Figure 11 This is a schematic diagram illustrating the principle of pointer offset spatiotemporal alignment and aerodynamic adjacency matrix generation according to an embodiment of the present invention. In the insulation aging early warning method provided by the present invention, step S90, after obtaining the time-delay indexes of each sensor under the thermal convection-dominated mode, the edge gateway executes a pointer offset addressing and directed graph topology construction mechanism to generate a graph data object conforming to fluid dynamics characteristics. Specifically, this may include the following steps:

[0160] S901, the edge gateway uses a ring buffer management mechanism to perform pointer offset operations and extract spatiotemporally aligned historical feature data;

[0161] In this embodiment, the edge gateway allocates contiguous memory space for each odor sensor node to construct a network with a fixed depth. A first-in, first-out circular buffer. To map physical time delays to underlying storage addresses, the edge gateway uses the offset index obtained in the aforementioned steps. The current write pointer is offset backwards. Considering the linear incrementing nature of memory addresses and the circular structure of the buffer, direct subtraction could lead to negative pointers exceeding the bounds. Therefore, the edge gateway uses modulo operations to calculate the actual read pointers for historical data at each node. The corresponding calculation formula is as follows:

[0162] ;

[0163] in, This represents the memory pointer position at the current sampling moment when the system writes the latest concentration data to the circular buffer, with `mod` representing the modulo operation. By introducing this offset and modulo mechanism, the program can effectively prevent memory addressing out of bounds and locate the time-compensated historical data register slot, thereby extracting the aligned feature concentration value from the underlying register. .

[0164] S903, the edge gateway calculates and generates a directed aerodynamic adjacency matrix based on the asymmetric diffusion physics law under thermal convection;

[0165] Unlike the static equivalent under natural diffusion, gas molecules in the thermal convection-dominated mode exhibit bottom-up directional migration due to thermal buoyancy. To enable the input topology of the graph neural network to characterize this asymmetric hydrodynamic feature, the edge gateway constructs a directed aerodynamic adjacency matrix based on the spatial relative height of the sensors. The elements in this matrix... Characterizing the first The gas released from the first node is directed towards the second node. The smoothness of transmission between individual nodes. The edge gateway calculates the vertical height difference between two nodes.

[0166] when When, it indicates the target node Located in the source node Above it, following the natural upward path of the hot airflow, its adjacency weight is calculated using the following formula:

[0167] ;

[0168] in, and The and the first The physical location of each odor sensor node in the internal three-dimensional coordinate system is determined and stored in the non-volatile memory of the edge gateway during initial device installation; λ is a preset vertical aerodynamic compressibility coefficient, used to characterize the lower transmission resistance in the vertical direction compared to the horizontal direction under convection. As a preferred method, The value range is usually set between 0.1 and 0.5; To prevent the small constant that causes computational overflow due to a denominator of zero, it is usually taken as... .

[0169] when This indicates that the gas needs to overcome thermal buoyancy for reverse or horizontal transport, significantly reducing its hydrodynamic transport probability. The edge gateway applies an attenuation penalty to connectivity in this direction, calculated as follows:

[0170] ;

[0171] in, The three-dimensional Euclidean distance between the two nodes. The preset reverse current penalty factor is typically between 0.01 and 0.1 to characterize the high impedance characteristics of reverse transmission.

[0172] S904, the edge gateway combines and encapsulates the feature matrix and the aerodynamic adjacency matrix to output a structured dynamic graph data object;

[0173] After the aforementioned asymmetric weighting operation, the edge gateway generates a dynamic aerodynamic adjacency matrix with clearly defined directional attributes. The dimension of this matrix is... Its numerical distribution directly maps the airflow channel characteristics under the current thermal conditions. Subsequently, the system will use the feature matrix reflecting the physical state of the nodes. Adjacency matrix reflecting fluid transport topology Package and combine the data to generate a directed graph data object suitable for the current convection conditions. .

[0174] For the memory address allocation and contiguous data block movement operations of the circular buffer, those skilled in the art can use conventional pointer management instructions and direct memory access mechanisms to implement them. The underlying hardware interface calls are well-known technologies in the field and will not be elaborated upon here. Through the aforementioned time-delay addressing and directed graph reconstruction mechanism, the algorithm logic not only reconstructs the airflow transmission channel driven by heat, but also provides directed topological support that conforms to the physical laws of the scene for the subsequent graph convolution operation of the graph neural network.

[0175] See attached document Figure 12 , Figure 12 This is a schematic diagram illustrating the static feature extraction and spatial topology degradation principle according to an embodiment of the present invention. In the insulation aging early warning method provided by the present invention, step S100, under the natural diffusion-dominated mode, involves the edge gateway inputting the generated static undirected graph data object into a pre-deployed graph convolutional neural network model to perform layer-by-layer feature aggregation and equipment insulation status assessment. Specifically, this may include the following steps:

[0176] S1001, the edge gateway performs self-loop compensation and symmetric normalization on the static undirected adjacency matrix to construct a Laplacian operator suitable for graph convolution operations;

[0177] In the natural diffusion-dominated mode, due to the lack of directional kinetic energy in gas molecules, the spatial topology of the system degenerates into an isotropic symmetric structure. To prevent the original concentration information of a node from being masked by its neighboring nodes during feature aggregation, the edge gateway introduces a self-connection mechanism into the static undirected adjacency matrix, constructing an adjacency matrix with self-loops. ,in For dimension The identity matrix. Based on this structure, the system calculates the corresponding degree matrix. Its diagonal elements are obtained by summing the rows of the adjacency matrix, i.e. Due to the introduction of self-loop matrices diagonal elements The value is always greater than or equal to 1, which avoids the risk of a dead zone with a zero denominator that might be triggered in subsequent inversion operations from the underlying algorithm logic. After completing the self-loop compensation, the edge gateway performs symmetric normalization on the matrix to generate the core Laplacian operator for graph convolution. This operator can smooth the feature propagation weights between nodes, helping to suppress drastic numerical fluctuations or gradient anomalies during the forward propagation process of the network.

[0178] S1002, the edge gateway performs local aggregation and high-dimensional mapping of static features on the degraded spatial topology through a multi-layer graph convolutional network structure;

[0179] In this embodiment, the pre-built graph convolutional neural network inside the edge gateway includes two spatial graph convolutional layers and a global pooling module. The system uses the node feature matrix representing the physical attributes of the nodes. (Its initial dimension is) and the calculated Laplace operator The first graph convolutional layer of the synchronous input network. This layer maps one-dimensional concentration features to a high-dimensional latent space through a learnable parameter matrix to extract richer local features. Its calculation formula is as follows:

[0180] ;

[0181] In the formula, For dimension The weight matrix, This is the output feature dimension of the second layer; The corresponding dimension is The bias vector. After two layers of convolution operations, all... High-dimensional feature matrix of sensor nodes (dimension is) Afterwards, to shield the global diagnosis from interference caused by changes in the number or arrangement of local sensor nodes, the system introduces a global average pooling (Readout) operation to reduce the dimensionality of the feature tensors. Specifically, the system... By node space dimension The formula for calculating the mean is as follows: Thus generating a dimension of Graph-level eigenvectors This vector set characterizes the feature distribution of the global insulation state within the entire electrical equipment.

[0182] S1003, the edge gateway uses the fully connected layer to perform linear transformation and probability normalization on the graph-level feature vector, and outputs the corresponding insulation aging status diagnosis results.

[0183] After feature aggregation, the data flows to the network's classification decision layer for physical state mapping. The edge gateway then maps the graph-level feature vectors... The input is fed into a single-layer fully connected network, and a normalized exponential function is used to calculate and map the classification probability.

[0184] ;

[0185] in, Here is the weight matrix of the fully connected layer, with dimension 1. , Represents the total number of preset insulation state categories; For dimension The bias term; The output dimension is The probability distribution vector is obtained. The different dimensions of this vector directly correspond to the specific operational physical state of the insulation material inside the electrical equipment, such as probability assessment indicators covering categories like normal operation, early pyrolysis aging, and partial discharge degradation. As a preferred method, the edge gateway extracts the probability distribution vector. The category label corresponding to the component with the largest median value is used as the final insulation aging early warning diagnosis result for output or reported to the host computer.

[0186] S1004: Researchers completed the offline training and parameter fixing of the graph convolutional neural network based on historical experimental samples and the cross-entropy loss function.

[0187] To improve classification accuracy during the online inference phase, the system requires thorough offline model training using historical data before equipment deployment. The sample set needed for model training comes from recorded data collected from the same type of electrical equipment under artificial accelerated aging test platforms and historical real-world operating scenarios. Each independent training sample contains a graph input tensor at the corresponding sampling time. Its corresponding truth label data This information originates from the synchronously executed insulating oil chromatographic analysis report or the actual aging status confirmed during power outage maintenance, and is represented at the program's underlying layer using a unique thermal encoding format.

[0188] During the offline training phase, the computing platform inputs multiple batches of samples into the model to perform forward propagation and obtain the predicted probability distribution. The system uses the cross-entropy loss function. To quantify the statistical difference between the predicted distribution and the true label:

[0189] ;

[0190] in, This represents the total number of insulation state categories. Based on the loss assessment value calculated in each round, the computing platform uses an adaptive moment estimation optimizer to execute the backpropagation algorithm, calculate the gradient, and iteratively update the weight parameters of each layer within the network. , , And the corresponding bias term. In this process, the initial learning rate of the Adam optimizer is preferably set within the range of 0.001 to 0.01. When the loss value... After several consecutive training epochs, the rate of decrease tends to level off, and the absolute value converges to a preset threshold range (e.g., 10). to Within a certain timeframe, the system determines that the model training has converged and triggers the termination mechanism. The finalized model weight matrix and network hyperparameters will be packaged and stored, and burned into the non-volatile memory of the edge gateway to provide underlying parameter support for subsequent real-time online inference.

[0191] See attached document Figure 13 , Figure 13 This is a schematic diagram of the forward inference and state mapping principle of a cross-modal graph neural network according to an embodiment of the present invention. In the insulation aging early warning method provided by the present invention, step S110, under the heat convection-dominated mode, the edge gateway performs cross-modal fusion of the spatiotemporally aligned odor concentration features and dynamic temperature rise features, and substitutes them into a pre-deployed directed graph neural network model to perform forward inference, so as to output the insulation aging diagnosis result for the current environmental and electrical load coupling conditions. Specifically, it may include the following steps:

[0192] S1101, the edge gateway performs tensor splicing and dimension unification operations to construct a multimodal node feature matrix that integrates thermal stress and characteristic gas distribution;

[0193] In this embodiment, a single odor concentration data point is insufficient to reflect the nonlinear accelerated aging characteristics of heat-sensitive insulating materials at different temperatures. From a physical perspective, the increase in temperature inside electrical equipment not only constitutes a thermal stress that accelerates insulation degradation, but also acts as a catalyst to significantly increase the release rate of decomposition gases from the insulating material. Take the estimated value of dynamic temperature rise of the insulation surface at the current sampling time calculated in step S40. Since this parameter was originally a system-level scalar reflecting the macroscopic state of the system, the edge gateway expands its dimensions to generate a dimension of... Temperature rise characteristic column vector The value of each element is equal to .

[0194] Considering the significant differences in physical dimensions and absolute scale between odor concentration data and temperature rise values, direct fusion would lead to network weights being dominated by large-scale features during training, causing small-scale odor features to be overlooked. Therefore, the edge gateway uses the Z-score normalization algorithm to normalize both data to zero mean and unit variance before splicing. Subsequently, the program converts the normalized node feature matrix... With temperature rise feature column vector The concatenation is performed along the feature dimension (i.e., the column direction), and the calculation formula is as follows:

[0195] ;

[0196] In the formula, This represents the column concatenation operation of the matrix. After this combination process, the output is a multimodal node feature matrix. The dimension has been changed to This allows the network to simultaneously perceive local concentration gradients and global thermal stress environment during subsequent convolution processes.

[0197] S1102, the edge gateway performs out-degree normalization processing on the directed aerodynamic adjacency matrix, which helps to suppress information transmission deviations caused by asymmetric topology;

[0198] After completing the multimodal combination of data features, the system needs to constrain the spatial topology transfer rules. Unlike undirected graphs in the natural diffusion state, the aerodynamic adjacency matrix in the thermal convection state... The system exhibits an asymmetric structure. If the original matrix is ​​directly used for feature propagation, nodes with large out-degrees will cause a sharp increase in the feature values ​​of the target node. To stabilize the feature scale, the edge gateway introduces an out-degree normalization mechanism. The system is a matrix... Adding self-loops preserves the node's own state characteristics, i.e. ,in for The identity matrix. Based on the matrix after introducing self-loops, the system calculates the out-degree matrix in diagonal form. Its diagonal elements are the sum of the row elements of the matrix, that is... Due to the identity matrix The existence of each diagonal element This avoids the potential calculation anomaly of a zero denominator that might occur during subsequent inversion operations. The edge gateway generates a directed transition matrix by multiplying the inverse of the out-degree matrix with the adjacency matrix containing self-loops. This transition matrix controls the information flow and weight distribution in the asymmetric network.

[0199] S1103, the hidden layer network of the edge gateway-driven graph neural network performs asymmetric spatial aggregation and nonlinear mapping of cross-modal features;

[0200] To extract deep joint features, the edge gateway will use a directed transition matrix. With multimodal node feature matrix The input is fed into a pre-trained directed graph convolutional network. The first layer of graph convolution utilizes directed edge relationships to aggregate multimodal information of neighboring nodes in the reverse direction along the airflow channels. The specific calculation formula is as follows:

[0201] ;

[0202] in, This is the first layer's learnable weight matrix. Since the input feature dimension is 2, its matrix dimension is set to... , is the hyperparameter dimension of the hidden layer node features. For dimension is The bias vector is denoted by ReLU, and ReLU is the activation function. After mapping, the output is the first hidden layer feature matrix. Dimension changed to .

[0203] The data stream then enters a second convolutional structure to obtain deeper spatial joint semantics:

[0204] ;

[0205] In the formula, The dimension is The dimension is This represents a high-order feature dimension. After two rounds of asymmetric feature transfer, the system extracts a node-level high-order representation that integrates electrothermal coupling stress and the spatiotemporal characteristics of aerodynamic diffusion. Its matrix dimension is

[0206] S1104, the edge gateway calculates a graph-level representation through pooling aggregation and full connectivity layer, mapping the probability distribution of the insulation state of the output device.

[0207] To meet overall diagnostic needs, the edge gateway invokes a global average pooling operation to perform mean aggregation and dimensionality reduction on node-level features along the spatial dimension. The calculation formula is as follows: .in, Characterization matrix The Middle The high-order feature row vectors corresponding to each sensor node are generated from this. This produces a dimension of... Graph-level eigenvectors This vector is input to the fully connected layer and, together with the Softmax function, completes the state transition to the business dimension.

[0208] ;

[0209] in, Here is the weight matrix of the fully connected layer, with dimension 1. , This represents the number of preset aging state categories (such as normal operation, slight overheating, moisture-induced degradation, partial discharge, etc.). For dimension The bias term. Calculate the output probability distribution vector. Dimensions The edge gateway filters out the component with the highest value in the vector and extracts the corresponding business physical status label as the final warning result.

[0210] S1105: Researchers collected multimodal joint operation samples, configured the cross-entropy loss function, and completed offline supervised training of cross-modal graph neural networks;

[0211] To improve the classification confidence of online inference, the model needs to be iteratively trained using sufficient joint historical samples before deployment. The training sample set comes from bench test data of homogeneous high-voltage equipment under multiple temperature zones and load conditions. Each sample encapsulates current-derived temperature rise features, measured concentration matrices, and aerodynamic adjacency matrices. (Label data) The results were obtained through actual oil and gas chromatography analysis and converted into unique thermal codes. The computational platform utilizes the probability output from forward propagation. Calculate cross-entropy loss with the true label :

[0212] ;

[0213] Based on the aforementioned loss metrics, the platform employs the Adam optimizer to execute the backpropagation mechanism. As a preferred approach, the initial learning rate is set between 0.001 and 0.005, and the system updates the network weights round by round. , and The convergence threshold is reached when the variance of the loss value drops below a preset threshold for ten consecutive training epochs (e.g., ...). When the iteration ends, the system terminates. The finalized network parameter model is then burned into the edge gateway to support real-time calculations in the field environment.

[0214] For the lifecycle management of multidimensional tensors in microprocessor memory and the underlying implementation of automatic chained differentiation in the backpropagation process, those skilled in the art can complete the development based on the existing deep learning framework compilation interface. The underlying computation graph scheduling is a well-known technology in this field and will not be elaborated here.

[0215] See attached document Figure 14 and attached Figure 15 , Figure 14 This is a visualization comparison curve of cross-modal feature splicing and graph-level representation extraction according to an embodiment of the present invention; Figure 15 This is a bar chart comparing the accuracy and false alarm rate of insulation state diagnosis for different models according to an embodiment of the present invention.

[0216] I. Specific Application Scenario Example: At a high-voltage substation during peak summer power supply, a 220kV oil-immersed main transformer was operating under heavy load for an extended period. The edge gateway in the system performed high-frequency online monitoring of an 8-node high-temperature odor sensor array deployed within the transformer (nodes covering the bottom area, the middle of the windings, and the top outgoing terminals). Due to the high temperatures of midsummer and the near-full load, significant thermodynamic forces were generated inside the equipment.

[0217] During the system monitoring phase, the edge gateway extracts the power frequency load current in real time and substitutes it into the first-order equivalent thermal resistance-capacitance network model. The calculated estimated dynamic temperature rise of the insulation surface, Trie, reaches 55℃. Further calculations show that its dimensionless convection dominance index has exceeded the upper limit phase change threshold. The system determines that the internal gas transmission state has switched to the thermal convection dominance mode (i.e., status flag Mt=1), and then triggers the proxy source anchoring and time-delay calculation mechanism.

[0218] Entering the cross-modal manifold solution stage, the edge gateway performs pointer offset addressing operations through a first-in-first-out circular buffer, performs time compensation on the top-positioned sensor nodes, and extracts aligned concentration data from historical moments, thus physically and logically smoothing out the physical time delay of the characteristic gas's upward transport with the hot oil flow. After completing spatiotemporal alignment, the system constructs a directed adjacency matrix characterizing asymmetric aerodynamic properties and enforces out-degree normalization. This effectively prevents feature value inflation caused by excessively large bottom nodes during network inference. Based on this, the system precisely concatenates the aligned multidimensional odor concentration features with the Z-score-normalized temperature rise features along the feature dimension. The cross-modal multidimensional tensor is simultaneously fed into a pre-programmed two-layer directed graph convolutional neural network for forward inference. After asymmetric spatial aggregation and nonlinear mapping are completed in the network's hidden layers, the edge gateway generates a graph-level feature vector through global average pooling and feeds it into a fully connected layer to calculate the probability distribution vector. The main control unit selects the highest probability component (e.g., reaching 0.92) and maps it to the corresponding early pyrolysis aging label. The edge gateway then generates an early warning message with specific aging type and location assessment and reports it to the control center, allowing the system to safely enter the next monitoring cycle.

[0219] To quantitatively evaluate the robustness of the time-delay calculation mechanism and cross-modal multidimensional graph network algorithm in this invention, the research team constructed a hardware-in-the-loop experimental platform within an environmental simulation test chamber. This platform was equipped with a microcurrent transmitter with high-frequency sampling capability and a high-precision chromatograph to monitor the temperature rise of the heat source inside the transformer and the microscopic release rate of multi-component characteristic gases in real time. The entire test chamber was set to a variable load and temperature mode to reproduce the real physical environment evolution of the main transformer from a low-load, light-load period to a high-load, high-temperature condition. The raw multimodal sensor data, hidden-layer tensor characteristic response data, and real oil chromatographic test labels collected during the experiment were fully exported for visualization processing and performance analysis (e.g., ...). Figure 14 (As shown). In Figure 14 In the graph, the X-axis represents the iterative computation level of the network's forward inference, and the Y-axis represents the relative characteristic response amplitude of the manifold within the hidden layers of the graph network. The waveform clearly shows the instantaneous response evolution trend of the cross-modal tensor as it traverses the graph convolutional structure under thermal convection conditions. This waveform, based on real sensor data, intuitively confirms that the Z-score dimension unification and asymmetric out-degree normalization strategy designed in this scheme can effectively suppress the nonlinear gradient spikes that occur when large-scale temperature rise data jumps through different levels.

[0220] III. Effect Comparison and Data Analysis To further verify the technological advantages, a control group was set up for effect comparison (e.g., ...). Figure 15 (As shown). The control group used a conventional graph network classification algorithm with undirected symmetric graph topology, single concentration mode input, and no pointer offset time delay calculation mechanism; the experimental group used the thermal convection spatiotemporal alignment, cross-modal temperature rise fusion, and asymmetric directed graph inference algorithm proposed in this invention. Under the simulated heavy-load operation of a 220kV transformer platform and the high dynamic temperature rise caused by the environment and load, 1000 fully automated state diagnosis tests were performed on each isomorphic miniature model.

[0221] After the test, the two sets of data were statistically compared. The comparison indicators mainly covered the accuracy of insulation condition diagnosis, false alarm rate, and feature response delay time for early faults. Based on the experimental data statistics, the control group, unable to adapt to the topological asymmetric gas transport caused by thermal buoyancy, frequently experienced spatial information mismatches in the feature extraction stage, resulting in a false alarm rate as high as 14.5%. In the time-series matching stage, misaligned top abrupt data interfered with the network judgment in advance, causing an average fault response delay of up to 45 minutes. Finally, the overall condition diagnosis accuracy output on the full sample test set barely reached 82.3%. In contrast, the experimental group of this invention, by using the quantitative assessment of the convection dominance index and the multi-dimensional perception coupling solution of spatially directed topology, endowed the diagnostic strategy with physical environment adaptability. Combined with the pointer reconstruction based on offset and the cross-modal feature splicing mechanism, the experimental group reduced the false alarm rate to 1.2%, shortened the early fault feature response delay time to the near-zero real-time range, and significantly increased the overall diagnostic accuracy to the high confidence range of 98.4%. The above quantitative comparison data fully demonstrate that the early warning method provided by this invention exhibits stronger feature extraction robustness and diagnostic accuracy when facing complex and variable electrothermal coupling physical fields. It effectively avoids the logical dead zone of traditional single-modal undirected graph algorithms under edge fluid conditions and truly meets the stringent technical requirements for high reliability of operation and maintenance early warning in the all-weather online monitoring scenario of high-voltage equipment in smart grids.

[0222] 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. An insulation aging early warning method based on an odor sensing graph neural network, characterized in that, Includes the following steps: By comparing the current convection dominance index of the electrical equipment with the critical threshold, it can be determined whether the gas transport mode inside the current electrical equipment is a thermal convection dominance mode or a natural diffusion dominance mode. When the gas transport mode is the thermal convection-dominated mode, the physical transport delay of the characteristic gas to each odor sensor node is calculated, and the physical transport delay is mapped to an addressing index to extract historical odor concentration data from the circular buffer. A directed aerodynamic adjacency matrix is ​​constructed based on the spatial relative height of each odor sensor node; the historical odor concentration data is spliced ​​with the dynamic temperature rise characteristics of the electrical equipment to construct a multimodal node feature matrix; The multimodal node feature matrix and the directed aerodynamic adjacency matrix are input into a pre-deployed directed graph convolutional neural network model to output insulation aging diagnosis results. When the gas transport mode is dominated by natural diffusion, the addressing indexes of each odor sensor node are reset to zero, and the real-time odor concentration data at the current sampling time is extracted. A static undirected adjacency matrix is ​​constructed based on the spatial Euclidean distance between each of the odor sensor nodes; The node feature matrix composed of the real-time odor concentration data and the static undirected adjacency matrix are input into a pre-deployed undirected graph convolutional neural network model to output the insulation aging diagnosis result.

2. The insulating aging early warning method based on the odor sensing map neural network according to claim 1, characterized in that, The step of determining whether the gas transport mode inside the current electrical equipment is dominated by thermal convection or natural diffusion by comparing the convection dominance index of the current electrical equipment with a critical threshold specifically includes: Configure a lower limit phase transition threshold and an upper limit phase transition threshold with a tolerance range; When the gas transport mode at the previous sampling time was the natural diffusion-dominated mode, if the currently calculated convection-dominated index is greater than or equal to the upper limit phase change threshold, the gas transport mode is converted to the thermal convection-dominated mode. If the gas transport mode at the previous sampling time was thermal convection-dominated, and the currently calculated convection dominance index is less than or equal to the lower limit phase transition threshold, then the gas transport mode is reset to natural diffusion-dominated.

3. The odor-sensing graph neural network-based insulation aging early warning method according to claim 1, characterized in that, The calculation of the physical transmission delay of the characteristic gas to each odor sensor node specifically includes: extracting the geometric center of the primary conductive circuit of the electrical equipment and constructing it as the physical location of the proxy source for the diffusion of the characteristic gas in the internal three-dimensional coordinate system; By combining the effective temperature difference inside the equipment with aerodynamic parameters, an empirical reduction factor for convection velocity is introduced to calculate the dynamic convection velocity of the characteristic gas in the vertical direction. By employing a spatial dimensionality reduction and decoupling strategy, the time for the characteristic gas to be transported vertically by the dynamic convection velocity and the time for it to diffuse naturally on the horizontal surface by the concentration gradient are independently calculated and superimposed to obtain the theoretical physical transmission delay from the physical location of the proxy source to each of the odor sensor nodes.

4. The insulating aging early warning method based on the odor sensing map neural network according to claim 1, characterized in that, The step of mapping the physical transmission delay to an addressing index to extract historical odor concentration data from the circular buffer specifically includes: The physical transmission delay of each odor sensor node is divided by a fixed sampling period and rounded to the nearest integer. After the maximum buffer depth exceeds the limit protection judgment, the corresponding addressing index is obtained. In the underlying memory space, the write pointer of the current data is subtracted from the address index, and the maximum depth of the buffer is introduced to perform a modulo operation to prevent memory overflow. The location is then determined to the register slot after time compensation alignment, and the corresponding historical odor concentration data is extracted.

5. The odor-sensing graph neural network-based insulation aging early warning method according to claim 1, characterized in that, The construction of the directed aerodynamic adjacency matrix based on the spatial relative height of each of the odor sensor nodes specifically includes: Calculate the height difference between the target odor sensor node and the source odor sensor node in the vertical direction; when the height difference is greater than zero, it is determined to be a forward thermal convection path, and a preset vertical aerodynamic compressibility coefficient and a small constant are introduced to calculate the forward adjacency weight; When the height difference is less than or equal to zero, it is determined to be a reverse or horizontal transmission path, and a preset reverse flow penalty factor and the three-dimensional Euclidean distance between nodes are introduced to calculate the reverse adjacency weight. The directed aerodynamic adjacency matrix with directional attributes is generated by constructing the forward and reverse adjacency weights of each node.

6. The odor-sensing graph neural network-based insulation aging early warning method according to claim 1, characterized in that, The step of concatenating the historical odor concentration data with the dynamic temperature rise characteristics of electrical equipment to construct a multimodal node feature matrix specifically includes: The dynamic temperature rise feature is expanded in dimension to generate a temperature rise feature column vector, and a standardization algorithm is called to normalize the historical odor concentration data and the temperature rise feature column vector to zero mean and unit variance respectively to unify the physical dimensions. The normalized historical odor concentration data and the temperature rise feature column vector are concatenated along the feature dimension to generate the multimodal node feature matrix.

7. The insulation aging early warning method based on odor sensing graph neural network according to claim 1, characterized in that, The step of inputting the multimodal node feature matrix and the directed aerodynamic adjacency matrix into a pre-deployed directed graph convolutional neural network model and outputting insulation aging diagnosis results specifically includes: Self-loop compensation is introduced into the directed aerodynamic adjacency matrix, and out-degree normalization is performed in conjunction with the out-degree matrix of the nodes to generate a directed transition matrix. The directed transition matrix and the multimodal node feature matrix are input into the hidden layer network of the directed graph convolutional neural network model, and asymmetric spatial aggregation and nonlinear mapping are performed along the airflow channel to obtain the node-level high-order expression matrix. The node-level high-order representation matrix is ​​aggregated and reduced in dimensionality along the spatial dimension by calling the global average pooling operation to generate a graph-level feature vector. This feature vector is then mapped by a fully connected layer and a normalized exponential function to a probability distribution vector reflecting the insulation state. The physical state label corresponding to the highest probability is extracted as the insulation aging diagnosis result.

8. The insulation aging early warning method based on odor sensing graph neural network according to claim 1, characterized in that, The extraction of real-time odor concentration data at the current sampling time specifically includes: During cold start, the filtered output is directly initialized to the original odor concentration data collected by each odor sensor node at the current sampling time; During non-cold start times, a first-order inertial low-pass filter algorithm is invoked, and the original odor concentration data is smoothly recursively derived by combining the filtered data stored in the previous sampling time to suppress background noise interference and obtain the real-time odor concentration data.

9. The insulation aging early warning method based on odor sensing graph neural network according to claim 1, characterized in that, The construction of a static undirected adjacency matrix based on the spatial Euclidean distance of each of the odor sensor nodes specifically includes: Obtain the physical Euclidean distance between the i-th odor sensor node and the j-th odor sensor node in three-dimensional space; Calculate the reciprocal of the physical Euclidean distance, add a small constant compensation to prevent the denominator from being zero, and use the calculation result as the connectivity weight element in the static undirected adjacency matrix to construct the static undirected adjacency matrix.

10. The insulation aging early warning method based on odor sensing graph neural network according to claim 1, characterized in that, Before determining the gas transport pattern inside the current electrical equipment, the process includes an offline training step for the directed graph convolutional neural network model and the undirected graph convolutional neural network model, specifically including: Acquire historical multimodal joint operation samples and synchronously represent real aging state labels through a unique encoding format; The running sample is input into the model to perform forward propagation to obtain the predicted probability distribution, and the statistical difference between the predicted probability distribution and the actual aging state label is calculated using the cross-entropy loss function. Based on the calculated cross-entropy loss, the adaptive moment estimation optimizer is used to execute the backpropagation algorithm to iteratively update the network layer weights and bias terms inside the model until the loss fluctuations converge to the preset threshold range, thus completing the model shaping and parameter solidification.

Images