Intelligent monitoring method and system for internal fault of box-type substation

Abstract

The application discloses a kind of box-type substation internal fault intelligent monitoring method and system, related field is related to point digital data processing technology, the method synchronously acquires temperature field, sound field, three-phase current and characteristic gas concentration multi-physical field time sequence signal;Based on transmission entropy, each physical field is analyzed for causal association, transmission entropy and causal delay time are calculated, and after replacement test, multi-physical field causal directed graph is automatically generated;Physical information graph neural network is constructed, and heat conduction equation, acoustic wave equation and current continuity equation are embedded in loss function as physical constraint;With causal directed graph guiding graph attention mechanism, the propagation probability and expected arrival time of fault between each monitoring node are calculated, and the probability weighted fault evolution path set and the diagnostic result containing fault causal chain, origin positioning, prediction evolution path and confidence evaluation are output.The application realizes the physical interpretable box-type substation internal fault intelligent monitoring, and improves accuracy and early warning foresight.

Description

Technical Field

[0001] This application relates to the field of electrical digital data processing technology, and in particular to an intelligent monitoring method and system for internal faults in a prefabricated substation. Background Technology

[0002] Prefabricated substations, as compact sets of power distribution equipment that integrate high-voltage switchgear, distribution transformers, and low-voltage distribution devices according to a specific wiring scheme, are widely used in urban power distribution networks, industrial parks, and new energy power plants. The operational reliability of their internal equipment directly affects power supply quality and electricity safety. Common fault types within prefabricated substations include overload overheating, poor contact, partial discharge, insulation degradation, and short-circuit arcing. In their early stages, these faults often manifest as weak anomalies in multiple physical fields (temperature, sound waves, current, and gas). If the source of the fault can be detected early and accurately located, and its propagation path traced, the expansion of the fault can be effectively prevented, and unplanned outage losses can be reduced.

[0003] In existing fault monitoring methods, various sensor signals are often collected and processed in isolation, or simply spliced ​​and statically weighted at the feature level, failing to reveal the causal transmission relationships between different physical quantities. Some solutions use deep learning to extract features from multi-source heterogeneous data collected by IoT sensors, but these solutions fail to solve the problem of automatically identifying causal driving relationships between multiple physical fields.

[0004] Existing technologies mostly construct purely data-driven models such as CNNs and LSTMs for fault diagnosis. These models suffer from decreased inference reliability when training data is sparse or operating conditions change, and their diagnostic results lack physical evidence, making it difficult to guide on-site maintenance personnel in understanding the fault mechanism. Furthermore, the diagnostic results of existing technologies are mostly binary judgments of whether a fault exists or simple classifications, unable to trace the causal source and propagation path of the fault. While a few solutions mention fault propagation simulation, they lack a complete technical chain from causal discovery to propagation prediction. Summary of the Invention

[0005] To address the technical problems of the prior art, this application provides an intelligent monitoring method and system for internal faults in prefabricated substations. By automatically discovering causal relationships in multi-physics fields based on transfer entropy and predicting fault propagation using physical information graph neural networks, the method enables causal tracing and dynamic deduction of internal faults in prefabricated substations, thereby improving the accuracy, interpretability, and reliability of fault diagnosis.

[0006] This application provides an intelligent monitoring method for internal faults in a prefabricated substation, including:

[0007] Step S10: Construct a physical field sensor network for the internal space of the box-type substation, and synchronously collect multi-physical field time-series signals, including temperature field time-series signals, sound field time-series signals, three-phase current time-series signals, and characteristic gas concentration time-series signals. Step S20: Based on the transfer entropy, perform causal correlation analysis on each physical field signal, calculate the transfer entropy value between each pair of physical fields and the causal delay time that makes the transfer entropy reach its maximum value, and automatically generate a multi-physics causal directed graph with time delay label after the significance verification by permutation test. Step S30: Construct a physical information graph neural network, discretize the internal space of the box-type substation into monitoring nodes and construct a spatial graph structure, and embed the heat conduction equation, sound wave equation and current continuity equation into the loss function as physical constraints. Step S40: Based on the multiphysics causal directed graph guidance graph attention mechanism, the attention coefficients are positively biased along the causal propagation direction; based on the current abnormal event, the propagation probability and expected arrival time of the fault between each monitoring node are calculated, and a probability-weighted set of fault evolution paths is output. Step S50: Based on the multiphysics causal directed graph and the fault propagation probability, output the diagnostic results including the fault causal chain, fault origin location, predicted evolution path and confidence assessment.

[0008] Further, in step S10, the temperature field timing signal The components are arranged in a spatial grid inside the prefabricated substation. Temperature data is collected by one temperature sensor, among which Indicates the first Spatial coordinates of each temperature monitoring point For time indexing; The sound field timing signal The components installed on the internal walls of the prefabricated substation enclosure and the surface of the transformer casing are... Data collected by a voiceprint sensor, among which Indicates the first Spatial coordinates of each voiceprint monitoring point; The three-phase current timing signal , and Data is collected by high-frequency current sensors installed at the entry and exit points of each phase power cable in the box-type substation. The characteristic gas concentration time-series signal Gas data is collected by gas sensors installed at high points inside the prefabricated substation enclosure, as well as in the cable compartment and transformer compartment. The gas types are numbered; the characteristic gases include , , and At least one of them; The temperature field time series signal, sound field time series signal, three-phase current time series signal and characteristic gas concentration time series signal are filtered, denoised and timestamped to construct a multiphysics spatiotemporal data matrix.

[0009] Furthermore, step S20 includes the following detailed steps: Step S201, for any two physical field time series and Calculate the physical field according to the following formula To the physical field Transmission entropy Used to characterize physical fields Past states of the physical field The degree to which it contributes to the ability to predict future states:

[0010] in, and Representing physical fields respectively and physical field At any moment The value of , For physical fields At any moment The value of , For time delay; For joint probability density, In order to be in and under conditions The conditional probability density, In order to be in under conditions The conditional probability density; calculate the time delays for each pairwise time-series signal between the temperature field, sound field, current field, and gas field. The propagation entropy value under ,in , They represent any two different physical fields; Step S202, for each pair of physical fields The delay time that maximizes the propagation entropy is taken as the causal delay time, which is expressed by the formula: ,in Represents physical field To the physical field causal delay time, Indicates taking The independent variable that achieves its maximum value The possible values ​​of ; Step S203: Use the permutation test to test the significance of the transfer entropy value of each pair of physical fields: randomly shuffle the time series of each physical field multiple times, recalculate the transfer entropy distribution, and if the original transfer entropy value is higher than the 95th percentile, the causal relationship is determined to be significant and false causal associations are eliminated. Step S204: Based on the significant causal relationships retained after significance testing, using temperature field anomaly event T, sound field anomaly event P, current field anomaly event I, and gas field anomaly event G as nodes, significant causal relationships as directed edges, and causal delay time... Using edge weights, automatically generate a multiphysics causal directed graph with time delay annotations.

[0011] Furthermore, step S30 includes the following detailed steps: Step S301: Discretize the internal space of the prefabricated substation into Q monitoring nodes and construct a spatial graph structure. This is used to organize the physical field data with spatial location attributes inside a prefabricated substation into graph-structured data, enabling fault characteristics to be transmitted and aggregated between spatially adjacent monitoring nodes. Representing a spatial diagram, and Represent the set of nodes and the set of edges, respectively; Step S302, set the nodes Each node The node features are represented as ,in , , and Representing nodes respectively Temperature, sound pressure, current, and characteristic gas concentration values ​​at the location. and It is obtained by spatial interpolation or nearest neighbor matching of the temperature field and sound field time-series signals according to the node spatial coordinates. The node is assigned a value based on the corresponding phase of the three-phase current timing signal. The value is assigned based on the concentration time-series signal collected by the gas sensor in the corresponding region where the node is located; Step S303: The edge features between nodes in the node set V are represented as physical distance and medium properties between nodes, which are used to characterize the spatial correlation strength between adjacent monitoring nodes and the physical conditions for fault propagation. Step S304: Embed a physical constraint loss term into the loss function of the physical information graph neural network. The total loss function of the physical information graph neural network is constructed by weighting three constraints: the heat conduction equation, the sound wave equation, and the current continuity equation. This can be expressed as a formula:

[0012] in, The task loss for fault classification and propagation prediction is used to measure the error between the predicted value and the true label. This is the total weight coefficient for physical constraints, used to adjust the proportion of physical constraints in the total loss.

[0013] Furthermore, in step S40, the process of the multiphysics-based causal directed graph guided graph attention mechanism includes: Graph attention mechanism in physical information graph neural networks at computation nodes For nodes When the attention coefficient exists in the multiphysics causal directed graph, For directed edges, a positive bias is applied to the attention coefficient, expressed by the formula:

[0014] in, For nodes For nodes The normalized attention coefficient, For nodes For nodes The original attention score, For causal bias weights, This is an indicator function; it takes the value 1 if the condition inside the parentheses is true, and 0 otherwise. Represents a node To the node The directed edges exist in the multiphysics causal directed graph. It is a normalized exponential function.

[0015] The calculation process for the fault propagation probability includes: The fault originated from the internal spatial nodes of the prefabricated substation. To the node The probability of propagation is calculated using the following formula:

[0016] in, Indicates at time Fault from node propagation to nodes The conditional probability, To predict the time step, For nodes For nodes The normalized attention coefficient, and They are nodes and nodes exist The hidden state representation at time step is obtained by forward propagation calculation of the physical information graph neural network; This represents the causal delay time estimated by the propagation entropy. This represents a vector concatenation operation. is the activation function of the physical information graph neural network, used to map the output value to the probability interval [0,1]. The physical information graph neural network finally outputs the failure propagation probability and expected arrival time of each propagation path in the future time window, forming a probability-weighted set of failure evolution paths.

[0017] This application also provides an intelligent monitoring system for internal faults in a prefabricated substation, including: Data acquisition and preprocessing module: used to construct a physical field sensor network inside the box-type substation, and synchronously acquire multi-physical field time-series signals, including temperature field time-series signals, sound field time-series signals, three-phase current time-series signals and characteristic gas concentration time-series signals; Causal correlation analysis module: It is used to perform causal correlation analysis on signals of various physical fields based on the transfer entropy, calculate the transfer entropy values ​​between each pair of physical fields and the causal delay time that makes the transfer entropy reach its maximum value, and automatically generate a multi-physics causal directed graph with time delay annotation after the significance verification by permutation test. Physical Information Graph Neural Network Construction Module: Used to construct a physical information graph neural network, discretize the internal space of the box-type substation into monitoring nodes and construct a spatial graph structure, and embed heat conduction equation, sound wave equation and current continuity equation as physical constraints in the loss function; Causal-guided inference module: used to guide the graph attention mechanism based on multi-physics causal directed graph, so that the attention coefficients are positively biased along the causal propagation direction; calculates the propagation probability and expected arrival time of the fault between each monitoring node based on the current abnormal event, and outputs a probability-weighted set of fault evolution paths; Diagnostic Result Output Module: Based on multiphysics causal directed graphs and fault propagation probabilities, this module outputs diagnostic results including fault causal chains, fault origin location, predicted evolution paths, and confidence assessments.

[0018] This application also provides an intelligent monitoring device for internal faults in a prefabricated substation. The intelligent monitoring device for internal faults in a prefabricated substation includes: a memory, a processor, and an intelligent monitoring program for internal faults in the prefabricated substation stored in the memory and executable on the processor. When the intelligent monitoring program for internal faults in the prefabricated substation is executed by the processor, it implements the above-mentioned method.

[0019] This application also provides a computer program product, which includes an intelligent monitoring program for internal faults in a prefabricated substation. When the intelligent monitoring program for internal faults in a prefabricated substation is executed by a processor, it implements the above-described method.

[0020] This application discloses the following technical effects: This application provides an intelligent monitoring method and system for internal faults in prefabricated substations. By transferring entropy, it automatically discovers the causal transmission direction and delay time between multiple physical fields such as heat, electricity, sound and air. This solves the fundamental defect of existing correlation analysis methods that cannot distinguish between causal relationships and spurious correlations, and does not require relying on human experience to preset the fault propagation sequence.

[0021] In this application, the loss function constraint term of the physical information graph neural network is applied to ensure that the reasoning process conforms to the three basic physical laws of heat conduction, sound wave fluctuation, and current continuity, avoiding the risk of misjudgment when the training data is sparse or the operating conditions change. The attention mechanism guided by the causal graph enables the network to prioritize inference along the causal transmission direction, and each prediction conclusion can be traced back to the physical causal chain.

[0022] The method proposed in this invention outputs a multiphysics causal directed graph in step S20, which provides physical prior guidance for the graph attention in step S40. The graph neural network in step S40 verifies and refines the causal propagation path under physical constraints. The two form a closed loop of "causal discovery → physical constraint deduction → path verification", which has higher diagnostic accuracy and interpretability than using either method alone. Attached Figure Description

[0023] Figure 1 This is a flowchart illustrating an intelligent monitoring method for internal faults in a prefabricated substation, as provided in an embodiment of this application.

[0024] Figure 2 This is a schematic diagram of the structure of an intelligent monitoring system for internal faults in a prefabricated substation, provided as an embodiment of this application. Detailed Implementation

[0025] To make the objectives, technical solutions, and advantages of this application clearer, the following detailed description of this application will be provided in conjunction with the accompanying drawings. The described embodiments should not be considered as limitations on this application. All other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of this application.

[0026] Example 1: This application provides an intelligent monitoring method for internal faults in a prefabricated substation, applied to a 10kV / 0.4kV prefabricated substation with a capacity of 1000kVA; the prefabricated substation contains three main functional areas: a high-voltage switch room, a transformer room, and a low-voltage distribution room; such as Figure 1 As shown, the method includes: Step S10: Construct a physical field sensor network for the internal space of the box-type substation, and synchronously collect multi-physical field time-series signals, including temperature field time-series signals, sound field time-series signals, three-phase current time-series signals, and characteristic gas concentration time-series signals.

[0027] In this embodiment, temperature sensors are deployed inside the prefabricated substation using a three-dimensional spatial grid. A total of 24 temperature sensors are deployed in the transformer room, cable room, and low-voltage room, forming a spatial grid of 4 (x-direction, along the length of the prefabricated structure, spaced 0.6 meters apart) × 3 (y-direction, along the width of the prefabricated structure, spaced 0.5 meters apart) × 2 (z-direction, along the height of the prefabricated structure, spaced 0.7 meters apart). Each temperature sensor is a PT100 platinum resistance temperature sensor with a measurement range of -40℃ to +200℃ and an accuracy of ±0.3℃. Each temperature sensor continuously collects temperature data at a sampling frequency of 1Hz to obtain the temperature field time-series signal. ,in =1,2,...,24 During the sampling period, temperature data is transmitted to the local data acquisition unit via an RS-485 bus.

[0028] Eight acoustic signature sensors are installed on the internal walls of the enclosure and the surface of the transformer casing. These sensors utilize a capacitive microphone array with a frequency response range of 20Hz to 20kHz and a sensitivity of 50mV / Pa. Each sensor collects sound pressure signals at a sampling frequency of 20kHz to obtain the sound field timing signal. ,in =1,2,...,8, The voiceprint data is transmitted to the data acquisition unit via a coaxial cable, and the data acquisition unit performs 16-bit analog-to-digital conversion on the voiceprint signal.

[0029] High-frequency current sensors are installed at the inlet and outlet of the three-phase power cables on the high-voltage side of the transformer. The high-frequency current sensors adopt the Rogowski coil type, with a measurement range of 0~1000A, a frequency response range of 50Hz~1MHz, and an accuracy of ±1%. Each phase current sensor synchronously acquires the three-phase current timing signal at a sampling frequency of 10kHz. , and The current data is transmitted to the data acquisition unit via optical fiber to isolate electromagnetic interference from the high-voltage side.

[0030] Gas sensors were installed at the highest point inside the enclosure (where gas tends to accumulate, i.e., the central area at the top of the enclosure), as well as in the cable compartment and transformer compartment. A total of four gas sensors were installed, each containing a CO sensor (measurement range 0~1000ppm). Sensor (measurement range 0~100ppm) Sensor (measurement range 0~200ppm) and Sensors (measurement range 0~5000ppm); each gas sensor acquires time-series signals of characteristic gas concentrations at a sampling frequency of 0.1Hz. ,in =1,2,3,4 correspond to CO, respectively , , The concentration data of the four gases are transmitted to the data acquisition unit via wireless communication.

[0031] The data acquisition unit performs filtering and noise reduction on the time-series signals of each physical field: the temperature signal is filtered using a moving average (window length of 5 sampling points); the acoustic signature signal is denoised using wavelet thresholding (selecting the db4 wavelet basis, decomposing into 5 layers, and using a soft thresholding method); the current signal is filtered using median (window length of 3 sampling points); and the gas concentration signal is filtered using an exponentially weighted moving average (smoothing coefficient of 0.3).

[0032] The filtered time series signals of each physical field are timestamped using GPS clock as the synchronization reference to ensure that the data of each physical field has a unified time coordinate. The aligned data are arranged in chronological order to construct a multi-physical field spatiotemporal data matrix. Each row records the data of all monitoring points of each physical field at the same time, and each column records the change sequence of a single monitoring point or a single physical quantity over time.

[0033] Understandably, this step achieves comprehensive synchronous acquisition of multi-physical field signals (thermal, electrical, acoustic, and gas) within the prefabricated substation by deploying temperature sensors in a three-dimensional spatial grid, deploying acoustic sensors at multiple locations, installing current sensors by phase, and installing gas sensors by region. Through filtering and denoising, and timestamp alignment, environmental noise and sensor clock deviations are eliminated, providing high-quality synchronous time-series data for subsequent causal correlation analysis and effectively avoiding false causal correlations caused by poor signal quality or time asynchrony.

[0034] Step S20: Based on the transfer entropy, perform causal correlation analysis on each physical field signal, calculate the transfer entropy value between each pair of physical fields and the causal delay time that makes the transfer entropy reach its maximum value, and automatically generate a multi-physics causal directed graph with time delay annotation after the significance verification by permutation test.

[0035] In this embodiment, the calculation process of transfer entropy is illustrated using the causal correlation analysis between the temperature field T and the sound field P as an example. The time series sequences of the temperature field and the sound field are extracted from the multiphysics spatiotemporal data matrix, and the time delay is taken. From 0 to 100 seconds (step size 1 second), for each Values, construct triples ,in and They are time points and time The sound pressure level, For a moment Temperature values; in this embodiment, the data collection time was 1 hour, and a total of 3600 data points were obtained.

[0036] Based on the collected data points, the joint probability density function is estimated using the kernel density estimation method. Conditional probability density and The kernel function uses a Gaussian kernel, and the bandwidth is automatically selected using the Silverman rule. Substitute these values ​​into the following formula for calculating the transfer entropy:

[0037] in, The transfer entropy value from the temperature field to the sound field is calculated. In each The numerical value under the given value.

[0038] Similarly, calculate the time delays for each pair of time-series signals between temperature field T and current field I, temperature field T and gas field G, sound field P and current field I, sound field P and gas field G, and current field I and gas field G. The propagation entropy value under ,in , For physical field index, and .

[0039] For the transfer entropy curve from temperature field T to sound field P The delay time that maximizes the propagation entropy is taken as the causal delay time: ,in This represents the causal delay time from the temperature field to the sound field. Indicates taking The independent variable that achieves its maximum value The value of ; in this embodiment, exist The maximum value is reached at 35 seconds, therefore =35 seconds.

[0040] The significance of the transfer entropy value was tested using a permutation test: the temperature time series was randomly shuffled 1000 times, and the value was recalculated after each shuffle. We obtained 1000 permutation transitivity entropy values; we sorted these 1000 values ​​in ascending order and took the 950th value (i.e., the 95th quantile) as the significance threshold. If the value is higher than the threshold, the causal relationship between the temperature field and the sound field is determined to be significant.

[0041] Similarly, for each pair of physical fields ( Causal delay time estimation and permutation tests are performed. If the transfer entropy value between a pair of physical fields fails the significance test, the causal association is removed. In this embodiment, significant causal associations include: T→P ( =35s), T→G ( =55s), P→G ( =20s), P→I ( =30s), G→I ( =15s), where → represents the direction of causal propagation; while the direct causal relationship T→I failed the significance test, indicating that the causal relationship from temperature to current is an indirect causal relationship transmitted through the medium of sound field or gas field.

[0042] Based on the significant causal relationships retained after significance testing, with temperature field anomaly event T, sound field anomaly event P, current field anomaly event I, and gas field anomaly event G as nodes, significant causal relationships as directed edges, and causal delay time as the causal delay time. A multiphysics causal directed graph with time delay annotations is automatically generated for the edge weights. The causal directed graph generated in this embodiment is as follows:

[0043] in, The graph represents the edges between nodes; it reveals the causal chain of the fault within the prefabricated substation: temperature anomalies (such as overload heating) occur first, followed by acoustic anomalies (such as acoustic emission signals generated by partial discharge) approximately 35 seconds later, and then characteristic gases (such as CO produced by the thermal decomposition of insulating materials) escape approximately 20 seconds later. (And other gases), and finally, about 15 seconds later, current harmonic distortion appears. This causal chain is consistent with the actual physical propagation law of faults inside the prefabricated substation: heat accumulation leads to thermal expansion of materials and insulation deterioration, which generates acoustic emission; insulation deterioration further produces decomposition gases; and when the gases accumulate to a certain extent, they affect the electric field distribution, leading to current waveform distortion.

[0044] Understandably, this step, through transfer entropy calculation and permutation testing, eliminates the need for manual experience to pre-determine the fault propagation sequence, objectively and automatically discovering the causal transmission direction and delay time between multiple physical fields (thermal, electrical, acoustic, and gaseous), upgrading from traditional time-series correlation analysis to causal-driven analysis. This embodiment distinguishes between direct causal relationships (e.g., T→P) and indirect causal relationships (e.g., T→I transmitted through intermediaries), overcoming the limitation of traditional correlation analysis methods in distinguishing between causal and spurious correlations. The time delay annotations in the causal directed graph provide crucial temporal prior information for subsequent fault evolution prediction.

[0045] Step S30: Construct a physical information graph neural network, discretize the internal space of the box-type substation into monitoring nodes and construct a spatial graph structure, and embed the heat conduction equation, sound wave equation and current continuity equation into the loss function as physical constraints.

[0046] In this embodiment, the internal space of the prefabricated substation is discretized into... =48 monitoring nodes; of which 24 nodes correspond to 24 temperature measurement points (including the actual location of temperature sensors and interpolation points between adjacent sensors), 8 nodes correspond to 8 acoustic sensor locations, 3 nodes correspond to 3 current monitoring points (cable access points of phases A, B, and C), and 13 nodes are densified interpolation points (arranged in key locations in the transformer oil area and high-voltage switch area) to ensure full coverage of fault-prone areas.

[0047] Constructing a spatial graph structure ,in It is a set of 48 nodes. This is the set of edges connecting nodes; the edges between nodes are determined according to the following rule: the physical distance between any two monitoring nodes is less than a threshold. When the distance is 0.8 meters, an undirected edge is established. In this embodiment, a total of 156 edges are established.

[0048] Node set Each node The node features are represented as ,in , , and Representing nodes respectively The temperature, sound pressure, current, and characteristic gas concentration values ​​at the location are assigned as follows: Assignment: For nodes at the location of temperature sensors, the actual temperature value measured by the sensor is used directly; for nodes at locations other than temperature sensors, the inverse distance weighted interpolation method is used, which is a weighted average of the actual values ​​measured by the three nearest temperature measurement points. The weight is inversely proportional to the square of the spatial distance from the measurement point to the node.

[0049] Assignment: For nodes at the location of the acoustic sensor, the measured sound pressure value of the sensor is used directly; for nodes at locations other than the location of the acoustic sensor, the Kriging interpolation method is used, and the propagation attenuation model of sound waves in the medium inside the box-type substation is used for interpolation estimation.

[0050] Assignment: For nodes belonging to phase A electrical circuits, assign the value as follows: Nodes belonging to phase B are assigned the value Nodes belonging to phase C are assigned the value .

[0051] Value assignment: Nodes located in the cable compartment area are assigned the concentration value measured by the gas sensor in the cable compartment; nodes located in the transformer compartment area are assigned the concentration value measured by the gas sensor in the transformer compartment; nodes located in the low-pressure compartment area are assigned the concentration value measured by the gas sensor at the center of the top of the enclosure.

[0052] Edge features include the Euclidean distance between nodes (unit: meters) and dielectric properties (0 for air regions, 1 for transformer oil regions, and 2 for solid insulating dielectric regions).

[0053] A physical constraint graph neural network is constructed, using spatial graph structure and multiphysics causal directed graphs as input features, and embedding a physical constraint loss term into the loss function. , represented as constraint terms in the heat conduction equation Constraint terms of the acoustic wave equation and current continuity equation constraint terms The weighted sum. Among them, the constraint terms of the heat conduction equation... Expressed as a formula:

[0054] in, It represents the temperature field and is an abbreviation for the temperature field time series signal; Let be the Laplace operator, representing the sum of the second-order partial derivatives of temperature in the three spatial directions; The thermal diffusivity is an inherent physical property of the medium. In this embodiment, it is taken as 0.086 mm² / s in the transformer oil region and 19.1 mm² / s in the air region. and This represents the first derivative of temperature with respect to time. The Frobenius norm is used to calculate the square root of the sum of squares of the residuals of the heat conduction equation at all monitoring nodes. This constraint requires that the spatiotemporal evolution of the temperature field output by the graph neural network satisfies the heat conduction equation, that is, the spatial second-order partial derivative of temperature should be equal to the rate of change of temperature over time divided by the thermal diffusivity, thereby ensuring that the temperature prediction conforms to the physical laws of heat diffusion.

[0055] Constraints of the acoustic wave equation Expressed as a formula:

[0056] in, It represents the sound field and is an abbreviation for the sound field timing signal; The speed of sound in a medium is denoted by m / s, and the speed of sound in transformer oil is 1420 m / s (at 25°C). and represents the second-order partial derivative of sound pressure with respect to time, and represents the acceleration of sound pressure. This constraint requires that the spatiotemporal evolution of the sound pressure field output by the graph neural network satisfies the sound wave equation, that is, the spatial second-order partial derivative of sound pressure should be equal to the temporal second-order partial derivative of sound pressure divided by the square of the speed of sound, thereby ensuring that the propagation of sound pressure conforms to the physical laws of mechanical waves.

[0057] Current continuity equation constraint terms Expressed as a formula:

[0058] in, It represents the current density vector, which represents the current flowing per unit area, and its direction is the direction of positive charge flow; It is a divergence operator used to quantify the net outflow of current at a point in space; Charge density represents the amount of charge per unit volume; For nodes The divergence discretization approximation of the current density vector field at a given point is calculated using the current difference between the node and its adjacent nodes. For nodes The first-order partial derivative of the charge density with respect to time is discretized and approximated using the central difference method. This constraint requires that the current field output by the graph neural network satisfies the law of charge conservation, that is, the sum of the divergence of the current density and the rate of change of the charge density is zero, thereby ensuring that the electric field distribution conforms to the basic physical laws of electromagnetism.

[0059] The total loss function for constructing a physical information graph neural network This can be expressed as a formula:

[0060] in, The task loss for fault classification and propagation prediction is used to measure the error between the predicted value and the true label. This represents the total weight coefficient for physical constraints, used to adjust the proportion of physical constraints in the total loss. In this embodiment, the physical information graph neural network adopts a graph attention network (GAT) architecture, containing three graph attention layers, each with a hidden dimension of 64 and four attention heads. Task Loss The cross-entropy loss function is used to measure the accuracy of the network in predicting fault type classification and propagation path, with the total weight of physical constraints as the benchmark. =0.1, used to adjust the proportion of physical constraints in the total loss. The network training uses the Adam optimizer with an initial learning rate of 0.001, 500 training epochs, and a batch size of 32. The training data comes from the historical operating data of this prefabricated substation over the past two years, including 15,000 normal operating condition samples and 5,000 samples of various fault conditions (including overload heating, poor contact, partial discharge, insulation degradation, and short circuit arc).

[0061] Understandably, this step discretizes the internal space of the prefabricated substation into a spatial graph structure, organizing the multi-physics data (thermal, electrical, acoustic, and gaseous) into a graph structure with spatial relationships. Fault features can be transmitted and aggregated between spatially adjacent monitoring nodes, providing a data foundation consistent with the physical spatial distribution for subsequent graph neural network inference. Embedding three types of physical constraints—heat conduction equations, acoustic wave equations, and current continuity equations—into the loss function ensures that the neural network's learning process is constrained by fundamental physical laws. Even in regions with sparse training data, the model can make reasonable inferences based on physical equations. The introduction of physical constraint terms guarantees the physical consistency of the prediction results, avoiding diagnostic conclusions that violate physical laws.

[0062] Step S40: Based on the multiphysics causal directed graph guidance graph attention mechanism, the attention coefficients are positively biased along the causal propagation direction; based on the current abnormal event, the propagation probability and expected arrival time of the fault between each monitoring node are calculated, and a probability-weighted set of fault evolution paths is output.

[0063] In this embodiment, the multiphysics causal directed graph generated in step S20 is used as prior knowledge, and a graph attention mechanism is injected into the computation nodes. For nodes Attention coefficient First, the original attention score is calculated based on the node features. :

[0064] in, and They are nodes and nodes Node characteristics; A learnable linear transformation weight matrix used to map four-dimensional input features to... The latent space of a dimension makes physical quantities comparable within the same characteristic space. This represents the hidden dimension of the graph attention layer; in this embodiment, it is set to 64. The transpose of the learnable attention weight vector, with a dimension of 2. ; This represents a vector concatenation operation. This is a modified linear activation function with leakage.

[0065] If a multiphysics causal directed graph exists For directed edges, a positive bias is applied to the attention coefficient, expressed by the formula:

[0066] in, The causal bias weight is set to 2.0 in this embodiment; This is an indicator function; it takes the value 1 if the condition inside the parentheses is true, and 0 otherwise. Represents a node To the node The directed edges exist in the multiphysics causal directed graph. This is a normalized exponential function. After softmax normalization, node pairs in the causal direction receive higher attention weights. For example, if there is a directed edge from the temperature node to the voiceprint node in the causal directed graph, a positive bias is applied when calculating the attention coefficient, causing the network to focus more on the information transmission from the temperature node to the voiceprint node when aggregating node features, rather than treating all adjacent nodes unequally.

[0067] The fault originated from the internal spatial nodes of the prefabricated substation. To the node The probability of propagation is calculated using the following formula:

[0068] in, Indicates at time Fault from node propagation to nodes The conditional probability, To predict the time step, For nodes For nodes The normalized attention coefficient, and They are nodes and nodes exist The hidden state representation at time step is obtained by forward propagation calculation of the physical information graph neural network; This represents the causal delay time estimated by the propagation entropy. For activation functions; In this embodiment, the hidden state representation of each node obtained in step S30 is used as a basis. and And the causal delay time estimated by the transfer entropy in step S20. (like , For temperature-soundprint node pairs, =35 seconds; if it is a voiceprint-gas node pair =20 seconds; if it is a gas-current node pair =15 seconds), after splicing, it is mapped to the probability interval [0,1] by the activation function σ.

[0069] for The three values ​​of 5 minutes, 15 minutes, and 60 minutes are used to calculate the fault propagation probability. Similarly, the propagation probability of the fault along each path (T→P→G→I) in the causal graph is calculated among all nodes, forming a complete probability distribution. For each propagation path, the expected arrival time is the sum of the causal delay times of each segment. The physical information graph neural network finally outputs the fault propagation probability and expected arrival time of each propagation path within the future time window, forming a probability-weighted set of fault evolution paths. In this embodiment, within 60 seconds after detecting a temperature anomaly, the model outputs the following probability-weighted path set: Path 1 (probability 0.82): T→P (35s)→G (55s)→I (70s); Path 2 (probability 0.13): T→P (35s)→I_direct (65s, causal edge from P directly to I); Path 3 (probability 0.05): T→G_direct (50s, causal edge from T directly to G)→I (65s).

[0070] Understandably, this step, based on a causal directed graph-guided attention mechanism, prioritizes attention allocation along the causal propagation direction, rather than calculating attention equally across all nodes. This reduces unnecessary computation and enhances the physical interpretability of the deduction process. Guided by the causal directed graph bias, fault propagation prediction automatically prioritizes the causal chain T→P→G→I, perfectly aligning with the actual physical propagation patterns of overload and overheating faults in prefabricated substations. The probability-weighted fault evolution path set provides multiple possibilities for fault development and their respective probabilities, offering maintenance personnel a quantitative basis for risk assessment.

[0071] Step S50: Based on the multiphysics causal directed graph and the fault propagation probability, output the diagnostic results including the fault causal chain, fault origin location, predicted evolution path and confidence assessment.

[0072] Based on the multiphysics causal directed graph generated in step S20, the causal chain of faults is generated by labeling the chronological propagation order of anomalous events in each physics field according to the causal delay time: In this embodiment, after detecting an abnormal temperature in the transformer room, the system automatically retrieves the causal directed graph and identifies the following causal chain: T→(35s)→P→(20s)→G→(15s)→I; the system outputs the fault causal chain as: "A hot spot appears in the transformer room → 35 seconds later, abnormal acoustic signature (a sudden increase in energy in the 20kHz frequency band was detected, indicating partial discharge characteristics) → 55 seconds later..." Gas concentration increases (from 5 ppm to 32 ppm, exceeding the warning threshold) → After 70 seconds, the harmonic distortion rate of phase B current increases to 4.2% (exceeding the normal operating limit of 2.5%).

[0073] The nodes with an in-degree of zero are traced in the causal directed graph. In this embodiment, in the causal directed graph T→P→G→I, the temperature field T has an in-degree of zero (no other physical field points to the temperature field), so the fault is determined to originate from the spatial location corresponding to the temperature field. The spatial coordinates of the temperature field are further mapped to the physical space inside the box-type substation, and the fault origin location is output as: "the upper part of the middle of the transformer room (coordinates x=1.2m, y=0.8m, z=1.5m), corresponding to the vicinity of the top of the B phase winding of the transformer".

[0074] Based on the probability-weighted fault evolution path set obtained in step S40, the system outputs the K fault propagation paths with the highest probability within each time window. In this embodiment, K=3, and the system outputs: Within a 5-minute window (the 3 paths with the highest probability): Path T→P: 72% probability, expected arrival time 35 seconds; Path T→P→G: 58% probability, expected arrival time 55 seconds; Path T→G: 31% probability, expected arrival time 50 seconds. Within a 15-minute window (the 3 paths with the highest probability): Path T→P→G→I: 89% probability, expected arrival time 70 seconds; Path T→P→I: 67% probability, expected arrival time 65 seconds; Path T→P→G: 42% probability, expected arrival time 55 seconds. Within a 60-minute window (the 3 paths with the highest probability): Path T→P→G→I: 96% probability, expected arrival time 70 seconds; Path T→P→G→I→box wall: 61% probability, expected arrival time 130 seconds; Path T→P→I→upstream switch: 44% probability, expected arrival time 105 seconds.

[0075] Finally, the confidence score is calculated based on the inverse normalized value of the prediction variance of the physical information graph neural network. In this embodiment, the prediction variance of each path is 0.015, and the confidence score is calculated to be 0.87 (within the interval [0,1]).

[0076] Understandably, this step integrates the causal directed graph from step S20 with the fault propagation probability prediction from step S40 to output a complete diagnostic report containing the fault causal chain, origin location, evolution path, and confidence assessment. The fault causal chain is presented in an intuitive "event → time interval → event" format, enabling operations and maintenance personnel to clearly understand the timeline of fault development; fault origin location traces the source node in the causal directed graph, achieving an improvement from fault type judgment to fault source location; the predicted evolution path provides fault situation prediction at multiple time scales, providing a forward-looking basis for operations and maintenance decisions; and the confidence assessment quantifies the credibility of the diagnostic conclusions, assisting operations and maintenance personnel in risk classification and handling.

[0077] Example 2: The intelligent monitoring system for internal faults in a prefabricated substation provided in this embodiment of the invention can execute the intelligent monitoring method for internal faults in a prefabricated substation provided in any embodiment of the invention, and has the corresponding functional modules and beneficial effects of the method, such as... Figure 2 As shown, it includes: Data acquisition and preprocessing module: used to construct a physical field sensor network inside the box-type substation, and synchronously acquire multi-physical field time-series signals, including temperature field time-series signals, sound field time-series signals, three-phase current time-series signals and characteristic gas concentration time-series signals; Causal correlation analysis module: It is used to perform causal correlation analysis on signals of various physical fields based on the transfer entropy, calculate the transfer entropy values ​​between each pair of physical fields and the causal delay time that makes the transfer entropy reach its maximum value, and automatically generate a multi-physics causal directed graph with time delay annotation after the significance verification by permutation test. Physical Information Graph Neural Network Construction Module: Used to construct a physical information graph neural network, discretize the internal space of the box-type substation into monitoring nodes and construct a spatial graph structure, and embed heat conduction equation, sound wave equation and current continuity equation as physical constraints in the loss function; Causal-guided inference module: used to guide the graph attention mechanism based on multi-physics causal directed graph, so that the attention coefficients are positively biased along the causal propagation direction; calculates the propagation probability and expected arrival time of the fault between each monitoring node based on the current abnormal event, and outputs a probability-weighted set of fault evolution paths; Diagnostic Result Output Module: Based on multiphysics causal directed graphs and fault propagation probabilities, this module outputs diagnostic results including fault causal chains, fault origin location, predicted evolution paths, and confidence assessments.

[0078] Although this application makes various references to certain modules in the system according to the embodiments of this application, any number of different modules can be used and run on user terminals and / or servers. The various units and modules included are only divided according to functional logic, but are not limited to the above division, as long as the corresponding functions can be achieved; in addition, the specific names of each functional unit are only for easy distinction between each other and are not used to limit the scope of protection of this invention.

[0079] Example 3: This application provides an intelligent monitoring device for internal faults in a prefabricated substation. The intelligent monitoring device for internal faults in a prefabricated substation includes: at least one processor; and a memory communicatively connected to the at least one processor; wherein the memory stores instructions executable by the at least one processor, and the instructions are executed by the at least one processor to enable the at least one processor to execute the intelligent monitoring method for internal faults in a prefabricated substation as described in Example 1 above.

[0080] Example 4: This application provides a computer program product including a computer program carried on a computer-readable medium, the computer program containing program code for performing the methods shown in the flowchart. In such an embodiment, the computer program can be downloaded and installed from a network via a communication system, or installed from a storage system. When the computer program is executed by a processing system, it performs the functions defined in the method of Example 1 of this application.

[0081] The specific embodiments described above do not constitute a limitation on the scope of protection of this application. Those skilled in the art should understand that various modifications, combinations, and substitutions can be made according to design requirements and other factors. Any modifications, equivalent substitutions, and improvements made within the spirit and principles of this application should be included within the scope of protection of this application. In some cases, the actions or steps described in this application can be performed in a different order than that shown in the embodiments and still achieve the desired results. Furthermore, the processes depicted in the accompanying drawings do not necessarily require a specific or sequential order to achieve the desired results. In some embodiments, multitasking and parallel processing are also possible or may be advantageous.

Claims

1. An intelligent monitoring method for internal faults in a prefabricated substation, characterized in that, The method includes: Step S10: Construct a physical field sensor network for the internal space of the box-type substation, and synchronously collect multi-physical field time-series signals, including temperature field time-series signals, sound field time-series signals, three-phase current time-series signals, and characteristic gas concentration time-series signals. Step S20: Based on the transfer entropy, perform causal correlation analysis on each physical field signal, calculate the transfer entropy value between each pair of physical fields and the causal delay time that makes the transfer entropy reach its maximum value, and automatically generate a multi-physics causal directed graph with time delay label after the significance verification by permutation test. Step S30: Construct a physical information graph neural network, discretize the internal space of the box-type substation into monitoring nodes and construct a spatial graph structure, and embed the heat conduction equation, sound wave equation and current continuity equation into the loss function as physical constraints. Step S40: Based on the multiphysics causal directed graph guidance graph attention mechanism, the attention coefficients are positively biased along the causal propagation direction; based on the current abnormal event, the propagation probability and expected arrival time of the fault between each monitoring node are calculated, and a probability-weighted set of fault evolution paths is output. Step S50: Based on the multiphysics causal directed graph and the fault propagation probability, output the diagnostic results including the fault causal chain, fault origin location, predicted evolution path and confidence assessment.

2. The intelligent monitoring method for internal faults in a prefabricated substation as described in claim 1, characterized in that, In step S10, the temperature field timing signal The components are arranged in a spatial grid inside the prefabricated substation. Temperature data is collected by one temperature sensor, among which Indicates the first Spatial coordinates of each temperature monitoring point For time indexing; The sound field timing signal The components installed on the internal walls of the prefabricated substation enclosure and the surface of the transformer casing are... Data collected by a voiceprint sensor, among which Indicates the first Spatial coordinates of each voiceprint monitoring point; The three-phase current timing signal , and Data is collected by high-frequency current sensors installed at the entry and exit points of each phase power cable in the box-type substation. The characteristic gas concentration time-series signal Gas data is collected by gas sensors installed at high points inside the prefabricated substation enclosure, as well as in the cable compartment and transformer compartment. The gas types are numbered; the characteristic gases include , , and At least one of them; The temperature field time series signal, sound field time series signal, three-phase current time series signal and characteristic gas concentration time series signal are filtered, denoised and timestamped to construct a multiphysics spatiotemporal data matrix.

3. The intelligent monitoring method for internal faults in a prefabricated substation as described in claim 1, characterized in that, Step S20 includes the following detailed steps: Step S201, for any two physical field time series and Calculate the physical field according to the following formula To the physical field Transmission entropy Used to characterize physical fields Past states of the physical field The degree to which the ability to predict future states contributes: in, and Representing physical fields respectively and physical field At any moment The value of , For physical fields At any moment The value of , For time delay; For joint probability density, In order to be in and under conditions The conditional probability density, In order to be in under conditions The conditional probability density; calculate the time delays for each pairwise time-series signal between the temperature field, sound field, current field, and gas field. The propagation entropy value under ,in , They represent any two different physical fields; Step S202, for each pair of physical fields The delay time that maximizes the propagation entropy is taken as the causal delay time, which is expressed by the formula: ,in Represents physical field To the physical field causal delay time, Indicates taking The independent variable that achieves its maximum value The value of ; Step S203: Use the permutation test to test the significance of the transfer entropy value of each pair of physical fields: randomly shuffle the time series of each physical field multiple times, recalculate the transfer entropy distribution, and if the original transfer entropy value is higher than the 95th percentile, the causal relationship is determined to be significant and false causal associations are eliminated. Step S204: Based on the significant causal relationships retained after significance testing, using temperature field anomaly event T, sound field anomaly event P, current field anomaly event I, and gas field anomaly event G as nodes, significant causal relationships as directed edges, and causal delay time... Using edge weights, automatically generate a multiphysics causal directed graph with time delay annotations.

4. The intelligent monitoring method for internal faults in a prefabricated substation as described in claim 1, characterized in that, Step S30 includes the following detailed steps: Step S301: Discretize the internal space of the prefabricated substation into Q monitoring nodes and construct a spatial graph structure. This is used to organize the physical field data with spatial location attributes inside a prefabricated substation into graph-structured data, enabling fault characteristics to be transmitted and aggregated between spatially adjacent monitoring nodes. Representing a spatial diagram, and Represent the set of nodes and the set of edges, respectively; Step S302, set the nodes Each node The node features are represented as ,in , , and Representing nodes respectively Temperature, sound pressure, current, and characteristic gas concentration values ​​at the location. and It is obtained by spatial interpolation or nearest neighbor matching of the temperature field and sound field time-series signals according to the spatial coordinates of the nodes. The node is assigned a value based on the corresponding phase of the three-phase current timing signal. The value is assigned based on the concentration time-series signal collected by the gas sensor in the corresponding region where the node is located; Step S303: The edge features between nodes in the node set V are represented as physical distance and medium properties between nodes, which are used to characterize the spatial correlation strength between adjacent monitoring nodes and the physical conditions for fault propagation. Step S304: Embed a physical constraint loss term into the loss function of the physical information graph neural network. The total loss function of the physical information graph neural network is constructed by weighting three constraints: the heat conduction equation, the sound wave equation, and the current continuity equation. This can be expressed as a formula: in, The task loss for fault classification and propagation prediction is used to measure the error between the predicted value and the true label. This is the total weight coefficient for physical constraints, used to adjust the proportion of physical constraints in the total loss.

5. The intelligent monitoring method for internal faults in a prefabricated substation as described in claim 1, characterized in that, In step S40, the process of the multiphysics-based causal directed graph guided graph attention mechanism includes: Graph attention mechanism in physical information graph neural networks at computation nodes For nodes When the attention coefficient exists in the multiphysics causal directed graph, For directed edges, a positive bias is applied to the attention coefficient, expressed by the formula: in, For nodes For nodes The normalized attention coefficient, For nodes For nodes The original attention score, For causal bias weights, This is an indicator function; it takes the value 1 if the condition inside the parentheses is true, and 0 otherwise. Represents a node To the node The directed edges exist in the multiphysics causal directed graph. It is a normalized exponential function.

6. The intelligent monitoring method for internal faults in a prefabricated substation as described in claim 1, characterized in that, In step S40, the calculation process for the fault propagation probability includes: The fault originated from the internal spatial nodes of the prefabricated substation. To the node The probability of propagation is calculated using the following formula: in, Indicates at time Fault from node propagation to nodes The conditional probability, To predict the time step, For nodes For nodes The normalized attention coefficient, and They are nodes and nodes exist The hidden state representation at time step is obtained by forward propagation calculation of the physical information graph neural network; This represents the causal delay time estimated by the propagation entropy. This represents a vector concatenation operation. is the activation function of the physical information graph neural network, used to map the output value to the probability interval [0,1]. The physical information graph neural network finally outputs the failure propagation probability and expected arrival time of each propagation path in the future time window, forming a probability-weighted set of failure evolution paths.

7. The intelligent monitoring method for internal faults in a prefabricated substation as described in claim 1, characterized in that, In step S50, the diagnostic results specifically include: Fault Causal Chain: Based on the following format, the chronological order of propagation of each physical field anomalous event is marked according to the causal delay time: Physical field event A → Second → Physical field event B → Second → Physical field event C, where → indicates the direction of causal propagation. Let be the causal delay time from an anomalous event in physical field A to an anomalous event in physical field B. The causal delay time from an abnormal event in physical field B to an abnormal event in physical field C; Fault origin location: trace the physical field and its spatial coordinates corresponding to the node with zero in-degree in the multiphysics causal directed graph; Predicted evolution paths: the K most probable fault propagation paths and their corresponding probability values ​​within the next three time windows of 5 minutes, 15 minutes and 60 minutes, where K is greater than or equal to 3; Confidence assessment: Based on the inverse normalized value of the variance predicted by the physical information graph neural network, a confidence score in the range of 0 to 1 is given.

8. An intelligent monitoring system for internal faults in a prefabricated substation, characterized in that, The system is used to implement the intelligent monitoring method for internal faults in a prefabricated substation as described in any one of claims 1-7, and the system comprises: Data acquisition and preprocessing module: used to construct a physical field sensor network inside the box-type substation, and synchronously acquire multi-physical field time-series signals, including temperature field time-series signals, sound field time-series signals, three-phase current time-series signals and characteristic gas concentration time-series signals; Causal correlation analysis module: It is used to perform causal correlation analysis on signals of various physical fields based on the transfer entropy, calculate the transfer entropy values ​​between each pair of physical fields and the causal delay time that makes the transfer entropy reach its maximum value, and automatically generate a multi-physics causal directed graph with time delay annotation after the significance verification by permutation test. Physical Information Graph Neural Network Construction Module: Used to construct a physical information graph neural network, discretize the internal space of the box-type substation into monitoring nodes and construct a spatial graph structure, and embed heat conduction equation, sound wave equation and current continuity equation as physical constraints in the loss function; Causal-guided inference module: used to guide the graph attention mechanism based on multi-physics causal directed graph, so that the attention coefficients are positively biased along the causal propagation direction; calculates the propagation probability and expected arrival time of the fault between each monitoring node based on the current abnormal event, and outputs a probability-weighted set of fault evolution paths; Diagnostic Result Output Module: Based on multiphysics causal directed graphs and fault propagation probabilities, this module outputs diagnostic results including fault causal chains, fault origin location, predicted evolution paths, and confidence assessments.

9. An intelligent monitoring device for internal faults in a prefabricated substation, characterized in that, The intelligent monitoring device for internal faults in the prefabricated substation includes: a memory, a processor, and an intelligent monitoring program for internal faults in the prefabricated substation stored in the memory and executable on the processor. When the intelligent monitoring program for internal faults in the prefabricated substation is executed by the processor, it implements an intelligent monitoring method for internal faults in a prefabricated substation as described in any one of claims 1 to 7.

10. A computer program product, characterized in that, The computer program product includes an intelligent monitoring program for internal faults in a prefabricated substation. When the intelligent monitoring program for internal faults in a prefabricated substation is executed by a processor, it implements an intelligent monitoring method for internal faults in a prefabricated substation as described in any one of claims 1 to 7.

Images