Intelligent traction substation secondary circuit modeling and fault diagnosis method

Abstract

The invention provides an intelligent traction substation secondary circuit modeling and fault diagnosis method, and relates to the field of automation system / software design, and the method comprises the following steps: monitoring secondary circuit parameters in a secondary circuit in real time through monitoring platform software, and recognizing abnormal data points through a sliding window statistical method; on the basis of the abnormal point position set, performing abnormal point matching by utilizing expert rules stored in monitoring platform software; acquiring a real-time state and historical alarm information of the secondary circuit equipment, and extracting a communication link state based on a secondary circuit equipment connection relationship described by the secondary circuit dynamic model; and constructing a secondary circuit fault diagnosis GAT model, inputting the fault graph data into the secondary circuit fault diagnosis GAT model, and obtaining a diagnosis result through reasoning calculation. The fault source and the fault link can be quickly positioned, and the reliability and the operation and maintenance efficiency of the power system are improved.

Description

Technical Field

[0001] This invention relates to the field of automation system / software design, and more specifically, to a method for modeling and diagnosing secondary circuits in intelligent traction substations. Background Technology

[0002] As a core facility of the high-speed rail power supply system, the intelligent traction substation has achieved intelligent primary equipment and networked secondary equipment during its development. Data in the intelligent traction substation is transmitted via network; conventional analog signals and cable connections are replaced by digital signals and fiber optic connections. Information exchange eliminates hard-wired methods, replacing them with switches and network cables. Secondary circuits between secondary equipment are also transformed into virtual circuits connected by different virtual terminals, making the entire virtual terminal secondary circuit connection relationship completely invisible. This transformation greatly improves the transparency and efficiency of information transmission, but it also brings new challenges. In the event of a fault, it is difficult to quickly locate the fault source, affecting operation and maintenance efficiency.

[0003] Traditional secondary circuit fault diagnosis methods mainly rely on expert experience, threshold judgment, or shallow machine learning models, which have the following shortcomings:

[0004] Multi-source data fusion is difficult: the input data comes from multi-modal data such as analog and digital input values, GOOSE / SV communication messages, network device status, and IED device self-test information, which are difficult to effectively fuse using existing methods;

[0005] Complex fault chain reaction: A single device failure (such as switch congestion) can trigger multiple problems such as protection malfunction and data synchronization failure. Traditional methods are difficult to achieve cross-device collaborative diagnosis.

[0006] For example, Chinese patent application CN201910709596.5 discloses a method for diagnosing faults in secondary circuits of relay protection in intelligent substations. It obtains physical circuit information and virtual circuit information and their mapping relationship by parsing SCD files. Based on this, it establishes a circuit and virtual circuit disconnection alarm information table. When a fault occurs, it uses the cross- and overlapping characteristics of each virtual circuit and the circuit alarm and virtual circuit disconnection alarm information table to exclude normal components for each suspected faulty secondary circuit, thereby obtaining a set of suspected components for that secondary circuit.

[0007] However, the aforementioned diagnostic method establishes a loop alarm and virtual loop disconnection alarm information table to obtain a set of suspected components in the secondary loop when a fault occurs. It relies solely on virtual loop fault alarm information to infer faulty components, resulting in a single data source that cannot handle complex scenarios such as multi-source faults, communication delays, and network device anomalies.

[0008] For example, Chinese patent application number CN201910011248.0 discloses a method for diagnosing secondary circuit faults in intelligent substations. It establishes a mapping relationship between logical links and physical (fiber optic) links, and uses link breakage alarm information, protection device and switch port status information, and report control block status sent by network analyzer to achieve intelligent diagnosis of secondary circuit faults. The beneficial effect of this invention is that most current secondary circuit diagnosis methods rely on station control layer link alarm information to perform single diagnosis of logical link faults. Once a link breakage occurs, it is impossible to quickly analyze and locate the fault point from numerous alarm information.

[0009] However, the aforementioned diagnostic methods collect link alarm information, port optical intensity information, and report control block status, and then effectively integrate and process this data to diagnose and locate faults. Their fault diagnosis strategies rely on fixed rules and can only identify known faults within those rules; they cannot provide effective diagnostic results for other unknown faults or faults outside the rules.

[0010] No effective solutions have yet been proposed to address the problems in the relevant technologies. Summary of the Invention

[0011] In view of this, the present invention provides a method for modeling and diagnosing the secondary circuit of an intelligent traction substation to solve the aforementioned problems.

[0012] To solve the above problems, the specific technical solution adopted by the present invention is as follows:

[0013] A method for modeling and fault diagnosis of secondary circuits in intelligent traction substations includes the following steps:

[0014] S1. Use monitoring platform software to monitor the secondary circuit parameters in the secondary circuit in real time, and use the sliding window statistical method to identify abnormal data points and obtain a set of abnormal points.

[0015] S2. Based on the set of abnormal points, use the expert rules stored in the monitoring platform software to match abnormal points. If the match is successful, obtain the diagnosis result based on the pre-set knowledge base and mark the fault node. Otherwise, proceed to step S3.

[0016] S3. Obtain the real-time status and historical alarm information of the secondary circuit equipment, and extract the communication link status based on the connection relationship of the secondary circuit equipment described by the secondary circuit dynamic model, and establish a fault data graph.

[0017] S4. Construct a secondary circuit fault diagnosis GAT model, input the fault diagram data into the secondary circuit fault diagnosis GAT model, and obtain the diagnosis results through reasoning calculation.

[0018] Preferably, the step of using monitoring platform software to monitor the secondary circuit parameters in real time and identifying abnormal data points through a sliding window statistical method to obtain a set of abnormal points includes the following steps:

[0019] S11. The monitoring platform software monitors the electrical parameter measurement data, equipment operating parameters and communication parameters in the secondary circuit in real time.

[0020] S12. For electrical parameter measurement data, equipment operating parameters and communication parameters, use the sliding window statistical method and calculate the window mean and standard deviation of the current sampling point based on the preset window length and sliding step size.

[0021] S13. Calculate the range of outliers based on the window mean and standard deviation of the sampling points. If the current sampling point values ​​of electrical parameter measurement data, equipment operating parameters and communication parameters exceed the range of outliers, then the sampling point is determined to be an outlier data point; otherwise, the sampling point is determined to be a normal data point.

[0022] S14. Based on the abnormal data points, record the locations of all abnormal data points to obtain a set of abnormal data points.

[0023] Preferably, the abnormal point matching is performed based on the abnormal point set using expert rules stored in the monitoring platform software. If the matching is successful, a diagnostic result is obtained based on a pre-set knowledge base, and the faulty node is marked. Otherwise, step S3 includes the following steps:

[0024] S21. Read the predefined fault diagnosis formula configuration file in the knowledge base of the monitoring platform software to obtain the fault diagnosis formula;

[0025] S22. Determine whether the set of data points in the fault diagnosis formula is a subset of the set of abnormal points. If so, input the subset of the set of abnormal points into the fault diagnosis formula, calculate the result, combine it with the formula result description predefined in the knowledge base, obtain the diagnosis result, and mark the fault node. Otherwise, execute step S3.

[0026] S23. Visualize the obtained diagnostic results to obtain visualized diagnostic results.

[0027] Preferably, the fault diagnosis formula includes logical operation formulas and numerical operation formulas, wherein,

[0028] The logical operation formula is:

[0029] A = P1 & P2 & ... & P n ;

[0030] The numerical calculation formula is as follows:

[0031] S = w1·P1 + w2·P2 + ... + w n ·P n ;

[0032] In the formula, A represents the result of the logical operation, {P1,P2,…,P…} n} represents the set of input points for the formula, and S represents the result of the data calculation. n} represents the weight of the input points in the formula.

[0033] Preferably, the steps of acquiring the real-time status and historical alarm information of the secondary circuit equipment, and extracting the communication link status and establishing a fault data graph based on the connection relationship of the secondary circuit equipment described by the secondary circuit dynamic model include the following steps:

[0034] S31. Extract the real-time status and historical alarm information of the secondary circuit equipment based on the electrical parameter measurement data, equipment operating parameters and communication parameters;

[0035] S32. Based on graph theory, construct a dynamic model of the secondary loop and use the dynamic model of the secondary loop to describe the connection relationship between each secondary loop device.

[0036] S33. Based on the connection relationship between each secondary circuit device, extract the communication link status and establish a fault data diagram according to the communication link status.

[0037] Preferably, the step of constructing a dynamic model of the secondary loop based on graph theory and using the dynamic model to describe the connection relationships between the secondary loop devices includes the following steps:

[0038] S321. Treat the secondary circuit equipment as nodes, and configure attribute labels for each node based on its functional type;

[0039] S322. Treat the communication links between secondary circuit devices as edges, and configure weights for each edge based on the link characteristics of the communication links.

[0040] S323. Based on the nodes and edges, construct a dynamic model of a quadratic loop in the graph structure, and establish a relational database table to store the attributes and connection relationships of the nodes and edges.

[0041] Preferably, the steps of constructing a secondary circuit fault diagnosis GAT model, inputting fault diagram data into the secondary circuit fault diagnosis GAT model, and obtaining the diagnosis result through reasoning calculation include the following steps:

[0042] S41. Construct and train a secondary circuit fault diagnosis GAT model based on the secondary circuit dynamic model;

[0043] S42. Input the fault diagram data into the pre-trained secondary circuit fault diagnosis GAT model, and obtain the fault type code of each secondary circuit device and the fault association weight between each secondary circuit device through reasoning calculation.

[0044] S43. Based on the fault type codes of each secondary circuit device and the fault association weights between each secondary circuit device, interpret the fault types of the secondary circuit devices, filter the propagation links that meet the weight conditions, mark the fault nodes and fault links, and obtain the diagnostic results.

[0045] S44. Visualize the obtained diagnostic results to obtain visualized diagnostic results.

[0046] Preferably, the construction and training of the secondary circuit fault diagnosis GAT model based on the secondary circuit dynamic model includes the following steps:

[0047] S411. Based on the dynamic model of the secondary circuit, extract the real-time attributes used to characterize the faults of the secondary circuit equipment, obtain the fault characteristics of the secondary circuit equipment, and construct the secondary circuit fault diagram structure based on the fault characteristics of the secondary circuit equipment.

[0048] S412. Based on the attention mechanism, calculate the attention score between each secondary loop device node in the secondary loop fault diagram structure;

[0049] S413. Normalize the attention score using the softmax function to obtain the attention weight;

[0050] S414. Based on the attention weight, the neighbor features of each secondary loop device node are weighted and aggregated to update the representation vector of each secondary loop device node, thereby obtaining the context feature embedding representation of each secondary loop device.

[0051] S415. Based on the cross-entropy loss function, the contextual features of the secondary circuit equipment are embedded to represent the input fault classifier for training, thus obtaining the GAT model of the secondary circuit equipment fault.

[0052] Preferably, the formula for calculating the attention score between each secondary loop device node in the secondary loop fault diagram structure based on the attention mechanism is as follows:

[0053]

[0054] In the formula, e ijLet represent the attention score between secondary loop device node i and secondary loop device node j, LeakyReLU represent the activation function, W represent the learnable weight matrix of the secondary loop device node, a represent the learnable attention vector, l represent the current GAT layer number, T represent the transpose of the vector, X represent the input feature vector of a single node, || represent the concatenation operation, and R ij This represents the communication link coefficient between secondary circuit device nodes.

[0055] Preferably, the normalization process of the attention score using the softmax function yields the following formula for calculating the attention weight:

[0056]

[0057] In the formula, α ij Represents attention weight, e ij Let N(i) represent the attention score between secondary loop device node i and secondary loop device node j, and let N(i) represent the neighbor set of secondary loop device node i. ik This represents the attention score between secondary loop device node i and secondary loop device node k.

[0058] Preferably, the calculation formula for weighted aggregation of neighbor features of each secondary loop device node based on attention weights is as follows:

[0059]

[0060] In the formula, Let W represent the layer l state of secondary loop device node i, W represent the learnable weight matrix, and σ represent the activation function. This indicates the state of the next-level neighbor node of the secondary loop device node i.

[0061] Preferably, the expression for the cross-entropy loss function is:

[0062]

[0063] In the formula, L represents the cross-entropy loss function, N represents the number of samples, C represents the number of classes, and y i,c p represents the one-hot encoding of the true label of sample i. i,c This represents the probability that sample i belongs to category c, as predicted by the GAT model for secondary circuit fault diagnosis.

[0064] Preferably, the step of inputting fault diagram data into a pre-trained secondary loop fault diagnosis GAT model and obtaining the fault type code of each secondary loop device and the fault association weight between each secondary loop device through inference calculation includes the following steps:

[0065] S421. Based on the fault graph data, determine the characteristics of all fault nodes, the adjacency matrix of the fault nodes, and the edge type matrix;

[0066] S422. Input all the fault node features, the adjacency matrix and edge type matrix of the fault node into the pre-trained secondary loop fault diagnosis GAT model, and obtain the output of each fault node after information transmission and operation through multiple GAT layers.

[0067] S423. Use the output of each fault node as the input of the classification layer in the secondary loop fault diagnosis GAT model to obtain the fault classification result, and output the fault type encoding vector through the secondary loop fault diagnosis GAT model.

[0068] S424. Interpret the fault type encoding vector as fault type and fault path, and combine it with the secondary loop dynamic model to obtain the fault association weight between each secondary loop device.

[0069] The beneficial effects of this invention are as follows:

[0070] 1. This invention models secondary circuits from a graph perspective. Based on the IEC61850 standard IED model, it extends the model by establishing a switch equipment model, a network port feature model, and a GOOSE / SV control block dynamic model, achieving integrated modeling of secondary physical communication links and secondary logical circuits. Physical devices in the secondary circuit are treated as nodes, and fiber optic connections and virtual circuit connections as edges, proposing a representation method for secondary circuit fault characteristics. A graph neural network secondary circuit fault diagnosis model is built, utilizing a pre-trained GAT model to achieve online fault diagnosis of abnormal states of devices such as current transformers, voltage transformers, merging units, intelligent terminals, protection devices, and switches in the secondary circuit. This allows for rapid location of fault sources and fault links, improving the reliability and operation and maintenance efficiency of the power system.

[0071] 2. This invention utilizes a fusion modeling approach combining physical and logical secondary circuits, along with data acquisition and a real-time library module, to generate a visualized dynamic model of secondary circuits that supports real-time updates and dynamic adjustments.

[0072] 3. The GAT model based on the graph attention mechanism in this invention can predict fault nodes and fault types, reflect fault propagation paths, reduce reliance on expert experience, and improve operation and maintenance efficiency and decision credibility.

[0073] 4. The multi-source data features such as IED device status, network device status, and communication link status in this invention are used as inputs simultaneously, which can adapt to complex scenarios and has a stronger generalization ability. Attached Figure Description

[0074] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the drawings used in the embodiments will be briefly described below. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort. In the drawings:

[0075] Figure 1 This is a flowchart of a method for modeling and diagnosing secondary circuits in an intelligent traction substation according to an embodiment of the present invention;

[0076] Figure 2 This is a schematic diagram of the architecture of a smart traction substation secondary circuit modeling and fault diagnosis method according to an embodiment of the present invention, applicable to a station control layer SCADA monitoring platform.

[0077] Figure 3 This is a schematic diagram of the secondary circuit connection relationship in a method for modeling and diagnosing secondary circuits in an intelligent traction substation according to an embodiment of the present invention.

[0078] Figure 4 This is a flowchart of secondary circuit fault diagnosis in a method for modeling and diagnosing secondary circuits of an intelligent traction substation according to an embodiment of the present invention. Detailed Implementation

[0079] To enable those skilled in the art to better understand the technical solutions in this application, the technical solutions in the embodiments of this application will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only a part of the embodiments of this application, and not all of the embodiments. Based on the embodiments in this application, all other embodiments obtained by those skilled in the art without creative effort should fall within the scope of protection of this application.

[0080] According to an embodiment of the present invention, a method for modeling and fault diagnosis of secondary circuits in intelligent traction substations is provided.

[0081] like Figure 2 As shown, the intelligent traction substation secondary circuit modeling and fault diagnosis method is applicable to the station control layer SCADA monitoring platform. The entire SCADA monitoring platform consists of a real-time data acquisition module, a real-time database, a secondary circuit model, and an online fault diagnosis module.

[0082] The real-time data acquisition module is responsible for the access and preprocessing of multi-source heterogeneous data, collecting various status data in the secondary system in real time, registering them to the real-time database through an interface, and updating the corresponding status data in real time. The real-time database enables millisecond-level access to time-series data management. The secondary circuit model is an abstract representation of the secondary equipment used for monitoring, protection, and control in the power system and its physical or virtual terminal connection circuits, containing dual mapping relationships. The online fault diagnosis module is used to realize real-time calculation and visualization of equipment-level fault types, and its core algorithms are the expert rule base and the GAT inference network.

[0083] The present invention will now be further described in conjunction with the accompanying drawings and specific embodiments, such as... Figure 1 and Figure 4 As shown, the intelligent traction substation secondary circuit modeling and fault diagnosis method according to an embodiment of the present invention includes the following steps:

[0084] S1. Use monitoring platform software to monitor the secondary circuit parameters in the secondary circuit in real time, and use the sliding window statistical method to identify abnormal data points and obtain a set of abnormal points.

[0085] In a preferred embodiment, the step of using monitoring platform software to monitor the secondary circuit parameters in the secondary circuit in real time and identifying abnormal data points through a sliding window statistical method to obtain a set of abnormal points includes the following steps:

[0086] S11. The monitoring platform software monitors the electrical parameter measurement data, equipment operating parameters and communication parameters in the secondary circuit in real time.

[0087] S12. For electrical parameter measurement data, equipment operating parameters and communication parameters, use the sliding window statistical method and calculate the window mean and standard deviation of the current sampling point based on the preset window length and sliding step size.

[0088] Specifically, a sliding window statistical method is used, with a window length of 50 sampling points and a sliding step of 1 sampling point. The window mean μ is calculated for the current sampling point t. t :

[0089]

[0090] In the formula, μ t X represents the window mean, N represents the window length, and X represents the window mean. i This represents the sampled value at the i-th sampling point, and then the window standard deviation σ is calculated. t :

[0091]

[0092] In the formula, σ t This represents the standard deviation of the window.

[0093] S13. Calculate the range of outliers based on the window mean and standard deviation of the sampling points. If the current sampling point values ​​of electrical parameter measurement data, equipment operating parameters and communication parameters exceed the range of outliers, then the sampling point is determined to be an outlier data point; otherwise, the sampling point is determined to be a normal data point.

[0094] Specifically, if the current sample value X i Beyond μ t ±3σ t If the range is not specified, then the sampling point is determined to be abnormal.

[0095] S14. Based on the abnormal data points, record the locations of all abnormal data points to obtain a set of abnormal data points.

[0096] S2. Based on the set of abnormal points, use the expert rules stored in the monitoring platform software to match abnormal points. If the match is successful, obtain the diagnosis result based on the pre-set knowledge base and mark the fault node. Otherwise, proceed to step S3.

[0097] In a preferred embodiment, the abnormal point matching is performed based on the abnormal point set using expert rules stored in the monitoring platform software. If the matching is successful, a diagnostic result is obtained based on a pre-set knowledge base, and the faulty node is marked. Otherwise, step S3 includes the following steps:

[0098] S21. Read the predefined fault diagnosis formula configuration file in the knowledge base of the monitoring platform software to obtain the fault diagnosis formula;

[0099] S22. Determine whether the set of data points in the fault diagnosis formula is a subset of the set of abnormal points. If so, input the subset of the set of abnormal points into the fault diagnosis formula, calculate the result, combine it with the formula result description predefined in the knowledge base, obtain the diagnosis result, and mark the fault node. Otherwise, execute step S3.

[0100] S23. Visualize the obtained diagnostic results to obtain visualized diagnostic results.

[0101] In a preferred embodiment, the fault diagnosis formula includes a logical operation formula and a numerical operation formula, wherein...

[0102] The logical operation formula is:

[0103] A = P1 & P2 & ... & P n ;

[0104] The numerical calculation formula is as follows:

[0105] S = w1·P1 + w2·P2 + ... + w n ·P n ;

[0106] In the formula, A represents the result of the logical operation, {P1,P2,…,P…} n} represents the set of input points for the formula, and S represents the result of the data calculation. n} represents the weight of the input points in the formula.

[0107] Specifically, if the set of input points for the formula is a subset of the set of abnormal signal points, then the formula is substituted and the real-time values ​​of other points are read simultaneously to calculate one or more operation results. Based on the formula result description defined in the knowledge base, the diagnostic results are obtained, the fault nodes are marked, and the diagnostic results are visualized.

[0108] S3. Obtain the real-time status and historical alarm information of the secondary circuit equipment, and extract the communication link status based on the connection relationship of the secondary circuit equipment described by the secondary circuit dynamic model, and establish a fault data graph.

[0109] It should be noted that, due to the limitations of the rules, when a match cannot be successfully established, the real-time status of the secondary circuit equipment and some historical alarm information are extracted, and the communication link status is extracted based on the equipment connection relationship described by the secondary circuit model to establish a fault data graph G=(V,E,X,R).

[0110] In a preferred embodiment, the steps of acquiring the real-time status and historical alarm information of the secondary circuit equipment, extracting the communication link status, and establishing a fault data graph based on the connection relationship of the secondary circuit equipment described by the secondary circuit dynamic model include the following steps:

[0111] S31. Extract the real-time status and historical alarm information of the secondary circuit equipment based on the electrical parameter measurement data, equipment operating parameters and communication parameters;

[0112] S32. Based on graph theory, construct a dynamic model of the secondary loop and use the dynamic model of the secondary loop to describe the connection relationship between each secondary loop device.

[0113] In a preferred embodiment, the step of constructing a dynamic model of a secondary loop based on graph theory and using the dynamic model to describe the connection relationships between the secondary loop devices includes the following steps:

[0114] S321. Treat the secondary circuit equipment as nodes, and configure attribute labels for each node based on its functional type;

[0115] S322. Treat the communication links between secondary circuit devices as edges, and configure weights for each edge based on the link characteristics of the communication links.

[0116] S323. Based on the nodes and edges, construct a dynamic model of a quadratic loop in the graph structure, and establish a relational database table to store the attributes and connection relationships of the nodes and edges.

[0117] It should be noted that the secondary system loop includes both physical connection loops and logical connection loops based on the GOOSE / SV publish-subscribe mechanism. The complete model construction of the secondary loop requires modeling not only the intelligent devices and switching devices, but also establishing a model describing the direct communication link connections between the devices. A graph theory approach is used to abstract the devices and communication links in the secondary loop. First, devices are represented as nodes in a graph model, with each node assigned attribute labels based on its function type (e.g., measurement, protection, control, switch, etc.). Second, communication links are represented as edges, with physical connections or logical signal transmission links abstracted as edges in the graph model, and weights added to represent link characteristics (e.g., link type, transmission delay, etc.). Finally, a directed graph model reflecting the physical connections and logical signal transmission of the devices is constructed. To store the attributes and connection relationships of nodes and edges and dynamically update their states, corresponding relational database tables also need to be established.

[0118] The overall steps are as follows:

[0119] Step 1: Define the database table for the intelligent device object IEDevice: Key fields Type, CPUSystemEx, MemSystemEx, Temperature, PSState, EthernetList, and AlmEvents represent the device type, CPU utilization, memory utilization, temperature, power module status, network port list, and cumulative number of abnormal events of the IED device, respectively.

[0120] Step 2: Define the Switch database table for the Switch object: Key fields CPUSystemEx, MemSystemEx, NMS_Temperature, PSState, and EthernetList, which represent the switch's CPU utilization, memory utilization, temperature, power module status, and network port list, respectively.

[0121] Step 3: Ethernet database table definition for communication network interface objects: Key fields ConnectedEthernet, InPktLossRate, OutPktLossRate, Status, and Traffic represent the peer network interface, receive packet loss rate, send packet loss rate, online status, and port traffic, respectively. Based on the list of network interfaces included in each physical device in the secondary loop and its corresponding peer network interface, the real-time physical link connection relationship can be obtained.

[0122] Step 4: GOOSE / SV Control Block Object ControlBlock Database Table Definition: Key fields include Type, SendIED, ReceiveIED, AppID, StaNum, SmpCnt, State, LossRate, and RTT, representing the control block type (GOOSE / SV), message sending IED device, message subscribing IED device, unique control block identifier, status number, sample count, connectivity status, packet loss rate, and communication latency, respectively. This model configuration information comes from the SCD file and can describe the logical connection relationships between IED devices.

[0123] Step 5: Generate a real-time dynamic model. Utilize the SCADA monitoring system's data acquisition module to collect and update the real-time parameters of the IEDevice, Switch, Ethernet, and ControlBlock object instances in the secondary loop. Collect and write the real-time running status of the IEDevice object via the MMS protocol. Use the SNMP protocol to poll and write the real-time running indicators of the Switch object and the peer network interface identifier of the Ethernet object. Network probes capture GOOSE / SV message text and parse out the StaNum and SmpCnt keywords, representing the sequence number of the GOOSE / SV message, respectively. Based on the time difference between the message sender and receiver, the number of packets sent by the message sender, and the number of packets received by the message receiver, obtain the communication latency and packet loss rate. Communication is considered interrupted when three consecutive packets are lost. Write the calculation results to the corresponding real-time attributes of the ControlBlock object.

[0124] Step 6: Generate a dynamic model of a graph structure with secondary loops. Mark the already established IEDevice and Switch objects as independent nodes in the graph structure; read the EthernetList property of the nodes, obtain the physical connection relationship based on the real-time ConnectedEthernet property of each network interface, and mark it as the physical connection edge of the graph structure; read the SendIED and ReceiveIED configuration properties of the ControlBlock object to obtain the GOOSE / SV logical connection relationship, and mark it as the logical connection edge of the graph structure. In this way, even if the network interface connection relationship of each device changes, the model structure can be updated in real time.

[0125] A well-constructed typical quadratic loop relationship diagram structure model is as follows: Figure 3 As shown, Figure 3 The middle node represents secondary equipment or switch equipment such as measurement and control unit, merging unit, protection device, and intelligent terminal, i.e., the established IED or Switch object; the bidirectional arrow represents the physical connection edge; the unidirectional arrow represents the GOOSE / SV logical connection edge.

[0126] S33. Based on the connection relationship between each secondary circuit device, extract the communication link status and establish a fault data diagram according to the communication link status.

[0127] S4. Construct a secondary circuit fault diagnosis GAT model, input the fault diagram data into the secondary circuit fault diagnosis GAT model, and obtain the diagnosis results through reasoning calculation.

[0128] Specifically, components in a secondary loop can be categorized by function into signal sources, relay devices, signal sinks, and connecting elements. When any device in a secondary loop fails, the power flow distribution of the secondary loop communication network changes. Furthermore, when a device fails, all related devices in the secondary loop generate abnormal signals, making fault localization difficult. The Gaussian Atlas (GAT) method maps the varied morphology and complex physical characteristics of secondary loops into node and edge connection information in a graph. Utilizing GAT technology, it achieves graph-domain representation of the characteristics of embedded nodes and fault propagation paths, enabling the organic integration of physical modeling and data-driven approaches. This improves the practical application and generalization capabilities of artificial intelligence methods in secondary loop fault diagnosis and localization.

[0129] Graph Attention Networks (GATs) introduce an attention mechanism, assigning different attention weights to the neighbors of each node, thereby dynamically learning the importance relationships between nodes. This weight allocation is based on the correlation between node features, enabling the model to adaptively focus on node information that is more critical to the current task. It can simultaneously identify abnormal nodes and edges, and obtain fault propagation paths.

[0130] The quadratic circuit is represented as a weighted heterogeneous graph G = (V, E, X, R), where V = (v1, v2, ..., v i () represents a set of i device nodes, including protection devices, smart terminals, switches, etc.; E = (e1, e2, ..., e) m () represents the set of connection relationships of m connecting edges; X∈R nxd R represents the d-dimensional feature matrix of n nodes; R represents the type of communication link between nodes, including physical link, GOOSE signal link or SV signal link.

[0131] The secondary loop fault location problem can be represented as a graph classification problem. The input parameters are the feature representations of nodes and edges in the graph data, and the output is a probability vector representation of the fault type generated by each node. The steps for constructing the fault location GAT model are as follows:

[0132] 1) Input Feature Construction

[0133] Based on the constructed dynamic model of the secondary loop diagram structure, real-time attributes that can characterize equipment faults are extracted, such as real-time collected values ​​of CPU, memory, temperature, power status, and number of abnormal alarm events of equipment nodes; packet loss rate, traffic, and online status of network ports; and latency, packet loss rate, and connectivity status of GOOSE / SV communication links. The mathematical expression of the fault characteristics of node v is obtained as follows:

[0134]

[0135] Among them, X v Encoding The one-hot encoding representing the device type (v) is used if there are 5 device types and the current type is type 3. v Encoding = [0,0,1,0,0];

[0136] Features representing the real-time state of device v are Z-score standardized for each feature of the device's real-time operating state:

[0137]

[0138] Where c, m, t, and p represent the real-time collected values ​​of CPU, memory, temperature, and power status, respectively, and μ and σ represent the mean and standard deviation of the corresponding training set data, respectively.

[0139] This represents the feature aggregation of the port status of device v, assuming the number of ports is n and the port traffic set is {f1, f2, ..., f...} n The port state set is {s1, s2, ..., s}. n Construct the following feature vector:

[0140]

[0141] Where, n norm μ represents the normalized value indicating the number of ports. f ,f max ,f min ,σ f These represent the mean, maximum, minimum, and standard deviation of port traffic, respectively. s Indicates the proportion of ports in normal state;

[0142] This indicates the status of the communication link associated with device v. The link type includes SV link set, GOOSE link, and fiber optic link, and its characteristics are expressed as follows:

[0143]

[0144] Where c∈{0,1} represents the link connectivity status, r c μ represents the on / off status ratio of a link type. d μ represents the average latency of a link type. l This represents the average packet loss rate for a link type; N∈[0,1] represents the normalized value of the number of alarms generated in the past hour; sv represents SV type link; of represents fiber type link; goose represents GOOSE type link.

[0145]

[0146] In the formula, n links This indicates the number of such communication links.

[0147] 2) Graph attention weight calculation

[0148] By introducing an attention mechanism, the attention score e between device node i and device node j in the secondary loop fault graph structure is calculated. ij Dynamically capturing the relationships between nodes can better explain the fault propagation chain. Only when a connection exists between nodes, i.e., E... ij The attention score is calculated when the value is 1; otherwise, the attention score is negative infinity.

[0149] Then, the attention score is normalized to a probability distribution by applying the softmax function, and the attention weight is obtained. The normalized value reflects the importance of neighbor j node to neighbor i node.

[0150] 3) Node information transmission operation

[0151] Based on the calculated attention weights, the node information aggregation function is defined as a weighted aggregation of neighbor information.

[0152] 4) Fault Classifier

[0153] After multiple GAT layers of information transmission and computation, the output state of the node at layer l is obtained. Then, a fully connected neural network layer is used for fault classification. The sigmoid function is chosen as the activation function, transforming the fault diagnosis problem into a multi-class classification problem. The model performance is measured using the cross-entropy loss function.

[0154] This completes the construction of the GAT fault diagnosis model. The model uses an end-to-end architecture of feature input layer → multiple GAT operation layers → classification output layer. After training, it can be used for accurate fault diagnosis of secondary circuits.

[0155] The model outputs a similar fault type matrix [0,1,0,0,1...]. If all values ​​in the matrix are 0, it indicates that the node is fault-free. When a value of 1 appears, the fault type corresponding to the faulty node can be obtained according to the one-hot encoding of the fault type. Furthermore, through backpropagation training, the model optimizes the weight matrix W and attention vector a, aligning high-weight edges with the actual fault paths. This allows the high-weight edges (e...) to be... ij The threshold is considered to be the critical path for fault propagation, and the attention weight of abnormal devices and their neighbors should be significantly higher than the normal value.

[0156] In a preferred embodiment, the steps of constructing a secondary circuit fault diagnosis GAT model, inputting fault diagram data into the secondary circuit fault diagnosis GAT model, and obtaining the diagnosis result through reasoning calculation include the following steps:

[0157] S41. Construct and train a secondary circuit fault diagnosis GAT model based on the secondary circuit dynamic model;

[0158] In a preferred embodiment, the construction and training of the secondary circuit fault diagnosis GAT model based on the secondary circuit dynamic model includes the following steps:

[0159] S411. Based on the dynamic model of the secondary circuit, extract the real-time attributes used to characterize the faults of the secondary circuit equipment, obtain the fault characteristics of the secondary circuit equipment, and construct the secondary circuit fault diagram structure based on the fault characteristics of the secondary circuit equipment.

[0160] Specifically, based on the secondary loop dynamic model, real-time attributes for characterizing secondary loop equipment faults are extracted to obtain secondary loop equipment fault characteristics. The construction of a secondary loop fault graph structure based on these characteristics includes the following steps:

[0161] S4111. Based on the dynamic model of the secondary circuit, a temporal convolutional network is used to extract real-time attributes used to characterize the faults of the secondary circuit equipment, thereby obtaining the fault characteristics of the secondary circuit equipment.

[0162] S4112. Based on the fault characteristics of secondary circuit equipment, the fault characteristics of secondary circuit equipment are selected by using the causal discovery method based on residual clustering to obtain key fault characteristics.

[0163] The process of selecting key fault features based on the fault characteristics of secondary circuit equipment, using a causal discovery method based on residual clustering, involves the following steps:

[0164] S41121. Preprocess the fault characteristics of the secondary circuit equipment, and use the optimal sample size analysis method to estimate the optimal sample size of the preprocessed secondary circuit equipment fault characteristics, and construct a stratified sampling data subset based on the optimal sample size estimation results.

[0165] It should be noted that the various fault characteristics of the secondary circuit equipment are preprocessed for normalization, such as standardizing the data to the [0,1] interval to eliminate the influence of different dimensions.

[0166] The formula for estimating the optimal sample size of the preprocessed secondary circuit equipment fault characteristics using the optimal sample size analysis method is as follows:

[0167]

[0168] In the formula, m represents the optimal sample size estimate, r1 and r2 represent the means of each fault characteristic variable, q1 and q2 represent the standard deviations of each fault characteristic variable, M represents the significance level, usually taken as 0.05, 1-B represents the statistical power, usually taken as 0.8, and Z... M / 2 The M / 2 quantile of the standard normal distribution, Z 1-B The first-B quantile of the standard normal distribution.

[0169] Based on the optimal sample size estimate, stratified sampling is performed on the preprocessed data. Stratified sampling involves dividing the population into several strata and then independently drawing samples from each stratum. This method ensures that representative samples are selected from each stratum, thereby improving sample diversity. The sampling steps are as follows:

[0170] Calculate the proportion of each layer (e.g., the proportion of the i-th layer w) i =n i / N,n i (where N is the number of samples within the stratum and N is the total population).

[0171] Based on the optimal sample size m and the stratum proportion, determine the number of samples m to be drawn from each stratum. i =m·w i .

[0172] Randomly select m from each layer i The samples are merged to form the final sample set.

[0173] S41122. Based on the Bayesian information criterion, construct an equilibrium relationship model for each stratified sampling data subset and calculate the residual set.

[0174] Specifically, the Bayesian Information Criterion (BIC) is a model selection criterion that chooses the optimal model by balancing goodness of fit (e.g., log-likelihood) and model complexity (e.g., number of parameters). Its formula is:

[0175] BIC = -2lnL(θ) + klnn;

[0176] Where: L(θ) represents the likelihood function of the model on the data, k represents the number of model parameters (complexity), and n represents the sample size.

[0177] In the equilibrium relationship model, it is assumed that there is a linear or nonlinear relationship between fault features (X) and faults (Y). Then, for each stratified sampling data subset, the corresponding BIC value is calculated using a linear model or a quadratic model. The model with the smallest BIC value is then selected as the optimal equilibrium relationship model for that subset.

[0178] Furthermore, after obtaining the optimal equilibrium relationship model, the optimal equilibrium relationship model is used to predict all samples in each stratified sampling subset to obtain predicted values. For each sample, the difference between the actual value and the predicted value is calculated to obtain the residual. The residuals of all samples are summed to form the residual set of that subset.

[0179] S41123. Summarize the residual sets of all data subsets to form residual clusters, and calculate the mean distribution of the residual clusters respectively.

[0180] Specifically, by arranging or merging the residuals of each subset into an array or list in sequence, a residual cluster can be formed. If the residual cluster is formed by merging the residuals of multiple hierarchical sampling sets, it can be re-divided into several subgroups according to the subset partitioning criteria (such as equipment type, fault type, etc.). Finally, the mean of the residuals is calculated for each subgroup to obtain the mean distribution.

[0181] S41124. By comparing the differences in the mean distribution of residual clusters and combining them with causal direction discrimination, key fault characteristics are selected.

[0182] It should be noted that, since the mean distribution of the residual clusters was calculated by subset partitioning in step S41123, the mean residual of each subset reflects the systematic bias of the model on that subset. Then, box plots, violin plots, or error bar plots are used to visually demonstrate the differences in the distribution of the mean residuals across different subsets. Hypothesis testing methods (such as t-tests and ANOVA tests) are then used to verify whether there are significant differences in the mean residuals among different subsets.

[0183] For example, if the mean residual of subset 1 is 0.5 and that of subset 2 is -0.3, and the statistical test shows a significant difference (p<0.05), it indicates that the model overestimates the impact of failure in subset 1 or underestimates it in subset 2.

[0184] This involves using a causal orientation discrimination method to determine the causal relationship between fault characteristics and faults, and to screen out characteristics that have a significant causal impact on the fault. Specifically, this includes the following steps:

[0185] Calculate the Pearson correlation coefficient or mutual information between residuals to identify significantly correlated feature pairs.

[0186] Use causal discovery algorithms (such as PC algorithm, FCI algorithm) to test the conditional independence of features and faults after controlling for other variables.

[0187] If the data has time series characteristics, use Granger causality test or convergent cross-mapping (CCM) to determine the causal direction.

[0188] Based on all the above analyses, features that simultaneously meet the following conditions are selected:

[0189] The residual mean differs significantly across different subsets; it is significantly correlated with the residual (Pearson or MI); it meets the causal discovery criteria and there is a clear causal path; if it is time series data, it also passes the Granger / CCM test.

[0190] S4113. Based on key fault characteristics, use fault tree analysis to construct a secondary loop fault diagram structure.

[0191] It should be noted that Fault Tree Analysis (FTA) is a top-down deductive analysis method used to identify the causes of system failures and their combinations. It constructs a fault tree, decomposing top-level failures (such as overall secondary circuit failures) into intermediate events and bottom-level events (such as key failure characteristics), thus visually demonstrating the propagation path and causal relationships of the failure.

[0192] Specifically, based on key fault characteristics, constructing a secondary loop fault diagram structure using fault tree analysis includes the following steps:

[0193] S41131. Define the overall failure of the secondary circuit as the top event of the fault tree, such as "secondary circuit equipment failure".

[0194] S41132. Based on the key failure characteristics, the top-level failure is decomposed into intermediate events. The intermediate events are the direct causes of the top-level failure and are combinations or interactions of certain key failure characteristics.

[0195] It should be noted that intermediate events are the direct causes of top-level failures and are combinations or interactions of certain key failure characteristics. For example, if key failure characteristics include "abnormal voltage," "current fluctuations," and "overheating," then intermediate events are combinations of these characteristics, such as "abnormal voltage and current fluctuations" or "overheating leading to equipment overheating."

[0196] S41133. Using logic gates (such as AND gates and OR gates) to connect intermediate events and underlying events represents the fault propagation logic: an AND gate indicates that the output event will only occur when all input events occur simultaneously. An OR gate indicates that the output event will occur when any one of the input events occurs.

[0197] S41134. Connect all event nodes in the hierarchical order of "top-level event → intermediate event → basic event" to construct a complete fault tree structure.

[0198] S412. Based on the attention mechanism, calculate the attention score between each secondary loop device node in the secondary loop fault diagram structure;

[0199] As a preferred embodiment, the formula for calculating the attention score between each secondary loop device node in the secondary loop fault diagram structure based on the attention mechanism is as follows:

[0200]

[0201] In the formula, e ij Let represent the attention score between secondary loop device node i and secondary loop device node j, LeakyReLU represent the activation function, W represent the learnable weight matrix of the secondary loop device node, a represent the learnable attention vector, l represent the current GAT layer number, T represent the transpose of the vector, X represent the input feature vector of a single node, || represent the concatenation operation, and R ij This represents the communication link coefficient between secondary circuit device nodes.

[0202] S413. Normalize the attention score using the softmax function to obtain the attention weight;

[0203] In a preferred embodiment, the attention score is normalized using the softmax function, and the formula for calculating the attention weight is as follows:

[0204]

[0205] In the formula, α ij Represents attention weight, e ij Let N(i) represent the attention score between secondary loop device node i and secondary loop device node j, and let N(i) represent the set of neighbors of secondary loop device node i (including itself). ik This represents the attention score between secondary loop device node i and secondary loop device node k.

[0206] S414. Based on the attention weight, the neighbor features of each secondary loop device node are weighted and aggregated to update the representation vector of each secondary loop device node, thereby obtaining the context feature embedding representation of each secondary loop device.

[0207] In a preferred embodiment, the calculation formula for weighted aggregation of neighbor features of each secondary loop device node based on attention weights is as follows:

[0208]

[0209] In the formula, Let W represent the layer l state of secondary loop device node i, W represent the learnable weight matrix, and σ represent the activation function. This indicates the state of the next-level neighbor node of the secondary loop device node i.

[0210] S415. Based on the cross-entropy loss function, the contextual features of the secondary circuit equipment are embedded to represent the input fault classifier for training, thus obtaining the GAT model of the secondary circuit equipment fault.

[0211] In a preferred embodiment, the expression for the cross-entropy loss function is:

[0212]

[0213] In the formula, L represents the cross-entropy loss function, N represents the number of samples, C represents the number of classes, and y i,c p represents the one-hot encoding of the true label of sample i. i,c This represents the probability that sample i belongs to category c, as predicted by the GAT model for secondary circuit fault diagnosis.

[0214] S42. Input the fault diagram data into the pre-trained secondary circuit fault diagnosis GAT model, and obtain the fault type code of each secondary circuit device and the fault association weight between each secondary circuit device through reasoning calculation.

[0215] S43. Based on the fault type codes of each secondary circuit device and the fault association weights between each secondary circuit device, interpret the fault types of the secondary circuit devices, filter the propagation links that meet the weight conditions, mark the fault nodes and fault links, and obtain the diagnostic results.

[0216] S44. Visualize the obtained diagnostic results to obtain visualized diagnostic results.

[0217] In a preferred embodiment, the step of inputting fault map data into a pre-trained secondary loop fault diagnosis GAT model and obtaining the fault type code of each secondary loop device and the fault association weight between each secondary loop device through inference calculation includes the following steps:

[0218] S421. Based on the fault graph data, determine the characteristics of all fault nodes, the adjacency matrix of the fault nodes, and the edge type matrix;

[0219] S422. Input all the fault node features, the adjacency matrix and edge type matrix of the fault node into the pre-trained secondary loop fault diagnosis GAT model, and obtain the output of each fault node after information transmission and operation through multiple GAT layers.

[0220] S423. Use the output of each fault node as the input of the classification layer in the secondary loop fault diagnosis GAT model to obtain the fault classification result, and output the fault type encoding vector through the secondary loop fault diagnosis GAT model.

[0221] S424. Interpret the fault type encoding vector as fault type and fault path, and combine it with the secondary loop dynamic model to obtain the fault association weight between each secondary loop device.

[0222] Specifically, after defining the fault diagnosis problem of the secondary circuit as a graph classification problem, for a certain fault graph data G=(V,E,X,R), the input of the graph neural network is the features X of all fault nodes in the graph data, the adjacency matrix E of the fault nodes, and the edge type matrix R. The features x of a certain fault node are processed by multiple GAT layers to obtain the output of each node. Finally, the nodes are used as the input of the classification layer to obtain the fault classification result. The output of the model is the fault type vector, which is interpreted as the specific fault type and fault path.

[0223] During fault interpretation, the GAT model outputs node v with feature Y. V ∈{0,1} D (D is the total number of fault types), the fault type is determined by the index of the maximum value: FaultType v =argmaxY V FaultType v Indicate the fault type; then, extract connections with edge weights higher than the threshold (default threshold = 0.7): E fault ={E ij =1|a ij >0.7};

[0224] Among them, E fault E represents a connection whose edge weight is higher than a threshold. ij This represents the connection between node i and node j, a ij This represents the weight of the edge connecting node i and node j; finally, the upstream devices are traced back from the faulty node, the propagation links that meet the weight conditions are selected, the faulty nodes and faulty links are labeled, and the diagnostic results are visualized.

[0225] In summary, by utilizing the above-mentioned technical solutions of this invention, the secondary circuit is modeled from a graph perspective. Based on the IEC61850 standard IED model, a switch equipment model, a network port feature model, and a GOOSE / SV control block dynamic model are extended, achieving integrated modeling of secondary physical communication links and secondary logical circuits. Physical devices in the secondary circuit are considered as nodes, and fiber optic connections and virtual circuit connections are considered as edges, thus proposing a representation method for secondary circuit fault characteristics. A graph neural network secondary circuit fault diagnosis model is built, utilizing a pre-trained GAT model to achieve online fault diagnosis of abnormal states of devices such as current transformers, voltage transformers, merging units, intelligent terminals, protection devices, and switches in the secondary circuit. This allows for rapid location of fault sources and fault links, improving the reliability and operation and maintenance efficiency of the power system. This invention utilizes a fusion modeling approach of physical and logical secondary circuits, combined with data acquisition and real-time library modules, to generate a visualized dynamic model of the secondary circuit, supporting real-time updates and dynamic adjustments. The GAT model based on the graph attention mechanism of this invention can predict fault nodes and fault types, reflect fault propagation paths, reduce reliance on expert experience, and improve operation and maintenance efficiency and decision-making credibility. The present invention uses multiple data features, such as IED device status, network device status, and communication link status, as inputs simultaneously, adapting to complex scenarios and having stronger generalization capabilities.

[0226] Those skilled in the art will understand that embodiments of the present invention can be provided as methods, systems, or computer program products. Therefore, the present invention can take the form of a completely hardware embodiment, a completely software embodiment, or an embodiment combining software and hardware aspects. Furthermore, the present invention can take the form of a computer program product embodied on one or more computer-usable storage media (including, but not limited to, disk storage, optical storage, etc.) containing computer-usable program code.

[0227] The specific embodiments described above further illustrate the objectives, technical solutions and beneficial effects of the present invention in detail. It should be understood that the above description is only a specific embodiment of the present invention and is not intended to limit the scope of protection of the present invention. Any modifications, equivalent substitutions, improvements, etc. made within the spirit and principles of the present invention should be included in the scope of protection of the present invention.

Claims

1. A method for modeling and fault diagnosis of secondary circuits in intelligent traction substations, characterized in that, Includes the following steps: S1. Use monitoring platform software to monitor the secondary circuit parameters in the secondary circuit in real time, and use the sliding window statistical method to identify abnormal data points and obtain a set of abnormal points. S2. Based on the set of abnormal points, use the expert rules stored in the monitoring platform software to match abnormal points. If the match is successful, obtain the diagnosis result based on the pre-set knowledge base and mark the fault node. Otherwise, proceed to step S3. S3. Obtain the real-time status and historical alarm information of the secondary circuit equipment, and extract the communication link status based on the connection relationship of the secondary circuit equipment described by the secondary circuit dynamic model, and establish a fault data graph. S4. Construct a secondary circuit fault diagnosis GAT model, input the fault diagram data into the secondary circuit fault diagnosis GAT model, and obtain the diagnosis results through reasoning calculation.

2. The method for modeling and fault diagnosis of secondary circuits in intelligent traction substations according to claim 1, characterized in that, The process of using monitoring platform software to monitor secondary circuit parameters in real time and identifying abnormal data points using a sliding window statistical method to obtain a set of abnormal points includes the following steps: S11. The monitoring platform software monitors the electrical parameter measurement data, equipment operating parameters and communication parameters in the secondary circuit in real time. S12. For electrical parameter measurement data, equipment operating parameters and communication parameters, use the sliding window statistical method and calculate the window mean and standard deviation of the current sampling point based on the preset window length and sliding step size. S13. Calculate the range of outliers based on the window mean and standard deviation of the sampling points. If the current sampling point values ​​of electrical parameter measurement data, equipment operating parameters and communication parameters exceed the range of outliers, then the sampling point is determined to be an outlier data point; otherwise, the sampling point is determined to be a normal data point. S14. Based on the abnormal data points, record the locations of all abnormal data points to obtain a set of abnormal data points.

3. The method for modeling and fault diagnosis of secondary circuits in intelligent traction substations according to claim 1, characterized in that, The step of matching anomalies based on the set of anomaly points using expert rules stored in the monitoring platform software is as follows: If a match is successful, a diagnostic result is obtained based on a pre-set knowledge base, and the faulty node is marked; otherwise, step S3 includes the following steps: S21. Read the predefined fault diagnosis formula configuration file in the knowledge base of the monitoring platform software to obtain the fault diagnosis formula; S22. Determine whether the set of data points in the fault diagnosis formula is a subset of the set of abnormal points. If so, input the subset of the set of abnormal points into the fault diagnosis formula, calculate the result, combine it with the formula result description predefined in the knowledge base, obtain the diagnosis result, and mark the fault node. Otherwise, execute step S3. S23. Visualize the obtained diagnostic results to obtain visualized diagnostic results.

4. The method for modeling and fault diagnosis of secondary circuits in intelligent traction substations according to claim 3, characterized in that, The fault diagnosis formula includes logical operation formulas and numerical operation formulas, wherein... The logical operation formula is: A=P1&P2&...&P n ; The numerical calculation formula is as follows: S=w1·P1+w2·P2+...+w n ·P n ; In the formula, A represents the result of the logical operation, {P1,P2,…,P…} n } represents the set of input points for the formula, and S represents the result of the data calculation. n } represents the weight of the input points in the formula.

5. The method for modeling and fault diagnosis of secondary circuits in intelligent traction substations according to claim 1, characterized in that, The process of acquiring real-time status and historical alarm information of secondary circuit equipment, extracting communication link status, and establishing a fault data graph based on the connection relationship of secondary circuit equipment described by the secondary circuit dynamic model includes the following steps: S31. Extract the real-time status and historical alarm information of the secondary circuit equipment based on the electrical parameter measurement data, equipment operating parameters and communication parameters; S32. Based on graph theory, construct a dynamic model of the secondary loop and use the dynamic model of the secondary loop to describe the connection relationship between each secondary loop device. S33. Based on the connection relationship between each secondary circuit device, extract the communication link status and establish a fault data diagram according to the communication link status.

6. The method for modeling and fault diagnosis of secondary circuits in intelligent traction substations according to claim 5, characterized in that, The method of constructing a dynamic model of a secondary loop based on graph theory, and using the dynamic model to describe the connection relationships between the secondary loop devices, includes the following steps: S321. Treat the secondary circuit equipment as nodes, and configure attribute labels for each node based on its functional type; S322. Treat the communication links between secondary circuit devices as edges, and configure weights for each edge based on the link characteristics of the communication links. S323. Based on the nodes and edges, construct a dynamic model of a quadratic loop in the graph structure, and establish a relational database table to store the attributes and connection relationships of the nodes and edges.

7. The method for modeling and fault diagnosis of secondary circuits in intelligent traction substations according to claim 1, characterized in that, The process of constructing a secondary circuit fault diagnosis GAT model, inputting fault diagram data into the secondary circuit fault diagnosis GAT model, and obtaining the diagnosis result through inference calculation includes the following steps: S41. Construct and train a secondary circuit fault diagnosis GAT model based on the secondary circuit dynamic model; S42. Input the fault diagram data into the pre-trained secondary circuit fault diagnosis GAT model, and obtain the fault type code of each secondary circuit device and the fault association weight between each secondary circuit device through reasoning calculation. S43. Based on the fault type codes of each secondary circuit device and the fault association weights between each secondary circuit device, interpret the fault types of the secondary circuit devices, filter the propagation links that meet the weight conditions, mark the fault nodes and fault links, and obtain the diagnostic results. S44. Visualize the obtained diagnostic results to obtain visualized diagnostic results.

8. The method for modeling and fault diagnosis of secondary circuits in intelligent traction substations according to claim 7, characterized in that, The construction and training of the secondary circuit fault diagnosis GAT model based on the secondary circuit dynamic model includes the following steps: S411. Based on the dynamic model of the secondary circuit, extract the real-time attributes used to characterize the faults of the secondary circuit equipment, obtain the fault characteristics of the secondary circuit equipment, and construct the secondary circuit fault diagram structure based on the fault characteristics of the secondary circuit equipment. S412. Based on the attention mechanism, calculate the attention score between each secondary loop device node in the secondary loop fault diagram structure; S413. Normalize the attention score using the softmax function to obtain the attention weight; S414. Based on the attention weight, the neighbor features of each secondary loop device node are weighted and aggregated to update the representation vector of each secondary loop device node, thereby obtaining the context feature embedding representation of each secondary loop device. S415. Based on the cross-entropy loss function, the contextual features of the secondary circuit equipment are embedded to represent the input fault classifier for training, thus obtaining the GAT model of the secondary circuit equipment fault.

9. The method for modeling and fault diagnosis of secondary circuits in an intelligent traction substation according to claim 8, characterized in that, The formula for calculating the attention score between secondary loop device nodes in the secondary loop fault graph structure based on the attention mechanism is as follows: In the formula, e ij Let represent the attention score between secondary loop device node i and secondary loop device node j, LeakyReLU represent the activation function, W represent the learnable weight matrix of the secondary loop device node, a represent the learnable attention vector, l represent the current GAT layer number, T represent the transpose of the vector, X represent the input feature vector of a single node, || represent the concatenation operation, and R ij This represents the communication link coefficient between secondary circuit device nodes.

10. The method for modeling and fault diagnosis of secondary circuits in an intelligent traction substation according to claim 8, characterized in that, The attention score is normalized using the softmax function, and the formula for calculating the attention weight is as follows: In the formula, α ij Represents attention weight, e ij Let N(i) represent the attention score between secondary loop device node i and secondary loop device node j, and let N(i) represent the neighbor set of secondary loop device node i. ik This represents the attention score between secondary loop device node i and secondary loop device node k.

11. The method for modeling and fault diagnosis of secondary circuits in intelligent traction substations according to claim 8, characterized in that, The calculation formula for weighted aggregation of neighbor features of each secondary loop device node based on attention weight is as follows: Where, Let W represent the layer l state of secondary loop device node i, W represent the learnable weight matrix, and σ represent the activation function. This indicates the state of the next-level neighbor node of the secondary loop device node i.

12. The method for modeling and fault diagnosis of secondary circuits in intelligent traction substations according to claim 8, characterized in that, The expression for the cross-entropy loss function is: In the formula, L represents the cross-entropy loss function, N represents the number of samples, C represents the number of classes, and y i,c p represents the one-hot encoding of the true label of sample i. i,c This represents the probability that sample i belongs to category c, as predicted by the GAT model for secondary circuit fault diagnosis.

13. The method for modeling and fault diagnosis of secondary circuits in an intelligent traction substation according to claim 8, characterized in that, The process of inputting fault diagram data into a pre-trained secondary loop fault diagnosis GAT model and obtaining the fault type code of each secondary loop device and the fault association weight between each secondary loop device through inference calculation includes the following steps: S421. Based on the fault graph data, determine the characteristics of all fault nodes, the adjacency matrix of the fault nodes, and the edge type matrix; S422. Input all the fault node features, the adjacency matrix and edge type matrix of the fault node into the pre-trained secondary loop fault diagnosis GAT model, and obtain the output of each fault node after information transmission and operation through multiple GAT layers. S423. Use the output of each fault node as the input of the classification layer in the secondary loop fault diagnosis GAT model to obtain the fault classification result, and output the fault type encoding vector through the secondary loop fault diagnosis GAT model. S424. Interpret the fault type encoding vector as fault type and fault path, and combine it with the secondary loop dynamic model to obtain the fault association weight between each secondary loop device.

Images