Fault detection method and device

By constructing the correlation knowledge graph and state clustering of high-voltage relay protection devices, the problems of low fault detection efficiency and insufficient accuracy are solved, and automated efficient fault detection is realized, and detection accuracy is improved.

CN119004065APending Publication Date: 2024-11-22GUANGDONG POWER GRID CO LTD +1

Patent Information

Application Number
CN202410912256.3
Authority / Receiving Office
CN · China
Patent Type
Applications(China)
Current Assignee / Owner
Filing Date
2024-07-09
Publication Date
2024-11-22

AI Technical Summary

Technical Problem

The fault detection efficiency of high-voltage relay protection devices is low and the accuracy is insufficient, which is mainly due to the reliance on manual interpretation and experience, resulting in low efficiency and poor accuracy.

Method used

By determining feature indicators based on the current sampled data sequence of the high-voltage relay protection device, building a correlation knowledge graph with static data and preset device model library and image library, and using diagnostic rules and status clustering to achieve automatic fault detection, improving detection efficiency and accuracy sex.

Benefits of technology

Automatic fault detection of high-voltage relay protection devices is realized, the fault detection efficiency and accuracy are improved, manual intervention is reduced, and accurate perception of faults is enhanced.

✦ Generated by Eureka AI based on patent content.

Smart Images

  • Figure CN119004065A_ABST
    Figure CN119004065A_ABST
Patent Text Reader

Abstract

The invention discloses a fault detection method and equipment, and the method comprises the steps: determining a characteristic index according to a current sampling data sequence of a high-voltage relay protection device; determining an associated knowledge graph of the high-voltage relay protection device according to the static data of the high-voltage relay protection device, the current sampling data sequence, a preset equipment model library and an image library; determining an intermediate diagnosis conclusion according to the feature index and a diagnosis rule corresponding to the feature index in the associated knowledge graph; determining a current classification result of the feature index according to the distance between the feature index and each classification result; and determining a fault detection result of the high-voltage relay protection device according to the intermediate diagnosis conclusion and the current classification result. According to the fault detection method, the fault detection efficiency and the fault detection accuracy are improved.
Need to check novelty before this filing date? Find Prior Art

Description

Technical Field

[0001] The present invention relates to the field of data processing technology, and in particular to a fault detection method and device. Background Art

[0002] High-voltage relay protection devices refer to automation devices used to monitor and protect major equipment such as transmission lines, transformers, and busbars in high-voltage power systems. In order to ensure the normal operation of high-voltage power systems, it is very important to perform fault detection on high-voltage relay protection devices.

[0003] At present, the operation and maintenance of high-voltage relay protection devices mainly rely on regular inspections and post-event analysis. The operation and maintenance personnel manually interpret the sampled data of the high-voltage relay protection device based on their work experience and the parameters of the high-voltage relay protection device to determine whether the high-voltage relay protection device has a fault. Therefore, the current fault detection method for high-voltage relay protection devices has the problems of low efficiency and low accuracy due to strong subjectivity. Summary of the invention

[0004] The present invention provides a fault detection method and device to solve the technical problems of low efficiency and low accuracy in related art when performing fault detection on a high-voltage relay protection device.

[0005] According to one aspect of the present invention, there is provided a fault detection method, comprising:

[0006] Determine characteristic indicators according to the current sampling data sequence of the high-voltage relay protection device;

[0007] Determine the associated knowledge graph of the high-voltage relay protection device according to the static data of the high-voltage relay protection device, the current sampling data sequence, and a preset device model library and image library;

[0008] Determine an intermediate diagnosis conclusion according to the characteristic indicator and the diagnosis rule corresponding to the characteristic indicator in the associated knowledge graph;

[0009] Determine the current classification result of the feature indicator according to the distance between the feature indicator and each classification result; wherein each classification result is a result obtained by clustering all states in the associated knowledge graph, and each classification result is used to characterize a working state of the high-voltage relay protection device;

[0010] The fault detection result of the high-voltage relay protection device is determined according to the intermediate diagnosis conclusion and the current classification result.

[0011] According to another aspect of the present invention, an electronic device is provided, the electronic device comprising:

[0012] at least one processor; and

[0013] a memory communicatively connected to the at least one processor; wherein,

[0014] The memory stores a computer program executable by the at least one processor, and the computer program is executed by the at least one processor so that the at least one processor can execute the fault detection method described in any embodiment of the present invention.

[0015] According to another aspect of the present invention, a computer-readable storage medium is provided, wherein the computer-readable storage medium stores a computer program, wherein the computer program is used to enable a processor to implement the fault detection method described in any embodiment of the present invention when executed.

[0016] According to another aspect of the present invention, a computer program product is provided, comprising a computer program, wherein when the computer program is executed by a processor, the fault detection method according to any embodiment of the present invention is implemented.

[0017] The technical scheme of the embodiment of the present invention includes: determining a characteristic index according to the current sampling data sequence of the high-voltage relay protection device; determining the associated knowledge graph of the high-voltage relay protection device according to the static data of the high-voltage relay protection device, the current sampling data sequence and the preset device model library and image library; determining the intermediate diagnosis conclusion according to the characteristic index and the diagnosis rule corresponding to the characteristic index in the associated knowledge graph; determining the current classification result of the characteristic index according to the distance between the characteristic index and each classification result, wherein each classification result is the result of clustering all states in the associated knowledge graph, and each classification result is used to characterize a working state of the high-voltage relay protection device; determining the fault detection result of the high-voltage relay protection device according to the intermediate diagnosis conclusion and the current classification result. It has the following technical effects: on the one hand, fault detection can be automatically realized based on the current sampling data sequence, thereby improving the efficiency of fault detection; on the other hand, in the fault detection process, the intermediate diagnosis conclusion based on the associated knowledge graph and the current classification result based on the state clustering are integrated, thereby realizing accurate perception of the fault and improving the accuracy of fault detection.

[0018] It should be understood that the contents described in this section are not intended to identify the key or important features of the embodiments of the present invention, nor are they intended to limit the scope of the present invention. Other features of the present invention will become easily understood through the following description. BRIEF DESCRIPTION OF THE DRAWINGS

[0019] In order to more clearly illustrate the technical solutions in the embodiments of the present invention, the following briefly introduces the drawings required for use in the description of the embodiments. Obviously, the drawings described below are only some embodiments of the present invention. For ordinary technicians in this field, other drawings can be obtained based on these drawings without creative work.

[0020] Figure 1 is a schematic diagram of an application scenario of the fault detection method provided by an embodiment of the present invention;

[0021] Figure 2 is a flow chart of a fault detection method provided by an embodiment of the present invention;

[0022] Figure 3 is a flow chart of another fault detection method provided by an embodiment of the present invention;

[0023] Figure 4 is a flow chart of another fault detection method provided by an embodiment of the present invention;

[0024] Figure 5 It is a structural schematic diagram of an electronic device for implementing the fault detection method according to an embodiment of the present invention. DETAILED DESCRIPTION

[0025] In order to enable those skilled in the art to better understand the scheme of the present invention, the technical scheme in the embodiments of the present invention will be clearly and completely described below in conjunction with the drawings in the embodiments of the present invention. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work should fall within the scope of protection of the present invention.

[0026] It should be noted that the terms "first", "second", etc. in the specification and claims of the present invention and the above-mentioned drawings are used to distinguish similar objects, and are not necessarily used to describe a specific order or sequence. It should be understood that the data used in this way can be interchangeable where appropriate, so that the embodiments of the present invention described herein can be implemented in an order other than those illustrated or described herein. In addition, the term "comprising" and any variation thereof are intended to cover non-exclusive inclusions, for example, a process, method, system, product or device comprising a series of steps or units is not necessarily limited to those steps or units clearly listed, but may include other steps or units that are not clearly listed or inherent to these processes, methods, products or devices. The acquisition, storage, use, processing, etc. of data in the embodiments of the present invention all comply with the relevant provisions of national laws and regulations.

[0027] High-voltage relay protection devices refer to automation devices used to monitor and protect major equipment such as transmission lines, transformers, and busbars in high-voltage power systems. When short circuits, grounding, overloads, and other faults occur in the power system, the high-voltage relay protection device can quickly cut off the fault current, isolate the faulty equipment, prevent the accident from expanding, and ensure the safe and stable operation of the power grid. High-voltage relay protection devices are generally composed of measurement units, logic units, control units, and communication units. Among them, the measurement unit collects the analog signal quantity on the primary side through current and voltage transformers and converts it into digital quantity. The logic unit processes and compares the digital quantity according to the set value and logical relationship of the protection to determine whether the protection action conditions are met. The control unit executes the trip command and controls the circuit breaker to open and close. The communication unit exchanges telesignaling and telemetry data and dispatching commands with the substation automation system. Common types of high-voltage relay protection include: distance protection, zero-sequence protection, differential protection, overcurrent protection, low-frequency and low-voltage protection, etc. Among them, the line protection of 110kV and above is mainly distance protection and zero-sequence protection, and the line protection of 110kV and above is mainly overcurrent protection.

[0028] Figure 1 FIG. 1 is a schematic diagram of an application scenario of the fault detection method provided by an embodiment of the present invention. Figure 1 As shown, the high-voltage relay protection device 21 collects original sampling data. The original sampling data here can be the state parameters of the equipment protected by the high-voltage relay protection device 21 and / or the state parameters of the high-voltage relay protection device 21. For example, a substation is equipped with a 110kV intelligent transformer protection device (i.e., a specific high-voltage relay protection device). The protection device is equipped with high-precision voltage transformers and current transformers, which can measure the three-phase voltage, three-phase current, frequency, temperature and other state parameters of the transformer in real time. The sampling unit of the protection device samples the voltage and current signals at a frequency of 4kHz, and obtains 96 sampling points in each sampling cycle. Each sampling point contains 7 measurement quantities, and the data length is 12 bytes, forming the original sampling data.

[0029] Considering the large volume of the original sampled data, in order to reduce the data transmission and storage overhead, the compression unit of the high-voltage relay protection device 21 uses a discrete wavelet transform algorithm to compress the original sampled data. Specifically, the Daubechies4 (db4) wavelet function can be selected to perform a 5-level wavelet decomposition on the original sampled data to obtain 1 low-frequency component and 5 high-frequency components. Then each component is quantized and encoded, and the quantization step is the standard deviation of the wavelet coefficient to obtain a discrete quantization coefficient. Finally, the quantization coefficient is run-length encoded and Huffman encoded to obtain the compressed sampled data. The length of the compressed data packet is 1 / 4 to 1 / 8 of the original data, which greatly reduces the amount of data. The high-voltage relay protection device 21 can transmit the compressed sampled data to the industrial switch uploaded to the substation through the Ethernet port, and then transmit it to the data center 221 of the operation and maintenance system 22 through the power dedicated network. After the original sampled data is compressed and transmitted, the bandwidth occupancy of data transmission can be reduced and the transmission efficiency can be improved.

[0030] The operation and maintenance system 22 represents an intelligent operation and maintenance management system for substation equipment, which can be a software system with a network (Web) architecture, including a data center 221, an equipment monitoring module 222, functional modules such as status assessment, auxiliary decision-making, intelligent inspection, and remote control, which can realize the full life cycle management of the primary and secondary equipment of the entire substation. The data center 221 is equipped with a high-performance server cluster and a distributed storage system, and a data processing and storage platform can be built using the Hadoop and HBase frameworks. The data center 221 decompresses the received compressed sampled data and pre-processes the decompressed sampled data. The data center 221 distributes the pre-processed sampled data to the equipment monitoring module 222 in the form of a real-time data stream through a message queue middleware based on a publishing mechanism.

[0031] Among them, the message queue middleware is a distributed software system based on message passing and asynchronous communication. It achieves decoupling and peak-shaving between applications by temporarily storing messages in queues, improving the concurrency, scalability and fault tolerance of the system. The message queue adopts a publish / subscribe communication model. The sender (producer) of the message only needs to publish the message to a specific topic or queue without caring about the receiver (consumer) of the message. The receiver of the message only needs to subscribe to the topic or queue of interest to receive messages from the sender in real time without knowing the source of the message. Common message queue middlewares include: Kafka, Rabbit Message Queue (MQ), Rocket MQ, Zero MQ, etc. Among them, Kafka adopts log-based sequential reading and writing, with the advantages of high throughput, low latency, and traceability, and is suitable for processing massive real-time data streams. In the operation and maintenance system 22, the data center 221 distributes the pre-processed sampled data in the form of real-time data stream to the equipment monitoring module 222 through message queue middleware such as Kafka. This publish / subscribe mechanism reduces the coupling between the data source and the data consumer, simplifies the logic of data distribution, and improves the real-time performance and reliability of the system. At the same time, the message queue can also buffer, cut peaks, and partition the data stream, thereby improving the system's concurrent processing capability and load balancing capability.

[0032] A possible data distribution process is as follows. Data center 221 first creates a topic named "Sample Data" in the Kafka cluster, and sets multiple (for example, 20) partitions and multiple (for example, 3) copies to carry the sample data stream of the high-voltage relay protection device. Then, data center 221 deploys a Kafka producer program, connects to the database storing the sample data, and queries the latest sample data. Whenever new data is written to the database, the producer program will immediately read the data, convert it into a preset format, such as JavaScript Object Notation (JSON), and publish it to the "Sample Data" topic in order of timestamp. When publishing, the producer program uses a round-robin algorithm to evenly distribute messages to each partition to avoid data skew. At the same time, the producer program sets the message life span, and messages that have not been consumed after the time limit will be automatically expired and deleted to prevent data accumulation.

[0033] The equipment monitoring module 222 is the core module of the operation and maintenance system 22, which can subscribe to the real-time sampling data distributed by the data center 221 and perform fault detection on the high-voltage relay protection device 21 according to the real-time sampling data.

[0034] A possible data subscription process is as follows. As a consumer of Kafka, the device monitoring module 222 subscribes to the data stream of the "Sample Data" topic. Considering the large data flow, the device monitoring module 222 uses a consumer group to start multiple (for example, 10) consumer instances. Each instance independently connects to the Kafka cluster and consumes data from different partitions in parallel. The consumer group ensures that each message will only be processed by one consumer instance, will not be consumed repeatedly, and the consumption progress between multiple instances is independent of each other and does not affect each other. After receiving the message, the consumer instance first checks the integrity and timing of the message. If out-of-order or missing data is found, a retransmission request is sent to the producer program. The verified message is parsed into a JSON object, and fields such as sampling timestamp, measurement item name, and sampling value are extracted. The consumer instance forwards the parsing result to the downstream data processing module. If the processing fails, the message is put back into Kafka and waits for the next retry.

[0035] The fault detection method provided in this embodiment can be run in an operation and maintenance system to achieve efficient and accurate fault detection.

[0036] The fault detection method of the present invention is further described in detail below through several specific embodiments.

[0037] Figure 2 is a flow chart of a fault detection method provided by an embodiment of the present invention. This embodiment can be applied to the scenario of fault detection of a high-voltage relay protection device. The method can be executed by a fault detection device, which can be implemented in the form of hardware and / or software. The fault detection device can be configured in an electronic device. For example, the electronic device can be Figure 1 The equipment monitoring module in the operation and maintenance system. In this embodiment, the electronic device is used as the equipment monitoring module for exemplary description. Figure 2 As shown, the method includes the following steps 201 to 205.

[0038] Step 201: Determine a characteristic index according to a current sampling data sequence of a high-voltage relay protection device.

[0039] The device monitoring module can subscribe to the real-time sampling data stream distributed by the data center. Optionally, the device monitoring module can obtain the current sampling data sequence through the message middleware. The current sampling data sequence in this embodiment includes sampling data at multiple sampling moments. The current sampling data sequence may include sampling data at the current sampling moment and sampling data at multiple historical sampling moments. Exemplarily, the number of historical sampling moments may be the 10 historical sampling moments closest to the current sampling moment.

[0040] In this embodiment, the equipment monitoring module can perform real-time calculations on the current sampling data sequence to obtain characteristic indicators reflecting the high-voltage relay protection device, such as three-phase voltage, three-phase current, active power, reactive power, apparent power, frequency and other characteristic indicators.

[0041] Step 202: Determine the associated knowledge graph of the high-voltage relay protection device based on the static data of the high-voltage relay protection device, the current sampling data sequence, and the preset device model library and image library.

[0042] In this embodiment, the device monitoring module can pre-build a device knowledge graph based on the ontology construction method. The device knowledge graph includes device type, device attribute, device topology relationship and device operation status.

[0043] Among them, the device topology relationship refers to the connection, composition and hierarchical relationship of devices in the system. It describes the physical and logical dependencies between devices and reflects the structural characteristics of the system. Common device topology relationships include the following relationships. Inclusion relationship: Indicates that a device is composed of multiple sub-devices or components, such as a transformer containing windings, cores and other components. Connection relationship: Indicates that there is a physical or signal connection between two devices, such as a busbar connecting multiple circuit breakers. Power supply relationship: Indicates that a device provides power to other devices, such as a generator supplying power to a transformer. Communication relationship: Indicates that there is data interaction between two devices, such as a trip signal between a protection device and a circuit breaker. Position relationship: Indicates the relative position of the device in space, such as upstream, downstream, left, right, etc.

[0044] A possible process of pre-building a device knowledge graph is as follows. The device monitoring module constructs a device knowledge graph for high-voltage relay protection devices based on the ontology construction method of the World Wide Web Ontology Language (OWL). First, the device monitoring module creates a new ontology project in the knowledge graph construction software, named "Prot Relay Onto", and imports the Institute of Electrical and Electronics Engineers (IEEE) standard vocabulary Common Format for Transient Data Exchange of Power Systems (COMTRADE) as the upper ontology. Then, the following core classes are defined in the Prot Relay Onto ontology: Protection Relay class: represents high-voltage relay protection devices and is the parent class of all protection device classes. Current Relay class: represents current protection, such as overcurrent protection, differential protection, etc., and is a subclass of Protection Relay. Voltage Relay class: represents voltage protection, such as undervoltage protection, overvoltage protection, etc., and is a subclass of Protection Relay. Distance Relay class: represents distance protection, such as impedance protection, directional protection, etc., and is a subclass of Protection Relay. Transformer Relay class: represents transformer protection, such as gas protection, pressure release protection, etc., and is a subclass of Protection Relay. Switch gear class: represents primary equipment, such as circuit breakers, disconnectors, etc., and is associated with the Protection Relay class. Measuring Point class: represents the measuring point of secondary equipment, such as CT, PT, etc., and is associated with the Protection Relay class. At the same time, the following core attributes are defined: has Type attribute: represents the type of equipment, such as the types of current protection, such as overcurrent, distance, etc. has Manufacturer attribute: represents the manufacturer of the equipment, and the value is a string. has Model attribute: represents the model specification of the equipment, and the value is a string or a numeric value. has Grade attribute: represents the voltage grade of the equipment, and the value is a numeric value in kV. has Set point attribute: represents the protection setting value, such as overcurrent setting value, delay time, etc., and the value is a string. connects To attribute: represents the topological connection relationship between devices, and the value is an instance of Switch gear. has Measurement attribute: represents the measurement input of the protection, and the value is an instance of Measuring Point. hasState attribute: indicates the operating status of the device. Its value is a string, such as "Normal", "Alarm" and "Trip".generates CBM attribute: represents the status monitoring data generated by the device, and its value is a string, such as "measurement point indication" and "waveform recording and broadcasting". Finally, create a device instance on the personal page of the knowledge graph construction software, fill in the attribute values, and form the basic framework of the device knowledge graph.

[0045] When the equipment monitoring module receives new sampled data from the Kafka message queue, it triggers the instance mapping program to associate the static data and sampled data of the high-voltage relay protection device with the pre-established equipment knowledge graph to obtain the instantiated knowledge graph. Then, based on the instantiated knowledge graph, the preset equipment model library and the image library, the complete associated knowledge graph of the high-voltage relay protection device is obtained.

[0046] In this embodiment, the associated knowledge graph of the high-voltage relay protection device can reflect the topological relationship, composition relationship, fault impact relationship, intrinsic characteristics and diagnostic rules of the high-voltage relay protection device.

[0047] Step 203: Determine an intermediate diagnosis conclusion based on the feature indicators and the diagnosis rules corresponding to the feature indicators in the associated knowledge graph.

[0048] The intermediate diagnosis conclusion in this embodiment is an intermediate conclusion of the state monitoring and abnormal diagnosis analysis of the high-voltage relay protection device, and is used to characterize whether the characteristic indicator satisfies the diagnosis rule corresponding to the characteristic indicator.

[0049] In this embodiment, the diagnosis rules corresponding to the characteristic indicators may be determined from the associated knowledge graph first.

[0050] Diagnostic rules are a knowledge representation method used to describe the correspondence between equipment status and fault type. It associates the characteristic indicators, working status, environmental conditions and other factors of the equipment with possible fault modes, causes, locations and other results to form a set of logical rules for diagnostic reasoning. Diagnostic rules are usually based on the summary of knowledge such as expert experience, historical cases, physical mechanisms, etc., and have clear causal logic and decision boundaries. Diagnostic rule representation methods include the following methods. Logic-based methods, such as propositional logic, first-order logic, description logic, etc., use logical relations and predicate formulas to describe diagnostic rules, such as "IF (feature A>threshold) AND (feature B<threshold) THEN (fault type C)". Rule-based methods, such as production rules, decision trees, decision tables, etc., use condition-action pairs or IF-THEN statements to describe diagnostic rules, such as "IF oil temperature is too high AND oil pressure is too low THEN lubrication system failure". Case-based methods, such as case-based reasoning (CBR), use historical fault case libraries to describe diagnostic rules, perform similarity retrieval and matching on new fault instances, and find the most similar case as the diagnostic result. Based on statistical methods, such as Bayesian networks and Markov logic networks, probabilistic graphical models are used to describe diagnostic rules, and the dependency between features and faults is characterized by conditional probability distribution, and probabilistic reasoning is performed to obtain diagnostic results.

[0051] In this embodiment, the diagnostic rules are expressed in the form of Horn clauses based on semantic reasoning, which combines the advantages of logical representation and semantic representation. It can simultaneously characterize the numerical relationship between device attributes and physical quantities, as well as the semantic association between concepts, forming a more comprehensive, flexible and explainable diagnostic knowledge base.

[0052] Horn clause is a special form of logical clause, which consists of a positive word (header) and zero or more negative words (body), and the words are connected by logical "and" (∧). The general form of Horn clause is: A:-B1, B2,..., Bn. Among them, A is a positive word, which represents the conclusion or goal; B1, B2,..., Bn are negative words, which represent the premise or condition; ":-" is an implication relationship, which means "if B1, B2,..., Bn are all true, then A is true". If there is no body, the Horn clause degenerates into a single positive word, which represents an unconditional fact or axiom. In this embodiment, the diagnostic rules are expressed in the form of Horn clauses, and each Horn clause is composed of multiple atomic predicates. The atomic predicates are used to describe the numerical relationship between device attributes and physical quantities, such as "temperature (x)>threshold (y)". Through the Horn clause form, the device diagnostic knowledge can be formalized in the form of logical rules, which is convenient for subsequent reasoning and matching.

[0053] Among them, the atomic predicate is a basic logical unit used to describe the properties or relationships of objects. It consists of a predicate symbol and a set of terms (constants, variables or functions), indicating whether a certain relationship or property between objects is established. Atomic predicates are the basic elements of first-order logic formulas, which are connected by logical connectives (such as ∧, ∨, → etc.) can form complex logical expressions by combining multiple atomic predicates to represent various complex knowledge and rules. In Horn clauses, atomic predicates are used to describe conditions and conclusions. Multiple conditional predicates are connected by logical "and" to form the body of the clause; the conclusion predicate is the head of the clause, indicating new attributes or relationships that can be deduced when the conditions are met.

[0054] In the field of equipment diagnosis, atomic predicates can be used to describe various characteristic attributes and physical quantities of equipment, such as voltage, current, temperature, pressure, etc., as well as the numerical relationships between them, such as greater than, less than, equal to, range, etc. Through these predicates, the quantitative information of the equipment status can be converted into symbolic form and matched and reasoned with the diagnosis rules. For example, the predicate "voltage(u)>220" means that the voltage value is greater than 220 volts; "temperature(t)in[60,80]" means that the temperature value is between 60 and 80 degrees Celsius. In this embodiment, the characteristic indicators extracted by the equipment module are represented as atomic predicates as the factual basis for diagnostic reasoning. At the same time, the conditions and conclusions in the diagnostic rules are also represented as atomic predicates to describe the logical relationship between equipment attributes and physical quantities. Through matching and replacement between predicates, the association and reasoning between equipment status information and diagnostic knowledge can be achieved.

[0055] A possible implementation process of step 203 is as follows: Through the unified algorithm, according to the characteristic index, the variables in the diagnosis rule corresponding to the characteristic index are matched and replaced to obtain the matched and replaced diagnosis rule; according to the matched and replaced diagnosis rule, the intermediate diagnosis conclusion is determined.

[0056] The unification algorithm is an algorithm used to determine whether two logical expressions can become equal through variable substitution. It is one of the core algorithms in the fields of automated reasoning, theorem proving, and logic programming. The basic idea of ​​the unification algorithm is to find the variables in the two expressions and replace them with appropriate terms so that the two expressions become equal after the substitution. The input of the unification algorithm is two logical expressions (such as atomic predicates, clauses, etc.), and the output is a most general substitution (Most General Unifier, MGU), that is, the most general variable substitution that can make the two expressions equal. If such a substitution does not exist, the two expressions are said to be ununifiable. The unification algorithm recursively compares the structure of the two expressions and updates the variable substitution until the expressions are completely matched or ununifiable.

[0057] In this embodiment, the unification algorithm is used to match the extracted feature indicators with the diagnostic rules. By taking the feature indicators as facts and unifying them with the atomic predicates in the rules, the rules that match the current device status can be found and the consistent replacement of variables can be obtained. This matching method based on logical expressions can effectively address the semantic gap between device status information and diagnostic knowledge, and realize flexible and generalized diagnostic reasoning.

[0058] The following is a specific example to illustrate the implementation process of step 203. For example, the diagnostic rule for determining whether the secondary side of the current transformer is broken can be expressed as: has Feature (CT, Secondary Open Circuit): - has Measurement (CT, Current), has Nominal Value (Current, I_n), less Than (Current, 0.05*I_n). The diagnostic rule consists of three atomic predicates, which respectively describe the current transformer attribute (has Measurement), the rated parameter (has Nominal Value) and the numerical relationship (less Than), which together express "When the current is less than 5% of the rated value, it is determined that the secondary side is broken". The equipment monitoring module takes the characteristic indicators as facts and matches them with the diagnostic rules. For example, at a certain moment, the measured A phase current I_a = 0.02*I_n, the B phase current I_b = 0.98*I_n, and the C phase current I_c = 1.05*I_n, where I_n = 5A. Then we can get the following facts: has Measurement(CT_A, 0.1).has Measurement(CT_B, 4.9).has Measurement(CT_C, 5.25).has Nominal Value(CT_A).has Nominal Value(CT_B).has Nominal Value(CT_C). The equipment monitoring module uses the unification algorithm to uniformly replace the variable Current in the diagnostic rule with 0.1, 4.9, and 5.25, and replaces I_n with 5, and then uses the forward link and reverse link reasoning algorithms for matching. For the A phase current, it can be deduced that: less Than(0.1, 0.05*5).hasFeature(CT_A, Secondary Open Circuit), that is, the intermediate diagnostic conclusion of "CT_A has a secondary side line break fault". For the B and C phase currents, the rule conditions cannot be met, indicating that they are normal.

[0059] In this embodiment, the diagnostic rules are expressed in the form of horn clauses based on semantic representation, and the interpretable analysis of the abnormal state of the equipment is achieved through the unified algorithm and link reasoning.

[0060] Step 204: Determine the current classification result of the feature indicator according to the distance between the feature indicator and each classification result.

[0061] Among them, each classification result is the result of clustering all states in the associated knowledge graph. Each classification result is used to characterize a working state of the high-voltage relay protection device.

[0062] Optionally, before step 204, this embodiment may further include the following steps: extracting all state-related weighted triples from the associated knowledge graph; inputting each weighted triple into the conversion model to obtain an embedding vector of the weighted triple output by the conversion model; clustering the embedding vectors of all state-related weighted triples to obtain various classification results. Among them, each classification result is used to characterize at least two of the following working states: normal state, overload state, underload state, abnormal state and fault state.

[0063] The conversion model in this embodiment is a pre-trained model that can convert weighted triples of states into embedded vectors.

[0064] A possible training process is as follows: obtain weighted triplets for training; divide the weighted triplets for training into a training set, a validation set, and a test set, use an initial conversion model to perform representation learning on the weighted triplets, train parameters through a stochastic gradient descent algorithm, and obtain a conversion model.

[0065] Each weighted triplet is input into the conversion model to obtain the embedding vector of the weighted triplet output by the conversion model.

[0066] Another possible training process is as follows: the device monitoring module extracts all state-related weighted triplets from the associated knowledge graph and divides them into training set, validation set and test set. Optionally, the ratio is 8:1:1. Then a conversion model (for example, translation model TransE, TransR, etc.) is used to represent the weighted triplets, and the embedding vector is trained using a stochastic gradient descent algorithm by minimizing the loss of the energy function. The energy function is defined as: E(h, r, t) = ||h+rt||_2. Where h, r, and t are the embedding vectors of the head entity, relationship, and tail entity in the triple (h, r, t), respectively. After 10 training iterations (epochs), the loss converges to below 0.01, and a k-dimensional embedding vector for each entity and relationship is obtained, that is, an embedding vector for each weighted triple is obtained.

[0067] Among them, Stochastic Gradient Descent (SGD) is an iterative algorithm for optimizing the objective function, which is particularly suitable for large-scale data sets and online learning scenarios. The basic idea is that in each iteration, a sample (or a small batch of samples) is randomly selected, its gradient is calculated, and the model parameters are updated in the opposite direction of the gradient until the convergence condition is reached. In this embodiment, the SGD algorithm is used to optimize the objective function of the conversion model, and the embedding vector representation of entities and relationships is learned by minimizing the similarity difference between positive and negative triples in the knowledge graph. The SGD algorithm can efficiently process large-scale knowledge graphs, achieve good convergence effects within a limited number of iterations, and obtain high-quality embedding vectors.

[0068] Embedding vector is a representation learning method that maps discrete variables to continuous low-dimensional vector space. The core idea is to learn an embedding function to map the original high-dimensional, sparse, discrete feature space to a low-dimensional, dense, continuous vector space while retaining the semantic structure and association pattern of the original features. The embedded vector is called an embedding vector, which is a distributed representation of the original feature. Embedding learning methods include: matrix decomposition-based methods, which decompose high-dimensional sparse matrices into low-dimensional dense matrices, and use the decomposed matrices to represent the embedding vectors of the original features. Neural network-based methods, which learn the embedding vectors of features by constructing shallow neural networks, taking the original features as input, and predicting their context or co-occurrence features. Random walk-based methods, which generate node sequences by random walk sampling on the graph, and use the sequences as context information to learn the embedding vectors of nodes. Knowledge graph-based methods, such as TransE and TransR, characterize the semantic compatibility of triples (h, r, t) by defining different scoring functions, and use embedding vector operations to fit the scoring function, thereby learning the embedding vectors of entities and relations.

[0069] Optionally, the equipment monitoring module uses a Gaussian mixture model (GMM) to perform unsupervised clustering on the embedding vectors of all state-related weighted triplets to obtain a clustering result reflecting the state mode of the high-voltage relay protection device. The clustering result is also called a classification result, and each classification result corresponds to a working state of the high-voltage relay protection device.

[0070] The Gaussian mixture model is a commonly used probability density estimation and clustering model used to describe and characterize the distribution characteristics of data. It assumes that the data is composed of multiple Gaussian distributions (normal distributions), each of which represents a subgroup or component. Through the weighted mixture of these subdistributions, it is possible to approximate any complex true distribution. The probability density function of GMM is expressed as: Where x is the d-dimensional observation data, K is the number of mixed components, and π k is the prior probability or mixing coefficient of the kth mixture component. is the Gaussian probability density function of the kth mixture component, μ k is the mean vector of the kth mixture component, and ∑k is the covariance matrix of the kth mixture component. The parameter estimation of GMM usually adopts the expectation maximization (EM) algorithm, alternating the E step and the M step until convergence. Among them, the E step is: calculate the posterior probability (responsibilities) of each data point belonging to each component based on the current parameters. The M step is: update the prior probability, mean and covariance of each component based on the posterior probability. The clustering process of GMM divides each data point into the mixture component with the largest posterior probability by making a hard decision (taking the maximum value) or a soft decision (weighted average) on the posterior probability of the data point. Based on the above process, multiple classification results can be obtained, each classification result corresponds to a Gaussian distribution, described by the mean vector and the covariance matrix.

[0071] The optimal clustering number K is selected according to the Bayesian Information Criterion (BIC). In this embodiment, K=5 can be set, and five typical working states of the high-voltage relay protection device are obtained: normal, overload, underload, abnormal and fault.

[0072] In step 204, the current classification result corresponding to the feature indicator is determined according to the distance between the feature indicator and each classification result.

[0073] Optionally, a possible implementation process of step 204 is: determining an embedding vector of the feature indicator; respectively calculating the Mahalanobis distance between the embedding vector of the feature indicator and the center of each classification result; and determining the current classification result of the feature indicator according to each Mahalanobis distance.

[0074] Mahalanobis distance is a measure of the relative distance from a sample point to the center of the distribution, taking into account the correlation and scale differences between the features. Compared with the Euclidean distance, the Mahalanobis distance not only takes into account the value of each feature, but also the covariance matrix between the features, so it can more accurately measure the similarity and abnormality of sample points in multidimensional space. In this embodiment, the Mahalanobis distance is used to determine the similarity between the current device state represented by the feature index and the historical state pattern.

[0075] When determining the embedding vector of the feature indicator, the feature indicator can be mapped to the embedding space, for example, the embedding space of the conversion model, to obtain a k-dimensional embedding vector. The Mahalanobis distance from the embedding vector to the center of each classification result is calculated respectively. The embedding vector is divided into the classification result with the smallest Mahalanobis distance to obtain the current classification result of the feature indicator. This indicates that the current device state is most similar to the current classification result.

[0076] Furthermore, if the minimum Mahalanobis distance between the embedding vector of the feature index and the center of each classification result exceeds a certain threshold, it indicates that the current state of the high-voltage relay protection device is an abnormal state.

[0077] According to the classification results to which the embedding vector corresponding to the feature index is divided and the size of the Mahalanobis distance, it is judged whether the high-voltage relay protection device is abnormal, as well as the type, cause and location of the abnormality.

[0078] By introducing the Mahalanobis distance, this embodiment can comprehensively consider the correlation and scale differences between multiple feature indicators, more accurately judge the similarity between the device status and the historical pattern, and improve the accuracy and reliability of abnormal diagnosis. At the same time, the threshold of the Mahalanobis distance can be adaptively adjusted according to historical data to improve the sensitivity and adaptability of fault detection.

[0079] In this embodiment, an unsupervised learning method combining a transformation model and a Gaussian mixture model is used to cluster and detect anomalies on the knowledge graph, which can automatically mine patterns of device status and overcome the limitations of traditional anomaly detection methods based on thresholds or statistics.

[0080] Step 205: Determine the fault detection result of the high-voltage relay protection device according to the intermediate diagnosis conclusion and the current classification result.

[0081] In this embodiment, the equipment monitoring module integrates the intermediate diagnosis conclusions and the current classification results to determine the status monitoring results and fault detection results used to characterize the high-voltage relay protection device, thereby realizing real-time evaluation, fault warning, cause analysis and auxiliary decision-making of the high-voltage relay protection device.

[0082] Optionally, the fault detection result includes: whether the high-voltage relay protection device is abnormal, the abnormality type, the abnormality cause and the abnormality location, etc.

[0083] A possible fault detection result is shown below. Taking 110kV distance protection as an example, the fault detection result is as follows: Device name: Distance Relay_1, Health score: 85, Real-time working condition: Normal, Fault alarm: CT_A, Secondary side disconnection, Fault location: Secondary side connection line of CT_A, Possible cause: Loose terminal or damaged cable, Repair suggestion: Check the connection line of CT_A secondary side, tighten the terminal or replace the cable, Optical fiber communication is interrupted, Fault location: Optical fiber transceiver module of Slot_3, Possible cause: Poor contact of optical fiber plug or damaged module, Repair suggestion: Reseat the optical fiber plug, or replace the optical fiber transceiver module of Slot_3.

[0084] In this embodiment, the integration of diagnostic reasoning and state clustering can achieve accurate perception of equipment status, rapid location of anomalies, and active learning of unknown working conditions, greatly improving the efficiency and intelligence level of abnormal diagnosis.

[0085] The fault detection method provided in this embodiment deeply integrates the massive real-time sampling data in the high-voltage relay protection device with the business process of the operation and maintenance system through data sharing and semantic association, thus realizing the release of data value and value-added application. At the same time, the abnormal diagnosis method based on the associated knowledge graph overcomes the shortcomings of the traditional diagnosis method based on experience and thresholds, making the judgment of the equipment status more objective and accurate, and the analysis of the abnormal cause more comprehensive and in-depth, which can reduce the false alarm and missed alarm rate, improve the reliability of the diagnosis results, and reduce unnecessary downtime and maintenance time and cost.

[0086] When performing equipment status monitoring and abnormal diagnosis, this fault detection method extracts feature indicators from the current sampling data sequence, matches and logically infers with the diagnostic rules in the associated knowledge graph, and simultaneously performs representation learning on the associated knowledge graph, and realizes the classification of equipment working status through clustering of embedded vectors, thus realizing the organic integration of data-driven feature analysis and knowledge-driven rule reasoning, thereby improving the accuracy of fault detection.

[0087] In this embodiment, the operation and maintenance system can be used as a node in the blockchain network. Specifically, the device monitoring module or the data center can be used as a node in the blockchain network.

[0088] When the equipment monitoring module is a node in a blockchain network, the method provided in this embodiment further includes the following steps: determining shared data to be uploaded to the blockchain network according to a preset sharing strategy, wherein the shared data includes at least one of the following: a target sampling data sequence in the current sampling data sequence to be uploaded to the blockchain network and / or a target detection result in the fault detection result to be uploaded to the blockchain network; encrypting the shared data using an asymmetric encryption algorithm to obtain encrypted shared data; and uploading the encrypted shared data to the blockchain network.

[0089] When the data center is a node in the blockchain network, the device monitoring module can generate the fault detection result into a detection report in a preset format and send it to the data center. The data center can determine the shared data in a similar manner to the device monitoring module. The shared data is encrypted using an asymmetric encryption algorithm to obtain the encrypted shared data, and the encrypted shared data is uploaded to the blockchain network.

[0090] When uploading the encrypted shared data to the blockchain network, it can be submitted to the blockchain network in the form of a transaction, waiting for consensus confirmation from other maintenance nodes.

[0091] Among them, the asymmetric encryption algorithm is also called the public key cryptographic algorithm, which is an algorithm that uses different encryption keys and decryption keys to achieve the encryption and decryption process. Unlike the traditional symmetric encryption algorithm, the encryption key (i.e., public key) of the asymmetric encryption algorithm can be made public, while the decryption key (i.e., private key) needs to be kept private. In this embodiment, the data center or the device monitoring module uses an asymmetric encryption algorithm to encrypt the shared data uploaded to the blockchain network. Other nodes in the blockchain network can decrypt the ciphertext with the public key, but cannot forge the data encrypted by the private key. This encryption mechanism not only ensures data sharing, but also provides digital signatures and anti-tampering capabilities, which meets the security requirements of the blockchain. In addition, asymmetric encryption can also be combined with a digital certificate system to achieve node identity authentication and access control, further enhancing the credibility of the blockchain.

[0092] Furthermore, the operation and maintenance system in this embodiment may also include a visualization module. The equipment monitoring module may send the fault detection result to the visualization module. The visualization module presents the fault detection result to the user in a visual manner. Exemplarily, the visualization module may be an augmented reality (AR) visualization model.

[0093] The fault detection method provided in this embodiment includes: determining a characteristic index according to the current sampling data sequence of the high-voltage relay protection device; determining the associated knowledge graph of the high-voltage relay protection device according to the static data of the high-voltage relay protection device, the current sampling data sequence, and the preset device model library and image library; determining an intermediate diagnosis conclusion according to the characteristic index and the diagnosis rule corresponding to the characteristic index in the associated knowledge graph; determining the current classification result of the characteristic index according to the distance between the characteristic index and each classification result, wherein each classification result is the result of clustering all states in the associated knowledge graph, and each classification result is used to characterize a working state of the high-voltage relay protection device; determining the fault detection result of the high-voltage relay protection device according to the intermediate diagnosis conclusion and the current classification result. It has the following technical effects: on the one hand, fault detection can be automatically realized based on the current sampling data sequence, thereby improving the efficiency of fault detection; on the other hand, in the fault detection process, the intermediate diagnosis conclusion based on the associated knowledge graph and the current classification result based on the state clustering are integrated, thereby realizing accurate perception of the fault and improving the accuracy of fault detection.

[0094] Figure 3 is a flow chart of another fault detection method provided by an embodiment of the present invention. Figure 2 Based on the illustrated embodiment and various optional implementations, the process of determining the associated knowledge graph is described in detail. For ease of description, this embodiment only shows the steps of determining the associated knowledge graph of the high-voltage relay protection device based on the static data of the high-voltage relay protection device, the current sampling data sequence, and the preset device model library and image library. Figure 3 As shown, in the fault detection method provided in this embodiment, the process of determining the associated knowledge graph includes the following steps 301 and 302.

[0095] Step 301: Determine an instantiated knowledge graph of the high-voltage relay protection device based on the static data of the high-voltage relay protection device and the current sampling data sequence.

[0096] In this embodiment, the equipment monitoring module obtains static data of the high-voltage relay protection device from the operation and maintenance system. The static information includes the nameplate parameters, factory configuration and structural image of the high-voltage relay protection device. The equipment monitoring module determines the instantiation knowledge graph of the high-voltage relay protection device based on the static data and the current sampling data sequence.

[0097] Optionally, the equipment monitoring module can access the equipment asset management system through the web service interface to obtain the static data of the high-voltage relay protection device. The equipment monitoring module first initiates an authorization request to the authentication service to obtain an access token. Then, it sends a request to the equipment asset management system with the token, passes in the equipment identifier as a query parameter, and reads the nameplate parameters, factory information, computer-aided design (CAD) drawings, historical maintenance records, test reports, and operation tickets of the equipment, etc. Structured data and unstructured data. Among them, the nameplate parameters include equipment model, capacity, voltage level, commissioning date, etc., which are returned in a preset format; factory information includes manufacturer, factory serial number, quality inspection report, etc., which are returned in the form of a preset file; CAD drawings include primary main wiring diagram, secondary relay schematic diagram, terminal layout diagram, etc., which are returned in Auto CAD DWG format; historical maintenance records include maintenance time, maintenance items, defect description, processing results, etc., which are returned in a table form.

[0098] After receiving the data fed back by the equipment asset management system, the equipment monitoring module parses and extracts the data in different formats to obtain static data. For example, the Python application program interface of Auto CAD is called to read the primitive information. The extracted results are associated with the equipment identification and stored in the database for storing equipment identification and static data. Indexes such as timestamp, equipment identification, and data type are created to facilitate subsequent query retrieval.

[0099] Optionally, a possible implementation process of step 301 includes the following steps 3011 and 3012.

[0100] Step 3011: Convert static data and current sampled data sequence into device instances.

[0101] In this embodiment, the device instance may be determined according to static data and a current sampling data sequence.

[0102] A device instance refers to a device node in the knowledge graph, which represents a specific device in reality. Each device instance has its own unique identifier, such as device number, asset number, etc. The device instance carries various attribute information of the device, including: Static attributes: such as device type, nameplate parameters, factory configuration, structural image, etc. These attributes generally do not change during the life cycle of the device. Dynamic attributes: such as the operating status of the equipment, measurement point indication, maintenance record, etc. These attributes change continuously as the equipment runs. Device instances are linked to other entities and concepts through attributes, forming an information aggregation point in the knowledge graph. Through device instances, you can quickly retrieve and browse various information about the device, discover the association rules between devices, and tap the potential value of the device.

[0103] The measuring point indication refers to the real-time value of each monitored quantity during the operation of the equipment. It reflects the current state and working condition of the equipment and is an important basis for equipment operation and maintenance. Common measuring point indications include: electrical quantities: such as voltage, current, power, power factor, etc. Physical quantities: such as temperature, pressure, flow, speed, vibration, etc. Switching quantities: such as the switch state of the equipment, protection action signals, etc. Alarm quantities: such as equipment fault alarms, over-limit alarms, etc. The current sampling data sequence in this embodiment is the measuring point indication.

[0104] Step 3012: Map the device instance to the pre-built device knowledge graph according to the ontology model to obtain an instantiated knowledge graph.

[0105] After obtaining the device instances, by defining the semantic relationship between the device instances, the device instances are mapped to the constructed device knowledge graph according to the ontology model to obtain the instantiated knowledge graph.

[0106] Among them, the ontology schema is the top-level design of the knowledge graph, which defines the core concepts, relationships and constraints of the graph. The ontology schema provides a consistent semantic foundation for the knowledge graph and guides the organization, representation and retrieval of knowledge. The ontology schema of the equipment knowledge graph usually contains the following contents: Class: represents the core concepts in the equipment field, such as equipment, components, measurement points, alarms, etc. Each class has its semantic attributes and hierarchical relationships. Relation: represents the semantic connection between classes, such as the topological relationship, composition relationship, fault impact relationship, etc. of the equipment. Relations have directionality and cardinality constraints. Attribute: represents the intrinsic characteristics of the class, such as the voltage level, rated capacity, manufacturer, etc. of the equipment. Attributes have their value types and constraints. Instance: represents the specific object of the class, such as a transformer, an alarm event, etc. The instance carries the attribute value of the class. Axiom: represents the logical rules and reasoning mechanism of the ontology, such as the fault propagation rules of the equipment, the association rules of the alarm, etc. The ontology model uses ontology languages ​​such as WebOntology Language (OWL) for formal representation to support knowledge sharing, reuse and automatic reasoning. In the process of building the device knowledge graph, the ontology model plays the role of a "type system" and standardizes the structure and semantics of the graph. Tasks such as data mapping, consistency verification, and logical reasoning are all based on the ontology model.

[0107] A specific process of obtaining an instantiated knowledge graph is as follows. When the equipment monitoring module receives new sampling data from the Kafka message queue, the instance mapping program is triggered to dynamically associate the sampling data with the equipment knowledge graph. The mapping program first extracts the equipment identification, sampling time, measurement point name, sampling value and other fields in the message. Then, with the equipment identification as the primary key, the database storing the equipment identification and static data is queried to obtain the static data of the equipment. According to the static data, the corresponding equipment class is searched in the Prot Relay Onto (i.e., the ontology project created in step 202) ontology. If found, a new equipment instance is created, otherwise the existing instance is reused. Next, the program traverses each measurement point in the current sampling data sequence and searches for the corresponding Measuring Point instance in the ontology; if found, the has State attribute of the instance is updated to convert the sampling data into the running state; if not found, a new Measuring Point instance is created and the has Measurement relationship is added. For the connection relationship, the program queries the equipment asset management system, obtains the primary equipment identification of the equipment, searches for the corresponding Switch gear instance in the ontology, and establishes a topological connection relationship.

[0108] After instance mapping, the static Prot Relay Onto ontology is associated with the dynamic sampled data stream to form a real-time knowledge graph that reflects the current device status. For example, for a 110kV distance protection device, its knowledge graph fragment is as follows: Distance Relay_1rdf: type: Distance Relay; has Type "Impedance Protection"; has Manufacturer "NR Electric"; has Model "PCS-931"; has Grade "110"^^xsd: int; has Setpoint "Z1 = 0.2, tZ1 = 0.1"; connects To: Switch gear_1; has Measurement: CT_1, CT_2, PT_1; has State "Normal"; CT_1rdf: type: Measuring Point; has State "Normal"; generates CBM "0.8"^^xsd: float.

[0109] Step 302: Determine the associated knowledge graph of the high-voltage relay protection device based on the instantiated knowledge graph, the preset device model library and the image library.

[0110] Optionally, the equipment monitoring module obtains model data of the high-voltage relay protection device from the equipment model library, obtains graphic image data of the high-voltage relay protection device from the image library, and uses the cross-modal attention mechanism to establish an association between model parameters and equipment instance attributes, and establishes an association between image features and equipment instance attributes in a common semantic space, and adds the above association to the instantiated knowledge graph to obtain a complete associated knowledge graph.

[0111] Optionally, step 302 may be implemented through steps 3021 to 30215 as follows.

[0112] Step 3021: Obtain the physical model and mechanism model corresponding to the device instance from the device model library.

[0113] Furthermore, the operation and maintenance system in this embodiment also includes a device model library and an image library.

[0114] In step 3021, the device monitoring module obtains the physical model and mechanism model corresponding to the device instance from the device model library. The physical model is constructed based on CAD drawings or 3D scanning data, representing the geometric structure and spatial layout of the high-voltage relay protection device. The mechanism model is constructed based on physical equations or simulation algorithms, representing the working mechanism and dynamic behavior of the high-voltage relay protection device.

[0115] A specific implementation process of obtaining the physical model is as follows. The equipment monitoring module accesses the physical modeling library, which is based on the Auto CAD platform and stores the CAD three-dimensional model of the high-voltage relay protection device. The equipment monitoring module obtains the physical model corresponding to the equipment instance. Taking the 110kV distance protection as an example, its physical model includes the following components. Cabinet: 2000mm long, 800mm wide, 2200mm high, made of aluminum alloy, with protection unit, measurement and control unit, power supply unit, communication unit, etc. installed inside. Panel: Located on the front of the cabinet, 600mm long, 800mm wide, made of acrylonitrile-butadiene-styrene copolymer (ABS plastic), inlaid with liquid crystal display (Liquid Crystal Display, LCD), alarm indicator light, control button, etc. Plug-in: Located at the back of the cabinet, it adopts a 19-inch standard structure with a total of 12 slots. Each slot is equipped with a functional board, such as a digital signal processing (DSP) processing board, an analog to digital (AD) sampling board, a switch input and output board, etc. Terminal block: Located at the top and bottom of the cabinet, it adopts imported copper busbars, with 48 terminals in each row, which are used to access current, voltage, and switch signals. The equipment monitoring module parses the CAD model into triangular mesh data in stereolithography (STL) format, extracts physical parameters such as size, material, position, etc., maps them to the physical component class of the Prot Relay Onto body, and creates a physical model instance.

[0116] A specific implementation process of obtaining the mechanism model is as follows. The equipment monitoring module accesses the simulation modeling library, which is based on the MATLAB / Simulink platform and stores the mechanism model of the high-voltage relay protection device. The equipment monitoring module obtains the mechanism model corresponding to the equipment instance. Taking the 110kV distance protection as an example, its mechanism model includes the following modules. Analog input module: receives three-phase current and three-phase voltage analog quantities, performs AD sampling, removes DC components, and performs full-wave rectification. Distance measurement module: calculates the apparent impedance from the fault point to the protection installation, and determines the fault area according to the distance constant. Direction determination module: calculates the phase difference between the fault current and the reference voltage to determine whether the fault direction is correct. Zero-sequence compensation module: calculates the zero-sequence component of the residual voltage and residual current, and corrects the grounding distance. Reliability criterion module: calculates reliability criteria such as fault duration, phase-to-phase / phase-to-ground fault type, and reclosing dead zone. Tripping logic module: outputs a tripping command based on the start signal and reliability criterion of each measurement module after a certain logical combination. The equipment monitoring module converts the Simulink model into an Extensible Markup Language (XML) file, extracts input and output parameters, function equations, logical conditions, etc., maps them to the functional algorithm class of the Prot Relay Onto ontology, and creates a mechanism model instance.

[0117] Step 3022: Obtain the device image and component image corresponding to the device instance from the image library.

[0118] The equipment monitoring module obtains equipment images and component images corresponding to equipment instances from the image library. The image library uses deep learning-based image recognition and segmentation algorithms to extract image data of equipment and components from equipment drawings, photos, and videos, and classify and annotate them.

[0119] A specific implementation process of obtaining the device image and component image corresponding to the device instance is as follows. The device monitoring module accesses the image library, which uses a deep learning algorithm based on convolutional neural networks to automatically extract the device image and component image corresponding to the device instance from the device drawings, photos and videos. Taking the 110kV distance protection as an example, its image data includes the following contents. Equipment image: color photos of the front, side and back, with a resolution of 1920*1080. Component close-up image: close-ups of LCD panels, button indicators, plug-in boards, terminal blocks, etc., with a resolution of 1280*720. Internal structure image: the internal structure diagram of the cabinet obtained by industrial computed tomography (CT), with a resolution of 2048*2048. Principle wiring image: the primary main wiring diagram and secondary relay logic diagram extracted from the CAD drawing library, with a resolution of 4096*4096.

[0120] Optionally, the equipment monitoring module may pre-process the image data, including image enhancement, noise removal, background segmentation, etc., to improve image quality.

[0121] Step 3023: Match the device instance attributes in the instantiated knowledge graph with the physical model parameters in the physical model to obtain a first semantic correspondence between the device instance attributes and the physical model parameters.

[0122] Optionally, the device monitoring module uses semantic mapping rules based on ontology dictionary matching to match device instance attributes in the instantiated knowledge graph with physical model parameters in the physical model, and establishes a first semantic correspondence between device instance attributes and physical model parameters by querying synonyms and hyponyms defined in the ontology dictionary. That is, through semantic mapping rules, the physical model parameters in the physical model are associated with the device instance attributes in the knowledge graph and integrated into the instantiated knowledge graph.

[0123] Step 3024: Match the device instance attributes in the instantiated knowledge graph with the mechanism model parameters in the mechanism model to obtain a second semantic correspondence between the device instance attributes and the mechanism model parameters.

[0124] Optionally, the device monitoring module uses semantic mapping rules based on ontology dictionary matching to match the device instance attributes in the instantiated knowledge graph with the mechanism model parameters in the mechanism model, and establishes a second semantic correspondence between the device instance attributes and the mechanism model parameters by querying the synonyms and hyponyms defined in the ontology dictionary. That is, through the semantic mapping rules, the mechanism model parameters in the mechanism model are associated with the device instance attributes in the knowledge graph and integrated into the instantiated knowledge graph.

[0125] Among them, the semantic mapping rule based on ontology dictionary matching is a method for concept matching and conversion using ontology dictionary (OntologyLexicon). Its basic idea is to establish the correspondence between ontology concepts and natural language vocabulary by pre-building domain ontology dictionary, and then use the semantic similarity of vocabulary to realize concept mapping between different data sources. In this embodiment, device instance attributes and model parameters (physical model parameters or mechanism model parameters) come from different data sources, and their naming and semantics may be different. In order to establish the semantic correspondence between them, it is necessary to introduce a shared semantic layer, namely, device domain ontology dictionary. The ontology dictionary defines semantic information such as standard concepts, synonyms, hyponyms and hyponyms in the device domain. For example: standard concepts: voltage, current, power, temperature, etc. Synonyms: voltage (voltage, U), current (current, I), temperature (temperature, temp), etc. Hyponyms: sensor includes temperature sensor, pressure sensor, etc.

[0126] Based on the device domain ontology dictionary, semantic mapping rules can be formulated to map device instance attributes to standard concepts, and then map standard concepts to model parameters. The specific steps are as follows: perform string matching on device instance attributes and model parameters with the vocabulary in the ontology dictionary to obtain candidate standard concepts. If the match fails, use the synonyms in the ontology dictionary for extended matching. For example, match "U" to "voltage". If the match still fails, use the hyponyms and hyponyms in the ontology dictionary for generalization or specialization matching. For example, generalize "pressure transmitter" to "sensor". According to the matched standard concepts, establish the mapping relationship between device instance attributes and model parameters. For example, map the device instance attribute "rated voltage" to the model parameter "nominal_voltage". For the case where the mapping relationship is uncertain or ambiguous, confidence, priority and other metrics can be introduced, or domain experts can perform manual annotation and verification. The semantic mapping rules based on ontology dictionary matching make full use of the semantic knowledge in the ontology, which can effectively solve problems such as inconsistent naming and semantic heterogeneity, and realize the automatic association between device instances and models. At the same time, the ontology dictionary and mapping rules are knowledge assets that can be reused and expanded. With the accumulation and iteration of knowledge, the accuracy and coverage of mapping will continue to improve.

[0127] In step 3023 and step 3024, in order to seamlessly integrate the physical model and the mechanism model into the knowledge graph, the equipment monitoring module predefines a set of semantic mapping rules for matching model parameters and ontology attributes. The rules are based on the ProtRelay Onto ontology dictionary, which uses the Simple Knowledge Organization System (SKOS) standard to define the preferred labels, synonyms, and hyponyms of common concepts in the power field. For example, the "length" parameter label in the physical model is "length", and there are synonyms such as "length", "long", and "L" in the ontology dictionary; in the ontology attribute, "length" may be marked as "has Length". Therefore, the semantic mapping rules automatically establish an equivalent relationship between the "length" parameter and the "has Length" attribute. Similarly, the "fault direction" parameter in the mechanism model can be matched with synonyms and hyponyms such as "forward", "reverse", "forward", and "reverse" in the ontology dictionary, and mapped to the ontology attribute "has Fault Direction". Finally, the physical model instance is associated with the corresponding node of the instantiated knowledge graph by matching the "has Physical Model" attribute, and the mechanism model instance is associated with the corresponding node of the instantiated knowledge graph by matching the "has Principal Model" attribute. In other words, it is equivalent to establishing the first semantic correspondence between the device instance attribute and the physical model parameter, and the second semantic correspondence between the device instance attribute and the mechanism model parameter.

[0128] Continuing with the example of 110kV distance protection, the fused knowledge graph fragment is as follows: Di stance Relay_1rdf: type: Di stance Relay; has Type "Impedance Protection"; has Manufacturer "NR Electric"; has Model "PCS-931"; has Grade "110"^^xsd: int; has Set point "Z1=0.2,tZ1=0.1"; connects To: Switch gear_1; has Measurement: CT_1,CT_2,PT_1; has State "Normal"; has Physical Model: Cabinet_1,Panel_1,Slot_1; has Principal Model: Analog Input_1,Distance Measure_1,Direction Judge_1。 :Cabinet_1rdf: type: Cabinet; has Length "2000" ^^xsd: int; has Width "800" ^^xsd: int; hasHeight "2200" ^^xsd: int; has Material "Aluminium"; Distance Measure_1rdf: type: Distance Measure; has Input "Ia, Ib, Ic, Ua, Ub, Uc"; has Output "Z"; has Equation "Z=U / I".

[0129] Step 3025: Map the device instance attributes in the instantiated knowledge graph to a low-dimensional semantic space through word embedding to obtain a distributed vector representation of the device instance attributes.

[0130] In this embodiment, the device monitoring module can adopt a transformer-based cross-modal attention mechanism to map the device instance attributes in the instantiated knowledge graph to a low-dimensional semantic space through word embedding to obtain a distributed vector representation of the device instance attributes.

[0131] Among them, the cross-modal attention mechanism is a method that uses the attention mechanism to achieve association and fusion between different modal data. The basic idea is to adaptively align and aggregate cross-modal information by learning the attention weights between different modal data, so as to obtain a more comprehensive and accurate semantic representation. Transformer is a deep learning model based on the self-attention mechanism, which was originally applied in the field of natural language processing to process sequence data. The core of Transformer is the attention layer, which can learn the dependencies between different positions in the sequence and generate the attention weights of the positions. By stacking multiple attention layers and feedforward layers, Transformer can build a powerful context encoder to capture the long-distance dependencies and global information of the sequence. Extending the attention mechanism of Transformer to cross-modal scenarios, a cross-modal attention mechanism based on Transformer is formed. It regards different modal data (such as text, images) as different sequences, and learns the interaction and alignment between them through the attention layer.

[0132] Among them, low-dimensional semantic space is a representation method that maps high-dimensional and sparse original features to low-dimensional and dense semantic vector space through dimensionality reduction learning technology. Its basic idea is to compress the original features into a low-dimensional subspace by learning the intrinsic semantic structure and association pattern of the data while retaining its core semantic information. Low-dimensional semantic space plays a vital role in knowledge graphs and cross-modal learning. The original knowledge graphs and multimodal data are usually high-dimensional, sparse, and heterogeneous. Direct association and fusion in the original feature space often face challenges such as dimensionality disaster, data noise, and semantic gap. By mapping to low-dimensional semantic space, these problems can be effectively alleviated, and more robust, efficient, and interpretable semantic representation and association can be achieved. Low-dimensional semantic space learning methods include: matrix decomposition-based methods, such as singular value decomposition (SVD) and non-negative matrix factorization (NMF), which decompose high-dimensional matrices into the product of low-dimensional factor matrices and use factor matrices to represent the semantic vectors of data. Neural network-based methods, such as auto encoder and word embedding, map input data to low-dimensional latent vectors by building multi-layer neural networks, and use latent vectors to represent the semantic vectors of data. Graph embedding-based methods sample node sequences on the knowledge graph through random walks, and use word embedding models to learn low-dimensional vector representations of nodes.

[0133] A possible implementation process of step 3025 is as follows: convert the device instance attributes and their values ​​into text sequences, such as "has Type: Distance Relay", "has Manufacturer: NR Electric", etc.; then, use the Token Embedding layer of the Bidirectional Encoder Representations from Transformers (BERT) to map each word into a 768-dimensional word vector, and then use the Transformer Encoder layer to learn the contextual relationship and semantic interaction between the attributes, and finally use the pooling layer to extract the semantic representation of the entire attribute, which is also a 768-dimensional feature vector.

[0134] Step 3026: Perform feature extraction and semantic encoding on the device image, map the encoded device image to the same semantic space as the device instance attribute, and obtain a distributed vector representation of the device image.

[0135] Step 3027: Perform feature extraction and semantic encoding on the component image, map the encoded component image to the same semantic space as the device instance attribute, and obtain a distributed vector representation of the component image.

[0136] In this embodiment, the equipment monitoring module can perform feature extraction and semantic encoding on the equipment images and component images in the image library, and adopt a transformer-based cross-modal attention mechanism to map the encoded equipment images to the same semantic space as the equipment instance attributes to obtain a distributed vector representation of the equipment image, and map the encoded component images to the same semantic space as the equipment instance attributes to obtain a distributed vector representation of the component images.

[0137] Optionally, the equipment monitoring module can use a pre-trained convolutional neural network (CNN) model to perform feature extraction and semantic encoding on equipment images and component images. The encoded image features are mapped to a low-dimensional semantic space through a fully connected layer to obtain a distributed vector representation of the equipment image and a distributed vector representation of the component image. For example, a device photo is mapped to a 512-dimensional feature vector, and an internal structure diagram is mapped to a 1024-dimensional feature vector.

[0138] Step 3028: Determine a first semantic association matrix between the distributed vector representation of the device instance attributes and the distributed vector representation of the device image.

[0139] Step 3029: Determine a second semantic association matrix between the distributed vector representation of the device instance attributes and the distributed vector representation of the component image.

[0140] In this embodiment, the device module uses a transformer-based cross-modal attention mechanism in the semantic space to establish a semantic association matrix between the distributed vector representation of device instance attributes and the distributed vector representation of the image.

[0141] In this embodiment, the device monitoring module uses word embedding technology to map device instance attributes to a low-dimensional semantic space to obtain a distributed vector representation of the device instance attributes. This distributed vector representation can characterize the semantic similarity and association patterns between attributes, providing convenience for subsequent cross-modal associations. At the same time, the image data is also mapped to the same semantic space, and the similarity between attributes and images can be measured in a unified vector space to establish a cross-modal semantic association matrix.

[0142] In this embodiment, the device monitoring module uses a Transformer-based cross-modal attention mechanism to map device instance attributes (structured data) and device images (unstructured data) into a shared semantic space, and establishes a semantic association matrix between them through attention weights. This cross-modal association method can fully mine multi-source heterogeneous information in the device knowledge graph and improve the richness and accuracy of semantic representation.

[0143] Optionally, the device monitoring module uses a transformer-based cross-modal attention mechanism in the semantic space to establish a first semantic association matrix between the distributed vector representation of device instance attributes and the distributed vector representation of device images, and a second semantic association matrix between the distributed vector representation of device instance attributes and the distributed vector representation of component images.

[0144] A possible implementation process of step 3028 is as follows. The device monitoring module uses the distributed vector representation of the device image as the query (Query), and the distributed vector representation of the device instance attribute as the key (Key) and the value (Value). The correlation score between the device image and the device instance attribute is calculated through the attention layer, and then the attribute-perceived representation of the device image is obtained through weighted summation. Similarly, the distributed vector representation of the device instance attribute is used as the query, and the distributed vector representation of the device image is used as the key and the value to obtain the device image-perceived representation of the device instance attribute. Finally, the attribute-perceived representation of the device image and the device image-perceived representation are spliced ​​together, mapped to the common semantic space through a feedforward neural network, and the first semantic association matrix is ​​established.

[0145] A possible implementation process of step 3029 is as follows. The device monitoring module uses the distributed vector representation of the component image as Query, and the distributed vector representation of the device instance attribute as Key and Value. The correlation score between the component image and the device instance attribute is calculated through the attention layer, and then the attribute-perceived representation of the component image is obtained through weighted summation. Similarly, the distributed vector representation of the device instance attribute is represented as Query, and the distributed vector representation of the component image is represented as Key and Value to obtain the component image-perceived representation of the device instance attribute. Finally, the attribute-perceived representation of the component image and the component image-perceived representation are spliced ​​together, mapped to the common semantic space through a feedforward neural network, and a second semantic association matrix is ​​established.

[0146] Step 30210: Determine the embedding vector of the device instance attribute and the physical model parameter triplet according to the first semantic correspondence.

[0147] In this embodiment, the device monitoring module determines the device instance attributes and physical model parameters as real-valued vectors through one-hot encoding according to the first semantic correspondence obtained. Specifically, the device instance attributes are one-hot encoded to obtain the attribute vector, and the physical model parameters are one-hot encoded to obtain the physical model parameter vector. Then, by minimizing the Euclidean distance between the attribute vector and the physical model parameter vector, that is, minimizing the loss function of the Euclidean distance, the embedding vector of the device instance attribute and the physical model parameter triple is obtained.

[0148] Step 30211: Determine the embedding vector of the triplet of device instance attribute and mechanism model parameter according to the second semantic correspondence.

[0149] In this embodiment, the device monitoring module determines the device instance attributes and mechanism model parameters as real-valued vectors through one-hot encoding according to the obtained second semantic correspondence. Specifically, the device instance attributes are one-hot encoded to obtain the attribute vector, and the mechanism model parameters are one-hot encoded to obtain the mechanism model parameter vector. Then, by minimizing the Euclidean distance between the attribute vector and the mechanism model parameter vector, the embedding vector of the device instance attribute and the mechanism model parameter triple is obtained.

[0150] Among them, one-hot encoding is a method of converting discrete features into numerical vectors. For a discrete feature with n possible values, one-hot encoding converts it into an n-dimensional binary vector, in which only one element is 1, indicating the current feature value, and the rest of the elements are 0. For example, for the value set of the feature "color" {red, green, blue}, the one-hot encoding corresponding to red is [1, 0, 0], green corresponds to [0, 1, 0], and blue corresponds to [0, 0, 1]. Among them, a real-valued vector refers to a vector whose element values ​​are real numbers. Unlike the binary vector obtained by one-hot encoding, the elements of a real-valued vector can be any real number, expressing more flexible and detailed numerical information. In knowledge graph embedding, discrete elements are mapped to a low-dimensional real-valued vector space through embedding learning, which can characterize the semantic similarity and association strength between elements. Therefore, embedding knowledge graph elements into real-valued vectors can significantly improve the ability of semantic representation and calculation, and support more intelligent and flexible knowledge association and reasoning.

[0151] Step 30212: Determine the embedding vector of the device instance attribute and the device image triplet according to the first semantic association matrix.

[0152] Optionally, in this embodiment, according to the obtained first semantic association matrix, the device instance attributes are converted into real-valued vectors through a node2vec algorithm, and the device image is subjected to a convolutional neural network to extract high-level semantic features to obtain a device image vector. By maximizing the cosine similarity between the real-valued vector of the device instance attribute and the device image vector, that is, by maximizing the loss function of the cosine similarity, the embedding vector of the device instance attribute and the device image triple is obtained.

[0153] Step 30213: Determine the embedding vector of the device instance attribute and component image triplet according to the second semantic association matrix.

[0154] Optionally, in this embodiment, according to the obtained second semantic association matrix, the device instance attributes are converted into real-valued vectors through the node vector node2vec algorithm, and the component image is extracted with high-level semantic features through a convolutional neural network to obtain a component image vector. By maximizing the cosine similarity between the real-valued vector of the device instance attribute and the component image vector, the embedding vector of the device instance attribute and the component image triple is obtained.

[0155] Among them, Node2vec is a graph embedding algorithm based on random walks, which is used to learn low-dimensional vector representations of nodes in the graph. The basic idea is to perform biased random walks on the graph to generate node sequences, and regard the sequences as the context information of the nodes. The Skip-gram model is used to learn the embedding vectors of the nodes so that nodes with high co-occurrence frequencies in the sequence have similar embedding vectors. The random walk strategy of Node2vec is controlled by two parameters: return parameter p: controls the probability of the walk returning to the previous node. The larger the p value, the more the walk tends to explore the local neighborhood, and the generated sequences are concentrated around the node. In and out parameter q: controls the switching probability between visited nodes. The larger the q value, the more the walk tends to explore new areas, and the generated sequences spread in the graph. By adjusting p and q, node2vec can balance between breadth-first (BFS) and depth-first (DFS), generate different node sequences, and capture the structural role and homogeneity of nodes in the graph.

[0156] Step 30214: Use the embedding vector of the device instance attribute and physical model parameter triplet, the embedding vector of the device instance attribute and mechanism model parameter triplet, the embedding vector of the device instance attribute and device image triplet, and the embedding vector of the device instance attribute and component image triplet as real-valued weights to determine the target weighted triplet.

[0157] Step 30215: Add the target weighted triple to the instantiated knowledge graph to obtain the associated knowledge graph.

[0158] In this embodiment, the equipment monitoring module uses the embedding vectors of the four triples obtained as real-valued weights to form weighted triples according to the subject, predicate and object. The specific process is: the embedding vectors of the four triples learned are used as the real-valued weights of the triples, and added to the original semantic relationship triples to form the target weighted triples. Then the weighted triples are added as semantic links to the instantiated knowledge graph to obtain the associated knowledge graph.

[0159] For example, the weight of “Distance Relay_1-hasPhysicalModel-Cabinet_1” is 0.85, and the weight of “Distance Relay_1is Shown In Image_1” is 0.92. After weighted fusion, the device instance attributes are semantically associated with the multimodal data. Continuing with the example of 110kV distance protection, the fragment of the fused knowledge graph (i.e. the associated knowledge graph) is as follows: Distance Relay_1rdf:type:Distance Relay;has Type“Impedance Protection”;has Manufacturer“NR Electric”;has Model“PCS-931”;has Grade“110”^^xsd:int;has Set point“Z1=0.2,tZ1=0.1”;connects To:Switch gear_1;has Measurement:CT_1,CT_2,PT_1;has State“Normal”;has Physical Model:Cabinet_1,Panel_1,Slot_1;hasPrincipal Model:Analog Input_1,Di stance Measure_1,Direction Judge_1;is ShownIn:Image_1,Image_2.:Image_1rdf:type:Image;has URL“http: / / www.example.com / images / Distance Relay_1.jpg”;has Feature "[0.12, 0.35, ..., 0.82]"^^xsd: float; hasPart: Panel_1, Slot_1.

[0160] In this embodiment, embedding learning is used to map different modal data such as device instance attributes, physical model parameters, mechanism model parameters, device images, and component images to the same low-dimensional semantic space to obtain their embedded vector representations. By minimizing the Euclidean distance or maximizing the cosine similarity to obtain the distance or similarity between different modal embedding vectors, the semantic correspondence between them can be learned to achieve cross-modal semantic association. As a continuous, low-dimensional, dense semantic representation, the embedded vector can effectively alleviate the symbolization and sparseness problems of the original knowledge graph and improve the effect of semantic calculation and association.

[0161] The fault detection method provided in this embodiment may further include step 201 and steps 203 to 205, which are not described in detail here.

[0162] The fault detection method provided in this embodiment determines the instantiation knowledge graph of the high-voltage relay protection device based on the static data of the high-voltage relay protection device and the current sampling data sequence, and determines the associated knowledge graph of the high-voltage relay protection device based on the instantiation knowledge graph, the preset device model library and the image library, thereby realizing the construction of an instantiation device knowledge graph based on the static data of the device and the dynamic current sampling data sequence, and then semantically associating the physical model, mechanism model and image data of the high-voltage relay protection device with the instantiation knowledge graph to form a unified semantic representation that integrates multi-source heterogeneous information. As a result, the accuracy of the associated knowledge graph is higher, further improving the accuracy of fault detection.

[0163] Figure 4 is a flowchart of another fault detection method provided by an embodiment of the present invention. Figure 2 Based on the illustrated embodiment and various optional implementations, the steps before determining the characteristic index are described in detail. For ease of description, this embodiment only shows the steps of how to obtain the current sampling data sequence. The electronic device used by the method provided in this embodiment can be Figure 1 The equipment monitoring module or data center in the operation and maintenance system. In this embodiment, the electronic device is used as a data center for exemplary description. Figure 4 As shown, in the fault detection method provided in this embodiment, before determining the characteristic index according to the current sampling data sequence of the high-voltage relay protection device, it also includes the following steps 401 and 406.

[0164] Step 401: Obtain the original current sampling data sequence.

[0165] In a possible implementation, the data center decompresses the received sampled data to obtain an original current sampled data sequence.

[0166] The data center receives a batch of compressed sampling data packets from the high-voltage relay protection device. For example, the size of each data packet can be 1KB. The data center first decompresses the data packets. If the high-voltage relay protection device uses Huffman coding and wavelet transform algorithms to compress the sampling data, the data center uses the same Huffman decoding and wavelet reconstruction algorithms as the compression end to restore the data packets to the original current sampling data sequence.

[0167] Exemplarily, the original current sampling data sequence may include 96 sampling points, that is, the original current sampling data sequence includes 96 sampling points, each point includes 7 measurement quantities. The data type is a 32-bit floating point number. Of course, the number of sampling points in the original current sampling data sequence may also be other numbers, which is not limited in this embodiment.

[0168] In another possible implementation, the data center obtains the original current sampling data at the current sampling moment, obtains the original historical sampling data corresponding to the current sampling moment from the historical data cache according to a preset time window length, and obtains the original current sampling data sequence.

[0169] When there are multiple measurement quantities of the sampling data, the following steps 402 to 406 are performed for each measurement quantity.

[0170] Step 402: Determine the characteristic statistics of the original current sampling data sequence.

[0171] Optionally, the feature statistics include at least one of the following: mean, variance, kurtosis and deviation.

[0172] Among them, the mean reflects the central tendency of the data. The variance reflects the degree of dispersion of the data. The skewness reflects the symmetry of the data distribution. Kurtosis is a statistic that reflects the sharpness or flatness of the probability distribution curve. It represents the ratio of the fourth-order moment of the random variable to the square of the variance, and describes the peak height of the distribution curve at the mean value. In this embodiment, the data center determines whether the data distribution is close to the normal distribution by calculating the kurtosis value of the current sampled data sequence. If the kurtosis value is within the preset threshold range (such as between -1 and 1), it means that the sharpness of the data is acceptable, and Gaussian filtering and other filtering algorithms based on the normal assumption can be used for noise removal. If the kurtosis exceeds the threshold, it is necessary to consider other filtering algorithms suitable for sharp or flat distributions, such as median filtering, bilateral filtering, etc. for noise removal.

[0173] Optionally, the characteristic statistics of the original current sampling data sequence can be determined based on the functions in the programming language computing library. For example, the mean is determined based on the mean(X) function, the variance is determined based on the var(X) function, the skewness is determined based on the skew(X) function, and the kurtness is determined based on the kurt(X) function. Where X represents the original current sampling data sequence.

[0174] According to the preset threshold and the calculated characteristic statistics, it is determined whether the original current sampling data sequence meets the requirements. The at least one characteristic statistic calculated is compared with its corresponding normal range threshold. If they all meet their corresponding normal range thresholds, it is determined to perform the following step 403. If any characteristic statistic does not meet its corresponding normal range threshold, the data cleaning process is triggered and resampling is performed.

[0175] Step 403: When it is determined that the original current sampling data sequence meets the requirements according to the preset feature statistic threshold and the feature statistic, the mean estimate and variance estimate of the original current sampling data sequence are determined.

[0176] Optionally, when it is determined that the original current sampling data sequence meets the requirements, in order to improve the accuracy, an exponentially weighted moving average algorithm (Exponentially Weighted Moving Average, EWMA for short) may be used to determine a mean estimate and a variance estimate of the original current sampling data sequence.

[0177] Among them, the exponentially weighted moving average algorithm is an online learning algorithm for estimating the mean and variance of time series data. It gives higher weights to the most recent observations by weighted averaging, while the weights of old observations decay exponentially over time, thereby adapting to the dynamic trend of data changes. In this embodiment, the data center uses the EWMA algorithm to estimate the mean estimate and variance estimate of the original current sampled data sequence as parameters for constructing the Kalman filter. This adaptive estimation method can track the statistical characteristics of the data and improve the robustness of the filter. At the same time, the EWMA algorithm only needs to store the estimated value of the previous moment, and the computational overhead is very small, which is suitable for processing massive real-time sampled data streams.

[0178] Optionally, the mean estimate is given by: t =α×x(t)+(1-α)×μ t-1 Among them, μ t Represents the mean estimate at the current sampling time t, μt -1 Represents the mean estimate of the previous sampling time t-1. α represents the forgetting factor, ranging from 0 to 1, which is used to control the decay weight of historical sampling data. x(t) represents a measurement at the current sampling time t.

[0179] Optionally, the variance estimate is calculated as: in, represents the variance estimate at the current sampling time t, Represents the variance estimate at the previous sampling time t-1.

[0180] Among them, the traditional EWMA algorithm uses a fixed smoothing coefficient α, whose value needs to be pre-set according to the characteristics of the data. Here, the forgetting factor α can be adaptively adjusted according to the dynamic changes of the data, so that the algorithm can better track the time-varying trend of the data. The value of the forgetting factor α is negatively correlated with the length of the time window. The larger the α, the shorter the time window, and the faster the algorithm forgets; the smaller the α, the longer the time window, and the longer the algorithm remembers the historical data. By adjusting α, the degree of dependence of the algorithm on the data history can be flexibly controlled to adapt to different time scales and data characteristics. For example, for data with strong periodicity, you can choose an α value that matches the period to capture the long-term trend of the data; for non-stationary data, you can choose a smaller α value to reduce the impact of past observations.

[0181] Step 404: construct a Kalman filter model according to the mean estimate and the variance estimate.

[0182] The calculated mean estimate and variance estimate are used to construct a Kalman filter model.

[0183] The Kalman filter model contains state equations and observation equations, which are used to describe the dynamic change law and observation relationship of the sampled data. Its implementation principle is as follows. The original sampled data at the current sampling moment is taken as input, and the constructed Kalman filter model is used to recursively estimate the optimal estimate of the original sampled data at the current sampling moment. Using the obtained optimal estimate of the sampled data, the upper and lower limits of the filter threshold at the current sampling moment are calculated: the upper threshold = the optimal estimate of the sampled data + k* variance estimate, the lower threshold = the optimal estimate of the sampled data - k* variance estimate, where k is the preset threshold coefficient. Determine whether the original sampled data at the current sampling moment exceeds the range of the upper and lower thresholds. If it exceeds, the corresponding data is eliminated. The filtered sampled data is stored in the historical data cache, and the current sampling moment is updated. Repeat the above steps to process the sampled data at the next sampling moment until the filtering of the sampled data at all sampling moments is completed.

[0184] The Kalman filter is a Bayesian filter used to estimate the state of a linear Gaussian dynamic system. It recursively estimates the optimal value of the system state variable and its uncertainty through the system's state equation and observation equation. Let x t is the system state vector at time t, z t is the observation vector at time t (i.e. x(t) in the above description), then the Kalman filter model can be expressed as: State equation: x t =A×x t-1 +B×u t +w t , observation equation: z t =H×x t +v t . Where A is the state transfer matrix, which describes the dynamic relationship between state variables. B is the control input matrix. t is the control input of the original sampling data at the known current sampling time. H is the observation matrix, which describes the relationship between the state variables and the observation values. t and v t are process noise and observation noise respectively. They are assumed to be independent Gaussian white noises with mean 0 and covariance matrices Q and R.

[0185] In this embodiment, the data center uses the mean and variance estimated by EWMA to construct a simple Kalman filter model. The state variable contains only one scalar xt , represents the true level of the sampled value; the observed value z t is the original sampling value; the state transfer matrix A and the observation matrix H are both 1; the process noise w t The variance of is the variance of the EWMA estimate, and the observation noise v t The variance of is empirically set to half of the variance estimate. Through Kalman filtering, the data center obtains the optimal estimate of each sample value, and then uses this estimate as a benchmark to determine whether the original sample value is abnormal based on the variance estimate and the preset threshold. This estimation-based anomaly detection method is more robust than simple threshold judgment and can adapt to the fluctuation trend of data. At the same time, the Kalman filter uses the information of historical observations to smooth short-term disturbances of the sample values, achieves smoothing and denoising, and makes the filtering results more stable and reliable.

[0186] Step 405: Input the original current sampling data sequence into the Kalman filter model to obtain the filtered original current sampling data sequence.

[0187] After the Kalman filter model is constructed, the original current sampling data sequence is input into the Kalman filter model to obtain the filtered original current sampling data sequence.

[0188] Step 406: normalize the filtered original current sampling data sequence to obtain a current sampling data sequence.

[0189] Optionally, in this embodiment, a minimum-maximum normalization algorithm may be used to normalize the filtered original current sampling data sequence to obtain the current sampling data sequence.

[0190] Furthermore, the data center stores the current sampling data sequence in a time series database in a time series manner, so as to facilitate the subsequent distribution to the device monitoring module in the form of a real-time data stream through the message queue middleware.

[0191] The following is a specific example to illustrate the above steps 401 to 406. The data center reads the original sampled data x(t) at the current time t and the historical data {x t ,x t-1 ,......,x t-10}, forming an original current sampling data sequence with a length of 11 (called a sliding window).

[0192] Assuming that the original current sampling data sequence meets the requirements, apply the exponentially weighted moving average algorithm to the original current sampling data sequence to calculate the mean estimate μ at the current moment t and variance estimation Setting the forgetting factor α = 0.2 means that the weight of historical data decays by 20% every period. Then the recursive formula for the mean estimate is: t =α×x(t)+(1-α)×μ t-1 =0.2×x(t)+0.8×μ t-1 , the recursive formula for the variance estimate is:

[0193]

[0194] Estimate μ using the mean t and variance estimation Construct a Kalman filter model. The state equation assumes that the sampled data follows a Gaussian distribution and the state variable is the true value x t , the observed variable is the observed value z(t) under noise interference, then: x(t) = x t-1 +w(t), w(t):N(0,Q), z(t)=x(t)+v(t), v(t):N(0,R), where w(t) and v(t) are process noise and observation noise, respectively, and obey a Gaussian distribution with mean 0. The covariance matrices Q and R are determined by trial and error.

[0195] Apply the Kalman filter, and use the optimal estimate x at time t-1 t-1 and the estimated variance P t-1 , predict the prior estimate x(t) and prior variance P(t) at time t, then combine the observed value z(t) at time t to calculate the Kalman gain K(t) and update the optimal estimate x at time t * (t) and estimated variance P(t): x * (t)=x(t)+K(t)×[z(t)-x(t)], P(t)=[1-K(t)]×P t-1 . By the optimal estimate x * (t) and variance estimates Calculate the filter threshold at time t: TH upper =x * (t)+3×σ(t),TH lower =x * (t)-3×σ(t). σ(t) represents the filter variance at time t. If the original sampled data x(t) exceeds [TH lower , TH upper ], it is considered an outlier and the optimal estimate x * (t) Replace; otherwise, keep the original value unchanged.

[0196] The filtered sampled data x'(t) is added to the sliding window, the database is updated, and the process is repeated until all data is filtered. After Kalman filtering, the abnormal data is repaired, the data curve becomes smooth and continuous, and the dynamic change trend of the data can be tracked, reflecting the adaptive characteristics.

[0197] Finally, the data center performs minimum-maximum normalization on the filtered original current sampling data sequence. i '(t) represents the sampling value of the i-th measurement at time t, x i ' min and x i ' max They represent the minimum and maximum values ​​of the i-th measurement in the entire monitoring period or in the original current sampling data sequence after filtering. The normalization formula is: After normalization, the numerical range of all measured quantities is scaled to the interval [0, 1], eliminating the differences in dimension and numerical scale, which is convenient for subsequent feature analysis and modeling. The normalized sampling data is stored in the time series database in the format of (timestamp, measurement, value) in a time series manner, providing a unified data interface service for the equipment monitoring module.

[0198] The fault detection method provided in this embodiment further includes steps 201 to 205 or steps 301 and 302, which will not be described in detail here.

[0199] The fault detection method provided in this embodiment realizes the processing of the original current sampling data sequence to obtain the current sampling data sequence. The processing process includes data quality judgment, filtering, normalization, etc., which can improve the availability and consistency of data and lay a good data foundation for subsequent data analysis and abnormal diagnosis, thereby further improving the efficiency and reliability of fault detection.

[0200] Figure 5 1 is a schematic diagram of the structure of an electronic device that implements the fault detection method of an embodiment of the present invention. The electronic device is intended to represent various forms of digital computers, such as laptop computers, desktop computers, workstations, personal digital assistants, servers, blade servers, mainframe computers, and other suitable computers. The components shown herein, their connections and relationships, and their functions are merely examples and are not intended to limit the implementation of the present invention described and / or claimed herein.

[0201] like Figure 5As shown, the electronic device 10 includes at least one processor 11, and a memory connected to the at least one processor 11, such as a read-only memory (ROM) 12, a random access memory (RAM) 13, etc., wherein the memory stores a computer program that can be executed by at least one processor, and the processor 11 can perform various appropriate actions and processes according to the computer program stored in the read-only memory (ROM) 12 or the computer program loaded from the storage unit 18 to the random access memory (RAM) 13. In the RAM 13, various programs and data required for the operation of the electronic device 10 can also be stored. The processor 11, the ROM 12, and the RAM 13 are connected to each other through a bus 14. An input / output (I / O) interface 15 is also connected to the bus 14.

[0202] A number of components in the electronic device 10 are connected to the I / O interface 15, including: an input unit 16, such as a keyboard, a mouse, etc.; an output unit 17, such as various types of displays, speakers, etc.; a storage unit 18, such as a disk, an optical disk, etc.; and a communication unit 19, such as a network card, a modem, a wireless communication transceiver, etc. The communication unit 19 allows the electronic device 10 to exchange information / data with other devices through a computer network such as the Internet and / or various telecommunication networks.

[0203] The processor 11 may be a variety of general and / or special processing components with processing and computing capabilities. Some examples of the processor 11 include, but are not limited to, a central processing unit (CPU), a graphics processing unit (GPU), various special artificial intelligence (AI) computing chips, various processors running machine learning model algorithms, a digital signal processor (DSP), and any appropriate processor, controller, microcontroller, etc. The processor 11 executes the various methods and processes described above, such as a fault detection method.

[0204] In some embodiments, the fault detection method may be implemented as a computer program, which is tangibly contained in a computer-readable storage medium, such as a storage unit 18. In some embodiments, part or all of the computer program may be loaded and / or installed on the electronic device 10 via the ROM 12 and / or the communication unit 19. When the computer program is loaded into the RAM 13 and executed by the processor 11, one or more steps of the fault detection method described above may be performed. Alternatively, in other embodiments, the processor 11 may be configured to perform the fault detection method in any other appropriate manner (e.g., by means of firmware).

[0205] Various implementations of the systems and techniques described above herein can be implemented in digital electronic circuit systems, integrated circuit systems, field programmable gate arrays (FPGAs), application specific integrated circuits (ASICs), application specific standard products (ASSPs), systems on chips (SOCs), load programmable logic devices (CPLDs), computer hardware, firmware, software, and / or combinations thereof. These various implementations can include: being implemented in one or more computer programs that can be executed and / or interpreted on a programmable system including at least one programmable processor, which can be a special purpose or general purpose programmable processor that can receive data and instructions from a storage system, at least one input device, and at least one output device, and transmit data and instructions to the storage system, the at least one input device, and the at least one output device.

[0206] Computer programs for implementing the methods of the present invention may be written in any combination of one or more programming languages. These computer programs may be provided to a processor of a general-purpose computer, a special-purpose computer, or other device, so that when the computer program is executed by the processor, the functions / operations specified in the flow chart and / or block diagram are implemented. The computer program may be executed entirely on the machine, partially on the machine, partially on the machine and partially on a remote machine as a stand-alone software package, or entirely on a remote machine or server.

[0207] In the context of the present invention, a computer-readable storage medium may be a tangible medium that may contain or store a computer program for use by or in combination with an instruction execution system, device or equipment. A computer-readable storage medium may include, but is not limited to, electronic, magnetic, optical, electromagnetic, infrared, or semiconductor systems, devices or equipment, or any suitable combination of the foregoing. Alternatively, a computer-readable storage medium may be a machine-readable signal medium. A more specific example of a machine-readable storage medium may include an electrical connection based on one or more lines, a portable computer disk, a hard disk, a random access memory (RAM), a read-only memory (ROM), an erasable programmable read-only memory (EPROM or flash memory), an optical fiber, a portable compact disk read-only memory (CD-ROM), an optical storage device, a magnetic storage device, or any suitable combination of the foregoing.

[0208] To provide interaction with a user, the systems and techniques described herein may be implemented on an electronic device having: a display device (e.g., a CRT (cathode ray tube) or LCD (liquid crystal display) monitor) for displaying information to the user; and a keyboard and a pointing device (e.g., a mouse or trackball) through which the user can provide input to the electronic device. Other types of devices may also be used to provide interaction with the user; for example, the feedback provided to the user may be any form of sensory feedback (e.g., visual feedback, auditory feedback, or tactile feedback); and input from the user may be received in any form (including acoustic input, voice input, or tactile input).

[0209] The systems and techniques described herein may be implemented in a computing system that includes backend components (e.g., as a data server), or a computing system that includes middleware components (e.g., an application server), or a computing system that includes frontend components (e.g., a user computer with a graphical user interface or a web browser through which a user can interact with implementations of the systems and techniques described herein), or a computing system that includes any combination of such backend components, middleware components, or frontend components. The components of the system may be interconnected by any form or medium of digital data communication (e.g., a communication network). Examples of communication networks include: a local area network (LAN), a wide area network (WAN), a blockchain network, and the Internet.

[0210] A computing system may include a client and a server. The client and the server are generally remote from each other and usually interact through a communication network. The client and server relationship is generated by computer programs running on the corresponding computers and having a client-server relationship with each other. The server may be a cloud server, also known as a cloud computing server or cloud host, which is a host product in the cloud computing service system to solve the defects of difficult management and weak business scalability in traditional physical hosts and VPS services.

[0211] An embodiment of the present invention further provides a computer program product, including a computer program, which, when executed by a processor, implements the fault detection method provided by any embodiment of the present invention.

[0212] In the process of implementation, the computer program product can be written in one or more programming languages ​​or a combination thereof to perform the computer program code of the present invention, including object-oriented programming languages, such as Java, Smalltalk, C++, and conventional procedural programming languages, such as "C" language or similar programming languages. The program code can be executed entirely on the user's computer, partially on the user's computer, as a separate software package, partially on the user's computer and partially on a remote computer, or entirely on a remote computer or server. In the case of a remote computer, the remote computer can be connected to the user's computer through any type of network, including a local area network (LAN) or a wide area network (WAN), or can be connected to an external computer (for example, using an Internet service provider to connect through the Internet).

[0213] It should be understood that the various forms of processes shown above can be used to reorder, add or delete steps. For example, the steps described in the present invention can be executed in parallel, sequentially or in different orders, as long as the desired results of the technical solution of the present invention can be achieved, and this document does not limit this.

[0214] The above specific implementations do not constitute a limitation on the protection scope of the present invention. It should be understood by those skilled in the art that various modifications, combinations, sub-combinations and substitutions can be made according to design requirements and other factors. Any modification, equivalent substitution and improvement made within the spirit and principle of the present invention should be included in the protection scope of the present invention.

Claims

1. A fault detection method, characterized in that: include: Determine characteristic indicators according to the current sampling data sequence of the high-voltage relay protection device; Determine the associated knowledge graph of the high-voltage relay protection device according to the static data of the high-voltage relay protection device, the current sampling data sequence, and a preset device model library and image library; Determine an intermediate diagnosis conclusion according to the characteristic indicator and the diagnosis rule corresponding to the characteristic indicator in the associated knowledge graph; Determine the current classification result of the feature indicator according to the distance between the feature indicator and each classification result; wherein each classification result is a result obtained by clustering all states in the associated knowledge graph, and each classification result is used to characterize a working state of the high-voltage relay protection device; The fault detection result of the high-voltage relay protection device is determined according to the intermediate diagnosis conclusion and the current classification result.

2. The method according to claim 1, characterized in that After determining the fault detection result of the high-voltage relay protection device according to the intermediate diagnosis conclusion and the current classification result, the method further includes: According to a preset sharing strategy, the shared data to be uploaded to the blockchain network is determined; wherein the shared data includes at least one of the following: a target sampling data sequence to be uploaded to the blockchain network in the current sampling data sequence and / or a target detection result to be uploaded to the blockchain network in the fault detection result; Encrypting the shared data using an asymmetric encryption algorithm to obtain encrypted shared data; The encrypted shared data is uploaded to the blockchain network.

3. The method according to claim 1, characterized in that Determining the associated knowledge graph of the high-voltage relay protection device according to the static data of the high-voltage relay protection device, the current sampling data sequence, and a preset device model library and image library includes: Determine an instantiated knowledge graph of the high-voltage relay protection device according to the static data of the high-voltage relay protection device and the current sampling data sequence; According to the instantiated knowledge graph, the preset equipment model library and the image library, the associated knowledge graph of the high-voltage relay protection device is determined.

4. The method according to claim 3, characterized in that The step of determining the instantiated knowledge graph of the high-voltage relay protection device according to the static data of the high-voltage relay protection device and the current sampling data sequence includes: Converting the static data and the current sampled data sequence into a device instance; The device instance is mapped to a pre-built device knowledge graph according to the ontology model to obtain the instantiated knowledge graph.

5. The method according to claim 4, characterized in that The determining of the associated knowledge graph of the high-voltage relay protection device according to the instantiated knowledge graph, the preset device model library and the image library includes: Acquire a physical model and a mechanism model corresponding to the device instance from the device model library; Acquire a device image and a component image corresponding to the device instance from the image library; Matching the device instance attributes in the instantiated knowledge graph with the physical model parameters in the physical model to obtain a first semantic correspondence between the device instance attributes and the physical model parameters; Matching the device instance attributes in the instantiated knowledge graph with the mechanism model parameters in the mechanism model to obtain a second semantic correspondence between the device instance attributes and the mechanism model parameters; Mapping the device instance attributes in the instantiated knowledge graph to a low-dimensional semantic space through word embedding to obtain a distributed vector representation of the device instance attributes; Performing feature extraction and semantic encoding on the device image, mapping the encoded device image to a semantic space identical to the device instance attribute, and obtaining a distributed vector representation of the device image; Performing feature extraction and semantic encoding on the component image, mapping the encoded component image to a semantic space identical to the device instance attribute, and obtaining a distributed vector representation of the component image; Determining a first semantic association matrix between the distributed vector representation of the device instance attributes and the distributed vector representation of the device image; determining a second semantic association matrix between the distributed vector representation of the device instance attributes and the distributed vector representation of the component image; Determine, according to the first semantic correspondence, an embedding vector of a triplet of the device instance attribute and the physical model parameter; Determining, according to the second semantic correspondence, an embedding vector of a triplet of the device instance attribute and the mechanism model parameter; Determining, according to the first semantic association matrix, an embedding vector of a triplet of the device instance attribute and the device image; Determining an embedding vector of a triplet of the device instance attribute and the component image according to the second semantic association matrix; Determine a target weighted triplet by using the embedding vector of the device instance attribute and physical model parameter triplet, the embedding vector of the device instance attribute and mechanism model parameter triplet, the embedding vector of the device instance attribute and device image triplet, and the embedding vector of the device instance attribute and component image triplet as real-valued weights; The target weighted triple is added to the instantiated knowledge graph to obtain the associated knowledge graph.

6. The method according to claim 1, characterized in that Determining an intermediate diagnosis conclusion according to the characteristic indicator and the diagnosis rule corresponding to the characteristic indicator in the associated knowledge graph includes: By using a unified algorithm, according to the characteristic index, the variable in the diagnosis rule corresponding to the characteristic index is matched and replaced to obtain the diagnosis rule after matching and replacement; The intermediate diagnosis conclusion is determined according to the matched replaced diagnosis rule.

7. The method according to claim 1, characterized in that Before determining the current classification result of the feature indicator according to the distance between the feature indicator and each classification result, the method further includes: Extracting all state-related weighted triples from the associated knowledge graph; Inputting each of the weighted triples into a conversion model to obtain an embedding vector of the weighted triplet output by the conversion model; Clustering is performed on the embedded vectors of the weighted triplets related to all the states to obtain various classification results; wherein the various classification results are used to characterize at least two of the following working states: normal state, overload state, underload state, abnormal state and fault state.

8. The method according to claim 1, characterized in that Determining the current classification result of the feature indicator according to the distance between the feature indicator and each classification result includes: Determining an embedding vector of the feature indicator; Calculating the Mahalanobis distance between the embedding vector of the feature index and the center of each classification result respectively; According to each Mahalanobis distance, the current classification result of the feature index is determined.

9. The method according to claim 1, characterized in that: Before determining the characteristic index according to the current sampling data sequence of the high-voltage relay protection device, the method further includes: Get the original current sampling data sequence; Determining characteristic statistics of the original current sampled data sequence; When it is determined that the original current sampled data sequence meets the requirements according to the preset feature statistic threshold and the feature statistic, a mean estimate and a variance estimate of the original current sampled data sequence are determined; Constructing a Kalman filter model according to the mean estimate and the variance estimate; Inputting the original current sampling data sequence into the Kalman filter model to obtain a filtered original current sampling data sequence; The filtered original current sampling data sequence is normalized to obtain the current sampling data sequence.

10. An electronic device, characterized in that: The electronic device comprises: at least one processor; and a memory communicatively connected to the at least one processor; wherein, The memory stores a computer program executable by the at least one processor, and the computer program is executed by the at least one processor so that the at least one processor can perform the fault detection method according to any one of claims 1 to 9.

Citation Information

Patent Citations

  • Equipment fault diagnosis and maintenance knowledge recommendation system based on knowledge graph

    CN114579875A

  • Relay protection hidden fault identification reasoning method based on knowledge graph

    CN117521804A

  • Systems and methods for ai continued learning in electrical power grid fault analysis

    US20230036218A1

Cited By

  • Dangerous chemical identification method and device

    CN119474832A

  • Intelligent substation secondary equipment fault positioning method based on improved Transform

    CN119807982A

  • Relay protection configuration information intelligent identification system and method based on natural language technology

    CN119849475A

  • Large-scale circuit partitioning method and device based on space filling curve

    CN119862844A

  • Large-scale circuit partitioning method and device based on space-filling curve

    CN119862844B